diff --git a/Makefile b/Makefile index 71a1d68d..535225cb 100644 --- a/Makefile +++ b/Makefile @@ -2005,6 +2005,7 @@ Q3OBJ += \ $(B)/client/opus/lin2log.o \ $(B)/client/opus/log2lin.o \ $(B)/client/opus/LPC_analysis_filter.o \ + $(B)/client/opus/LPC_fit.o \ $(B)/client/opus/LPC_inv_pred_gain.o \ $(B)/client/opus/table_LSF_cos.o \ $(B)/client/opus/NLSF2A.o \ @@ -2038,11 +2039,9 @@ Q3OBJ += \ $(B)/client/opus/LTP_analysis_filter_FLP.o \ $(B)/client/opus/LTP_scale_ctrl_FLP.o \ $(B)/client/opus/noise_shape_analysis_FLP.o \ - $(B)/client/opus/prefilter_FLP.o \ $(B)/client/opus/process_gains_FLP.o \ $(B)/client/opus/regularize_correlations_FLP.o \ $(B)/client/opus/residual_energy_FLP.o \ - $(B)/client/opus/solve_LS_FLP.o \ $(B)/client/opus/warped_autocorrelation_FLP.o \ $(B)/client/opus/wrappers_FLP.o \ $(B)/client/opus/autocorrelation_FLP.o \ @@ -2051,7 +2050,6 @@ Q3OBJ += \ $(B)/client/opus/energy_FLP.o \ $(B)/client/opus/inner_product_FLP.o \ $(B)/client/opus/k2a_FLP.o \ - $(B)/client/opus/levinsondurbin_FLP.o \ $(B)/client/opus/LPC_inv_pred_gain_FLP.o \ $(B)/client/opus/pitch_analysis_core_FLP.o \ $(B)/client/opus/scale_copy_vector_FLP.o \ diff --git a/code/opus-1.1.4/celt/_kiss_fft_guts.h b/code/opus-1.1.4/celt/_kiss_fft_guts.h index 5e3d58fd..17392b3e 100644 --- a/code/opus-1.1.4/celt/_kiss_fft_guts.h +++ b/code/opus-1.1.4/celt/_kiss_fft_guts.h @@ -58,12 +58,12 @@ # define S_MUL(a,b) MULT16_32_Q15(b, a) # define C_MUL(m,a,b) \ - do{ (m).r = SUB32(S_MUL((a).r,(b).r) , S_MUL((a).i,(b).i)); \ - (m).i = ADD32(S_MUL((a).r,(b).i) , S_MUL((a).i,(b).r)); }while(0) + do{ (m).r = SUB32_ovflw(S_MUL((a).r,(b).r) , S_MUL((a).i,(b).i)); \ + (m).i = ADD32_ovflw(S_MUL((a).r,(b).i) , S_MUL((a).i,(b).r)); }while(0) # define C_MULC(m,a,b) \ - do{ (m).r = ADD32(S_MUL((a).r,(b).r) , S_MUL((a).i,(b).i)); \ - (m).i = SUB32(S_MUL((a).i,(b).r) , S_MUL((a).r,(b).i)); }while(0) + do{ (m).r = ADD32_ovflw(S_MUL((a).r,(b).r) , S_MUL((a).i,(b).i)); \ + (m).i = SUB32_ovflw(S_MUL((a).i,(b).r) , S_MUL((a).r,(b).i)); }while(0) # define C_MULBYSCALAR( c, s ) \ do{ (c).r = S_MUL( (c).r , s ) ;\ @@ -77,17 +77,17 @@ DIVSCALAR( (c).i , div); }while (0) #define C_ADD( res, a,b)\ - do {(res).r=ADD32((a).r,(b).r); (res).i=ADD32((a).i,(b).i); \ + do {(res).r=ADD32_ovflw((a).r,(b).r); (res).i=ADD32_ovflw((a).i,(b).i); \ }while(0) #define C_SUB( res, a,b)\ - do {(res).r=SUB32((a).r,(b).r); (res).i=SUB32((a).i,(b).i); \ + do {(res).r=SUB32_ovflw((a).r,(b).r); (res).i=SUB32_ovflw((a).i,(b).i); \ }while(0) #define C_ADDTO( res , a)\ - do {(res).r = ADD32((res).r, (a).r); (res).i = ADD32((res).i,(a).i);\ + do {(res).r = ADD32_ovflw((res).r, (a).r); (res).i = ADD32_ovflw((res).i,(a).i);\ }while(0) #define C_SUBFROM( res , a)\ - do {(res).r = ADD32((res).r,(a).r); (res).i = SUB32((res).i,(a).i); \ + do {(res).r = ADD32_ovflw((res).r,(a).r); (res).i = SUB32_ovflw((res).i,(a).i); \ }while(0) #if defined(OPUS_ARM_INLINE_ASM) diff --git a/code/opus-1.1.4/celt/arch.h b/code/opus-1.1.4/celt/arch.h index 8ceab5fe..d1e6457c 100644 --- a/code/opus-1.1.4/celt/arch.h +++ b/code/opus-1.1.4/celt/arch.h @@ -46,6 +46,14 @@ # endif # endif +#if OPUS_GNUC_PREREQ(3, 0) +#define opus_likely(x) (__builtin_expect(!!(x), 1)) +#define opus_unlikely(x) (__builtin_expect(!!(x), 0)) +#else +#define opus_likely(x) (!!(x)) +#define opus_unlikely(x) (!!(x)) +#endif + #define CELT_SIG_SCALE 32768.f #define celt_fatal(str) _celt_fatal(str, __FILE__, __LINE__); @@ -93,6 +101,7 @@ static OPUS_INLINE void _celt_fatal(const char *str, const char *file, int line) typedef opus_int16 opus_val16; typedef opus_int32 opus_val32; +typedef opus_int64 opus_val64; typedef opus_val32 celt_sig; typedef opus_val16 celt_norm; @@ -101,6 +110,9 @@ typedef opus_val32 celt_ener; #define Q15ONE 32767 #define SIG_SHIFT 12 +/* Safe saturation value for 32-bit signals. Should be less than + 2^31*(1-0.85) to avoid blowing up on DC at deemphasis.*/ +#define SIG_SAT (300000000) #define NORM_SCALING 16384 @@ -147,6 +159,7 @@ static OPUS_INLINE opus_int16 SAT16(opus_int32 x) { typedef float opus_val16; typedef float opus_val32; +typedef float opus_val64; typedef float celt_sig; typedef float celt_norm; @@ -186,6 +199,7 @@ static OPUS_INLINE int celt_isnan(float x) #define NEG16(x) (-(x)) #define NEG32(x) (-(x)) +#define NEG32_ovflw(x) (-(x)) #define EXTRACT16(x) (x) #define EXTEND32(x) (x) #define SHR16(a,shift) (a) @@ -202,6 +216,7 @@ static OPUS_INLINE int celt_isnan(float x) #define SATURATE16(x) (x) #define ROUND16(a,shift) (a) +#define SROUND16(a,shift) (a) #define HALF16(x) (.5f*(x)) #define HALF32(x) (.5f*(x)) @@ -209,6 +224,8 @@ static OPUS_INLINE int celt_isnan(float x) #define SUB16(a,b) ((a)-(b)) #define ADD32(a,b) ((a)+(b)) #define SUB32(a,b) ((a)-(b)) +#define ADD32_ovflw(a,b) ((a)+(b)) +#define SUB32_ovflw(a,b) ((a)-(b)) #define MULT16_16_16(a,b) ((a)*(b)) #define MULT16_16(a,b) ((opus_val32)(a)*(opus_val32)(b)) #define MAC16_16(c,a,b) ((c)+(opus_val32)(a)*(opus_val32)(b)) @@ -243,9 +260,9 @@ static OPUS_INLINE int celt_isnan(float x) #ifndef GLOBAL_STACK_SIZE #ifdef FIXED_POINT -#define GLOBAL_STACK_SIZE 100000 +#define GLOBAL_STACK_SIZE 120000 #else -#define GLOBAL_STACK_SIZE 100000 +#define GLOBAL_STACK_SIZE 120000 #endif #endif diff --git a/code/opus-1.1.4/celt/arm/arm2gnu.pl b/code/opus-1.1.4/celt/arm/arm2gnu.pl index 6c922ac8..a2895f74 100755 --- a/code/opus-1.1.4/celt/arm/arm2gnu.pl +++ b/code/opus-1.1.4/celt/arm/arm2gnu.pl @@ -164,11 +164,11 @@ while (<>) { $prefix = ""; if ($proc) { - $prefix = $prefix.sprintf("\t.type\t%s, %%function; ",$proc) unless ($apple); + $prefix = $prefix.sprintf("\t.type\t%s, %%function", $proc) unless ($apple); # Make sure we $prefix isn't empty here (for the $apple case). # We handle mangling the label here, make sure it doesn't match # the label handling below (if $prefix would be empty). - $prefix = "; "; + $prefix = $prefix."; "; push(@proc_stack, $proc); s/^[A-Za-z_\.]\w+/$symprefix$&:/; } diff --git a/code/opus-1.1.4/celt/arm/arm_celt_map.c b/code/opus-1.1.4/celt/arm/arm_celt_map.c index 4d4d069a..ca988b66 100644 --- a/code/opus-1.1.4/celt/arm/arm_celt_map.c +++ b/code/opus-1.1.4/celt/arm/arm_celt_map.c @@ -35,12 +35,29 @@ #if defined(OPUS_HAVE_RTCD) +# if defined(OPUS_ARM_MAY_HAVE_NEON_INTR) && !defined(OPUS_ARM_PRESUME_NEON_INTR) +opus_val32 (*const CELT_INNER_PROD_IMPL[OPUS_ARCHMASK+1])(const opus_val16 *x, const opus_val16 *y, int N) = { + celt_inner_prod_c, /* ARMv4 */ + celt_inner_prod_c, /* EDSP */ + celt_inner_prod_c, /* Media */ + celt_inner_prod_neon /* NEON */ +}; + +void (*const DUAL_INNER_PROD_IMPL[OPUS_ARCHMASK+1])(const opus_val16 *x, const opus_val16 *y01, const opus_val16 *y02, + int N, opus_val32 *xy1, opus_val32 *xy2) = { + dual_inner_prod_c, /* ARMv4 */ + dual_inner_prod_c, /* EDSP */ + dual_inner_prod_c, /* Media */ + dual_inner_prod_neon /* NEON */ +}; +# endif + # if defined(FIXED_POINT) # if ((defined(OPUS_ARM_MAY_HAVE_NEON) && !defined(OPUS_ARM_PRESUME_NEON)) || \ (defined(OPUS_ARM_MAY_HAVE_MEDIA) && !defined(OPUS_ARM_PRESUME_MEDIA)) || \ (defined(OPUS_ARM_MAY_HAVE_EDSP) && !defined(OPUS_ARM_PRESUME_EDSP))) opus_val32 (*const CELT_PITCH_XCORR_IMPL[OPUS_ARCHMASK+1])(const opus_val16 *, - const opus_val16 *, opus_val32 *, int , int) = { + const opus_val16 *, opus_val32 *, int, int, int) = { celt_pitch_xcorr_c, /* ARMv4 */ MAY_HAVE_EDSP(celt_pitch_xcorr), /* EDSP */ MAY_HAVE_MEDIA(celt_pitch_xcorr), /* Media */ @@ -51,7 +68,7 @@ opus_val32 (*const CELT_PITCH_XCORR_IMPL[OPUS_ARCHMASK+1])(const opus_val16 *, # else /* !FIXED_POINT */ # if defined(OPUS_ARM_MAY_HAVE_NEON_INTR) && !defined(OPUS_ARM_PRESUME_NEON_INTR) void (*const CELT_PITCH_XCORR_IMPL[OPUS_ARCHMASK+1])(const opus_val16 *, - const opus_val16 *, opus_val32 *, int, int) = { + const opus_val16 *, opus_val32 *, int, int, int) = { celt_pitch_xcorr_c, /* ARMv4 */ celt_pitch_xcorr_c, /* EDSP */ celt_pitch_xcorr_c, /* Media */ diff --git a/code/opus-1.1.4/celt/arm/celt_ne10_fft.c b/code/opus-1.1.4/celt/arm/celt_ne10_fft.c index 42d96a71..b8c60abe 100644 --- a/code/opus-1.1.4/celt/arm/celt_ne10_fft.c +++ b/code/opus-1.1.4/celt/arm/celt_ne10_fft.c @@ -36,7 +36,6 @@ #endif #endif -#include #include #include "os_support.h" #include "kiss_fft.h" diff --git a/code/opus-1.1.4/celt/arm/celt_neon_intr.c b/code/opus-1.1.4/celt/arm/celt_neon_intr.c index 47bbe3dc..cf443988 100644 --- a/code/opus-1.1.4/celt/arm/celt_neon_intr.c +++ b/code/opus-1.1.4/celt/arm/celt_neon_intr.c @@ -191,107 +191,10 @@ static void xcorr_kernel_neon_float(const float32_t *x, const float32_t *y, vst1q_f32(sum, SUMM); } -/* - * Function: xcorr_kernel_neon_float_process1 - * --------------------------------- - * Computes single correlation values and stores in *sum - */ -static void xcorr_kernel_neon_float_process1(const float32_t *x, - const float32_t *y, float32_t *sum, int len) { - float32x4_t XX[4]; - float32x4_t YY[4]; - float32x2_t XX_2; - float32x2_t YY_2; - float32x4_t SUMM; - float32x2_t SUMM_2[2]; - const float32_t *xi = x; - const float32_t *yi = y; - - SUMM = vdupq_n_f32(0); - - /* Work on 16 values per iteration */ - while (len >= 16) { - XX[0] = vld1q_f32(xi); - xi += 4; - XX[1] = vld1q_f32(xi); - xi += 4; - XX[2] = vld1q_f32(xi); - xi += 4; - XX[3] = vld1q_f32(xi); - xi += 4; - - YY[0] = vld1q_f32(yi); - yi += 4; - YY[1] = vld1q_f32(yi); - yi += 4; - YY[2] = vld1q_f32(yi); - yi += 4; - YY[3] = vld1q_f32(yi); - yi += 4; - - SUMM = vmlaq_f32(SUMM, YY[0], XX[0]); - SUMM = vmlaq_f32(SUMM, YY[1], XX[1]); - SUMM = vmlaq_f32(SUMM, YY[2], XX[2]); - SUMM = vmlaq_f32(SUMM, YY[3], XX[3]); - len -= 16; - } - - /* Work on 8 values */ - if (len >= 8) { - XX[0] = vld1q_f32(xi); - xi += 4; - XX[1] = vld1q_f32(xi); - xi += 4; - - YY[0] = vld1q_f32(yi); - yi += 4; - YY[1] = vld1q_f32(yi); - yi += 4; - - SUMM = vmlaq_f32(SUMM, YY[0], XX[0]); - SUMM = vmlaq_f32(SUMM, YY[1], XX[1]); - len -= 8; - } - - /* Work on 4 values */ - if (len >= 4) { - XX[0] = vld1q_f32(xi); - xi += 4; - YY[0] = vld1q_f32(yi); - yi += 4; - SUMM = vmlaq_f32(SUMM, YY[0], XX[0]); - len -= 4; - } - - /* Start accumulating results */ - SUMM_2[0] = vget_low_f32(SUMM); - if (len >= 2) { - /* While at it, consume 2 more values if available */ - XX_2 = vld1_f32(xi); - xi += 2; - YY_2 = vld1_f32(yi); - yi += 2; - SUMM_2[0] = vmla_f32(SUMM_2[0], YY_2, XX_2); - len -= 2; - } - SUMM_2[1] = vget_high_f32(SUMM); - SUMM_2[0] = vadd_f32(SUMM_2[0], SUMM_2[1]); - SUMM_2[0] = vpadd_f32(SUMM_2[0], SUMM_2[0]); - /* Ok, now we have result accumulated in SUMM_2[0].0 */ - - if (len > 0) { - /* Case when you have one value left */ - XX_2 = vld1_dup_f32(xi); - YY_2 = vld1_dup_f32(yi); - SUMM_2[0] = vmla_f32(SUMM_2[0], XX_2, YY_2); - } - - vst1_lane_f32(sum, SUMM_2[0], 0); -} - void celt_pitch_xcorr_float_neon(const opus_val16 *_x, const opus_val16 *_y, - opus_val32 *xcorr, int len, int max_pitch) { + opus_val32 *xcorr, int len, int max_pitch, int arch) { int i; + (void)arch; celt_assert(max_pitch > 0); celt_assert((((unsigned char *)_x-(unsigned char *)NULL)&3)==0); @@ -300,12 +203,9 @@ void celt_pitch_xcorr_float_neon(const opus_val16 *_x, const opus_val16 *_y, (float32_t *)xcorr+i, len); } - /* In case max_pitch isn't multiple of 4 - * compute single correlation value per iteration - */ + /* In case max_pitch isn't a multiple of 4, do non-unrolled version. */ for (; i < max_pitch; i++) { - xcorr_kernel_neon_float_process1((const float32_t *)_x, - (const float32_t *)_y+i, (float32_t *)xcorr+i, len); + xcorr[i] = celt_inner_prod_neon(_x, _y+i, len); } } #endif diff --git a/code/opus-1.1.4/celt/arm/celt_pitch_xcorr_arm-gnu.S b/code/opus-1.1.4/celt/arm/celt_pitch_xcorr_arm-gnu.S index 5b2ee55a..10668e54 100644 --- a/code/opus-1.1.4/celt/arm/celt_pitch_xcorr_arm-gnu.S +++ b/code/opus-1.1.4/celt/arm/celt_pitch_xcorr_arm-gnu.S @@ -44,7 +44,7 @@ .if OPUS_ARM_MAY_HAVE_NEON @ Compute sum[k]=sum(x[j]*y[j+k],j=0...len-1), k=0...3 -; xcorr_kernel_neon: @ PROC + .type xcorr_kernel_neon, %function; xcorr_kernel_neon: @ PROC xcorr_kernel_neon_start: @ input: @ r3 = int len @@ -156,8 +156,8 @@ xcorr_kernel_neon_process1: .size xcorr_kernel_neon, .-xcorr_kernel_neon @ ENDP @ opus_val32 celt_pitch_xcorr_neon(opus_val16 *_x, opus_val16 *_y, -@ opus_val32 *xcorr, int len, int max_pitch) -; celt_pitch_xcorr_neon: @ PROC +@ opus_val32 *xcorr, int len, int max_pitch, int arch) + .type celt_pitch_xcorr_neon, %function; celt_pitch_xcorr_neon: @ PROC @ input: @ r0 = opus_val16 *_x @ r1 = opus_val16 *_y @@ -171,6 +171,8 @@ xcorr_kernel_neon_process1: @ r6 = int max_pitch @ r12 = int j @ q15 = int maxcorr[4] (q15 is not used by xcorr_kernel_neon()) + @ ignored: + @ int arch STMFD sp!, {r4-r6, lr} LDR r6, [sp, #16] VMOV.S32 q15, #1 @@ -260,7 +262,7 @@ celt_pitch_xcorr_neon_done: @ This will get used on ARMv7 devices without NEON, so it has been optimized @ to take advantage of dual-issuing where possible. -; xcorr_kernel_edsp: @ PROC + .type xcorr_kernel_edsp, %function; xcorr_kernel_edsp: @ PROC xcorr_kernel_edsp_start: @ input: @ r3 = int len @@ -344,7 +346,7 @@ xcorr_kernel_edsp_done: LDMFD sp!, {r2,r4,r5,pc} .size xcorr_kernel_edsp, .-xcorr_kernel_edsp @ ENDP -; celt_pitch_xcorr_edsp: @ PROC + .type celt_pitch_xcorr_edsp, %function; celt_pitch_xcorr_edsp: @ PROC @ input: @ r0 = opus_val16 *_x (must be 32-bit aligned) @ r1 = opus_val16 *_y (only needs to be 16-bit aligned) @@ -361,6 +363,8 @@ xcorr_kernel_edsp_done: @ r9 = opus_val32 sum3 @ r1 = int max_pitch @ r12 = int j + @ ignored: + @ int arch STMFD sp!, {r4-r11, lr} MOV r5, r1 LDR r1, [sp, #36] diff --git a/code/opus-1.1.4/celt/arm/celt_pitch_xcorr_arm.s b/code/opus-1.1.4/celt/arm/celt_pitch_xcorr_arm.s index f96e0a88..6e873afc 100644 --- a/code/opus-1.1.4/celt/arm/celt_pitch_xcorr_arm.s +++ b/code/opus-1.1.4/celt/arm/celt_pitch_xcorr_arm.s @@ -153,7 +153,7 @@ xcorr_kernel_neon_process1 ENDP ; opus_val32 celt_pitch_xcorr_neon(opus_val16 *_x, opus_val16 *_y, -; opus_val32 *xcorr, int len, int max_pitch) +; opus_val32 *xcorr, int len, int max_pitch, int arch) celt_pitch_xcorr_neon PROC ; input: ; r0 = opus_val16 *_x @@ -168,6 +168,8 @@ celt_pitch_xcorr_neon PROC ; r6 = int max_pitch ; r12 = int j ; q15 = int maxcorr[4] (q15 is not used by xcorr_kernel_neon()) + ; ignored: + ; int arch STMFD sp!, {r4-r6, lr} LDR r6, [sp, #16] VMOV.S32 q15, #1 @@ -358,6 +360,8 @@ celt_pitch_xcorr_edsp PROC ; r9 = opus_val32 sum3 ; r1 = int max_pitch ; r12 = int j + ; ignored: + ; int arch STMFD sp!, {r4-r11, lr} MOV r5, r1 LDR r1, [sp, #36] diff --git a/code/opus-1.1.4/celt/arm/fft_arm.h b/code/opus-1.1.4/celt/arm/fft_arm.h index 0cb55d8e..0b78175f 100644 --- a/code/opus-1.1.4/celt/arm/fft_arm.h +++ b/code/opus-1.1.4/celt/arm/fft_arm.h @@ -34,7 +34,6 @@ #if !defined(FFT_ARM_H) #define FFT_ARM_H -#include "config.h" #include "kiss_fft.h" #if defined(HAVE_ARM_NE10) diff --git a/code/opus-1.1.4/celt/arm/fixed_armv4.h b/code/opus-1.1.4/celt/arm/fixed_armv4.h index efb3b189..d84888a7 100644 --- a/code/opus-1.1.4/celt/arm/fixed_armv4.h +++ b/code/opus-1.1.4/celt/arm/fixed_armv4.h @@ -37,7 +37,7 @@ static OPUS_INLINE opus_val32 MULT16_32_Q16_armv4(opus_val16 a, opus_val32 b) "#MULT16_32_Q16\n\t" "smull %0, %1, %2, %3\n\t" : "=&r"(rd_lo), "=&r"(rd_hi) - : "%r"(b),"r"(a<<16) + : "%r"(b),"r"(SHL32(a,16)) ); return rd_hi; } @@ -54,10 +54,10 @@ static OPUS_INLINE opus_val32 MULT16_32_Q15_armv4(opus_val16 a, opus_val32 b) "#MULT16_32_Q15\n\t" "smull %0, %1, %2, %3\n\t" : "=&r"(rd_lo), "=&r"(rd_hi) - : "%r"(b), "r"(a<<16) + : "%r"(b), "r"(SHL32(a,16)) ); /*We intentionally don't OR in the high bit of rd_lo for speed.*/ - return rd_hi<<1; + return SHL32(rd_hi,1); } #define MULT16_32_Q15(a, b) (MULT16_32_Q15_armv4(a, b)) diff --git a/code/opus-1.1.4/celt/arm/fixed_armv5e.h b/code/opus-1.1.4/celt/arm/fixed_armv5e.h index 36a63211..6bf73cba 100644 --- a/code/opus-1.1.4/celt/arm/fixed_armv5e.h +++ b/code/opus-1.1.4/celt/arm/fixed_armv5e.h @@ -59,7 +59,7 @@ static OPUS_INLINE opus_val32 MULT16_32_Q15_armv5e(opus_val16 a, opus_val32 b) : "=r"(res) : "r"(b), "r"(a) ); - return res<<1; + return SHL32(res,1); } #define MULT16_32_Q15(a, b) (MULT16_32_Q15_armv5e(a, b)) @@ -76,7 +76,7 @@ static OPUS_INLINE opus_val32 MAC16_32_Q15_armv5e(opus_val32 c, opus_val16 a, "#MAC16_32_Q15\n\t" "smlawb %0, %1, %2, %3;\n" : "=r"(res) - : "r"(b<<1), "r"(a), "r"(c) + : "r"(SHL32(b,1)), "r"(a), "r"(c) ); return res; } diff --git a/code/opus-1.1.4/celt/arm/mdct_arm.h b/code/opus-1.1.4/celt/arm/mdct_arm.h index 49cbb445..14200bac 100644 --- a/code/opus-1.1.4/celt/arm/mdct_arm.h +++ b/code/opus-1.1.4/celt/arm/mdct_arm.h @@ -33,7 +33,6 @@ #if !defined(MDCT_ARM_H) #define MDCT_ARM_H -#include "config.h" #include "mdct.h" #if defined(HAVE_ARM_NE10) diff --git a/code/opus-1.1.4/celt/arm/pitch_arm.h b/code/opus-1.1.4/celt/arm/pitch_arm.h index 14331169..4ee13bd0 100644 --- a/code/opus-1.1.4/celt/arm/pitch_arm.h +++ b/code/opus-1.1.4/celt/arm/pitch_arm.h @@ -30,11 +30,47 @@ # include "armcpu.h" +# if defined(OPUS_ARM_MAY_HAVE_NEON_INTR) +opus_val32 celt_inner_prod_neon(const opus_val16 *x, const opus_val16 *y, int N); +void dual_inner_prod_neon(const opus_val16 *x, const opus_val16 *y01, + const opus_val16 *y02, int N, opus_val32 *xy1, opus_val32 *xy2); + +# if !defined(OPUS_HAVE_RTCD) && defined(OPUS_ARM_PRESUME_NEON) +# define OVERRIDE_CELT_INNER_PROD (1) +# define OVERRIDE_DUAL_INNER_PROD (1) +# define celt_inner_prod(x, y, N, arch) ((void)(arch), PRESUME_NEON(celt_inner_prod)(x, y, N)) +# define dual_inner_prod(x, y01, y02, N, xy1, xy2, arch) ((void)(arch), PRESUME_NEON(dual_inner_prod)(x, y01, y02, N, xy1, xy2)) +# endif +# endif + +# if !defined(OVERRIDE_CELT_INNER_PROD) +# if defined(OPUS_HAVE_RTCD) && (defined(OPUS_ARM_MAY_HAVE_NEON_INTR) && !defined(OPUS_ARM_PRESUME_NEON_INTR)) +extern opus_val32 (*const CELT_INNER_PROD_IMPL[OPUS_ARCHMASK+1])(const opus_val16 *x, const opus_val16 *y, int N); +# define OVERRIDE_CELT_INNER_PROD (1) +# define celt_inner_prod(x, y, N, arch) ((*CELT_INNER_PROD_IMPL[(arch)&OPUS_ARCHMASK])(x, y, N)) +# elif defined(OPUS_ARM_PRESUME_NEON_INTR) +# define OVERRIDE_CELT_INNER_PROD (1) +# define celt_inner_prod(x, y, N, arch) ((void)(arch), celt_inner_prod_neon(x, y, N)) +# endif +# endif + +# if !defined(OVERRIDE_DUAL_INNER_PROD) +# if defined(OPUS_HAVE_RTCD) && (defined(OPUS_ARM_MAY_HAVE_NEON_INTR) && !defined(OPUS_ARM_PRESUME_NEON_INTR)) +extern void (*const DUAL_INNER_PROD_IMPL[OPUS_ARCHMASK+1])(const opus_val16 *x, + const opus_val16 *y01, const opus_val16 *y02, int N, opus_val32 *xy1, opus_val32 *xy2); +# define OVERRIDE_DUAL_INNER_PROD (1) +# define dual_inner_prod(x, y01, y02, N, xy1, xy2, arch) ((*DUAL_INNER_PROD_IMPL[(arch)&OPUS_ARCHMASK])(x, y01, y02, N, xy1, xy2)) +# elif defined(OPUS_ARM_PRESUME_NEON_INTR) +# define OVERRIDE_DUAL_INNER_PROD (1) +# define dual_inner_prod(x, y01, y02, N, xy1, xy2, arch) ((void)(arch), dual_inner_prod_neon(x, y01, y02, N, xy1, xy2)) +# endif +# endif + # if defined(FIXED_POINT) # if defined(OPUS_ARM_MAY_HAVE_NEON) opus_val32 celt_pitch_xcorr_neon(const opus_val16 *_x, const opus_val16 *_y, - opus_val32 *xcorr, int len, int max_pitch); + opus_val32 *xcorr, int len, int max_pitch, int arch); # endif # if defined(OPUS_ARM_MAY_HAVE_MEDIA) @@ -43,7 +79,7 @@ opus_val32 celt_pitch_xcorr_neon(const opus_val16 *_x, const opus_val16 *_y, # if defined(OPUS_ARM_MAY_HAVE_EDSP) opus_val32 celt_pitch_xcorr_edsp(const opus_val16 *_x, const opus_val16 *_y, - opus_val32 *xcorr, int len, int max_pitch); + opus_val32 *xcorr, int len, int max_pitch, int arch); # endif # if defined(OPUS_HAVE_RTCD) && \ @@ -52,18 +88,15 @@ opus_val32 celt_pitch_xcorr_edsp(const opus_val16 *_x, const opus_val16 *_y, (defined(OPUS_ARM_MAY_HAVE_EDSP) && !defined(OPUS_ARM_PRESUME_EDSP))) extern opus_val32 (*const CELT_PITCH_XCORR_IMPL[OPUS_ARCHMASK+1])(const opus_val16 *, - const opus_val16 *, opus_val32 *, int, int); + const opus_val16 *, opus_val32 *, int, int, int); # define OVERRIDE_PITCH_XCORR (1) -# define celt_pitch_xcorr(_x, _y, xcorr, len, max_pitch, arch) \ - ((*CELT_PITCH_XCORR_IMPL[(arch)&OPUS_ARCHMASK])(_x, _y, \ - xcorr, len, max_pitch)) +# define celt_pitch_xcorr (*CELT_PITCH_XCORR_IMPL[(arch)&OPUS_ARCHMASK]) # elif defined(OPUS_ARM_PRESUME_EDSP) || \ defined(OPUS_ARM_PRESUME_MEDIA) || \ defined(OPUS_ARM_PRESUME_NEON) # define OVERRIDE_PITCH_XCORR (1) -# define celt_pitch_xcorr(_x, _y, xcorr, len, max_pitch, arch) \ - ((void)(arch),PRESUME_NEON(celt_pitch_xcorr)(_x, _y, xcorr, len, max_pitch)) +# define celt_pitch_xcorr (PRESUME_NEON(celt_pitch_xcorr)) # endif @@ -99,25 +132,22 @@ extern void (*const XCORR_KERNEL_IMPL[OPUS_ARCHMASK + 1])( /* Float case */ #if defined(OPUS_ARM_MAY_HAVE_NEON_INTR) void celt_pitch_xcorr_float_neon(const opus_val16 *_x, const opus_val16 *_y, - opus_val32 *xcorr, int len, int max_pitch); + opus_val32 *xcorr, int len, int max_pitch, int arch); #endif # if defined(OPUS_HAVE_RTCD) && \ (defined(OPUS_ARM_MAY_HAVE_NEON_INTR) && !defined(OPUS_ARM_PRESUME_NEON_INTR)) extern void (*const CELT_PITCH_XCORR_IMPL[OPUS_ARCHMASK+1])(const opus_val16 *, - const opus_val16 *, opus_val32 *, int, int); + const opus_val16 *, opus_val32 *, int, int, int); # define OVERRIDE_PITCH_XCORR (1) -# define celt_pitch_xcorr(_x, _y, xcorr, len, max_pitch, arch) \ - ((*CELT_PITCH_XCORR_IMPL[(arch)&OPUS_ARCHMASK])(_x, _y, \ - xcorr, len, max_pitch)) +# define celt_pitch_xcorr (*CELT_PITCH_XCORR_IMPL[(arch)&OPUS_ARCHMASK]) # elif defined(OPUS_ARM_PRESUME_NEON_INTR) # define OVERRIDE_PITCH_XCORR (1) -# define celt_pitch_xcorr(_x, _y, xcorr, len, max_pitch, arch) \ - ((void)(arch),celt_pitch_xcorr_float_neon(_x, _y, xcorr, len, max_pitch)) +# define celt_pitch_xcorr celt_pitch_xcorr_float_neon # endif diff --git a/code/opus-1.1.4/celt/arm/pitch_neon_intr.c b/code/opus-1.1.4/celt/arm/pitch_neon_intr.c new file mode 100644 index 00000000..1ac38c43 --- /dev/null +++ b/code/opus-1.1.4/celt/arm/pitch_neon_intr.c @@ -0,0 +1,290 @@ +/*********************************************************************** +Copyright (c) 2017 Google Inc. +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions +are met: +- Redistributions of source code must retain the above copyright notice, +this list of conditions and the following disclaimer. +- Redistributions in binary form must reproduce the above copyright +notice, this list of conditions and the following disclaimer in the +documentation and/or other materials provided with the distribution. +- Neither the name of Internet Society, IETF or IETF Trust, nor the +names of specific contributors, may be used to endorse or promote +products derived from this software without specific prior written +permission. +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE +ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE +LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR +CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF +SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS +INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN +CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) +ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +POSSIBILITY OF SUCH DAMAGE. +***********************************************************************/ + +#ifdef HAVE_CONFIG_H +#include "config.h" +#endif + +#include +#include "pitch.h" + +#ifdef FIXED_POINT + +opus_val32 celt_inner_prod_neon(const opus_val16 *x, const opus_val16 *y, int N) +{ + int i; + opus_val32 xy; + int16x8_t x_s16x8, y_s16x8; + int32x4_t xy_s32x4 = vdupq_n_s32(0); + int64x2_t xy_s64x2; + int64x1_t xy_s64x1; + + for (i = 0; i < N - 7; i += 8) { + x_s16x8 = vld1q_s16(&x[i]); + y_s16x8 = vld1q_s16(&y[i]); + xy_s32x4 = vmlal_s16(xy_s32x4, vget_low_s16 (x_s16x8), vget_low_s16 (y_s16x8)); + xy_s32x4 = vmlal_s16(xy_s32x4, vget_high_s16(x_s16x8), vget_high_s16(y_s16x8)); + } + + if (N - i >= 4) { + const int16x4_t x_s16x4 = vld1_s16(&x[i]); + const int16x4_t y_s16x4 = vld1_s16(&y[i]); + xy_s32x4 = vmlal_s16(xy_s32x4, x_s16x4, y_s16x4); + i += 4; + } + + xy_s64x2 = vpaddlq_s32(xy_s32x4); + xy_s64x1 = vadd_s64(vget_low_s64(xy_s64x2), vget_high_s64(xy_s64x2)); + xy = vget_lane_s32(vreinterpret_s32_s64(xy_s64x1), 0); + + for (; i < N; i++) { + xy = MAC16_16(xy, x[i], y[i]); + } + +#ifdef OPUS_CHECK_ASM + celt_assert(celt_inner_prod_c(x, y, N) == xy); +#endif + + return xy; +} + +void dual_inner_prod_neon(const opus_val16 *x, const opus_val16 *y01, const opus_val16 *y02, + int N, opus_val32 *xy1, opus_val32 *xy2) +{ + int i; + opus_val32 xy01, xy02; + int16x8_t x_s16x8, y01_s16x8, y02_s16x8; + int32x4_t xy01_s32x4 = vdupq_n_s32(0); + int32x4_t xy02_s32x4 = vdupq_n_s32(0); + int64x2_t xy01_s64x2, xy02_s64x2; + int64x1_t xy01_s64x1, xy02_s64x1; + + for (i = 0; i < N - 7; i += 8) { + x_s16x8 = vld1q_s16(&x[i]); + y01_s16x8 = vld1q_s16(&y01[i]); + y02_s16x8 = vld1q_s16(&y02[i]); + xy01_s32x4 = vmlal_s16(xy01_s32x4, vget_low_s16 (x_s16x8), vget_low_s16 (y01_s16x8)); + xy02_s32x4 = vmlal_s16(xy02_s32x4, vget_low_s16 (x_s16x8), vget_low_s16 (y02_s16x8)); + xy01_s32x4 = vmlal_s16(xy01_s32x4, vget_high_s16(x_s16x8), vget_high_s16(y01_s16x8)); + xy02_s32x4 = vmlal_s16(xy02_s32x4, vget_high_s16(x_s16x8), vget_high_s16(y02_s16x8)); + } + + if (N - i >= 4) { + const int16x4_t x_s16x4 = vld1_s16(&x[i]); + const int16x4_t y01_s16x4 = vld1_s16(&y01[i]); + const int16x4_t y02_s16x4 = vld1_s16(&y02[i]); + xy01_s32x4 = vmlal_s16(xy01_s32x4, x_s16x4, y01_s16x4); + xy02_s32x4 = vmlal_s16(xy02_s32x4, x_s16x4, y02_s16x4); + i += 4; + } + + xy01_s64x2 = vpaddlq_s32(xy01_s32x4); + xy02_s64x2 = vpaddlq_s32(xy02_s32x4); + xy01_s64x1 = vadd_s64(vget_low_s64(xy01_s64x2), vget_high_s64(xy01_s64x2)); + xy02_s64x1 = vadd_s64(vget_low_s64(xy02_s64x2), vget_high_s64(xy02_s64x2)); + xy01 = vget_lane_s32(vreinterpret_s32_s64(xy01_s64x1), 0); + xy02 = vget_lane_s32(vreinterpret_s32_s64(xy02_s64x1), 0); + + for (; i < N; i++) { + xy01 = MAC16_16(xy01, x[i], y01[i]); + xy02 = MAC16_16(xy02, x[i], y02[i]); + } + *xy1 = xy01; + *xy2 = xy02; + +#ifdef OPUS_CHECK_ASM + { + opus_val32 xy1_c, xy2_c; + dual_inner_prod_c(x, y01, y02, N, &xy1_c, &xy2_c); + celt_assert(xy1_c == *xy1); + celt_assert(xy2_c == *xy2); + } +#endif +} + +#else /* !FIXED_POINT */ + +/* ========================================================================== */ + +#ifdef OPUS_CHECK_ASM + +/* This part of code simulates floating-point NEON operations. */ + +/* celt_inner_prod_neon_float_c_simulation() simulates the floating-point */ +/* operations of celt_inner_prod_neon(), and both functions should have bit */ +/* exact output. */ +static opus_val32 celt_inner_prod_neon_float_c_simulation(const opus_val16 *x, const opus_val16 *y, int N) +{ + int i; + opus_val32 xy, xy0 = 0, xy1 = 0, xy2 = 0, xy3 = 0; + for (i = 0; i < N - 3; i += 4) { + xy0 = MAC16_16(xy0, x[i + 0], y[i + 0]); + xy1 = MAC16_16(xy1, x[i + 1], y[i + 1]); + xy2 = MAC16_16(xy2, x[i + 2], y[i + 2]); + xy3 = MAC16_16(xy3, x[i + 3], y[i + 3]); + } + xy0 += xy2; + xy1 += xy3; + xy = xy0 + xy1; + for (; i < N; i++) { + xy = MAC16_16(xy, x[i], y[i]); + } + return xy; +} + +/* dual_inner_prod_neon_float_c_simulation() simulates the floating-point */ +/* operations of dual_inner_prod_neon(), and both functions should have bit */ +/* exact output. */ +static void dual_inner_prod_neon_float_c_simulation(const opus_val16 *x, const opus_val16 *y01, const opus_val16 *y02, + int N, opus_val32 *xy1, opus_val32 *xy2) +{ + int i; + opus_val32 xy01, xy02, xy01_0 = 0, xy01_1 = 0, xy01_2 = 0, xy01_3 = 0, xy02_0 = 0, xy02_1 = 0, xy02_2 = 0, xy02_3 = 0; + for (i = 0; i < N - 3; i += 4) { + xy01_0 = MAC16_16(xy01_0, x[i + 0], y01[i + 0]); + xy01_1 = MAC16_16(xy01_1, x[i + 1], y01[i + 1]); + xy01_2 = MAC16_16(xy01_2, x[i + 2], y01[i + 2]); + xy01_3 = MAC16_16(xy01_3, x[i + 3], y01[i + 3]); + xy02_0 = MAC16_16(xy02_0, x[i + 0], y02[i + 0]); + xy02_1 = MAC16_16(xy02_1, x[i + 1], y02[i + 1]); + xy02_2 = MAC16_16(xy02_2, x[i + 2], y02[i + 2]); + xy02_3 = MAC16_16(xy02_3, x[i + 3], y02[i + 3]); + } + xy01_0 += xy01_2; + xy02_0 += xy02_2; + xy01_1 += xy01_3; + xy02_1 += xy02_3; + xy01 = xy01_0 + xy01_1; + xy02 = xy02_0 + xy02_1; + for (; i < N; i++) { + xy01 = MAC16_16(xy01, x[i], y01[i]); + xy02 = MAC16_16(xy02, x[i], y02[i]); + } + *xy1 = xy01; + *xy2 = xy02; +} + +#endif /* OPUS_CHECK_ASM */ + +/* ========================================================================== */ + +opus_val32 celt_inner_prod_neon(const opus_val16 *x, const opus_val16 *y, int N) +{ + int i; + opus_val32 xy; + float32x4_t xy_f32x4 = vdupq_n_f32(0); + float32x2_t xy_f32x2; + + for (i = 0; i < N - 7; i += 8) { + float32x4_t x_f32x4, y_f32x4; + x_f32x4 = vld1q_f32(&x[i]); + y_f32x4 = vld1q_f32(&y[i]); + xy_f32x4 = vmlaq_f32(xy_f32x4, x_f32x4, y_f32x4); + x_f32x4 = vld1q_f32(&x[i + 4]); + y_f32x4 = vld1q_f32(&y[i + 4]); + xy_f32x4 = vmlaq_f32(xy_f32x4, x_f32x4, y_f32x4); + } + + if (N - i >= 4) { + const float32x4_t x_f32x4 = vld1q_f32(&x[i]); + const float32x4_t y_f32x4 = vld1q_f32(&y[i]); + xy_f32x4 = vmlaq_f32(xy_f32x4, x_f32x4, y_f32x4); + i += 4; + } + + xy_f32x2 = vadd_f32(vget_low_f32(xy_f32x4), vget_high_f32(xy_f32x4)); + xy_f32x2 = vpadd_f32(xy_f32x2, xy_f32x2); + xy = vget_lane_f32(xy_f32x2, 0); + + for (; i < N; i++) { + xy = MAC16_16(xy, x[i], y[i]); + } + +#ifdef OPUS_CHECK_ASM + celt_assert(ABS32(celt_inner_prod_neon_float_c_simulation(x, y, N) - xy) <= VERY_SMALL); +#endif + + return xy; +} + +void dual_inner_prod_neon(const opus_val16 *x, const opus_val16 *y01, const opus_val16 *y02, + int N, opus_val32 *xy1, opus_val32 *xy2) +{ + int i; + opus_val32 xy01, xy02; + float32x4_t xy01_f32x4 = vdupq_n_f32(0); + float32x4_t xy02_f32x4 = vdupq_n_f32(0); + float32x2_t xy01_f32x2, xy02_f32x2; + + for (i = 0; i < N - 7; i += 8) { + float32x4_t x_f32x4, y01_f32x4, y02_f32x4; + x_f32x4 = vld1q_f32(&x[i]); + y01_f32x4 = vld1q_f32(&y01[i]); + y02_f32x4 = vld1q_f32(&y02[i]); + xy01_f32x4 = vmlaq_f32(xy01_f32x4, x_f32x4, y01_f32x4); + xy02_f32x4 = vmlaq_f32(xy02_f32x4, x_f32x4, y02_f32x4); + x_f32x4 = vld1q_f32(&x[i + 4]); + y01_f32x4 = vld1q_f32(&y01[i + 4]); + y02_f32x4 = vld1q_f32(&y02[i + 4]); + xy01_f32x4 = vmlaq_f32(xy01_f32x4, x_f32x4, y01_f32x4); + xy02_f32x4 = vmlaq_f32(xy02_f32x4, x_f32x4, y02_f32x4); + } + + if (N - i >= 4) { + const float32x4_t x_f32x4 = vld1q_f32(&x[i]); + const float32x4_t y01_f32x4 = vld1q_f32(&y01[i]); + const float32x4_t y02_f32x4 = vld1q_f32(&y02[i]); + xy01_f32x4 = vmlaq_f32(xy01_f32x4, x_f32x4, y01_f32x4); + xy02_f32x4 = vmlaq_f32(xy02_f32x4, x_f32x4, y02_f32x4); + i += 4; + } + + xy01_f32x2 = vadd_f32(vget_low_f32(xy01_f32x4), vget_high_f32(xy01_f32x4)); + xy02_f32x2 = vadd_f32(vget_low_f32(xy02_f32x4), vget_high_f32(xy02_f32x4)); + xy01_f32x2 = vpadd_f32(xy01_f32x2, xy01_f32x2); + xy02_f32x2 = vpadd_f32(xy02_f32x2, xy02_f32x2); + xy01 = vget_lane_f32(xy01_f32x2, 0); + xy02 = vget_lane_f32(xy02_f32x2, 0); + + for (; i < N; i++) { + xy01 = MAC16_16(xy01, x[i], y01[i]); + xy02 = MAC16_16(xy02, x[i], y02[i]); + } + *xy1 = xy01; + *xy2 = xy02; + +#ifdef OPUS_CHECK_ASM + { + opus_val32 xy1_c, xy2_c; + dual_inner_prod_neon_float_c_simulation(x, y01, y02, N, &xy1_c, &xy2_c); + celt_assert(ABS32(xy1_c - *xy1) <= VERY_SMALL); + celt_assert(ABS32(xy2_c - *xy2) <= VERY_SMALL); + } +#endif +} + +#endif /* FIXED_POINT */ diff --git a/code/opus-1.1.4/celt/bands.c b/code/opus-1.1.4/celt/bands.c index 87eaa6c0..3b1f5cfb 100644 --- a/code/opus-1.1.4/celt/bands.c +++ b/code/opus-1.1.4/celt/bands.c @@ -65,7 +65,7 @@ opus_uint32 celt_lcg_rand(opus_uint32 seed) /* This is a cos() approximation designed to be bit-exact on any platform. Bit exactness with this approximation is important because it has an impact on the bit allocation */ -static opus_int16 bitexact_cos(opus_int16 x) +opus_int16 bitexact_cos(opus_int16 x) { opus_int32 tmp; opus_int16 x2; @@ -77,7 +77,7 @@ static opus_int16 bitexact_cos(opus_int16 x) return 1+x2; } -static int bitexact_log2tan(int isin,int icos) +int bitexact_log2tan(int isin,int icos) { int lc; int ls; @@ -92,10 +92,11 @@ static int bitexact_log2tan(int isin,int icos) #ifdef FIXED_POINT /* Compute the amplitude (sqrt energy) in each of the bands */ -void compute_band_energies(const CELTMode *m, const celt_sig *X, celt_ener *bandE, int end, int C, int LM) +void compute_band_energies(const CELTMode *m, const celt_sig *X, celt_ener *bandE, int end, int C, int LM, int arch) { int i, c, N; const opus_int16 *eBands = m->eBands; + (void)arch; N = m->shortMdctSize<eBands; @@ -164,7 +165,7 @@ void compute_band_energies(const CELTMode *m, const celt_sig *X, celt_ener *band for (i=0;inbEBands] = celt_sqrt(sum); /*printf ("%f ", bandE[i+c*m->nbEBands]);*/ } @@ -224,9 +225,9 @@ void denormalise_bands(const CELTMode *m, const celt_norm * OPUS_RESTRICT X, #endif j=M*eBands[i]; band_end = M*eBands[i+1]; - lg = ADD16(bandLogE[i], SHL16((opus_val16)eMeans[i],6)); + lg = SATURATE16(ADD32(bandLogE[i], SHL32((opus_val32)eMeans[i],6))); #ifndef FIXED_POINT - g = celt_exp2(lg); + g = celt_exp2(MIN32(32.f, lg)); #else /* Handle the integer part of the log energy */ shift = 16-(lg>>DB_SHIFT); @@ -241,12 +242,12 @@ void denormalise_bands(const CELTMode *m, const celt_norm * OPUS_RESTRICT X, /* Handle extreme gains with negative shift. */ if (shift<0) { - /* For shift < -2 we'd be likely to overflow, so we're capping - the gain here. This shouldn't happen unless the bitstream is - already corrupted. */ - if (shift < -2) + /* For shift <= -2 and g > 16384 we'd be likely to overflow, so we're + capping the gain here, which is equivalent to a cap of 18 on lg. + This shouldn't trigger unless the bitstream is already corrupted. */ + if (shift <= -2) { - g = 32767; + g = 16384; shift = -2; } do { @@ -360,6 +361,30 @@ void anti_collapse(const CELTMode *m, celt_norm *X_, unsigned char *collapse_mas } } +/* Compute the weights to use for optimizing normalized distortion across + channels. We use the amplitude to weight square distortion, which means + that we use the square root of the value we would have been using if we + wanted to minimize the MSE in the non-normalized domain. This roughly + corresponds to some quick-and-dirty perceptual experiments I ran to + measure inter-aural masking (there doesn't seem to be any published data + on the topic). */ +static void compute_channel_weights(celt_ener Ex, celt_ener Ey, opus_val16 w[2]) +{ + celt_ener minE; +#if FIXED_POINT + int shift; +#endif + minE = MIN32(Ex, Ey); + /* Adjustment to make the weights a bit more conservative. */ + Ex = ADD32(Ex, minE/3); + Ey = ADD32(Ey, minE/3); +#if FIXED_POINT + shift = celt_ilog2(EPSILON+MAX32(Ex, Ey))-14; +#endif + w[0] = VSHR32(Ex, shift); + w[1] = VSHR32(Ey, shift); +} + static void intensity_stereo(const CELTMode *m, celt_norm * OPUS_RESTRICT X, const celt_norm * OPUS_RESTRICT Y, const celt_ener *bandE, int bandID, int N) { int i = bandID; @@ -647,6 +672,7 @@ static int compute_qn(int N, int b, int offset, int pulse_cap, int stereo) struct band_ctx { int encode; + int resynth; const CELTMode *m; int i; int intensity; @@ -657,6 +683,9 @@ struct band_ctx { const celt_ener *bandE; opus_uint32 seed; int arch; + int theta_round; + int disable_inv; + int avoid_split_noise; }; struct split_ctx { @@ -714,8 +743,35 @@ static void compute_theta(struct band_ctx *ctx, struct split_ctx *sctx, if (qn!=1) { if (encode) - itheta = (itheta*(opus_int32)qn+8192)>>14; - + { + if (!stereo || ctx->theta_round == 0) + { + itheta = (itheta*(opus_int32)qn+8192)>>14; + if (!stereo && ctx->avoid_split_noise && itheta > 0 && itheta < qn) + { + /* Check if the selected value of theta will cause the bit allocation + to inject noise on one side. If so, make sure the energy of that side + is zero. */ + int unquantized = celt_udiv((opus_int32)itheta*16384, qn); + imid = bitexact_cos((opus_int16)unquantized); + iside = bitexact_cos((opus_int16)(16384-unquantized)); + delta = FRAC_MUL16((N-1)<<7,bitexact_log2tan(iside,imid)); + if (delta > *b) + itheta = qn; + else if (delta < -*b) + itheta = 0; + } + } else { + int down; + /* Bias quantization towards itheta=0 and itheta=16384. */ + int bias = itheta > 8192 ? 32767/qn : -32767/qn; + down = IMIN(qn-1, IMAX(0, (itheta*(opus_int32)qn + bias)>>14)); + if (ctx->theta_round < 0) + itheta = down; + else + itheta = down+1; + } + } /* Entropy coding of the angle. We use a uniform pdf for the time split, a step for stereo, and a triangular one for the rest. */ if (stereo && N>2) @@ -793,7 +849,7 @@ static void compute_theta(struct band_ctx *ctx, struct split_ctx *sctx, } else if (stereo) { if (encode) { - inv = itheta > 8192; + inv = itheta > 8192 && !ctx->disable_inv; if (inv) { int j; @@ -810,6 +866,9 @@ static void compute_theta(struct band_ctx *ctx, struct split_ctx *sctx, inv = ec_dec_bit_logp(ec, 2); } else inv = 0; + /* inv flag override to avoid problems with downmixing. */ + if (ctx->disable_inv) + inv = 0; itheta = 0; } qalloc = ec_tell_frac(ec) - tell; @@ -845,11 +904,6 @@ static void compute_theta(struct band_ctx *ctx, struct split_ctx *sctx, static unsigned quant_band_n1(struct band_ctx *ctx, celt_norm *X, celt_norm *Y, int b, celt_norm *lowband_out) { -#ifdef RESYNTH - int resynth = 1; -#else - int resynth = !ctx->encode; -#endif int c; int stereo; celt_norm *x = X; @@ -874,7 +928,7 @@ static unsigned quant_band_n1(struct band_ctx *ctx, celt_norm *X, celt_norm *Y, ctx->remaining_bits -= 1<resynth) x[0] = sign ? -NORM_SCALING : NORM_SCALING; x = Y; } while (++c<1+stereo); @@ -899,11 +953,6 @@ static unsigned quant_partition(struct band_ctx *ctx, celt_norm *X, int B0=B; opus_val16 mid=0, side=0; unsigned cm=0; -#ifdef RESYNTH - int resynth = 1; -#else - int resynth = !ctx->encode; -#endif celt_norm *Y=NULL; int encode; const CELTMode *m; @@ -935,8 +984,7 @@ static unsigned quant_partition(struct band_ctx *ctx, celt_norm *X, fill = (fill&1)|(fill<<1); B = (B+1)>>1; - compute_theta(ctx, &sctx, X, Y, N, &b, B, B0, - LM, 0, &fill); + compute_theta(ctx, &sctx, X, Y, N, &b, B, B0, LM, 0, &fill); imid = sctx.imid; iside = sctx.iside; delta = sctx.delta; @@ -970,24 +1018,20 @@ static unsigned quant_partition(struct band_ctx *ctx, celt_norm *X, rebalance = ctx->remaining_bits; if (mbits >= sbits) { - cm = quant_partition(ctx, X, N, mbits, B, - lowband, LM, + cm = quant_partition(ctx, X, N, mbits, B, lowband, LM, MULT16_16_P15(gain,mid), fill); rebalance = mbits - (rebalance-ctx->remaining_bits); if (rebalance > 3<>B)<<(B0>>1); } else { - cm = quant_partition(ctx, Y, N, sbits, B, - next_lowband2, LM, + cm = quant_partition(ctx, Y, N, sbits, B, next_lowband2, LM, MULT16_16_P15(gain,side), fill>>B)<<(B0>>1); rebalance = sbits - (rebalance-ctx->remaining_bits); if (rebalance > 3<resynth, ctx->arch); } else { cm = alg_unquant(X, N, K, spread, B, ec, gain); } } else { /* If there's no pulse, fill the band anyway */ int j; - if (resynth) + if (ctx->resynth) { unsigned cm_mask; /* B can be as large as 16, so this shift might overflow an int on a @@ -1080,11 +1120,6 @@ static unsigned quant_band(struct band_ctx *ctx, celt_norm *X, int recombine=0; int longBlocks; unsigned cm=0; -#ifdef RESYNTH - int resynth = 1; -#else - int resynth = !ctx->encode; -#endif int k; int encode; int tf_change; @@ -1151,11 +1186,10 @@ static unsigned quant_band(struct band_ctx *ctx, celt_norm *X, deinterleave_hadamard(lowband, N_B>>recombine, B0<resynth) { /* Undo the sample reorganization going from time order to frequency order */ if (B0>1) @@ -1208,11 +1242,6 @@ static unsigned quant_band_stereo(struct band_ctx *ctx, celt_norm *X, celt_norm int inv = 0; opus_val16 mid=0, side=0; unsigned cm=0; -#ifdef RESYNTH - int resynth = 1; -#else - int resynth = !ctx->encode; -#endif int mbits, sbits, delta; int itheta; int qalloc; @@ -1232,8 +1261,7 @@ static unsigned quant_band_stereo(struct band_ctx *ctx, celt_norm *X, celt_norm orig_fill = fill; - compute_theta(ctx, &sctx, X, Y, N, &b, B, B, - LM, 1, &fill); + compute_theta(ctx, &sctx, X, Y, N, &b, B, B, LM, 1, &fill); inv = sctx.inv; imid = sctx.imid; iside = sctx.iside; @@ -1281,13 +1309,13 @@ static unsigned quant_band_stereo(struct band_ctx *ctx, celt_norm *X, celt_norm sign = 1-2*sign; /* We use orig_fill here because we want to fold the side, but if itheta==16384, we'll have cleared the low bits of fill. */ - cm = quant_band(ctx, x2, N, mbits, B, lowband, - LM, lowband_out, Q15ONE, lowband_scratch, orig_fill); + cm = quant_band(ctx, x2, N, mbits, B, lowband, LM, lowband_out, Q15ONE, + lowband_scratch, orig_fill); /* We don't split N=2 bands, so cm is either 1 or 0 (for a fold-collapse), and there's no need to worry about mixing with the other channel. */ y2[0] = -sign*x2[1]; y2[1] = sign*x2[0]; - if (resynth) + if (ctx->resynth) { celt_norm tmp; X[0] = MULT16_16_Q15(mid, X[0]); @@ -1314,38 +1342,32 @@ static unsigned quant_band_stereo(struct band_ctx *ctx, celt_norm *X, celt_norm { /* In stereo mode, we do not apply a scaling to the mid because we need the normalized mid for folding later. */ - cm = quant_band(ctx, X, N, mbits, B, - lowband, LM, lowband_out, - Q15ONE, lowband_scratch, fill); + cm = quant_band(ctx, X, N, mbits, B, lowband, LM, lowband_out, Q15ONE, + lowband_scratch, fill); rebalance = mbits - (rebalance-ctx->remaining_bits); if (rebalance > 3<>B); + cm |= quant_band(ctx, Y, N, sbits, B, NULL, LM, NULL, side, NULL, fill>>B); } else { /* For a stereo split, the high bits of fill are always zero, so no folding will be done to the side. */ - cm = quant_band(ctx, Y, N, sbits, B, - NULL, LM, NULL, - side, NULL, fill>>B); + cm = quant_band(ctx, Y, N, sbits, B, NULL, LM, NULL, side, NULL, fill>>B); rebalance = sbits - (rebalance-ctx->remaining_bits); if (rebalance > 3<resynth) { if (N!=2) stereo_merge(X, Y, mid, N, ctx->arch); @@ -1359,19 +1381,38 @@ static unsigned quant_band_stereo(struct band_ctx *ctx, celt_norm *X, celt_norm return cm; } +static void special_hybrid_folding(const CELTMode *m, celt_norm *norm, celt_norm *norm2, int start, int M, int dual_stereo) +{ + int n1, n2; + const opus_int16 * OPUS_RESTRICT eBands = m->eBands; + n1 = M*(eBands[start+1]-eBands[start]); + n2 = M*(eBands[start+2]-eBands[start+1]); + /* Duplicate enough of the first band folding data to be able to fold the second band. + Copies no data for CELT-only mode. */ + OPUS_COPY(&norm[n1], &norm[2*n1 - n2], n2-n1); + if (dual_stereo) + OPUS_COPY(&norm2[n1], &norm2[2*n1 - n2], n2-n1); +} void quant_all_bands(int encode, const CELTMode *m, int start, int end, celt_norm *X_, celt_norm *Y_, unsigned char *collapse_masks, const celt_ener *bandE, int *pulses, int shortBlocks, int spread, int dual_stereo, int intensity, int *tf_res, opus_int32 total_bits, opus_int32 balance, ec_ctx *ec, int LM, int codedBands, - opus_uint32 *seed, int arch) + opus_uint32 *seed, int complexity, int arch, int disable_inv) { int i; opus_int32 remaining_bits; const opus_int16 * OPUS_RESTRICT eBands = m->eBands; celt_norm * OPUS_RESTRICT norm, * OPUS_RESTRICT norm2; VARDECL(celt_norm, _norm); + VARDECL(celt_norm, _lowband_scratch); + VARDECL(celt_norm, X_save); + VARDECL(celt_norm, Y_save); + VARDECL(celt_norm, X_save2); + VARDECL(celt_norm, Y_save2); + VARDECL(celt_norm, norm_save2); + int resynth_alloc; celt_norm *lowband_scratch; int B; int M; @@ -1379,10 +1420,11 @@ void quant_all_bands(int encode, const CELTMode *m, int start, int end, int update_lowband = 1; int C = Y_ != NULL ? 2 : 1; int norm_offset; + int theta_rdo = encode && Y_!=NULL && !dual_stereo && complexity>=8; #ifdef RESYNTH int resynth = 1; #else - int resynth = !encode; + int resynth = !encode || theta_rdo; #endif struct band_ctx ctx; SAVE_STACK; @@ -1395,9 +1437,24 @@ void quant_all_bands(int encode, const CELTMode *m, int start, int end, ALLOC(_norm, C*(M*eBands[m->nbEBands-1]-norm_offset), celt_norm); norm = _norm; norm2 = norm + M*eBands[m->nbEBands-1]-norm_offset; - /* We can use the last band as scratch space because we don't need that - scratch space for the last band. */ - lowband_scratch = X_+M*eBands[m->nbEBands-1]; + + /* For decoding, we can use the last band as scratch space because we don't need that + scratch space for the last band and we don't care about the data there until we're + decoding the last band. */ + if (encode && resynth) + resynth_alloc = M*(eBands[m->nbEBands]-eBands[m->nbEBands-1]); + else + resynth_alloc = ALLOC_NONE; + ALLOC(_lowband_scratch, resynth_alloc, celt_norm); + if (encode && resynth) + lowband_scratch = _lowband_scratch; + else + lowband_scratch = X_+M*eBands[m->nbEBands-1]; + ALLOC(X_save, resynth_alloc, celt_norm); + ALLOC(Y_save, resynth_alloc, celt_norm); + ALLOC(X_save2, resynth_alloc, celt_norm); + ALLOC(Y_save2, resynth_alloc, celt_norm); + ALLOC(norm_save2, resynth_alloc, celt_norm); lowband_offset = 0; ctx.bandE = bandE; @@ -1408,6 +1465,11 @@ void quant_all_bands(int encode, const CELTMode *m, int start, int end, ctx.seed = *seed; ctx.spread = spread; ctx.arch = arch; + ctx.disable_inv = disable_inv; + ctx.resynth = resynth; + ctx.theta_round = 0; + /* Avoid injecting noise in the first band on transients. */ + ctx.avoid_split_noise = B > 1; for (i=start;i= M*eBands[start] || i==start+1) && (update_lowband || lowband_offset==0)) + lowband_offset = i; + if (i == start+1) + special_hybrid_folding(m, norm, norm2, start, M, dual_stereo); +#else if (resynth && M*eBands[i]-N >= M*eBands[start] && (update_lowband || lowband_offset==0)) lowband_offset = i; +#endif tf_change = tf_res[i]; ctx.tf_change = tf_change; @@ -1457,7 +1526,7 @@ void quant_all_bands(int encode, const CELTMode *m, int start, int end, Y = norm; lowband_scratch = NULL; } - if (i==end-1) + if (last && !theta_rdo) lowband_scratch = NULL; /* Get a conservative estimate of the collapse_mask's for the bands we're @@ -1472,7 +1541,11 @@ void quant_all_bands(int encode, const CELTMode *m, int start, int end, fold_start = lowband_offset; while(M*eBands[--fold_start] > effective_lowband+norm_offset); fold_end = lowband_offset-1; +#ifdef ENABLE_UPDATE_DRAFT + while(++fold_end < i && M*eBands[fold_end] < effective_lowband+norm_offset+N); +#else while(M*eBands[++fold_end] < effective_lowband+norm_offset+N); +#endif x_cm = y_cm = 0; fold_i = fold_start; do { x_cm |= collapse_masks[fold_i*C+0]; @@ -1505,13 +1578,77 @@ void quant_all_bands(int encode, const CELTMode *m, int start, int end, } else { if (Y!=NULL) { - x_cm = quant_band_stereo(&ctx, X, Y, N, b, B, - effective_lowband != -1 ? norm+effective_lowband : NULL, LM, - last?NULL:norm+M*eBands[i]-norm_offset, lowband_scratch, x_cm|y_cm); + if (theta_rdo && i < intensity) + { + ec_ctx ec_save, ec_save2; + struct band_ctx ctx_save, ctx_save2; + opus_val32 dist0, dist1; + unsigned cm, cm2; + int nstart_bytes, nend_bytes, save_bytes; + unsigned char *bytes_buf; + unsigned char bytes_save[1275]; + opus_val16 w[2]; + compute_channel_weights(bandE[i], bandE[i+m->nbEBands], w); + /* Make a copy. */ + cm = x_cm|y_cm; + ec_save = *ec; + ctx_save = ctx; + OPUS_COPY(X_save, X, N); + OPUS_COPY(Y_save, Y, N); + /* Encode and round down. */ + ctx.theta_round = -1; + x_cm = quant_band_stereo(&ctx, X, Y, N, b, B, + effective_lowband != -1 ? norm+effective_lowband : NULL, LM, + last?NULL:norm+M*eBands[i]-norm_offset, lowband_scratch, cm); + dist0 = MULT16_32_Q15(w[0], celt_inner_prod(X_save, X, N, arch)) + MULT16_32_Q15(w[1], celt_inner_prod(Y_save, Y, N, arch)); + + /* Save first result. */ + cm2 = x_cm; + ec_save2 = *ec; + ctx_save2 = ctx; + OPUS_COPY(X_save2, X, N); + OPUS_COPY(Y_save2, Y, N); + if (!last) + OPUS_COPY(norm_save2, norm+M*eBands[i]-norm_offset, N); + nstart_bytes = ec_save.offs; + nend_bytes = ec_save.storage; + bytes_buf = ec_save.buf+nstart_bytes; + save_bytes = nend_bytes-nstart_bytes; + OPUS_COPY(bytes_save, bytes_buf, save_bytes); + + /* Restore */ + *ec = ec_save; + ctx = ctx_save; + OPUS_COPY(X, X_save, N); + OPUS_COPY(Y, Y_save, N); + if (i == start+1) + special_hybrid_folding(m, norm, norm2, start, M, dual_stereo); + /* Encode and round up. */ + ctx.theta_round = 1; + x_cm = quant_band_stereo(&ctx, X, Y, N, b, B, + effective_lowband != -1 ? norm+effective_lowband : NULL, LM, + last?NULL:norm+M*eBands[i]-norm_offset, lowband_scratch, cm); + dist1 = MULT16_32_Q15(w[0], celt_inner_prod(X_save, X, N, arch)) + MULT16_32_Q15(w[1], celt_inner_prod(Y_save, Y, N, arch)); + if (dist0 >= dist1) { + x_cm = cm2; + *ec = ec_save2; + ctx = ctx_save2; + OPUS_COPY(X, X_save2, N); + OPUS_COPY(Y, Y_save2, N); + if (!last) + OPUS_COPY(norm+M*eBands[i]-norm_offset, norm_save2, N); + OPUS_COPY(bytes_buf, bytes_save, save_bytes); + } + } else { + ctx.theta_round = 0; + x_cm = quant_band_stereo(&ctx, X, Y, N, b, B, + effective_lowband != -1 ? norm+effective_lowband : NULL, LM, + last?NULL:norm+M*eBands[i]-norm_offset, lowband_scratch, x_cm|y_cm); + } } else { x_cm = quant_band(&ctx, X, N, b, B, effective_lowband != -1 ? norm+effective_lowband : NULL, LM, - last?NULL:norm+M*eBands[i]-norm_offset, Q15ONE, lowband_scratch, x_cm|y_cm); + last?NULL:norm+M*eBands[i]-norm_offset, Q15ONE, lowband_scratch, x_cm|y_cm); } y_cm = x_cm; } @@ -1521,6 +1658,9 @@ void quant_all_bands(int encode, const CELTMode *m, int start, int end, /* Update the folding position only as long as we have 1 bit/sample depth. */ update_lowband = b>(N<start = 0; st->end = st->mode->effEBands; st->signalling = 1; +#ifdef ENABLE_UPDATE_DRAFT + st->disable_inv = channels == 1; +#else + st->disable_inv = 0; +#endif st->arch = opus_select_arch(); opus_custom_decoder_ctl(st, OPUS_RESET_STATE); @@ -177,6 +183,36 @@ void opus_custom_decoder_destroy(CELTDecoder *st) } #endif /* CUSTOM_MODES */ +#ifndef CUSTOM_MODES +/* Special case for stereo with no downsampling and no accumulation. This is + quite common and we can make it faster by processing both channels in the + same loop, reducing overhead due to the dependency loop in the IIR filter. */ +static void deemphasis_stereo_simple(celt_sig *in[], opus_val16 *pcm, int N, const opus_val16 coef0, + celt_sig *mem) +{ + celt_sig * OPUS_RESTRICT x0; + celt_sig * OPUS_RESTRICT x1; + celt_sig m0, m1; + int j; + x0=in[0]; + x1=in[1]; + m0 = mem[0]; + m1 = mem[1]; + for (j=0;jmdct, &freq[b], out_syn[0]+NB*b, mode->window, overlap, shift, B, arch); } else { @@ -345,6 +389,12 @@ void celt_synthesis(const CELTMode *mode, celt_norm *X, celt_sig * out_syn[], clt_mdct_backward(&mode->mdct, &freq[b], out_syn[c]+NB*b, mode->window, overlap, shift, B, arch); } while (++cwindow; c=0; do { opus_val16 decay; @@ -561,6 +613,23 @@ static void celt_decode_lost(CELTDecoder * OPUS_RESTRICT st, int N, int LM) #endif } _celt_lpc(lpc+c*LPC_ORDER, ac, LPC_ORDER); +#ifdef FIXED_POINT + /* For fixed-point, apply bandwidth expansion until we can guarantee that + no overflow can happen in the IIR filter. This means: + 32768*sum(abs(filter)) < 2^31 */ + while (1) { + opus_val16 tmp=Q15ONE; + opus_val32 sum=QCONST16(1., SIG_SHIFT); + for (i=0;iarch); + exc+MAX_PERIOD-exc_length, exc_length, LPC_ORDER, st->arch); } /* Check if the waveform is decaying, and if so how fast. @@ -630,9 +698,8 @@ static void celt_decode_lost(CELTDecoder * OPUS_RESTRICT st, int N, int LM) tmp = ROUND16( buf[DECODE_BUFFER_SIZE-MAX_PERIOD-N+extrapolation_offset+j], SIG_SHIFT); - S1 += SHR32(MULT16_16(tmp, tmp), 8); + S1 += SHR32(MULT16_16(tmp, tmp), 10); } - { opus_val16 lpc_mem[LPC_ORDER]; /* Copy the last decoded samples (prior to the overlap region) to @@ -644,6 +711,10 @@ static void celt_decode_lost(CELTDecoder * OPUS_RESTRICT st, int N, int LM) celt_iir(buf+DECODE_BUFFER_SIZE-N, lpc+c*LPC_ORDER, buf+DECODE_BUFFER_SIZE-N, extrapolation_len, LPC_ORDER, lpc_mem, st->arch); +#ifdef FIXED_POINT + for (i=0; i < extrapolation_len; i++) + buf[DECODE_BUFFER_SIZE-N+i] = SATURATE(buf[DECODE_BUFFER_SIZE-N+i], SIG_SAT); +#endif } /* Check if the synthesis energy is higher than expected, which can @@ -654,7 +725,7 @@ static void celt_decode_lost(CELTDecoder * OPUS_RESTRICT st, int N, int LM) for (i=0;irng, st->arch); + len*(8<rng, 0, + st->arch, st->disable_inv); if (anti_collapse_rsv > 0) { @@ -1234,6 +1306,26 @@ int opus_custom_decoder_ctl(CELTDecoder * OPUS_RESTRICT st, int request, ...) *value=st->rng; } break; + case OPUS_SET_PHASE_INVERSION_DISABLED_REQUEST: + { + opus_int32 value = va_arg(ap, opus_int32); + if(value<0 || value>1) + { + goto bad_arg; + } + st->disable_inv = value; + } + break; + case OPUS_GET_PHASE_INVERSION_DISABLED_REQUEST: + { + opus_int32 *value = va_arg(ap, opus_int32*); + if (!value) + { + goto bad_arg; + } + *value = st->disable_inv; + } + break; default: goto bad_request; } diff --git a/code/opus-1.1.4/celt/celt_encoder.c b/code/opus-1.1.4/celt/celt_encoder.c index 3ee7a4d3..053e5a3b 100644 --- a/code/opus-1.1.4/celt/celt_encoder.c +++ b/code/opus-1.1.4/celt/celt_encoder.c @@ -73,8 +73,8 @@ struct OpusCustomEncoder { int constrained_vbr; /* If zero, VBR can do whatever it likes with the rate */ int loss_rate; int lsb_depth; - int variable_duration; int lfe; + int disable_inv; int arch; /* Everything beyond this point gets cleared on a reset */ @@ -98,6 +98,7 @@ struct OpusCustomEncoder { #endif int consec_transient; AnalysisInfo analysis; + SILKInfo silk_info; opus_val32 preemph_memE[2]; opus_val32 preemph_memD[2]; @@ -123,6 +124,7 @@ struct OpusCustomEncoder { /* opus_val16 oldBandE[], Size = channels*mode->nbEBands */ /* opus_val16 oldLogE[], Size = channels*mode->nbEBands */ /* opus_val16 oldLogE2[], Size = channels*mode->nbEBands */ + /* opus_val16 energyError[], Size = channels*mode->nbEBands */ }; int celt_encoder_get_size(int channels) @@ -136,9 +138,10 @@ OPUS_CUSTOM_NOSTATIC int opus_custom_encoder_get_size(const CELTMode *mode, int int size = sizeof(struct CELTEncoder) + (channels*mode->overlap-1)*sizeof(celt_sig) /* celt_sig in_mem[channels*mode->overlap]; */ + channels*COMBFILTER_MAXPERIOD*sizeof(celt_sig) /* celt_sig prefilter_mem[channels*COMBFILTER_MAXPERIOD]; */ - + 3*channels*mode->nbEBands*sizeof(opus_val16); /* opus_val16 oldBandE[channels*mode->nbEBands]; */ + + 4*channels*mode->nbEBands*sizeof(opus_val16); /* opus_val16 oldBandE[channels*mode->nbEBands]; */ /* opus_val16 oldLogE[channels*mode->nbEBands]; */ /* opus_val16 oldLogE2[channels*mode->nbEBands]; */ + /* opus_val16 energyError[channels*mode->nbEBands]; */ return size; } @@ -178,7 +181,6 @@ static int opus_custom_encoder_init_arch(CELTEncoder *st, const CELTMode *mode, st->start = 0; st->end = st->mode->effEBands; st->signalling = 1; - st->arch = arch; st->constrained_vbr = 1; @@ -223,7 +225,8 @@ void opus_custom_encoder_destroy(CELTEncoder *st) static int transient_analysis(const opus_val32 * OPUS_RESTRICT in, int len, int C, - opus_val16 *tf_estimate, int *tf_chan) + opus_val16 *tf_estimate, int *tf_chan, int allow_weak_transients, + int *weak_transient) { int i; VARDECL(opus_val16, tmp); @@ -233,6 +236,12 @@ static int transient_analysis(const opus_val32 * OPUS_RESTRICT in, int len, int int c; opus_val16 tf_max; int len2; + /* Forward masking: 6.7 dB/ms. */ +#ifdef FIXED_POINT + int forward_shift = 4; +#else + opus_val16 forward_decay = QCONST16(.0625f,15); +#endif /* Table of 6*64/x, trained on real data to minimize the average error */ static const unsigned char inv_table[128] = { 255,255,156,110, 86, 70, 59, 51, 45, 40, 37, 33, 31, 28, 26, 25, @@ -247,6 +256,19 @@ static int transient_analysis(const opus_val32 * OPUS_RESTRICT in, int len, int SAVE_STACK; ALLOC(tmp, len, opus_val16); + *weak_transient = 0; + /* For lower bitrates, let's be more conservative and have a forward masking + decay of 3.3 dB/ms. This avoids having to code transients at very low + bitrate (mostly for hybrid), which can result in unstable energy and/or + partial collapse. */ + if (allow_weak_transients) + { +#ifdef FIXED_POINT + forward_shift = 5; +#else + forward_decay = QCONST16(.03125f,15); +#endif + } len2=len/2; for (c=0;c=0;i--) { + /* Backward masking: 13.9 dB/ms. */ #ifdef FIXED_POINT /* FIXME: Use PSHR16() instead */ tmp[i] = mem0 + PSHR32(tmp[i]-mem0,3); @@ -359,7 +382,12 @@ static int transient_analysis(const opus_val32 * OPUS_RESTRICT in, int len, int } } is_transient = mask_metric>200; - + /* For low bitrates, define "weak transients" that need to be + handled differently to avoid partial collapse. */ + if (allow_weak_transients && is_transient && mask_metric<600) { + is_transient = 0; + *weak_transient = 1; + } /* Arbitrary metric for VBR boost */ tf_max = MAX16(0,celt_sqrt(27*mask_metric)-42); /* *tf_estimate = 1 + MIN16(1, sqrt(MAX16(0, tf_max-30))/20); */ @@ -549,7 +577,7 @@ static opus_val32 l1_metric(const celt_norm *tmp, int N, int LM, opus_val16 bias static int tf_analysis(const CELTMode *m, int len, int isTransient, int *tf_res, int lambda, celt_norm *X, int N0, int LM, - int *tf_sum, opus_val16 tf_estimate, int tf_chan) + opus_val16 tf_estimate, int tf_chan) { int i; VARDECL(int, metric); @@ -574,7 +602,6 @@ static int tf_analysis(const CELTMode *m, int len, int isTransient, ALLOC(path0, len, int); ALLOC(path1, len, int); - *tf_sum = 0; for (i=0;i> 10; + trim = QCONST16(4.f, 8) + QCONST16(1.f/16.f, 8)*frac; + } if (C==2) { opus_val16 sum = 0; /* Q10 */ @@ -809,7 +843,7 @@ static int alloc_trim_analysis(const CELTMode *m, const celt_norm *X, } while (++c=12) follower[i] = HALF16(follower[i]); + } +#ifdef DISABLE_FLOAT_API + (void)analysis; +#else + if (analysis->valid) + { + for (i=start;ileak_boost[i]; + } +#endif + for (i=start;i>BITRES>>3 > effectiveBytes/4) + && (tot_boost+boost_bits)>>BITRES>>3 > 2*effectiveBytes/3) { - opus_int32 cap = ((effectiveBytes/4)<tonality-.15f)-0.09f; + tonal = MAX16(0.f,analysis->tonality-.15f)-0.12f; tonal_target = target + (opus_int32)((coded_bins<channels; @@ -1343,7 +1378,6 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, int end; int effEnd; int codedBands; - int tf_sum; int alloc_trim; int pitch_index=COMBFILTER_MINPERIOD; opus_val16 gain1 = 0; @@ -1355,6 +1389,7 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, opus_int32 total_boost; opus_int32 balance; opus_int32 tell; + opus_int32 tell0_frac; int prefilter_tapset=0; int pf_on; int anti_collapse_rsv; @@ -1376,7 +1411,9 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, opus_val16 surround_masking=0; opus_val16 temporal_vbr=0; opus_val16 surround_trim = 0; - opus_int32 equiv_rate = 510000; + opus_int32 equiv_rate; + int hybrid; + int weak_transient = 0; VARDECL(opus_val16, surround_dynalloc); ALLOC_STACK; @@ -1386,6 +1423,7 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, eBands = mode->eBands; start = st->start; end = st->end; + hybrid = start != 0; tf_estimate = 0; if (nbCompressedBytes<2 || pcm==NULL) { @@ -1409,12 +1447,14 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, oldBandE = (opus_val16*)(st->in_mem+CC*(overlap+COMBFILTER_MAXPERIOD)); oldLogE = oldBandE + CC*nbEBands; oldLogE2 = oldLogE + CC*nbEBands; + energyError = oldLogE2 + CC*nbEBands; if (enc==NULL) { - tell=1; + tell0_frac=tell=1; nbFilledBytes=0; } else { + tell0_frac=tell=ec_tell_frac(enc); tell=ec_tell(enc); nbFilledBytes=(tell+4)>>3; } @@ -1467,10 +1507,11 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, if (st->bitrate!=OPUS_BITRATE_MAX) nbCompressedBytes = IMAX(2, IMIN(nbCompressedBytes, (tmp+4*mode->Fs)/(8*mode->Fs)-!!st->signalling)); - effectiveBytes = nbCompressedBytes; + effectiveBytes = nbCompressedBytes - nbFilledBytes; } + equiv_rate = ((opus_int32)nbCompressedBytes*8*50 >> (3-LM)) - (40*C+20)*((400>>LM) - 50); if (st->bitrate != OPUS_BITRATE_MAX) - equiv_rate = st->bitrate - (40*C+20)*((400>>LM) - 50); + equiv_rate = IMIN(equiv_rate, st->bitrate - (40*C+20)*((400>>LM) - 50)); if (enc==NULL) { @@ -1558,8 +1599,8 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, { int enabled; int qg; - enabled = ((st->lfe&&nbAvailableBytes>3) || nbAvailableBytes>12*C) && start==0 && !silence && !st->disable_pf - && st->complexity >= 5 && !(st->consec_transient && LM!=3 && st->variable_duration==OPUS_FRAMESIZE_VARIABLE); + enabled = ((st->lfe&&nbAvailableBytes>3) || nbAvailableBytes>12*C) && !hybrid && !silence && !st->disable_pf + && st->complexity >= 5; prefilter_tapset = st->tapset_decision; pf_on = run_prefilter(st, in, prefilter_mem, CC, N, prefilter_tapset, &pitch_index, &gain1, &qg, enabled, nbAvailableBytes); @@ -1568,7 +1609,7 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, pitch_change = 1; if (pf_on==0) { - if(start==0 && tell+16<=total_bits) + if(!hybrid && tell+16<=total_bits) ec_enc_bit_logp(enc, 0, 1); } else { /*This block is not gated by a total bits check only because @@ -1589,8 +1630,12 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, shortBlocks = 0; if (st->complexity >= 1 && !st->lfe) { + /* Reduces the likelihood of energy instability on fricatives at low bitrate + in hybrid mode. It seems like we still want to have real transients on vowels + though (small SILK quantization offset value). */ + int allow_weak_transients = hybrid && effectiveBytes<15 && st->silk_info.offset >= 100; isTransient = transient_analysis(in, N+overlap, CC, - &tf_estimate, &tf_chan); + &tf_estimate, &tf_chan, allow_weak_transients, &weak_transient); } if (LM>0 && ec_tell(enc)+3<=total_bits) { @@ -1610,7 +1655,7 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, if (secondMdct) { compute_mdcts(mode, 0, in, freq, C, CC, LM, st->upsample, st->arch); - compute_band_energies(mode, freq, bandE, effEnd, C, LM); + compute_band_energies(mode, freq, bandE, effEnd, C, LM, st->arch); amp2Log2(mode, effEnd, end, bandE, bandLogE2, C); for (i=0;iupsample, st->arch); if (CC==2&&C==1) tf_chan = 0; - compute_band_energies(mode, freq, bandE, effEnd, C, LM); + compute_band_energies(mode, freq, bandE, effEnd, C, LM, st->arch); if (st->lfe) { @@ -1634,7 +1679,7 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, ALLOC(surround_dynalloc, C*nbEBands, opus_val16); OPUS_CLEAR(surround_dynalloc, end); /* This computes how much masking takes place between surround channels */ - if (start==0&&st->energy_mask&&!st->lfe) + if (!hybrid&&st->energy_mask&&!st->lfe) { int mask_end; int midband; @@ -1736,14 +1781,14 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, /* Last chance to catch any transient we might have missed in the time-domain analysis */ - if (LM>0 && ec_tell(enc)+3<=total_bits && !isTransient && st->complexity>=5 && !st->lfe) + if (LM>0 && ec_tell(enc)+3<=total_bits && !isTransient && st->complexity>=5 && !st->lfe && !hybrid) { if (patch_transient_decision(bandLogE, oldBandE, nbEBands, start, end, C)) { isTransient = 1; shortBlocks = M; compute_mdcts(mode, shortBlocks, in, freq, C, CC, LM, st->upsample, st->arch); - compute_band_energies(mode, freq, bandE, effEnd, C, LM); + compute_band_energies(mode, freq, bandE, effEnd, C, LM, st->arch); amp2Log2(mode, effEnd, end, bandE, bandLogE, C); /* Compensate for the scaling of short vs long mdcts */ for (i=0;i=15*C && start==0 && st->complexity>=2 && !st->lfe) + if (effectiveBytes>=15*C && !hybrid && st->complexity>=2 && !st->lfe) { int lambda; - if (effectiveBytes<40) - lambda = 12; - else if (effectiveBytes<60) - lambda = 6; - else if (effectiveBytes<100) - lambda = 4; - else - lambda = 3; - lambda*=2; - tf_select = tf_analysis(mode, effEnd, isTransient, tf_res, lambda, X, N, LM, &tf_sum, tf_estimate, tf_chan); + lambda = IMAX(5, 1280/effectiveBytes + 2); + tf_select = tf_analysis(mode, effEnd, isTransient, tf_res, lambda, X, N, LM, tf_estimate, tf_chan); for (i=effEnd;iforce_intra, @@ -1798,7 +1861,15 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, { st->tapset_decision = 0; st->spread_decision = SPREAD_NORMAL; - } else if (shortBlocks || st->complexity < 3 || nbAvailableBytes < 10*C || start != 0) + } else if (hybrid) + { + if (st->complexity == 0) + st->spread_decision = SPREAD_NONE; + else if (isTransient) + st->spread_decision = SPREAD_NORMAL; + else + st->spread_decision = SPREAD_AGGRESSIVE; + } else if (shortBlocks || st->complexity < 3 || nbAvailableBytes < 10*C) { if (st->complexity == 0) st->spread_decision = SPREAD_NONE; @@ -1834,7 +1905,7 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, maxDepth = dynalloc_analysis(bandLogE, bandLogE2, nbEBands, start, end, C, offsets, st->lsb_depth, mode->logN, isTransient, st->vbr, st->constrained_vbr, - eBands, LM, effectiveBytes, &tot_boost, st->lfe, surround_dynalloc); + eBands, LM, effectiveBytes, &tot_boost, st->lfe, surround_dynalloc, &st->analysis); /* For LFE, everything interesting is in the first band */ if (st->lfe) offsets[0] = IMIN(8, effectiveBytes/3); @@ -1896,12 +1967,15 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, alloc_trim = 5; if (tell+(6<lfe) + if (start > 0 || st->lfe) + { + st->stereo_saving = 0; alloc_trim = 5; - else + } else { alloc_trim = alloc_trim_analysis(mode, X, bandLogE, end, LM, C, N, &st->analysis, &st->stereo_saving, tf_estimate, - st->intensity, surround_trim, st->arch); + st->intensity, surround_trim, equiv_rate, st->arch); + } ec_enc_icdf(enc, alloc_trim, trim_icdf, 7); tell = ec_tell_frac(enc); } @@ -1919,17 +1993,36 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, /* Don't attempt to use more than 510 kb/s, even for frames smaller than 20 ms. The CELT allocator will just not be able to use more than that anyway. */ nbCompressedBytes = IMIN(nbCompressedBytes,1275>>(3-LM)); - base_target = vbr_rate - ((40*C+20)<constrained_vbr) base_target += (st->vbr_offset>>lm_diff); - target = compute_vbr(mode, &st->analysis, base_target, LM, equiv_rate, + if (!hybrid) + { + target = compute_vbr(mode, &st->analysis, base_target, LM, equiv_rate, st->lastCodedBands, C, st->intensity, st->constrained_vbr, st->stereo_saving, tot_boost, tf_estimate, pitch_change, maxDepth, - st->variable_duration, st->lfe, st->energy_mask!=NULL, surround_masking, + st->lfe, st->energy_mask!=NULL, surround_masking, temporal_vbr); - + } else { + target = base_target; + /* Tonal frames (offset<100) need more bits than noisy (offset>100) ones. */ + if (st->silk_info.offset < 100) target += 12 << BITRES >> (3-LM); + if (st->silk_info.offset > 100) target -= 18 << BITRES >> (3-LM); + /* Boosting bitrate on transients and vowels with significant temporal + spikes. */ + target += (opus_int32)MULT16_16_Q14(tf_estimate-QCONST16(.25f,14), (50< QCONST16(.7f,14)) + target = IMAX(target, 50<>(BITRES+3)) + 2 - nbFilledBytes; + min_allowed = ((tell+total_boost+(1<<(BITRES+3))-1)>>(BITRES+3)) + 2; + /* Take into account the 37 bits we need to have left in the packet to + signal a redundant frame in hybrid mode. Creating a shorter packet would + create an entropy coder desync. */ + if (hybrid) + min_allowed = IMAX(min_allowed, (tell0_frac+(37<>(BITRES+3)); nbAvailableBytes = (target+(1<<(BITRES+2)))>>(BITRES+3); nbAvailableBytes = IMAX(min_allowed,nbAvailableBytes); - nbAvailableBytes = IMIN(nbCompressedBytes,nbAvailableBytes+nbFilledBytes) - nbFilledBytes; + nbAvailableBytes = IMIN(nbCompressedBytes,nbAvailableBytes); /* By how much did we "miss" the target on that frame */ delta = target - vbr_rate; @@ -1988,7 +2086,7 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, st->vbr_reservoir = 0; /*printf ("+%d\n", adjust);*/ } - nbCompressedBytes = IMIN(nbCompressedBytes,nbAvailableBytes+nbFilledBytes); + nbCompressedBytes = IMIN(nbCompressedBytes,nbAvailableBytes); /*printf("%d\n", nbCompressedBytes*50*8);*/ /* This moves the raw bits to take into account the new compressed size */ ec_enc_shrink(enc, nbCompressedBytes); @@ -2038,7 +2136,7 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, quant_all_bands(1, mode, start, end, X, C==2 ? X+N : NULL, collapse_masks, bandE, pulses, shortBlocks, st->spread_decision, dual_stereo, st->intensity, tf_res, nbCompressedBytes*(8<rng, st->arch); + balance, enc, LM, codedBands, &st->rng, st->complexity, st->arch, st->disable_inv); if (anti_collapse_rsv > 0) { @@ -2049,6 +2147,14 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, ec_enc_bits(enc, anti_collapse_on, 1); } quant_energy_finalise(mode, start, end, oldBandE, error, fine_quant, fine_priority, nbCompressedBytes*8-ec_tell(enc), enc, C); + OPUS_CLEAR(energyError, nbEBands*CC); + c=0; + do { + for (i=start;ilsb_depth; } break; - case OPUS_SET_EXPERT_FRAME_DURATION_REQUEST: + case OPUS_SET_PHASE_INVERSION_DISABLED_REQUEST: { opus_int32 value = va_arg(ap, opus_int32); - st->variable_duration = value; + if(value<0 || value>1) + { + goto bad_arg; + } + st->disable_inv = value; + } + break; + case OPUS_GET_PHASE_INVERSION_DISABLED_REQUEST: + { + opus_int32 *value = va_arg(ap, opus_int32*); + if (!value) + { + goto bad_arg; + } + *value = st->disable_inv; } break; case OPUS_RESET_STATE: @@ -2368,6 +2488,13 @@ int opus_custom_encoder_ctl(CELTEncoder * OPUS_RESTRICT st, int request, ...) OPUS_COPY(&st->analysis, info, 1); } break; + case CELT_SET_SILK_INFO_REQUEST: + { + SILKInfo *info = va_arg(ap, SILKInfo *); + if (info) + OPUS_COPY(&st->silk_info, info, 1); + } + break; case CELT_GET_MODE_REQUEST: { const CELTMode ** value = va_arg(ap, const CELTMode**); diff --git a/code/opus-1.1.4/celt/celt_lpc.c b/code/opus-1.1.4/celt/celt_lpc.c index b410a21c..57bc5f3d 100644 --- a/code/opus-1.1.4/celt/celt_lpc.c +++ b/code/opus-1.1.4/celt/celt_lpc.c @@ -89,58 +89,40 @@ int p void celt_fir_c( - const opus_val16 *_x, + const opus_val16 *x, const opus_val16 *num, - opus_val16 *_y, + opus_val16 *y, int N, int ord, - opus_val16 *mem, int arch) { int i,j; VARDECL(opus_val16, rnum); - VARDECL(opus_val16, x); SAVE_STACK; ALLOC(rnum, ord, opus_val16); - ALLOC(x, N+ord, opus_val16); for(i=0;i0) ? SHR32(a, shift) : SHL32(a, -(shift))) #define ROUND16(x,a) (celt_mips--,EXTRACT16(PSHR32((x),(a)))) +#define SROUND16(x,a) (celt_mips--,EXTRACT16(SATURATE(PSHR32(x,a), 32767))); + #define HALF16(x) (SHR16(x,1)) #define HALF32(x) (SHR32(x,1)) -//#define SHR(a,shift) ((a) >> (shift)) -//#define SHL(a,shift) ((a) << (shift)) - #define ADD16(a, b) ADD16_(a, b, __FILE__, __LINE__) static OPUS_INLINE short ADD16_(int a, int b, char *file, int line) { diff --git a/code/opus-1.1.4/celt/fixed_generic.h b/code/opus-1.1.4/celt/fixed_generic.h index 1cfd6d69..5f4abda7 100644 --- a/code/opus-1.1.4/celt/fixed_generic.h +++ b/code/opus-1.1.4/celt/fixed_generic.h @@ -104,6 +104,9 @@ /** Shift by a and round-to-neareast 32-bit value. Result is a 16-bit value */ #define ROUND16(x,a) (EXTRACT16(PSHR32((x),(a)))) +/** Shift by a and round-to-neareast 32-bit value. Result is a saturated 16-bit value */ +#define SROUND16(x,a) EXTRACT16(SATURATE(PSHR32(x,a), 32767)); + /** Divide by two */ #define HALF16(x) (SHR16(x,1)) #define HALF32(x) (SHR32(x,1)) @@ -117,6 +120,14 @@ /** Subtract two 32-bit values */ #define SUB32(a,b) ((opus_val32)(a)-(opus_val32)(b)) +/** Add two 32-bit values, ignore any overflows */ +#define ADD32_ovflw(a,b) ((opus_val32)((opus_uint32)(a)+(opus_uint32)(b))) +/** Subtract two 32-bit values, ignore any overflows */ +#define SUB32_ovflw(a,b) ((opus_val32)((opus_uint32)(a)-(opus_uint32)(b))) +/* Avoid MSVC warning C4146: unary minus operator applied to unsigned type */ +/** Negate 32-bit value, ignore any overflows */ +#define NEG32_ovflw(a) ((opus_val32)(0-(opus_uint32)(a))) + /** 16x16 multiplication where the result fits in 16 bits */ #define MULT16_16_16(a,b) ((((opus_val16)(a))*((opus_val16)(b)))) diff --git a/code/opus-1.1.4/celt/float_cast.h b/code/opus-1.1.4/celt/float_cast.h index ed5a39b5..98b40abc 100644 --- a/code/opus-1.1.4/celt/float_cast.h +++ b/code/opus-1.1.4/celt/float_cast.h @@ -61,7 +61,13 @@ ** the config.h file. */ -#if (HAVE_LRINTF) +/* With GCC, when SSE is available, the fastest conversion is cvtss2si. */ +#if defined(__GNUC__) && defined(__SSE__) + +#include +static OPUS_INLINE opus_int32 float2int(float x) {return _mm_cvt_ss2si(_mm_set_ss(x));} + +#elif defined(HAVE_LRINTF) /* These defines enable functionality introduced with the 1999 ISO C ** standard. They must be defined before the inclusion of math.h to diff --git a/code/opus-1.1.4/celt/kiss_fft.c b/code/opus-1.1.4/celt/kiss_fft.c index 1f8fd053..83775165 100644 --- a/code/opus-1.1.4/celt/kiss_fft.c +++ b/code/opus-1.1.4/celt/kiss_fft.c @@ -82,8 +82,8 @@ static void kf_bfly2( C_SUB( Fout2[0] , Fout[0] , t ); C_ADDTO( Fout[0] , t ); - t.r = S_MUL(Fout2[1].r+Fout2[1].i, tw); - t.i = S_MUL(Fout2[1].i-Fout2[1].r, tw); + t.r = S_MUL(ADD32_ovflw(Fout2[1].r, Fout2[1].i), tw); + t.i = S_MUL(SUB32_ovflw(Fout2[1].i, Fout2[1].r), tw); C_SUB( Fout2[1] , Fout[1] , t ); C_ADDTO( Fout[1] , t ); @@ -92,8 +92,8 @@ static void kf_bfly2( C_SUB( Fout2[2] , Fout[2] , t ); C_ADDTO( Fout[2] , t ); - t.r = S_MUL(Fout2[3].i-Fout2[3].r, tw); - t.i = S_MUL(-Fout2[3].i-Fout2[3].r, tw); + t.r = S_MUL(SUB32_ovflw(Fout2[3].i, Fout2[3].r), tw); + t.i = S_MUL(NEG32_ovflw(ADD32_ovflw(Fout2[3].i, Fout2[3].r)), tw); C_SUB( Fout2[3] , Fout[3] , t ); C_ADDTO( Fout[3] , t ); Fout += 8; @@ -126,10 +126,10 @@ static void kf_bfly4( C_ADDTO( *Fout , scratch1 ); C_SUB( scratch1 , Fout[1] , Fout[3] ); - Fout[1].r = scratch0.r + scratch1.i; - Fout[1].i = scratch0.i - scratch1.r; - Fout[3].r = scratch0.r - scratch1.i; - Fout[3].i = scratch0.i + scratch1.r; + Fout[1].r = ADD32_ovflw(scratch0.r, scratch1.i); + Fout[1].i = SUB32_ovflw(scratch0.i, scratch1.r); + Fout[3].r = SUB32_ovflw(scratch0.r, scratch1.i); + Fout[3].i = ADD32_ovflw(scratch0.i, scratch1.r); Fout+=4; } } else { @@ -160,10 +160,10 @@ static void kf_bfly4( tw3 += fstride*3; C_ADDTO( *Fout , scratch[3] ); - Fout[m].r = scratch[5].r + scratch[4].i; - Fout[m].i = scratch[5].i - scratch[4].r; - Fout[m3].r = scratch[5].r - scratch[4].i; - Fout[m3].i = scratch[5].i + scratch[4].r; + Fout[m].r = ADD32_ovflw(scratch[5].r, scratch[4].i); + Fout[m].i = SUB32_ovflw(scratch[5].i, scratch[4].r); + Fout[m3].r = SUB32_ovflw(scratch[5].r, scratch[4].i); + Fout[m3].i = ADD32_ovflw(scratch[5].i, scratch[4].r); ++Fout; } } @@ -212,18 +212,18 @@ static void kf_bfly3( tw1 += fstride; tw2 += fstride*2; - Fout[m].r = Fout->r - HALF_OF(scratch[3].r); - Fout[m].i = Fout->i - HALF_OF(scratch[3].i); + Fout[m].r = SUB32_ovflw(Fout->r, HALF_OF(scratch[3].r)); + Fout[m].i = SUB32_ovflw(Fout->i, HALF_OF(scratch[3].i)); C_MULBYSCALAR( scratch[0] , epi3.i ); C_ADDTO(*Fout,scratch[3]); - Fout[m2].r = Fout[m].r + scratch[0].i; - Fout[m2].i = Fout[m].i - scratch[0].r; + Fout[m2].r = ADD32_ovflw(Fout[m].r, scratch[0].i); + Fout[m2].i = SUB32_ovflw(Fout[m].i, scratch[0].r); - Fout[m].r -= scratch[0].i; - Fout[m].i += scratch[0].r; + Fout[m].r = SUB32_ovflw(Fout[m].r, scratch[0].i); + Fout[m].i = ADD32_ovflw(Fout[m].i, scratch[0].r); ++Fout; } while(--k); @@ -282,22 +282,22 @@ static void kf_bfly5( C_ADD( scratch[8],scratch[2],scratch[3]); C_SUB( scratch[9],scratch[2],scratch[3]); - Fout0->r += scratch[7].r + scratch[8].r; - Fout0->i += scratch[7].i + scratch[8].i; + Fout0->r = ADD32_ovflw(Fout0->r, ADD32_ovflw(scratch[7].r, scratch[8].r)); + Fout0->i = ADD32_ovflw(Fout0->i, ADD32_ovflw(scratch[7].i, scratch[8].i)); - scratch[5].r = scratch[0].r + S_MUL(scratch[7].r,ya.r) + S_MUL(scratch[8].r,yb.r); - scratch[5].i = scratch[0].i + S_MUL(scratch[7].i,ya.r) + S_MUL(scratch[8].i,yb.r); + scratch[5].r = ADD32_ovflw(scratch[0].r, ADD32_ovflw(S_MUL(scratch[7].r,ya.r), S_MUL(scratch[8].r,yb.r))); + scratch[5].i = ADD32_ovflw(scratch[0].i, ADD32_ovflw(S_MUL(scratch[7].i,ya.r), S_MUL(scratch[8].i,yb.r))); - scratch[6].r = S_MUL(scratch[10].i,ya.i) + S_MUL(scratch[9].i,yb.i); - scratch[6].i = -S_MUL(scratch[10].r,ya.i) - S_MUL(scratch[9].r,yb.i); + scratch[6].r = ADD32_ovflw(S_MUL(scratch[10].i,ya.i), S_MUL(scratch[9].i,yb.i)); + scratch[6].i = NEG32_ovflw(ADD32_ovflw(S_MUL(scratch[10].r,ya.i), S_MUL(scratch[9].r,yb.i))); C_SUB(*Fout1,scratch[5],scratch[6]); C_ADD(*Fout4,scratch[5],scratch[6]); - scratch[11].r = scratch[0].r + S_MUL(scratch[7].r,yb.r) + S_MUL(scratch[8].r,ya.r); - scratch[11].i = scratch[0].i + S_MUL(scratch[7].i,yb.r) + S_MUL(scratch[8].i,ya.r); - scratch[12].r = - S_MUL(scratch[10].i,yb.i) + S_MUL(scratch[9].i,ya.i); - scratch[12].i = S_MUL(scratch[10].r,yb.i) - S_MUL(scratch[9].r,ya.i); + scratch[11].r = ADD32_ovflw(scratch[0].r, ADD32_ovflw(S_MUL(scratch[7].r,yb.r), S_MUL(scratch[8].r,ya.r))); + scratch[11].i = ADD32_ovflw(scratch[0].i, ADD32_ovflw(S_MUL(scratch[7].i,yb.r), S_MUL(scratch[8].i,ya.r))); + scratch[12].r = SUB32_ovflw(S_MUL(scratch[9].i,ya.i), S_MUL(scratch[10].i,yb.i)); + scratch[12].i = SUB32_ovflw(S_MUL(scratch[10].r,yb.i), S_MUL(scratch[9].r,ya.i)); C_ADD(*Fout2,scratch[11],scratch[12]); C_SUB(*Fout3,scratch[11],scratch[12]); diff --git a/code/opus-1.1.4/celt/mathops.h b/code/opus-1.1.4/celt/mathops.h index a0525a96..1f8a20cb 100644 --- a/code/opus-1.1.4/celt/mathops.h +++ b/code/opus-1.1.4/celt/mathops.h @@ -38,11 +38,44 @@ #include "entcode.h" #include "os_support.h" +#define PI 3.141592653f + /* Multiplies two 16-bit fractional values. Bit-exactness of this macro is important */ #define FRAC_MUL16(a,b) ((16384+((opus_int32)(opus_int16)(a)*(opus_int16)(b)))>>15) unsigned isqrt32(opus_uint32 _val); +/* CELT doesn't need it for fixed-point, by analysis.c does. */ +#if !defined(FIXED_POINT) || defined(ANALYSIS_C) +#define cA 0.43157974f +#define cB 0.67848403f +#define cC 0.08595542f +#define cE ((float)PI/2) +static OPUS_INLINE float fast_atan2f(float y, float x) { + float x2, y2; + x2 = x*x; + y2 = y*y; + /* For very small values, we don't care about the answer, so + we can just return 0. */ + if (x2 + y2 < 1e-18f) + { + return 0; + } + if(x2eBands); - opus_free((opus_int16*)mode->allocVectors); + opus_free((unsigned char*)mode->allocVectors); opus_free((opus_val16*)mode->window); opus_free((opus_int16*)mode->logN); diff --git a/code/opus-1.1.4/celt/pitch.c b/code/opus-1.1.4/celt/pitch.c index bf46e7d5..5b061d86 100644 --- a/code/opus-1.1.4/celt/pitch.c +++ b/code/opus-1.1.4/celt/pitch.c @@ -220,13 +220,8 @@ opus_val32 #else void #endif -#if defined(OVERRIDE_PITCH_XCORR) celt_pitch_xcorr_c(const opus_val16 *_x, const opus_val16 *_y, - opus_val32 *xcorr, int len, int max_pitch) -#else -celt_pitch_xcorr(const opus_val16 *_x, const opus_val16 *_y, opus_val32 *xcorr, int len, int max_pitch, int arch) -#endif { #if 0 /* This is a simple version of the pitch correlation that should work @@ -265,11 +260,7 @@ celt_pitch_xcorr(const opus_val16 *_x, const opus_val16 *_y, for (i=0;i>1;j++) sum += SHR32(MULT16_16(x_lp[j],y[i+j]), shift); #else - sum = celt_inner_prod_c(x_lp, y+i, len>>1); + sum = celt_inner_prod(x_lp, y+i, len>>1, arch); #endif xcorr[i] = MAX32(-1, sum); #ifdef FIXED_POINT @@ -424,7 +411,7 @@ static opus_val16 compute_pitch_gain(opus_val32 xy, opus_val32 xx, opus_val32 yy sx = celt_ilog2(xx)-14; sy = celt_ilog2(yy)-14; shift = sx + sy; - x2y2 = MULT16_16_Q14(VSHR32(xx, sx), VSHR32(yy, sy)); + x2y2 = SHR32(MULT16_16(VSHR32(xx, sx), VSHR32(yy, sy)), 14); if (shift & 1) { if (x2y2 < 32768) { diff --git a/code/opus-1.1.4/celt/pitch.h b/code/opus-1.1.4/celt/pitch.h index d3503532..e425f56a 100644 --- a/code/opus-1.1.4/celt/pitch.h +++ b/code/opus-1.1.4/celt/pitch.h @@ -46,8 +46,7 @@ #include "mips/pitch_mipsr1.h" #endif -#if ((defined(OPUS_ARM_ASM) && defined(FIXED_POINT)) \ - || defined(OPUS_ARM_MAY_HAVE_NEON_INTR)) +#if (defined(OPUS_ARM_ASM) || defined(OPUS_ARM_MAY_HAVE_NEON_INTR)) # include "arm/pitch_arm.h" #endif @@ -184,17 +183,10 @@ opus_val32 void #endif celt_pitch_xcorr_c(const opus_val16 *_x, const opus_val16 *_y, - opus_val32 *xcorr, int len, int max_pitch); - -#if !defined(OVERRIDE_PITCH_XCORR) -#ifdef FIXED_POINT -opus_val32 -#else -void -#endif -celt_pitch_xcorr(const opus_val16 *_x, const opus_val16 *_y, opus_val32 *xcorr, int len, int max_pitch, int arch); +#ifndef OVERRIDE_PITCH_XCORR +# define celt_pitch_xcorr celt_pitch_xcorr_c #endif #endif diff --git a/code/opus-1.1.4/celt/quant_bands.c b/code/opus-1.1.4/celt/quant_bands.c index 95076e0a..56101b1c 100644 --- a/code/opus-1.1.4/celt/quant_bands.c +++ b/code/opus-1.1.4/celt/quant_bands.c @@ -418,6 +418,7 @@ void quant_energy_finalise(const CELTMode *m, int start, int end, opus_val16 *ol offset = (q2-.5f)*(1<<(14-fine_quant[i]-1))*(1.f/16384); #endif oldEBands[i+c*m->nbEBands] += offset; + error[i+c*m->nbEBands] -= offset; bits_left--; } while (++c < C); } @@ -547,9 +548,15 @@ void amp2Log2(const CELTMode *m, int effEnd, int end, c=0; do { for (i=0;inbEBands] = - celt_log2(SHL32(bandE[i+c*m->nbEBands],2)) + celt_log2(bandE[i+c*m->nbEBands]) - SHL16((opus_val16)eMeans[i],6); +#ifdef FIXED_POINT + /* Compensate for bandE[] being Q12 but celt_log2() taking a Q14 input. */ + bandLogE[i+c*m->nbEBands] += QCONST16(2.f, DB_SHIFT); +#endif + } for (i=effEnd;inbEBands+i] = -QCONST16(14.f,DB_SHIFT); } while (++c < C); diff --git a/code/opus-1.1.4/celt/rate.c b/code/opus-1.1.4/celt/rate.c index 7dfa5be8..ca4cc870 100644 --- a/code/opus-1.1.4/celt/rate.c +++ b/code/opus-1.1.4/celt/rate.c @@ -348,12 +348,17 @@ static OPUS_INLINE int interp_bits2pulses(const CELTMode *m, int start, int end, /*This if() block is the only part of the allocation function that is not a mandatory part of the bitstream: any bands we choose to skip here must be explicitly signaled.*/ - /*Choose a threshold with some hysteresis to keep bands from - fluctuating in and out.*/ + int depth_threshold; + /*We choose a threshold with some hysteresis to keep bands from + fluctuating in and out, but we try not to fold below a certain point. */ + if (codedBands > 17) + depth_threshold = j ((j>4 && j<=signalBandwidth)) + if (codedBands<=start+2 || (band_bits > (depth_threshold*band_width<>4 && j<=signalBandwidth)) #endif { ec_enc_bit_logp(ec, 1, 1); diff --git a/code/opus-1.1.4/celt/static_modes_fixed_arm_ne10.h b/code/opus-1.1.4/celt/static_modes_fixed_arm_ne10.h index b8ef0cee..76230921 100644 --- a/code/opus-1.1.4/celt/static_modes_fixed_arm_ne10.h +++ b/code/opus-1.1.4/celt/static_modes_fixed_arm_ne10.h @@ -1,7 +1,7 @@ /* The contents of this file was automatically generated by * dump_mode_arm_ne10.c with arguments: 48000 960 * It contains static definitions for some pre-defined modes. */ -#include +#include #ifndef NE10_FFT_PARAMS48000_960 #define NE10_FFT_PARAMS48000_960 diff --git a/code/opus-1.1.4/celt/static_modes_float_arm_ne10.h b/code/opus-1.1.4/celt/static_modes_float_arm_ne10.h index 934a82a4..66e1abb1 100644 --- a/code/opus-1.1.4/celt/static_modes_float_arm_ne10.h +++ b/code/opus-1.1.4/celt/static_modes_float_arm_ne10.h @@ -1,7 +1,7 @@ /* The contents of this file was automatically generated by * dump_mode_arm_ne10.c with arguments: 48000 960 * It contains static definitions for some pre-defined modes. */ -#include +#include #ifndef NE10_FFT_PARAMS48000_960 #define NE10_FFT_PARAMS48000_960 diff --git a/code/opus-1.1.4/celt/vq.c b/code/opus-1.1.4/celt/vq.c index d29f38fd..8ef80e50 100644 --- a/code/opus-1.1.4/celt/vq.c +++ b/code/opus-1.1.4/celt/vq.c @@ -67,7 +67,7 @@ static void exp_rotation1(celt_norm *X, int len, int stride, opus_val16 c, opus_ } #endif /* OVERRIDE_vq_exp_rotation1 */ -static void exp_rotation(celt_norm *X, int len, int dir, int stride, int K, int spread) +void exp_rotation(celt_norm *X, int len, int dir, int stride, int K, int spread) { static const int SPREAD_FACTOR[3]={15,10,5}; int i; @@ -158,42 +158,27 @@ static unsigned extract_collapse_mask(int *iy, int N, int B) return collapse_mask; } -unsigned alg_quant(celt_norm *X, int N, int K, int spread, int B, ec_enc *enc -#ifdef RESYNTH - , opus_val16 gain -#endif - ) +opus_val16 op_pvq_search_c(celt_norm *X, int *iy, int K, int N, int arch) { VARDECL(celt_norm, y); - VARDECL(int, iy); - VARDECL(opus_val16, signx); + VARDECL(int, signx); int i, j; - opus_val16 s; int pulsesLeft; opus_val32 sum; opus_val32 xy; opus_val16 yy; - unsigned collapse_mask; SAVE_STACK; - celt_assert2(K>0, "alg_quant() needs at least one pulse"); - celt_assert2(N>1, "alg_quant() needs at least two dimensions"); - + (void)arch; ALLOC(y, N, celt_norm); - ALLOC(iy, N, int); - ALLOC(signx, N, opus_val16); - - exp_rotation(X, N, 1, B, K, spread); + ALLOC(signx, N, int); /* Get rid of the sign */ sum = 0; j=0; do { - if (X[j]>0) - signx[j]=1; - else { - signx[j]=-1; - X[j]=-X[j]; - } + signx[j] = X[j]<0; + /* OPT: Make sure the compiler doesn't use a branch on ABS16(). */ + X[j] = ABS16(X[j]); iy[j] = 0; y[j] = 0; } while (++j=1, "Allocated too many pulses in the quick pass"); + celt_assert2(pulsesLeft>=0, "Allocated too many pulses in the quick pass"); /* This should never happen, but just in case it does (e.g. on silence) we fill the first bin with pulses. */ @@ -256,12 +246,12 @@ unsigned alg_quant(celt_norm *X, int N, int K, int spread, int B, ec_enc *enc pulsesLeft=0; } - s = 1; for (i=0;i= best_num/best_den, but that way we can do it without any division */ - /* OPT: Make sure to use conditional moves here */ - if (MULT16_16(best_den, Rxy) > MULT16_16(Ryy, best_num)) + /* OPT: It's not clear whether a cmov is faster than a branch here + since the condition is more often false than true and using + a cmov introduces data dependencies across iterations. The optimal + choice may be architecture-dependent. */ + if (opus_unlikely(MULT16_16(best_den, Rxy) > MULT16_16(Ryy, best_num))) { best_den = Ryy; best_num = Rxy; @@ -301,23 +307,47 @@ unsigned alg_quant(celt_norm *X, int N, int K, int spread, int B, ec_enc *enc /* Only now that we've made the final choice, update y/iy */ /* Multiplying y[j] by 2 so we don't have to do it everywhere else */ - y[best_id] += 2*s; + y[best_id] += 2; iy[best_id]++; } /* Put the original sign back */ j=0; do { - X[j] = MULT16_16(signx[j],X[j]); - if (signx[j] < 0) - iy[j] = -iy[j]; + /*iy[j] = signx[j] ? -iy[j] : iy[j];*/ + /* OPT: The is more likely to be compiled without a branch than the code above + but has the same performance otherwise. */ + iy[j] = (iy[j]^-signx[j]) + signx[j]; } while (++j0, "alg_quant() needs at least one pulse"); + celt_assert2(N>1, "alg_quant() needs at least two dimensions"); + + /* Covers vectorization by up to 4. */ + ALLOC(iy, N+3, int); + + exp_rotation(X, N, 1, B, K, spread); + + yy = op_pvq_search(X, iy, K, N, arch); + encode_pulses(iy, N, K, enc); -#ifdef RESYNTH - normalise_residual(iy, X, N, yy, gain); - exp_rotation(X, N, -1, B, K, spread); -#endif + if (resynth) + { + normalise_residual(iy, X, N, yy, gain); + exp_rotation(X, N, -1, B, K, spread); + } collapse_mask = extract_collapse_mask(iy, N, B); RESTORE_STACK; @@ -401,7 +431,7 @@ int stereo_itheta(const celt_norm *X, const celt_norm *Y, int stereo, int N, int /* 0.63662 = 2/pi */ itheta = MULT16_16_Q15(QCONST16(0.63662f,15),celt_atan2p(side, mid)); #else - itheta = (int)floor(.5f+16384*0.63662f*atan2(side,mid)); + itheta = (int)floor(.5f+16384*0.63662f*fast_atan2f(side,mid)); #endif return itheta; diff --git a/code/opus-1.1.4/celt/vq.h b/code/opus-1.1.4/celt/vq.h index 5cfcbe50..0dfe6af0 100644 --- a/code/opus-1.1.4/celt/vq.h +++ b/code/opus-1.1.4/celt/vq.h @@ -37,10 +37,22 @@ #include "entdec.h" #include "modes.h" +#if (defined(OPUS_X86_MAY_HAVE_SSE2) && !defined(FIXED_POINT)) +#include "x86/vq_sse.h" +#endif + #if defined(MIPSr1_ASM) #include "mips/vq_mipsr1.h" #endif +void exp_rotation(celt_norm *X, int len, int dir, int stride, int K, int spread); + +opus_val16 op_pvq_search_c(celt_norm *X, int *iy, int K, int N, int arch); + +#if !defined(OVERRIDE_OP_PVQ_SEARCH) +#define op_pvq_search(x, iy, K, N, arch) \ + (op_pvq_search_c(x, iy, K, N, arch)) +#endif /** Algebraic pulse-vector quantiser. The signal x is replaced by the sum of * the pitch and a combination of pulses such that its norm is still equal @@ -51,12 +63,8 @@ * @param enc Entropy encoder state * @ret A mask indicating which blocks in the band received pulses */ -unsigned alg_quant(celt_norm *X, int N, int K, int spread, int B, - ec_enc *enc -#ifdef RESYNTH - , opus_val16 gain -#endif - ); +unsigned alg_quant(celt_norm *X, int N, int K, int spread, int B, ec_enc *enc, + opus_val16 gain, int resynth, int arch); /** Algebraic pulse decoder * @param X Decoded normalised spectrum (returned) diff --git a/code/opus-1.1.4/celt/x86/celt_lpc_sse.c b/code/opus-1.1.4/celt/x86/celt_lpc_sse.c index 67e5592a..54785688 100644 --- a/code/opus-1.1.4/celt/x86/celt_lpc_sse.c +++ b/code/opus-1.1.4/celt/x86/celt_lpc_sse.c @@ -40,65 +40,23 @@ #if defined(FIXED_POINT) -void celt_fir_sse4_1(const opus_val16 *_x, +void celt_fir_sse4_1(const opus_val16 *x, const opus_val16 *num, - opus_val16 *_y, + opus_val16 *y, int N, int ord, - opus_val16 *mem, int arch) { int i,j; VARDECL(opus_val16, rnum); - VARDECL(opus_val16, x); __m128i vecNoA; opus_int32 noA ; SAVE_STACK; ALLOC(rnum, ord, opus_val16); - ALLOC(x, N+ord, opus_val16); for(i=0;i> 1; vecNoA = _mm_set_epi32(noA, noA, noA, noA); @@ -107,25 +65,24 @@ void celt_fir_sse4_1(const opus_val16 *_x, opus_val32 sums[4] = {0}; __m128i vecSum, vecX; - xcorr_kernel(rnum, x+i, sums, ord, arch); + xcorr_kernel(rnum, x+i-ord, sums, ord, arch); vecSum = _mm_loadu_si128((__m128i *)sums); vecSum = _mm_add_epi32(vecSum, vecNoA); vecSum = _mm_srai_epi32(vecSum, SIG_SHIFT); - vecX = OP_CVTEPI16_EPI32_M64(_x + i); + vecX = OP_CVTEPI16_EPI32_M64(x + i); vecSum = _mm_add_epi32(vecSum, vecX); vecSum = _mm_packs_epi32(vecSum, vecSum); - _mm_storel_epi64((__m128i *)(_y + i), vecSum); + _mm_storel_epi64((__m128i *)(y + i), vecSum); } for (;i +#include +#include "celt_lpc.h" +#include "stack_alloc.h" +#include "mathops.h" +#include "vq.h" +#include "x86cpu.h" + + +#ifndef FIXED_POINT + +opus_val16 op_pvq_search_sse2(celt_norm *_X, int *iy, int K, int N, int arch) +{ + int i, j; + int pulsesLeft; + float xy, yy; + VARDECL(celt_norm, y); + VARDECL(celt_norm, X); + VARDECL(float, signy); + __m128 signmask; + __m128 sums; + __m128i fours; + SAVE_STACK; + + (void)arch; + /* All bits set to zero, except for the sign bit. */ + signmask = _mm_set_ps1(-0.f); + fours = _mm_set_epi32(4, 4, 4, 4); + ALLOC(y, N+3, celt_norm); + ALLOC(X, N+3, celt_norm); + ALLOC(signy, N+3, float); + + OPUS_COPY(X, _X, N); + X[N] = X[N+1] = X[N+2] = 0; + sums = _mm_setzero_ps(); + for (j=0;j (N>>1)) + { + __m128i pulses_sum; + __m128 yy4, xy4; + __m128 rcp4; + opus_val32 sum = _mm_cvtss_f32(sums); + /* If X is too small, just replace it with a pulse at 0 */ + /* Prevents infinities and NaNs from causing too many pulses + to be allocated. 64 is an approximation of infinity here. */ + if (!(sum > EPSILON && sum < 64)) + { + X[0] = QCONST16(1.f,14); + j=1; do + X[j]=0; + while (++j=0, "Allocated too many pulses in the quick pass"); + + /* This should never happen, but just in case it does (e.g. on silence) + we fill the first bin with pulses. */ + if (pulsesLeft > N+3) + { + opus_val16 tmp = (opus_val16)pulsesLeft; + yy = MAC16_16(yy, tmp, tmp); + yy = MAC16_16(yy, tmp, y[0]); + iy[0] += pulsesLeft; + pulsesLeft=0; + } + + for (i=0;iOPUS_FRAMESIZE_20_MS
Use 20 ms frames.
*
OPUS_FRAMESIZE_40_MS
Use 40 ms frames.
*
OPUS_FRAMESIZE_60_MS
Use 60 ms frames.
- *
OPUS_FRAMESIZE_VARIABLE
Optimize the frame size dynamically.
+ *
OPUS_FRAMESIZE_80_MS
Use 80 ms frames.
+ *
OPUS_FRAMESIZE_100_MS
Use 100 ms frames.
+ *
OPUS_FRAMESIZE_120_MS
Use 120 ms frames.
* * @hideinitializer */ #define OPUS_SET_EXPERT_FRAME_DURATION(x) OPUS_SET_EXPERT_FRAME_DURATION_REQUEST, __opus_check_int(x) @@ -581,7 +587,9 @@ extern "C" { *
OPUS_FRAMESIZE_20_MS
Use 20 ms frames.
*
OPUS_FRAMESIZE_40_MS
Use 40 ms frames.
*
OPUS_FRAMESIZE_60_MS
Use 60 ms frames.
- *
OPUS_FRAMESIZE_VARIABLE
Optimize the frame size dynamically.
+ *
OPUS_FRAMESIZE_80_MS
Use 80 ms frames.
+ *
OPUS_FRAMESIZE_100_MS
Use 100 ms frames.
+ *
OPUS_FRAMESIZE_120_MS
Use 120 ms frames.
* * @hideinitializer */ #define OPUS_GET_EXPERT_FRAME_DURATION(x) OPUS_GET_EXPERT_FRAME_DURATION_REQUEST, __opus_check_int_ptr(x) @@ -681,6 +689,30 @@ extern "C" { */ #define OPUS_GET_SAMPLE_RATE(x) OPUS_GET_SAMPLE_RATE_REQUEST, __opus_check_int_ptr(x) +/** If set to 1, disables the use of phase inversion for intensity stereo, + * improving the quality of mono downmixes, but slightly reducing normal + * stereo quality. Disabling phase inversion in the decoder does not comply + * with RFC 6716, although it does not cause any interoperability issue and + * is expected to become part of the Opus standard once RFC 6716 is updated + * by draft-ietf-codec-opus-update. + * @see OPUS_GET_PHASE_INVERSION_DISABLED + * @param[in] x opus_int32: Allowed values: + *
+ *
0
Enable phase inversion (default).
+ *
1
Disable phase inversion.
+ *
+ * @hideinitializer */ +#define OPUS_SET_PHASE_INVERSION_DISABLED(x) OPUS_SET_PHASE_INVERSION_DISABLED_REQUEST, __opus_check_int(x) +/** Gets the encoder's configured phase inversion status. + * @see OPUS_SET_PHASE_INVERSION_DISABLED + * @param[out] x opus_int32 *: Returns one of the following values: + *
+ *
0
Stereo phase inversion enabled (default).
+ *
1
Stereo phase inversion disabled.
+ *
+ * @hideinitializer */ +#define OPUS_GET_PHASE_INVERSION_DISABLED(x) OPUS_GET_PHASE_INVERSION_DISABLED_REQUEST, __opus_check_int_ptr(x) + /**@}*/ /** @defgroup opus_decoderctls Decoder related CTLs diff --git a/code/opus-1.1.4/include/opus_multistream.h b/code/opus-1.1.4/include/opus_multistream.h index 3622e009..babcee69 100644 --- a/code/opus-1.1.4/include/opus_multistream.h +++ b/code/opus-1.1.4/include/opus_multistream.h @@ -273,7 +273,7 @@ OPUS_EXPORT OPUS_WARN_UNUSED_RESULT OpusMSEncoder *opus_multistream_surround_enc unsigned char *mapping, int application, int *error -) OPUS_ARG_NONNULL(5); +) OPUS_ARG_NONNULL(4) OPUS_ARG_NONNULL(5) OPUS_ARG_NONNULL(6); /** Initialize a previously allocated multistream encoder state. * The memory pointed to by \a st must be at least the size returned by @@ -342,7 +342,7 @@ OPUS_EXPORT int opus_multistream_surround_encoder_init( int *coupled_streams, unsigned char *mapping, int application -) OPUS_ARG_NONNULL(1) OPUS_ARG_NONNULL(6); +) OPUS_ARG_NONNULL(1) OPUS_ARG_NONNULL(5) OPUS_ARG_NONNULL(6) OPUS_ARG_NONNULL(7); /** Encodes a multistream Opus frame. * @param st OpusMSEncoder*: Multistream encoder state. diff --git a/code/opus-1.1.4/include/opus_types.h b/code/opus-1.1.4/include/opus_types.h index b28e03ae..71808266 100644 --- a/code/opus-1.1.4/include/opus_types.h +++ b/code/opus-1.1.4/include/opus_types.h @@ -34,7 +34,7 @@ #define OPUS_TYPES_H /* Use the real stdint.h if it's there (taken from Paul Hsieh's pstdint.h) */ -#if (defined(__STDC__) && __STDC__ && __STDC_VERSION__ >= 199901L) || (defined(__GNUC__) && (defined(_STDINT_H) || defined(_STDINT_H_)) || defined (HAVE_STDINT_H)) +#if (defined(__STDC__) && __STDC__ && defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L) || (defined(__GNUC__) && (defined(_STDINT_H) || defined(_STDINT_H_)) || defined (HAVE_STDINT_H)) #include typedef int16_t opus_int16; diff --git a/code/opus-1.1.4/silk/A2NLSF.c b/code/opus-1.1.4/silk/A2NLSF.c index b6e9e5ff..b487686f 100644 --- a/code/opus-1.1.4/silk/A2NLSF.c +++ b/code/opus-1.1.4/silk/A2NLSF.c @@ -40,7 +40,7 @@ POSSIBILITY OF SUCH DAMAGE. /* Number of binary divisions, when not in low complexity mode */ #define BIN_DIV_STEPS_A2NLSF_FIX 3 /* must be no higher than 16 - log2( LSF_COS_TAB_SZ_FIX ) */ -#define MAX_ITERATIONS_A2NLSF_FIX 30 +#define MAX_ITERATIONS_A2NLSF_FIX 16 /* Helper function for A2NLSF(..) */ /* Transforms polynomials from cos(n*f) to cos(f)^n */ @@ -130,7 +130,7 @@ void silk_A2NLSF( const opus_int d /* I Filter order (must be even) */ ) { - opus_int i, k, m, dd, root_ix, ffrac; + opus_int i, k, m, dd, root_ix, ffrac; opus_int32 xlo, xhi, xmid; opus_int32 ylo, yhi, ymid, thr; opus_int32 nom, den; @@ -239,13 +239,13 @@ void silk_A2NLSF( /* Set NLSFs to white spectrum and exit */ NLSF[ 0 ] = (opus_int16)silk_DIV32_16( 1 << 15, d + 1 ); for( k = 1; k < d; k++ ) { - NLSF[ k ] = (opus_int16)silk_SMULBB( k + 1, NLSF[ 0 ] ); + NLSF[ k ] = (opus_int16)silk_ADD16( NLSF[ k-1 ], NLSF[ 0 ] ); } return; } /* Error: Apply progressively more bandwidth expansion and run again */ - silk_bwexpander_32( a_Q16, d, 65536 - silk_SMULBB( 10 + i, i ) ); /* 10_Q16 = 0.00015*/ + silk_bwexpander_32( a_Q16, d, 65536 - silk_LSHIFT( 1, i ) ); silk_A2NLSF_init( a_Q16, P, Q, dd ); p = P; /* Pointer to polynomial */ diff --git a/code/opus-1.1.4/silk/CNG.c b/code/opus-1.1.4/silk/CNG.c index 8443ad63..e6d9b861 100644 --- a/code/opus-1.1.4/silk/CNG.c +++ b/code/opus-1.1.4/silk/CNG.c @@ -138,11 +138,11 @@ void silk_CNG( gain_Q16 = silk_LSHIFT32( silk_SQRT_APPROX( gain_Q16 ), 8 ); } gain_Q10 = silk_RSHIFT( gain_Q16, 6 ); - + silk_CNG_exc( CNG_sig_Q14 + MAX_LPC_ORDER, psCNG->CNG_exc_buf_Q14, length, &psCNG->rand_seed ); /* Convert CNG NLSF to filter representation */ - silk_NLSF2A( A_Q12, psCNG->CNG_smth_NLSF_Q15, psDec->LPC_order ); + silk_NLSF2A( A_Q12, psCNG->CNG_smth_NLSF_Q15, psDec->LPC_order, psDec->arch ); /* Generate CNG signal, by synthesis filtering */ silk_memcpy( CNG_sig_Q14, psCNG->CNG_synth_state, MAX_LPC_ORDER * sizeof( opus_int32 ) ); @@ -170,11 +170,11 @@ void silk_CNG( } /* Update states */ - CNG_sig_Q14[ MAX_LPC_ORDER + i ] = silk_ADD_LSHIFT( CNG_sig_Q14[ MAX_LPC_ORDER + i ], LPC_pred_Q10, 4 ); - + CNG_sig_Q14[ MAX_LPC_ORDER + i ] = silk_ADD_SAT32( CNG_sig_Q14[ MAX_LPC_ORDER + i ], silk_LSHIFT_SAT32( LPC_pred_Q10, 4 ) ); + /* Scale with Gain and add to input signal */ frame[ i ] = (opus_int16)silk_ADD_SAT16( frame[ i ], silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( CNG_sig_Q14[ MAX_LPC_ORDER + i ], gain_Q10 ), 8 ) ) ); - + } silk_memcpy( psCNG->CNG_synth_state, &CNG_sig_Q14[ length ], MAX_LPC_ORDER * sizeof( opus_int32 ) ); } else { diff --git a/code/opus-1.1.4/silk/LPC_analysis_filter.c b/code/opus-1.1.4/silk/LPC_analysis_filter.c index 20906673..7715f70f 100644 --- a/code/opus-1.1.4/silk/LPC_analysis_filter.c +++ b/code/opus-1.1.4/silk/LPC_analysis_filter.c @@ -39,6 +39,13 @@ POSSIBILITY OF SUCH DAMAGE. /* first d output samples are set to zero */ /*******************************************/ +/* OPT: Using celt_fir() for this function should be faster, but it may cause + integer overflows in intermediate values (not final results), which the + current implementation silences by casting to unsigned. Enabling + this should be safe in pretty much all cases, even though it is not technically + C89-compliant. */ +#define USE_CELT_FIR 0 + void silk_LPC_analysis_filter( opus_int16 *out, /* O Output signal */ const opus_int16 *in, /* I Input signal */ @@ -49,8 +56,7 @@ void silk_LPC_analysis_filter( ) { opus_int j; -#ifdef FIXED_POINT - opus_int16 mem[SILK_MAX_ORDER_LPC]; +#if defined(FIXED_POINT) && USE_CELT_FIR opus_int16 num[SILK_MAX_ORDER_LPC]; #else int ix; @@ -62,15 +68,12 @@ void silk_LPC_analysis_filter( silk_assert( (d & 1) == 0 ); silk_assert( d <= len ); -#ifdef FIXED_POINT +#if defined(FIXED_POINT) && USE_CELT_FIR silk_assert( d <= SILK_MAX_ORDER_LPC ); for ( j = 0; j < d; j++ ) { num[ j ] = -B[ j ]; } - for (j=0;j maxabs ) { + maxabs = absval; + idx = k; + } + } + maxabs = silk_RSHIFT_ROUND( maxabs, QIN - QOUT ); + + if( maxabs > silk_int16_MAX ) { + /* Reduce magnitude of prediction coefficients */ + maxabs = silk_min( maxabs, 163838 ); /* ( silk_int32_MAX >> 14 ) + silk_int16_MAX = 163838 */ + chirp_Q16 = SILK_FIX_CONST( 0.999, 16 ) - silk_DIV32( silk_LSHIFT( maxabs - silk_int16_MAX, 14 ), + silk_RSHIFT32( silk_MUL( maxabs, idx + 1), 2 ) ); + silk_bwexpander_32( a_QIN, d, chirp_Q16 ); + } else { + break; + } + } + + if( i == 10 ) { + /* Reached the last iteration, clip the coefficients */ + for( k = 0; k < d; k++ ) { + a_QOUT[ k ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( a_QIN[ k ], QIN - QOUT ) ); + a_QIN[ k ] = silk_LSHIFT( (opus_int32)a_QOUT[ k ], QIN - QOUT ); + } + } else { + for( k = 0; k < d; k++ ) { + a_QOUT[ k ] = (opus_int16)silk_RSHIFT_ROUND( a_QIN[ k ], QIN - QOUT ); + } + } +} diff --git a/code/opus-1.1.4/silk/LPC_inv_pred_gain.c b/code/opus-1.1.4/silk/LPC_inv_pred_gain.c index 4af89aa5..a3746a6e 100644 --- a/code/opus-1.1.4/silk/LPC_inv_pred_gain.c +++ b/code/opus-1.1.4/silk/LPC_inv_pred_gain.c @@ -30,6 +30,7 @@ POSSIBILITY OF SUCH DAMAGE. #endif #include "SigProc_FIX.h" +#include "define.h" #define QA 24 #define A_LIMIT SILK_FIX_CONST( 0.99975, QA ) @@ -38,117 +39,103 @@ POSSIBILITY OF SUCH DAMAGE. /* Compute inverse of LPC prediction gain, and */ /* test if LPC coefficients are stable (all poles within unit circle) */ -static opus_int32 LPC_inverse_pred_gain_QA( /* O Returns inverse prediction gain in energy domain, Q30 */ - opus_int32 A_QA[ 2 ][ SILK_MAX_ORDER_LPC ], /* I Prediction coefficients */ +static opus_int32 LPC_inverse_pred_gain_QA_c( /* O Returns inverse prediction gain in energy domain, Q30 */ + opus_int32 A_QA[ SILK_MAX_ORDER_LPC ], /* I Prediction coefficients */ const opus_int order /* I Prediction order */ ) { opus_int k, n, mult2Q; - opus_int32 invGain_Q30, rc_Q31, rc_mult1_Q30, rc_mult2, tmp_QA; - opus_int32 *Aold_QA, *Anew_QA; + opus_int32 invGain_Q30, rc_Q31, rc_mult1_Q30, rc_mult2, tmp1, tmp2; - Anew_QA = A_QA[ order & 1 ]; - - invGain_Q30 = (opus_int32)1 << 30; + invGain_Q30 = SILK_FIX_CONST( 1, 30 ); for( k = order - 1; k > 0; k-- ) { /* Check for stability */ - if( ( Anew_QA[ k ] > A_LIMIT ) || ( Anew_QA[ k ] < -A_LIMIT ) ) { + if( ( A_QA[ k ] > A_LIMIT ) || ( A_QA[ k ] < -A_LIMIT ) ) { return 0; } /* Set RC equal to negated AR coef */ - rc_Q31 = -silk_LSHIFT( Anew_QA[ k ], 31 - QA ); + rc_Q31 = -silk_LSHIFT( A_QA[ k ], 31 - QA ); /* rc_mult1_Q30 range: [ 1 : 2^30 ] */ - rc_mult1_Q30 = ( (opus_int32)1 << 30 ) - silk_SMMUL( rc_Q31, rc_Q31 ); + rc_mult1_Q30 = silk_SUB32( SILK_FIX_CONST( 1, 30 ), silk_SMMUL( rc_Q31, rc_Q31 ) ); silk_assert( rc_mult1_Q30 > ( 1 << 15 ) ); /* reduce A_LIMIT if fails */ silk_assert( rc_mult1_Q30 <= ( 1 << 30 ) ); - /* rc_mult2 range: [ 2^30 : silk_int32_MAX ] */ - mult2Q = 32 - silk_CLZ32( silk_abs( rc_mult1_Q30 ) ); - rc_mult2 = silk_INVERSE32_varQ( rc_mult1_Q30, mult2Q + 30 ); - /* Update inverse gain */ /* invGain_Q30 range: [ 0 : 2^30 ] */ invGain_Q30 = silk_LSHIFT( silk_SMMUL( invGain_Q30, rc_mult1_Q30 ), 2 ); silk_assert( invGain_Q30 >= 0 ); silk_assert( invGain_Q30 <= ( 1 << 30 ) ); + if( invGain_Q30 < SILK_FIX_CONST( 1.0f / MAX_PREDICTION_POWER_GAIN, 30 ) ) { + return 0; + } - /* Swap pointers */ - Aold_QA = Anew_QA; - Anew_QA = A_QA[ k & 1 ]; + /* rc_mult2 range: [ 2^30 : silk_int32_MAX ] */ + mult2Q = 32 - silk_CLZ32( silk_abs( rc_mult1_Q30 ) ); + rc_mult2 = silk_INVERSE32_varQ( rc_mult1_Q30, mult2Q + 30 ); /* Update AR coefficient */ - for( n = 0; n < k; n++ ) { - tmp_QA = Aold_QA[ n ] - MUL32_FRAC_Q( Aold_QA[ k - n - 1 ], rc_Q31, 31 ); - Anew_QA[ n ] = MUL32_FRAC_Q( tmp_QA, rc_mult2 , mult2Q ); + for( n = 0; n < (k + 1) >> 1; n++ ) { + opus_int64 tmp64; + tmp1 = A_QA[ n ]; + tmp2 = A_QA[ k - n - 1 ]; + tmp64 = silk_RSHIFT_ROUND64( silk_SMULL( silk_SUB_SAT32(tmp1, + MUL32_FRAC_Q( tmp2, rc_Q31, 31 ) ), rc_mult2 ), mult2Q); + if( tmp64 > silk_int32_MAX || tmp64 < silk_int32_MIN ) { + return 0; + } + A_QA[ n ] = ( opus_int32 )tmp64; + tmp64 = silk_RSHIFT_ROUND64( silk_SMULL( silk_SUB_SAT32(tmp2, + MUL32_FRAC_Q( tmp1, rc_Q31, 31 ) ), rc_mult2), mult2Q); + if( tmp64 > silk_int32_MAX || tmp64 < silk_int32_MIN ) { + return 0; + } + A_QA[ k - n - 1 ] = ( opus_int32 )tmp64; } } /* Check for stability */ - if( ( Anew_QA[ 0 ] > A_LIMIT ) || ( Anew_QA[ 0 ] < -A_LIMIT ) ) { + if( ( A_QA[ k ] > A_LIMIT ) || ( A_QA[ k ] < -A_LIMIT ) ) { return 0; } /* Set RC equal to negated AR coef */ - rc_Q31 = -silk_LSHIFT( Anew_QA[ 0 ], 31 - QA ); + rc_Q31 = -silk_LSHIFT( A_QA[ 0 ], 31 - QA ); /* Range: [ 1 : 2^30 ] */ - rc_mult1_Q30 = ( (opus_int32)1 << 30 ) - silk_SMMUL( rc_Q31, rc_Q31 ); + rc_mult1_Q30 = silk_SUB32( SILK_FIX_CONST( 1, 30 ), silk_SMMUL( rc_Q31, rc_Q31 ) ); /* Update inverse gain */ /* Range: [ 0 : 2^30 ] */ invGain_Q30 = silk_LSHIFT( silk_SMMUL( invGain_Q30, rc_mult1_Q30 ), 2 ); - silk_assert( invGain_Q30 >= 0 ); - silk_assert( invGain_Q30 <= 1<<30 ); + silk_assert( invGain_Q30 >= 0 ); + silk_assert( invGain_Q30 <= ( 1 << 30 ) ); + if( invGain_Q30 < SILK_FIX_CONST( 1.0f / MAX_PREDICTION_POWER_GAIN, 30 ) ) { + return 0; + } return invGain_Q30; } /* For input in Q12 domain */ -opus_int32 silk_LPC_inverse_pred_gain( /* O Returns inverse prediction gain in energy domain, Q30 */ +opus_int32 silk_LPC_inverse_pred_gain_c( /* O Returns inverse prediction gain in energy domain, Q30 */ const opus_int16 *A_Q12, /* I Prediction coefficients, Q12 [order] */ const opus_int order /* I Prediction order */ ) { opus_int k; - opus_int32 Atmp_QA[ 2 ][ SILK_MAX_ORDER_LPC ]; - opus_int32 *Anew_QA; + opus_int32 Atmp_QA[ SILK_MAX_ORDER_LPC ]; opus_int32 DC_resp = 0; - Anew_QA = Atmp_QA[ order & 1 ]; - /* Increase Q domain of the AR coefficients */ for( k = 0; k < order; k++ ) { DC_resp += (opus_int32)A_Q12[ k ]; - Anew_QA[ k ] = silk_LSHIFT32( (opus_int32)A_Q12[ k ], QA - 12 ); + Atmp_QA[ k ] = silk_LSHIFT32( (opus_int32)A_Q12[ k ], QA - 12 ); } /* If the DC is unstable, we don't even need to do the full calculations */ if( DC_resp >= 4096 ) { return 0; } - return LPC_inverse_pred_gain_QA( Atmp_QA, order ); + return LPC_inverse_pred_gain_QA_c( Atmp_QA, order ); } - -#ifdef FIXED_POINT - -/* For input in Q24 domain */ -opus_int32 silk_LPC_inverse_pred_gain_Q24( /* O Returns inverse prediction gain in energy domain, Q30 */ - const opus_int32 *A_Q24, /* I Prediction coefficients [order] */ - const opus_int order /* I Prediction order */ -) -{ - opus_int k; - opus_int32 Atmp_QA[ 2 ][ SILK_MAX_ORDER_LPC ]; - opus_int32 *Anew_QA; - - Anew_QA = Atmp_QA[ order & 1 ]; - - /* Increase Q domain of the AR coefficients */ - for( k = 0; k < order; k++ ) { - Anew_QA[ k ] = silk_RSHIFT32( A_Q24[ k ], 24 - QA ); - } - - return LPC_inverse_pred_gain_QA( Atmp_QA, order ); -} -#endif diff --git a/code/opus-1.1.4/silk/LP_variable_cutoff.c b/code/opus-1.1.4/silk/LP_variable_cutoff.c index f639e1f8..79112ad3 100644 --- a/code/opus-1.1.4/silk/LP_variable_cutoff.c +++ b/code/opus-1.1.4/silk/LP_variable_cutoff.c @@ -130,6 +130,6 @@ void silk_LP_variable_cutoff( /* ARMA low-pass filtering */ silk_assert( TRANSITION_NB == 3 && TRANSITION_NA == 2 ); - silk_biquad_alt( frame, B_Q28, A_Q28, psLP->In_LP_State, frame, frame_length, 1); + silk_biquad_alt_stride1( frame, B_Q28, A_Q28, psLP->In_LP_State, frame, frame_length); } } diff --git a/code/opus-1.1.4/silk/MacroCount.h b/code/opus-1.1.4/silk/MacroCount.h index 834817d0..78100ffe 100644 --- a/code/opus-1.1.4/silk/MacroCount.h +++ b/code/opus-1.1.4/silk/MacroCount.h @@ -319,14 +319,6 @@ static OPUS_INLINE opus_int32 silk_ADD_POS_SAT32(opus_int64 a, opus_int64 b){ return(tmp); } -#undef silk_ADD_POS_SAT64 -static OPUS_INLINE opus_int64 silk_ADD_POS_SAT64(opus_int64 a, opus_int64 b){ - opus_int64 tmp; - ops_count += 1; - tmp = ((((a)+(b)) & 0x8000000000000000LL) ? silk_int64_MAX : ((a)+(b))); - return(tmp); -} - #undef silk_LSHIFT8 static OPUS_INLINE opus_int8 silk_LSHIFT8(opus_int8 a, opus_int32 shift){ opus_int8 ret; @@ -699,7 +691,7 @@ return(ret); #undef silk_LIMIT_32 -static OPUS_INLINE opus_int silk_LIMIT_32(opus_int32 a, opus_int32 limit1, opus_int32 limit2) +static OPUS_INLINE opus_int32 silk_LIMIT_32(opus_int32 a, opus_int32 limit1, opus_int32 limit2) { opus_int32 ret; ops_count += 6; diff --git a/code/opus-1.1.4/silk/MacroDebug.h b/code/opus-1.1.4/silk/MacroDebug.h index 35aedc5c..8dd4ce2e 100644 --- a/code/opus-1.1.4/silk/MacroDebug.h +++ b/code/opus-1.1.4/silk/MacroDebug.h @@ -539,8 +539,7 @@ static OPUS_INLINE opus_int32 silk_DIV32_16_(opus_int32 a32, opus_int32 b32, cha no checking needed for silk_POS_SAT32 no checking needed for silk_ADD_POS_SAT8 no checking needed for silk_ADD_POS_SAT16 - no checking needed for silk_ADD_POS_SAT32 - no checking needed for silk_ADD_POS_SAT64 */ + no checking needed for silk_ADD_POS_SAT32 */ #undef silk_LSHIFT8 #define silk_LSHIFT8(a,b) silk_LSHIFT8_((a), (b), __FILE__, __LINE__) diff --git a/code/opus-1.1.4/silk/NLSF2A.c b/code/opus-1.1.4/silk/NLSF2A.c index b1c559ea..116b465b 100644 --- a/code/opus-1.1.4/silk/NLSF2A.c +++ b/code/opus-1.1.4/silk/NLSF2A.c @@ -66,7 +66,8 @@ static OPUS_INLINE void silk_NLSF2A_find_poly( void silk_NLSF2A( opus_int16 *a_Q12, /* O monic whitening filter coefficients in Q12, [ d ] */ const opus_int16 *NLSF, /* I normalized line spectral frequencies in Q15, [ d ] */ - const opus_int d /* I filter order (should be even) */ + const opus_int d, /* I filter order (should be even) */ + int arch /* I Run-time architecture */ ) { /* This ordering was found to maximize quality. It improves numerical accuracy of @@ -83,15 +84,14 @@ void silk_NLSF2A( opus_int32 P[ SILK_MAX_ORDER_LPC / 2 + 1 ], Q[ SILK_MAX_ORDER_LPC / 2 + 1 ]; opus_int32 Ptmp, Qtmp, f_int, f_frac, cos_val, delta; opus_int32 a32_QA1[ SILK_MAX_ORDER_LPC ]; - opus_int32 maxabs, absval, idx=0, sc_Q16; silk_assert( LSF_COS_TAB_SZ_FIX == 128 ); - silk_assert( d==10||d==16 ); + silk_assert( d==10 || d==16 ); /* convert LSFs to 2*cos(LSF), using piecewise linear curve from table */ ordering = d == 16 ? ordering16 : ordering10; for( k = 0; k < d; k++ ) { - silk_assert(NLSF[k] >= 0 ); + silk_assert( NLSF[k] >= 0 ); /* f_int on a scale 0-127 (rounded down) */ f_int = silk_RSHIFT( NLSF[k], 15 - 7 ); @@ -126,52 +126,15 @@ void silk_NLSF2A( a32_QA1[ d-k-1 ] = Qtmp - Ptmp; /* QA+1 */ } - /* Limit the maximum absolute value of the prediction coefficients, so that they'll fit in int16 */ - for( i = 0; i < 10; i++ ) { - /* Find maximum absolute value and its index */ - maxabs = 0; - for( k = 0; k < d; k++ ) { - absval = silk_abs( a32_QA1[k] ); - if( absval > maxabs ) { - maxabs = absval; - idx = k; - } - } - maxabs = silk_RSHIFT_ROUND( maxabs, QA + 1 - 12 ); /* QA+1 -> Q12 */ + /* Convert int32 coefficients to Q12 int16 coefs */ + silk_LPC_fit( a_Q12, a32_QA1, 12, QA + 1, d ); - if( maxabs > silk_int16_MAX ) { - /* Reduce magnitude of prediction coefficients */ - maxabs = silk_min( maxabs, 163838 ); /* ( silk_int32_MAX >> 14 ) + silk_int16_MAX = 163838 */ - sc_Q16 = SILK_FIX_CONST( 0.999, 16 ) - silk_DIV32( silk_LSHIFT( maxabs - silk_int16_MAX, 14 ), - silk_RSHIFT32( silk_MUL( maxabs, idx + 1), 2 ) ); - silk_bwexpander_32( a32_QA1, d, sc_Q16 ); - } else { - break; - } - } - - if( i == 10 ) { - /* Reached the last iteration, clip the coefficients */ + for( i = 0; silk_LPC_inverse_pred_gain( a_Q12, d, arch ) == 0 && i < MAX_LPC_STABILIZE_ITERATIONS; i++ ) { + /* Prediction coefficients are (too close to) unstable; apply bandwidth expansion */ + /* on the unscaled coefficients, convert to Q12 and measure again */ + silk_bwexpander_32( a32_QA1, d, 65536 - silk_LSHIFT( 2, i ) ); for( k = 0; k < d; k++ ) { - a_Q12[ k ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( a32_QA1[ k ], QA + 1 - 12 ) ); /* QA+1 -> Q12 */ - a32_QA1[ k ] = silk_LSHIFT( (opus_int32)a_Q12[ k ], QA + 1 - 12 ); - } - } else { - for( k = 0; k < d; k++ ) { - a_Q12[ k ] = (opus_int16)silk_RSHIFT_ROUND( a32_QA1[ k ], QA + 1 - 12 ); /* QA+1 -> Q12 */ - } - } - - for( i = 0; i < MAX_LPC_STABILIZE_ITERATIONS; i++ ) { - if( silk_LPC_inverse_pred_gain( a_Q12, d ) < SILK_FIX_CONST( 1.0 / MAX_PREDICTION_POWER_GAIN, 30 ) ) { - /* Prediction coefficients are (too close to) unstable; apply bandwidth expansion */ - /* on the unscaled coefficients, convert to Q12 and measure again */ - silk_bwexpander_32( a32_QA1, d, 65536 - silk_LSHIFT( 2, i ) ); - for( k = 0; k < d; k++ ) { - a_Q12[ k ] = (opus_int16)silk_RSHIFT_ROUND( a32_QA1[ k ], QA + 1 - 12 ); /* QA+1 -> Q12 */ - } - } else { - break; + a_Q12[ k ] = (opus_int16)silk_RSHIFT_ROUND( a32_QA1[ k ], QA + 1 - 12 ); /* QA+1 -> Q12 */ } } } diff --git a/code/opus-1.1.4/silk/NLSF_VQ.c b/code/opus-1.1.4/silk/NLSF_VQ.c index 69b6e22e..452f3dcb 100644 --- a/code/opus-1.1.4/silk/NLSF_VQ.c +++ b/code/opus-1.1.4/silk/NLSF_VQ.c @@ -33,36 +33,44 @@ POSSIBILITY OF SUCH DAMAGE. /* Compute quantization errors for an LPC_order element input vector for a VQ codebook */ void silk_NLSF_VQ( - opus_int32 err_Q26[], /* O Quantization errors [K] */ + opus_int32 err_Q24[], /* O Quantization errors [K] */ const opus_int16 in_Q15[], /* I Input vectors to be quantized [LPC_order] */ const opus_uint8 pCB_Q8[], /* I Codebook vectors [K*LPC_order] */ + const opus_int16 pWght_Q9[], /* I Codebook weights [K*LPC_order] */ const opus_int K, /* I Number of codebook vectors */ const opus_int LPC_order /* I Number of LPCs */ ) { - opus_int i, m; - opus_int32 diff_Q15, sum_error_Q30, sum_error_Q26; + opus_int i, m; + opus_int32 diff_Q15, diffw_Q24, sum_error_Q24, pred_Q24; + const opus_int16 *w_Q9_ptr; + const opus_uint8 *cb_Q8_ptr; - silk_assert( LPC_order <= 16 ); silk_assert( ( LPC_order & 1 ) == 0 ); /* Loop over codebook */ + cb_Q8_ptr = pCB_Q8; + w_Q9_ptr = pWght_Q9; for( i = 0; i < K; i++ ) { - sum_error_Q26 = 0; - for( m = 0; m < LPC_order; m += 2 ) { - /* Compute weighted squared quantization error for index m */ - diff_Q15 = silk_SUB_LSHIFT32( in_Q15[ m ], (opus_int32)*pCB_Q8++, 7 ); /* range: [ -32767 : 32767 ]*/ - sum_error_Q30 = silk_SMULBB( diff_Q15, diff_Q15 ); + sum_error_Q24 = 0; + pred_Q24 = 0; + for( m = LPC_order-2; m >= 0; m -= 2 ) { + /* Compute weighted absolute predictive quantization error for index m + 1 */ + diff_Q15 = silk_SUB_LSHIFT32( in_Q15[ m + 1 ], (opus_int32)cb_Q8_ptr[ m + 1 ], 7 ); /* range: [ -32767 : 32767 ]*/ + diffw_Q24 = silk_SMULBB( diff_Q15, w_Q9_ptr[ m + 1 ] ); + sum_error_Q24 = silk_ADD32( sum_error_Q24, silk_abs( silk_SUB_RSHIFT32( diffw_Q24, pred_Q24, 1 ) ) ); + pred_Q24 = diffw_Q24; - /* Compute weighted squared quantization error for index m + 1 */ - diff_Q15 = silk_SUB_LSHIFT32( in_Q15[m + 1], (opus_int32)*pCB_Q8++, 7 ); /* range: [ -32767 : 32767 ]*/ - sum_error_Q30 = silk_SMLABB( sum_error_Q30, diff_Q15, diff_Q15 ); + /* Compute weighted absolute predictive quantization error for index m */ + diff_Q15 = silk_SUB_LSHIFT32( in_Q15[ m ], (opus_int32)cb_Q8_ptr[ m ], 7 ); /* range: [ -32767 : 32767 ]*/ + diffw_Q24 = silk_SMULBB( diff_Q15, w_Q9_ptr[ m ] ); + sum_error_Q24 = silk_ADD32( sum_error_Q24, silk_abs( silk_SUB_RSHIFT32( diffw_Q24, pred_Q24, 1 ) ) ); + pred_Q24 = diffw_Q24; - sum_error_Q26 = silk_ADD_RSHIFT32( sum_error_Q26, sum_error_Q30, 4 ); - - silk_assert( sum_error_Q26 >= 0 ); - silk_assert( sum_error_Q30 >= 0 ); + silk_assert( sum_error_Q24 >= 0 ); } - err_Q26[ i ] = sum_error_Q26; + err_Q24[ i ] = sum_error_Q24; + cb_Q8_ptr += LPC_order; + w_Q9_ptr += LPC_order; } } diff --git a/code/opus-1.1.4/silk/NLSF_decode.c b/code/opus-1.1.4/silk/NLSF_decode.c index 9f715060..eeb0ba8c 100644 --- a/code/opus-1.1.4/silk/NLSF_decode.c +++ b/code/opus-1.1.4/silk/NLSF_decode.c @@ -32,7 +32,7 @@ POSSIBILITY OF SUCH DAMAGE. #include "main.h" /* Predictive dequantizer for NLSF residuals */ -static OPUS_INLINE void silk_NLSF_residual_dequant( /* O Returns RD value in Q30 */ +static OPUS_INLINE void silk_NLSF_residual_dequant( /* O Returns RD value in Q30 */ opus_int16 x_Q10[], /* O Output [ order ] */ const opus_int8 indices[], /* I Quantization indices [ order ] */ const opus_uint8 pred_coef_Q8[], /* I Backward predictor coefs [ order ] */ @@ -70,15 +70,9 @@ void silk_NLSF_decode( opus_uint8 pred_Q8[ MAX_LPC_ORDER ]; opus_int16 ec_ix[ MAX_LPC_ORDER ]; opus_int16 res_Q10[ MAX_LPC_ORDER ]; - opus_int16 W_tmp_QW[ MAX_LPC_ORDER ]; - opus_int32 W_tmp_Q9, NLSF_Q15_tmp; + opus_int32 NLSF_Q15_tmp; const opus_uint8 *pCB_element; - - /* Decode first stage */ - pCB_element = &psNLSF_CB->CB1_NLSF_Q8[ NLSFIndices[ 0 ] * psNLSF_CB->order ]; - for( i = 0; i < psNLSF_CB->order; i++ ) { - pNLSF_Q15[ i ] = silk_LSHIFT( (opus_int16)pCB_element[ i ], 7 ); - } + const opus_int16 *pCB_Wght_Q9; /* Unpack entropy table indices and predictor for current CB1 index */ silk_NLSF_unpack( ec_ix, pred_Q8, psNLSF_CB, NLSFIndices[ 0 ] ); @@ -86,13 +80,11 @@ void silk_NLSF_decode( /* Predictive residual dequantizer */ silk_NLSF_residual_dequant( res_Q10, &NLSFIndices[ 1 ], pred_Q8, psNLSF_CB->quantStepSize_Q16, psNLSF_CB->order ); - /* Weights from codebook vector */ - silk_NLSF_VQ_weights_laroia( W_tmp_QW, pNLSF_Q15, psNLSF_CB->order ); - - /* Apply inverse square-rooted weights and add to output */ + /* Apply inverse square-rooted weights to first stage and add to output */ + pCB_element = &psNLSF_CB->CB1_NLSF_Q8[ NLSFIndices[ 0 ] * psNLSF_CB->order ]; + pCB_Wght_Q9 = &psNLSF_CB->CB1_Wght_Q9[ NLSFIndices[ 0 ] * psNLSF_CB->order ]; for( i = 0; i < psNLSF_CB->order; i++ ) { - W_tmp_Q9 = silk_SQRT_APPROX( silk_LSHIFT( (opus_int32)W_tmp_QW[ i ], 18 - NLSF_W_Q ) ); - NLSF_Q15_tmp = silk_ADD32( pNLSF_Q15[ i ], silk_DIV32_16( silk_LSHIFT( (opus_int32)res_Q10[ i ], 14 ), W_tmp_Q9 ) ); + NLSF_Q15_tmp = silk_ADD_LSHIFT32( silk_DIV32_16( silk_LSHIFT( (opus_int32)res_Q10[ i ], 14 ), pCB_Wght_Q9[ i ] ), (opus_int16)pCB_element[ i ], 7 ); pNLSF_Q15[ i ] = (opus_int16)silk_LIMIT( NLSF_Q15_tmp, 0, 32767 ); } diff --git a/code/opus-1.1.4/silk/NLSF_del_dec_quant.c b/code/opus-1.1.4/silk/NLSF_del_dec_quant.c index de88fee0..44a16acd 100644 --- a/code/opus-1.1.4/silk/NLSF_del_dec_quant.c +++ b/code/opus-1.1.4/silk/NLSF_del_dec_quant.c @@ -84,7 +84,7 @@ opus_int32 silk_NLSF_del_dec_quant( /* O Returns nStates = 1; RD_Q25[ 0 ] = 0; prev_out_Q10[ 0 ] = 0; - for( i = order - 1; ; i-- ) { + for( i = order - 1; i >= 0; i-- ) { rates_Q5 = &ec_rates_Q5[ ec_ix[ i ] ]; in_Q10 = x_Q10[ i ]; for( j = 0; j < nStates; j++ ) { @@ -131,7 +131,7 @@ opus_int32 silk_NLSF_del_dec_quant( /* O Returns RD_Q25[ j + nStates ] = silk_SMLABB( silk_MLA( RD_tmp_Q25, silk_SMULBB( diff_Q10, diff_Q10 ), w_Q5[ i ] ), mu_Q20, rate1_Q5 ); } - if( nStates <= ( NLSF_QUANT_DEL_DEC_STATES >> 1 ) ) { + if( nStates <= NLSF_QUANT_DEL_DEC_STATES/2 ) { /* double number of states and copy */ for( j = 0; j < nStates; j++ ) { ind[ j + nStates ][ i ] = ind[ j ][ i ] + 1; @@ -140,7 +140,7 @@ opus_int32 silk_NLSF_del_dec_quant( /* O Returns for( j = nStates; j < NLSF_QUANT_DEL_DEC_STATES; j++ ) { ind[ j ][ i ] = ind[ j - nStates ][ i ]; } - } else if( i > 0 ) { + } else { /* sort lower and upper half of RD_Q25, pairwise */ for( j = 0; j < NLSF_QUANT_DEL_DEC_STATES; j++ ) { if( RD_Q25[ j ] > RD_Q25[ j + NLSF_QUANT_DEL_DEC_STATES ] ) { @@ -191,8 +191,6 @@ opus_int32 silk_NLSF_del_dec_quant( /* O Returns for( j = 0; j < NLSF_QUANT_DEL_DEC_STATES; j++ ) { ind[ j ][ i ] += silk_RSHIFT( ind_sort[ j ], NLSF_QUANT_DEL_DEC_STATES_LOG2 ); } - } else { /* i == 0 */ - break; } } diff --git a/code/opus-1.1.4/silk/NLSF_encode.c b/code/opus-1.1.4/silk/NLSF_encode.c index f03c3f1c..268b9a19 100644 --- a/code/opus-1.1.4/silk/NLSF_encode.c +++ b/code/opus-1.1.4/silk/NLSF_encode.c @@ -37,9 +37,9 @@ POSSIBILITY OF SUCH DAMAGE. /***********************/ opus_int32 silk_NLSF_encode( /* O Returns RD value in Q25 */ opus_int8 *NLSFIndices, /* I Codebook path vector [ LPC_ORDER + 1 ] */ - opus_int16 *pNLSF_Q15, /* I/O Quantized NLSF vector [ LPC_ORDER ] */ + opus_int16 *pNLSF_Q15, /* I/O (Un)quantized NLSF vector [ LPC_ORDER ] */ const silk_NLSF_CB_struct *psNLSF_CB, /* I Codebook object */ - const opus_int16 *pW_QW, /* I NLSF weight vector [ LPC_ORDER ] */ + const opus_int16 *pW_Q2, /* I NLSF weight vector [ LPC_ORDER ] */ const opus_int NLSF_mu_Q20, /* I Rate weight for the RD optimization */ const opus_int nSurvivors, /* I Max survivors after first stage */ const opus_int signalType /* I Signal type: 0/1/2 */ @@ -47,21 +47,19 @@ opus_int32 silk_NLSF_encode( /* O Returns { opus_int i, s, ind1, bestIndex, prob_Q8, bits_q7; opus_int32 W_tmp_Q9, ret; - VARDECL( opus_int32, err_Q26 ); + VARDECL( opus_int32, err_Q24 ); VARDECL( opus_int32, RD_Q25 ); VARDECL( opus_int, tempIndices1 ); VARDECL( opus_int8, tempIndices2 ); - opus_int16 res_Q15[ MAX_LPC_ORDER ]; opus_int16 res_Q10[ MAX_LPC_ORDER ]; opus_int16 NLSF_tmp_Q15[ MAX_LPC_ORDER ]; - opus_int16 W_tmp_QW[ MAX_LPC_ORDER ]; opus_int16 W_adj_Q5[ MAX_LPC_ORDER ]; opus_uint8 pred_Q8[ MAX_LPC_ORDER ]; opus_int16 ec_ix[ MAX_LPC_ORDER ]; const opus_uint8 *pCB_element, *iCDF_ptr; + const opus_int16 *pCB_Wght_Q9; SAVE_STACK; - silk_assert( nSurvivors <= NLSF_VQ_MAX_SURVIVORS ); silk_assert( signalType >= 0 && signalType <= 2 ); silk_assert( NLSF_mu_Q20 <= 32767 && NLSF_mu_Q20 >= 0 ); @@ -69,12 +67,12 @@ opus_int32 silk_NLSF_encode( /* O Returns silk_NLSF_stabilize( pNLSF_Q15, psNLSF_CB->deltaMin_Q15, psNLSF_CB->order ); /* First stage: VQ */ - ALLOC( err_Q26, psNLSF_CB->nVectors, opus_int32 ); - silk_NLSF_VQ( err_Q26, pNLSF_Q15, psNLSF_CB->CB1_NLSF_Q8, psNLSF_CB->nVectors, psNLSF_CB->order ); + ALLOC( err_Q24, psNLSF_CB->nVectors, opus_int32 ); + silk_NLSF_VQ( err_Q24, pNLSF_Q15, psNLSF_CB->CB1_NLSF_Q8, psNLSF_CB->CB1_Wght_Q9, psNLSF_CB->nVectors, psNLSF_CB->order ); /* Sort the quantization errors */ ALLOC( tempIndices1, nSurvivors, opus_int ); - silk_insertion_sort_increasing( err_Q26, tempIndices1, psNLSF_CB->nVectors, nSurvivors ); + silk_insertion_sort_increasing( err_Q24, tempIndices1, psNLSF_CB->nVectors, nSurvivors ); ALLOC( RD_Q25, nSurvivors, opus_int32 ); ALLOC( tempIndices2, nSurvivors * MAX_LPC_ORDER, opus_int8 ); @@ -85,23 +83,12 @@ opus_int32 silk_NLSF_encode( /* O Returns /* Residual after first stage */ pCB_element = &psNLSF_CB->CB1_NLSF_Q8[ ind1 * psNLSF_CB->order ]; + pCB_Wght_Q9 = &psNLSF_CB->CB1_Wght_Q9[ ind1 * psNLSF_CB->order ]; for( i = 0; i < psNLSF_CB->order; i++ ) { NLSF_tmp_Q15[ i ] = silk_LSHIFT16( (opus_int16)pCB_element[ i ], 7 ); - res_Q15[ i ] = pNLSF_Q15[ i ] - NLSF_tmp_Q15[ i ]; - } - - /* Weights from codebook vector */ - silk_NLSF_VQ_weights_laroia( W_tmp_QW, NLSF_tmp_Q15, psNLSF_CB->order ); - - /* Apply square-rooted weights */ - for( i = 0; i < psNLSF_CB->order; i++ ) { - W_tmp_Q9 = silk_SQRT_APPROX( silk_LSHIFT( (opus_int32)W_tmp_QW[ i ], 18 - NLSF_W_Q ) ); - res_Q10[ i ] = (opus_int16)silk_RSHIFT( silk_SMULBB( res_Q15[ i ], W_tmp_Q9 ), 14 ); - } - - /* Modify input weights accordingly */ - for( i = 0; i < psNLSF_CB->order; i++ ) { - W_adj_Q5[ i ] = silk_DIV32_16( silk_LSHIFT( (opus_int32)pW_QW[ i ], 5 ), W_tmp_QW[ i ] ); + W_tmp_Q9 = pCB_Wght_Q9[ i ]; + res_Q10[ i ] = (opus_int16)silk_RSHIFT( silk_SMULBB( pNLSF_Q15[ i ] - NLSF_tmp_Q15[ i ], W_tmp_Q9 ), 14 ); + W_adj_Q5[ i ] = silk_DIV32_varQ( (opus_int32)pW_Q2[ i ], silk_SMULBB( W_tmp_Q9, W_tmp_Q9 ), 21 ); } /* Unpack entropy table indices and predictor for current CB1 index */ diff --git a/code/opus-1.1.4/silk/NSQ.c b/code/opus-1.1.4/silk/NSQ.c index 43e3fee7..617a19fe 100644 --- a/code/opus-1.1.4/silk/NSQ.c +++ b/code/opus-1.1.4/silk/NSQ.c @@ -37,7 +37,7 @@ POSSIBILITY OF SUCH DAMAGE. static OPUS_INLINE void silk_nsq_scale_states( const silk_encoder_state *psEncC, /* I Encoder State */ silk_nsq_state *NSQ, /* I/O NSQ state */ - const opus_int32 x_Q3[], /* I input in Q3 */ + const opus_int16 x16[], /* I input */ opus_int32 x_sc_Q10[], /* O input scaled with 1/Gain */ const opus_int16 sLTP[], /* I re-whitened LTP state in Q0 */ opus_int32 sLTP_Q15[], /* O LTP state matching scaled input */ @@ -75,14 +75,14 @@ static OPUS_INLINE void silk_noise_shape_quantizer( void silk_NSQ_c ( - const silk_encoder_state *psEncC, /* I/O Encoder State */ + const silk_encoder_state *psEncC, /* I Encoder State */ silk_nsq_state *NSQ, /* I/O NSQ state */ SideInfoIndices *psIndices, /* I/O Quantization Indices */ - const opus_int32 x_Q3[], /* I Prefiltered input signal */ + const opus_int16 x16[], /* I Input */ opus_int8 pulses[], /* O Quantized pulse signal */ const opus_int16 PredCoef_Q12[ 2 * MAX_LPC_ORDER ], /* I Short term prediction coefs */ const opus_int16 LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ], /* I Long term prediction coefs */ - const opus_int16 AR2_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs */ + const opus_int16 AR_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs */ const opus_int HarmShapeGain_Q14[ MAX_NB_SUBFR ], /* I Long term shaping coefs */ const opus_int Tilt_Q14[ MAX_NB_SUBFR ], /* I Spectral tilt */ const opus_int32 LF_shp_Q14[ MAX_NB_SUBFR ], /* I Low frequency shaping coefs */ @@ -117,8 +117,7 @@ void silk_NSQ_c LSF_interpolation_flag = 1; } - ALLOC( sLTP_Q15, - psEncC->ltp_mem_length + psEncC->frame_length, opus_int32 ); + ALLOC( sLTP_Q15, psEncC->ltp_mem_length + psEncC->frame_length, opus_int32 ); ALLOC( sLTP, psEncC->ltp_mem_length + psEncC->frame_length, opus_int16 ); ALLOC( x_sc_Q10, psEncC->subfr_length, opus_int32 ); /* Set up pointers to start of sub frame */ @@ -128,7 +127,7 @@ void silk_NSQ_c for( k = 0; k < psEncC->nb_subfr; k++ ) { A_Q12 = &PredCoef_Q12[ (( k >> 1 ) | ( 1 - LSF_interpolation_flag )) * MAX_LPC_ORDER ]; B_Q14 = <PCoef_Q14[ k * LTP_ORDER ]; - AR_shp_Q13 = &AR2_Q13[ k * MAX_SHAPE_LPC_ORDER ]; + AR_shp_Q13 = &AR_Q13[ k * MAX_SHAPE_LPC_ORDER ]; /* Noise shape parameters */ silk_assert( HarmShapeGain_Q14[ k ] >= 0 ); @@ -154,13 +153,13 @@ void silk_NSQ_c } } - silk_nsq_scale_states( psEncC, NSQ, x_Q3, x_sc_Q10, sLTP, sLTP_Q15, k, LTP_scale_Q14, Gains_Q16, pitchL, psIndices->signalType ); + silk_nsq_scale_states( psEncC, NSQ, x16, x_sc_Q10, sLTP, sLTP_Q15, k, LTP_scale_Q14, Gains_Q16, pitchL, psIndices->signalType ); silk_noise_shape_quantizer( NSQ, psIndices->signalType, x_sc_Q10, pulses, pxq, sLTP_Q15, A_Q12, B_Q14, AR_shp_Q13, lag, HarmShapeFIRPacked_Q14, Tilt_Q14[ k ], LF_shp_Q14[ k ], Gains_Q16[ k ], Lambda_Q10, offset_Q10, psEncC->subfr_length, psEncC->shapingLPCOrder, psEncC->predictLPCOrder, psEncC->arch ); - x_Q3 += psEncC->subfr_length; + x16 += psEncC->subfr_length; pulses += psEncC->subfr_length; pxq += psEncC->subfr_length; } @@ -169,7 +168,6 @@ void silk_NSQ_c NSQ->lagPrev = pitchL[ psEncC->nb_subfr - 1 ]; /* Save quantized speech and noise shaping signals */ - /* DEBUG_STORE_DATA( enc.pcm, &NSQ->xq[ psEncC->ltp_mem_length ], psEncC->frame_length * sizeof( opus_int16 ) ) */ silk_memmove( NSQ->xq, &NSQ->xq[ psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int16 ) ); silk_memmove( NSQ->sLTP_shp_Q14, &NSQ->sLTP_shp_Q14[ psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int32 ) ); RESTORE_STACK; @@ -250,7 +248,7 @@ void silk_noise_shape_quantizer( /* Noise shape feedback */ silk_assert( ( shapingLPCOrder & 1 ) == 0 ); /* check that order is even */ - n_AR_Q12 = silk_NSQ_noise_shape_feedback_loop(psLPC_Q14, NSQ->sAR2_Q14, AR_shp_Q13, shapingLPCOrder, arch); + n_AR_Q12 = silk_NSQ_noise_shape_feedback_loop(&NSQ->sDiff_shp_Q14, NSQ->sAR2_Q14, AR_shp_Q13, shapingLPCOrder, arch); n_AR_Q12 = silk_SMLAWB( n_AR_Q12, NSQ->sLF_AR_shp_Q14, Tilt_Q14 ); @@ -279,14 +277,27 @@ void silk_noise_shape_quantizer( r_Q10 = silk_SUB32( x_sc_Q10[ i ], tmp1 ); /* residual error Q10 */ /* Flip sign depending on dither */ - if ( NSQ->rand_seed < 0 ) { - r_Q10 = -r_Q10; + if( NSQ->rand_seed < 0 ) { + r_Q10 = -r_Q10; } r_Q10 = silk_LIMIT_32( r_Q10, -(31 << 10), 30 << 10 ); /* Find two quantization level candidates and measure their rate-distortion */ q1_Q10 = silk_SUB32( r_Q10, offset_Q10 ); q1_Q0 = silk_RSHIFT( q1_Q10, 10 ); + if (Lambda_Q10 > 2048) { + /* For aggressive RDO, the bias becomes more than one pulse. */ + int rdo_offset = Lambda_Q10/2 - 512; + if (q1_Q10 > rdo_offset) { + q1_Q0 = silk_RSHIFT( q1_Q10 - rdo_offset, 10 ); + } else if (q1_Q10 < -rdo_offset) { + q1_Q0 = silk_RSHIFT( q1_Q10 + rdo_offset, 10 ); + } else if (q1_Q10 < 0) { + q1_Q0 = -1; + } else { + q1_Q0 = 0; + } + } if( q1_Q0 > 0 ) { q1_Q10 = silk_SUB32( silk_LSHIFT( q1_Q0, 10 ), QUANT_LEVEL_ADJUST_Q10 ); q1_Q10 = silk_ADD32( q1_Q10, offset_Q10 ); @@ -337,7 +348,8 @@ void silk_noise_shape_quantizer( /* Update states */ psLPC_Q14++; *psLPC_Q14 = xq_Q14; - sLF_AR_shp_Q14 = silk_SUB_LSHIFT32( xq_Q14, n_AR_Q12, 2 ); + NSQ->sDiff_shp_Q14 = silk_SUB_LSHIFT32( xq_Q14, x_sc_Q10[ i ], 4 ); + sLF_AR_shp_Q14 = silk_SUB_LSHIFT32( NSQ->sDiff_shp_Q14, n_AR_Q12, 2 ); NSQ->sLF_AR_shp_Q14 = sLF_AR_shp_Q14; NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx ] = silk_SUB_LSHIFT32( sLF_AR_shp_Q14, n_LF_Q12, 2 ); @@ -356,7 +368,7 @@ void silk_noise_shape_quantizer( static OPUS_INLINE void silk_nsq_scale_states( const silk_encoder_state *psEncC, /* I Encoder State */ silk_nsq_state *NSQ, /* I/O NSQ state */ - const opus_int32 x_Q3[], /* I input in Q3 */ + const opus_int16 x16[], /* I input */ opus_int32 x_sc_Q10[], /* O input scaled with 1/Gain */ const opus_int16 sLTP[], /* I re-whitened LTP state in Q0 */ opus_int32 sLTP_Q15[], /* O LTP state matching scaled input */ @@ -368,28 +380,18 @@ static OPUS_INLINE void silk_nsq_scale_states( ) { opus_int i, lag; - opus_int32 gain_adj_Q16, inv_gain_Q31, inv_gain_Q23; + opus_int32 gain_adj_Q16, inv_gain_Q31, inv_gain_Q26; lag = pitchL[ subfr ]; inv_gain_Q31 = silk_INVERSE32_varQ( silk_max( Gains_Q16[ subfr ], 1 ), 47 ); silk_assert( inv_gain_Q31 != 0 ); - /* Calculate gain adjustment factor */ - if( Gains_Q16[ subfr ] != NSQ->prev_gain_Q16 ) { - gain_adj_Q16 = silk_DIV32_varQ( NSQ->prev_gain_Q16, Gains_Q16[ subfr ], 16 ); - } else { - gain_adj_Q16 = (opus_int32)1 << 16; - } - /* Scale input */ - inv_gain_Q23 = silk_RSHIFT_ROUND( inv_gain_Q31, 8 ); + inv_gain_Q26 = silk_RSHIFT_ROUND( inv_gain_Q31, 5 ); for( i = 0; i < psEncC->subfr_length; i++ ) { - x_sc_Q10[ i ] = silk_SMULWW( x_Q3[ i ], inv_gain_Q23 ); + x_sc_Q10[ i ] = silk_SMULWW( x16[ i ], inv_gain_Q26 ); } - /* Save inverse gain */ - NSQ->prev_gain_Q16 = Gains_Q16[ subfr ]; - /* After rewhitening the LTP state is un-scaled, so scale with inv_gain_Q16 */ if( NSQ->rewhite_flag ) { if( subfr == 0 ) { @@ -403,7 +405,9 @@ static OPUS_INLINE void silk_nsq_scale_states( } /* Adjust for changing gain */ - if( gain_adj_Q16 != (opus_int32)1 << 16 ) { + if( Gains_Q16[ subfr ] != NSQ->prev_gain_Q16 ) { + gain_adj_Q16 = silk_DIV32_varQ( NSQ->prev_gain_Q16, Gains_Q16[ subfr ], 16 ); + /* Scale long-term shaping state */ for( i = NSQ->sLTP_shp_buf_idx - psEncC->ltp_mem_length; i < NSQ->sLTP_shp_buf_idx; i++ ) { NSQ->sLTP_shp_Q14[ i ] = silk_SMULWW( gain_adj_Q16, NSQ->sLTP_shp_Q14[ i ] ); @@ -417,6 +421,7 @@ static OPUS_INLINE void silk_nsq_scale_states( } NSQ->sLF_AR_shp_Q14 = silk_SMULWW( gain_adj_Q16, NSQ->sLF_AR_shp_Q14 ); + NSQ->sDiff_shp_Q14 = silk_SMULWW( gain_adj_Q16, NSQ->sDiff_shp_Q14 ); /* Scale short-term prediction and shaping states */ for( i = 0; i < NSQ_LPC_BUF_LENGTH; i++ ) { @@ -425,5 +430,8 @@ static OPUS_INLINE void silk_nsq_scale_states( for( i = 0; i < MAX_SHAPE_LPC_ORDER; i++ ) { NSQ->sAR2_Q14[ i ] = silk_SMULWW( gain_adj_Q16, NSQ->sAR2_Q14[ i ] ); } + + /* Save inverse gain */ + NSQ->prev_gain_Q16 = Gains_Q16[ subfr ]; } } diff --git a/code/opus-1.1.4/silk/NSQ_del_dec.c b/code/opus-1.1.4/silk/NSQ_del_dec.c index ab6feeac..1cd29d96 100644 --- a/code/opus-1.1.4/silk/NSQ_del_dec.c +++ b/code/opus-1.1.4/silk/NSQ_del_dec.c @@ -43,6 +43,7 @@ typedef struct { opus_int32 Shape_Q14[ DECISION_DELAY ]; opus_int32 sAR2_Q14[ MAX_SHAPE_LPC_ORDER ]; opus_int32 LF_AR_Q14; + opus_int32 Diff_Q14; opus_int32 Seed; opus_int32 SeedInit; opus_int32 RD_Q10; @@ -53,6 +54,7 @@ typedef struct { opus_int32 RD_Q10; opus_int32 xq_Q14; opus_int32 LF_AR_Q14; + opus_int32 Diff_Q14; opus_int32 sLTP_shp_Q14; opus_int32 LPC_exc_Q14; } NSQ_sample_struct; @@ -66,7 +68,7 @@ static OPUS_INLINE void silk_nsq_del_dec_scale_states( const silk_encoder_state *psEncC, /* I Encoder State */ silk_nsq_state *NSQ, /* I/O NSQ state */ NSQ_del_dec_struct psDelDec[], /* I/O Delayed decision states */ - const opus_int32 x_Q3[], /* I Input in Q3 */ + const opus_int16 x16[], /* I Input */ opus_int32 x_sc_Q10[], /* O Input scaled with 1/Gain in Q10 */ const opus_int16 sLTP[], /* I Re-whitened LTP state in Q0 */ opus_int32 sLTP_Q15[], /* O LTP state matching scaled input */ @@ -107,20 +109,20 @@ static OPUS_INLINE void silk_noise_shape_quantizer_del_dec( opus_int predictLPCOrder, /* I Prediction filter order */ opus_int warping_Q16, /* I */ opus_int nStatesDelayedDecision, /* I Number of states in decision tree */ - opus_int *smpl_buf_idx, /* I Index to newest samples in buffers */ + opus_int *smpl_buf_idx, /* I/O Index to newest samples in buffers */ opus_int decisionDelay, /* I */ int arch /* I */ ); void silk_NSQ_del_dec_c( - const silk_encoder_state *psEncC, /* I/O Encoder State */ + const silk_encoder_state *psEncC, /* I Encoder State */ silk_nsq_state *NSQ, /* I/O NSQ state */ SideInfoIndices *psIndices, /* I/O Quantization Indices */ - const opus_int32 x_Q3[], /* I Prefiltered input signal */ + const opus_int16 x16[], /* I Input */ opus_int8 pulses[], /* O Quantized pulse signal */ const opus_int16 PredCoef_Q12[ 2 * MAX_LPC_ORDER ], /* I Short term prediction coefs */ const opus_int16 LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ], /* I Long term prediction coefs */ - const opus_int16 AR2_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs */ + const opus_int16 AR_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs */ const opus_int HarmShapeGain_Q14[ MAX_NB_SUBFR ], /* I Long term shaping coefs */ const opus_int Tilt_Q14[ MAX_NB_SUBFR ], /* I Spectral tilt */ const opus_int32 LF_shp_Q14[ MAX_NB_SUBFR ], /* I Low frequency shaping coefs */ @@ -159,6 +161,7 @@ void silk_NSQ_del_dec_c( psDD->SeedInit = psDD->Seed; psDD->RD_Q10 = 0; psDD->LF_AR_Q14 = NSQ->sLF_AR_shp_Q14; + psDD->Diff_Q14 = NSQ->sDiff_shp_Q14; psDD->Shape_Q14[ 0 ] = NSQ->sLTP_shp_Q14[ psEncC->ltp_mem_length - 1 ]; silk_memcpy( psDD->sLPC_Q14, NSQ->sLPC_Q14, NSQ_LPC_BUF_LENGTH * sizeof( opus_int32 ) ); silk_memcpy( psDD->sAR2_Q14, NSQ->sAR2_Q14, sizeof( NSQ->sAR2_Q14 ) ); @@ -186,8 +189,7 @@ void silk_NSQ_del_dec_c( LSF_interpolation_flag = 1; } - ALLOC( sLTP_Q15, - psEncC->ltp_mem_length + psEncC->frame_length, opus_int32 ); + ALLOC( sLTP_Q15, psEncC->ltp_mem_length + psEncC->frame_length, opus_int32 ); ALLOC( sLTP, psEncC->ltp_mem_length + psEncC->frame_length, opus_int16 ); ALLOC( x_sc_Q10, psEncC->subfr_length, opus_int32 ); ALLOC( delayedGain_Q10, DECISION_DELAY, opus_int32 ); @@ -199,7 +201,7 @@ void silk_NSQ_del_dec_c( for( k = 0; k < psEncC->nb_subfr; k++ ) { A_Q12 = &PredCoef_Q12[ ( ( k >> 1 ) | ( 1 - LSF_interpolation_flag ) ) * MAX_LPC_ORDER ]; B_Q14 = <PCoef_Q14[ k * LTP_ORDER ]; - AR_shp_Q13 = &AR2_Q13[ k * MAX_SHAPE_LPC_ORDER ]; + AR_shp_Q13 = &AR_Q13[ k * MAX_SHAPE_LPC_ORDER ]; /* Noise shape parameters */ silk_assert( HarmShapeGain_Q14[ k ] >= 0 ); @@ -235,7 +237,8 @@ void silk_NSQ_del_dec_c( psDD = &psDelDec[ Winner_ind ]; last_smple_idx = smpl_buf_idx + decisionDelay; for( i = 0; i < decisionDelay; i++ ) { - last_smple_idx = ( last_smple_idx - 1 ) & DECISION_DELAY_MASK; + last_smple_idx = ( last_smple_idx - 1 ) % DECISION_DELAY; + if( last_smple_idx < 0 ) last_smple_idx += DECISION_DELAY; pulses[ i - decisionDelay ] = (opus_int8)silk_RSHIFT_ROUND( psDD->Q_Q10[ last_smple_idx ], 10 ); pxq[ i - decisionDelay ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( psDD->Xq_Q14[ last_smple_idx ], Gains_Q16[ 1 ] ), 14 ) ); @@ -257,7 +260,7 @@ void silk_NSQ_del_dec_c( } } - silk_nsq_del_dec_scale_states( psEncC, NSQ, psDelDec, x_Q3, x_sc_Q10, sLTP, sLTP_Q15, k, + silk_nsq_del_dec_scale_states( psEncC, NSQ, psDelDec, x16, x_sc_Q10, sLTP, sLTP_Q15, k, psEncC->nStatesDelayedDecision, LTP_scale_Q14, Gains_Q16, pitchL, psIndices->signalType, decisionDelay ); silk_noise_shape_quantizer_del_dec( NSQ, psDelDec, psIndices->signalType, x_sc_Q10, pulses, pxq, sLTP_Q15, @@ -265,7 +268,7 @@ void silk_NSQ_del_dec_c( Gains_Q16[ k ], Lambda_Q10, offset_Q10, psEncC->subfr_length, subfr++, psEncC->shapingLPCOrder, psEncC->predictLPCOrder, psEncC->warping_Q16, psEncC->nStatesDelayedDecision, &smpl_buf_idx, decisionDelay, psEncC->arch ); - x_Q3 += psEncC->subfr_length; + x16 += psEncC->subfr_length; pulses += psEncC->subfr_length; pxq += psEncC->subfr_length; } @@ -286,7 +289,9 @@ void silk_NSQ_del_dec_c( last_smple_idx = smpl_buf_idx + decisionDelay; Gain_Q10 = silk_RSHIFT32( Gains_Q16[ psEncC->nb_subfr - 1 ], 6 ); for( i = 0; i < decisionDelay; i++ ) { - last_smple_idx = ( last_smple_idx - 1 ) & DECISION_DELAY_MASK; + last_smple_idx = ( last_smple_idx - 1 ) % DECISION_DELAY; + if( last_smple_idx < 0 ) last_smple_idx += DECISION_DELAY; + pulses[ i - decisionDelay ] = (opus_int8)silk_RSHIFT_ROUND( psDD->Q_Q10[ last_smple_idx ], 10 ); pxq[ i - decisionDelay ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( psDD->Xq_Q14[ last_smple_idx ], Gain_Q10 ), 8 ) ); @@ -297,10 +302,10 @@ void silk_NSQ_del_dec_c( /* Update states */ NSQ->sLF_AR_shp_Q14 = psDD->LF_AR_Q14; + NSQ->sDiff_shp_Q14 = psDD->Diff_Q14; NSQ->lagPrev = pitchL[ psEncC->nb_subfr - 1 ]; /* Save quantized speech signal */ - /* DEBUG_STORE_DATA( enc.pcm, &NSQ->xq[psEncC->ltp_mem_length], psEncC->frame_length * sizeof( opus_int16 ) ) */ silk_memmove( NSQ->xq, &NSQ->xq[ psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int16 ) ); silk_memmove( NSQ->sLTP_shp_Q14, &NSQ->sLTP_shp_Q14[ psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int32 ) ); RESTORE_STACK; @@ -335,7 +340,7 @@ static OPUS_INLINE void silk_noise_shape_quantizer_del_dec( opus_int predictLPCOrder, /* I Prediction filter order */ opus_int warping_Q16, /* I */ opus_int nStatesDelayedDecision, /* I Number of states in decision tree */ - opus_int *smpl_buf_idx, /* I Index to newest samples in buffers */ + opus_int *smpl_buf_idx, /* I/O Index to newest samples in buffers */ opus_int decisionDelay, /* I */ int arch /* I */ ) @@ -416,7 +421,7 @@ static OPUS_INLINE void silk_noise_shape_quantizer_del_dec( /* Noise shape feedback */ silk_assert( ( shapingLPCOrder & 1 ) == 0 ); /* check that order is even */ /* Output of lowpass section */ - tmp2 = silk_SMLAWB( psLPC_Q14[ 0 ], psDD->sAR2_Q14[ 0 ], warping_Q16 ); + tmp2 = silk_SMLAWB( psDD->Diff_Q14, psDD->sAR2_Q14[ 0 ], warping_Q16 ); /* Output of allpass section */ tmp1 = silk_SMLAWB( psDD->sAR2_Q14[ 0 ], psDD->sAR2_Q14[ 1 ] - tmp2, warping_Q16 ); psDD->sAR2_Q14[ 0 ] = tmp2; @@ -462,6 +467,19 @@ static OPUS_INLINE void silk_noise_shape_quantizer_del_dec( /* Find two quantization level candidates and measure their rate-distortion */ q1_Q10 = silk_SUB32( r_Q10, offset_Q10 ); q1_Q0 = silk_RSHIFT( q1_Q10, 10 ); + if (Lambda_Q10 > 2048) { + /* For aggressive RDO, the bias becomes more than one pulse. */ + int rdo_offset = Lambda_Q10/2 - 512; + if (q1_Q10 > rdo_offset) { + q1_Q0 = silk_RSHIFT( q1_Q10 - rdo_offset, 10 ); + } else if (q1_Q10 < -rdo_offset) { + q1_Q0 = silk_RSHIFT( q1_Q10 + rdo_offset, 10 ); + } else if (q1_Q10 < 0) { + q1_Q0 = -1; + } else { + q1_Q0 = 0; + } + } if( q1_Q0 > 0 ) { q1_Q10 = silk_SUB32( silk_LSHIFT( q1_Q0, 10 ), QUANT_LEVEL_ADJUST_Q10 ); q1_Q10 = silk_ADD32( q1_Q10, offset_Q10 ); @@ -515,7 +533,8 @@ static OPUS_INLINE void silk_noise_shape_quantizer_del_dec( xq_Q14 = silk_ADD32( LPC_exc_Q14, LPC_pred_Q14 ); /* Update states */ - sLF_AR_shp_Q14 = silk_SUB32( xq_Q14, n_AR_Q14 ); + psSS[ 0 ].Diff_Q14 = silk_SUB_LSHIFT32( xq_Q14, x_Q10[ i ], 4 ); + sLF_AR_shp_Q14 = silk_SUB32( psSS[ 0 ].Diff_Q14, n_AR_Q14 ); psSS[ 0 ].sLTP_shp_Q14 = silk_SUB32( sLF_AR_shp_Q14, n_LF_Q14 ); psSS[ 0 ].LF_AR_Q14 = sLF_AR_shp_Q14; psSS[ 0 ].LPC_exc_Q14 = LPC_exc_Q14; @@ -529,21 +548,22 @@ static OPUS_INLINE void silk_noise_shape_quantizer_del_dec( exc_Q14 = -exc_Q14; } - /* Add predictions */ LPC_exc_Q14 = silk_ADD32( exc_Q14, LTP_pred_Q14 ); xq_Q14 = silk_ADD32( LPC_exc_Q14, LPC_pred_Q14 ); /* Update states */ - sLF_AR_shp_Q14 = silk_SUB32( xq_Q14, n_AR_Q14 ); + psSS[ 1 ].Diff_Q14 = silk_SUB_LSHIFT32( xq_Q14, x_Q10[ i ], 4 ); + sLF_AR_shp_Q14 = silk_SUB32( psSS[ 1 ].Diff_Q14, n_AR_Q14 ); psSS[ 1 ].sLTP_shp_Q14 = silk_SUB32( sLF_AR_shp_Q14, n_LF_Q14 ); psSS[ 1 ].LF_AR_Q14 = sLF_AR_shp_Q14; psSS[ 1 ].LPC_exc_Q14 = LPC_exc_Q14; psSS[ 1 ].xq_Q14 = xq_Q14; } - *smpl_buf_idx = ( *smpl_buf_idx - 1 ) & DECISION_DELAY_MASK; /* Index to newest samples */ - last_smple_idx = ( *smpl_buf_idx + decisionDelay ) & DECISION_DELAY_MASK; /* Index to decisionDelay old samples */ + *smpl_buf_idx = ( *smpl_buf_idx - 1 ) % DECISION_DELAY; + if( *smpl_buf_idx < 0 ) *smpl_buf_idx += DECISION_DELAY; + last_smple_idx = ( *smpl_buf_idx + decisionDelay ) % DECISION_DELAY; /* Find winner */ RDmin_Q10 = psSampleState[ 0 ][ 0 ].RD_Q10; @@ -607,6 +627,7 @@ static OPUS_INLINE void silk_noise_shape_quantizer_del_dec( psDD = &psDelDec[ k ]; psSS = &psSampleState[ k ][ 0 ]; psDD->LF_AR_Q14 = psSS->LF_AR_Q14; + psDD->Diff_Q14 = psSS->Diff_Q14; psDD->sLPC_Q14[ NSQ_LPC_BUF_LENGTH + i ] = psSS->xq_Q14; psDD->Xq_Q14[ *smpl_buf_idx ] = psSS->xq_Q14; psDD->Q_Q10[ *smpl_buf_idx ] = psSS->Q_Q10; @@ -631,7 +652,7 @@ static OPUS_INLINE void silk_nsq_del_dec_scale_states( const silk_encoder_state *psEncC, /* I Encoder State */ silk_nsq_state *NSQ, /* I/O NSQ state */ NSQ_del_dec_struct psDelDec[], /* I/O Delayed decision states */ - const opus_int32 x_Q3[], /* I Input in Q3 */ + const opus_int16 x16[], /* I Input */ opus_int32 x_sc_Q10[], /* O Input scaled with 1/Gain in Q10 */ const opus_int16 sLTP[], /* I Re-whitened LTP state in Q0 */ opus_int32 sLTP_Q15[], /* O LTP state matching scaled input */ @@ -645,29 +666,19 @@ static OPUS_INLINE void silk_nsq_del_dec_scale_states( ) { opus_int i, k, lag; - opus_int32 gain_adj_Q16, inv_gain_Q31, inv_gain_Q23; + opus_int32 gain_adj_Q16, inv_gain_Q31, inv_gain_Q26; NSQ_del_dec_struct *psDD; lag = pitchL[ subfr ]; inv_gain_Q31 = silk_INVERSE32_varQ( silk_max( Gains_Q16[ subfr ], 1 ), 47 ); silk_assert( inv_gain_Q31 != 0 ); - /* Calculate gain adjustment factor */ - if( Gains_Q16[ subfr ] != NSQ->prev_gain_Q16 ) { - gain_adj_Q16 = silk_DIV32_varQ( NSQ->prev_gain_Q16, Gains_Q16[ subfr ], 16 ); - } else { - gain_adj_Q16 = (opus_int32)1 << 16; - } - /* Scale input */ - inv_gain_Q23 = silk_RSHIFT_ROUND( inv_gain_Q31, 8 ); + inv_gain_Q26 = silk_RSHIFT_ROUND( inv_gain_Q31, 5 ); for( i = 0; i < psEncC->subfr_length; i++ ) { - x_sc_Q10[ i ] = silk_SMULWW( x_Q3[ i ], inv_gain_Q23 ); + x_sc_Q10[ i ] = silk_SMULWW( x16[ i ], inv_gain_Q26 ); } - /* Save inverse gain */ - NSQ->prev_gain_Q16 = Gains_Q16[ subfr ]; - /* After rewhitening the LTP state is un-scaled, so scale with inv_gain_Q16 */ if( NSQ->rewhite_flag ) { if( subfr == 0 ) { @@ -681,7 +692,9 @@ static OPUS_INLINE void silk_nsq_del_dec_scale_states( } /* Adjust for changing gain */ - if( gain_adj_Q16 != (opus_int32)1 << 16 ) { + if( Gains_Q16[ subfr ] != NSQ->prev_gain_Q16 ) { + gain_adj_Q16 = silk_DIV32_varQ( NSQ->prev_gain_Q16, Gains_Q16[ subfr ], 16 ); + /* Scale long-term shaping state */ for( i = NSQ->sLTP_shp_buf_idx - psEncC->ltp_mem_length; i < NSQ->sLTP_shp_buf_idx; i++ ) { NSQ->sLTP_shp_Q14[ i ] = silk_SMULWW( gain_adj_Q16, NSQ->sLTP_shp_Q14[ i ] ); @@ -699,6 +712,7 @@ static OPUS_INLINE void silk_nsq_del_dec_scale_states( /* Scale scalar states */ psDD->LF_AR_Q14 = silk_SMULWW( gain_adj_Q16, psDD->LF_AR_Q14 ); + psDD->Diff_Q14 = silk_SMULWW( gain_adj_Q16, psDD->Diff_Q14 ); /* Scale short-term prediction and shaping states */ for( i = 0; i < NSQ_LPC_BUF_LENGTH; i++ ) { @@ -712,5 +726,8 @@ static OPUS_INLINE void silk_nsq_del_dec_scale_states( psDD->Shape_Q14[ i ] = silk_SMULWW( gain_adj_Q16, psDD->Shape_Q14[ i ] ); } } + + /* Save inverse gain */ + NSQ->prev_gain_Q16 = Gains_Q16[ subfr ]; } } diff --git a/code/opus-1.1.4/silk/PLC.c b/code/opus-1.1.4/silk/PLC.c index fb6ea887..a3e55ea7 100644 --- a/code/opus-1.1.4/silk/PLC.c +++ b/code/opus-1.1.4/silk/PLC.c @@ -275,7 +275,7 @@ static OPUS_INLINE void silk_PLC_conceal( /* Reduce random noise for unvoiced frames with high LPC gain */ opus_int32 invGain_Q30, down_scale_Q30; - invGain_Q30 = silk_LPC_inverse_pred_gain( psPLC->prevLPC_Q12, psDec->LPC_order ); + invGain_Q30 = silk_LPC_inverse_pred_gain( psPLC->prevLPC_Q12, psDec->LPC_order, arch ); down_scale_Q30 = silk_min_32( silk_RSHIFT( (opus_int32)1 << 30, LOG2_INV_LPC_GAIN_HIGH_THRES ), invGain_Q30 ); down_scale_Q30 = silk_max_32( silk_RSHIFT( (opus_int32)1 << 30, LOG2_INV_LPC_GAIN_LOW_THRES ), down_scale_Q30 ); @@ -328,8 +328,10 @@ static OPUS_INLINE void silk_PLC_conceal( for( j = 0; j < LTP_ORDER; j++ ) { B_Q14[ j ] = silk_RSHIFT( silk_SMULBB( harm_Gain_Q15, B_Q14[ j ] ), 15 ); } - /* Gradually reduce excitation gain */ - rand_scale_Q14 = silk_RSHIFT( silk_SMULBB( rand_scale_Q14, rand_Gain_Q15 ), 15 ); + if ( psDec->indices.signalType != TYPE_NO_VOICE_ACTIVITY ) { + /* Gradually reduce excitation gain */ + rand_scale_Q14 = silk_RSHIFT( silk_SMULBB( rand_scale_Q14, rand_Gain_Q15 ), 15 ); + } /* Slowly increase pitch lag */ psPLC->pitchL_Q8 = silk_SMLAWB( psPLC->pitchL_Q8, psPLC->pitchL_Q8, PITCH_DRIFT_FAC_Q16 ); diff --git a/code/opus-1.1.4/silk/SigProc_FIX.h b/code/opus-1.1.4/silk/SigProc_FIX.h index b6329944..f9ae3263 100644 --- a/code/opus-1.1.4/silk/SigProc_FIX.h +++ b/code/opus-1.1.4/silk/SigProc_FIX.h @@ -35,7 +35,7 @@ extern "C" /*#define silk_MACRO_COUNT */ /* Used to enable WMOPS counting */ -#define SILK_MAX_ORDER_LPC 16 /* max order of the LPC analysis in schur() and k2a() */ +#define SILK_MAX_ORDER_LPC 24 /* max order of the LPC analysis in schur() and k2a() */ #include /* for memset(), memcpy(), memmove() */ #include "typedef.h" @@ -47,6 +47,11 @@ extern "C" #include "x86/SigProc_FIX_sse.h" #endif +#if (defined(OPUS_ARM_ASM) || defined(OPUS_ARM_MAY_HAVE_NEON_INTR)) +#include "arm/biquad_alt_arm.h" +#include "arm/LPC_inv_pred_gain_arm.h" +#endif + /********************************************************************/ /* SIGNAL PROCESSING FUNCTIONS */ /********************************************************************/ @@ -96,14 +101,22 @@ void silk_resampler_down2_3( * slower than biquad() but uses more precise coefficients * can handle (slowly) varying coefficients */ -void silk_biquad_alt( +void silk_biquad_alt_stride1( const opus_int16 *in, /* I input signal */ const opus_int32 *B_Q28, /* I MA coefficients [3] */ const opus_int32 *A_Q28, /* I AR coefficients [2] */ opus_int32 *S, /* I/O State vector [2] */ opus_int16 *out, /* O output signal */ - const opus_int32 len, /* I signal length (must be even) */ - opus_int stride /* I Operate on interleaved signal if > 1 */ + const opus_int32 len /* I signal length (must be even) */ +); + +void silk_biquad_alt_stride2_c( + const opus_int16 *in, /* I input signal */ + const opus_int32 *B_Q28, /* I MA coefficients [3] */ + const opus_int32 *A_Q28, /* I AR coefficients [2] */ + opus_int32 *S, /* I/O State vector [4] */ + opus_int16 *out, /* O output signal */ + const opus_int32 len /* I signal length (must be even) */ ); /* Variable order MA prediction error filter. */ @@ -132,17 +145,11 @@ void silk_bwexpander_32( /* Compute inverse of LPC prediction gain, and */ /* test if LPC coefficients are stable (all poles within unit circle) */ -opus_int32 silk_LPC_inverse_pred_gain( /* O Returns inverse prediction gain in energy domain, Q30 */ +opus_int32 silk_LPC_inverse_pred_gain_c( /* O Returns inverse prediction gain in energy domain, Q30 */ const opus_int16 *A_Q12, /* I Prediction coefficients, Q12 [order] */ const opus_int order /* I Prediction order */ ); -/* For input in Q24 domain */ -opus_int32 silk_LPC_inverse_pred_gain_Q24( /* O Returns inverse prediction gain in energy domain, Q30 */ - const opus_int32 *A_Q24, /* I Prediction coefficients [order] */ - const opus_int order /* I Prediction order */ -); - /* Split signal in two decimated bands using first-order allpass filters */ void silk_ana_filt_bank_1( const opus_int16 *in, /* I Input signal [N] */ @@ -152,6 +159,14 @@ void silk_ana_filt_bank_1( const opus_int32 N /* I Number of input samples */ ); +#if !defined(OVERRIDE_silk_biquad_alt_stride2) +#define silk_biquad_alt_stride2(in, B_Q28, A_Q28, S, out, len, arch) ((void)(arch), silk_biquad_alt_stride2_c(in, B_Q28, A_Q28, S, out, len)) +#endif + +#if !defined(OVERRIDE_silk_LPC_inverse_pred_gain) +#define silk_LPC_inverse_pred_gain(A_Q12, order, arch) ((void)(arch), silk_LPC_inverse_pred_gain_c(A_Q12, order)) +#endif + /********************************************************************/ /* SCALAR FUNCTIONS */ /********************************************************************/ @@ -271,7 +286,17 @@ void silk_A2NLSF( void silk_NLSF2A( opus_int16 *a_Q12, /* O monic whitening filter coefficients in Q12, [ d ] */ const opus_int16 *NLSF, /* I normalized line spectral frequencies in Q15, [ d ] */ - const opus_int d /* I filter order (should be even) */ + const opus_int d, /* I filter order (should be even) */ + int arch /* I Run-time architecture */ +); + +/* Convert int32 coefficients to int16 coefs and make sure there's no wrap-around */ +void silk_LPC_fit( + opus_int16 *a_QOUT, /* O Output signal */ + opus_int32 *a_QIN, /* I/O Input signal */ + const opus_int QOUT, /* I Input Q domain */ + const opus_int QIN, /* I Input Q domain */ + const opus_int d /* I Filter order */ ); void silk_insertion_sort_increasing( @@ -471,8 +496,7 @@ static OPUS_INLINE opus_int32 silk_ROR32( opus_int32 a32, opus_int rot ) /* Add with saturation for positive input values */ #define silk_ADD_POS_SAT8(a, b) ((((a)+(b)) & 0x80) ? silk_int8_MAX : ((a)+(b))) #define silk_ADD_POS_SAT16(a, b) ((((a)+(b)) & 0x8000) ? silk_int16_MAX : ((a)+(b))) -#define silk_ADD_POS_SAT32(a, b) ((((a)+(b)) & 0x80000000) ? silk_int32_MAX : ((a)+(b))) -#define silk_ADD_POS_SAT64(a, b) ((((a)+(b)) & 0x8000000000000000LL) ? silk_int64_MAX : ((a)+(b))) +#define silk_ADD_POS_SAT32(a, b) ((((opus_uint32)(a)+(opus_uint32)(b)) & 0x80000000) ? silk_int32_MAX : ((a)+(b))) #define silk_LSHIFT8(a, shift) ((opus_int8)((opus_uint8)(a)<<(shift))) /* shift >= 0, shift < 8 */ #define silk_LSHIFT16(a, shift) ((opus_int16)((opus_uint16)(a)<<(shift))) /* shift >= 0, shift < 16 */ @@ -572,7 +596,9 @@ static OPUS_INLINE opus_int64 silk_max_64(opus_int64 a, opus_int64 b) /* Make sure to store the result as the seed for the next call (also in between */ /* frames), otherwise result won't be random at all. When only using some of the */ /* bits, take the most significant bits by right-shifting. */ -#define silk_RAND(seed) (silk_MLA_ovflw(907633515, (seed), 196314165)) +#define RAND_MULTIPLIER 196314165 +#define RAND_INCREMENT 907633515 +#define silk_RAND(seed) (silk_MLA_ovflw((RAND_INCREMENT), (seed), (RAND_MULTIPLIER))) /* Add some multiplication functions that can be easily mapped to ARM. */ diff --git a/code/opus-1.1.4/silk/VQ_WMat_EC.c b/code/opus-1.1.4/silk/VQ_WMat_EC.c index 7983f1db..0f3d545c 100644 --- a/code/opus-1.1.4/silk/VQ_WMat_EC.c +++ b/code/opus-1.1.4/silk/VQ_WMat_EC.c @@ -34,84 +34,95 @@ POSSIBILITY OF SUCH DAMAGE. /* Entropy constrained matrix-weighted VQ, hard-coded to 5-element vectors, for a single input data vector */ void silk_VQ_WMat_EC_c( opus_int8 *ind, /* O index of best codebook vector */ - opus_int32 *rate_dist_Q14, /* O best weighted quant error + mu * rate */ + opus_int32 *res_nrg_Q15, /* O best residual energy */ + opus_int32 *rate_dist_Q8, /* O best total bitrate */ opus_int *gain_Q7, /* O sum of absolute LTP coefficients */ - const opus_int16 *in_Q14, /* I input vector to be quantized */ - const opus_int32 *W_Q18, /* I weighting matrix */ + const opus_int32 *XX_Q17, /* I correlation matrix */ + const opus_int32 *xX_Q17, /* I correlation vector */ const opus_int8 *cb_Q7, /* I codebook */ const opus_uint8 *cb_gain_Q7, /* I codebook effective gain */ const opus_uint8 *cl_Q5, /* I code length for each codebook vector */ - const opus_int mu_Q9, /* I tradeoff betw. weighted error and rate */ + const opus_int subfr_len, /* I number of samples per subframe */ const opus_int32 max_gain_Q7, /* I maximum sum of absolute LTP coefficients */ - opus_int L /* I number of vectors in codebook */ + const opus_int L /* I number of vectors in codebook */ ) { opus_int k, gain_tmp_Q7; const opus_int8 *cb_row_Q7; - opus_int16 diff_Q14[ 5 ]; - opus_int32 sum1_Q14, sum2_Q16; + opus_int32 neg_xX_Q24[ 5 ]; + opus_int32 sum1_Q15, sum2_Q24; + opus_int32 bits_res_Q8, bits_tot_Q8; + + /* Negate and convert to new Q domain */ + neg_xX_Q24[ 0 ] = -silk_LSHIFT32( xX_Q17[ 0 ], 7 ); + neg_xX_Q24[ 1 ] = -silk_LSHIFT32( xX_Q17[ 1 ], 7 ); + neg_xX_Q24[ 2 ] = -silk_LSHIFT32( xX_Q17[ 2 ], 7 ); + neg_xX_Q24[ 3 ] = -silk_LSHIFT32( xX_Q17[ 3 ], 7 ); + neg_xX_Q24[ 4 ] = -silk_LSHIFT32( xX_Q17[ 4 ], 7 ); /* Loop over codebook */ - *rate_dist_Q14 = silk_int32_MAX; + *rate_dist_Q8 = silk_int32_MAX; + *res_nrg_Q15 = silk_int32_MAX; cb_row_Q7 = cb_Q7; + /* In things go really bad, at least *ind is set to something safe. */ + *ind = 0; for( k = 0; k < L; k++ ) { + opus_int32 penalty; gain_tmp_Q7 = cb_gain_Q7[k]; - - diff_Q14[ 0 ] = in_Q14[ 0 ] - silk_LSHIFT( cb_row_Q7[ 0 ], 7 ); - diff_Q14[ 1 ] = in_Q14[ 1 ] - silk_LSHIFT( cb_row_Q7[ 1 ], 7 ); - diff_Q14[ 2 ] = in_Q14[ 2 ] - silk_LSHIFT( cb_row_Q7[ 2 ], 7 ); - diff_Q14[ 3 ] = in_Q14[ 3 ] - silk_LSHIFT( cb_row_Q7[ 3 ], 7 ); - diff_Q14[ 4 ] = in_Q14[ 4 ] - silk_LSHIFT( cb_row_Q7[ 4 ], 7 ); - /* Weighted rate */ - sum1_Q14 = silk_SMULBB( mu_Q9, cl_Q5[ k ] ); + /* Quantization error: 1 - 2 * xX * cb + cb' * XX * cb */ + sum1_Q15 = SILK_FIX_CONST( 1.001, 15 ); /* Penalty for too large gain */ - sum1_Q14 = silk_ADD_LSHIFT32( sum1_Q14, silk_max( silk_SUB32( gain_tmp_Q7, max_gain_Q7 ), 0 ), 10 ); + penalty = silk_LSHIFT32( silk_max( silk_SUB32( gain_tmp_Q7, max_gain_Q7 ), 0 ), 11 ); - silk_assert( sum1_Q14 >= 0 ); + /* first row of XX_Q17 */ + sum2_Q24 = silk_MLA( neg_xX_Q24[ 0 ], XX_Q17[ 1 ], cb_row_Q7[ 1 ] ); + sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 2 ], cb_row_Q7[ 2 ] ); + sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 3 ], cb_row_Q7[ 3 ] ); + sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 4 ], cb_row_Q7[ 4 ] ); + sum2_Q24 = silk_LSHIFT32( sum2_Q24, 1 ); + sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 0 ], cb_row_Q7[ 0 ] ); + sum1_Q15 = silk_SMLAWB( sum1_Q15, sum2_Q24, cb_row_Q7[ 0 ] ); - /* first row of W_Q18 */ - sum2_Q16 = silk_SMULWB( W_Q18[ 1 ], diff_Q14[ 1 ] ); - sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 2 ], diff_Q14[ 2 ] ); - sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 3 ], diff_Q14[ 3 ] ); - sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 4 ], diff_Q14[ 4 ] ); - sum2_Q16 = silk_LSHIFT( sum2_Q16, 1 ); - sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 0 ], diff_Q14[ 0 ] ); - sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16, diff_Q14[ 0 ] ); + /* second row of XX_Q17 */ + sum2_Q24 = silk_MLA( neg_xX_Q24[ 1 ], XX_Q17[ 7 ], cb_row_Q7[ 2 ] ); + sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 8 ], cb_row_Q7[ 3 ] ); + sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 9 ], cb_row_Q7[ 4 ] ); + sum2_Q24 = silk_LSHIFT32( sum2_Q24, 1 ); + sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 6 ], cb_row_Q7[ 1 ] ); + sum1_Q15 = silk_SMLAWB( sum1_Q15, sum2_Q24, cb_row_Q7[ 1 ] ); - /* second row of W_Q18 */ - sum2_Q16 = silk_SMULWB( W_Q18[ 7 ], diff_Q14[ 2 ] ); - sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 8 ], diff_Q14[ 3 ] ); - sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 9 ], diff_Q14[ 4 ] ); - sum2_Q16 = silk_LSHIFT( sum2_Q16, 1 ); - sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 6 ], diff_Q14[ 1 ] ); - sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16, diff_Q14[ 1 ] ); + /* third row of XX_Q17 */ + sum2_Q24 = silk_MLA( neg_xX_Q24[ 2 ], XX_Q17[ 13 ], cb_row_Q7[ 3 ] ); + sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 14 ], cb_row_Q7[ 4 ] ); + sum2_Q24 = silk_LSHIFT32( sum2_Q24, 1 ); + sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 12 ], cb_row_Q7[ 2 ] ); + sum1_Q15 = silk_SMLAWB( sum1_Q15, sum2_Q24, cb_row_Q7[ 2 ] ); - /* third row of W_Q18 */ - sum2_Q16 = silk_SMULWB( W_Q18[ 13 ], diff_Q14[ 3 ] ); - sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 14 ], diff_Q14[ 4 ] ); - sum2_Q16 = silk_LSHIFT( sum2_Q16, 1 ); - sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 12 ], diff_Q14[ 2 ] ); - sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16, diff_Q14[ 2 ] ); + /* fourth row of XX_Q17 */ + sum2_Q24 = silk_MLA( neg_xX_Q24[ 3 ], XX_Q17[ 19 ], cb_row_Q7[ 4 ] ); + sum2_Q24 = silk_LSHIFT32( sum2_Q24, 1 ); + sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 18 ], cb_row_Q7[ 3 ] ); + sum1_Q15 = silk_SMLAWB( sum1_Q15, sum2_Q24, cb_row_Q7[ 3 ] ); - /* fourth row of W_Q18 */ - sum2_Q16 = silk_SMULWB( W_Q18[ 19 ], diff_Q14[ 4 ] ); - sum2_Q16 = silk_LSHIFT( sum2_Q16, 1 ); - sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 18 ], diff_Q14[ 3 ] ); - sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16, diff_Q14[ 3 ] ); - - /* last row of W_Q18 */ - sum2_Q16 = silk_SMULWB( W_Q18[ 24 ], diff_Q14[ 4 ] ); - sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16, diff_Q14[ 4 ] ); - - silk_assert( sum1_Q14 >= 0 ); + /* last row of XX_Q17 */ + sum2_Q24 = silk_LSHIFT32( neg_xX_Q24[ 4 ], 1 ); + sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 24 ], cb_row_Q7[ 4 ] ); + sum1_Q15 = silk_SMLAWB( sum1_Q15, sum2_Q24, cb_row_Q7[ 4 ] ); /* find best */ - if( sum1_Q14 < *rate_dist_Q14 ) { - *rate_dist_Q14 = sum1_Q14; - *ind = (opus_int8)k; - *gain_Q7 = gain_tmp_Q7; + if( sum1_Q15 >= 0 ) { + /* Translate residual energy to bits using high-rate assumption (6 dB ==> 1 bit/sample) */ + bits_res_Q8 = silk_SMULBB( subfr_len, silk_lin2log( sum1_Q15 + penalty) - (15 << 7) ); + /* In the following line we reduce the codelength component by half ("-1"); seems to slghtly improve quality */ + bits_tot_Q8 = silk_ADD_LSHIFT32( bits_res_Q8, cl_Q5[ k ], 3-1 ); + if( bits_tot_Q8 <= *rate_dist_Q8 ) { + *rate_dist_Q8 = bits_tot_Q8; + *res_nrg_Q15 = sum1_Q15 + penalty; + *ind = (opus_int8)k; + *gain_Q7 = gain_tmp_Q7; + } } /* Go to next cbk vector */ diff --git a/code/opus-1.1.4/silk/arm/LPC_inv_pred_gain_arm.h b/code/opus-1.1.4/silk/arm/LPC_inv_pred_gain_arm.h new file mode 100644 index 00000000..9895b555 --- /dev/null +++ b/code/opus-1.1.4/silk/arm/LPC_inv_pred_gain_arm.h @@ -0,0 +1,57 @@ +/*********************************************************************** +Copyright (c) 2017 Google Inc. +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions +are met: +- Redistributions of source code must retain the above copyright notice, +this list of conditions and the following disclaimer. +- Redistributions in binary form must reproduce the above copyright +notice, this list of conditions and the following disclaimer in the +documentation and/or other materials provided with the distribution. +- Neither the name of Internet Society, IETF or IETF Trust, nor the +names of specific contributors, may be used to endorse or promote +products derived from this software without specific prior written +permission. +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE +ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE +LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR +CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF +SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS +INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN +CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) +ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +POSSIBILITY OF SUCH DAMAGE. +***********************************************************************/ + +#ifndef SILK_LPC_INV_PRED_GAIN_ARM_H +# define SILK_LPC_INV_PRED_GAIN_ARM_H + +# include "celt/arm/armcpu.h" + +# if defined(OPUS_ARM_MAY_HAVE_NEON_INTR) +opus_int32 silk_LPC_inverse_pred_gain_neon( /* O Returns inverse prediction gain in energy domain, Q30 */ + const opus_int16 *A_Q12, /* I Prediction coefficients, Q12 [order] */ + const opus_int order /* I Prediction order */ +); + +# if !defined(OPUS_HAVE_RTCD) && defined(OPUS_ARM_PRESUME_NEON) +# define OVERRIDE_silk_LPC_inverse_pred_gain (1) +# define silk_LPC_inverse_pred_gain(A_Q12, order, arch) ((void)(arch), PRESUME_NEON(silk_LPC_inverse_pred_gain)(A_Q12, order)) +# endif +# endif + +# if !defined(OVERRIDE_silk_LPC_inverse_pred_gain) +/*Is run-time CPU detection enabled on this platform?*/ +# if defined(OPUS_HAVE_RTCD) && (defined(OPUS_ARM_MAY_HAVE_NEON_INTR) && !defined(OPUS_ARM_PRESUME_NEON_INTR)) +extern opus_int32 (*const SILK_LPC_INVERSE_PRED_GAIN_IMPL[OPUS_ARCHMASK+1])(const opus_int16 *A_Q12, const opus_int order); +# define OVERRIDE_silk_LPC_inverse_pred_gain (1) +# define silk_LPC_inverse_pred_gain(A_Q12, order, arch) ((*SILK_LPC_INVERSE_PRED_GAIN_IMPL[(arch)&OPUS_ARCHMASK])(A_Q12, order)) +# elif defined(OPUS_ARM_PRESUME_NEON_INTR) +# define OVERRIDE_silk_LPC_inverse_pred_gain (1) +# define silk_LPC_inverse_pred_gain(A_Q12, order, arch) ((void)(arch), silk_LPC_inverse_pred_gain_neon(A_Q12, order)) +# endif +# endif + +#endif /* end SILK_LPC_INV_PRED_GAIN_ARM_H */ diff --git a/code/opus-1.1.4/silk/arm/LPC_inv_pred_gain_neon_intr.c b/code/opus-1.1.4/silk/arm/LPC_inv_pred_gain_neon_intr.c new file mode 100644 index 00000000..27142f34 --- /dev/null +++ b/code/opus-1.1.4/silk/arm/LPC_inv_pred_gain_neon_intr.c @@ -0,0 +1,280 @@ +/*********************************************************************** +Copyright (c) 2017 Google Inc. +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions +are met: +- Redistributions of source code must retain the above copyright notice, +this list of conditions and the following disclaimer. +- Redistributions in binary form must reproduce the above copyright +notice, this list of conditions and the following disclaimer in the +documentation and/or other materials provided with the distribution. +- Neither the name of Internet Society, IETF or IETF Trust, nor the +names of specific contributors, may be used to endorse or promote +products derived from this software without specific prior written +permission. +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE +ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE +LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR +CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF +SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS +INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN +CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) +ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +POSSIBILITY OF SUCH DAMAGE. +***********************************************************************/ + +#ifdef HAVE_CONFIG_H +#include "config.h" +#endif + +#include +#include "SigProc_FIX.h" +#include "define.h" + +#define QA 24 +#define A_LIMIT SILK_FIX_CONST( 0.99975, QA ) + +#define MUL32_FRAC_Q(a32, b32, Q) ((opus_int32)(silk_RSHIFT_ROUND64(silk_SMULL(a32, b32), Q))) + +/* The difficulty is how to judge a 64-bit signed integer tmp64 is 32-bit overflowed, + * since NEON has no 64-bit min, max or comparison instructions. + * A failed idea is to compare the results of vmovn(tmp64) and vqmovn(tmp64) whether they are equal or not. + * However, this idea fails when the tmp64 is something like 0xFFFFFFF980000000. + * Here we know that mult2Q >= 1, so the highest bit (bit 63, sign bit) of tmp64 must equal to bit 62. + * tmp64 was shifted left by 1 and we got tmp64'. If high_half(tmp64') != 0 and high_half(tmp64') != -1, + * then we know that bit 31 to bit 63 of tmp64 can not all be the sign bit, and therefore tmp64 is 32-bit overflowed. + * That is, we judge if tmp64' > 0x00000000FFFFFFFF, or tmp64' <= 0xFFFFFFFF00000000. + * We use narrowing shift right 31 bits to tmp32' to save data bandwidth and instructions. + * That is, we judge if tmp32' > 0x00000000, or tmp32' <= 0xFFFFFFFF. + */ + +/* Compute inverse of LPC prediction gain, and */ +/* test if LPC coefficients are stable (all poles within unit circle) */ +static OPUS_INLINE opus_int32 LPC_inverse_pred_gain_QA_neon( /* O Returns inverse prediction gain in energy domain, Q30 */ + opus_int32 A_QA[ SILK_MAX_ORDER_LPC ], /* I Prediction coefficients */ + const opus_int order /* I Prediction order */ +) +{ + opus_int k, n, mult2Q; + opus_int32 invGain_Q30, rc_Q31, rc_mult1_Q30, rc_mult2, tmp1, tmp2; + opus_int32 max, min; + int32x4_t max_s32x4, min_s32x4; + int32x2_t max_s32x2, min_s32x2; + + max_s32x4 = vdupq_n_s32( silk_int32_MIN ); + min_s32x4 = vdupq_n_s32( silk_int32_MAX ); + invGain_Q30 = SILK_FIX_CONST( 1, 30 ); + for( k = order - 1; k > 0; k-- ) { + int32x2_t rc_Q31_s32x2, rc_mult2_s32x2; + int64x2_t mult2Q_s64x2; + + /* Check for stability */ + if( ( A_QA[ k ] > A_LIMIT ) || ( A_QA[ k ] < -A_LIMIT ) ) { + return 0; + } + + /* Set RC equal to negated AR coef */ + rc_Q31 = -silk_LSHIFT( A_QA[ k ], 31 - QA ); + + /* rc_mult1_Q30 range: [ 1 : 2^30 ] */ + rc_mult1_Q30 = silk_SUB32( SILK_FIX_CONST( 1, 30 ), silk_SMMUL( rc_Q31, rc_Q31 ) ); + silk_assert( rc_mult1_Q30 > ( 1 << 15 ) ); /* reduce A_LIMIT if fails */ + silk_assert( rc_mult1_Q30 <= ( 1 << 30 ) ); + + /* Update inverse gain */ + /* invGain_Q30 range: [ 0 : 2^30 ] */ + invGain_Q30 = silk_LSHIFT( silk_SMMUL( invGain_Q30, rc_mult1_Q30 ), 2 ); + silk_assert( invGain_Q30 >= 0 ); + silk_assert( invGain_Q30 <= ( 1 << 30 ) ); + if( invGain_Q30 < SILK_FIX_CONST( 1.0f / MAX_PREDICTION_POWER_GAIN, 30 ) ) { + return 0; + } + + /* rc_mult2 range: [ 2^30 : silk_int32_MAX ] */ + mult2Q = 32 - silk_CLZ32( silk_abs( rc_mult1_Q30 ) ); + rc_mult2 = silk_INVERSE32_varQ( rc_mult1_Q30, mult2Q + 30 ); + + /* Update AR coefficient */ + rc_Q31_s32x2 = vdup_n_s32( rc_Q31 ); + mult2Q_s64x2 = vdupq_n_s64( -mult2Q ); + rc_mult2_s32x2 = vdup_n_s32( rc_mult2 ); + + for( n = 0; n < ( ( k + 1 ) >> 1 ) - 3; n += 4 ) { + /* We always calculate extra elements of A_QA buffer when ( k % 4 ) != 0, to take the advantage of SIMD parallelization. */ + int32x4_t tmp1_s32x4, tmp2_s32x4, t0_s32x4, t1_s32x4, s0_s32x4, s1_s32x4, t_QA0_s32x4, t_QA1_s32x4; + int64x2_t t0_s64x2, t1_s64x2, t2_s64x2, t3_s64x2; + tmp1_s32x4 = vld1q_s32( A_QA + n ); + tmp2_s32x4 = vld1q_s32( A_QA + k - n - 4 ); + tmp2_s32x4 = vrev64q_s32( tmp2_s32x4 ); + tmp2_s32x4 = vcombine_s32( vget_high_s32( tmp2_s32x4 ), vget_low_s32( tmp2_s32x4 ) ); + t0_s32x4 = vqrdmulhq_lane_s32( tmp2_s32x4, rc_Q31_s32x2, 0 ); + t1_s32x4 = vqrdmulhq_lane_s32( tmp1_s32x4, rc_Q31_s32x2, 0 ); + t_QA0_s32x4 = vqsubq_s32( tmp1_s32x4, t0_s32x4 ); + t_QA1_s32x4 = vqsubq_s32( tmp2_s32x4, t1_s32x4 ); + t0_s64x2 = vmull_s32( vget_low_s32 ( t_QA0_s32x4 ), rc_mult2_s32x2 ); + t1_s64x2 = vmull_s32( vget_high_s32( t_QA0_s32x4 ), rc_mult2_s32x2 ); + t2_s64x2 = vmull_s32( vget_low_s32 ( t_QA1_s32x4 ), rc_mult2_s32x2 ); + t3_s64x2 = vmull_s32( vget_high_s32( t_QA1_s32x4 ), rc_mult2_s32x2 ); + t0_s64x2 = vrshlq_s64( t0_s64x2, mult2Q_s64x2 ); + t1_s64x2 = vrshlq_s64( t1_s64x2, mult2Q_s64x2 ); + t2_s64x2 = vrshlq_s64( t2_s64x2, mult2Q_s64x2 ); + t3_s64x2 = vrshlq_s64( t3_s64x2, mult2Q_s64x2 ); + t0_s32x4 = vcombine_s32( vmovn_s64( t0_s64x2 ), vmovn_s64( t1_s64x2 ) ); + t1_s32x4 = vcombine_s32( vmovn_s64( t2_s64x2 ), vmovn_s64( t3_s64x2 ) ); + s0_s32x4 = vcombine_s32( vshrn_n_s64( t0_s64x2, 31 ), vshrn_n_s64( t1_s64x2, 31 ) ); + s1_s32x4 = vcombine_s32( vshrn_n_s64( t2_s64x2, 31 ), vshrn_n_s64( t3_s64x2, 31 ) ); + max_s32x4 = vmaxq_s32( max_s32x4, s0_s32x4 ); + min_s32x4 = vminq_s32( min_s32x4, s0_s32x4 ); + max_s32x4 = vmaxq_s32( max_s32x4, s1_s32x4 ); + min_s32x4 = vminq_s32( min_s32x4, s1_s32x4 ); + t1_s32x4 = vrev64q_s32( t1_s32x4 ); + t1_s32x4 = vcombine_s32( vget_high_s32( t1_s32x4 ), vget_low_s32( t1_s32x4 ) ); + vst1q_s32( A_QA + n, t0_s32x4 ); + vst1q_s32( A_QA + k - n - 4, t1_s32x4 ); + } + for( ; n < (k + 1) >> 1; n++ ) { + opus_int64 tmp64; + tmp1 = A_QA[ n ]; + tmp2 = A_QA[ k - n - 1 ]; + tmp64 = silk_RSHIFT_ROUND64( silk_SMULL( silk_SUB_SAT32(tmp1, + MUL32_FRAC_Q( tmp2, rc_Q31, 31 ) ), rc_mult2 ), mult2Q); + if( tmp64 > silk_int32_MAX || tmp64 < silk_int32_MIN ) { + return 0; + } + A_QA[ n ] = ( opus_int32 )tmp64; + tmp64 = silk_RSHIFT_ROUND64( silk_SMULL( silk_SUB_SAT32(tmp2, + MUL32_FRAC_Q( tmp1, rc_Q31, 31 ) ), rc_mult2), mult2Q); + if( tmp64 > silk_int32_MAX || tmp64 < silk_int32_MIN ) { + return 0; + } + A_QA[ k - n - 1 ] = ( opus_int32 )tmp64; + } + } + + /* Check for stability */ + if( ( A_QA[ k ] > A_LIMIT ) || ( A_QA[ k ] < -A_LIMIT ) ) { + return 0; + } + + max_s32x2 = vmax_s32( vget_low_s32( max_s32x4 ), vget_high_s32( max_s32x4 ) ); + min_s32x2 = vmin_s32( vget_low_s32( min_s32x4 ), vget_high_s32( min_s32x4 ) ); + max_s32x2 = vmax_s32( max_s32x2, vreinterpret_s32_s64( vshr_n_s64( vreinterpret_s64_s32( max_s32x2 ), 32 ) ) ); + min_s32x2 = vmin_s32( min_s32x2, vreinterpret_s32_s64( vshr_n_s64( vreinterpret_s64_s32( min_s32x2 ), 32 ) ) ); + max = vget_lane_s32( max_s32x2, 0 ); + min = vget_lane_s32( min_s32x2, 0 ); + if( ( max > 0 ) || ( min < -1 ) ) { + return 0; + } + + /* Set RC equal to negated AR coef */ + rc_Q31 = -silk_LSHIFT( A_QA[ 0 ], 31 - QA ); + + /* Range: [ 1 : 2^30 ] */ + rc_mult1_Q30 = silk_SUB32( SILK_FIX_CONST( 1, 30 ), silk_SMMUL( rc_Q31, rc_Q31 ) ); + + /* Update inverse gain */ + /* Range: [ 0 : 2^30 ] */ + invGain_Q30 = silk_LSHIFT( silk_SMMUL( invGain_Q30, rc_mult1_Q30 ), 2 ); + silk_assert( invGain_Q30 >= 0 ); + silk_assert( invGain_Q30 <= ( 1 << 30 ) ); + if( invGain_Q30 < SILK_FIX_CONST( 1.0f / MAX_PREDICTION_POWER_GAIN, 30 ) ) { + return 0; + } + + return invGain_Q30; +} + +/* For input in Q12 domain */ +opus_int32 silk_LPC_inverse_pred_gain_neon( /* O Returns inverse prediction gain in energy domain, Q30 */ + const opus_int16 *A_Q12, /* I Prediction coefficients, Q12 [order] */ + const opus_int order /* I Prediction order */ +) +{ +#ifdef OPUS_CHECK_ASM + const opus_int32 invGain_Q30_c = silk_LPC_inverse_pred_gain_c( A_Q12, order ); +#endif + + opus_int32 invGain_Q30; + if( ( SILK_MAX_ORDER_LPC != 24 ) || ( order & 1 )) { + invGain_Q30 = silk_LPC_inverse_pred_gain_c( A_Q12, order ); + } + else { + opus_int32 Atmp_QA[ SILK_MAX_ORDER_LPC ]; + opus_int32 DC_resp; + int16x8_t t0_s16x8, t1_s16x8, t2_s16x8; + int32x4_t t0_s32x4; + const opus_int leftover = order & 7; + + /* Increase Q domain of the AR coefficients */ + t0_s16x8 = vld1q_s16( A_Q12 + 0 ); + t1_s16x8 = vld1q_s16( A_Q12 + 8 ); + t2_s16x8 = vld1q_s16( A_Q12 + 16 ); + t0_s32x4 = vpaddlq_s16( t0_s16x8 ); + + switch( order - leftover ) + { + case 24: + t0_s32x4 = vpadalq_s16( t0_s32x4, t2_s16x8 ); + /* Intend to fall through */ + + case 16: + t0_s32x4 = vpadalq_s16( t0_s32x4, t1_s16x8 ); + vst1q_s32( Atmp_QA + 16, vshll_n_s16( vget_low_s16 ( t2_s16x8 ), QA - 12 ) ); + vst1q_s32( Atmp_QA + 20, vshll_n_s16( vget_high_s16( t2_s16x8 ), QA - 12 ) ); + /* Intend to fall through */ + + case 8: + { + const int32x2_t t_s32x2 = vpadd_s32( vget_low_s32( t0_s32x4 ), vget_high_s32( t0_s32x4 ) ); + const int64x1_t t_s64x1 = vpaddl_s32( t_s32x2 ); + DC_resp = vget_lane_s32( vreinterpret_s32_s64( t_s64x1 ), 0 ); + vst1q_s32( Atmp_QA + 8, vshll_n_s16( vget_low_s16 ( t1_s16x8 ), QA - 12 ) ); + vst1q_s32( Atmp_QA + 12, vshll_n_s16( vget_high_s16( t1_s16x8 ), QA - 12 ) ); + } + break; + + default: + DC_resp = 0; + break; + } + A_Q12 += order - leftover; + + switch( leftover ) + { + case 6: + DC_resp += (opus_int32)A_Q12[ 5 ]; + DC_resp += (opus_int32)A_Q12[ 4 ]; + /* Intend to fall through */ + + case 4: + DC_resp += (opus_int32)A_Q12[ 3 ]; + DC_resp += (opus_int32)A_Q12[ 2 ]; + /* Intend to fall through */ + + case 2: + DC_resp += (opus_int32)A_Q12[ 1 ]; + DC_resp += (opus_int32)A_Q12[ 0 ]; + /* Intend to fall through */ + + default: + break; + } + + /* If the DC is unstable, we don't even need to do the full calculations */ + if( DC_resp >= 4096 ) { + invGain_Q30 = 0; + } else { + vst1q_s32( Atmp_QA + 0, vshll_n_s16( vget_low_s16 ( t0_s16x8 ), QA - 12 ) ); + vst1q_s32( Atmp_QA + 4, vshll_n_s16( vget_high_s16( t0_s16x8 ), QA - 12 ) ); + invGain_Q30 = LPC_inverse_pred_gain_QA_neon( Atmp_QA, order ); + } + } + +#ifdef OPUS_CHECK_ASM + silk_assert( invGain_Q30_c == invGain_Q30 ); +#endif + + return invGain_Q30; +} diff --git a/code/opus-1.1.4/silk/arm/NSQ_del_dec_arm.h b/code/opus-1.1.4/silk/arm/NSQ_del_dec_arm.h new file mode 100644 index 00000000..9e76e169 --- /dev/null +++ b/code/opus-1.1.4/silk/arm/NSQ_del_dec_arm.h @@ -0,0 +1,100 @@ +/*********************************************************************** +Copyright (c) 2017 Google Inc. +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions +are met: +- Redistributions of source code must retain the above copyright notice, +this list of conditions and the following disclaimer. +- Redistributions in binary form must reproduce the above copyright +notice, this list of conditions and the following disclaimer in the +documentation and/or other materials provided with the distribution. +- Neither the name of Internet Society, IETF or IETF Trust, nor the +names of specific contributors, may be used to endorse or promote +products derived from this software without specific prior written +permission. +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE +ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE +LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR +CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF +SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS +INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN +CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) +ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +POSSIBILITY OF SUCH DAMAGE. +***********************************************************************/ + +#ifndef SILK_NSQ_DEL_DEC_ARM_H +#define SILK_NSQ_DEL_DEC_ARM_H + +#include "celt/arm/armcpu.h" + +#if defined(OPUS_ARM_MAY_HAVE_NEON_INTR) +void silk_NSQ_del_dec_neon( + const silk_encoder_state *psEncC, silk_nsq_state *NSQ, + SideInfoIndices *psIndices, const opus_int16 x16[], opus_int8 pulses[], + const opus_int16 PredCoef_Q12[2 * MAX_LPC_ORDER], + const opus_int16 LTPCoef_Q14[LTP_ORDER * MAX_NB_SUBFR], + const opus_int16 AR_Q13[MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER], + const opus_int HarmShapeGain_Q14[MAX_NB_SUBFR], + const opus_int Tilt_Q14[MAX_NB_SUBFR], + const opus_int32 LF_shp_Q14[MAX_NB_SUBFR], + const opus_int32 Gains_Q16[MAX_NB_SUBFR], + const opus_int pitchL[MAX_NB_SUBFR], const opus_int Lambda_Q10, + const opus_int LTP_scale_Q14); + +#if !defined(OPUS_HAVE_RTCD) +#define OVERRIDE_silk_NSQ_del_dec (1) +#define silk_NSQ_del_dec(psEncC, NSQ, psIndices, x16, pulses, PredCoef_Q12, \ + LTPCoef_Q14, AR_Q13, HarmShapeGain_Q14, Tilt_Q14, \ + LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, \ + LTP_scale_Q14, arch) \ + ((void)(arch), \ + PRESUME_NEON(silk_NSQ_del_dec)( \ + psEncC, NSQ, psIndices, x16, pulses, PredCoef_Q12, LTPCoef_Q14, \ + AR_Q13, HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, pitchL, \ + Lambda_Q10, LTP_scale_Q14)) +#endif +#endif + +#if !defined(OVERRIDE_silk_NSQ_del_dec) +/*Is run-time CPU detection enabled on this platform?*/ +#if defined(OPUS_HAVE_RTCD) && (defined(OPUS_ARM_MAY_HAVE_NEON_INTR) && \ + !defined(OPUS_ARM_PRESUME_NEON_INTR)) +extern void (*const SILK_NSQ_DEL_DEC_IMPL[OPUS_ARCHMASK + 1])( + const silk_encoder_state *psEncC, silk_nsq_state *NSQ, + SideInfoIndices *psIndices, const opus_int16 x16[], opus_int8 pulses[], + const opus_int16 PredCoef_Q12[2 * MAX_LPC_ORDER], + const opus_int16 LTPCoef_Q14[LTP_ORDER * MAX_NB_SUBFR], + const opus_int16 AR_Q13[MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER], + const opus_int HarmShapeGain_Q14[MAX_NB_SUBFR], + const opus_int Tilt_Q14[MAX_NB_SUBFR], + const opus_int32 LF_shp_Q14[MAX_NB_SUBFR], + const opus_int32 Gains_Q16[MAX_NB_SUBFR], + const opus_int pitchL[MAX_NB_SUBFR], const opus_int Lambda_Q10, + const opus_int LTP_scale_Q14); +#define OVERRIDE_silk_NSQ_del_dec (1) +#define silk_NSQ_del_dec(psEncC, NSQ, psIndices, x16, pulses, PredCoef_Q12, \ + LTPCoef_Q14, AR_Q13, HarmShapeGain_Q14, Tilt_Q14, \ + LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, \ + LTP_scale_Q14, arch) \ + ((*SILK_NSQ_DEL_DEC_IMPL[(arch)&OPUS_ARCHMASK])( \ + psEncC, NSQ, psIndices, x16, pulses, PredCoef_Q12, LTPCoef_Q14, \ + AR_Q13, HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, pitchL, \ + Lambda_Q10, LTP_scale_Q14)) +#elif defined(OPUS_ARM_PRESUME_NEON_INTR) +#define OVERRIDE_silk_NSQ_del_dec (1) +#define silk_NSQ_del_dec(psEncC, NSQ, psIndices, x16, pulses, PredCoef_Q12, \ + LTPCoef_Q14, AR_Q13, HarmShapeGain_Q14, Tilt_Q14, \ + LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, \ + LTP_scale_Q14, arch) \ + ((void)(arch), \ + silk_NSQ_del_dec_neon(psEncC, NSQ, psIndices, x16, pulses, PredCoef_Q12, \ + LTPCoef_Q14, AR_Q13, HarmShapeGain_Q14, Tilt_Q14, \ + LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, \ + LTP_scale_Q14)) +#endif +#endif + +#endif /* end SILK_NSQ_DEL_DEC_ARM_H */ diff --git a/code/opus-1.1.4/silk/arm/NSQ_del_dec_neon_intr.c b/code/opus-1.1.4/silk/arm/NSQ_del_dec_neon_intr.c new file mode 100644 index 00000000..212410f3 --- /dev/null +++ b/code/opus-1.1.4/silk/arm/NSQ_del_dec_neon_intr.c @@ -0,0 +1,1124 @@ +/*********************************************************************** +Copyright (c) 2017 Google Inc. +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions +are met: +- Redistributions of source code must retain the above copyright notice, +this list of conditions and the following disclaimer. +- Redistributions in binary form must reproduce the above copyright +notice, this list of conditions and the following disclaimer in the +documentation and/or other materials provided with the distribution. +- Neither the name of Internet Society, IETF or IETF Trust, nor the +names of specific contributors, may be used to endorse or promote +products derived from this software without specific prior written +permission. +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE +ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE +LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR +CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF +SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS +INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN +CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) +ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +POSSIBILITY OF SUCH DAMAGE. +***********************************************************************/ + +#ifdef HAVE_CONFIG_H +#include "config.h" +#endif + +#include +#ifdef OPUS_CHECK_ASM +# include +#endif +#include "main.h" +#include "stack_alloc.h" + +/* NEON intrinsics optimization now can only parallelize up to 4 delay decision states. */ +/* If there are more states, C function is called, and this optimization must be expanded. */ +#define NEON_MAX_DEL_DEC_STATES 4 + +typedef struct { + opus_int32 sLPC_Q14[ MAX_SUB_FRAME_LENGTH + NSQ_LPC_BUF_LENGTH ][ NEON_MAX_DEL_DEC_STATES ]; + opus_int32 RandState[ DECISION_DELAY ][ NEON_MAX_DEL_DEC_STATES ]; + opus_int32 Q_Q10[ DECISION_DELAY ][ NEON_MAX_DEL_DEC_STATES ]; + opus_int32 Xq_Q14[ DECISION_DELAY ][ NEON_MAX_DEL_DEC_STATES ]; + opus_int32 Pred_Q15[ DECISION_DELAY ][ NEON_MAX_DEL_DEC_STATES ]; + opus_int32 Shape_Q14[ DECISION_DELAY ][ NEON_MAX_DEL_DEC_STATES ]; + opus_int32 sAR2_Q14[ MAX_SHAPE_LPC_ORDER ][ NEON_MAX_DEL_DEC_STATES ]; + opus_int32 LF_AR_Q14[ NEON_MAX_DEL_DEC_STATES ]; + opus_int32 Diff_Q14[ NEON_MAX_DEL_DEC_STATES ]; + opus_int32 Seed[ NEON_MAX_DEL_DEC_STATES ]; + opus_int32 SeedInit[ NEON_MAX_DEL_DEC_STATES ]; + opus_int32 RD_Q10[ NEON_MAX_DEL_DEC_STATES ]; +} NSQ_del_decs_struct; + +typedef struct { + opus_int32 Q_Q10[ NEON_MAX_DEL_DEC_STATES ]; + opus_int32 RD_Q10[ NEON_MAX_DEL_DEC_STATES ]; + opus_int32 xq_Q14[ NEON_MAX_DEL_DEC_STATES ]; + opus_int32 LF_AR_Q14[ NEON_MAX_DEL_DEC_STATES ]; + opus_int32 Diff_Q14[ NEON_MAX_DEL_DEC_STATES ]; + opus_int32 sLTP_shp_Q14[ NEON_MAX_DEL_DEC_STATES ]; + opus_int32 LPC_exc_Q14[ NEON_MAX_DEL_DEC_STATES ]; +} NSQ_samples_struct; + +static OPUS_INLINE void silk_nsq_del_dec_scale_states_neon( + const silk_encoder_state *psEncC, /* I Encoder State */ + silk_nsq_state *NSQ, /* I/O NSQ state */ + NSQ_del_decs_struct psDelDec[], /* I/O Delayed decision states */ + const opus_int16 x16[], /* I Input */ + opus_int32 x_sc_Q10[], /* O Input scaled with 1/Gain in Q10 */ + const opus_int16 sLTP[], /* I Re-whitened LTP state in Q0 */ + opus_int32 sLTP_Q15[], /* O LTP state matching scaled input */ + opus_int subfr, /* I Subframe number */ + const opus_int LTP_scale_Q14, /* I LTP state scaling */ + const opus_int32 Gains_Q16[ MAX_NB_SUBFR ], /* I */ + const opus_int pitchL[ MAX_NB_SUBFR ], /* I Pitch lag */ + const opus_int signal_type, /* I Signal type */ + const opus_int decisionDelay /* I Decision delay */ +); + +/******************************************/ +/* Noise shape quantizer for one subframe */ +/******************************************/ +static OPUS_INLINE void silk_noise_shape_quantizer_del_dec_neon( + silk_nsq_state *NSQ, /* I/O NSQ state */ + NSQ_del_decs_struct psDelDec[], /* I/O Delayed decision states */ + opus_int signalType, /* I Signal type */ + const opus_int32 x_Q10[], /* I */ + opus_int8 pulses[], /* O */ + opus_int16 xq[], /* O */ + opus_int32 sLTP_Q15[], /* I/O LTP filter state */ + opus_int32 delayedGain_Q10[], /* I/O Gain delay buffer */ + const opus_int16 a_Q12[], /* I Short term prediction coefs */ + const opus_int16 b_Q14[], /* I Long term prediction coefs */ + const opus_int16 AR_shp_Q13[], /* I Noise shaping coefs */ + opus_int lag, /* I Pitch lag */ + opus_int32 HarmShapeFIRPacked_Q14, /* I */ + opus_int Tilt_Q14, /* I Spectral tilt */ + opus_int32 LF_shp_Q14, /* I */ + opus_int32 Gain_Q16, /* I */ + opus_int Lambda_Q10, /* I */ + opus_int offset_Q10, /* I */ + opus_int length, /* I Input length */ + opus_int subfr, /* I Subframe number */ + opus_int shapingLPCOrder, /* I Shaping LPC filter order */ + opus_int predictLPCOrder, /* I Prediction filter order */ + opus_int warping_Q16, /* I */ + opus_int nStatesDelayedDecision, /* I Number of states in decision tree */ + opus_int *smpl_buf_idx, /* I/O Index to newest samples in buffers */ + opus_int decisionDelay /* I */ +); + +static OPUS_INLINE void copy_winner_state_kernel( + const NSQ_del_decs_struct *psDelDec, + const opus_int offset, + const opus_int last_smple_idx, + const opus_int Winner_ind, + const int32x2_t gain_lo_s32x2, + const int32x2_t gain_hi_s32x2, + const int32x4_t shift_s32x4, + int32x4_t t0_s32x4, + int32x4_t t1_s32x4, + opus_int8 *const pulses, + opus_int16 *pxq, + silk_nsq_state *NSQ +) +{ + int16x8_t t_s16x8; + int32x4_t o0_s32x4, o1_s32x4; + + t0_s32x4 = vld1q_lane_s32( &psDelDec->Q_Q10[ last_smple_idx - 0 ][ Winner_ind ], t0_s32x4, 0 ); + t0_s32x4 = vld1q_lane_s32( &psDelDec->Q_Q10[ last_smple_idx - 1 ][ Winner_ind ], t0_s32x4, 1 ); + t0_s32x4 = vld1q_lane_s32( &psDelDec->Q_Q10[ last_smple_idx - 2 ][ Winner_ind ], t0_s32x4, 2 ); + t0_s32x4 = vld1q_lane_s32( &psDelDec->Q_Q10[ last_smple_idx - 3 ][ Winner_ind ], t0_s32x4, 3 ); + t1_s32x4 = vld1q_lane_s32( &psDelDec->Q_Q10[ last_smple_idx - 4 ][ Winner_ind ], t1_s32x4, 0 ); + t1_s32x4 = vld1q_lane_s32( &psDelDec->Q_Q10[ last_smple_idx - 5 ][ Winner_ind ], t1_s32x4, 1 ); + t1_s32x4 = vld1q_lane_s32( &psDelDec->Q_Q10[ last_smple_idx - 6 ][ Winner_ind ], t1_s32x4, 2 ); + t1_s32x4 = vld1q_lane_s32( &psDelDec->Q_Q10[ last_smple_idx - 7 ][ Winner_ind ], t1_s32x4, 3 ); + t_s16x8 = vcombine_s16( vrshrn_n_s32( t0_s32x4, 10 ), vrshrn_n_s32( t1_s32x4, 10 ) ); + vst1_s8( &pulses[ offset ], vmovn_s16( t_s16x8 ) ); + + t0_s32x4 = vld1q_lane_s32( &psDelDec->Xq_Q14[ last_smple_idx - 0 ][ Winner_ind ], t0_s32x4, 0 ); + t0_s32x4 = vld1q_lane_s32( &psDelDec->Xq_Q14[ last_smple_idx - 1 ][ Winner_ind ], t0_s32x4, 1 ); + t0_s32x4 = vld1q_lane_s32( &psDelDec->Xq_Q14[ last_smple_idx - 2 ][ Winner_ind ], t0_s32x4, 2 ); + t0_s32x4 = vld1q_lane_s32( &psDelDec->Xq_Q14[ last_smple_idx - 3 ][ Winner_ind ], t0_s32x4, 3 ); + t1_s32x4 = vld1q_lane_s32( &psDelDec->Xq_Q14[ last_smple_idx - 4 ][ Winner_ind ], t1_s32x4, 0 ); + t1_s32x4 = vld1q_lane_s32( &psDelDec->Xq_Q14[ last_smple_idx - 5 ][ Winner_ind ], t1_s32x4, 1 ); + t1_s32x4 = vld1q_lane_s32( &psDelDec->Xq_Q14[ last_smple_idx - 6 ][ Winner_ind ], t1_s32x4, 2 ); + t1_s32x4 = vld1q_lane_s32( &psDelDec->Xq_Q14[ last_smple_idx - 7 ][ Winner_ind ], t1_s32x4, 3 ); + o0_s32x4 = vqdmulhq_lane_s32( t0_s32x4, gain_lo_s32x2, 0 ); + o1_s32x4 = vqdmulhq_lane_s32( t1_s32x4, gain_lo_s32x2, 0 ); + o0_s32x4 = vmlaq_lane_s32( o0_s32x4, t0_s32x4, gain_hi_s32x2, 0 ); + o1_s32x4 = vmlaq_lane_s32( o1_s32x4, t1_s32x4, gain_hi_s32x2, 0 ); + o0_s32x4 = vrshlq_s32( o0_s32x4, shift_s32x4 ); + o1_s32x4 = vrshlq_s32( o1_s32x4, shift_s32x4 ); + vst1_s16( &pxq[ offset + 0 ], vqmovn_s32( o0_s32x4 ) ); + vst1_s16( &pxq[ offset + 4 ], vqmovn_s32( o1_s32x4 ) ); + + t0_s32x4 = vld1q_lane_s32( &psDelDec->Shape_Q14[ last_smple_idx - 0 ][ Winner_ind ], t0_s32x4, 0 ); + t0_s32x4 = vld1q_lane_s32( &psDelDec->Shape_Q14[ last_smple_idx - 1 ][ Winner_ind ], t0_s32x4, 1 ); + t0_s32x4 = vld1q_lane_s32( &psDelDec->Shape_Q14[ last_smple_idx - 2 ][ Winner_ind ], t0_s32x4, 2 ); + t0_s32x4 = vld1q_lane_s32( &psDelDec->Shape_Q14[ last_smple_idx - 3 ][ Winner_ind ], t0_s32x4, 3 ); + t1_s32x4 = vld1q_lane_s32( &psDelDec->Shape_Q14[ last_smple_idx - 4 ][ Winner_ind ], t1_s32x4, 0 ); + t1_s32x4 = vld1q_lane_s32( &psDelDec->Shape_Q14[ last_smple_idx - 5 ][ Winner_ind ], t1_s32x4, 1 ); + t1_s32x4 = vld1q_lane_s32( &psDelDec->Shape_Q14[ last_smple_idx - 6 ][ Winner_ind ], t1_s32x4, 2 ); + t1_s32x4 = vld1q_lane_s32( &psDelDec->Shape_Q14[ last_smple_idx - 7 ][ Winner_ind ], t1_s32x4, 3 ); + vst1q_s32( &NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx + offset + 0 ], t0_s32x4 ); + vst1q_s32( &NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx + offset + 4 ], t1_s32x4 ); +} + +static OPUS_INLINE void copy_winner_state( + const NSQ_del_decs_struct *psDelDec, + const opus_int decisionDelay, + const opus_int smpl_buf_idx, + const opus_int Winner_ind, + const opus_int32 gain, + const opus_int32 shift, + opus_int8 *const pulses, + opus_int16 *pxq, + silk_nsq_state *NSQ +) +{ + opus_int i, last_smple_idx; + const int32x2_t gain_lo_s32x2 = vdup_n_s32( silk_LSHIFT32( gain & 0x0000FFFF, 15 ) ); + const int32x2_t gain_hi_s32x2 = vdup_n_s32( gain >> 16 ); + const int32x4_t shift_s32x4 = vdupq_n_s32( -shift ); + int32x4_t t0_s32x4, t1_s32x4; + + t0_s32x4 = t1_s32x4 = vdupq_n_s32( 0 ); /* initialization */ + last_smple_idx = smpl_buf_idx + decisionDelay - 1 + DECISION_DELAY; + if( last_smple_idx >= DECISION_DELAY ) last_smple_idx -= DECISION_DELAY; + if( last_smple_idx >= DECISION_DELAY ) last_smple_idx -= DECISION_DELAY; + + for( i = 0; ( i < ( decisionDelay - 7 ) ) && ( last_smple_idx >= 7 ); i += 8, last_smple_idx -= 8 ) { + copy_winner_state_kernel( psDelDec, i - decisionDelay, last_smple_idx, Winner_ind, gain_lo_s32x2, gain_hi_s32x2, shift_s32x4, t0_s32x4, t1_s32x4, pulses, pxq, NSQ ); + } + for( ; ( i < decisionDelay ) && ( last_smple_idx >= 0 ); i++, last_smple_idx-- ) { + pulses[ i - decisionDelay ] = (opus_int8)silk_RSHIFT_ROUND( psDelDec->Q_Q10[ last_smple_idx ][ Winner_ind ], 10 ); + pxq[ i - decisionDelay ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( psDelDec->Xq_Q14[ last_smple_idx ][ Winner_ind ], gain ), shift ) ); + NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx - decisionDelay + i ] = psDelDec->Shape_Q14[ last_smple_idx ][ Winner_ind ]; + } + + last_smple_idx += DECISION_DELAY; + for( ; i < ( decisionDelay - 7 ); i++, last_smple_idx-- ) { + copy_winner_state_kernel( psDelDec, i - decisionDelay, last_smple_idx, Winner_ind, gain_lo_s32x2, gain_hi_s32x2, shift_s32x4, t0_s32x4, t1_s32x4, pulses, pxq, NSQ ); + } + for( ; i < decisionDelay; i++, last_smple_idx-- ) { + pulses[ i - decisionDelay ] = (opus_int8)silk_RSHIFT_ROUND( psDelDec->Q_Q10[ last_smple_idx ][ Winner_ind ], 10 ); + pxq[ i - decisionDelay ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( psDelDec->Xq_Q14[ last_smple_idx ][ Winner_ind ], gain ), shift ) ); + NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx - decisionDelay + i ] = psDelDec->Shape_Q14[ last_smple_idx ][ Winner_ind ]; + } +} + +void silk_NSQ_del_dec_neon( + const silk_encoder_state *psEncC, /* I Encoder State */ + silk_nsq_state *NSQ, /* I/O NSQ state */ + SideInfoIndices *psIndices, /* I/O Quantization Indices */ + const opus_int16 x16[], /* I Input */ + opus_int8 pulses[], /* O Quantized pulse signal */ + const opus_int16 PredCoef_Q12[ 2 * MAX_LPC_ORDER ], /* I Short term prediction coefs */ + const opus_int16 LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ], /* I Long term prediction coefs */ + const opus_int16 AR_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs */ + const opus_int HarmShapeGain_Q14[ MAX_NB_SUBFR ], /* I Long term shaping coefs */ + const opus_int Tilt_Q14[ MAX_NB_SUBFR ], /* I Spectral tilt */ + const opus_int32 LF_shp_Q14[ MAX_NB_SUBFR ], /* I Low frequency shaping coefs */ + const opus_int32 Gains_Q16[ MAX_NB_SUBFR ], /* I Quantization step sizes */ + const opus_int pitchL[ MAX_NB_SUBFR ], /* I Pitch lags */ + const opus_int Lambda_Q10, /* I Rate/distortion tradeoff */ + const opus_int LTP_scale_Q14 /* I LTP state scaling */ +) +{ +#ifdef OPUS_CHECK_ASM + silk_nsq_state NSQ_c; + SideInfoIndices psIndices_c; + opus_int8 pulses_c[ MAX_FRAME_LENGTH ]; + const opus_int8 *const pulses_a = pulses; + + ( void )pulses_a; + silk_memcpy( &NSQ_c, NSQ, sizeof( NSQ_c ) ); + silk_memcpy( &psIndices_c, psIndices, sizeof( psIndices_c ) ); + silk_memcpy( pulses_c, pulses, sizeof( pulses_c ) ); + silk_NSQ_del_dec_c( psEncC, &NSQ_c, &psIndices_c, x16, pulses_c, PredCoef_Q12, LTPCoef_Q14, AR_Q13, HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, + pitchL, Lambda_Q10, LTP_scale_Q14 ); +#endif + + /* The optimization parallelizes the different delay decision states. */ + if(( psEncC->nStatesDelayedDecision > NEON_MAX_DEL_DEC_STATES ) || ( psEncC->nStatesDelayedDecision <= 2 )) { + /* NEON intrinsics optimization now can only parallelize up to 4 delay decision states. */ + /* If there are more states, C function is called, and this optimization must be expanded. */ + /* When the number of delay decision states is less than 3, there are penalties using this */ + /* optimization, and C function is called. */ + /* When the number of delay decision states is 2, it's better to specialize another */ + /* structure NSQ_del_dec2_struct and optimize with shorter NEON registers. (Low priority) */ + silk_NSQ_del_dec_c( psEncC, NSQ, psIndices, x16, pulses, PredCoef_Q12, LTPCoef_Q14, AR_Q13, HarmShapeGain_Q14, + Tilt_Q14, LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, LTP_scale_Q14 ); + } else { + opus_int i, k, lag, start_idx, LSF_interpolation_flag, Winner_ind, subfr; + opus_int smpl_buf_idx, decisionDelay; + const opus_int16 *A_Q12, *B_Q14, *AR_shp_Q13; + opus_int16 *pxq; + VARDECL( opus_int32, sLTP_Q15 ); + VARDECL( opus_int16, sLTP ); + opus_int32 HarmShapeFIRPacked_Q14; + opus_int offset_Q10; + opus_int32 RDmin_Q10, Gain_Q10; + VARDECL( opus_int32, x_sc_Q10 ); + VARDECL( opus_int32, delayedGain_Q10 ); + VARDECL( NSQ_del_decs_struct, psDelDec ); + int32x4_t t_s32x4; + SAVE_STACK; + + /* Set unvoiced lag to the previous one, overwrite later for voiced */ + lag = NSQ->lagPrev; + + silk_assert( NSQ->prev_gain_Q16 != 0 ); + + /* Initialize delayed decision states */ + ALLOC( psDelDec, 1, NSQ_del_decs_struct ); + /* Only RandState and RD_Q10 need to be initialized to 0. */ + silk_memset( psDelDec->RandState, 0, sizeof( psDelDec->RandState ) ); + vst1q_s32( psDelDec->RD_Q10, vdupq_n_s32( 0 ) ); + + for( k = 0; k < psEncC->nStatesDelayedDecision; k++ ) { + psDelDec->SeedInit[ k ] = psDelDec->Seed[ k ] = ( k + psIndices->Seed ) & 3; + } + vst1q_s32( psDelDec->LF_AR_Q14, vld1q_dup_s32( &NSQ->sLF_AR_shp_Q14 ) ); + vst1q_s32( psDelDec->Diff_Q14, vld1q_dup_s32( &NSQ->sDiff_shp_Q14 ) ); + vst1q_s32( psDelDec->Shape_Q14[ 0 ], vld1q_dup_s32( &NSQ->sLTP_shp_Q14[ psEncC->ltp_mem_length - 1 ] ) ); + for( i = 0; i < NSQ_LPC_BUF_LENGTH; i++ ) { + vst1q_s32( psDelDec->sLPC_Q14[ i ], vld1q_dup_s32( &NSQ->sLPC_Q14[ i ] ) ); + } + for( i = 0; i < (opus_int)( sizeof( NSQ->sAR2_Q14 ) / sizeof( NSQ->sAR2_Q14[ 0 ] ) ); i++ ) { + vst1q_s32( psDelDec->sAR2_Q14[ i ], vld1q_dup_s32( &NSQ->sAR2_Q14[ i ] ) ); + } + + offset_Q10 = silk_Quantization_Offsets_Q10[ psIndices->signalType >> 1 ][ psIndices->quantOffsetType ]; + smpl_buf_idx = 0; /* index of oldest samples */ + + decisionDelay = silk_min_int( DECISION_DELAY, psEncC->subfr_length ); + + /* For voiced frames limit the decision delay to lower than the pitch lag */ + if( psIndices->signalType == TYPE_VOICED ) { + opus_int pitch_min = pitchL[ 0 ]; + for( k = 1; k < psEncC->nb_subfr; k++ ) { + pitch_min = silk_min_int( pitch_min, pitchL[ k ] ); + } + decisionDelay = silk_min_int( decisionDelay, pitch_min - LTP_ORDER / 2 - 1 ); + } else { + if( lag > 0 ) { + decisionDelay = silk_min_int( decisionDelay, lag - LTP_ORDER / 2 - 1 ); + } + } + + if( psIndices->NLSFInterpCoef_Q2 == 4 ) { + LSF_interpolation_flag = 0; + } else { + LSF_interpolation_flag = 1; + } + + ALLOC( sLTP_Q15, psEncC->ltp_mem_length + psEncC->frame_length, opus_int32 ); + ALLOC( sLTP, psEncC->ltp_mem_length + psEncC->frame_length, opus_int16 ); + ALLOC( x_sc_Q10, psEncC->subfr_length, opus_int32 ); + ALLOC( delayedGain_Q10, DECISION_DELAY, opus_int32 ); + /* Set up pointers to start of sub frame */ + pxq = &NSQ->xq[ psEncC->ltp_mem_length ]; + NSQ->sLTP_shp_buf_idx = psEncC->ltp_mem_length; + NSQ->sLTP_buf_idx = psEncC->ltp_mem_length; + subfr = 0; + for( k = 0; k < psEncC->nb_subfr; k++ ) { + A_Q12 = &PredCoef_Q12[ ( ( k >> 1 ) | ( 1 - LSF_interpolation_flag ) ) * MAX_LPC_ORDER ]; + B_Q14 = <PCoef_Q14[ k * LTP_ORDER ]; + AR_shp_Q13 = &AR_Q13[ k * MAX_SHAPE_LPC_ORDER ]; + + /* Noise shape parameters */ + silk_assert( HarmShapeGain_Q14[ k ] >= 0 ); + HarmShapeFIRPacked_Q14 = silk_RSHIFT( HarmShapeGain_Q14[ k ], 2 ); + HarmShapeFIRPacked_Q14 |= silk_LSHIFT( (opus_int32)silk_RSHIFT( HarmShapeGain_Q14[ k ], 1 ), 16 ); + + NSQ->rewhite_flag = 0; + if( psIndices->signalType == TYPE_VOICED ) { + /* Voiced */ + lag = pitchL[ k ]; + + /* Re-whitening */ + if( ( k & ( 3 - silk_LSHIFT( LSF_interpolation_flag, 1 ) ) ) == 0 ) { + if( k == 2 ) { + /* RESET DELAYED DECISIONS */ + /* Find winner */ + int32x4_t RD_Q10_s32x4; + RDmin_Q10 = psDelDec->RD_Q10[ 0 ]; + Winner_ind = 0; + for( i = 1; i < psEncC->nStatesDelayedDecision; i++ ) { + if( psDelDec->RD_Q10[ i ] < RDmin_Q10 ) { + RDmin_Q10 = psDelDec->RD_Q10[ i ]; + Winner_ind = i; + } + } + psDelDec->RD_Q10[ Winner_ind ] -= ( silk_int32_MAX >> 4 ); + RD_Q10_s32x4 = vld1q_s32( psDelDec->RD_Q10 ); + RD_Q10_s32x4 = vaddq_s32( RD_Q10_s32x4, vdupq_n_s32( silk_int32_MAX >> 4 ) ); + vst1q_s32( psDelDec->RD_Q10, RD_Q10_s32x4 ); + + /* Copy final part of signals from winner state to output and long-term filter states */ + copy_winner_state( psDelDec, decisionDelay, smpl_buf_idx, Winner_ind, Gains_Q16[ 1 ], 14, pulses, pxq, NSQ ); + + subfr = 0; + } + + /* Rewhiten with new A coefs */ + start_idx = psEncC->ltp_mem_length - lag - psEncC->predictLPCOrder - LTP_ORDER / 2; + silk_assert( start_idx > 0 ); + + silk_LPC_analysis_filter( &sLTP[ start_idx ], &NSQ->xq[ start_idx + k * psEncC->subfr_length ], + A_Q12, psEncC->ltp_mem_length - start_idx, psEncC->predictLPCOrder, psEncC->arch ); + + NSQ->sLTP_buf_idx = psEncC->ltp_mem_length; + NSQ->rewhite_flag = 1; + } + } + + silk_nsq_del_dec_scale_states_neon( psEncC, NSQ, psDelDec, x16, x_sc_Q10, sLTP, sLTP_Q15, k, + LTP_scale_Q14, Gains_Q16, pitchL, psIndices->signalType, decisionDelay ); + + silk_noise_shape_quantizer_del_dec_neon( NSQ, psDelDec, psIndices->signalType, x_sc_Q10, pulses, pxq, sLTP_Q15, + delayedGain_Q10, A_Q12, B_Q14, AR_shp_Q13, lag, HarmShapeFIRPacked_Q14, Tilt_Q14[ k ], LF_shp_Q14[ k ], + Gains_Q16[ k ], Lambda_Q10, offset_Q10, psEncC->subfr_length, subfr++, psEncC->shapingLPCOrder, + psEncC->predictLPCOrder, psEncC->warping_Q16, psEncC->nStatesDelayedDecision, &smpl_buf_idx, decisionDelay ); + + x16 += psEncC->subfr_length; + pulses += psEncC->subfr_length; + pxq += psEncC->subfr_length; + } + + /* Find winner */ + RDmin_Q10 = psDelDec->RD_Q10[ 0 ]; + Winner_ind = 0; + for( k = 1; k < psEncC->nStatesDelayedDecision; k++ ) { + if( psDelDec->RD_Q10[ k ] < RDmin_Q10 ) { + RDmin_Q10 = psDelDec->RD_Q10[ k ]; + Winner_ind = k; + } + } + + /* Copy final part of signals from winner state to output and long-term filter states */ + psIndices->Seed = psDelDec->SeedInit[ Winner_ind ]; + Gain_Q10 = silk_RSHIFT32( Gains_Q16[ psEncC->nb_subfr - 1 ], 6 ); + copy_winner_state( psDelDec, decisionDelay, smpl_buf_idx, Winner_ind, Gain_Q10, 8, pulses, pxq, NSQ ); + + t_s32x4 = vdupq_n_s32( 0 ); /* initialization */ + for( i = 0; i < ( NSQ_LPC_BUF_LENGTH - 3 ); i += 4 ) { + t_s32x4 = vld1q_lane_s32( &psDelDec->sLPC_Q14[ i + 0 ][ Winner_ind ], t_s32x4, 0 ); + t_s32x4 = vld1q_lane_s32( &psDelDec->sLPC_Q14[ i + 1 ][ Winner_ind ], t_s32x4, 1 ); + t_s32x4 = vld1q_lane_s32( &psDelDec->sLPC_Q14[ i + 2 ][ Winner_ind ], t_s32x4, 2 ); + t_s32x4 = vld1q_lane_s32( &psDelDec->sLPC_Q14[ i + 3 ][ Winner_ind ], t_s32x4, 3 ); + vst1q_s32( &NSQ->sLPC_Q14[ i ], t_s32x4 ); + } + + for( ; i < NSQ_LPC_BUF_LENGTH; i++ ) { + NSQ->sLPC_Q14[ i ] = psDelDec->sLPC_Q14[ i ][ Winner_ind ]; + } + + for( i = 0; i < (opus_int)( sizeof( NSQ->sAR2_Q14 ) / sizeof( NSQ->sAR2_Q14[ 0 ] ) - 3 ); i += 4 ) { + t_s32x4 = vld1q_lane_s32( &psDelDec->sAR2_Q14[ i + 0 ][ Winner_ind ], t_s32x4, 0 ); + t_s32x4 = vld1q_lane_s32( &psDelDec->sAR2_Q14[ i + 1 ][ Winner_ind ], t_s32x4, 1 ); + t_s32x4 = vld1q_lane_s32( &psDelDec->sAR2_Q14[ i + 2 ][ Winner_ind ], t_s32x4, 2 ); + t_s32x4 = vld1q_lane_s32( &psDelDec->sAR2_Q14[ i + 3 ][ Winner_ind ], t_s32x4, 3 ); + vst1q_s32( &NSQ->sAR2_Q14[ i ], t_s32x4 ); + } + + for( ; i < (opus_int)( sizeof( NSQ->sAR2_Q14 ) / sizeof( NSQ->sAR2_Q14[ 0 ] ) ); i++ ) { + NSQ->sAR2_Q14[ i ] = psDelDec->sAR2_Q14[ i ][ Winner_ind ]; + } + + /* Update states */ + NSQ->sLF_AR_shp_Q14 = psDelDec->LF_AR_Q14[ Winner_ind ]; + NSQ->sDiff_shp_Q14 = psDelDec->Diff_Q14[ Winner_ind ]; + NSQ->lagPrev = pitchL[ psEncC->nb_subfr - 1 ]; + + /* Save quantized speech signal */ + silk_memmove( NSQ->xq, &NSQ->xq[ psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int16 ) ); + silk_memmove( NSQ->sLTP_shp_Q14, &NSQ->sLTP_shp_Q14[ psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int32 ) ); + RESTORE_STACK; + } + +#ifdef OPUS_CHECK_ASM + silk_assert( !memcmp( &NSQ_c, NSQ, sizeof( NSQ_c ) ) ); + silk_assert( !memcmp( &psIndices_c, psIndices, sizeof( psIndices_c ) ) ); + silk_assert( !memcmp( pulses_c, pulses_a, sizeof( pulses_c ) ) ); +#endif +} + +/******************************************/ +/* Noise shape quantizer for one subframe */ +/******************************************/ +/* Note: Function silk_short_prediction_create_arch_coef_neon() defined in NSQ_neon.h is actually a hacking C function. */ +/* Therefore here we append "_local" to the NEON function name to avoid confusion. */ +static OPUS_INLINE void silk_short_prediction_create_arch_coef_neon_local(opus_int32 *out, const opus_int16 *in, opus_int order) +{ + int16x8_t t_s16x8; + int32x4_t t0_s32x4, t1_s32x4, t2_s32x4, t3_s32x4; + silk_assert( order == 10 || order == 16 ); + + t_s16x8 = vld1q_s16( in + 0 ); /* 7 6 5 4 3 2 1 0 */ + t_s16x8 = vrev64q_s16( t_s16x8 ); /* 4 5 6 7 0 1 2 3 */ + t2_s32x4 = vshll_n_s16( vget_high_s16( t_s16x8 ), 15 ); /* 4 5 6 7 */ + t3_s32x4 = vshll_n_s16( vget_low_s16( t_s16x8 ), 15 ); /* 0 1 2 3 */ + + if( order == 16 ) { + t_s16x8 = vld1q_s16( in + 8 ); /* F E D C B A 9 8 */ + t_s16x8 = vrev64q_s16( t_s16x8 ); /* C D E F 8 9 A B */ + t0_s32x4 = vshll_n_s16( vget_high_s16( t_s16x8 ), 15 ); /* C D E F */ + t1_s32x4 = vshll_n_s16( vget_low_s16( t_s16x8 ), 15 ); /* 8 9 A B */ + } else { + int16x4_t t_s16x4; + + t0_s32x4 = vdupq_n_s32( 0 ); /* zero zero zero zero */ + t_s16x4 = vld1_s16( in + 6 ); /* 9 8 7 6 */ + t_s16x4 = vrev64_s16( t_s16x4 ); /* 6 7 8 9 */ + t1_s32x4 = vshll_n_s16( t_s16x4, 15 ); + t1_s32x4 = vcombine_s32( vget_low_s32(t0_s32x4), vget_low_s32( t1_s32x4 ) ); /* 8 9 zero zero */ + } + vst1q_s32( out + 0, t0_s32x4 ); + vst1q_s32( out + 4, t1_s32x4 ); + vst1q_s32( out + 8, t2_s32x4 ); + vst1q_s32( out + 12, t3_s32x4 ); +} + +static OPUS_INLINE int32x4_t silk_SMLAWB_lane0_neon( + const int32x4_t out_s32x4, + const int32x4_t in_s32x4, + const int32x2_t coef_s32x2 +) +{ + return vaddq_s32( out_s32x4, vqdmulhq_lane_s32( in_s32x4, coef_s32x2, 0 ) ); +} + +static OPUS_INLINE int32x4_t silk_SMLAWB_lane1_neon( + const int32x4_t out_s32x4, + const int32x4_t in_s32x4, + const int32x2_t coef_s32x2 +) +{ + return vaddq_s32( out_s32x4, vqdmulhq_lane_s32( in_s32x4, coef_s32x2, 1 ) ); +} + +/* Note: This function has different return value than silk_noise_shape_quantizer_short_prediction_neon(). */ +/* Therefore here we append "_local" to the function name to avoid confusion. */ +static OPUS_INLINE int32x4_t silk_noise_shape_quantizer_short_prediction_neon_local(const opus_int32 *buf32, const opus_int32 *a_Q12_arch, opus_int order) +{ + const int32x4_t a_Q12_arch0_s32x4 = vld1q_s32( a_Q12_arch + 0 ); + const int32x4_t a_Q12_arch1_s32x4 = vld1q_s32( a_Q12_arch + 4 ); + const int32x4_t a_Q12_arch2_s32x4 = vld1q_s32( a_Q12_arch + 8 ); + const int32x4_t a_Q12_arch3_s32x4 = vld1q_s32( a_Q12_arch + 12 ); + int32x4_t LPC_pred_Q14_s32x4; + + silk_assert( order == 10 || order == 16 ); + /* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */ + LPC_pred_Q14_s32x4 = vdupq_n_s32( silk_RSHIFT( order, 1 ) ); + LPC_pred_Q14_s32x4 = silk_SMLAWB_lane0_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 + 0 * NEON_MAX_DEL_DEC_STATES ), vget_low_s32( a_Q12_arch0_s32x4 ) ); + LPC_pred_Q14_s32x4 = silk_SMLAWB_lane1_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 + 1 * NEON_MAX_DEL_DEC_STATES ), vget_low_s32( a_Q12_arch0_s32x4 ) ); + LPC_pred_Q14_s32x4 = silk_SMLAWB_lane0_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 + 2 * NEON_MAX_DEL_DEC_STATES ), vget_high_s32( a_Q12_arch0_s32x4 ) ); + LPC_pred_Q14_s32x4 = silk_SMLAWB_lane1_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 + 3 * NEON_MAX_DEL_DEC_STATES ), vget_high_s32( a_Q12_arch0_s32x4 ) ); + LPC_pred_Q14_s32x4 = silk_SMLAWB_lane0_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 + 4 * NEON_MAX_DEL_DEC_STATES ), vget_low_s32( a_Q12_arch1_s32x4 ) ); + LPC_pred_Q14_s32x4 = silk_SMLAWB_lane1_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 + 5 * NEON_MAX_DEL_DEC_STATES ), vget_low_s32( a_Q12_arch1_s32x4 ) ); + LPC_pred_Q14_s32x4 = silk_SMLAWB_lane0_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 + 6 * NEON_MAX_DEL_DEC_STATES ), vget_high_s32( a_Q12_arch1_s32x4 ) ); + LPC_pred_Q14_s32x4 = silk_SMLAWB_lane1_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 + 7 * NEON_MAX_DEL_DEC_STATES ), vget_high_s32( a_Q12_arch1_s32x4 ) ); + LPC_pred_Q14_s32x4 = silk_SMLAWB_lane0_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 + 8 * NEON_MAX_DEL_DEC_STATES ), vget_low_s32( a_Q12_arch2_s32x4 ) ); + LPC_pred_Q14_s32x4 = silk_SMLAWB_lane1_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 + 9 * NEON_MAX_DEL_DEC_STATES ), vget_low_s32( a_Q12_arch2_s32x4 ) ); + LPC_pred_Q14_s32x4 = silk_SMLAWB_lane0_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 + 10 * NEON_MAX_DEL_DEC_STATES ), vget_high_s32( a_Q12_arch2_s32x4 ) ); + LPC_pred_Q14_s32x4 = silk_SMLAWB_lane1_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 + 11 * NEON_MAX_DEL_DEC_STATES ), vget_high_s32( a_Q12_arch2_s32x4 ) ); + LPC_pred_Q14_s32x4 = silk_SMLAWB_lane0_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 + 12 * NEON_MAX_DEL_DEC_STATES ), vget_low_s32( a_Q12_arch3_s32x4 ) ); + LPC_pred_Q14_s32x4 = silk_SMLAWB_lane1_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 + 13 * NEON_MAX_DEL_DEC_STATES ), vget_low_s32( a_Q12_arch3_s32x4 ) ); + LPC_pred_Q14_s32x4 = silk_SMLAWB_lane0_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 + 14 * NEON_MAX_DEL_DEC_STATES ), vget_high_s32( a_Q12_arch3_s32x4 ) ); + LPC_pred_Q14_s32x4 = silk_SMLAWB_lane1_neon( LPC_pred_Q14_s32x4, vld1q_s32( buf32 + 15 * NEON_MAX_DEL_DEC_STATES ), vget_high_s32( a_Q12_arch3_s32x4 ) ); + + return LPC_pred_Q14_s32x4; +} + +static OPUS_INLINE void silk_noise_shape_quantizer_del_dec_neon( + silk_nsq_state *NSQ, /* I/O NSQ state */ + NSQ_del_decs_struct psDelDec[], /* I/O Delayed decision states */ + opus_int signalType, /* I Signal type */ + const opus_int32 x_Q10[], /* I */ + opus_int8 pulses[], /* O */ + opus_int16 xq[], /* O */ + opus_int32 sLTP_Q15[], /* I/O LTP filter state */ + opus_int32 delayedGain_Q10[], /* I/O Gain delay buffer */ + const opus_int16 a_Q12[], /* I Short term prediction coefs */ + const opus_int16 b_Q14[], /* I Long term prediction coefs */ + const opus_int16 AR_shp_Q13[], /* I Noise shaping coefs */ + opus_int lag, /* I Pitch lag */ + opus_int32 HarmShapeFIRPacked_Q14, /* I */ + opus_int Tilt_Q14, /* I Spectral tilt */ + opus_int32 LF_shp_Q14, /* I */ + opus_int32 Gain_Q16, /* I */ + opus_int Lambda_Q10, /* I */ + opus_int offset_Q10, /* I */ + opus_int length, /* I Input length */ + opus_int subfr, /* I Subframe number */ + opus_int shapingLPCOrder, /* I Shaping LPC filter order */ + opus_int predictLPCOrder, /* I Prediction filter order */ + opus_int warping_Q16, /* I */ + opus_int nStatesDelayedDecision, /* I Number of states in decision tree */ + opus_int *smpl_buf_idx, /* I/O Index to newest samples in buffers */ + opus_int decisionDelay /* I */ +) +{ + opus_int i, j, k, Winner_ind, RDmin_ind, RDmax_ind, last_smple_idx; + opus_int32 Winner_rand_state; + opus_int32 LTP_pred_Q14, n_LTP_Q14; + opus_int32 RDmin_Q10, RDmax_Q10; + opus_int32 Gain_Q10; + opus_int32 *pred_lag_ptr, *shp_lag_ptr; + opus_int32 a_Q12_arch[MAX_LPC_ORDER]; + const int32x2_t warping_Q16_s32x2 = vdup_n_s32( silk_LSHIFT32( warping_Q16, 16 ) >> 1 ); + const opus_int32 LF_shp_Q29 = silk_LSHIFT32( LF_shp_Q14, 16 ) >> 1; + opus_int32 AR_shp_Q28[ MAX_SHAPE_LPC_ORDER ]; + const uint32x4_t rand_multiplier_u32x4 = vdupq_n_u32( RAND_MULTIPLIER ); + const uint32x4_t rand_increment_u32x4 = vdupq_n_u32( RAND_INCREMENT ); + + VARDECL( NSQ_samples_struct, psSampleState ); + SAVE_STACK; + + silk_assert( nStatesDelayedDecision > 0 ); + silk_assert( ( shapingLPCOrder & 1 ) == 0 ); /* check that order is even */ + ALLOC( psSampleState, 2, NSQ_samples_struct ); + + shp_lag_ptr = &NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx - lag + HARM_SHAPE_FIR_TAPS / 2 ]; + pred_lag_ptr = &sLTP_Q15[ NSQ->sLTP_buf_idx - lag + LTP_ORDER / 2 ]; + Gain_Q10 = silk_RSHIFT( Gain_Q16, 6 ); + + for( i = 0; i < ( MAX_SHAPE_LPC_ORDER - 7 ); i += 8 ) { + const int16x8_t t_s16x8 = vld1q_s16( AR_shp_Q13 + i ); + vst1q_s32( AR_shp_Q28 + i + 0, vshll_n_s16( vget_low_s16( t_s16x8 ), 15 ) ); + vst1q_s32( AR_shp_Q28 + i + 4, vshll_n_s16( vget_high_s16( t_s16x8 ), 15 ) ); + } + + for( ; i < MAX_SHAPE_LPC_ORDER; i++ ) { + AR_shp_Q28[i] = silk_LSHIFT32( AR_shp_Q13[i], 15 ); + } + + silk_short_prediction_create_arch_coef_neon_local( a_Q12_arch, a_Q12, predictLPCOrder ); + + for( i = 0; i < length; i++ ) { + int32x4_t Seed_s32x4, LPC_pred_Q14_s32x4; + int32x4_t sign_s32x4, tmp1_s32x4, tmp2_s32x4; + int32x4_t n_AR_Q14_s32x4, n_LF_Q14_s32x4; + int32x2_t AR_shp_Q28_s32x2; + int16x4_t r_Q10_s16x4, rr_Q10_s16x4; + + /* Perform common calculations used in all states */ + + /* Long-term prediction */ + if( signalType == TYPE_VOICED ) { + /* Unrolled loop */ + /* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */ + LTP_pred_Q14 = 2; + LTP_pred_Q14 = silk_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ 0 ], b_Q14[ 0 ] ); + LTP_pred_Q14 = silk_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ -1 ], b_Q14[ 1 ] ); + LTP_pred_Q14 = silk_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ -2 ], b_Q14[ 2 ] ); + LTP_pred_Q14 = silk_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ -3 ], b_Q14[ 3 ] ); + LTP_pred_Q14 = silk_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ -4 ], b_Q14[ 4 ] ); + LTP_pred_Q14 = silk_LSHIFT( LTP_pred_Q14, 1 ); /* Q13 -> Q14 */ + pred_lag_ptr++; + } else { + LTP_pred_Q14 = 0; + } + + /* Long-term shaping */ + if( lag > 0 ) { + /* Symmetric, packed FIR coefficients */ + n_LTP_Q14 = silk_SMULWB( silk_ADD32( shp_lag_ptr[ 0 ], shp_lag_ptr[ -2 ] ), HarmShapeFIRPacked_Q14 ); + n_LTP_Q14 = silk_SMLAWT( n_LTP_Q14, shp_lag_ptr[ -1 ], HarmShapeFIRPacked_Q14 ); + n_LTP_Q14 = silk_SUB_LSHIFT32( LTP_pred_Q14, n_LTP_Q14, 2 ); /* Q12 -> Q14 */ + shp_lag_ptr++; + } else { + n_LTP_Q14 = 0; + } + + /* Generate dither */ + Seed_s32x4 = vld1q_s32( psDelDec->Seed ); + Seed_s32x4 = vreinterpretq_s32_u32( vmlaq_u32( rand_increment_u32x4, vreinterpretq_u32_s32( Seed_s32x4 ), rand_multiplier_u32x4 ) ); + vst1q_s32( psDelDec->Seed, Seed_s32x4 ); + + /* Short-term prediction */ + LPC_pred_Q14_s32x4 = silk_noise_shape_quantizer_short_prediction_neon_local(psDelDec->sLPC_Q14[ NSQ_LPC_BUF_LENGTH - 16 + i ], a_Q12_arch, predictLPCOrder); + LPC_pred_Q14_s32x4 = vshlq_n_s32( LPC_pred_Q14_s32x4, 4 ); /* Q10 -> Q14 */ + + /* Noise shape feedback */ + /* Output of lowpass section */ + tmp2_s32x4 = silk_SMLAWB_lane0_neon( vld1q_s32( psDelDec->Diff_Q14 ), vld1q_s32( psDelDec->sAR2_Q14[ 0 ] ), warping_Q16_s32x2 ); + /* Output of allpass section */ + tmp1_s32x4 = vsubq_s32( vld1q_s32( psDelDec->sAR2_Q14[ 1 ] ), tmp2_s32x4 ); + tmp1_s32x4 = silk_SMLAWB_lane0_neon( vld1q_s32( psDelDec->sAR2_Q14[ 0 ] ), tmp1_s32x4, warping_Q16_s32x2 ); + vst1q_s32( psDelDec->sAR2_Q14[ 0 ], tmp2_s32x4 ); + AR_shp_Q28_s32x2 = vld1_s32( AR_shp_Q28 ); + n_AR_Q14_s32x4 = vaddq_s32( vdupq_n_s32( silk_RSHIFT( shapingLPCOrder, 1 ) ), vqdmulhq_lane_s32( tmp2_s32x4, AR_shp_Q28_s32x2, 0 ) ); + + /* Loop over allpass sections */ + for( j = 2; j < shapingLPCOrder; j += 2 ) { + /* Output of allpass section */ + tmp2_s32x4 = vsubq_s32( vld1q_s32( psDelDec->sAR2_Q14[ j + 0 ] ), tmp1_s32x4 ); + tmp2_s32x4 = silk_SMLAWB_lane0_neon( vld1q_s32( psDelDec->sAR2_Q14[ j - 1 ] ), tmp2_s32x4, warping_Q16_s32x2 ); + vst1q_s32( psDelDec->sAR2_Q14[ j - 1 ], tmp1_s32x4 ); + n_AR_Q14_s32x4 = vaddq_s32( n_AR_Q14_s32x4, vqdmulhq_lane_s32( tmp1_s32x4, AR_shp_Q28_s32x2, 1 ) ); + /* Output of allpass section */ + tmp1_s32x4 = vsubq_s32( vld1q_s32( psDelDec->sAR2_Q14[ j + 1 ] ), tmp2_s32x4 ); + tmp1_s32x4 = silk_SMLAWB_lane0_neon( vld1q_s32( psDelDec->sAR2_Q14[ j + 0 ] ), tmp1_s32x4, warping_Q16_s32x2 ); + vst1q_s32( psDelDec->sAR2_Q14[ j + 0 ], tmp2_s32x4 ); + AR_shp_Q28_s32x2 = vld1_s32( &AR_shp_Q28[ j ] ); + n_AR_Q14_s32x4 = vaddq_s32( n_AR_Q14_s32x4, vqdmulhq_lane_s32( tmp2_s32x4, AR_shp_Q28_s32x2, 0 ) ); + } + vst1q_s32( psDelDec->sAR2_Q14[ shapingLPCOrder - 1 ], tmp1_s32x4 ); + n_AR_Q14_s32x4 = vaddq_s32( n_AR_Q14_s32x4, vqdmulhq_lane_s32( tmp1_s32x4, AR_shp_Q28_s32x2, 1 ) ); + n_AR_Q14_s32x4 = vshlq_n_s32( n_AR_Q14_s32x4, 1 ); /* Q11 -> Q12 */ + n_AR_Q14_s32x4 = vaddq_s32( n_AR_Q14_s32x4, vqdmulhq_n_s32( vld1q_s32( psDelDec->LF_AR_Q14 ), silk_LSHIFT32( Tilt_Q14, 16 ) >> 1 ) ); /* Q12 */ + n_AR_Q14_s32x4 = vshlq_n_s32( n_AR_Q14_s32x4, 2 ); /* Q12 -> Q14 */ + n_LF_Q14_s32x4 = vqdmulhq_n_s32( vld1q_s32( psDelDec->Shape_Q14[ *smpl_buf_idx ] ), LF_shp_Q29 ); /* Q12 */ + n_LF_Q14_s32x4 = vaddq_s32( n_LF_Q14_s32x4, vqdmulhq_n_s32( vld1q_s32( psDelDec->LF_AR_Q14 ), silk_LSHIFT32( LF_shp_Q14 >> 16 , 15 ) ) ); /* Q12 */ + n_LF_Q14_s32x4 = vshlq_n_s32( n_LF_Q14_s32x4, 2 ); /* Q12 -> Q14 */ + + /* Input minus prediction plus noise feedback */ + /* r = x[ i ] - LTP_pred - LPC_pred + n_AR + n_Tilt + n_LF + n_LTP */ + tmp1_s32x4 = vaddq_s32( n_AR_Q14_s32x4, n_LF_Q14_s32x4 ); /* Q14 */ + tmp2_s32x4 = vaddq_s32( vdupq_n_s32( n_LTP_Q14 ), LPC_pred_Q14_s32x4 ); /* Q13 */ + tmp1_s32x4 = vsubq_s32( tmp2_s32x4, tmp1_s32x4 ); /* Q13 */ + tmp1_s32x4 = vrshrq_n_s32( tmp1_s32x4, 4 ); /* Q10 */ + tmp1_s32x4 = vsubq_s32( vdupq_n_s32( x_Q10[ i ] ), tmp1_s32x4 ); /* residual error Q10 */ + + /* Flip sign depending on dither */ + sign_s32x4 = vreinterpretq_s32_u32( vcltq_s32( Seed_s32x4, vdupq_n_s32( 0 ) ) ); + tmp1_s32x4 = veorq_s32( tmp1_s32x4, sign_s32x4 ); + tmp1_s32x4 = vsubq_s32( tmp1_s32x4, sign_s32x4 ); + tmp1_s32x4 = vmaxq_s32( tmp1_s32x4, vdupq_n_s32( -( 31 << 10 ) ) ); + tmp1_s32x4 = vminq_s32( tmp1_s32x4, vdupq_n_s32( 30 << 10 ) ); + r_Q10_s16x4 = vmovn_s32( tmp1_s32x4 ); + + /* Find two quantization level candidates and measure their rate-distortion */ + { + int16x4_t q1_Q10_s16x4 = vsub_s16( r_Q10_s16x4, vdup_n_s16( offset_Q10 ) ); + int16x4_t q1_Q0_s16x4 = vshr_n_s16( q1_Q10_s16x4, 10 ); + int16x4_t q2_Q10_s16x4; + int32x4_t rd1_Q10_s32x4, rd2_Q10_s32x4; + uint32x4_t t_u32x4; + + if( Lambda_Q10 > 2048 ) { + /* For aggressive RDO, the bias becomes more than one pulse. */ + const int rdo_offset = Lambda_Q10/2 - 512; + const uint16x4_t greaterThanRdo = vcgt_s16( q1_Q10_s16x4, vdup_n_s16( rdo_offset ) ); + const uint16x4_t lessThanMinusRdo = vclt_s16( q1_Q10_s16x4, vdup_n_s16( -rdo_offset ) ); + /* If Lambda_Q10 > 32767, then q1_Q0, q1_Q10 and q2_Q10 must change to 32-bit. */ + silk_assert( Lambda_Q10 <= 32767 ); + + q1_Q0_s16x4 = vreinterpret_s16_u16( vclt_s16( q1_Q10_s16x4, vdup_n_s16( 0 ) ) ); + q1_Q0_s16x4 = vbsl_s16( greaterThanRdo, vsub_s16( q1_Q10_s16x4, vdup_n_s16( rdo_offset ) ), q1_Q0_s16x4 ); + q1_Q0_s16x4 = vbsl_s16( lessThanMinusRdo, vadd_s16( q1_Q10_s16x4, vdup_n_s16( rdo_offset ) ), q1_Q0_s16x4 ); + q1_Q0_s16x4 = vshr_n_s16( q1_Q0_s16x4, 10 ); + } + { + const uint16x4_t equal0_u16x4 = vceq_s16( q1_Q0_s16x4, vdup_n_s16( 0 ) ); + const uint16x4_t equalMinus1_u16x4 = vceq_s16( q1_Q0_s16x4, vdup_n_s16( -1 ) ); + const uint16x4_t lessThanMinus1_u16x4 = vclt_s16( q1_Q0_s16x4, vdup_n_s16( -1 ) ); + int16x4_t tmp1_s16x4, tmp2_s16x4; + + q1_Q10_s16x4 = vshl_n_s16( q1_Q0_s16x4, 10 ); + tmp1_s16x4 = vadd_s16( q1_Q10_s16x4, vdup_n_s16( offset_Q10 - QUANT_LEVEL_ADJUST_Q10 ) ); + q1_Q10_s16x4 = vadd_s16( q1_Q10_s16x4, vdup_n_s16( offset_Q10 + QUANT_LEVEL_ADJUST_Q10 ) ); + q1_Q10_s16x4 = vbsl_s16( lessThanMinus1_u16x4, q1_Q10_s16x4, tmp1_s16x4 ); + q1_Q10_s16x4 = vbsl_s16( equal0_u16x4, vdup_n_s16( offset_Q10 ), q1_Q10_s16x4 ); + q1_Q10_s16x4 = vbsl_s16( equalMinus1_u16x4, vdup_n_s16( offset_Q10 - ( 1024 - QUANT_LEVEL_ADJUST_Q10 ) ), q1_Q10_s16x4 ); + q2_Q10_s16x4 = vadd_s16( q1_Q10_s16x4, vdup_n_s16( 1024 ) ); + q2_Q10_s16x4 = vbsl_s16( equal0_u16x4, vdup_n_s16( offset_Q10 + 1024 - QUANT_LEVEL_ADJUST_Q10 ), q2_Q10_s16x4 ); + q2_Q10_s16x4 = vbsl_s16( equalMinus1_u16x4, vdup_n_s16( offset_Q10 ), q2_Q10_s16x4 ); + tmp1_s16x4 = q1_Q10_s16x4; + tmp2_s16x4 = q2_Q10_s16x4; + tmp1_s16x4 = vbsl_s16( vorr_u16( equalMinus1_u16x4, lessThanMinus1_u16x4 ), vneg_s16( tmp1_s16x4 ), tmp1_s16x4 ); + tmp2_s16x4 = vbsl_s16( lessThanMinus1_u16x4, vneg_s16( tmp2_s16x4 ), tmp2_s16x4 ); + rd1_Q10_s32x4 = vmull_s16( tmp1_s16x4, vdup_n_s16( Lambda_Q10 ) ); + rd2_Q10_s32x4 = vmull_s16( tmp2_s16x4, vdup_n_s16( Lambda_Q10 ) ); + } + + rr_Q10_s16x4 = vsub_s16( r_Q10_s16x4, q1_Q10_s16x4 ); + rd1_Q10_s32x4 = vmlal_s16( rd1_Q10_s32x4, rr_Q10_s16x4, rr_Q10_s16x4 ); + rd1_Q10_s32x4 = vshrq_n_s32( rd1_Q10_s32x4, 10 ); + + rr_Q10_s16x4 = vsub_s16( r_Q10_s16x4, q2_Q10_s16x4 ); + rd2_Q10_s32x4 = vmlal_s16( rd2_Q10_s32x4, rr_Q10_s16x4, rr_Q10_s16x4 ); + rd2_Q10_s32x4 = vshrq_n_s32( rd2_Q10_s32x4, 10 ); + + tmp2_s32x4 = vld1q_s32( psDelDec->RD_Q10 ); + tmp1_s32x4 = vaddq_s32( tmp2_s32x4, vminq_s32( rd1_Q10_s32x4, rd2_Q10_s32x4 ) ); + tmp2_s32x4 = vaddq_s32( tmp2_s32x4, vmaxq_s32( rd1_Q10_s32x4, rd2_Q10_s32x4 ) ); + vst1q_s32( psSampleState[ 0 ].RD_Q10, tmp1_s32x4 ); + vst1q_s32( psSampleState[ 1 ].RD_Q10, tmp2_s32x4 ); + t_u32x4 = vcltq_s32( rd1_Q10_s32x4, rd2_Q10_s32x4 ); + tmp1_s32x4 = vbslq_s32( t_u32x4, vmovl_s16( q1_Q10_s16x4 ), vmovl_s16( q2_Q10_s16x4 ) ); + tmp2_s32x4 = vbslq_s32( t_u32x4, vmovl_s16( q2_Q10_s16x4 ), vmovl_s16( q1_Q10_s16x4 ) ); + vst1q_s32( psSampleState[ 0 ].Q_Q10, tmp1_s32x4 ); + vst1q_s32( psSampleState[ 1 ].Q_Q10, tmp2_s32x4 ); + } + + { + /* Update states for best quantization */ + int32x4_t exc_Q14_s32x4, LPC_exc_Q14_s32x4, xq_Q14_s32x4, sLF_AR_shp_Q14_s32x4; + + /* Quantized excitation */ + exc_Q14_s32x4 = vshlq_n_s32( tmp1_s32x4, 4 ); + exc_Q14_s32x4 = veorq_s32( exc_Q14_s32x4, sign_s32x4 ); + exc_Q14_s32x4 = vsubq_s32( exc_Q14_s32x4, sign_s32x4 ); + + /* Add predictions */ + LPC_exc_Q14_s32x4 = vaddq_s32( exc_Q14_s32x4, vdupq_n_s32( LTP_pred_Q14 ) ); + xq_Q14_s32x4 = vaddq_s32( LPC_exc_Q14_s32x4, LPC_pred_Q14_s32x4 ); + + /* Update states */ + tmp1_s32x4 = vsubq_s32( xq_Q14_s32x4, vshlq_n_s32( vdupq_n_s32( x_Q10[ i ] ), 4 ) ); + vst1q_s32( psSampleState[ 0 ].Diff_Q14, tmp1_s32x4 ); + sLF_AR_shp_Q14_s32x4 = vsubq_s32( tmp1_s32x4, n_AR_Q14_s32x4 ); + vst1q_s32( psSampleState[ 0 ].sLTP_shp_Q14, vsubq_s32( sLF_AR_shp_Q14_s32x4, n_LF_Q14_s32x4 ) ); + vst1q_s32( psSampleState[ 0 ].LF_AR_Q14, sLF_AR_shp_Q14_s32x4 ); + vst1q_s32( psSampleState[ 0 ].LPC_exc_Q14, LPC_exc_Q14_s32x4 ); + vst1q_s32( psSampleState[ 0 ].xq_Q14, xq_Q14_s32x4 ); + + /* Quantized excitation */ + exc_Q14_s32x4 = vshlq_n_s32( tmp2_s32x4, 4 ); + exc_Q14_s32x4 = veorq_s32( exc_Q14_s32x4, sign_s32x4 ); + exc_Q14_s32x4 = vsubq_s32( exc_Q14_s32x4, sign_s32x4 ); + + /* Add predictions */ + LPC_exc_Q14_s32x4 = vaddq_s32( exc_Q14_s32x4, vdupq_n_s32( LTP_pred_Q14 ) ); + xq_Q14_s32x4 = vaddq_s32( LPC_exc_Q14_s32x4, LPC_pred_Q14_s32x4 ); + + /* Update states */ + tmp1_s32x4 = vsubq_s32( xq_Q14_s32x4, vshlq_n_s32( vdupq_n_s32( x_Q10[ i ] ), 4 ) ); + vst1q_s32( psSampleState[ 1 ].Diff_Q14, tmp1_s32x4 ); + sLF_AR_shp_Q14_s32x4 = vsubq_s32( tmp1_s32x4, n_AR_Q14_s32x4 ); + vst1q_s32( psSampleState[ 1 ].sLTP_shp_Q14, vsubq_s32( sLF_AR_shp_Q14_s32x4, n_LF_Q14_s32x4 ) ); + vst1q_s32( psSampleState[ 1 ].LF_AR_Q14, sLF_AR_shp_Q14_s32x4 ); + vst1q_s32( psSampleState[ 1 ].LPC_exc_Q14, LPC_exc_Q14_s32x4 ); + vst1q_s32( psSampleState[ 1 ].xq_Q14, xq_Q14_s32x4 ); + } + + *smpl_buf_idx = *smpl_buf_idx ? ( *smpl_buf_idx - 1 ) : ( DECISION_DELAY - 1); + last_smple_idx = *smpl_buf_idx + decisionDelay + DECISION_DELAY; + if( last_smple_idx >= DECISION_DELAY ) last_smple_idx -= DECISION_DELAY; + if( last_smple_idx >= DECISION_DELAY ) last_smple_idx -= DECISION_DELAY; + + /* Find winner */ + RDmin_Q10 = psSampleState[ 0 ].RD_Q10[ 0 ]; + Winner_ind = 0; + for( k = 1; k < nStatesDelayedDecision; k++ ) { + if( psSampleState[ 0 ].RD_Q10[ k ] < RDmin_Q10 ) { + RDmin_Q10 = psSampleState[ 0 ].RD_Q10[ k ]; + Winner_ind = k; + } + } + + /* Increase RD values of expired states */ + { + uint32x4_t t_u32x4; + Winner_rand_state = psDelDec->RandState[ last_smple_idx ][ Winner_ind ]; + t_u32x4 = vceqq_s32( vld1q_s32( psDelDec->RandState[ last_smple_idx ] ), vdupq_n_s32( Winner_rand_state ) ); + t_u32x4 = vmvnq_u32( t_u32x4 ); + t_u32x4 = vshrq_n_u32( t_u32x4, 5 ); + tmp1_s32x4 = vld1q_s32( psSampleState[ 0 ].RD_Q10 ); + tmp2_s32x4 = vld1q_s32( psSampleState[ 1 ].RD_Q10 ); + tmp1_s32x4 = vaddq_s32( tmp1_s32x4, vreinterpretq_s32_u32( t_u32x4 ) ); + tmp2_s32x4 = vaddq_s32( tmp2_s32x4, vreinterpretq_s32_u32( t_u32x4 ) ); + vst1q_s32( psSampleState[ 0 ].RD_Q10, tmp1_s32x4 ); + vst1q_s32( psSampleState[ 1 ].RD_Q10, tmp2_s32x4 ); + + /* Find worst in first set and best in second set */ + RDmax_Q10 = psSampleState[ 0 ].RD_Q10[ 0 ]; + RDmin_Q10 = psSampleState[ 1 ].RD_Q10[ 0 ]; + RDmax_ind = 0; + RDmin_ind = 0; + for( k = 1; k < nStatesDelayedDecision; k++ ) { + /* find worst in first set */ + if( psSampleState[ 0 ].RD_Q10[ k ] > RDmax_Q10 ) { + RDmax_Q10 = psSampleState[ 0 ].RD_Q10[ k ]; + RDmax_ind = k; + } + /* find best in second set */ + if( psSampleState[ 1 ].RD_Q10[ k ] < RDmin_Q10 ) { + RDmin_Q10 = psSampleState[ 1 ].RD_Q10[ k ]; + RDmin_ind = k; + } + } + } + + /* Replace a state if best from second set outperforms worst in first set */ + if( RDmin_Q10 < RDmax_Q10 ) { + opus_int32 (*ptr)[NEON_MAX_DEL_DEC_STATES] = psDelDec->RandState; + const int numOthers = (int)( ( sizeof( NSQ_del_decs_struct ) - sizeof( ( (NSQ_del_decs_struct *)0 )->sLPC_Q14 ) ) + / ( NEON_MAX_DEL_DEC_STATES * sizeof( opus_int32 ) ) ); + /* Only ( predictLPCOrder - 1 ) of sLPC_Q14 buffer need to be updated, though the first several */ + /* useless sLPC_Q14[] will be different comparing with C when predictLPCOrder < NSQ_LPC_BUF_LENGTH. */ + /* Here just update constant ( NSQ_LPC_BUF_LENGTH - 1 ) for simplicity. */ + for( j = i + 1; j < i + NSQ_LPC_BUF_LENGTH; j++ ) { + psDelDec->sLPC_Q14[ j ][ RDmax_ind ] = psDelDec->sLPC_Q14[ j ][ RDmin_ind ]; + } + for( j = 0; j < numOthers; j++ ) { + ptr[ j ][ RDmax_ind ] = ptr[ j ][ RDmin_ind ]; + } + + psSampleState[ 0 ].Q_Q10[ RDmax_ind ] = psSampleState[ 1 ].Q_Q10[ RDmin_ind ]; + psSampleState[ 0 ].RD_Q10[ RDmax_ind ] = psSampleState[ 1 ].RD_Q10[ RDmin_ind ]; + psSampleState[ 0 ].xq_Q14[ RDmax_ind ] = psSampleState[ 1 ].xq_Q14[ RDmin_ind ]; + psSampleState[ 0 ].LF_AR_Q14[ RDmax_ind ] = psSampleState[ 1 ].LF_AR_Q14[ RDmin_ind ]; + psSampleState[ 0 ].Diff_Q14[ RDmax_ind ] = psSampleState[ 1 ].Diff_Q14[ RDmin_ind ]; + psSampleState[ 0 ].sLTP_shp_Q14[ RDmax_ind ] = psSampleState[ 1 ].sLTP_shp_Q14[ RDmin_ind ]; + psSampleState[ 0 ].LPC_exc_Q14[ RDmax_ind ] = psSampleState[ 1 ].LPC_exc_Q14[ RDmin_ind ]; + } + + /* Write samples from winner to output and long-term filter states */ + if( subfr > 0 || i >= decisionDelay ) { + pulses[ i - decisionDelay ] = (opus_int8)silk_RSHIFT_ROUND( psDelDec->Q_Q10[ last_smple_idx ][ Winner_ind ], 10 ); + xq[ i - decisionDelay ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( + silk_SMULWW( psDelDec->Xq_Q14[ last_smple_idx ][ Winner_ind ], delayedGain_Q10[ last_smple_idx ] ), 8 ) ); + NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx - decisionDelay ] = psDelDec->Shape_Q14[ last_smple_idx ][ Winner_ind ]; + sLTP_Q15[ NSQ->sLTP_buf_idx - decisionDelay ] = psDelDec->Pred_Q15[ last_smple_idx ][ Winner_ind ]; + } + NSQ->sLTP_shp_buf_idx++; + NSQ->sLTP_buf_idx++; + + /* Update states */ + vst1q_s32( psDelDec->LF_AR_Q14, vld1q_s32( psSampleState[ 0 ].LF_AR_Q14 ) ); + vst1q_s32( psDelDec->Diff_Q14, vld1q_s32( psSampleState[ 0 ].Diff_Q14 ) ); + vst1q_s32( psDelDec->sLPC_Q14[ NSQ_LPC_BUF_LENGTH + i ], vld1q_s32( psSampleState[ 0 ].xq_Q14 ) ); + vst1q_s32( psDelDec->Xq_Q14[ *smpl_buf_idx ], vld1q_s32( psSampleState[ 0 ].xq_Q14 ) ); + tmp1_s32x4 = vld1q_s32( psSampleState[ 0 ].Q_Q10 ); + vst1q_s32( psDelDec->Q_Q10[ *smpl_buf_idx ], tmp1_s32x4 ); + vst1q_s32( psDelDec->Pred_Q15[ *smpl_buf_idx ], vshlq_n_s32( vld1q_s32( psSampleState[ 0 ].LPC_exc_Q14 ), 1 ) ); + vst1q_s32( psDelDec->Shape_Q14[ *smpl_buf_idx ], vld1q_s32( psSampleState[ 0 ].sLTP_shp_Q14 ) ); + tmp1_s32x4 = vrshrq_n_s32( tmp1_s32x4, 10 ); + tmp1_s32x4 = vaddq_s32( vld1q_s32( psDelDec->Seed ), tmp1_s32x4 ); + vst1q_s32( psDelDec->Seed, tmp1_s32x4 ); + vst1q_s32( psDelDec->RandState[ *smpl_buf_idx ], tmp1_s32x4 ); + vst1q_s32( psDelDec->RD_Q10, vld1q_s32( psSampleState[ 0 ].RD_Q10 ) ); + delayedGain_Q10[ *smpl_buf_idx ] = Gain_Q10; + } + /* Update LPC states */ + silk_memcpy( psDelDec->sLPC_Q14[ 0 ], psDelDec->sLPC_Q14[ length ], NEON_MAX_DEL_DEC_STATES * NSQ_LPC_BUF_LENGTH * sizeof( opus_int32 ) ); + + RESTORE_STACK; +} + +static OPUS_INLINE void silk_SMULWB_8_neon( + const opus_int16 *a, + const int32x2_t b, + opus_int32 *o +) +{ + const int16x8_t a_s16x8 = vld1q_s16( a ); + int32x4_t o0_s32x4, o1_s32x4; + + o0_s32x4 = vshll_n_s16( vget_low_s16( a_s16x8 ), 15 ); + o1_s32x4 = vshll_n_s16( vget_high_s16( a_s16x8 ), 15 ); + o0_s32x4 = vqdmulhq_lane_s32( o0_s32x4, b, 0 ); + o1_s32x4 = vqdmulhq_lane_s32( o1_s32x4, b, 0 ); + vst1q_s32( o, o0_s32x4 ); + vst1q_s32( o + 4, o1_s32x4 ); +} + +/* Only works when ( b >= -65536 ) && ( b < 65536 ). */ +static OPUS_INLINE void silk_SMULWW_small_b_4_neon( + opus_int32 *a, + const int32x2_t b_s32x2) +{ + int32x4_t o_s32x4; + + o_s32x4 = vld1q_s32( a ); + o_s32x4 = vqdmulhq_lane_s32( o_s32x4, b_s32x2, 0 ); + vst1q_s32( a, o_s32x4 ); +} + +/* Only works when ( b >= -65536 ) && ( b < 65536 ). */ +static OPUS_INLINE void silk_SMULWW_small_b_8_neon( + opus_int32 *a, + const int32x2_t b_s32x2 +) +{ + int32x4_t o0_s32x4, o1_s32x4; + + o0_s32x4 = vld1q_s32( a ); + o1_s32x4 = vld1q_s32( a + 4 ); + o0_s32x4 = vqdmulhq_lane_s32( o0_s32x4, b_s32x2, 0 ); + o1_s32x4 = vqdmulhq_lane_s32( o1_s32x4, b_s32x2, 0 ); + vst1q_s32( a, o0_s32x4 ); + vst1q_s32( a + 4, o1_s32x4 ); +} + +static OPUS_INLINE void silk_SMULWW_4_neon( + opus_int32 *a, + const int32x2_t b_s32x2) +{ + int32x4_t a_s32x4, o_s32x4; + + a_s32x4 = vld1q_s32( a ); + o_s32x4 = vqdmulhq_lane_s32( a_s32x4, b_s32x2, 0 ); + o_s32x4 = vmlaq_lane_s32( o_s32x4, a_s32x4, b_s32x2, 1 ); + vst1q_s32( a, o_s32x4 ); +} + +static OPUS_INLINE void silk_SMULWW_8_neon( + opus_int32 *a, + const int32x2_t b_s32x2 +) +{ + int32x4_t a0_s32x4, a1_s32x4, o0_s32x4, o1_s32x4; + + a0_s32x4 = vld1q_s32( a ); + a1_s32x4 = vld1q_s32( a + 4 ); + o0_s32x4 = vqdmulhq_lane_s32( a0_s32x4, b_s32x2, 0 ); + o1_s32x4 = vqdmulhq_lane_s32( a1_s32x4, b_s32x2, 0 ); + o0_s32x4 = vmlaq_lane_s32( o0_s32x4, a0_s32x4, b_s32x2, 1 ); + o1_s32x4 = vmlaq_lane_s32( o1_s32x4, a1_s32x4, b_s32x2, 1 ); + vst1q_s32( a, o0_s32x4 ); + vst1q_s32( a + 4, o1_s32x4 ); +} + +static OPUS_INLINE void silk_SMULWW_loop_neon( + const opus_int16 *a, + const opus_int32 b, + opus_int32 *o, + const opus_int loop_num +) +{ + opus_int i; + int32x2_t b_s32x2; + + b_s32x2 = vdup_n_s32( b ); + for( i = 0; i < loop_num - 7; i += 8 ) { + silk_SMULWB_8_neon( a + i, b_s32x2, o + i ); + } + for( ; i < loop_num; i++ ) { + o[ i ] = silk_SMULWW( a[ i ], b ); + } +} + +static OPUS_INLINE void silk_nsq_del_dec_scale_states_neon( + const silk_encoder_state *psEncC, /* I Encoder State */ + silk_nsq_state *NSQ, /* I/O NSQ state */ + NSQ_del_decs_struct psDelDec[], /* I/O Delayed decision states */ + const opus_int16 x16[], /* I Input */ + opus_int32 x_sc_Q10[], /* O Input scaled with 1/Gain in Q10 */ + const opus_int16 sLTP[], /* I Re-whitened LTP state in Q0 */ + opus_int32 sLTP_Q15[], /* O LTP state matching scaled input */ + opus_int subfr, /* I Subframe number */ + const opus_int LTP_scale_Q14, /* I LTP state scaling */ + const opus_int32 Gains_Q16[ MAX_NB_SUBFR ], /* I */ + const opus_int pitchL[ MAX_NB_SUBFR ], /* I Pitch lag */ + const opus_int signal_type, /* I Signal type */ + const opus_int decisionDelay /* I Decision delay */ +) +{ + opus_int i, lag; + opus_int32 gain_adj_Q16, inv_gain_Q31, inv_gain_Q26; + + lag = pitchL[ subfr ]; + inv_gain_Q31 = silk_INVERSE32_varQ( silk_max( Gains_Q16[ subfr ], 1 ), 47 ); + silk_assert( inv_gain_Q31 != 0 ); + + /* Scale input */ + inv_gain_Q26 = silk_RSHIFT_ROUND( inv_gain_Q31, 5 ); + silk_SMULWW_loop_neon( x16, inv_gain_Q26, x_sc_Q10, psEncC->subfr_length ); + + /* After rewhitening the LTP state is un-scaled, so scale with inv_gain_Q16 */ + if( NSQ->rewhite_flag ) { + if( subfr == 0 ) { + /* Do LTP downscaling */ + inv_gain_Q31 = silk_LSHIFT( silk_SMULWB( inv_gain_Q31, LTP_scale_Q14 ), 2 ); + } + silk_SMULWW_loop_neon( sLTP + NSQ->sLTP_buf_idx - lag - LTP_ORDER / 2, inv_gain_Q31, sLTP_Q15 + NSQ->sLTP_buf_idx - lag - LTP_ORDER / 2, lag + LTP_ORDER / 2 ); + } + + /* Adjust for changing gain */ + if( Gains_Q16[ subfr ] != NSQ->prev_gain_Q16 ) { + int32x2_t gain_adj_Q16_s32x2; + gain_adj_Q16 = silk_DIV32_varQ( NSQ->prev_gain_Q16, Gains_Q16[ subfr ], 16 ); + + /* Scale long-term shaping state */ + if( ( gain_adj_Q16 >= -65536 ) && ( gain_adj_Q16 < 65536 ) ) { + gain_adj_Q16_s32x2 = vdup_n_s32( silk_LSHIFT32( gain_adj_Q16, 15 ) ); + for( i = NSQ->sLTP_shp_buf_idx - psEncC->ltp_mem_length; i < NSQ->sLTP_shp_buf_idx - 7; i += 8 ) { + silk_SMULWW_small_b_8_neon( NSQ->sLTP_shp_Q14 + i, gain_adj_Q16_s32x2 ); + } + for( ; i < NSQ->sLTP_shp_buf_idx; i++ ) { + NSQ->sLTP_shp_Q14[ i ] = silk_SMULWW( gain_adj_Q16, NSQ->sLTP_shp_Q14[ i ] ); + } + + /* Scale long-term prediction state */ + if( signal_type == TYPE_VOICED && NSQ->rewhite_flag == 0 ) { + for( i = NSQ->sLTP_buf_idx - lag - LTP_ORDER / 2; i < NSQ->sLTP_buf_idx - decisionDelay - 7; i += 8 ) { + silk_SMULWW_small_b_8_neon( sLTP_Q15 + i, gain_adj_Q16_s32x2 ); + } + for( ; i < NSQ->sLTP_buf_idx - decisionDelay; i++ ) { + sLTP_Q15[ i ] = silk_SMULWW( gain_adj_Q16, sLTP_Q15[ i ] ); + } + } + + /* Scale scalar states */ + silk_SMULWW_small_b_4_neon( psDelDec->LF_AR_Q14, gain_adj_Q16_s32x2 ); + silk_SMULWW_small_b_4_neon( psDelDec->Diff_Q14, gain_adj_Q16_s32x2 ); + + /* Scale short-term prediction and shaping states */ + for( i = 0; i < NSQ_LPC_BUF_LENGTH; i++ ) { + silk_SMULWW_small_b_4_neon( psDelDec->sLPC_Q14[ i ], gain_adj_Q16_s32x2 ); + } + + for( i = 0; i < MAX_SHAPE_LPC_ORDER; i++ ) { + silk_SMULWW_small_b_4_neon( psDelDec->sAR2_Q14[ i ], gain_adj_Q16_s32x2 ); + } + + for( i = 0; i < DECISION_DELAY; i++ ) { + silk_SMULWW_small_b_4_neon( psDelDec->Pred_Q15[ i ], gain_adj_Q16_s32x2 ); + silk_SMULWW_small_b_4_neon( psDelDec->Shape_Q14[ i ], gain_adj_Q16_s32x2 ); + } + } else { + gain_adj_Q16_s32x2 = vdup_n_s32( silk_LSHIFT32( gain_adj_Q16 & 0x0000FFFF, 15 ) ); + gain_adj_Q16_s32x2 = vset_lane_s32( gain_adj_Q16 >> 16, gain_adj_Q16_s32x2, 1 ); + for( i = NSQ->sLTP_shp_buf_idx - psEncC->ltp_mem_length; i < NSQ->sLTP_shp_buf_idx - 7; i += 8 ) { + silk_SMULWW_8_neon( NSQ->sLTP_shp_Q14 + i, gain_adj_Q16_s32x2 ); + } + for( ; i < NSQ->sLTP_shp_buf_idx; i++ ) { + NSQ->sLTP_shp_Q14[ i ] = silk_SMULWW( gain_adj_Q16, NSQ->sLTP_shp_Q14[ i ] ); + } + + /* Scale long-term prediction state */ + if( signal_type == TYPE_VOICED && NSQ->rewhite_flag == 0 ) { + for( i = NSQ->sLTP_buf_idx - lag - LTP_ORDER / 2; i < NSQ->sLTP_buf_idx - decisionDelay - 7; i += 8 ) { + silk_SMULWW_8_neon( sLTP_Q15 + i, gain_adj_Q16_s32x2 ); + } + for( ; i < NSQ->sLTP_buf_idx - decisionDelay; i++ ) { + sLTP_Q15[ i ] = silk_SMULWW( gain_adj_Q16, sLTP_Q15[ i ] ); + } + } + + /* Scale scalar states */ + silk_SMULWW_4_neon( psDelDec->LF_AR_Q14, gain_adj_Q16_s32x2 ); + silk_SMULWW_4_neon( psDelDec->Diff_Q14, gain_adj_Q16_s32x2 ); + + /* Scale short-term prediction and shaping states */ + for( i = 0; i < NSQ_LPC_BUF_LENGTH; i++ ) { + silk_SMULWW_4_neon( psDelDec->sLPC_Q14[ i ], gain_adj_Q16_s32x2 ); + } + + for( i = 0; i < MAX_SHAPE_LPC_ORDER; i++ ) { + silk_SMULWW_4_neon( psDelDec->sAR2_Q14[ i ], gain_adj_Q16_s32x2 ); + } + + for( i = 0; i < DECISION_DELAY; i++ ) { + silk_SMULWW_4_neon( psDelDec->Pred_Q15[ i ], gain_adj_Q16_s32x2 ); + silk_SMULWW_4_neon( psDelDec->Shape_Q14[ i ], gain_adj_Q16_s32x2 ); + } + } + + /* Save inverse gain */ + NSQ->prev_gain_Q16 = Gains_Q16[ subfr ]; + } +} diff --git a/code/opus-1.1.4/silk/arm/NSQ_neon.h b/code/opus-1.1.4/silk/arm/NSQ_neon.h index 77c946af..b31d9442 100644 --- a/code/opus-1.1.4/silk/arm/NSQ_neon.h +++ b/code/opus-1.1.4/silk/arm/NSQ_neon.h @@ -28,30 +28,31 @@ POSSIBILITY OF SUCH DAMAGE. #define SILK_NSQ_NEON_H #include "cpu_support.h" +#include "SigProc_FIX.h" #undef silk_short_prediction_create_arch_coef /* For vectorized calc, reverse a_Q12 coefs, convert to 32-bit, and shift for vqdmulhq_s32. */ static OPUS_INLINE void silk_short_prediction_create_arch_coef_neon(opus_int32 *out, const opus_int16 *in, opus_int order) { - out[15] = in[0] << 15; - out[14] = in[1] << 15; - out[13] = in[2] << 15; - out[12] = in[3] << 15; - out[11] = in[4] << 15; - out[10] = in[5] << 15; - out[9] = in[6] << 15; - out[8] = in[7] << 15; - out[7] = in[8] << 15; - out[6] = in[9] << 15; + out[15] = silk_LSHIFT32(in[0], 15); + out[14] = silk_LSHIFT32(in[1], 15); + out[13] = silk_LSHIFT32(in[2], 15); + out[12] = silk_LSHIFT32(in[3], 15); + out[11] = silk_LSHIFT32(in[4], 15); + out[10] = silk_LSHIFT32(in[5], 15); + out[9] = silk_LSHIFT32(in[6], 15); + out[8] = silk_LSHIFT32(in[7], 15); + out[7] = silk_LSHIFT32(in[8], 15); + out[6] = silk_LSHIFT32(in[9], 15); if (order == 16) { - out[5] = in[10] << 15; - out[4] = in[11] << 15; - out[3] = in[12] << 15; - out[2] = in[13] << 15; - out[1] = in[14] << 15; - out[0] = in[15] << 15; + out[5] = silk_LSHIFT32(in[10], 15); + out[4] = silk_LSHIFT32(in[11], 15); + out[3] = silk_LSHIFT32(in[12], 15); + out[2] = silk_LSHIFT32(in[13], 15); + out[1] = silk_LSHIFT32(in[14], 15); + out[0] = silk_LSHIFT32(in[15], 15); } else { diff --git a/code/opus-1.1.4/silk/arm/arm_silk_map.c b/code/opus-1.1.4/silk/arm/arm_silk_map.c index 9bd86a7b..0b9bfec2 100644 --- a/code/opus-1.1.4/silk/arm/arm_silk_map.c +++ b/code/opus-1.1.4/silk/arm/arm_silk_map.c @@ -28,13 +28,62 @@ POSSIBILITY OF SUCH DAMAGE. # include "config.h" #endif +#include "main_FIX.h" #include "NSQ.h" +#include "SigProc_FIX.h" #if defined(OPUS_HAVE_RTCD) # if (defined(OPUS_ARM_MAY_HAVE_NEON_INTR) && \ !defined(OPUS_ARM_PRESUME_NEON_INTR)) +void (*const SILK_BIQUAD_ALT_STRIDE2_IMPL[OPUS_ARCHMASK + 1])( + const opus_int16 *in, /* I input signal */ + const opus_int32 *B_Q28, /* I MA coefficients [3] */ + const opus_int32 *A_Q28, /* I AR coefficients [2] */ + opus_int32 *S, /* I/O State vector [4] */ + opus_int16 *out, /* O output signal */ + const opus_int32 len /* I signal length (must be even) */ +) = { + silk_biquad_alt_stride2_c, /* ARMv4 */ + silk_biquad_alt_stride2_c, /* EDSP */ + silk_biquad_alt_stride2_c, /* Media */ + silk_biquad_alt_stride2_neon, /* Neon */ +}; + +opus_int32 (*const SILK_LPC_INVERSE_PRED_GAIN_IMPL[OPUS_ARCHMASK + 1])( /* O Returns inverse prediction gain in energy domain, Q30 */ + const opus_int16 *A_Q12, /* I Prediction coefficients, Q12 [order] */ + const opus_int order /* I Prediction order */ +) = { + silk_LPC_inverse_pred_gain_c, /* ARMv4 */ + silk_LPC_inverse_pred_gain_c, /* EDSP */ + silk_LPC_inverse_pred_gain_c, /* Media */ + silk_LPC_inverse_pred_gain_neon, /* Neon */ +}; + +void (*const SILK_NSQ_DEL_DEC_IMPL[OPUS_ARCHMASK + 1])( + const silk_encoder_state *psEncC, /* I Encoder State */ + silk_nsq_state *NSQ, /* I/O NSQ state */ + SideInfoIndices *psIndices, /* I/O Quantization Indices */ + const opus_int16 x16[], /* I Input */ + opus_int8 pulses[], /* O Quantized pulse signal */ + const opus_int16 PredCoef_Q12[ 2 * MAX_LPC_ORDER ], /* I Short term prediction coefs */ + const opus_int16 LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ], /* I Long term prediction coefs */ + const opus_int16 AR_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs */ + const opus_int HarmShapeGain_Q14[ MAX_NB_SUBFR ], /* I Long term shaping coefs */ + const opus_int Tilt_Q14[ MAX_NB_SUBFR ], /* I Spectral tilt */ + const opus_int32 LF_shp_Q14[ MAX_NB_SUBFR ], /* I Low frequency shaping coefs */ + const opus_int32 Gains_Q16[ MAX_NB_SUBFR ], /* I Quantization step sizes */ + const opus_int pitchL[ MAX_NB_SUBFR ], /* I Pitch lags */ + const opus_int Lambda_Q10, /* I Rate/distortion tradeoff */ + const opus_int LTP_scale_Q14 /* I LTP state scaling */ +) = { + silk_NSQ_del_dec_c, /* ARMv4 */ + silk_NSQ_del_dec_c, /* EDSP */ + silk_NSQ_del_dec_c, /* Media */ + silk_NSQ_del_dec_neon, /* Neon */ +}; + /*There is no table for silk_noise_shape_quantizer_short_prediction because the NEON version takes different parameters than the C version. Instead RTCD is done via if statements at the call sites. @@ -52,4 +101,23 @@ opus_int32 # endif +# if defined(FIXED_POINT) && \ + defined(OPUS_ARM_MAY_HAVE_NEON_INTR) && !defined(OPUS_ARM_PRESUME_NEON_INTR) + +void (*const SILK_WARPED_AUTOCORRELATION_FIX_IMPL[OPUS_ARCHMASK + 1])( + opus_int32 *corr, /* O Result [order + 1] */ + opus_int *scale, /* O Scaling of the correlation vector */ + const opus_int16 *input, /* I Input data to correlate */ + const opus_int warping_Q16, /* I Warping coefficient */ + const opus_int length, /* I Length of input */ + const opus_int order /* I Correlation order (even) */ +) = { + silk_warped_autocorrelation_FIX_c, /* ARMv4 */ + silk_warped_autocorrelation_FIX_c, /* EDSP */ + silk_warped_autocorrelation_FIX_c, /* Media */ + silk_warped_autocorrelation_FIX_neon, /* Neon */ +}; + +# endif + #endif /* OPUS_HAVE_RTCD */ diff --git a/code/opus-1.1.4/silk/arm/biquad_alt_arm.h b/code/opus-1.1.4/silk/arm/biquad_alt_arm.h new file mode 100644 index 00000000..66ea9f43 --- /dev/null +++ b/code/opus-1.1.4/silk/arm/biquad_alt_arm.h @@ -0,0 +1,68 @@ +/*********************************************************************** +Copyright (c) 2017 Google Inc. +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions +are met: +- Redistributions of source code must retain the above copyright notice, +this list of conditions and the following disclaimer. +- Redistributions in binary form must reproduce the above copyright +notice, this list of conditions and the following disclaimer in the +documentation and/or other materials provided with the distribution. +- Neither the name of Internet Society, IETF or IETF Trust, nor the +names of specific contributors, may be used to endorse or promote +products derived from this software without specific prior written +permission. +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE +ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE +LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR +CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF +SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS +INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN +CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) +ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +POSSIBILITY OF SUCH DAMAGE. +***********************************************************************/ + +#ifndef SILK_BIQUAD_ALT_ARM_H +# define SILK_BIQUAD_ALT_ARM_H + +# include "celt/arm/armcpu.h" + +# if defined(OPUS_ARM_MAY_HAVE_NEON_INTR) +void silk_biquad_alt_stride2_neon( + const opus_int16 *in, /* I input signal */ + const opus_int32 *B_Q28, /* I MA coefficients [3] */ + const opus_int32 *A_Q28, /* I AR coefficients [2] */ + opus_int32 *S, /* I/O State vector [4] */ + opus_int16 *out, /* O output signal */ + const opus_int32 len /* I signal length (must be even) */ +); + +# if !defined(OPUS_HAVE_RTCD) && defined(OPUS_ARM_PRESUME_NEON) +# define OVERRIDE_silk_biquad_alt_stride2 (1) +# define silk_biquad_alt_stride2(in, B_Q28, A_Q28, S, out, len, arch) ((void)(arch), PRESUME_NEON(silk_biquad_alt_stride2)(in, B_Q28, A_Q28, S, out, len)) +# endif +# endif + +# if !defined(OVERRIDE_silk_biquad_alt_stride2) +/*Is run-time CPU detection enabled on this platform?*/ +# if defined(OPUS_HAVE_RTCD) && (defined(OPUS_ARM_MAY_HAVE_NEON_INTR) && !defined(OPUS_ARM_PRESUME_NEON_INTR)) +extern void (*const SILK_BIQUAD_ALT_STRIDE2_IMPL[OPUS_ARCHMASK+1])( + const opus_int16 *in, /* I input signal */ + const opus_int32 *B_Q28, /* I MA coefficients [3] */ + const opus_int32 *A_Q28, /* I AR coefficients [2] */ + opus_int32 *S, /* I/O State vector [4] */ + opus_int16 *out, /* O output signal */ + const opus_int32 len /* I signal length (must be even) */ + ); +# define OVERRIDE_silk_biquad_alt_stride2 (1) +# define silk_biquad_alt_stride2(in, B_Q28, A_Q28, S, out, len, arch) ((*SILK_BIQUAD_ALT_STRIDE2_IMPL[(arch)&OPUS_ARCHMASK])(in, B_Q28, A_Q28, S, out, len)) +# elif defined(OPUS_ARM_PRESUME_NEON_INTR) +# define OVERRIDE_silk_biquad_alt_stride2 (1) +# define silk_biquad_alt_stride2(in, B_Q28, A_Q28, S, out, len, arch) ((void)(arch), silk_biquad_alt_stride2_neon(in, B_Q28, A_Q28, S, out, len)) +# endif +# endif + +#endif /* end SILK_BIQUAD_ALT_ARM_H */ diff --git a/code/opus-1.1.4/silk/arm/biquad_alt_neon_intr.c b/code/opus-1.1.4/silk/arm/biquad_alt_neon_intr.c new file mode 100644 index 00000000..97157331 --- /dev/null +++ b/code/opus-1.1.4/silk/arm/biquad_alt_neon_intr.c @@ -0,0 +1,156 @@ +/*********************************************************************** +Copyright (c) 2017 Google Inc. +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions +are met: +- Redistributions of source code must retain the above copyright notice, +this list of conditions and the following disclaimer. +- Redistributions in binary form must reproduce the above copyright +notice, this list of conditions and the following disclaimer in the +documentation and/or other materials provided with the distribution. +- Neither the name of Internet Society, IETF or IETF Trust, nor the +names of specific contributors, may be used to endorse or promote +products derived from this software without specific prior written +permission. +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE +ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE +LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR +CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF +SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS +INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN +CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) +ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +POSSIBILITY OF SUCH DAMAGE. +***********************************************************************/ + +#ifdef HAVE_CONFIG_H +#include "config.h" +#endif + +#include +#ifdef OPUS_CHECK_ASM +# include +# include "stack_alloc.h" +#endif +#include "SigProc_FIX.h" + +static inline void silk_biquad_alt_stride2_kernel( const int32x4_t A_L_s32x4, const int32x4_t A_U_s32x4, const int32x4_t B_Q28_s32x4, const int32x2_t t_s32x2, const int32x4_t in_s32x4, int32x4_t *S_s32x4, int32x2_t *out32_Q14_s32x2 ) +{ + int32x4_t t_s32x4, out32_Q14_s32x4; + + *out32_Q14_s32x2 = vadd_s32( vget_low_s32( *S_s32x4 ), t_s32x2 ); /* silk_SMLAWB( S{0,1}, B_Q28[ 0 ], in{0,1} ) */ + *S_s32x4 = vcombine_s32( vget_high_s32( *S_s32x4 ), vdup_n_s32( 0 ) ); /* S{0,1} = S{2,3}; S{2,3} = 0; */ + *out32_Q14_s32x2 = vshl_n_s32( *out32_Q14_s32x2, 2 ); /* out32_Q14_{0,1} = silk_LSHIFT( silk_SMLAWB( S{0,1}, B_Q28[ 0 ], in{0,1} ), 2 ); */ + out32_Q14_s32x4 = vcombine_s32( *out32_Q14_s32x2, *out32_Q14_s32x2 ); /* out32_Q14_{0,1,0,1} */ + t_s32x4 = vqdmulhq_s32( out32_Q14_s32x4, A_L_s32x4 ); /* silk_SMULWB( out32_Q14_{0,1,0,1}, A{0,0,1,1}_L_Q28 ) */ + *S_s32x4 = vrsraq_n_s32( *S_s32x4, t_s32x4, 14 ); /* S{0,1} = S{2,3} + silk_RSHIFT_ROUND(); S{2,3} = silk_RSHIFT_ROUND(); */ + t_s32x4 = vqdmulhq_s32( out32_Q14_s32x4, A_U_s32x4 ); /* silk_SMULWB( out32_Q14_{0,1,0,1}, A{0,0,1,1}_U_Q28 ) */ + *S_s32x4 = vaddq_s32( *S_s32x4, t_s32x4 ); /* S0 = silk_SMLAWB( S{0,1,2,3}, out32_Q14_{0,1,0,1}, A{0,0,1,1}_U_Q28 ); */ + t_s32x4 = vqdmulhq_s32( in_s32x4, B_Q28_s32x4 ); /* silk_SMULWB( B_Q28[ {1,1,2,2} ], in{0,1,0,1} ) */ + *S_s32x4 = vaddq_s32( *S_s32x4, t_s32x4 ); /* S0 = silk_SMLAWB( S0, B_Q28[ {1,1,2,2} ], in{0,1,0,1} ); */ +} + +void silk_biquad_alt_stride2_neon( + const opus_int16 *in, /* I input signal */ + const opus_int32 *B_Q28, /* I MA coefficients [3] */ + const opus_int32 *A_Q28, /* I AR coefficients [2] */ + opus_int32 *S, /* I/O State vector [4] */ + opus_int16 *out, /* O output signal */ + const opus_int32 len /* I signal length (must be even) */ +) +{ + /* DIRECT FORM II TRANSPOSED (uses 2 element state vector) */ + opus_int k = 0; + const int32x2_t offset_s32x2 = vdup_n_s32( (1<<14) - 1 ); + const int32x4_t offset_s32x4 = vcombine_s32( offset_s32x2, offset_s32x2 ); + int16x4_t in_s16x4 = vdup_n_s16( 0 ); + int16x4_t out_s16x4; + int32x2_t A_Q28_s32x2, A_L_s32x2, A_U_s32x2, B_Q28_s32x2, t_s32x2; + int32x4_t A_L_s32x4, A_U_s32x4, B_Q28_s32x4, S_s32x4, out32_Q14_s32x4; + int32x2x2_t t0_s32x2x2, t1_s32x2x2, t2_s32x2x2, S_s32x2x2; + +#ifdef OPUS_CHECK_ASM + opus_int32 S_c[ 4 ]; + VARDECL( opus_int16, out_c ); + SAVE_STACK; + ALLOC( out_c, 2 * len, opus_int16 ); + + silk_memcpy( &S_c, S, sizeof( S_c ) ); + silk_biquad_alt_stride2_c( in, B_Q28, A_Q28, S_c, out_c, len ); +#endif + + /* Negate A_Q28 values and split in two parts */ + A_Q28_s32x2 = vld1_s32( A_Q28 ); + A_Q28_s32x2 = vneg_s32( A_Q28_s32x2 ); + A_L_s32x2 = vshl_n_s32( A_Q28_s32x2, 18 ); /* ( -A_Q28[] & 0x00003FFF ) << 18 */ + A_L_s32x2 = vreinterpret_s32_u32( vshr_n_u32( vreinterpret_u32_s32( A_L_s32x2 ), 3 ) ); /* ( -A_Q28[] & 0x00003FFF ) << 15 */ + A_U_s32x2 = vshr_n_s32( A_Q28_s32x2, 14 ); /* silk_RSHIFT( -A_Q28[], 14 ) */ + A_U_s32x2 = vshl_n_s32( A_U_s32x2, 16 ); /* silk_RSHIFT( -A_Q28[], 14 ) << 16 (Clip two leading bits to conform to C function.) */ + A_U_s32x2 = vshr_n_s32( A_U_s32x2, 1 ); /* silk_RSHIFT( -A_Q28[], 14 ) << 15 */ + + B_Q28_s32x2 = vld1_s32( B_Q28 ); + t_s32x2 = vld1_s32( B_Q28 + 1 ); + t0_s32x2x2 = vzip_s32( A_L_s32x2, A_L_s32x2 ); + t1_s32x2x2 = vzip_s32( A_U_s32x2, A_U_s32x2 ); + t2_s32x2x2 = vzip_s32( t_s32x2, t_s32x2 ); + A_L_s32x4 = vcombine_s32( t0_s32x2x2.val[ 0 ], t0_s32x2x2.val[ 1 ] ); /* A{0,0,1,1}_L_Q28 */ + A_U_s32x4 = vcombine_s32( t1_s32x2x2.val[ 0 ], t1_s32x2x2.val[ 1 ] ); /* A{0,0,1,1}_U_Q28 */ + B_Q28_s32x4 = vcombine_s32( t2_s32x2x2.val[ 0 ], t2_s32x2x2.val[ 1 ] ); /* B_Q28[ {1,1,2,2} ] */ + S_s32x4 = vld1q_s32( S ); /* S0 = S[ 0 ]; S3 = S[ 3 ]; */ + S_s32x2x2 = vtrn_s32( vget_low_s32( S_s32x4 ), vget_high_s32( S_s32x4 ) ); /* S2 = S[ 1 ]; S1 = S[ 2 ]; */ + S_s32x4 = vcombine_s32( S_s32x2x2.val[ 0 ], S_s32x2x2.val[ 1 ] ); + + for( ; k < len - 1; k += 2 ) { + int32x4_t in_s32x4[ 2 ], t_s32x4; + int32x2_t out32_Q14_s32x2[ 2 ]; + + /* S[ 2 * i + 0 ], S[ 2 * i + 1 ], S[ 2 * i + 2 ], S[ 2 * i + 3 ]: Q12 */ + in_s16x4 = vld1_s16( &in[ 2 * k ] ); /* in{0,1,2,3} = in[ 2 * k + {0,1,2,3} ]; */ + in_s32x4[ 0 ] = vshll_n_s16( in_s16x4, 15 ); /* in{0,1,2,3} << 15 */ + t_s32x4 = vqdmulhq_lane_s32( in_s32x4[ 0 ], B_Q28_s32x2, 0 ); /* silk_SMULWB( B_Q28[ 0 ], in{0,1,2,3} ) */ + in_s32x4[ 1 ] = vcombine_s32( vget_high_s32( in_s32x4[ 0 ] ), vget_high_s32( in_s32x4[ 0 ] ) ); /* in{2,3,2,3} << 15 */ + in_s32x4[ 0 ] = vcombine_s32( vget_low_s32 ( in_s32x4[ 0 ] ), vget_low_s32 ( in_s32x4[ 0 ] ) ); /* in{0,1,0,1} << 15 */ + silk_biquad_alt_stride2_kernel( A_L_s32x4, A_U_s32x4, B_Q28_s32x4, vget_low_s32 ( t_s32x4 ), in_s32x4[ 0 ], &S_s32x4, &out32_Q14_s32x2[ 0 ] ); + silk_biquad_alt_stride2_kernel( A_L_s32x4, A_U_s32x4, B_Q28_s32x4, vget_high_s32( t_s32x4 ), in_s32x4[ 1 ], &S_s32x4, &out32_Q14_s32x2[ 1 ] ); + + /* Scale back to Q0 and saturate */ + out32_Q14_s32x4 = vcombine_s32( out32_Q14_s32x2[ 0 ], out32_Q14_s32x2[ 1 ] ); /* out32_Q14_{0,1,2,3} */ + out32_Q14_s32x4 = vaddq_s32( out32_Q14_s32x4, offset_s32x4 ); /* out32_Q14_{0,1,2,3} + (1<<14) - 1 */ + out_s16x4 = vqshrn_n_s32( out32_Q14_s32x4, 14 ); /* (opus_int16)silk_SAT16( silk_RSHIFT( out32_Q14_{0,1,2,3} + (1<<14) - 1, 14 ) ) */ + vst1_s16( &out[ 2 * k ], out_s16x4 ); /* out[ 2 * k + {0,1,2,3} ] = (opus_int16)silk_SAT16( silk_RSHIFT( out32_Q14_{0,1,2,3} + (1<<14) - 1, 14 ) ); */ + } + + /* Process leftover. */ + if( k < len ) { + int32x4_t in_s32x4; + int32x2_t out32_Q14_s32x2; + + /* S[ 2 * i + 0 ], S[ 2 * i + 1 ]: Q12 */ + in_s16x4 = vld1_lane_s16( &in[ 2 * k + 0 ], in_s16x4, 0 ); /* in{0,1} = in[ 2 * k + {0,1} ]; */ + in_s16x4 = vld1_lane_s16( &in[ 2 * k + 1 ], in_s16x4, 1 ); /* in{0,1} = in[ 2 * k + {0,1} ]; */ + in_s32x4 = vshll_n_s16( in_s16x4, 15 ); /* in{0,1} << 15 */ + t_s32x2 = vqdmulh_lane_s32( vget_low_s32( in_s32x4 ), B_Q28_s32x2, 0 ); /* silk_SMULWB( B_Q28[ 0 ], in{0,1} ) */ + in_s32x4 = vcombine_s32( vget_low_s32( in_s32x4 ), vget_low_s32( in_s32x4 ) ); /* in{0,1,0,1} << 15 */ + silk_biquad_alt_stride2_kernel( A_L_s32x4, A_U_s32x4, B_Q28_s32x4, t_s32x2, in_s32x4, &S_s32x4, &out32_Q14_s32x2 ); + + /* Scale back to Q0 and saturate */ + out32_Q14_s32x2 = vadd_s32( out32_Q14_s32x2, offset_s32x2 ); /* out32_Q14_{0,1} + (1<<14) - 1 */ + out32_Q14_s32x4 = vcombine_s32( out32_Q14_s32x2, out32_Q14_s32x2 ); /* out32_Q14_{0,1,0,1} + (1<<14) - 1 */ + out_s16x4 = vqshrn_n_s32( out32_Q14_s32x4, 14 ); /* (opus_int16)silk_SAT16( silk_RSHIFT( out32_Q14_{0,1,0,1} + (1<<14) - 1, 14 ) ) */ + vst1_lane_s16( &out[ 2 * k + 0 ], out_s16x4, 0 ); /* out[ 2 * k + 0 ] = (opus_int16)silk_SAT16( silk_RSHIFT( out32_Q14_0 + (1<<14) - 1, 14 ) ); */ + vst1_lane_s16( &out[ 2 * k + 1 ], out_s16x4, 1 ); /* out[ 2 * k + 1 ] = (opus_int16)silk_SAT16( silk_RSHIFT( out32_Q14_1 + (1<<14) - 1, 14 ) ); */ + } + + vst1q_lane_s32( &S[ 0 ], S_s32x4, 0 ); /* S[ 0 ] = S0; */ + vst1q_lane_s32( &S[ 1 ], S_s32x4, 2 ); /* S[ 1 ] = S2; */ + vst1q_lane_s32( &S[ 2 ], S_s32x4, 1 ); /* S[ 2 ] = S1; */ + vst1q_lane_s32( &S[ 3 ], S_s32x4, 3 ); /* S[ 3 ] = S3; */ + +#ifdef OPUS_CHECK_ASM + silk_assert( !memcmp( S_c, S, sizeof( S_c ) ) ); + silk_assert( !memcmp( out_c, out, 2 * len * sizeof( opus_int16 ) ) ); + RESTORE_STACK; +#endif +} diff --git a/code/opus-1.1.4/silk/arm/macros_armv4.h b/code/opus-1.1.4/silk/arm/macros_armv4.h index 3f30e972..877eb18d 100644 --- a/code/opus-1.1.4/silk/arm/macros_armv4.h +++ b/code/opus-1.1.4/silk/arm/macros_armv4.h @@ -28,6 +28,11 @@ POSSIBILITY OF SUCH DAMAGE. #ifndef SILK_MACROS_ARMv4_H #define SILK_MACROS_ARMv4_H +/* This macro only avoids the undefined behaviour from a left shift of + a negative value. It should only be used in macros that can't include + SigProc_FIX.h. In other cases, use silk_LSHIFT32(). */ +#define SAFE_SHL(a,b) ((opus_int32)((opus_uint32)(a) << (b))) + /* (a32 * (opus_int32)((opus_int16)(b32))) >> 16 output have to be 32bit int */ #undef silk_SMULWB static OPUS_INLINE opus_int32 silk_SMULWB_armv4(opus_int32 a, opus_int16 b) @@ -38,7 +43,7 @@ static OPUS_INLINE opus_int32 silk_SMULWB_armv4(opus_int32 a, opus_int16 b) "#silk_SMULWB\n\t" "smull %0, %1, %2, %3\n\t" : "=&r"(rd_lo), "=&r"(rd_hi) - : "%r"(a), "r"(b<<16) + : "%r"(a), "r"(SAFE_SHL(b,16)) ); return rd_hi; } @@ -80,7 +85,7 @@ static OPUS_INLINE opus_int32 silk_SMULWW_armv4(opus_int32 a, opus_int32 b) : "=&r"(rd_lo), "=&r"(rd_hi) : "%r"(a), "r"(b) ); - return (rd_hi<<16)+(rd_lo>>16); + return SAFE_SHL(rd_hi,16)+(rd_lo>>16); } #define silk_SMULWW(a, b) (silk_SMULWW_armv4(a, b)) @@ -96,8 +101,10 @@ static OPUS_INLINE opus_int32 silk_SMLAWW_armv4(opus_int32 a, opus_int32 b, : "=&r"(rd_lo), "=&r"(rd_hi) : "%r"(b), "r"(c) ); - return a+(rd_hi<<16)+(rd_lo>>16); + return a+SAFE_SHL(rd_hi,16)+(rd_lo>>16); } #define silk_SMLAWW(a, b, c) (silk_SMLAWW_armv4(a, b, c)) +#undef SAFE_SHL + #endif /* SILK_MACROS_ARMv4_H */ diff --git a/code/opus-1.1.4/silk/arm/macros_armv5e.h b/code/opus-1.1.4/silk/arm/macros_armv5e.h index aad4117e..b14ec65d 100644 --- a/code/opus-1.1.4/silk/arm/macros_armv5e.h +++ b/code/opus-1.1.4/silk/arm/macros_armv5e.h @@ -29,6 +29,11 @@ POSSIBILITY OF SUCH DAMAGE. #ifndef SILK_MACROS_ARMv5E_H #define SILK_MACROS_ARMv5E_H +/* This macro only avoids the undefined behaviour from a left shift of + a negative value. It should only be used in macros that can't include + SigProc_FIX.h. In other cases, use silk_LSHIFT32(). */ +#define SAFE_SHL(a,b) ((opus_int32)((opus_uint32)(a) << (b))) + /* (a32 * (opus_int32)((opus_int16)(b32))) >> 16 output have to be 32bit int */ #undef silk_SMULWB static OPUS_INLINE opus_int32 silk_SMULWB_armv5e(opus_int32 a, opus_int16 b) @@ -190,7 +195,7 @@ static OPUS_INLINE opus_int32 silk_CLZ16_armv5(opus_int16 in16) "#silk_CLZ16\n\t" "clz %0, %1;\n" : "=r"(res) - : "r"(in16<<16|0x8000) + : "r"(SAFE_SHL(in16,16)|0x8000) ); return res; } @@ -210,4 +215,6 @@ static OPUS_INLINE opus_int32 silk_CLZ32_armv5(opus_int32 in32) } #define silk_CLZ32(in32) (silk_CLZ32_armv5(in32)) +#undef SAFE_SHL + #endif /* SILK_MACROS_ARMv5E_H */ diff --git a/code/opus-1.1.4/silk/biquad_alt.c b/code/opus-1.1.4/silk/biquad_alt.c index d55f5ee9..54566a43 100644 --- a/code/opus-1.1.4/silk/biquad_alt.c +++ b/code/opus-1.1.4/silk/biquad_alt.c @@ -39,14 +39,13 @@ POSSIBILITY OF SUCH DAMAGE. #include "SigProc_FIX.h" /* Second order ARMA filter, alternative implementation */ -void silk_biquad_alt( +void silk_biquad_alt_stride1( const opus_int16 *in, /* I input signal */ const opus_int32 *B_Q28, /* I MA coefficients [3] */ const opus_int32 *A_Q28, /* I AR coefficients [2] */ opus_int32 *S, /* I/O State vector [2] */ opus_int16 *out, /* O output signal */ - const opus_int32 len, /* I signal length (must be even) */ - opus_int stride /* I Operate on interleaved signal if > 1 */ + const opus_int32 len /* I signal length (must be even) */ ) { /* DIRECT FORM II TRANSPOSED (uses 2 element state vector) */ @@ -61,7 +60,7 @@ void silk_biquad_alt( for( k = 0; k < len; k++ ) { /* S[ 0 ], S[ 1 ]: Q12 */ - inval = in[ k * stride ]; + inval = in[ k ]; out32_Q14 = silk_LSHIFT( silk_SMLAWB( S[ 0 ], B_Q28[ 0 ], inval ), 2 ); S[ 0 ] = S[1] + silk_RSHIFT_ROUND( silk_SMULWB( out32_Q14, A0_L_Q28 ), 14 ); @@ -73,6 +72,50 @@ void silk_biquad_alt( S[ 1 ] = silk_SMLAWB( S[ 1 ], B_Q28[ 2 ], inval ); /* Scale back to Q0 and saturate */ - out[ k * stride ] = (opus_int16)silk_SAT16( silk_RSHIFT( out32_Q14 + (1<<14) - 1, 14 ) ); + out[ k ] = (opus_int16)silk_SAT16( silk_RSHIFT( out32_Q14 + (1<<14) - 1, 14 ) ); + } +} + +void silk_biquad_alt_stride2_c( + const opus_int16 *in, /* I input signal */ + const opus_int32 *B_Q28, /* I MA coefficients [3] */ + const opus_int32 *A_Q28, /* I AR coefficients [2] */ + opus_int32 *S, /* I/O State vector [4] */ + opus_int16 *out, /* O output signal */ + const opus_int32 len /* I signal length (must be even) */ +) +{ + /* DIRECT FORM II TRANSPOSED (uses 2 element state vector) */ + opus_int k; + opus_int32 A0_U_Q28, A0_L_Q28, A1_U_Q28, A1_L_Q28, out32_Q14[ 2 ]; + + /* Negate A_Q28 values and split in two parts */ + A0_L_Q28 = ( -A_Q28[ 0 ] ) & 0x00003FFF; /* lower part */ + A0_U_Q28 = silk_RSHIFT( -A_Q28[ 0 ], 14 ); /* upper part */ + A1_L_Q28 = ( -A_Q28[ 1 ] ) & 0x00003FFF; /* lower part */ + A1_U_Q28 = silk_RSHIFT( -A_Q28[ 1 ], 14 ); /* upper part */ + + for( k = 0; k < len; k++ ) { + /* S[ 0 ], S[ 1 ], S[ 2 ], S[ 3 ]: Q12 */ + out32_Q14[ 0 ] = silk_LSHIFT( silk_SMLAWB( S[ 0 ], B_Q28[ 0 ], in[ 2 * k + 0 ] ), 2 ); + out32_Q14[ 1 ] = silk_LSHIFT( silk_SMLAWB( S[ 2 ], B_Q28[ 0 ], in[ 2 * k + 1 ] ), 2 ); + + S[ 0 ] = S[ 1 ] + silk_RSHIFT_ROUND( silk_SMULWB( out32_Q14[ 0 ], A0_L_Q28 ), 14 ); + S[ 2 ] = S[ 3 ] + silk_RSHIFT_ROUND( silk_SMULWB( out32_Q14[ 1 ], A0_L_Q28 ), 14 ); + S[ 0 ] = silk_SMLAWB( S[ 0 ], out32_Q14[ 0 ], A0_U_Q28 ); + S[ 2 ] = silk_SMLAWB( S[ 2 ], out32_Q14[ 1 ], A0_U_Q28 ); + S[ 0 ] = silk_SMLAWB( S[ 0 ], B_Q28[ 1 ], in[ 2 * k + 0 ] ); + S[ 2 ] = silk_SMLAWB( S[ 2 ], B_Q28[ 1 ], in[ 2 * k + 1 ] ); + + S[ 1 ] = silk_RSHIFT_ROUND( silk_SMULWB( out32_Q14[ 0 ], A1_L_Q28 ), 14 ); + S[ 3 ] = silk_RSHIFT_ROUND( silk_SMULWB( out32_Q14[ 1 ], A1_L_Q28 ), 14 ); + S[ 1 ] = silk_SMLAWB( S[ 1 ], out32_Q14[ 0 ], A1_U_Q28 ); + S[ 3 ] = silk_SMLAWB( S[ 3 ], out32_Q14[ 1 ], A1_U_Q28 ); + S[ 1 ] = silk_SMLAWB( S[ 1 ], B_Q28[ 2 ], in[ 2 * k + 0 ] ); + S[ 3 ] = silk_SMLAWB( S[ 3 ], B_Q28[ 2 ], in[ 2 * k + 1 ] ); + + /* Scale back to Q0 and saturate */ + out[ 2 * k + 0 ] = (opus_int16)silk_SAT16( silk_RSHIFT( out32_Q14[ 0 ] + (1<<14) - 1, 14 ) ); + out[ 2 * k + 1 ] = (opus_int16)silk_SAT16( silk_RSHIFT( out32_Q14[ 1 ] + (1<<14) - 1, 14 ) ); } } diff --git a/code/opus-1.1.4/silk/bwexpander.c b/code/opus-1.1.4/silk/bwexpander.c index 2eb44566..afa97907 100644 --- a/code/opus-1.1.4/silk/bwexpander.c +++ b/code/opus-1.1.4/silk/bwexpander.c @@ -45,7 +45,7 @@ void silk_bwexpander( /* Bias in silk_SMULWB can lead to unstable filters */ for( i = 0; i < d - 1; i++ ) { ar[ i ] = (opus_int16)silk_RSHIFT_ROUND( silk_MUL( chirp_Q16, ar[ i ] ), 16 ); - chirp_Q16 += silk_RSHIFT_ROUND( silk_MUL( chirp_Q16, chirp_minus_one_Q16 ), 16 ); + chirp_Q16 += silk_RSHIFT_ROUND( silk_MUL( chirp_Q16, chirp_minus_one_Q16 ), 16 ); } ar[ d - 1 ] = (opus_int16)silk_RSHIFT_ROUND( silk_MUL( chirp_Q16, ar[ d - 1 ] ), 16 ); } diff --git a/code/opus-1.1.4/silk/control.h b/code/opus-1.1.4/silk/control.h index 747e5426..b76ec33c 100644 --- a/code/opus-1.1.4/silk/control.h +++ b/code/opus-1.1.4/silk/control.h @@ -77,6 +77,9 @@ typedef struct { /* I: Flag to enable in-band Forward Error Correction (FEC); 0/1 */ opus_int useInBandFEC; + /* I: Flag to actually code in-band Forward Error Correction (FEC) in the current packet; 0/1 */ + opus_int LBRR_coded; + /* I: Flag to enable discontinuous transmission (DTX); 0/1 */ opus_int useDTX; @@ -110,6 +113,11 @@ typedef struct { /* O: Tells the Opus encoder we're ready to switch */ opus_int switchReady; + /* O: SILK Signal type */ + opus_int signalType; + + /* O: SILK offset (dithering) */ + opus_int offset; } silk_EncControlStruct; /**************************************************************************/ diff --git a/code/opus-1.1.4/silk/control_SNR.c b/code/opus-1.1.4/silk/control_SNR.c index cee87eb0..464c1acf 100644 --- a/code/opus-1.1.4/silk/control_SNR.c +++ b/code/opus-1.1.4/silk/control_SNR.c @@ -64,8 +64,7 @@ opus_int silk_control_SNR( /* Find bitrate interval in table and interpolate */ for( k = 1; k < TARGET_RATE_TAB_SZ; k++ ) { if( TargetRate_bps <= rateTable[ k ] ) { - frac_Q6 = silk_DIV32( silk_LSHIFT( TargetRate_bps - rateTable[ k - 1 ], 6 ), - rateTable[ k ] - rateTable[ k - 1 ] ); + frac_Q6 = silk_DIV32( silk_LSHIFT( TargetRate_bps - rateTable[ k - 1 ], 6 ), rateTable[ k ] - rateTable[ k - 1 ] ); psEncC->SNR_dB_Q7 = silk_LSHIFT( silk_SNR_table_Q1[ k - 1 ], 6 ) + silk_MUL( frac_Q6, silk_SNR_table_Q1[ k ] - silk_SNR_table_Q1[ k - 1 ] ); break; } diff --git a/code/opus-1.1.4/silk/control_codec.c b/code/opus-1.1.4/silk/control_codec.c index 044eea3f..9350fd4b 100644 --- a/code/opus-1.1.4/silk/control_codec.c +++ b/code/opus-1.1.4/silk/control_codec.c @@ -57,7 +57,7 @@ static opus_int silk_setup_complexity( static OPUS_INLINE opus_int silk_setup_LBRR( silk_encoder_state *psEncC, /* I/O */ - const opus_int32 TargetRate_bps /* I */ + const silk_EncControlStruct *encControl /* I */ ); @@ -65,7 +65,6 @@ static OPUS_INLINE opus_int silk_setup_LBRR( opus_int silk_control_encoder( silk_encoder_state_Fxx *psEnc, /* I/O Pointer to Silk encoder state */ silk_EncControlStruct *encControl, /* I Control structure */ - const opus_int32 TargetRate_bps, /* I Target max bitrate (bps) */ const opus_int allow_bw_switch, /* I Flag to allow switching audio bandwidth */ const opus_int channelNb, /* I Channel number */ const opus_int force_fs_kHz @@ -125,7 +124,7 @@ opus_int silk_control_encoder( /********************************************/ /* Set LBRR usage */ /********************************************/ - ret += silk_setup_LBRR( &psEnc->sCmn, TargetRate_bps ); + ret += silk_setup_LBRR( &psEnc->sCmn, encControl ); psEnc->sCmn.controlled_since_last_payload = 1; @@ -244,7 +243,6 @@ static opus_int silk_setup_fs( if( psEnc->sCmn.fs_kHz != fs_kHz ) { /* reset part of the state */ silk_memset( &psEnc->sShape, 0, sizeof( psEnc->sShape ) ); - silk_memset( &psEnc->sPrefilt, 0, sizeof( psEnc->sPrefilt ) ); silk_memset( &psEnc->sCmn.sNSQ, 0, sizeof( psEnc->sCmn.sNSQ ) ); silk_memset( psEnc->sCmn.prev_NLSFq_Q15, 0, sizeof( psEnc->sCmn.prev_NLSFq_Q15 ) ); silk_memset( &psEnc->sCmn.sLP.In_LP_State, 0, sizeof( psEnc->sCmn.sLP.In_LP_State ) ); @@ -255,7 +253,6 @@ static opus_int silk_setup_fs( /* Initialize non-zero parameters */ psEnc->sCmn.prevLag = 100; psEnc->sCmn.first_frame_after_reset = 1; - psEnc->sPrefilt.lagPrev = 100; psEnc->sShape.LastGainIndex = 10; psEnc->sCmn.sNSQ.lagPrev = 100; psEnc->sCmn.sNSQ.prev_gain_Q16 = 65536; @@ -293,13 +290,10 @@ static opus_int silk_setup_fs( psEnc->sCmn.pitch_LPC_win_length = silk_SMULBB( FIND_PITCH_LPC_WIN_MS_2_SF, fs_kHz ); } if( psEnc->sCmn.fs_kHz == 16 ) { - psEnc->sCmn.mu_LTP_Q9 = SILK_FIX_CONST( MU_LTP_QUANT_WB, 9 ); psEnc->sCmn.pitch_lag_low_bits_iCDF = silk_uniform8_iCDF; } else if( psEnc->sCmn.fs_kHz == 12 ) { - psEnc->sCmn.mu_LTP_Q9 = SILK_FIX_CONST( MU_LTP_QUANT_MB, 9 ); psEnc->sCmn.pitch_lag_low_bits_iCDF = silk_uniform6_iCDF; } else { - psEnc->sCmn.mu_LTP_Q9 = SILK_FIX_CONST( MU_LTP_QUANT_NB, 9 ); psEnc->sCmn.pitch_lag_low_bits_iCDF = silk_uniform4_iCDF; } } @@ -319,60 +313,75 @@ static opus_int silk_setup_complexity( /* Set encoding complexity */ silk_assert( Complexity >= 0 && Complexity <= 10 ); - if( Complexity < 2 ) { + if( Complexity < 1 ) { psEncC->pitchEstimationComplexity = SILK_PE_MIN_COMPLEX; psEncC->pitchEstimationThreshold_Q16 = SILK_FIX_CONST( 0.8, 16 ); psEncC->pitchEstimationLPCOrder = 6; - psEncC->shapingLPCOrder = 8; + psEncC->shapingLPCOrder = 12; psEncC->la_shape = 3 * psEncC->fs_kHz; psEncC->nStatesDelayedDecision = 1; psEncC->useInterpolatedNLSFs = 0; - psEncC->LTPQuantLowComplexity = 1; + psEncC->NLSF_MSVQ_Survivors = 2; + psEncC->warping_Q16 = 0; + } else if( Complexity < 2 ) { + psEncC->pitchEstimationComplexity = SILK_PE_MID_COMPLEX; + psEncC->pitchEstimationThreshold_Q16 = SILK_FIX_CONST( 0.76, 16 ); + psEncC->pitchEstimationLPCOrder = 8; + psEncC->shapingLPCOrder = 14; + psEncC->la_shape = 5 * psEncC->fs_kHz; + psEncC->nStatesDelayedDecision = 1; + psEncC->useInterpolatedNLSFs = 0; + psEncC->NLSF_MSVQ_Survivors = 3; + psEncC->warping_Q16 = 0; + } else if( Complexity < 3 ) { + psEncC->pitchEstimationComplexity = SILK_PE_MIN_COMPLEX; + psEncC->pitchEstimationThreshold_Q16 = SILK_FIX_CONST( 0.8, 16 ); + psEncC->pitchEstimationLPCOrder = 6; + psEncC->shapingLPCOrder = 12; + psEncC->la_shape = 3 * psEncC->fs_kHz; + psEncC->nStatesDelayedDecision = 2; + psEncC->useInterpolatedNLSFs = 0; psEncC->NLSF_MSVQ_Survivors = 2; psEncC->warping_Q16 = 0; } else if( Complexity < 4 ) { psEncC->pitchEstimationComplexity = SILK_PE_MID_COMPLEX; psEncC->pitchEstimationThreshold_Q16 = SILK_FIX_CONST( 0.76, 16 ); psEncC->pitchEstimationLPCOrder = 8; - psEncC->shapingLPCOrder = 10; + psEncC->shapingLPCOrder = 14; psEncC->la_shape = 5 * psEncC->fs_kHz; - psEncC->nStatesDelayedDecision = 1; + psEncC->nStatesDelayedDecision = 2; psEncC->useInterpolatedNLSFs = 0; - psEncC->LTPQuantLowComplexity = 0; psEncC->NLSF_MSVQ_Survivors = 4; psEncC->warping_Q16 = 0; } else if( Complexity < 6 ) { psEncC->pitchEstimationComplexity = SILK_PE_MID_COMPLEX; psEncC->pitchEstimationThreshold_Q16 = SILK_FIX_CONST( 0.74, 16 ); psEncC->pitchEstimationLPCOrder = 10; - psEncC->shapingLPCOrder = 12; + psEncC->shapingLPCOrder = 16; psEncC->la_shape = 5 * psEncC->fs_kHz; psEncC->nStatesDelayedDecision = 2; psEncC->useInterpolatedNLSFs = 1; - psEncC->LTPQuantLowComplexity = 0; - psEncC->NLSF_MSVQ_Survivors = 8; + psEncC->NLSF_MSVQ_Survivors = 6; psEncC->warping_Q16 = psEncC->fs_kHz * SILK_FIX_CONST( WARPING_MULTIPLIER, 16 ); } else if( Complexity < 8 ) { psEncC->pitchEstimationComplexity = SILK_PE_MID_COMPLEX; psEncC->pitchEstimationThreshold_Q16 = SILK_FIX_CONST( 0.72, 16 ); psEncC->pitchEstimationLPCOrder = 12; - psEncC->shapingLPCOrder = 14; + psEncC->shapingLPCOrder = 20; psEncC->la_shape = 5 * psEncC->fs_kHz; psEncC->nStatesDelayedDecision = 3; psEncC->useInterpolatedNLSFs = 1; - psEncC->LTPQuantLowComplexity = 0; - psEncC->NLSF_MSVQ_Survivors = 16; + psEncC->NLSF_MSVQ_Survivors = 8; psEncC->warping_Q16 = psEncC->fs_kHz * SILK_FIX_CONST( WARPING_MULTIPLIER, 16 ); } else { psEncC->pitchEstimationComplexity = SILK_PE_MAX_COMPLEX; psEncC->pitchEstimationThreshold_Q16 = SILK_FIX_CONST( 0.7, 16 ); psEncC->pitchEstimationLPCOrder = 16; - psEncC->shapingLPCOrder = 16; + psEncC->shapingLPCOrder = 24; psEncC->la_shape = 5 * psEncC->fs_kHz; psEncC->nStatesDelayedDecision = MAX_DEL_DEC_STATES; psEncC->useInterpolatedNLSFs = 1; - psEncC->LTPQuantLowComplexity = 0; - psEncC->NLSF_MSVQ_Survivors = 32; + psEncC->NLSF_MSVQ_Survivors = 16; psEncC->warping_Q16 = psEncC->fs_kHz * SILK_FIX_CONST( WARPING_MULTIPLIER, 16 ); } @@ -387,40 +396,26 @@ static opus_int silk_setup_complexity( silk_assert( psEncC->warping_Q16 <= 32767 ); silk_assert( psEncC->la_shape <= LA_SHAPE_MAX ); silk_assert( psEncC->shapeWinLength <= SHAPE_LPC_WIN_MAX ); - silk_assert( psEncC->NLSF_MSVQ_Survivors <= NLSF_VQ_MAX_SURVIVORS ); return ret; } static OPUS_INLINE opus_int silk_setup_LBRR( silk_encoder_state *psEncC, /* I/O */ - const opus_int32 TargetRate_bps /* I */ + const silk_EncControlStruct *encControl /* I */ ) { opus_int LBRR_in_previous_packet, ret = SILK_NO_ERROR; - opus_int32 LBRR_rate_thres_bps; LBRR_in_previous_packet = psEncC->LBRR_enabled; - psEncC->LBRR_enabled = 0; - if( psEncC->useInBandFEC && psEncC->PacketLoss_perc > 0 ) { - if( psEncC->fs_kHz == 8 ) { - LBRR_rate_thres_bps = LBRR_NB_MIN_RATE_BPS; - } else if( psEncC->fs_kHz == 12 ) { - LBRR_rate_thres_bps = LBRR_MB_MIN_RATE_BPS; + psEncC->LBRR_enabled = encControl->LBRR_coded; + if( psEncC->LBRR_enabled ) { + /* Set gain increase for coding LBRR excitation */ + if( LBRR_in_previous_packet == 0 ) { + /* Previous packet did not have LBRR, and was therefore coded at a higher bitrate */ + psEncC->LBRR_GainIncreases = 7; } else { - LBRR_rate_thres_bps = LBRR_WB_MIN_RATE_BPS; - } - LBRR_rate_thres_bps = silk_SMULWB( silk_MUL( LBRR_rate_thres_bps, 125 - silk_min( psEncC->PacketLoss_perc, 25 ) ), SILK_FIX_CONST( 0.01, 16 ) ); - - if( TargetRate_bps > LBRR_rate_thres_bps ) { - /* Set gain increase for coding LBRR excitation */ - if( LBRR_in_previous_packet == 0 ) { - /* Previous packet did not have LBRR, and was therefore coded at a higher bitrate */ - psEncC->LBRR_GainIncreases = 7; - } else { - psEncC->LBRR_GainIncreases = silk_max_int( 7 - silk_SMULWB( (opus_int32)psEncC->PacketLoss_perc, SILK_FIX_CONST( 0.4, 16 ) ), 2 ); - } - psEncC->LBRR_enabled = 1; + psEncC->LBRR_GainIncreases = silk_max_int( 7 - silk_SMULWB( (opus_int32)psEncC->PacketLoss_perc, SILK_FIX_CONST( 0.4, 16 ) ), 2 ); } } diff --git a/code/opus-1.1.4/silk/debug.h b/code/opus-1.1.4/silk/debug.h index efb6d3e9..6f68c1ca 100644 --- a/code/opus-1.1.4/silk/debug.h +++ b/code/opus-1.1.4/silk/debug.h @@ -39,23 +39,10 @@ extern "C" unsigned long GetHighResolutionTime(void); /* O time in usec*/ -/* make SILK_DEBUG dependent on compiler's _DEBUG */ -#if defined _WIN32 - #ifdef _DEBUG - #define SILK_DEBUG 1 - #else - #define SILK_DEBUG 0 - #endif - - /* overrule the above */ - #if 0 - /* #define NO_ASSERTS*/ - #undef SILK_DEBUG - #define SILK_DEBUG 1 - #endif -#else - #define SILK_DEBUG 0 -#endif +/* Set to 1 to enable DEBUG_STORE_DATA() macros for dumping + * intermediate signals from the codec. + */ +#define SILK_DEBUG 0 /* Flag for using timers */ #define SILK_TIC_TOC 0 diff --git a/code/opus-1.1.4/silk/decode_core.c b/code/opus-1.1.4/silk/decode_core.c index e569c0e7..fd4e857a 100644 --- a/code/opus-1.1.4/silk/decode_core.c +++ b/code/opus-1.1.4/silk/decode_core.c @@ -225,8 +225,6 @@ void silk_decode_core( pxq[ i ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( sLPC_Q14[ MAX_LPC_ORDER + i ], Gain_Q10 ), 8 ) ); } - /* DEBUG_STORE_DATA( dec.pcm, pxq, psDec->subfr_length * sizeof( opus_int16 ) ) */ - /* Update LPC filter state */ silk_memcpy( sLPC_Q14, &sLPC_Q14[ psDec->subfr_length ], MAX_LPC_ORDER * sizeof( opus_int32 ) ); pexc_Q14 += psDec->subfr_length; diff --git a/code/opus-1.1.4/silk/decode_frame.c b/code/opus-1.1.4/silk/decode_frame.c index a605d95a..dfa73c4a 100644 --- a/code/opus-1.1.4/silk/decode_frame.c +++ b/code/opus-1.1.4/silk/decode_frame.c @@ -97,6 +97,7 @@ opus_int silk_decode_frame( psDec->first_frame_after_reset = 0; } else { /* Handle packet loss by extrapolation */ + psDec->indices.signalType = psDec->prevSignalType; silk_PLC( psDec, psDecCtrl, pOut, 1, arch ); } diff --git a/code/opus-1.1.4/silk/decode_parameters.c b/code/opus-1.1.4/silk/decode_parameters.c index e345b1dc..a56a4098 100644 --- a/code/opus-1.1.4/silk/decode_parameters.c +++ b/code/opus-1.1.4/silk/decode_parameters.c @@ -52,7 +52,7 @@ void silk_decode_parameters( silk_NLSF_decode( pNLSF_Q15, psDec->indices.NLSFIndices, psDec->psNLSF_CB ); /* Convert NLSF parameters to AR prediction filter coefficients */ - silk_NLSF2A( psDecCtrl->PredCoef_Q12[ 1 ], pNLSF_Q15, psDec->LPC_order ); + silk_NLSF2A( psDecCtrl->PredCoef_Q12[ 1 ], pNLSF_Q15, psDec->LPC_order, psDec->arch ); /* If just reset, e.g., because internal Fs changed, do not allow interpolation */ /* improves the case of packet loss in the first frame after a switch */ @@ -69,7 +69,7 @@ void silk_decode_parameters( } /* Convert NLSF parameters to AR prediction filter coefficients */ - silk_NLSF2A( psDecCtrl->PredCoef_Q12[ 0 ], pNLSF0_Q15, psDec->LPC_order ); + silk_NLSF2A( psDecCtrl->PredCoef_Q12[ 0 ], pNLSF0_Q15, psDec->LPC_order, psDec->arch ); } else { /* Copy LPC coefficients for first half from second half */ silk_memcpy( psDecCtrl->PredCoef_Q12[ 0 ], psDecCtrl->PredCoef_Q12[ 1 ], psDec->LPC_order * sizeof( opus_int16 ) ); diff --git a/code/opus-1.1.4/silk/define.h b/code/opus-1.1.4/silk/define.h index 19c9b00e..1286048e 100644 --- a/code/opus-1.1.4/silk/define.h +++ b/code/opus-1.1.4/silk/define.h @@ -56,6 +56,7 @@ extern "C" /* DTX settings */ #define NB_SPEECH_FRAMES_BEFORE_DTX 10 /* eq 200 ms */ #define MAX_CONSECUTIVE_DTX 20 /* eq 400 ms */ +#define DTX_ACTIVITY_THRESHOLD 0.1f /* Maximum sampling frequency */ #define MAX_FS_KHZ 16 @@ -147,7 +148,7 @@ extern "C" #define USE_HARM_SHAPING 1 /* Max LPC order of noise shaping filters */ -#define MAX_SHAPE_LPC_ORDER 16 +#define MAX_SHAPE_LPC_ORDER 24 #define HARM_SHAPE_FIR_TAPS 3 @@ -157,8 +158,7 @@ extern "C" #define LTP_BUF_LENGTH 512 #define LTP_MASK ( LTP_BUF_LENGTH - 1 ) -#define DECISION_DELAY 32 -#define DECISION_DELAY_MASK ( DECISION_DELAY - 1 ) +#define DECISION_DELAY 40 /* Number of subframes for excitation entropy coding */ #define SHELL_CODEC_FRAME_LENGTH 16 @@ -173,11 +173,7 @@ extern "C" #define MAX_MATRIX_SIZE MAX_LPC_ORDER /* Max of LPC Order and LTP order */ -#if( MAX_LPC_ORDER > DECISION_DELAY ) # define NSQ_LPC_BUF_LENGTH MAX_LPC_ORDER -#else -# define NSQ_LPC_BUF_LENGTH DECISION_DELAY -#endif /***************************/ /* Voice activity detector */ @@ -205,7 +201,6 @@ extern "C" /******************/ #define NLSF_W_Q 2 #define NLSF_VQ_MAX_VECTORS 32 -#define NLSF_VQ_MAX_SURVIVORS 32 #define NLSF_QUANT_MAX_AMPLITUDE 4 #define NLSF_QUANT_MAX_AMPLITUDE_EXT 10 #define NLSF_QUANT_LEVEL_ADJ 0.1 diff --git a/code/opus-1.1.4/silk/enc_API.c b/code/opus-1.1.4/silk/enc_API.c index f8060286..701c2905 100644 --- a/code/opus-1.1.4/silk/enc_API.c +++ b/code/opus-1.1.4/silk/enc_API.c @@ -233,11 +233,10 @@ opus_int silk_Encode( /* O Returns error co } } - TargetRate_bps = silk_RSHIFT32( encControl->bitRate, encControl->nChannelsInternal - 1 ); for( n = 0; n < encControl->nChannelsInternal; n++ ) { /* Force the side channel to the same rate as the mid */ opus_int force_fs_kHz = (n==1) ? psEnc->state_Fxx[0].sCmn.fs_kHz : 0; - if( ( ret = silk_control_encoder( &psEnc->state_Fxx[ n ], encControl, TargetRate_bps, psEnc->allowBandwidthSwitch, n, force_fs_kHz ) ) != 0 ) { + if( ( ret = silk_control_encoder( &psEnc->state_Fxx[ n ], encControl, psEnc->allowBandwidthSwitch, n, force_fs_kHz ) ) != 0 ) { silk_assert( 0 ); RESTORE_STACK; return ret; @@ -416,7 +415,6 @@ opus_int silk_Encode( /* O Returns error co /* Reset side channel encoder memory for first frame with side coding */ if( psEnc->prev_decode_only_middle == 1 ) { silk_memset( &psEnc->state_Fxx[ 1 ].sShape, 0, sizeof( psEnc->state_Fxx[ 1 ].sShape ) ); - silk_memset( &psEnc->state_Fxx[ 1 ].sPrefilt, 0, sizeof( psEnc->state_Fxx[ 1 ].sPrefilt ) ); silk_memset( &psEnc->state_Fxx[ 1 ].sCmn.sNSQ, 0, sizeof( psEnc->state_Fxx[ 1 ].sCmn.sNSQ ) ); silk_memset( psEnc->state_Fxx[ 1 ].sCmn.prev_NLSFq_Q15, 0, sizeof( psEnc->state_Fxx[ 1 ].sCmn.prev_NLSFq_Q15 ) ); silk_memset( &psEnc->state_Fxx[ 1 ].sCmn.sLP.In_LP_State, 0, sizeof( psEnc->state_Fxx[ 1 ].sCmn.sLP.In_LP_State ) ); @@ -557,6 +555,10 @@ opus_int silk_Encode( /* O Returns error co } } + encControl->signalType = psEnc->state_Fxx[0].sCmn.indices.signalType; + encControl->offset = silk_Quantization_Offsets_Q10 + [ psEnc->state_Fxx[0].sCmn.indices.signalType >> 1 ] + [ psEnc->state_Fxx[0].sCmn.indices.quantOffsetType ]; RESTORE_STACK; return ret; } diff --git a/code/opus-1.1.4/silk/fixed/LTP_analysis_filter_FIX.c b/code/opus-1.1.4/silk/fixed/LTP_analysis_filter_FIX.c index a9419080..5574e706 100644 --- a/code/opus-1.1.4/silk/fixed/LTP_analysis_filter_FIX.c +++ b/code/opus-1.1.4/silk/fixed/LTP_analysis_filter_FIX.c @@ -45,7 +45,7 @@ void silk_LTP_analysis_filter_FIX( const opus_int16 *x_ptr, *x_lag_ptr; opus_int16 Btmp_Q14[ LTP_ORDER ]; opus_int16 *LTP_res_ptr; - opus_int k, i, j; + opus_int k, i; opus_int32 LTP_est; x_ptr = x; @@ -53,9 +53,12 @@ void silk_LTP_analysis_filter_FIX( for( k = 0; k < nb_subfr; k++ ) { x_lag_ptr = x_ptr - pitchL[ k ]; - for( i = 0; i < LTP_ORDER; i++ ) { - Btmp_Q14[ i ] = LTPCoef_Q14[ k * LTP_ORDER + i ]; - } + + Btmp_Q14[ 0 ] = LTPCoef_Q14[ k * LTP_ORDER ]; + Btmp_Q14[ 1 ] = LTPCoef_Q14[ k * LTP_ORDER + 1 ]; + Btmp_Q14[ 2 ] = LTPCoef_Q14[ k * LTP_ORDER + 2 ]; + Btmp_Q14[ 3 ] = LTPCoef_Q14[ k * LTP_ORDER + 3 ]; + Btmp_Q14[ 4 ] = LTPCoef_Q14[ k * LTP_ORDER + 4 ]; /* LTP analysis FIR filter */ for( i = 0; i < subfr_length + pre_length; i++ ) { @@ -63,9 +66,11 @@ void silk_LTP_analysis_filter_FIX( /* Long-term prediction */ LTP_est = silk_SMULBB( x_lag_ptr[ LTP_ORDER / 2 ], Btmp_Q14[ 0 ] ); - for( j = 1; j < LTP_ORDER; j++ ) { - LTP_est = silk_SMLABB_ovflw( LTP_est, x_lag_ptr[ LTP_ORDER / 2 - j ], Btmp_Q14[ j ] ); - } + LTP_est = silk_SMLABB_ovflw( LTP_est, x_lag_ptr[ 1 ], Btmp_Q14[ 1 ] ); + LTP_est = silk_SMLABB_ovflw( LTP_est, x_lag_ptr[ 0 ], Btmp_Q14[ 2 ] ); + LTP_est = silk_SMLABB_ovflw( LTP_est, x_lag_ptr[ -1 ], Btmp_Q14[ 3 ] ); + LTP_est = silk_SMLABB_ovflw( LTP_est, x_lag_ptr[ -2 ], Btmp_Q14[ 4 ] ); + LTP_est = silk_RSHIFT_ROUND( LTP_est, 14 ); /* round and -> Q0*/ /* Subtract long-term prediction */ diff --git a/code/opus-1.1.4/silk/fixed/burg_modified_FIX.c b/code/opus-1.1.4/silk/fixed/burg_modified_FIX.c index db348295..80edbdfa 100644 --- a/code/opus-1.1.4/silk/fixed/burg_modified_FIX.c +++ b/code/opus-1.1.4/silk/fixed/burg_modified_FIX.c @@ -37,12 +37,12 @@ POSSIBILITY OF SUCH DAMAGE. #define MAX_FRAME_SIZE 384 /* subfr_length * nb_subfr = ( 0.005 * 16000 + 16 ) * 4 = 384 */ #define QA 25 -#define N_BITS_HEAD_ROOM 2 +#define N_BITS_HEAD_ROOM 3 #define MIN_RSHIFTS -16 #define MAX_RSHIFTS (32 - QA) /* Compute reflection coefficients from input signal */ -void silk_burg_modified( +void silk_burg_modified_c( opus_int32 *res_nrg, /* O Residual energy */ opus_int *res_nrg_Q, /* O Residual energy Q value */ opus_int32 A_Q16[], /* O Prediction coefficients (length order) */ @@ -54,7 +54,7 @@ void silk_burg_modified( int arch /* I Run-time architecture */ ) { - opus_int k, n, s, lz, rshifts, rshifts_extra, reached_max_gain; + opus_int k, n, s, lz, rshifts, reached_max_gain; opus_int32 C0, num, nrg, rc_Q31, invGain_Q30, Atmp_QA, Atmp1, tmp1, tmp2, x1, x2; const opus_int16 *x_ptr; opus_int32 C_first_row[ SILK_MAX_ORDER_LPC ]; @@ -63,27 +63,23 @@ void silk_burg_modified( opus_int32 CAf[ SILK_MAX_ORDER_LPC + 1 ]; opus_int32 CAb[ SILK_MAX_ORDER_LPC + 1 ]; opus_int32 xcorr[ SILK_MAX_ORDER_LPC ]; + opus_int64 C0_64; silk_assert( subfr_length * nb_subfr <= MAX_FRAME_SIZE ); /* Compute autocorrelations, added over subframes */ - silk_sum_sqr_shift( &C0, &rshifts, x, nb_subfr * subfr_length ); - if( rshifts > MAX_RSHIFTS ) { - C0 = silk_LSHIFT32( C0, rshifts - MAX_RSHIFTS ); - silk_assert( C0 > 0 ); - rshifts = MAX_RSHIFTS; + C0_64 = silk_inner_prod16_aligned_64( x, x, subfr_length*nb_subfr, arch ); + lz = silk_CLZ64(C0_64); + rshifts = 32 + 1 + N_BITS_HEAD_ROOM - lz; + if (rshifts > MAX_RSHIFTS) rshifts = MAX_RSHIFTS; + if (rshifts < MIN_RSHIFTS) rshifts = MIN_RSHIFTS; + + if (rshifts > 0) { + C0 = (opus_int32)silk_RSHIFT64(C0_64, rshifts ); } else { - lz = silk_CLZ32( C0 ) - 1; - rshifts_extra = N_BITS_HEAD_ROOM - lz; - if( rshifts_extra > 0 ) { - rshifts_extra = silk_min( rshifts_extra, MAX_RSHIFTS - rshifts ); - C0 = silk_RSHIFT32( C0, rshifts_extra ); - } else { - rshifts_extra = silk_max( rshifts_extra, MIN_RSHIFTS - rshifts ); - C0 = silk_LSHIFT32( C0, -rshifts_extra ); - } - rshifts += rshifts_extra; + C0 = silk_LSHIFT32((opus_int32)C0_64, -rshifts ); } + CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL( SILK_FIX_CONST( FIND_LPC_COND_FAC, 32 ), C0 ) + 1; /* Q(-rshifts) */ silk_memset( C_first_row, 0, SILK_MAX_ORDER_LPC * sizeof( opus_int32 ) ); if( rshifts > 0 ) { @@ -91,7 +87,7 @@ void silk_burg_modified( x_ptr = x + s * subfr_length; for( n = 1; n < D + 1; n++ ) { C_first_row[ n - 1 ] += (opus_int32)silk_RSHIFT64( - silk_inner_prod16_aligned_64( x_ptr, x_ptr + n, subfr_length - n ), rshifts ); + silk_inner_prod16_aligned_64( x_ptr, x_ptr + n, subfr_length - n, arch ), rshifts ); } } } else { @@ -154,8 +150,11 @@ void silk_burg_modified( C_first_row[ k ] = silk_MLA( C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); /* Q( -rshifts ) */ C_last_row[ k ] = silk_MLA( C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ] ); /* Q( -rshifts ) */ Atmp1 = silk_RSHIFT_ROUND( Af_QA[ k ], QA - 17 ); /* Q17 */ - tmp1 = silk_MLA( tmp1, x_ptr[ n - k - 1 ], Atmp1 ); /* Q17 */ - tmp2 = silk_MLA( tmp2, x_ptr[ subfr_length - n + k ], Atmp1 ); /* Q17 */ + /* We sometimes have get overflows in the multiplications (even beyond +/- 2^32), + but they cancel each other and the real result seems to always fit in a 32-bit + signed integer. This was determined experimentally, not theoretically (unfortunately). */ + tmp1 = silk_MLA_ovflw( tmp1, x_ptr[ n - k - 1 ], Atmp1 ); /* Q17 */ + tmp2 = silk_MLA_ovflw( tmp2, x_ptr[ subfr_length - n + k ], Atmp1 ); /* Q17 */ } tmp1 = -tmp1; /* Q17 */ tmp2 = -tmp2; /* Q17 */ @@ -204,12 +203,14 @@ void silk_burg_modified( /* Max prediction gain exceeded; set reflection coefficient such that max prediction gain is exactly hit */ tmp2 = ( (opus_int32)1 << 30 ) - silk_DIV32_varQ( minInvGain_Q30, invGain_Q30, 30 ); /* Q30 */ rc_Q31 = silk_SQRT_APPROX( tmp2 ); /* Q15 */ - /* Newton-Raphson iteration */ - rc_Q31 = silk_RSHIFT32( rc_Q31 + silk_DIV32( tmp2, rc_Q31 ), 1 ); /* Q15 */ - rc_Q31 = silk_LSHIFT32( rc_Q31, 16 ); /* Q31 */ - if( num < 0 ) { - /* Ensure adjusted reflection coefficients has the original sign */ - rc_Q31 = -rc_Q31; + if( rc_Q31 > 0 ) { + /* Newton-Raphson iteration */ + rc_Q31 = silk_RSHIFT32( rc_Q31 + silk_DIV32( tmp2, rc_Q31 ), 1 ); /* Q15 */ + rc_Q31 = silk_LSHIFT32( rc_Q31, 16 ); /* Q31 */ + if( num < 0 ) { + /* Ensure adjusted reflection coefficients has the original sign */ + rc_Q31 = -rc_Q31; + } } invGain_Q30 = minInvGain_Q30; reached_max_gain = 1; @@ -252,12 +253,12 @@ void silk_burg_modified( if( rshifts > 0 ) { for( s = 0; s < nb_subfr; s++ ) { x_ptr = x + s * subfr_length; - C0 -= (opus_int32)silk_RSHIFT64( silk_inner_prod16_aligned_64( x_ptr, x_ptr, D ), rshifts ); + C0 -= (opus_int32)silk_RSHIFT64( silk_inner_prod16_aligned_64( x_ptr, x_ptr, D, arch ), rshifts ); } } else { for( s = 0; s < nb_subfr; s++ ) { x_ptr = x + s * subfr_length; - C0 -= silk_LSHIFT32( silk_inner_prod_aligned( x_ptr, x_ptr, D ), -rshifts ); + C0 -= silk_LSHIFT32( silk_inner_prod_aligned( x_ptr, x_ptr, D, arch), -rshifts); } } /* Approximate residual energy */ diff --git a/code/opus-1.1.4/silk/fixed/corrMatrix_FIX.c b/code/opus-1.1.4/silk/fixed/corrMatrix_FIX.c index c6172705..1b4a29c2 100644 --- a/code/opus-1.1.4/silk/fixed/corrMatrix_FIX.c +++ b/code/opus-1.1.4/silk/fixed/corrMatrix_FIX.c @@ -42,7 +42,8 @@ void silk_corrVector_FIX( const opus_int L, /* I Length of vectors */ const opus_int order, /* I Max lag for correlation */ opus_int32 *Xt, /* O Pointer to X'*t correlation vector [order] */ - const opus_int rshifts /* I Right shifts of correlations */ + const opus_int rshifts, /* I Right shifts of correlations */ + int arch /* I Run-time architecture */ ) { opus_int lag, i; @@ -57,7 +58,7 @@ void silk_corrVector_FIX( for( lag = 0; lag < order; lag++ ) { inner_prod = 0; for( i = 0; i < L; i++ ) { - inner_prod += silk_RSHIFT32( silk_SMULBB( ptr1[ i ], ptr2[i] ), rshifts ); + inner_prod = silk_ADD_RSHIFT32( inner_prod, silk_SMULBB( ptr1[ i ], ptr2[i] ), rshifts ); } Xt[ lag ] = inner_prod; /* X[:,lag]'*t */ ptr1--; /* Go to next column of X */ @@ -65,7 +66,7 @@ void silk_corrVector_FIX( } else { silk_assert( rshifts == 0 ); for( lag = 0; lag < order; lag++ ) { - Xt[ lag ] = silk_inner_prod_aligned( ptr1, ptr2, L ); /* X[:,lag]'*t */ + Xt[ lag ] = silk_inner_prod_aligned( ptr1, ptr2, L, arch ); /* X[:,lag]'*t */ ptr1--; /* Go to next column of X */ } } @@ -76,60 +77,54 @@ void silk_corrMatrix_FIX( const opus_int16 *x, /* I x vector [L + order - 1] used to form data matrix X */ const opus_int L, /* I Length of vectors */ const opus_int order, /* I Max lag for correlation */ - const opus_int head_room, /* I Desired headroom */ opus_int32 *XX, /* O Pointer to X'*X correlation matrix [ order x order ] */ - opus_int *rshifts /* I/O Right shifts of correlations */ + opus_int32 *nrg, /* O Energy of x vector */ + opus_int *rshifts, /* O Right shifts of correlations and energy */ + int arch /* I Run-time architecture */ ) { - opus_int i, j, lag, rshifts_local, head_room_rshifts; + opus_int i, j, lag; opus_int32 energy; const opus_int16 *ptr1, *ptr2; /* Calculate energy to find shift used to fit in 32 bits */ - silk_sum_sqr_shift( &energy, &rshifts_local, x, L + order - 1 ); - /* Add shifts to get the desired head room */ - head_room_rshifts = silk_max( head_room - silk_CLZ32( energy ), 0 ); - - energy = silk_RSHIFT32( energy, head_room_rshifts ); - rshifts_local += head_room_rshifts; + silk_sum_sqr_shift( nrg, rshifts, x, L + order - 1 ); + energy = *nrg; /* Calculate energy of first column (0) of X: X[:,0]'*X[:,0] */ /* Remove contribution of first order - 1 samples */ for( i = 0; i < order - 1; i++ ) { - energy -= silk_RSHIFT32( silk_SMULBB( x[ i ], x[ i ] ), rshifts_local ); - } - if( rshifts_local < *rshifts ) { - /* Adjust energy */ - energy = silk_RSHIFT32( energy, *rshifts - rshifts_local ); - rshifts_local = *rshifts; + energy -= silk_RSHIFT32( silk_SMULBB( x[ i ], x[ i ] ), *rshifts ); } /* Calculate energy of remaining columns of X: X[:,j]'*X[:,j] */ /* Fill out the diagonal of the correlation matrix */ matrix_ptr( XX, 0, 0, order ) = energy; + silk_assert( energy >= 0 ); ptr1 = &x[ order - 1 ]; /* First sample of column 0 of X */ for( j = 1; j < order; j++ ) { - energy = silk_SUB32( energy, silk_RSHIFT32( silk_SMULBB( ptr1[ L - j ], ptr1[ L - j ] ), rshifts_local ) ); - energy = silk_ADD32( energy, silk_RSHIFT32( silk_SMULBB( ptr1[ -j ], ptr1[ -j ] ), rshifts_local ) ); + energy = silk_SUB32( energy, silk_RSHIFT32( silk_SMULBB( ptr1[ L - j ], ptr1[ L - j ] ), *rshifts ) ); + energy = silk_ADD32( energy, silk_RSHIFT32( silk_SMULBB( ptr1[ -j ], ptr1[ -j ] ), *rshifts ) ); matrix_ptr( XX, j, j, order ) = energy; + silk_assert( energy >= 0 ); } ptr2 = &x[ order - 2 ]; /* First sample of column 1 of X */ /* Calculate the remaining elements of the correlation matrix */ - if( rshifts_local > 0 ) { + if( *rshifts > 0 ) { /* Right shifting used */ for( lag = 1; lag < order; lag++ ) { /* Inner product of column 0 and column lag: X[:,0]'*X[:,lag] */ energy = 0; for( i = 0; i < L; i++ ) { - energy += silk_RSHIFT32( silk_SMULBB( ptr1[ i ], ptr2[i] ), rshifts_local ); + energy += silk_RSHIFT32( silk_SMULBB( ptr1[ i ], ptr2[i] ), *rshifts ); } /* Calculate remaining off diagonal: X[:,j]'*X[:,j + lag] */ matrix_ptr( XX, lag, 0, order ) = energy; matrix_ptr( XX, 0, lag, order ) = energy; for( j = 1; j < ( order - lag ); j++ ) { - energy = silk_SUB32( energy, silk_RSHIFT32( silk_SMULBB( ptr1[ L - j ], ptr2[ L - j ] ), rshifts_local ) ); - energy = silk_ADD32( energy, silk_RSHIFT32( silk_SMULBB( ptr1[ -j ], ptr2[ -j ] ), rshifts_local ) ); + energy = silk_SUB32( energy, silk_RSHIFT32( silk_SMULBB( ptr1[ L - j ], ptr2[ L - j ] ), *rshifts ) ); + energy = silk_ADD32( energy, silk_RSHIFT32( silk_SMULBB( ptr1[ -j ], ptr2[ -j ] ), *rshifts ) ); matrix_ptr( XX, lag + j, j, order ) = energy; matrix_ptr( XX, j, lag + j, order ) = energy; } @@ -138,7 +133,7 @@ void silk_corrMatrix_FIX( } else { for( lag = 1; lag < order; lag++ ) { /* Inner product of column 0 and column lag: X[:,0]'*X[:,lag] */ - energy = silk_inner_prod_aligned( ptr1, ptr2, L ); + energy = silk_inner_prod_aligned( ptr1, ptr2, L, arch ); matrix_ptr( XX, lag, 0, order ) = energy; matrix_ptr( XX, 0, lag, order ) = energy; /* Calculate remaining off diagonal: X[:,j]'*X[:,j + lag] */ @@ -151,6 +146,5 @@ void silk_corrMatrix_FIX( ptr2--;/* Update pointer to first sample of next column (lag) in X */ } } - *rshifts = rshifts_local; } diff --git a/code/opus-1.1.4/silk/fixed/encode_frame_FIX.c b/code/opus-1.1.4/silk/fixed/encode_frame_FIX.c index b490986b..4f9e0861 100644 --- a/code/opus-1.1.4/silk/fixed/encode_frame_FIX.c +++ b/code/opus-1.1.4/silk/fixed/encode_frame_FIX.c @@ -29,6 +29,7 @@ POSSIBILITY OF SUCH DAMAGE. #include "config.h" #endif +#include #include "main_FIX.h" #include "stack_alloc.h" #include "tuning_parameters.h" @@ -37,7 +38,7 @@ POSSIBILITY OF SUCH DAMAGE. static OPUS_INLINE void silk_LBRR_encode_FIX( silk_encoder_state_FIX *psEnc, /* I/O Pointer to Silk FIX encoder state */ silk_encoder_control_FIX *psEncCtrl, /* I/O Pointer to Silk FIX encoder control struct */ - const opus_int32 xfw_Q3[], /* I Input signal */ + const opus_int16 x16[], /* I Input signal */ opus_int condCoding /* I The type of conditional coding used so far for this frame */ ); @@ -48,7 +49,7 @@ void silk_encode_do_VAD_FIX( /****************************/ /* Voice Activity Detection */ /****************************/ - silk_VAD_GetSA_Q8( &psEnc->sCmn, psEnc->sCmn.inputBuf + 1 ); + silk_VAD_GetSA_Q8( &psEnc->sCmn, psEnc->sCmn.inputBuf + 1, psEnc->sCmn.arch ); /**************************************************/ /* Convert speech activity into VAD and DTX flags */ @@ -94,6 +95,9 @@ opus_int silk_encode_frame_FIX( opus_int16 ec_prevLagIndex_copy; opus_int ec_prevSignalType_copy; opus_int8 LastGainIndex_copy2; + opus_int gain_lock[ MAX_NB_SUBFR ] = {0}; + opus_int16 best_gain_mult[ MAX_NB_SUBFR ]; + opus_int best_sum[ MAX_NB_SUBFR ]; SAVE_STACK; /* This is totally unnecessary but many compilers (including gcc) are too dumb to realise it */ @@ -118,7 +122,6 @@ opus_int silk_encode_frame_FIX( silk_memcpy( x_frame + LA_SHAPE_MS * psEnc->sCmn.fs_kHz, psEnc->sCmn.inputBuf + 1, psEnc->sCmn.frame_length * sizeof( opus_int16 ) ); if( !psEnc->sCmn.prefillFlag ) { - VARDECL( opus_int32, xfw_Q3 ); VARDECL( opus_int16, res_pitch ); VARDECL( opus_uint8, ec_buf_copy ); opus_int16 *res_pitch_frame; @@ -132,7 +135,7 @@ opus_int silk_encode_frame_FIX( /*****************************************/ /* Find pitch lags, initial LPC analysis */ /*****************************************/ - silk_find_pitch_lags_FIX( psEnc, &sEncCtrl, res_pitch, x_frame, psEnc->sCmn.arch ); + silk_find_pitch_lags_FIX( psEnc, &sEncCtrl, res_pitch, x_frame - psEnc->sCmn.ltp_mem_length, psEnc->sCmn.arch ); /************************/ /* Noise shape analysis */ @@ -142,23 +145,17 @@ opus_int silk_encode_frame_FIX( /***************************************************/ /* Find linear prediction coefficients (LPC + LTP) */ /***************************************************/ - silk_find_pred_coefs_FIX( psEnc, &sEncCtrl, res_pitch, x_frame, condCoding ); + silk_find_pred_coefs_FIX( psEnc, &sEncCtrl, res_pitch_frame, x_frame, condCoding ); /****************************************/ /* Process gains */ /****************************************/ silk_process_gains_FIX( psEnc, &sEncCtrl, condCoding ); - /*****************************************/ - /* Prefiltering for noise shaper */ - /*****************************************/ - ALLOC( xfw_Q3, psEnc->sCmn.frame_length, opus_int32 ); - silk_prefilter_FIX( psEnc, &sEncCtrl, xfw_Q3, x_frame ); - /****************************************/ /* Low Bitrate Redundant Encoding */ /****************************************/ - silk_LBRR_encode_FIX( psEnc, &sEncCtrl, xfw_Q3, condCoding ); + silk_LBRR_encode_FIX( psEnc, &sEncCtrl, x_frame, condCoding ); /* Loop over quantizer and entropy coding to control bitrate */ maxIter = 6; @@ -194,13 +191,19 @@ opus_int silk_encode_frame_FIX( /* Noise shaping quantization */ /*****************************************/ if( psEnc->sCmn.nStatesDelayedDecision > 1 || psEnc->sCmn.warping_Q16 > 0 ) { - silk_NSQ_del_dec( &psEnc->sCmn, &psEnc->sCmn.sNSQ, &psEnc->sCmn.indices, xfw_Q3, psEnc->sCmn.pulses, - sEncCtrl.PredCoef_Q12[ 0 ], sEncCtrl.LTPCoef_Q14, sEncCtrl.AR2_Q13, sEncCtrl.HarmShapeGain_Q14, - sEncCtrl.Tilt_Q14, sEncCtrl.LF_shp_Q14, sEncCtrl.Gains_Q16, sEncCtrl.pitchL, sEncCtrl.Lambda_Q10, sEncCtrl.LTP_scale_Q14 ); + silk_NSQ_del_dec( &psEnc->sCmn, &psEnc->sCmn.sNSQ, &psEnc->sCmn.indices, x_frame, psEnc->sCmn.pulses, + sEncCtrl.PredCoef_Q12[ 0 ], sEncCtrl.LTPCoef_Q14, sEncCtrl.AR_Q13, sEncCtrl.HarmShapeGain_Q14, + sEncCtrl.Tilt_Q14, sEncCtrl.LF_shp_Q14, sEncCtrl.Gains_Q16, sEncCtrl.pitchL, sEncCtrl.Lambda_Q10, sEncCtrl.LTP_scale_Q14, + psEnc->sCmn.arch ); } else { - silk_NSQ( &psEnc->sCmn, &psEnc->sCmn.sNSQ, &psEnc->sCmn.indices, xfw_Q3, psEnc->sCmn.pulses, - sEncCtrl.PredCoef_Q12[ 0 ], sEncCtrl.LTPCoef_Q14, sEncCtrl.AR2_Q13, sEncCtrl.HarmShapeGain_Q14, - sEncCtrl.Tilt_Q14, sEncCtrl.LF_shp_Q14, sEncCtrl.Gains_Q16, sEncCtrl.pitchL, sEncCtrl.Lambda_Q10, sEncCtrl.LTP_scale_Q14 ); + silk_NSQ( &psEnc->sCmn, &psEnc->sCmn.sNSQ, &psEnc->sCmn.indices, x_frame, psEnc->sCmn.pulses, + sEncCtrl.PredCoef_Q12[ 0 ], sEncCtrl.LTPCoef_Q14, sEncCtrl.AR_Q13, sEncCtrl.HarmShapeGain_Q14, + sEncCtrl.Tilt_Q14, sEncCtrl.LF_shp_Q14, sEncCtrl.Gains_Q16, sEncCtrl.pitchL, sEncCtrl.Lambda_Q10, sEncCtrl.LTP_scale_Q14, + psEnc->sCmn.arch); + } + + if ( iter == maxIter && !found_lower ) { + silk_memcpy( &sRangeEnc_copy2, psRangeEnc, sizeof( ec_enc ) ); } /****************************************/ @@ -216,6 +219,33 @@ opus_int silk_encode_frame_FIX( nBits = ec_tell( psRangeEnc ); + /* If we still bust after the last iteration, do some damage control. */ + if ( iter == maxIter && !found_lower && nBits > maxBits ) { + silk_memcpy( psRangeEnc, &sRangeEnc_copy2, sizeof( ec_enc ) ); + + /* Keep gains the same as the last frame. */ + psEnc->sShape.LastGainIndex = sEncCtrl.lastGainIndexPrev; + for ( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) { + psEnc->sCmn.indices.GainsIndices[ i ] = 4; + } + if (condCoding != CODE_CONDITIONALLY) { + psEnc->sCmn.indices.GainsIndices[ 0 ] = sEncCtrl.lastGainIndexPrev; + } + psEnc->sCmn.ec_prevLagIndex = ec_prevLagIndex_copy; + psEnc->sCmn.ec_prevSignalType = ec_prevSignalType_copy; + /* Clear all pulses. */ + for ( i = 0; i < psEnc->sCmn.frame_length; i++ ) { + psEnc->sCmn.pulses[ i ] = 0; + } + + silk_encode_indices( &psEnc->sCmn, psRangeEnc, psEnc->sCmn.nFramesEncoded, 0, condCoding ); + + silk_encode_pulses( psRangeEnc, psEnc->sCmn.indices.signalType, psEnc->sCmn.indices.quantOffsetType, + psEnc->sCmn.pulses, psEnc->sCmn.frame_length ); + + nBits = ec_tell( psRangeEnc ); + } + if( useCBR == 0 && iter == 0 && nBits <= maxBits ) { break; } @@ -263,15 +293,35 @@ opus_int silk_encode_frame_FIX( break; } + if ( !found_lower && nBits > maxBits ) { + int j; + for ( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) { + int sum=0; + for ( j = i*psEnc->sCmn.subfr_length; j < (i+1)*psEnc->sCmn.subfr_length; j++ ) { + sum += abs( psEnc->sCmn.pulses[j] ); + } + if ( iter == 0 || (sum < best_sum[i] && !gain_lock[i]) ) { + best_sum[i] = sum; + best_gain_mult[i] = gainMult_Q8; + } else { + gain_lock[i] = 1; + } + } + } if( ( found_lower & found_upper ) == 0 ) { /* Adjust gain according to high-rate rate/distortion curve */ - opus_int32 gain_factor_Q16; - gain_factor_Q16 = silk_log2lin( silk_LSHIFT( nBits - maxBits, 7 ) / psEnc->sCmn.frame_length + SILK_FIX_CONST( 16, 7 ) ); - gain_factor_Q16 = silk_min_32( gain_factor_Q16, SILK_FIX_CONST( 2, 16 ) ); if( nBits > maxBits ) { - gain_factor_Q16 = silk_max_32( gain_factor_Q16, SILK_FIX_CONST( 1.3, 16 ) ); + if (gainMult_Q8 < 16384) { + gainMult_Q8 *= 2; + } else { + gainMult_Q8 = 32767; + } + } else { + opus_int32 gain_factor_Q16; + gain_factor_Q16 = silk_log2lin( silk_LSHIFT( nBits - maxBits, 7 ) / psEnc->sCmn.frame_length + SILK_FIX_CONST( 16, 7 ) ); + gainMult_Q8 = silk_SMULWB( gain_factor_Q16, gainMult_Q8 ); } - gainMult_Q8 = silk_SMULWB( gain_factor_Q16, gainMult_Q8 ); + } else { /* Adjust gain by interpolating */ gainMult_Q8 = gainMult_lower + silk_DIV32_16( silk_MUL( gainMult_upper - gainMult_lower, maxBits - nBits_lower ), nBits_upper - nBits_lower ); @@ -285,9 +335,15 @@ opus_int silk_encode_frame_FIX( } for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) { - sEncCtrl.Gains_Q16[ i ] = silk_LSHIFT_SAT32( silk_SMULWB( sEncCtrl.GainsUnq_Q16[ i ], gainMult_Q8 ), 8 ); + opus_int16 tmp; + if ( gain_lock[i] ) { + tmp = best_gain_mult[i]; + } else { + tmp = gainMult_Q8; + } + sEncCtrl.Gains_Q16[ i ] = silk_LSHIFT_SAT32( silk_SMULWB( sEncCtrl.GainsUnq_Q16[ i ], tmp ), 8 ); } - + /* Quantize gains */ psEnc->sShape.LastGainIndex = sEncCtrl.lastGainIndexPrev; silk_gains_quant( psEnc->sCmn.indices.GainsIndices, sEncCtrl.Gains_Q16, @@ -329,7 +385,7 @@ opus_int silk_encode_frame_FIX( static OPUS_INLINE void silk_LBRR_encode_FIX( silk_encoder_state_FIX *psEnc, /* I/O Pointer to Silk FIX encoder state */ silk_encoder_control_FIX *psEncCtrl, /* I/O Pointer to Silk FIX encoder control struct */ - const opus_int32 xfw_Q3[], /* I Input signal */ + const opus_int16 x16[], /* I Input signal */ opus_int condCoding /* I The type of conditional coding used so far for this frame */ ) { @@ -368,15 +424,15 @@ static OPUS_INLINE void silk_LBRR_encode_FIX( /* Noise shaping quantization */ /*****************************************/ if( psEnc->sCmn.nStatesDelayedDecision > 1 || psEnc->sCmn.warping_Q16 > 0 ) { - silk_NSQ_del_dec( &psEnc->sCmn, &sNSQ_LBRR, psIndices_LBRR, xfw_Q3, + silk_NSQ_del_dec( &psEnc->sCmn, &sNSQ_LBRR, psIndices_LBRR, x16, psEnc->sCmn.pulses_LBRR[ psEnc->sCmn.nFramesEncoded ], psEncCtrl->PredCoef_Q12[ 0 ], psEncCtrl->LTPCoef_Q14, - psEncCtrl->AR2_Q13, psEncCtrl->HarmShapeGain_Q14, psEncCtrl->Tilt_Q14, psEncCtrl->LF_shp_Q14, - psEncCtrl->Gains_Q16, psEncCtrl->pitchL, psEncCtrl->Lambda_Q10, psEncCtrl->LTP_scale_Q14 ); + psEncCtrl->AR_Q13, psEncCtrl->HarmShapeGain_Q14, psEncCtrl->Tilt_Q14, psEncCtrl->LF_shp_Q14, + psEncCtrl->Gains_Q16, psEncCtrl->pitchL, psEncCtrl->Lambda_Q10, psEncCtrl->LTP_scale_Q14, psEnc->sCmn.arch ); } else { - silk_NSQ( &psEnc->sCmn, &sNSQ_LBRR, psIndices_LBRR, xfw_Q3, + silk_NSQ( &psEnc->sCmn, &sNSQ_LBRR, psIndices_LBRR, x16, psEnc->sCmn.pulses_LBRR[ psEnc->sCmn.nFramesEncoded ], psEncCtrl->PredCoef_Q12[ 0 ], psEncCtrl->LTPCoef_Q14, - psEncCtrl->AR2_Q13, psEncCtrl->HarmShapeGain_Q14, psEncCtrl->Tilt_Q14, psEncCtrl->LF_shp_Q14, - psEncCtrl->Gains_Q16, psEncCtrl->pitchL, psEncCtrl->Lambda_Q10, psEncCtrl->LTP_scale_Q14 ); + psEncCtrl->AR_Q13, psEncCtrl->HarmShapeGain_Q14, psEncCtrl->Tilt_Q14, psEncCtrl->LF_shp_Q14, + psEncCtrl->Gains_Q16, psEncCtrl->pitchL, psEncCtrl->Lambda_Q10, psEncCtrl->LTP_scale_Q14, psEnc->sCmn.arch ); } /* Restore original gains */ diff --git a/code/opus-1.1.4/silk/fixed/find_LPC_FIX.c b/code/opus-1.1.4/silk/fixed/find_LPC_FIX.c index 783d32e2..e55b63ac 100644 --- a/code/opus-1.1.4/silk/fixed/find_LPC_FIX.c +++ b/code/opus-1.1.4/silk/fixed/find_LPC_FIX.c @@ -92,10 +92,10 @@ void silk_find_LPC_FIX( silk_interpolate( NLSF0_Q15, psEncC->prev_NLSFq_Q15, NLSF_Q15, k, psEncC->predictLPCOrder ); /* Convert to LPC for residual energy evaluation */ - silk_NLSF2A( a_tmp_Q12, NLSF0_Q15, psEncC->predictLPCOrder ); + silk_NLSF2A( a_tmp_Q12, NLSF0_Q15, psEncC->predictLPCOrder, psEncC->arch ); /* Calculate residual energy with NLSF interpolation */ - silk_LPC_analysis_filter( LPC_res, x, a_tmp_Q12, 2 * subfr_length, psEncC->predictLPCOrder ); + silk_LPC_analysis_filter( LPC_res, x, a_tmp_Q12, 2 * subfr_length, psEncC->predictLPCOrder, psEncC->arch ); silk_sum_sqr_shift( &res_nrg0, &rshift0, LPC_res + psEncC->predictLPCOrder, subfr_length - psEncC->predictLPCOrder ); silk_sum_sqr_shift( &res_nrg1, &rshift1, LPC_res + psEncC->predictLPCOrder + subfr_length, subfr_length - psEncC->predictLPCOrder ); diff --git a/code/opus-1.1.4/silk/fixed/find_LTP_FIX.c b/code/opus-1.1.4/silk/fixed/find_LTP_FIX.c index 8c4d7037..62d4afb2 100644 --- a/code/opus-1.1.4/silk/fixed/find_LTP_FIX.c +++ b/code/opus-1.1.4/silk/fixed/find_LTP_FIX.c @@ -32,213 +32,68 @@ POSSIBILITY OF SUCH DAMAGE. #include "main_FIX.h" #include "tuning_parameters.h" -/* Head room for correlations */ -#define LTP_CORRS_HEAD_ROOM 2 - -void silk_fit_LTP( - opus_int32 LTP_coefs_Q16[ LTP_ORDER ], - opus_int16 LTP_coefs_Q14[ LTP_ORDER ] -); - void silk_find_LTP_FIX( - opus_int16 b_Q14[ MAX_NB_SUBFR * LTP_ORDER ], /* O LTP coefs */ - opus_int32 WLTP[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ], /* O Weight for LTP quantization */ - opus_int *LTPredCodGain_Q7, /* O LTP coding gain */ - const opus_int16 r_lpc[], /* I residual signal after LPC signal + state for first 10 ms */ + opus_int32 XXLTP_Q17[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ], /* O Correlation matrix */ + opus_int32 xXLTP_Q17[ MAX_NB_SUBFR * LTP_ORDER ], /* O Correlation vector */ + const opus_int16 r_ptr[], /* I Residual signal after LPC */ const opus_int lag[ MAX_NB_SUBFR ], /* I LTP lags */ - const opus_int32 Wght_Q15[ MAX_NB_SUBFR ], /* I weights */ - const opus_int subfr_length, /* I subframe length */ - const opus_int nb_subfr, /* I number of subframes */ - const opus_int mem_offset, /* I number of samples in LTP memory */ - opus_int corr_rshifts[ MAX_NB_SUBFR ] /* O right shifts applied to correlations */ + const opus_int subfr_length, /* I Subframe length */ + const opus_int nb_subfr, /* I Number of subframes */ + int arch /* I Run-time architecture */ ) { - opus_int i, k, lshift; - const opus_int16 *r_ptr, *lag_ptr; - opus_int16 *b_Q14_ptr; + opus_int i, k, extra_shifts; + opus_int xx_shifts, xX_shifts, XX_shifts; + const opus_int16 *lag_ptr; + opus_int32 *XXLTP_Q17_ptr, *xXLTP_Q17_ptr; + opus_int32 xx, nrg, temp; - opus_int32 regu; - opus_int32 *WLTP_ptr; - opus_int32 b_Q16[ LTP_ORDER ], delta_b_Q14[ LTP_ORDER ], d_Q14[ MAX_NB_SUBFR ], nrg[ MAX_NB_SUBFR ], g_Q26; - opus_int32 w[ MAX_NB_SUBFR ], WLTP_max, max_abs_d_Q14, max_w_bits; - - opus_int32 temp32, denom32; - opus_int extra_shifts; - opus_int rr_shifts, maxRshifts, maxRshifts_wxtra, LZs; - opus_int32 LPC_res_nrg, LPC_LTP_res_nrg, div_Q16; - opus_int32 Rr[ LTP_ORDER ], rr[ MAX_NB_SUBFR ]; - opus_int32 wd, m_Q12; - - b_Q14_ptr = b_Q14; - WLTP_ptr = WLTP; - r_ptr = &r_lpc[ mem_offset ]; + xXLTP_Q17_ptr = xXLTP_Q17; + XXLTP_Q17_ptr = XXLTP_Q17; for( k = 0; k < nb_subfr; k++ ) { lag_ptr = r_ptr - ( lag[ k ] + LTP_ORDER / 2 ); - silk_sum_sqr_shift( &rr[ k ], &rr_shifts, r_ptr, subfr_length ); /* rr[ k ] in Q( -rr_shifts ) */ - - /* Assure headroom */ - LZs = silk_CLZ32( rr[k] ); - if( LZs < LTP_CORRS_HEAD_ROOM ) { - rr[ k ] = silk_RSHIFT_ROUND( rr[ k ], LTP_CORRS_HEAD_ROOM - LZs ); - rr_shifts += ( LTP_CORRS_HEAD_ROOM - LZs ); - } - corr_rshifts[ k ] = rr_shifts; - silk_corrMatrix_FIX( lag_ptr, subfr_length, LTP_ORDER, LTP_CORRS_HEAD_ROOM, WLTP_ptr, &corr_rshifts[ k ] ); /* WLTP_fix_ptr in Q( -corr_rshifts[ k ] ) */ - - /* The correlation vector always has lower max abs value than rr and/or RR so head room is assured */ - silk_corrVector_FIX( lag_ptr, r_ptr, subfr_length, LTP_ORDER, Rr, corr_rshifts[ k ] ); /* Rr_fix_ptr in Q( -corr_rshifts[ k ] ) */ - if( corr_rshifts[ k ] > rr_shifts ) { - rr[ k ] = silk_RSHIFT( rr[ k ], corr_rshifts[ k ] - rr_shifts ); /* rr[ k ] in Q( -corr_rshifts[ k ] ) */ - } - silk_assert( rr[ k ] >= 0 ); - - regu = 1; - regu = silk_SMLAWB( regu, rr[ k ], SILK_FIX_CONST( LTP_DAMPING/3, 16 ) ); - regu = silk_SMLAWB( regu, matrix_ptr( WLTP_ptr, 0, 0, LTP_ORDER ), SILK_FIX_CONST( LTP_DAMPING/3, 16 ) ); - regu = silk_SMLAWB( regu, matrix_ptr( WLTP_ptr, LTP_ORDER-1, LTP_ORDER-1, LTP_ORDER ), SILK_FIX_CONST( LTP_DAMPING/3, 16 ) ); - silk_regularize_correlations_FIX( WLTP_ptr, &rr[k], regu, LTP_ORDER ); - - silk_solve_LDL_FIX( WLTP_ptr, LTP_ORDER, Rr, b_Q16 ); /* WLTP_fix_ptr and Rr_fix_ptr both in Q(-corr_rshifts[k]) */ - - /* Limit and store in Q14 */ - silk_fit_LTP( b_Q16, b_Q14_ptr ); - - /* Calculate residual energy */ - nrg[ k ] = silk_residual_energy16_covar_FIX( b_Q14_ptr, WLTP_ptr, Rr, rr[ k ], LTP_ORDER, 14 ); /* nrg_fix in Q( -corr_rshifts[ k ] ) */ - - /* temp = Wght[ k ] / ( nrg[ k ] * Wght[ k ] + 0.01f * subfr_length ); */ - extra_shifts = silk_min_int( corr_rshifts[ k ], LTP_CORRS_HEAD_ROOM ); - denom32 = silk_LSHIFT_SAT32( silk_SMULWB( nrg[ k ], Wght_Q15[ k ] ), 1 + extra_shifts ) + /* Q( -corr_rshifts[ k ] + extra_shifts ) */ - silk_RSHIFT( silk_SMULWB( (opus_int32)subfr_length, 655 ), corr_rshifts[ k ] - extra_shifts ); /* Q( -corr_rshifts[ k ] + extra_shifts ) */ - denom32 = silk_max( denom32, 1 ); - silk_assert( ((opus_int64)Wght_Q15[ k ] << 16 ) < silk_int32_MAX ); /* Wght always < 0.5 in Q0 */ - temp32 = silk_DIV32( silk_LSHIFT( (opus_int32)Wght_Q15[ k ], 16 ), denom32 ); /* Q( 15 + 16 + corr_rshifts[k] - extra_shifts ) */ - temp32 = silk_RSHIFT( temp32, 31 + corr_rshifts[ k ] - extra_shifts - 26 ); /* Q26 */ - - /* Limit temp such that the below scaling never wraps around */ - WLTP_max = 0; - for( i = 0; i < LTP_ORDER * LTP_ORDER; i++ ) { - WLTP_max = silk_max( WLTP_ptr[ i ], WLTP_max ); - } - lshift = silk_CLZ32( WLTP_max ) - 1 - 3; /* keep 3 bits free for vq_nearest_neighbor_fix */ - silk_assert( 26 - 18 + lshift >= 0 ); - if( 26 - 18 + lshift < 31 ) { - temp32 = silk_min_32( temp32, silk_LSHIFT( (opus_int32)1, 26 - 18 + lshift ) ); - } - - silk_scale_vector32_Q26_lshift_18( WLTP_ptr, temp32, LTP_ORDER * LTP_ORDER ); /* WLTP_ptr in Q( 18 - corr_rshifts[ k ] ) */ - - w[ k ] = matrix_ptr( WLTP_ptr, LTP_ORDER/2, LTP_ORDER/2, LTP_ORDER ); /* w in Q( 18 - corr_rshifts[ k ] ) */ - silk_assert( w[k] >= 0 ); - - r_ptr += subfr_length; - b_Q14_ptr += LTP_ORDER; - WLTP_ptr += LTP_ORDER * LTP_ORDER; - } - - maxRshifts = 0; - for( k = 0; k < nb_subfr; k++ ) { - maxRshifts = silk_max_int( corr_rshifts[ k ], maxRshifts ); - } - - /* Compute LTP coding gain */ - if( LTPredCodGain_Q7 != NULL ) { - LPC_LTP_res_nrg = 0; - LPC_res_nrg = 0; - silk_assert( LTP_CORRS_HEAD_ROOM >= 2 ); /* Check that no overflow will happen when adding */ - for( k = 0; k < nb_subfr; k++ ) { - LPC_res_nrg = silk_ADD32( LPC_res_nrg, silk_RSHIFT( silk_ADD32( silk_SMULWB( rr[ k ], Wght_Q15[ k ] ), 1 ), 1 + ( maxRshifts - corr_rshifts[ k ] ) ) ); /* Q( -maxRshifts ) */ - LPC_LTP_res_nrg = silk_ADD32( LPC_LTP_res_nrg, silk_RSHIFT( silk_ADD32( silk_SMULWB( nrg[ k ], Wght_Q15[ k ] ), 1 ), 1 + ( maxRshifts - corr_rshifts[ k ] ) ) ); /* Q( -maxRshifts ) */ - } - LPC_LTP_res_nrg = silk_max( LPC_LTP_res_nrg, 1 ); /* avoid division by zero */ - - div_Q16 = silk_DIV32_varQ( LPC_res_nrg, LPC_LTP_res_nrg, 16 ); - *LTPredCodGain_Q7 = ( opus_int )silk_SMULBB( 3, silk_lin2log( div_Q16 ) - ( 16 << 7 ) ); - - silk_assert( *LTPredCodGain_Q7 == ( opus_int )silk_SAT16( silk_MUL( 3, silk_lin2log( div_Q16 ) - ( 16 << 7 ) ) ) ); - } - - /* smoothing */ - /* d = sum( B, 1 ); */ - b_Q14_ptr = b_Q14; - for( k = 0; k < nb_subfr; k++ ) { - d_Q14[ k ] = 0; - for( i = 0; i < LTP_ORDER; i++ ) { - d_Q14[ k ] += b_Q14_ptr[ i ]; - } - b_Q14_ptr += LTP_ORDER; - } - - /* m = ( w * d' ) / ( sum( w ) + 1e-3 ); */ - - /* Find maximum absolute value of d_Q14 and the bits used by w in Q0 */ - max_abs_d_Q14 = 0; - max_w_bits = 0; - for( k = 0; k < nb_subfr; k++ ) { - max_abs_d_Q14 = silk_max_32( max_abs_d_Q14, silk_abs( d_Q14[ k ] ) ); - /* w[ k ] is in Q( 18 - corr_rshifts[ k ] ) */ - /* Find bits needed in Q( 18 - maxRshifts ) */ - max_w_bits = silk_max_32( max_w_bits, 32 - silk_CLZ32( w[ k ] ) + corr_rshifts[ k ] - maxRshifts ); - } - - /* max_abs_d_Q14 = (5 << 15); worst case, i.e. LTP_ORDER * -silk_int16_MIN */ - silk_assert( max_abs_d_Q14 <= ( 5 << 15 ) ); - - /* How many bits is needed for w*d' in Q( 18 - maxRshifts ) in the worst case, of all d_Q14's being equal to max_abs_d_Q14 */ - extra_shifts = max_w_bits + 32 - silk_CLZ32( max_abs_d_Q14 ) - 14; - - /* Subtract what we got available; bits in output var plus maxRshifts */ - extra_shifts -= ( 32 - 1 - 2 + maxRshifts ); /* Keep sign bit free as well as 2 bits for accumulation */ - extra_shifts = silk_max_int( extra_shifts, 0 ); - - maxRshifts_wxtra = maxRshifts + extra_shifts; - - temp32 = silk_RSHIFT( 262, maxRshifts + extra_shifts ) + 1; /* 1e-3f in Q( 18 - (maxRshifts + extra_shifts) ) */ - wd = 0; - for( k = 0; k < nb_subfr; k++ ) { - /* w has at least 2 bits of headroom so no overflow should happen */ - temp32 = silk_ADD32( temp32, silk_RSHIFT( w[ k ], maxRshifts_wxtra - corr_rshifts[ k ] ) ); /* Q( 18 - maxRshifts_wxtra ) */ - wd = silk_ADD32( wd, silk_LSHIFT( silk_SMULWW( silk_RSHIFT( w[ k ], maxRshifts_wxtra - corr_rshifts[ k ] ), d_Q14[ k ] ), 2 ) ); /* Q( 18 - maxRshifts_wxtra ) */ - } - m_Q12 = silk_DIV32_varQ( wd, temp32, 12 ); - - b_Q14_ptr = b_Q14; - for( k = 0; k < nb_subfr; k++ ) { - /* w_fix[ k ] from Q( 18 - corr_rshifts[ k ] ) to Q( 16 ) */ - if( 2 - corr_rshifts[k] > 0 ) { - temp32 = silk_RSHIFT( w[ k ], 2 - corr_rshifts[ k ] ); + silk_sum_sqr_shift( &xx, &xx_shifts, r_ptr, subfr_length + LTP_ORDER ); /* xx in Q( -xx_shifts ) */ + silk_corrMatrix_FIX( lag_ptr, subfr_length, LTP_ORDER, XXLTP_Q17_ptr, &nrg, &XX_shifts, arch ); /* XXLTP_Q17_ptr and nrg in Q( -XX_shifts ) */ + extra_shifts = xx_shifts - XX_shifts; + if( extra_shifts > 0 ) { + /* Shift XX */ + xX_shifts = xx_shifts; + for( i = 0; i < LTP_ORDER * LTP_ORDER; i++ ) { + XXLTP_Q17_ptr[ i ] = silk_RSHIFT32( XXLTP_Q17_ptr[ i ], extra_shifts ); /* Q( -xX_shifts ) */ + } + nrg = silk_RSHIFT32( nrg, extra_shifts ); /* Q( -xX_shifts ) */ + } else if( extra_shifts < 0 ) { + /* Shift xx */ + xX_shifts = XX_shifts; + xx = silk_RSHIFT32( xx, -extra_shifts ); /* Q( -xX_shifts ) */ } else { - temp32 = silk_LSHIFT_SAT32( w[ k ], corr_rshifts[ k ] - 2 ); + xX_shifts = xx_shifts; } + silk_corrVector_FIX( lag_ptr, r_ptr, subfr_length, LTP_ORDER, xXLTP_Q17_ptr, xX_shifts, arch ); /* xXLTP_Q17_ptr in Q( -xX_shifts ) */ - g_Q26 = silk_MUL( - silk_DIV32( - SILK_FIX_CONST( LTP_SMOOTHING, 26 ), - silk_RSHIFT( SILK_FIX_CONST( LTP_SMOOTHING, 26 ), 10 ) + temp32 ), /* Q10 */ - silk_LSHIFT_SAT32( silk_SUB_SAT32( (opus_int32)m_Q12, silk_RSHIFT( d_Q14[ k ], 2 ) ), 4 ) ); /* Q16 */ - - temp32 = 0; - for( i = 0; i < LTP_ORDER; i++ ) { - delta_b_Q14[ i ] = silk_max_16( b_Q14_ptr[ i ], 1638 ); /* 1638_Q14 = 0.1_Q0 */ - temp32 += delta_b_Q14[ i ]; /* Q14 */ + /* At this point all correlations are in Q(-xX_shifts) */ + temp = silk_SMLAWB( 1, nrg, SILK_FIX_CONST( LTP_CORR_INV_MAX, 16 ) ); + temp = silk_max( temp, xx ); +TIC(div) +#if 0 + for( i = 0; i < LTP_ORDER * LTP_ORDER; i++ ) { + XXLTP_Q17_ptr[ i ] = silk_DIV32_varQ( XXLTP_Q17_ptr[ i ], temp, 17 ); } - temp32 = silk_DIV32( g_Q26, temp32 ); /* Q14 -> Q12 */ for( i = 0; i < LTP_ORDER; i++ ) { - b_Q14_ptr[ i ] = silk_LIMIT_32( (opus_int32)b_Q14_ptr[ i ] + silk_SMULWB( silk_LSHIFT_SAT32( temp32, 4 ), delta_b_Q14[ i ] ), -16000, 28000 ); + xXLTP_Q17_ptr[ i ] = silk_DIV32_varQ( xXLTP_Q17_ptr[ i ], temp, 17 ); } - b_Q14_ptr += LTP_ORDER; - } -} - -void silk_fit_LTP( - opus_int32 LTP_coefs_Q16[ LTP_ORDER ], - opus_int16 LTP_coefs_Q14[ LTP_ORDER ] -) -{ - opus_int i; - - for( i = 0; i < LTP_ORDER; i++ ) { - LTP_coefs_Q14[ i ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( LTP_coefs_Q16[ i ], 2 ) ); +#else + for( i = 0; i < LTP_ORDER * LTP_ORDER; i++ ) { + XXLTP_Q17_ptr[ i ] = (opus_int32)( silk_LSHIFT64( (opus_int64)XXLTP_Q17_ptr[ i ], 17 ) / temp ); + } + for( i = 0; i < LTP_ORDER; i++ ) { + xXLTP_Q17_ptr[ i ] = (opus_int32)( silk_LSHIFT64( (opus_int64)xXLTP_Q17_ptr[ i ], 17 ) / temp ); + } +#endif +TOC(div) + r_ptr += subfr_length; + XXLTP_Q17_ptr += LTP_ORDER * LTP_ORDER; + xXLTP_Q17_ptr += LTP_ORDER; } } diff --git a/code/opus-1.1.4/silk/fixed/find_pitch_lags_FIX.c b/code/opus-1.1.4/silk/fixed/find_pitch_lags_FIX.c index 620f8dcd..9303e9db 100644 --- a/code/opus-1.1.4/silk/fixed/find_pitch_lags_FIX.c +++ b/code/opus-1.1.4/silk/fixed/find_pitch_lags_FIX.c @@ -44,7 +44,7 @@ void silk_find_pitch_lags_FIX( { opus_int buf_len, i, scale; opus_int32 thrhld_Q13, res_nrg; - const opus_int16 *x_buf, *x_buf_ptr; + const opus_int16 *x_ptr; VARDECL( opus_int16, Wsig ); opus_int16 *Wsig_ptr; opus_int32 auto_corr[ MAX_FIND_PITCH_LPC_ORDER + 1 ]; @@ -61,8 +61,6 @@ void silk_find_pitch_lags_FIX( /* Safety check */ silk_assert( buf_len >= psEnc->sCmn.pitch_LPC_win_length ); - x_buf = x - psEnc->sCmn.ltp_mem_length; - /*************************************/ /* Estimate LPC AR coefficients */ /*************************************/ @@ -72,19 +70,19 @@ void silk_find_pitch_lags_FIX( ALLOC( Wsig, psEnc->sCmn.pitch_LPC_win_length, opus_int16 ); /* First LA_LTP samples */ - x_buf_ptr = x_buf + buf_len - psEnc->sCmn.pitch_LPC_win_length; + x_ptr = x + buf_len - psEnc->sCmn.pitch_LPC_win_length; Wsig_ptr = Wsig; - silk_apply_sine_window( Wsig_ptr, x_buf_ptr, 1, psEnc->sCmn.la_pitch ); + silk_apply_sine_window( Wsig_ptr, x_ptr, 1, psEnc->sCmn.la_pitch ); /* Middle un - windowed samples */ Wsig_ptr += psEnc->sCmn.la_pitch; - x_buf_ptr += psEnc->sCmn.la_pitch; - silk_memcpy( Wsig_ptr, x_buf_ptr, ( psEnc->sCmn.pitch_LPC_win_length - silk_LSHIFT( psEnc->sCmn.la_pitch, 1 ) ) * sizeof( opus_int16 ) ); + x_ptr += psEnc->sCmn.la_pitch; + silk_memcpy( Wsig_ptr, x_ptr, ( psEnc->sCmn.pitch_LPC_win_length - silk_LSHIFT( psEnc->sCmn.la_pitch, 1 ) ) * sizeof( opus_int16 ) ); /* Last LA_LTP samples */ Wsig_ptr += psEnc->sCmn.pitch_LPC_win_length - silk_LSHIFT( psEnc->sCmn.la_pitch, 1 ); - x_buf_ptr += psEnc->sCmn.pitch_LPC_win_length - silk_LSHIFT( psEnc->sCmn.la_pitch, 1 ); - silk_apply_sine_window( Wsig_ptr, x_buf_ptr, 2, psEnc->sCmn.la_pitch ); + x_ptr += psEnc->sCmn.pitch_LPC_win_length - silk_LSHIFT( psEnc->sCmn.la_pitch, 1 ); + silk_apply_sine_window( Wsig_ptr, x_ptr, 2, psEnc->sCmn.la_pitch ); /* Calculate autocorrelation sequence */ silk_autocorr( auto_corr, &scale, Wsig, psEnc->sCmn.pitch_LPC_win_length, psEnc->sCmn.pitchEstimationLPCOrder + 1, arch ); @@ -112,7 +110,7 @@ void silk_find_pitch_lags_FIX( /*****************************************/ /* LPC analysis filtering */ /*****************************************/ - silk_LPC_analysis_filter( res, x_buf, A_Q12, buf_len, psEnc->sCmn.pitchEstimationLPCOrder ); + silk_LPC_analysis_filter( res, x, A_Q12, buf_len, psEnc->sCmn.pitchEstimationLPCOrder, psEnc->sCmn.arch ); if( psEnc->sCmn.indices.signalType != TYPE_NO_VOICE_ACTIVITY && psEnc->sCmn.first_frame_after_reset == 0 ) { /* Threshold for pitch estimator */ diff --git a/code/opus-1.1.4/silk/fixed/find_pred_coefs_FIX.c b/code/opus-1.1.4/silk/fixed/find_pred_coefs_FIX.c index 5c22f828..24c6aab3 100644 --- a/code/opus-1.1.4/silk/fixed/find_pred_coefs_FIX.c +++ b/code/opus-1.1.4/silk/fixed/find_pred_coefs_FIX.c @@ -41,13 +41,12 @@ void silk_find_pred_coefs_FIX( ) { opus_int i; - opus_int32 invGains_Q16[ MAX_NB_SUBFR ], local_gains[ MAX_NB_SUBFR ], Wght_Q15[ MAX_NB_SUBFR ]; + opus_int32 invGains_Q16[ MAX_NB_SUBFR ], local_gains[ MAX_NB_SUBFR ]; opus_int16 NLSF_Q15[ MAX_LPC_ORDER ]; const opus_int16 *x_ptr; opus_int16 *x_pre_ptr; VARDECL( opus_int16, LPC_in_pre ); - opus_int32 tmp, min_gain_Q16, minInvGain_Q30; - opus_int LTP_corrs_rshift[ MAX_NB_SUBFR ]; + opus_int32 min_gain_Q16, minInvGain_Q30; SAVE_STACK; /* weighting for weighted least squares */ @@ -61,13 +60,11 @@ void silk_find_pred_coefs_FIX( /* Invert and normalize gains, and ensure that maximum invGains_Q16 is within range of a 16 bit int */ invGains_Q16[ i ] = silk_DIV32_varQ( min_gain_Q16, psEncCtrl->Gains_Q16[ i ], 16 - 2 ); - /* Ensure Wght_Q15 a minimum value 1 */ - invGains_Q16[ i ] = silk_max( invGains_Q16[ i ], 363 ); + /* Limit inverse */ + invGains_Q16[ i ] = silk_max( invGains_Q16[ i ], 100 ); /* Square the inverted gains */ silk_assert( invGains_Q16[ i ] == silk_SAT16( invGains_Q16[ i ] ) ); - tmp = silk_SMULWB( invGains_Q16[ i ], invGains_Q16[ i ] ); - Wght_Q15[ i ] = silk_RSHIFT( tmp, 1 ); /* Invert the inverted and normalized gains */ local_gains[ i ] = silk_DIV32( ( (opus_int32)1 << 16 ), invGains_Q16[ i ] ); @@ -77,23 +74,24 @@ void silk_find_pred_coefs_FIX( psEnc->sCmn.nb_subfr * psEnc->sCmn.predictLPCOrder + psEnc->sCmn.frame_length, opus_int16 ); if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) { - VARDECL( opus_int32, WLTP ); + VARDECL( opus_int32, xXLTP_Q17 ); + VARDECL( opus_int32, XXLTP_Q17 ); /**********/ /* VOICED */ /**********/ silk_assert( psEnc->sCmn.ltp_mem_length - psEnc->sCmn.predictLPCOrder >= psEncCtrl->pitchL[ 0 ] + LTP_ORDER / 2 ); - ALLOC( WLTP, psEnc->sCmn.nb_subfr * LTP_ORDER * LTP_ORDER, opus_int32 ); + ALLOC( xXLTP_Q17, psEnc->sCmn.nb_subfr * LTP_ORDER, opus_int32 ); + ALLOC( XXLTP_Q17, psEnc->sCmn.nb_subfr * LTP_ORDER * LTP_ORDER, opus_int32 ); /* LTP analysis */ - silk_find_LTP_FIX( psEncCtrl->LTPCoef_Q14, WLTP, &psEncCtrl->LTPredCodGain_Q7, - res_pitch, psEncCtrl->pitchL, Wght_Q15, psEnc->sCmn.subfr_length, - psEnc->sCmn.nb_subfr, psEnc->sCmn.ltp_mem_length, LTP_corrs_rshift ); + silk_find_LTP_FIX( XXLTP_Q17, xXLTP_Q17, res_pitch, + psEncCtrl->pitchL, psEnc->sCmn.subfr_length, psEnc->sCmn.nb_subfr, psEnc->sCmn.arch ); /* Quantize LTP gain parameters */ silk_quant_LTP_gains( psEncCtrl->LTPCoef_Q14, psEnc->sCmn.indices.LTPIndex, &psEnc->sCmn.indices.PERIndex, - &psEnc->sCmn.sum_log_gain_Q7, WLTP, psEnc->sCmn.mu_LTP_Q9, psEnc->sCmn.LTPQuantLowComplexity, psEnc->sCmn.nb_subfr); + &psEnc->sCmn.sum_log_gain_Q7, &psEncCtrl->LTPredCodGain_Q7, XXLTP_Q17, xXLTP_Q17, psEnc->sCmn.subfr_length, psEnc->sCmn.nb_subfr, psEnc->sCmn.arch ); /* Control LTP scaling */ silk_LTP_scale_ctrl_FIX( psEnc, psEncCtrl, condCoding ); @@ -118,16 +116,16 @@ void silk_find_pred_coefs_FIX( silk_memset( psEncCtrl->LTPCoef_Q14, 0, psEnc->sCmn.nb_subfr * LTP_ORDER * sizeof( opus_int16 ) ); psEncCtrl->LTPredCodGain_Q7 = 0; - psEnc->sCmn.sum_log_gain_Q7 = 0; + psEnc->sCmn.sum_log_gain_Q7 = 0; } /* Limit on total predictive coding gain */ if( psEnc->sCmn.first_frame_after_reset ) { minInvGain_Q30 = SILK_FIX_CONST( 1.0f / MAX_PREDICTION_POWER_GAIN_AFTER_RESET, 30 ); - } else { + } else { minInvGain_Q30 = silk_log2lin( silk_SMLAWB( 16 << 7, (opus_int32)psEncCtrl->LTPredCodGain_Q7, SILK_FIX_CONST( 1.0 / 3, 16 ) ) ); /* Q16 */ - minInvGain_Q30 = silk_DIV32_varQ( minInvGain_Q30, - silk_SMULWW( SILK_FIX_CONST( MAX_PREDICTION_POWER_GAIN, 0 ), + minInvGain_Q30 = silk_DIV32_varQ( minInvGain_Q30, + silk_SMULWW( SILK_FIX_CONST( MAX_PREDICTION_POWER_GAIN, 0 ), silk_SMLAWB( SILK_FIX_CONST( 0.25, 18 ), SILK_FIX_CONST( 0.75, 18 ), psEncCtrl->coding_quality_Q14 ) ), 14 ); } @@ -139,7 +137,7 @@ void silk_find_pred_coefs_FIX( /* Calculate residual energy using quantized LPC coefficients */ silk_residual_energy_FIX( psEncCtrl->ResNrg, psEncCtrl->ResNrgQ, LPC_in_pre, psEncCtrl->PredCoef_Q12, local_gains, - psEnc->sCmn.subfr_length, psEnc->sCmn.nb_subfr, psEnc->sCmn.predictLPCOrder ); + psEnc->sCmn.subfr_length, psEnc->sCmn.nb_subfr, psEnc->sCmn.predictLPCOrder, psEnc->sCmn.arch ); /* Copy to prediction struct for use in next frame for interpolation */ silk_memcpy( psEnc->sCmn.prev_NLSFq_Q15, NLSF_Q15, sizeof( psEnc->sCmn.prev_NLSFq_Q15 ) ); diff --git a/code/opus-1.1.4/silk/fixed/k2a_FIX.c b/code/opus-1.1.4/silk/fixed/k2a_FIX.c index 5fee599b..549f6ead 100644 --- a/code/opus-1.1.4/silk/fixed/k2a_FIX.c +++ b/code/opus-1.1.4/silk/fixed/k2a_FIX.c @@ -39,14 +39,15 @@ void silk_k2a( ) { opus_int k, n; - opus_int32 Atmp[ SILK_MAX_ORDER_LPC ]; + opus_int32 rc, tmp1, tmp2; for( k = 0; k < order; k++ ) { - for( n = 0; n < k; n++ ) { - Atmp[ n ] = A_Q24[ n ]; - } - for( n = 0; n < k; n++ ) { - A_Q24[ n ] = silk_SMLAWB( A_Q24[ n ], silk_LSHIFT( Atmp[ k - n - 1 ], 1 ), rc_Q15[ k ] ); + rc = rc_Q15[ k ]; + for( n = 0; n < (k + 1) >> 1; n++ ) { + tmp1 = A_Q24[ n ]; + tmp2 = A_Q24[ k - n - 1 ]; + A_Q24[ n ] = silk_SMLAWB( tmp1, silk_LSHIFT( tmp2, 1 ), rc ); + A_Q24[ k - n - 1 ] = silk_SMLAWB( tmp2, silk_LSHIFT( tmp1, 1 ), rc ); } A_Q24[ k ] = -silk_LSHIFT( (opus_int32)rc_Q15[ k ], 9 ); } diff --git a/code/opus-1.1.4/silk/fixed/k2a_Q16_FIX.c b/code/opus-1.1.4/silk/fixed/k2a_Q16_FIX.c index 3b039875..1595aa62 100644 --- a/code/opus-1.1.4/silk/fixed/k2a_Q16_FIX.c +++ b/code/opus-1.1.4/silk/fixed/k2a_Q16_FIX.c @@ -39,15 +39,16 @@ void silk_k2a_Q16( ) { opus_int k, n; - opus_int32 Atmp[ SILK_MAX_ORDER_LPC ]; + opus_int32 rc, tmp1, tmp2; for( k = 0; k < order; k++ ) { - for( n = 0; n < k; n++ ) { - Atmp[ n ] = A_Q24[ n ]; + rc = rc_Q16[ k ]; + for( n = 0; n < (k + 1) >> 1; n++ ) { + tmp1 = A_Q24[ n ]; + tmp2 = A_Q24[ k - n - 1 ]; + A_Q24[ n ] = silk_SMLAWW( tmp1, tmp2, rc ); + A_Q24[ k - n - 1 ] = silk_SMLAWW( tmp2, tmp1, rc ); } - for( n = 0; n < k; n++ ) { - A_Q24[ n ] = silk_SMLAWW( A_Q24[ n ], Atmp[ k - n - 1 ], rc_Q16[ k ] ); - } - A_Q24[ k ] = -silk_LSHIFT( rc_Q16[ k ], 8 ); + A_Q24[ k ] = -silk_LSHIFT( rc, 8 ); } } diff --git a/code/opus-1.1.4/silk/fixed/main_FIX.h b/code/opus-1.1.4/silk/fixed/main_FIX.h index a56ca07a..780afa39 100644 --- a/code/opus-1.1.4/silk/fixed/main_FIX.h +++ b/code/opus-1.1.4/silk/fixed/main_FIX.h @@ -36,6 +36,11 @@ POSSIBILITY OF SUCH DAMAGE. #include "debug.h" #include "entenc.h" +#if ((defined(OPUS_ARM_ASM) && defined(FIXED_POINT)) \ + || defined(OPUS_ARM_MAY_HAVE_NEON_INTR)) +#include "fixed/arm/warped_autocorrelation_FIX_arm.h" +#endif + #ifndef FORCE_CPP_BUILD #ifdef __cplusplus extern "C" @@ -47,6 +52,9 @@ extern "C" #define silk_encode_do_VAD_Fxx silk_encode_do_VAD_FIX #define silk_encode_frame_Fxx silk_encode_frame_FIX +#define QC 10 +#define QS 13 + /*********************/ /* Encoder Functions */ /*********************/ @@ -81,22 +89,11 @@ opus_int silk_init_encoder( opus_int silk_control_encoder( silk_encoder_state_Fxx *psEnc, /* I/O Pointer to Silk encoder state */ silk_EncControlStruct *encControl, /* I Control structure */ - const opus_int32 TargetRate_bps, /* I Target max bitrate (bps) */ const opus_int allow_bw_switch, /* I Flag to allow switching audio bandwidth */ const opus_int channelNb, /* I Channel number */ const opus_int force_fs_kHz ); -/****************/ -/* Prefiltering */ -/****************/ -void silk_prefilter_FIX( - silk_encoder_state_FIX *psEnc, /* I/O Encoder state */ - const silk_encoder_control_FIX *psEncCtrl, /* I Encoder control */ - opus_int32 xw_Q10[], /* O Weighted signal */ - const opus_int16 x[] /* I Speech signal */ -); - /**************************/ /* Noise shaping analysis */ /**************************/ @@ -110,7 +107,7 @@ void silk_noise_shape_analysis_FIX( ); /* Autocorrelations for a warped frequency axis */ -void silk_warped_autocorrelation_FIX( +void silk_warped_autocorrelation_FIX_c( opus_int32 *corr, /* O Result [order + 1] */ opus_int *scale, /* O Scaling of the correlation vector */ const opus_int16 *input, /* I Input data to correlate */ @@ -119,6 +116,11 @@ void silk_warped_autocorrelation_FIX( const opus_int order /* I Correlation order (even) */ ); +#if !defined(OVERRIDE_silk_warped_autocorrelation_FIX) +#define silk_warped_autocorrelation_FIX(corr, scale, input, warping_Q16, length, order, arch) \ + ((void)(arch), silk_warped_autocorrelation_FIX_c(corr, scale, input, warping_Q16, length, order)) +#endif + /* Calculation of LTP state scaling */ void silk_LTP_scale_ctrl_FIX( silk_encoder_state_FIX *psEnc, /* I/O encoder state */ @@ -157,16 +159,13 @@ void silk_find_LPC_FIX( /* LTP analysis */ void silk_find_LTP_FIX( - opus_int16 b_Q14[ MAX_NB_SUBFR * LTP_ORDER ], /* O LTP coefs */ - opus_int32 WLTP[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ], /* O Weight for LTP quantization */ - opus_int *LTPredCodGain_Q7, /* O LTP coding gain */ - const opus_int16 r_lpc[], /* I residual signal after LPC signal + state for first 10 ms */ + opus_int32 XXLTP_Q17[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ], /* O Correlation matrix */ + opus_int32 xXLTP_Q17[ MAX_NB_SUBFR * LTP_ORDER ], /* O Correlation vector */ + const opus_int16 r_lpc[], /* I Residual signal after LPC */ const opus_int lag[ MAX_NB_SUBFR ], /* I LTP lags */ - const opus_int32 Wght_Q15[ MAX_NB_SUBFR ], /* I weights */ - const opus_int subfr_length, /* I subframe length */ - const opus_int nb_subfr, /* I number of subframes */ - const opus_int mem_offset, /* I number of samples in LTP memory */ - opus_int corr_rshifts[ MAX_NB_SUBFR ] /* O right shifts applied to correlations */ + const opus_int subfr_length, /* I Subframe length */ + const opus_int nb_subfr, /* I Number of subframes */ + int arch /* I Run-time architecture */ ); void silk_LTP_analysis_filter_FIX( @@ -190,7 +189,8 @@ void silk_residual_energy_FIX( const opus_int32 gains[ MAX_NB_SUBFR ], /* I Quantization gains */ const opus_int subfr_length, /* I Subframe length */ const opus_int nb_subfr, /* I Number of subframes */ - const opus_int LPC_order /* I LPC order */ + const opus_int LPC_order, /* I LPC order */ + int arch /* I Run-time architecture */ ); /* Residual energy: nrg = wxx - 2 * wXx * c + c' * wXX * c */ @@ -218,9 +218,10 @@ void silk_corrMatrix_FIX( const opus_int16 *x, /* I x vector [L + order - 1] used to form data matrix X */ const opus_int L, /* I Length of vectors */ const opus_int order, /* I Max lag for correlation */ - const opus_int head_room, /* I Desired headroom */ opus_int32 *XX, /* O Pointer to X'*X correlation matrix [ order x order ] */ - opus_int *rshifts /* I/O Right shifts of correlations */ + opus_int32 *nrg, /* O Energy of x vector */ + opus_int *rshifts, /* O Right shifts of correlations */ + int arch /* I Run-time architecture */ ); /* Calculates correlation vector X'*t */ @@ -230,23 +231,8 @@ void silk_corrVector_FIX( const opus_int L, /* I Length of vectors */ const opus_int order, /* I Max lag for correlation */ opus_int32 *Xt, /* O Pointer to X'*t correlation vector [order] */ - const opus_int rshifts /* I Right shifts of correlations */ -); - -/* Add noise to matrix diagonal */ -void silk_regularize_correlations_FIX( - opus_int32 *XX, /* I/O Correlation matrices */ - opus_int32 *xx, /* I/O Correlation values */ - opus_int32 noise, /* I Noise to add */ - opus_int D /* I Dimension of XX */ -); - -/* Solves Ax = b, assuming A is symmetric */ -void silk_solve_LDL_FIX( - opus_int32 *A, /* I Pointer to symetric square matrix A */ - opus_int M, /* I Size of matrix */ - const opus_int32 *b, /* I Pointer to b vector */ - opus_int32 *x_Q16 /* O Pointer to x solution vector */ + const opus_int rshifts, /* I Right shifts of correlations */ + int arch /* I Run-time architecture */ ); #ifndef FORCE_CPP_BUILD diff --git a/code/opus-1.1.4/silk/fixed/noise_shape_analysis_FIX.c b/code/opus-1.1.4/silk/fixed/noise_shape_analysis_FIX.c index e24d2e9d..85fea0bf 100644 --- a/code/opus-1.1.4/silk/fixed/noise_shape_analysis_FIX.c +++ b/code/opus-1.1.4/silk/fixed/noise_shape_analysis_FIX.c @@ -57,90 +57,86 @@ static OPUS_INLINE opus_int32 warped_gain( /* gain in Q16*/ /* Convert warped filter coefficients to monic pseudo-warped coefficients and limit maximum */ /* amplitude of monic warped coefficients by using bandwidth expansion on the true coefficients */ static OPUS_INLINE void limit_warped_coefs( - opus_int32 *coefs_syn_Q24, - opus_int32 *coefs_ana_Q24, + opus_int32 *coefs_Q24, opus_int lambda_Q16, opus_int32 limit_Q24, opus_int order ) { opus_int i, iter, ind = 0; - opus_int32 tmp, maxabs_Q24, chirp_Q16, gain_syn_Q16, gain_ana_Q16; + opus_int32 tmp, maxabs_Q24, chirp_Q16, gain_Q16; opus_int32 nom_Q16, den_Q24; + opus_int32 limit_Q20, maxabs_Q20; /* Convert to monic coefficients */ lambda_Q16 = -lambda_Q16; for( i = order - 1; i > 0; i-- ) { - coefs_syn_Q24[ i - 1 ] = silk_SMLAWB( coefs_syn_Q24[ i - 1 ], coefs_syn_Q24[ i ], lambda_Q16 ); - coefs_ana_Q24[ i - 1 ] = silk_SMLAWB( coefs_ana_Q24[ i - 1 ], coefs_ana_Q24[ i ], lambda_Q16 ); + coefs_Q24[ i - 1 ] = silk_SMLAWB( coefs_Q24[ i - 1 ], coefs_Q24[ i ], lambda_Q16 ); } lambda_Q16 = -lambda_Q16; - nom_Q16 = silk_SMLAWB( SILK_FIX_CONST( 1.0, 16 ), -(opus_int32)lambda_Q16, lambda_Q16 ); - den_Q24 = silk_SMLAWB( SILK_FIX_CONST( 1.0, 24 ), coefs_syn_Q24[ 0 ], lambda_Q16 ); - gain_syn_Q16 = silk_DIV32_varQ( nom_Q16, den_Q24, 24 ); - den_Q24 = silk_SMLAWB( SILK_FIX_CONST( 1.0, 24 ), coefs_ana_Q24[ 0 ], lambda_Q16 ); - gain_ana_Q16 = silk_DIV32_varQ( nom_Q16, den_Q24, 24 ); + nom_Q16 = silk_SMLAWB( SILK_FIX_CONST( 1.0, 16 ), -(opus_int32)lambda_Q16, lambda_Q16 ); + den_Q24 = silk_SMLAWB( SILK_FIX_CONST( 1.0, 24 ), coefs_Q24[ 0 ], lambda_Q16 ); + gain_Q16 = silk_DIV32_varQ( nom_Q16, den_Q24, 24 ); for( i = 0; i < order; i++ ) { - coefs_syn_Q24[ i ] = silk_SMULWW( gain_syn_Q16, coefs_syn_Q24[ i ] ); - coefs_ana_Q24[ i ] = silk_SMULWW( gain_ana_Q16, coefs_ana_Q24[ i ] ); + coefs_Q24[ i ] = silk_SMULWW( gain_Q16, coefs_Q24[ i ] ); } - + limit_Q20 = silk_RSHIFT(limit_Q24, 4); for( iter = 0; iter < 10; iter++ ) { /* Find maximum absolute value */ maxabs_Q24 = -1; for( i = 0; i < order; i++ ) { - tmp = silk_max( silk_abs_int32( coefs_syn_Q24[ i ] ), silk_abs_int32( coefs_ana_Q24[ i ] ) ); + tmp = silk_abs_int32( coefs_Q24[ i ] ); if( tmp > maxabs_Q24 ) { maxabs_Q24 = tmp; ind = i; } } - if( maxabs_Q24 <= limit_Q24 ) { + /* Use Q20 to avoid any overflow when multiplying by (ind + 1) later. */ + maxabs_Q20 = silk_RSHIFT(maxabs_Q24, 4); + if( maxabs_Q20 <= limit_Q20 ) { /* Coefficients are within range - done */ return; } /* Convert back to true warped coefficients */ for( i = 1; i < order; i++ ) { - coefs_syn_Q24[ i - 1 ] = silk_SMLAWB( coefs_syn_Q24[ i - 1 ], coefs_syn_Q24[ i ], lambda_Q16 ); - coefs_ana_Q24[ i - 1 ] = silk_SMLAWB( coefs_ana_Q24[ i - 1 ], coefs_ana_Q24[ i ], lambda_Q16 ); + coefs_Q24[ i - 1 ] = silk_SMLAWB( coefs_Q24[ i - 1 ], coefs_Q24[ i ], lambda_Q16 ); } - gain_syn_Q16 = silk_INVERSE32_varQ( gain_syn_Q16, 32 ); - gain_ana_Q16 = silk_INVERSE32_varQ( gain_ana_Q16, 32 ); + gain_Q16 = silk_INVERSE32_varQ( gain_Q16, 32 ); for( i = 0; i < order; i++ ) { - coefs_syn_Q24[ i ] = silk_SMULWW( gain_syn_Q16, coefs_syn_Q24[ i ] ); - coefs_ana_Q24[ i ] = silk_SMULWW( gain_ana_Q16, coefs_ana_Q24[ i ] ); + coefs_Q24[ i ] = silk_SMULWW( gain_Q16, coefs_Q24[ i ] ); } /* Apply bandwidth expansion */ chirp_Q16 = SILK_FIX_CONST( 0.99, 16 ) - silk_DIV32_varQ( - silk_SMULWB( maxabs_Q24 - limit_Q24, silk_SMLABB( SILK_FIX_CONST( 0.8, 10 ), SILK_FIX_CONST( 0.1, 10 ), iter ) ), - silk_MUL( maxabs_Q24, ind + 1 ), 22 ); - silk_bwexpander_32( coefs_syn_Q24, order, chirp_Q16 ); - silk_bwexpander_32( coefs_ana_Q24, order, chirp_Q16 ); + silk_SMULWB( maxabs_Q20 - limit_Q20, silk_SMLABB( SILK_FIX_CONST( 0.8, 10 ), SILK_FIX_CONST( 0.1, 10 ), iter ) ), + silk_MUL( maxabs_Q20, ind + 1 ), 22 ); + silk_bwexpander_32( coefs_Q24, order, chirp_Q16 ); /* Convert to monic warped coefficients */ lambda_Q16 = -lambda_Q16; for( i = order - 1; i > 0; i-- ) { - coefs_syn_Q24[ i - 1 ] = silk_SMLAWB( coefs_syn_Q24[ i - 1 ], coefs_syn_Q24[ i ], lambda_Q16 ); - coefs_ana_Q24[ i - 1 ] = silk_SMLAWB( coefs_ana_Q24[ i - 1 ], coefs_ana_Q24[ i ], lambda_Q16 ); + coefs_Q24[ i - 1 ] = silk_SMLAWB( coefs_Q24[ i - 1 ], coefs_Q24[ i ], lambda_Q16 ); } lambda_Q16 = -lambda_Q16; nom_Q16 = silk_SMLAWB( SILK_FIX_CONST( 1.0, 16 ), -(opus_int32)lambda_Q16, lambda_Q16 ); - den_Q24 = silk_SMLAWB( SILK_FIX_CONST( 1.0, 24 ), coefs_syn_Q24[ 0 ], lambda_Q16 ); - gain_syn_Q16 = silk_DIV32_varQ( nom_Q16, den_Q24, 24 ); - den_Q24 = silk_SMLAWB( SILK_FIX_CONST( 1.0, 24 ), coefs_ana_Q24[ 0 ], lambda_Q16 ); - gain_ana_Q16 = silk_DIV32_varQ( nom_Q16, den_Q24, 24 ); + den_Q24 = silk_SMLAWB( SILK_FIX_CONST( 1.0, 24 ), coefs_Q24[ 0 ], lambda_Q16 ); + gain_Q16 = silk_DIV32_varQ( nom_Q16, den_Q24, 24 ); for( i = 0; i < order; i++ ) { - coefs_syn_Q24[ i ] = silk_SMULWW( gain_syn_Q16, coefs_syn_Q24[ i ] ); - coefs_ana_Q24[ i ] = silk_SMULWW( gain_ana_Q16, coefs_ana_Q24[ i ] ); + coefs_Q24[ i ] = silk_SMULWW( gain_Q16, coefs_Q24[ i ] ); } } silk_assert( 0 ); } +/* Disable MIPS version until it's updated. */ +#if 0 && defined(MIPSr1_ASM) +#include "mips/noise_shape_analysis_FIX_mipsr1.h" +#endif + /**************************************************************/ /* Compute noise shaping coefficients and initial gain values */ /**************************************************************/ +#ifndef OVERRIDE_silk_noise_shape_analysis_FIX void silk_noise_shape_analysis_FIX( silk_encoder_state_FIX *psEnc, /* I/O Encoder state FIX */ silk_encoder_control_FIX *psEncCtrl, /* I/O Encoder control FIX */ @@ -150,14 +146,13 @@ void silk_noise_shape_analysis_FIX( ) { silk_shape_state_FIX *psShapeSt = &psEnc->sShape; - opus_int k, i, nSamples, Qnrg, b_Q14, warping_Q16, scale = 0; - opus_int32 SNR_adj_dB_Q7, HarmBoost_Q16, HarmShapeGain_Q16, Tilt_Q16, tmp32; - opus_int32 nrg, pre_nrg_Q30, log_energy_Q7, log_energy_prev_Q7, energy_variation_Q7; - opus_int32 delta_Q16, BWExp1_Q16, BWExp2_Q16, gain_mult_Q16, gain_add_Q16, strength_Q16, b_Q8; + opus_int k, i, nSamples, nSegs, Qnrg, b_Q14, warping_Q16, scale = 0; + opus_int32 SNR_adj_dB_Q7, HarmShapeGain_Q16, Tilt_Q16, tmp32; + opus_int32 nrg, log_energy_Q7, log_energy_prev_Q7, energy_variation_Q7; + opus_int32 BWExp_Q16, gain_mult_Q16, gain_add_Q16, strength_Q16, b_Q8; opus_int32 auto_corr[ MAX_SHAPE_LPC_ORDER + 1 ]; opus_int32 refl_coef_Q16[ MAX_SHAPE_LPC_ORDER ]; - opus_int32 AR1_Q24[ MAX_SHAPE_LPC_ORDER ]; - opus_int32 AR2_Q24[ MAX_SHAPE_LPC_ORDER ]; + opus_int32 AR_Q24[ MAX_SHAPE_LPC_ORDER ]; VARDECL( opus_int16, x_windowed ); const opus_int16 *x_ptr, *pitch_res_ptr; SAVE_STACK; @@ -204,14 +199,14 @@ void silk_noise_shape_analysis_FIX( if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) { /* Initially set to 0; may be overruled in process_gains(..) */ psEnc->sCmn.indices.quantOffsetType = 0; - psEncCtrl->sparseness_Q8 = 0; } else { /* Sparseness measure, based on relative fluctuations of energy per 2 milliseconds */ nSamples = silk_LSHIFT( psEnc->sCmn.fs_kHz, 1 ); energy_variation_Q7 = 0; log_energy_prev_Q7 = 0; pitch_res_ptr = pitch_res; - for( k = 0; k < silk_SMULBB( SUB_FRAME_LENGTH_MS, psEnc->sCmn.nb_subfr ) / 2; k++ ) { + nSegs = silk_SMULBB( SUB_FRAME_LENGTH_MS, psEnc->sCmn.nb_subfr ) / 2; + for( k = 0; k < nSegs; k++ ) { silk_sum_sqr_shift( &nrg, &scale, pitch_res_ptr, nSamples ); nrg += silk_RSHIFT( nSamples, scale ); /* Q(-scale)*/ @@ -223,18 +218,12 @@ void silk_noise_shape_analysis_FIX( pitch_res_ptr += nSamples; } - psEncCtrl->sparseness_Q8 = silk_RSHIFT( silk_sigm_Q15( silk_SMULWB( energy_variation_Q7 - - SILK_FIX_CONST( 5.0, 7 ), SILK_FIX_CONST( 0.1, 16 ) ) ), 7 ); - /* Set quantization offset depending on sparseness measure */ - if( psEncCtrl->sparseness_Q8 > SILK_FIX_CONST( SPARSENESS_THRESHOLD_QNT_OFFSET, 8 ) ) { + if( energy_variation_Q7 > SILK_FIX_CONST( ENERGY_VARIATION_THRESHOLD_QNT_OFFSET, 7 ) * (nSegs-1) ) { psEnc->sCmn.indices.quantOffsetType = 0; } else { psEnc->sCmn.indices.quantOffsetType = 1; } - - /* Increase coding SNR for sparse signals */ - SNR_adj_dB_Q7 = silk_SMLAWB( SNR_adj_dB_Q7, SILK_FIX_CONST( SPARSE_SNR_INCR_dB, 15 ), psEncCtrl->sparseness_Q8 - SILK_FIX_CONST( 0.5, 8 ) ); } /*******************************/ @@ -242,14 +231,8 @@ void silk_noise_shape_analysis_FIX( /*******************************/ /* More BWE for signals with high prediction gain */ strength_Q16 = silk_SMULWB( psEncCtrl->predGain_Q16, SILK_FIX_CONST( FIND_PITCH_WHITE_NOISE_FRACTION, 16 ) ); - BWExp1_Q16 = BWExp2_Q16 = silk_DIV32_varQ( SILK_FIX_CONST( BANDWIDTH_EXPANSION, 16 ), + BWExp_Q16 = silk_DIV32_varQ( SILK_FIX_CONST( BANDWIDTH_EXPANSION, 16 ), silk_SMLAWW( SILK_FIX_CONST( 1.0, 16 ), strength_Q16, strength_Q16 ), 16 ); - delta_Q16 = silk_SMULWB( SILK_FIX_CONST( 1.0, 16 ) - silk_SMULBB( 3, psEncCtrl->coding_quality_Q14 ), - SILK_FIX_CONST( LOW_RATE_BANDWIDTH_EXPANSION_DELTA, 16 ) ); - BWExp1_Q16 = silk_SUB32( BWExp1_Q16, delta_Q16 ); - BWExp2_Q16 = silk_ADD32( BWExp2_Q16, delta_Q16 ); - /* BWExp1 will be applied after BWExp2, so make it relative */ - BWExp1_Q16 = silk_DIV32_16( silk_LSHIFT( BWExp1_Q16, 14 ), silk_RSHIFT( BWExp2_Q16, 2 ) ); if( psEnc->sCmn.warping_Q16 > 0 ) { /* Slightly more warping in analysis will move quantization noise up in frequency, where it's better masked */ @@ -279,7 +262,7 @@ void silk_noise_shape_analysis_FIX( if( psEnc->sCmn.warping_Q16 > 0 ) { /* Calculate warped auto correlation */ - silk_warped_autocorrelation_FIX( auto_corr, &scale, x_windowed, warping_Q16, psEnc->sCmn.shapeWinLength, psEnc->sCmn.shapingLPCOrder ); + silk_warped_autocorrelation_FIX( auto_corr, &scale, x_windowed, warping_Q16, psEnc->sCmn.shapeWinLength, psEnc->sCmn.shapingLPCOrder, arch ); } else { /* Calculate regular auto correlation */ silk_autocorr( auto_corr, &scale, x_windowed, psEnc->sCmn.shapeWinLength, psEnc->sCmn.shapingLPCOrder + 1, arch ); @@ -294,7 +277,7 @@ void silk_noise_shape_analysis_FIX( silk_assert( nrg >= 0 ); /* Convert reflection coefficients to prediction coefficients */ - silk_k2a_Q16( AR2_Q24, refl_coef_Q16, psEnc->sCmn.shapingLPCOrder ); + silk_k2a_Q16( AR_Q24, refl_coef_Q16, psEnc->sCmn.shapingLPCOrder ); Qnrg = -scale; /* range: -12...30*/ silk_assert( Qnrg >= -12 ); @@ -313,40 +296,34 @@ void silk_noise_shape_analysis_FIX( if( psEnc->sCmn.warping_Q16 > 0 ) { /* Adjust gain for warping */ - gain_mult_Q16 = warped_gain( AR2_Q24, warping_Q16, psEnc->sCmn.shapingLPCOrder ); - silk_assert( psEncCtrl->Gains_Q16[ k ] >= 0 ); - if ( silk_SMULWW( silk_RSHIFT_ROUND( psEncCtrl->Gains_Q16[ k ], 1 ), gain_mult_Q16 ) >= ( silk_int32_MAX >> 1 ) ) { - psEncCtrl->Gains_Q16[ k ] = silk_int32_MAX; + gain_mult_Q16 = warped_gain( AR_Q24, warping_Q16, psEnc->sCmn.shapingLPCOrder ); + silk_assert( psEncCtrl->Gains_Q16[ k ] > 0 ); + if( psEncCtrl->Gains_Q16[ k ] < SILK_FIX_CONST( 0.25, 16 ) ) { + psEncCtrl->Gains_Q16[ k ] = silk_SMULWW( psEncCtrl->Gains_Q16[ k ], gain_mult_Q16 ); } else { - psEncCtrl->Gains_Q16[ k ] = silk_SMULWW( psEncCtrl->Gains_Q16[ k ], gain_mult_Q16 ); + psEncCtrl->Gains_Q16[ k ] = silk_SMULWW( silk_RSHIFT_ROUND( psEncCtrl->Gains_Q16[ k ], 1 ), gain_mult_Q16 ); + if ( psEncCtrl->Gains_Q16[ k ] >= ( silk_int32_MAX >> 1 ) ) { + psEncCtrl->Gains_Q16[ k ] = silk_int32_MAX; + } else { + psEncCtrl->Gains_Q16[ k ] = silk_LSHIFT32( psEncCtrl->Gains_Q16[ k ], 1 ); + } } + silk_assert( psEncCtrl->Gains_Q16[ k ] > 0 ); } - /* Bandwidth expansion for synthesis filter shaping */ - silk_bwexpander_32( AR2_Q24, psEnc->sCmn.shapingLPCOrder, BWExp2_Q16 ); + /* Bandwidth expansion */ + silk_bwexpander_32( AR_Q24, psEnc->sCmn.shapingLPCOrder, BWExp_Q16 ); - /* Compute noise shaping filter coefficients */ - silk_memcpy( AR1_Q24, AR2_Q24, psEnc->sCmn.shapingLPCOrder * sizeof( opus_int32 ) ); + if( psEnc->sCmn.warping_Q16 > 0 ) { + /* Convert to monic warped prediction coefficients and limit absolute values */ + limit_warped_coefs( AR_Q24, warping_Q16, SILK_FIX_CONST( 3.999, 24 ), psEnc->sCmn.shapingLPCOrder ); - /* Bandwidth expansion for analysis filter shaping */ - silk_assert( BWExp1_Q16 <= SILK_FIX_CONST( 1.0, 16 ) ); - silk_bwexpander_32( AR1_Q24, psEnc->sCmn.shapingLPCOrder, BWExp1_Q16 ); - - /* Ratio of prediction gains, in energy domain */ - pre_nrg_Q30 = silk_LPC_inverse_pred_gain_Q24( AR2_Q24, psEnc->sCmn.shapingLPCOrder ); - nrg = silk_LPC_inverse_pred_gain_Q24( AR1_Q24, psEnc->sCmn.shapingLPCOrder ); - - /*psEncCtrl->GainsPre[ k ] = 1.0f - 0.7f * ( 1.0f - pre_nrg / nrg ) = 0.3f + 0.7f * pre_nrg / nrg;*/ - pre_nrg_Q30 = silk_LSHIFT32( silk_SMULWB( pre_nrg_Q30, SILK_FIX_CONST( 0.7, 15 ) ), 1 ); - psEncCtrl->GainsPre_Q14[ k ] = ( opus_int ) SILK_FIX_CONST( 0.3, 14 ) + silk_DIV32_varQ( pre_nrg_Q30, nrg, 14 ); - - /* Convert to monic warped prediction coefficients and limit absolute values */ - limit_warped_coefs( AR2_Q24, AR1_Q24, warping_Q16, SILK_FIX_CONST( 3.999, 24 ), psEnc->sCmn.shapingLPCOrder ); - - /* Convert from Q24 to Q13 and store in int16 */ - for( i = 0; i < psEnc->sCmn.shapingLPCOrder; i++ ) { - psEncCtrl->AR1_Q13[ k * MAX_SHAPE_LPC_ORDER + i ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( AR1_Q24[ i ], 11 ) ); - psEncCtrl->AR2_Q13[ k * MAX_SHAPE_LPC_ORDER + i ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( AR2_Q24[ i ], 11 ) ); + /* Convert from Q24 to Q13 and store in int16 */ + for( i = 0; i < psEnc->sCmn.shapingLPCOrder; i++ ) { + psEncCtrl->AR_Q13[ k * MAX_SHAPE_LPC_ORDER + i ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( AR_Q24[ i ], 11 ) ); + } + } else { + silk_LPC_fit( &psEncCtrl->AR_Q13[ k * MAX_SHAPE_LPC_ORDER ], AR_Q24, 13, 24, psEnc->sCmn.shapingLPCOrder ); } } @@ -363,11 +340,6 @@ void silk_noise_shape_analysis_FIX( psEncCtrl->Gains_Q16[ k ] = silk_ADD_POS_SAT32( psEncCtrl->Gains_Q16[ k ], gain_add_Q16 ); } - gain_mult_Q16 = SILK_FIX_CONST( 1.0, 16 ) + silk_RSHIFT_ROUND( silk_MLA( SILK_FIX_CONST( INPUT_TILT, 26 ), - psEncCtrl->coding_quality_Q14, SILK_FIX_CONST( HIGH_RATE_INPUT_TILT, 12 ) ), 10 ); - for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) { - psEncCtrl->GainsPre_Q14[ k ] = silk_SMULWB( gain_mult_Q16, psEncCtrl->GainsPre_Q14[ k ] ); - } /************************************************/ /* Control low-frequency shaping and noise tilt */ @@ -405,14 +377,6 @@ void silk_noise_shape_analysis_FIX( /****************************/ /* HARMONIC SHAPING CONTROL */ /****************************/ - /* Control boosting of harmonic frequencies */ - HarmBoost_Q16 = silk_SMULWB( silk_SMULWB( SILK_FIX_CONST( 1.0, 17 ) - silk_LSHIFT( psEncCtrl->coding_quality_Q14, 3 ), - psEnc->LTPCorr_Q15 ), SILK_FIX_CONST( LOW_RATE_HARMONIC_BOOST, 16 ) ); - - /* More harmonic boost for noisy input signals */ - HarmBoost_Q16 = silk_SMLAWB( HarmBoost_Q16, - SILK_FIX_CONST( 1.0, 16 ) - silk_LSHIFT( psEncCtrl->input_quality_Q14, 2 ), SILK_FIX_CONST( LOW_INPUT_QUALITY_HARMONIC_BOOST, 16 ) ); - if( USE_HARM_SHAPING && psEnc->sCmn.indices.signalType == TYPE_VOICED ) { /* More harmonic noise shaping for high bitrates or noisy input */ HarmShapeGain_Q16 = silk_SMLAWB( SILK_FIX_CONST( HARMONIC_SHAPING, 16 ), @@ -430,16 +394,14 @@ void silk_noise_shape_analysis_FIX( /* Smooth over subframes */ /*************************/ for( k = 0; k < MAX_NB_SUBFR; k++ ) { - psShapeSt->HarmBoost_smth_Q16 = - silk_SMLAWB( psShapeSt->HarmBoost_smth_Q16, HarmBoost_Q16 - psShapeSt->HarmBoost_smth_Q16, SILK_FIX_CONST( SUBFR_SMTH_COEF, 16 ) ); psShapeSt->HarmShapeGain_smth_Q16 = silk_SMLAWB( psShapeSt->HarmShapeGain_smth_Q16, HarmShapeGain_Q16 - psShapeSt->HarmShapeGain_smth_Q16, SILK_FIX_CONST( SUBFR_SMTH_COEF, 16 ) ); psShapeSt->Tilt_smth_Q16 = silk_SMLAWB( psShapeSt->Tilt_smth_Q16, Tilt_Q16 - psShapeSt->Tilt_smth_Q16, SILK_FIX_CONST( SUBFR_SMTH_COEF, 16 ) ); - psEncCtrl->HarmBoost_Q14[ k ] = ( opus_int )silk_RSHIFT_ROUND( psShapeSt->HarmBoost_smth_Q16, 2 ); psEncCtrl->HarmShapeGain_Q14[ k ] = ( opus_int )silk_RSHIFT_ROUND( psShapeSt->HarmShapeGain_smth_Q16, 2 ); psEncCtrl->Tilt_Q14[ k ] = ( opus_int )silk_RSHIFT_ROUND( psShapeSt->Tilt_smth_Q16, 2 ); } RESTORE_STACK; } +#endif /* OVERRIDE_silk_noise_shape_analysis_FIX */ diff --git a/code/opus-1.1.4/silk/fixed/pitch_analysis_core_FIX.c b/code/opus-1.1.4/silk/fixed/pitch_analysis_core_FIX.c index 1641a0fb..8df109e6 100644 --- a/code/opus-1.1.4/silk/fixed/pitch_analysis_core_FIX.c +++ b/code/opus-1.1.4/silk/fixed/pitch_analysis_core_FIX.c @@ -72,14 +72,15 @@ static void silk_P_Ana_calc_energy_st3( opus_int start_lag, /* I lag offset to search around */ opus_int sf_length, /* I length of one 5 ms subframe */ opus_int nb_subfr, /* I number of subframes */ - opus_int complexity /* I Complexity setting */ + opus_int complexity, /* I Complexity setting */ + int arch /* I Run-time architecture */ ); /*************************************************************/ /* FIXED POINT CORE PITCH ANALYSIS FUNCTION */ /*************************************************************/ opus_int silk_pitch_analysis_core( /* O Voicing estimate: 0 voiced, 1 unvoiced */ - const opus_int16 *frame, /* I Signal of length PE_FRAME_LENGTH_MS*Fs_kHz */ + const opus_int16 *frame_unscaled, /* I Signal of length PE_FRAME_LENGTH_MS*Fs_kHz */ opus_int *pitch_out, /* O 4 pitch lag values */ opus_int16 *lagIndex, /* O Lag Index */ opus_int8 *contourIndex, /* O Pitch contour Index */ @@ -93,16 +94,17 @@ opus_int silk_pitch_analysis_core( /* O Voicing estimate: 0 int arch /* I Run-time architecture */ ) { - VARDECL( opus_int16, frame_8kHz ); + VARDECL( opus_int16, frame_8kHz_buf ); VARDECL( opus_int16, frame_4kHz ); + VARDECL( opus_int16, frame_scaled ); opus_int32 filt_state[ 6 ]; - const opus_int16 *input_frame_ptr; + const opus_int16 *frame, *frame_8kHz; opus_int i, k, d, j; VARDECL( opus_int16, C ); VARDECL( opus_int32, xcorr32 ); const opus_int16 *target_ptr, *basis_ptr; - opus_int32 cross_corr, normalizer, energy, shift, energy_basis, energy_target; - opus_int d_srch[ PE_D_SRCH_LENGTH ], Cmax, length_d_srch, length_d_comp; + opus_int32 cross_corr, normalizer, energy, energy_basis, energy_target; + opus_int d_srch[ PE_D_SRCH_LENGTH ], Cmax, length_d_srch, length_d_comp, shift; VARDECL( opus_int16, d_comp ); opus_int32 sum, threshold, lag_counter; opus_int CBimax, CBimax_new, CBimax_old, lag, start_lag, end_lag, lag_new; @@ -118,6 +120,7 @@ opus_int silk_pitch_analysis_core( /* O Voicing estimate: 0 opus_int32 delta_lag_log2_sqr_Q7, lag_log2_Q7, prevLag_log2_Q7, prev_lag_bias_Q13; const opus_int8 *Lag_CB_ptr; SAVE_STACK; + /* Check for valid sampling frequency */ silk_assert( Fs_kHz == 8 || Fs_kHz == 12 || Fs_kHz == 16 ); @@ -136,17 +139,33 @@ opus_int silk_pitch_analysis_core( /* O Voicing estimate: 0 min_lag = PE_MIN_LAG_MS * Fs_kHz; max_lag = PE_MAX_LAG_MS * Fs_kHz - 1; + /* Downscale input if necessary */ + silk_sum_sqr_shift( &energy, &shift, frame_unscaled, frame_length ); + shift += 3 - silk_CLZ32( energy ); /* at least two bits headroom */ + ALLOC( frame_scaled, frame_length, opus_int16 ); + if( shift > 0 ) { + shift = silk_RSHIFT( shift + 1, 1 ); + for( i = 0; i < frame_length; i++ ) { + frame_scaled[ i ] = silk_RSHIFT( frame_unscaled[ i ], shift ); + } + frame = frame_scaled; + } else { + frame = frame_unscaled; + } + + ALLOC( frame_8kHz_buf, ( Fs_kHz == 8 ) ? 1 : frame_length_8kHz, opus_int16 ); /* Resample from input sampled at Fs_kHz to 8 kHz */ - ALLOC( frame_8kHz, frame_length_8kHz, opus_int16 ); if( Fs_kHz == 16 ) { silk_memset( filt_state, 0, 2 * sizeof( opus_int32 ) ); - silk_resampler_down2( filt_state, frame_8kHz, frame, frame_length ); + silk_resampler_down2( filt_state, frame_8kHz_buf, frame, frame_length ); + frame_8kHz = frame_8kHz_buf; } else if( Fs_kHz == 12 ) { silk_memset( filt_state, 0, 6 * sizeof( opus_int32 ) ); - silk_resampler_down2_3( filt_state, frame_8kHz, frame, frame_length ); + silk_resampler_down2_3( filt_state, frame_8kHz_buf, frame, frame_length ); + frame_8kHz = frame_8kHz_buf; } else { silk_assert( Fs_kHz == 8 ); - silk_memcpy( frame_8kHz, frame, frame_length_8kHz * sizeof(opus_int16) ); + frame_8kHz = frame; } /* Decimate again to 4 kHz */ @@ -159,19 +178,6 @@ opus_int silk_pitch_analysis_core( /* O Voicing estimate: 0 frame_4kHz[ i ] = silk_ADD_SAT16( frame_4kHz[ i ], frame_4kHz[ i - 1 ] ); } - /******************************************************************************* - ** Scale 4 kHz signal down to prevent correlations measures from overflowing - ** find scaling as max scaling for each 8kHz(?) subframe - *******************************************************************************/ - - /* Inner product is calculated with different lengths, so scale for the worst case */ - silk_sum_sqr_shift( &energy, &shift, frame_4kHz, frame_length_4kHz ); - if( shift > 0 ) { - shift = silk_RSHIFT( shift, 1 ); - for( i = 0; i < frame_length_4kHz; i++ ) { - frame_4kHz[ i ] = silk_RSHIFT( frame_4kHz[ i ], shift ); - } - } /****************************************************************************** * FIRST STAGE, operating in 4 khz @@ -195,8 +201,8 @@ opus_int silk_pitch_analysis_core( /* O Voicing estimate: 0 /* Calculate first vector products before loop */ cross_corr = xcorr32[ MAX_LAG_4KHZ - MIN_LAG_4KHZ ]; - normalizer = silk_inner_prod_aligned( target_ptr, target_ptr, SF_LENGTH_8KHZ ); - normalizer = silk_ADD32( normalizer, silk_inner_prod_aligned( basis_ptr, basis_ptr, SF_LENGTH_8KHZ ) ); + normalizer = silk_inner_prod_aligned( target_ptr, target_ptr, SF_LENGTH_8KHZ, arch ); + normalizer = silk_ADD32( normalizer, silk_inner_prod_aligned( basis_ptr, basis_ptr, SF_LENGTH_8KHZ, arch ) ); normalizer = silk_ADD32( normalizer, silk_SMULBB( SF_LENGTH_8KHZ, 4000 ) ); matrix_ptr( C, k, 0, CSTRIDE_4KHZ ) = @@ -310,18 +316,6 @@ opus_int silk_pitch_analysis_core( /* O Voicing estimate: 0 ** SECOND STAGE, operating at 8 kHz, on lag sections with high correlation *************************************************************************************/ - /****************************************************************************** - ** Scale signal down to avoid correlations measures from overflowing - *******************************************************************************/ - /* find scaling as max scaling for each subframe */ - silk_sum_sqr_shift( &energy, &shift, frame_8kHz, frame_length_8kHz ); - if( shift > 0 ) { - shift = silk_RSHIFT( shift, 1 ); - for( i = 0; i < frame_length_8kHz; i++ ) { - frame_8kHz[ i ] = silk_RSHIFT( frame_8kHz[ i ], shift ); - } - } - /********************************************************************************* * Find energy of each subframe projected onto its history, for a range of delays *********************************************************************************/ @@ -334,7 +328,7 @@ opus_int silk_pitch_analysis_core( /* O Voicing estimate: 0 silk_assert( target_ptr >= frame_8kHz ); silk_assert( target_ptr + SF_LENGTH_8KHZ <= frame_8kHz + frame_length_8kHz ); - energy_target = silk_ADD32( silk_inner_prod_aligned( target_ptr, target_ptr, SF_LENGTH_8KHZ ), 1 ); + energy_target = silk_ADD32( silk_inner_prod_aligned( target_ptr, target_ptr, SF_LENGTH_8KHZ, arch ), 1 ); for( j = 0; j < length_d_comp; j++ ) { d = d_comp[ j ]; basis_ptr = target_ptr - d; @@ -343,9 +337,9 @@ opus_int silk_pitch_analysis_core( /* O Voicing estimate: 0 silk_assert( basis_ptr >= frame_8kHz ); silk_assert( basis_ptr + SF_LENGTH_8KHZ <= frame_8kHz + frame_length_8kHz ); - cross_corr = silk_inner_prod_aligned( target_ptr, basis_ptr, SF_LENGTH_8KHZ ); + cross_corr = silk_inner_prod_aligned( target_ptr, basis_ptr, SF_LENGTH_8KHZ, arch ); if( cross_corr > 0 ) { - energy_basis = silk_inner_prod_aligned( basis_ptr, basis_ptr, SF_LENGTH_8KHZ ); + energy_basis = silk_inner_prod_aligned( basis_ptr, basis_ptr, SF_LENGTH_8KHZ, arch ); matrix_ptr( C, k, d - ( MIN_LAG_8KHZ - 2 ), CSTRIDE_8KHZ ) = (opus_int16)silk_DIV32_varQ( cross_corr, silk_ADD32( energy_target, @@ -461,24 +455,6 @@ opus_int silk_pitch_analysis_core( /* O Voicing estimate: 0 silk_assert( *LTPCorr_Q15 >= 0 ); if( Fs_kHz > 8 ) { - VARDECL( opus_int16, scratch_mem ); - /***************************************************************************/ - /* Scale input signal down to avoid correlations measures from overflowing */ - /***************************************************************************/ - /* find scaling as max scaling for each subframe */ - silk_sum_sqr_shift( &energy, &shift, frame, frame_length ); - ALLOC( scratch_mem, shift > 0 ? frame_length : ALLOC_NONE, opus_int16 ); - if( shift > 0 ) { - /* Move signal to scratch mem because the input signal should be unchanged */ - shift = silk_RSHIFT( shift, 1 ); - for( i = 0; i < frame_length; i++ ) { - scratch_mem[ i ] = silk_RSHIFT( frame[ i ], shift ); - } - input_frame_ptr = scratch_mem; - } else { - input_frame_ptr = frame; - } - /* Search in original signal */ CBimax_old = CBimax; @@ -518,15 +494,15 @@ opus_int silk_pitch_analysis_core( /* O Voicing estimate: 0 /* Calculate the correlations and energies needed in stage 3 */ ALLOC( energies_st3, nb_subfr * nb_cbk_search, silk_pe_stage3_vals ); ALLOC( cross_corr_st3, nb_subfr * nb_cbk_search, silk_pe_stage3_vals ); - silk_P_Ana_calc_corr_st3( cross_corr_st3, input_frame_ptr, start_lag, sf_length, nb_subfr, complexity, arch ); - silk_P_Ana_calc_energy_st3( energies_st3, input_frame_ptr, start_lag, sf_length, nb_subfr, complexity ); + silk_P_Ana_calc_corr_st3( cross_corr_st3, frame, start_lag, sf_length, nb_subfr, complexity, arch ); + silk_P_Ana_calc_energy_st3( energies_st3, frame, start_lag, sf_length, nb_subfr, complexity, arch ); lag_counter = 0; silk_assert( lag == silk_SAT16( lag ) ); contour_bias_Q15 = silk_DIV32_16( SILK_FIX_CONST( PE_FLATCONTOUR_BIAS, 15 ), lag ); - target_ptr = &input_frame_ptr[ PE_LTP_MEM_LENGTH_MS * Fs_kHz ]; - energy_target = silk_ADD32( silk_inner_prod_aligned( target_ptr, target_ptr, nb_subfr * sf_length ), 1 ); + target_ptr = &frame[ PE_LTP_MEM_LENGTH_MS * Fs_kHz ]; + energy_target = silk_ADD32( silk_inner_prod_aligned( target_ptr, target_ptr, nb_subfr * sf_length, arch ), 1 ); for( d = start_lag; d <= end_lag; d++ ) { for( j = 0; j < nb_cbk_search; j++ ) { cross_corr = 0; @@ -671,7 +647,8 @@ static void silk_P_Ana_calc_energy_st3( opus_int start_lag, /* I lag offset to search around */ opus_int sf_length, /* I length of one 5 ms subframe */ opus_int nb_subfr, /* I number of subframes */ - opus_int complexity /* I Complexity setting */ + opus_int complexity, /* I Complexity setting */ + int arch /* I Run-time architecture */ ) { const opus_int16 *target_ptr, *basis_ptr; @@ -705,7 +682,7 @@ static void silk_P_Ana_calc_energy_st3( /* Calculate the energy for first lag */ basis_ptr = target_ptr - ( start_lag + matrix_ptr( Lag_range_ptr, k, 0, 2 ) ); - energy = silk_inner_prod_aligned( basis_ptr, basis_ptr, sf_length ); + energy = silk_inner_prod_aligned( basis_ptr, basis_ptr, sf_length, arch ); silk_assert( energy >= 0 ); scratch_mem[ lag_counter ] = energy; lag_counter++; diff --git a/code/opus-1.1.4/silk/fixed/prefilter_FIX.c b/code/opus-1.1.4/silk/fixed/prefilter_FIX.c deleted file mode 100644 index d381730c..00000000 --- a/code/opus-1.1.4/silk/fixed/prefilter_FIX.c +++ /dev/null @@ -1,209 +0,0 @@ -/*********************************************************************** -Copyright (c) 2006-2011, Skype Limited. All rights reserved. -Redistribution and use in source and binary forms, with or without -modification, are permitted provided that the following conditions -are met: -- Redistributions of source code must retain the above copyright notice, -this list of conditions and the following disclaimer. -- Redistributions in binary form must reproduce the above copyright -notice, this list of conditions and the following disclaimer in the -documentation and/or other materials provided with the distribution. -- Neither the name of Internet Society, IETF or IETF Trust, nor the -names of specific contributors, may be used to endorse or promote -products derived from this software without specific prior written -permission. -THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" -AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE -IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE -ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE -LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR -CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF -SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS -INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN -CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) -ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE -POSSIBILITY OF SUCH DAMAGE. -***********************************************************************/ - -#ifdef HAVE_CONFIG_H -#include "config.h" -#endif - -#include "main_FIX.h" -#include "stack_alloc.h" -#include "tuning_parameters.h" - -/* Prefilter for finding Quantizer input signal */ -static OPUS_INLINE void silk_prefilt_FIX( - silk_prefilter_state_FIX *P, /* I/O state */ - opus_int32 st_res_Q12[], /* I short term residual signal */ - opus_int32 xw_Q3[], /* O prefiltered signal */ - opus_int32 HarmShapeFIRPacked_Q12, /* I Harmonic shaping coeficients */ - opus_int Tilt_Q14, /* I Tilt shaping coeficient */ - opus_int32 LF_shp_Q14, /* I Low-frequancy shaping coeficients */ - opus_int lag, /* I Lag for harmonic shaping */ - opus_int length /* I Length of signals */ -); - -void silk_warped_LPC_analysis_filter_FIX( - opus_int32 state[], /* I/O State [order + 1] */ - opus_int32 res_Q2[], /* O Residual signal [length] */ - const opus_int16 coef_Q13[], /* I Coefficients [order] */ - const opus_int16 input[], /* I Input signal [length] */ - const opus_int16 lambda_Q16, /* I Warping factor */ - const opus_int length, /* I Length of input signal */ - const opus_int order /* I Filter order (even) */ -) -{ - opus_int n, i; - opus_int32 acc_Q11, tmp1, tmp2; - - /* Order must be even */ - silk_assert( ( order & 1 ) == 0 ); - - for( n = 0; n < length; n++ ) { - /* Output of lowpass section */ - tmp2 = silk_SMLAWB( state[ 0 ], state[ 1 ], lambda_Q16 ); - state[ 0 ] = silk_LSHIFT( input[ n ], 14 ); - /* Output of allpass section */ - tmp1 = silk_SMLAWB( state[ 1 ], state[ 2 ] - tmp2, lambda_Q16 ); - state[ 1 ] = tmp2; - acc_Q11 = silk_RSHIFT( order, 1 ); - acc_Q11 = silk_SMLAWB( acc_Q11, tmp2, coef_Q13[ 0 ] ); - /* Loop over allpass sections */ - for( i = 2; i < order; i += 2 ) { - /* Output of allpass section */ - tmp2 = silk_SMLAWB( state[ i ], state[ i + 1 ] - tmp1, lambda_Q16 ); - state[ i ] = tmp1; - acc_Q11 = silk_SMLAWB( acc_Q11, tmp1, coef_Q13[ i - 1 ] ); - /* Output of allpass section */ - tmp1 = silk_SMLAWB( state[ i + 1 ], state[ i + 2 ] - tmp2, lambda_Q16 ); - state[ i + 1 ] = tmp2; - acc_Q11 = silk_SMLAWB( acc_Q11, tmp2, coef_Q13[ i ] ); - } - state[ order ] = tmp1; - acc_Q11 = silk_SMLAWB( acc_Q11, tmp1, coef_Q13[ order - 1 ] ); - res_Q2[ n ] = silk_LSHIFT( (opus_int32)input[ n ], 2 ) - silk_RSHIFT_ROUND( acc_Q11, 9 ); - } -} - -void silk_prefilter_FIX( - silk_encoder_state_FIX *psEnc, /* I/O Encoder state */ - const silk_encoder_control_FIX *psEncCtrl, /* I Encoder control */ - opus_int32 xw_Q3[], /* O Weighted signal */ - const opus_int16 x[] /* I Speech signal */ -) -{ - silk_prefilter_state_FIX *P = &psEnc->sPrefilt; - opus_int j, k, lag; - opus_int32 tmp_32; - const opus_int16 *AR1_shp_Q13; - const opus_int16 *px; - opus_int32 *pxw_Q3; - opus_int HarmShapeGain_Q12, Tilt_Q14; - opus_int32 HarmShapeFIRPacked_Q12, LF_shp_Q14; - VARDECL( opus_int32, x_filt_Q12 ); - VARDECL( opus_int32, st_res_Q2 ); - opus_int16 B_Q10[ 2 ]; - SAVE_STACK; - - /* Set up pointers */ - px = x; - pxw_Q3 = xw_Q3; - lag = P->lagPrev; - ALLOC( x_filt_Q12, psEnc->sCmn.subfr_length, opus_int32 ); - ALLOC( st_res_Q2, psEnc->sCmn.subfr_length, opus_int32 ); - for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) { - /* Update Variables that change per sub frame */ - if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) { - lag = psEncCtrl->pitchL[ k ]; - } - - /* Noise shape parameters */ - HarmShapeGain_Q12 = silk_SMULWB( (opus_int32)psEncCtrl->HarmShapeGain_Q14[ k ], 16384 - psEncCtrl->HarmBoost_Q14[ k ] ); - silk_assert( HarmShapeGain_Q12 >= 0 ); - HarmShapeFIRPacked_Q12 = silk_RSHIFT( HarmShapeGain_Q12, 2 ); - HarmShapeFIRPacked_Q12 |= silk_LSHIFT( (opus_int32)silk_RSHIFT( HarmShapeGain_Q12, 1 ), 16 ); - Tilt_Q14 = psEncCtrl->Tilt_Q14[ k ]; - LF_shp_Q14 = psEncCtrl->LF_shp_Q14[ k ]; - AR1_shp_Q13 = &psEncCtrl->AR1_Q13[ k * MAX_SHAPE_LPC_ORDER ]; - - /* Short term FIR filtering*/ - silk_warped_LPC_analysis_filter_FIX( P->sAR_shp, st_res_Q2, AR1_shp_Q13, px, - psEnc->sCmn.warping_Q16, psEnc->sCmn.subfr_length, psEnc->sCmn.shapingLPCOrder ); - - /* Reduce (mainly) low frequencies during harmonic emphasis */ - B_Q10[ 0 ] = silk_RSHIFT_ROUND( psEncCtrl->GainsPre_Q14[ k ], 4 ); - tmp_32 = silk_SMLABB( SILK_FIX_CONST( INPUT_TILT, 26 ), psEncCtrl->HarmBoost_Q14[ k ], HarmShapeGain_Q12 ); /* Q26 */ - tmp_32 = silk_SMLABB( tmp_32, psEncCtrl->coding_quality_Q14, SILK_FIX_CONST( HIGH_RATE_INPUT_TILT, 12 ) ); /* Q26 */ - tmp_32 = silk_SMULWB( tmp_32, -psEncCtrl->GainsPre_Q14[ k ] ); /* Q24 */ - tmp_32 = silk_RSHIFT_ROUND( tmp_32, 14 ); /* Q10 */ - B_Q10[ 1 ]= silk_SAT16( tmp_32 ); - x_filt_Q12[ 0 ] = silk_MLA( silk_MUL( st_res_Q2[ 0 ], B_Q10[ 0 ] ), P->sHarmHP_Q2, B_Q10[ 1 ] ); - for( j = 1; j < psEnc->sCmn.subfr_length; j++ ) { - x_filt_Q12[ j ] = silk_MLA( silk_MUL( st_res_Q2[ j ], B_Q10[ 0 ] ), st_res_Q2[ j - 1 ], B_Q10[ 1 ] ); - } - P->sHarmHP_Q2 = st_res_Q2[ psEnc->sCmn.subfr_length - 1 ]; - - silk_prefilt_FIX( P, x_filt_Q12, pxw_Q3, HarmShapeFIRPacked_Q12, Tilt_Q14, LF_shp_Q14, lag, psEnc->sCmn.subfr_length ); - - px += psEnc->sCmn.subfr_length; - pxw_Q3 += psEnc->sCmn.subfr_length; - } - - P->lagPrev = psEncCtrl->pitchL[ psEnc->sCmn.nb_subfr - 1 ]; - RESTORE_STACK; -} - -/* Prefilter for finding Quantizer input signal */ -static OPUS_INLINE void silk_prefilt_FIX( - silk_prefilter_state_FIX *P, /* I/O state */ - opus_int32 st_res_Q12[], /* I short term residual signal */ - opus_int32 xw_Q3[], /* O prefiltered signal */ - opus_int32 HarmShapeFIRPacked_Q12, /* I Harmonic shaping coeficients */ - opus_int Tilt_Q14, /* I Tilt shaping coeficient */ - opus_int32 LF_shp_Q14, /* I Low-frequancy shaping coeficients */ - opus_int lag, /* I Lag for harmonic shaping */ - opus_int length /* I Length of signals */ -) -{ - opus_int i, idx, LTP_shp_buf_idx; - opus_int32 n_LTP_Q12, n_Tilt_Q10, n_LF_Q10; - opus_int32 sLF_MA_shp_Q12, sLF_AR_shp_Q12; - opus_int16 *LTP_shp_buf; - - /* To speed up use temp variables instead of using the struct */ - LTP_shp_buf = P->sLTP_shp; - LTP_shp_buf_idx = P->sLTP_shp_buf_idx; - sLF_AR_shp_Q12 = P->sLF_AR_shp_Q12; - sLF_MA_shp_Q12 = P->sLF_MA_shp_Q12; - - for( i = 0; i < length; i++ ) { - if( lag > 0 ) { - /* unrolled loop */ - silk_assert( HARM_SHAPE_FIR_TAPS == 3 ); - idx = lag + LTP_shp_buf_idx; - n_LTP_Q12 = silk_SMULBB( LTP_shp_buf[ ( idx - HARM_SHAPE_FIR_TAPS / 2 - 1) & LTP_MASK ], HarmShapeFIRPacked_Q12 ); - n_LTP_Q12 = silk_SMLABT( n_LTP_Q12, LTP_shp_buf[ ( idx - HARM_SHAPE_FIR_TAPS / 2 ) & LTP_MASK ], HarmShapeFIRPacked_Q12 ); - n_LTP_Q12 = silk_SMLABB( n_LTP_Q12, LTP_shp_buf[ ( idx - HARM_SHAPE_FIR_TAPS / 2 + 1) & LTP_MASK ], HarmShapeFIRPacked_Q12 ); - } else { - n_LTP_Q12 = 0; - } - - n_Tilt_Q10 = silk_SMULWB( sLF_AR_shp_Q12, Tilt_Q14 ); - n_LF_Q10 = silk_SMLAWB( silk_SMULWT( sLF_AR_shp_Q12, LF_shp_Q14 ), sLF_MA_shp_Q12, LF_shp_Q14 ); - - sLF_AR_shp_Q12 = silk_SUB32( st_res_Q12[ i ], silk_LSHIFT( n_Tilt_Q10, 2 ) ); - sLF_MA_shp_Q12 = silk_SUB32( sLF_AR_shp_Q12, silk_LSHIFT( n_LF_Q10, 2 ) ); - - LTP_shp_buf_idx = ( LTP_shp_buf_idx - 1 ) & LTP_MASK; - LTP_shp_buf[ LTP_shp_buf_idx ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( sLF_MA_shp_Q12, 12 ) ); - - xw_Q3[i] = silk_RSHIFT_ROUND( silk_SUB32( sLF_MA_shp_Q12, n_LTP_Q12 ), 9 ); - } - - /* Copy temp variable back to state */ - P->sLF_AR_shp_Q12 = sLF_AR_shp_Q12; - P->sLF_MA_shp_Q12 = sLF_MA_shp_Q12; - P->sLTP_shp_buf_idx = LTP_shp_buf_idx; -} diff --git a/code/opus-1.1.4/silk/fixed/residual_energy_FIX.c b/code/opus-1.1.4/silk/fixed/residual_energy_FIX.c index 105ae318..41f74778 100644 --- a/code/opus-1.1.4/silk/fixed/residual_energy_FIX.c +++ b/code/opus-1.1.4/silk/fixed/residual_energy_FIX.c @@ -42,7 +42,8 @@ void silk_residual_energy_FIX( const opus_int32 gains[ MAX_NB_SUBFR ], /* I Quantization gains */ const opus_int subfr_length, /* I Subframe length */ const opus_int nb_subfr, /* I Number of subframes */ - const opus_int LPC_order /* I LPC order */ + const opus_int LPC_order, /* I LPC order */ + int arch /* I Run-time architecture */ ) { opus_int offset, i, j, rshift, lz1, lz2; @@ -60,7 +61,7 @@ void silk_residual_energy_FIX( silk_assert( ( nb_subfr >> 1 ) * ( MAX_NB_SUBFR >> 1 ) == nb_subfr ); for( i = 0; i < nb_subfr >> 1; i++ ) { /* Calculate half frame LPC residual signal including preceding samples */ - silk_LPC_analysis_filter( LPC_res, x_ptr, a_Q12[ i ], ( MAX_NB_SUBFR >> 1 ) * offset, LPC_order ); + silk_LPC_analysis_filter( LPC_res, x_ptr, a_Q12[ i ], ( MAX_NB_SUBFR >> 1 ) * offset, LPC_order, arch ); /* Point to first subframe of the just calculated LPC residual signal */ LPC_res_ptr = LPC_res + LPC_order; diff --git a/code/opus-1.1.4/silk/fixed/schur64_FIX.c b/code/opus-1.1.4/silk/fixed/schur64_FIX.c index 764a10ef..b2cb12d9 100644 --- a/code/opus-1.1.4/silk/fixed/schur64_FIX.c +++ b/code/opus-1.1.4/silk/fixed/schur64_FIX.c @@ -43,7 +43,7 @@ opus_int32 silk_schur64( /* O returns residual ene opus_int32 C[ SILK_MAX_ORDER_LPC + 1 ][ 2 ]; opus_int32 Ctmp1_Q30, Ctmp2_Q30, rc_tmp_Q31; - silk_assert( order==6||order==8||order==10||order==12||order==14||order==16 ); + silk_assert( order >= 0 && order <= SILK_MAX_ORDER_LPC ); /* Check for invalid input */ if( c[ 0 ] <= 0 ) { @@ -51,9 +51,10 @@ opus_int32 silk_schur64( /* O returns residual ene return 0; } - for( k = 0; k < order + 1; k++ ) { + k = 0; + do { C[ k ][ 0 ] = C[ k ][ 1 ] = c[ k ]; - } + } while( ++k <= order ); for( k = 0; k < order; k++ ) { /* Check that we won't be getting an unstable rc, otherwise stop here. */ diff --git a/code/opus-1.1.4/silk/fixed/schur_FIX.c b/code/opus-1.1.4/silk/fixed/schur_FIX.c index c4c0ef23..59d44a6f 100644 --- a/code/opus-1.1.4/silk/fixed/schur_FIX.c +++ b/code/opus-1.1.4/silk/fixed/schur_FIX.c @@ -43,28 +43,29 @@ opus_int32 silk_schur( /* O Returns residual ene opus_int32 C[ SILK_MAX_ORDER_LPC + 1 ][ 2 ]; opus_int32 Ctmp1, Ctmp2, rc_tmp_Q15; - silk_assert( order==6||order==8||order==10||order==12||order==14||order==16 ); + silk_assert( order >= 0 && order <= SILK_MAX_ORDER_LPC ); /* Get number of leading zeros */ lz = silk_CLZ32( c[ 0 ] ); /* Copy correlations and adjust level to Q30 */ + k = 0; if( lz < 2 ) { /* lz must be 1, so shift one to the right */ - for( k = 0; k < order + 1; k++ ) { + do { C[ k ][ 0 ] = C[ k ][ 1 ] = silk_RSHIFT( c[ k ], 1 ); - } + } while( ++k <= order ); } else if( lz > 2 ) { /* Shift to the left */ lz -= 2; - for( k = 0; k < order + 1; k++ ) { + do { C[ k ][ 0 ] = C[ k ][ 1 ] = silk_LSHIFT( c[ k ], lz ); - } + } while( ++k <= order ); } else { /* No need to shift */ - for( k = 0; k < order + 1; k++ ) { + do { C[ k ][ 0 ] = C[ k ][ 1 ] = c[ k ]; - } + } while( ++k <= order ); } for( k = 0; k < order; k++ ) { diff --git a/code/opus-1.1.4/silk/fixed/solve_LS_FIX.c b/code/opus-1.1.4/silk/fixed/solve_LS_FIX.c deleted file mode 100644 index 51d7d49d..00000000 --- a/code/opus-1.1.4/silk/fixed/solve_LS_FIX.c +++ /dev/null @@ -1,249 +0,0 @@ -/*********************************************************************** -Copyright (c) 2006-2011, Skype Limited. All rights reserved. -Redistribution and use in source and binary forms, with or without -modification, are permitted provided that the following conditions -are met: -- Redistributions of source code must retain the above copyright notice, -this list of conditions and the following disclaimer. -- Redistributions in binary form must reproduce the above copyright -notice, this list of conditions and the following disclaimer in the -documentation and/or other materials provided with the distribution. -- Neither the name of Internet Society, IETF or IETF Trust, nor the -names of specific contributors, may be used to endorse or promote -products derived from this software without specific prior written -permission. -THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" -AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE -IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE -ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE -LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR -CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF -SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS -INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN -CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) -ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE -POSSIBILITY OF SUCH DAMAGE. -***********************************************************************/ - -#ifdef HAVE_CONFIG_H -#include "config.h" -#endif - -#include "main_FIX.h" -#include "stack_alloc.h" -#include "tuning_parameters.h" - -/*****************************/ -/* Internal function headers */ -/*****************************/ - -typedef struct { - opus_int32 Q36_part; - opus_int32 Q48_part; -} inv_D_t; - -/* Factorize square matrix A into LDL form */ -static OPUS_INLINE void silk_LDL_factorize_FIX( - opus_int32 *A, /* I/O Pointer to Symetric Square Matrix */ - opus_int M, /* I Size of Matrix */ - opus_int32 *L_Q16, /* I/O Pointer to Square Upper triangular Matrix */ - inv_D_t *inv_D /* I/O Pointer to vector holding inverted diagonal elements of D */ -); - -/* Solve Lx = b, when L is lower triangular and has ones on the diagonal */ -static OPUS_INLINE void silk_LS_SolveFirst_FIX( - const opus_int32 *L_Q16, /* I Pointer to Lower Triangular Matrix */ - opus_int M, /* I Dim of Matrix equation */ - const opus_int32 *b, /* I b Vector */ - opus_int32 *x_Q16 /* O x Vector */ -); - -/* Solve L^t*x = b, where L is lower triangular with ones on the diagonal */ -static OPUS_INLINE void silk_LS_SolveLast_FIX( - const opus_int32 *L_Q16, /* I Pointer to Lower Triangular Matrix */ - const opus_int M, /* I Dim of Matrix equation */ - const opus_int32 *b, /* I b Vector */ - opus_int32 *x_Q16 /* O x Vector */ -); - -static OPUS_INLINE void silk_LS_divide_Q16_FIX( - opus_int32 T[], /* I/O Numenator vector */ - inv_D_t *inv_D, /* I 1 / D vector */ - opus_int M /* I dimension */ -); - -/* Solves Ax = b, assuming A is symmetric */ -void silk_solve_LDL_FIX( - opus_int32 *A, /* I Pointer to symetric square matrix A */ - opus_int M, /* I Size of matrix */ - const opus_int32 *b, /* I Pointer to b vector */ - opus_int32 *x_Q16 /* O Pointer to x solution vector */ -) -{ - VARDECL( opus_int32, L_Q16 ); - opus_int32 Y[ MAX_MATRIX_SIZE ]; - inv_D_t inv_D[ MAX_MATRIX_SIZE ]; - SAVE_STACK; - - silk_assert( M <= MAX_MATRIX_SIZE ); - ALLOC( L_Q16, M * M, opus_int32 ); - - /*************************************************** - Factorize A by LDL such that A = L*D*L', - where L is lower triangular with ones on diagonal - ****************************************************/ - silk_LDL_factorize_FIX( A, M, L_Q16, inv_D ); - - /**************************************************** - * substitute D*L'*x = Y. ie: - L*D*L'*x = b => L*Y = b <=> Y = inv(L)*b - ******************************************************/ - silk_LS_SolveFirst_FIX( L_Q16, M, b, Y ); - - /**************************************************** - D*L'*x = Y <=> L'*x = inv(D)*Y, because D is - diagonal just multiply with 1/d_i - ****************************************************/ - silk_LS_divide_Q16_FIX( Y, inv_D, M ); - - /**************************************************** - x = inv(L') * inv(D) * Y - *****************************************************/ - silk_LS_SolveLast_FIX( L_Q16, M, Y, x_Q16 ); - RESTORE_STACK; -} - -static OPUS_INLINE void silk_LDL_factorize_FIX( - opus_int32 *A, /* I/O Pointer to Symetric Square Matrix */ - opus_int M, /* I Size of Matrix */ - opus_int32 *L_Q16, /* I/O Pointer to Square Upper triangular Matrix */ - inv_D_t *inv_D /* I/O Pointer to vector holding inverted diagonal elements of D */ -) -{ - opus_int i, j, k, status, loop_count; - const opus_int32 *ptr1, *ptr2; - opus_int32 diag_min_value, tmp_32, err; - opus_int32 v_Q0[ MAX_MATRIX_SIZE ], D_Q0[ MAX_MATRIX_SIZE ]; - opus_int32 one_div_diag_Q36, one_div_diag_Q40, one_div_diag_Q48; - - silk_assert( M <= MAX_MATRIX_SIZE ); - - status = 1; - diag_min_value = silk_max_32( silk_SMMUL( silk_ADD_SAT32( A[ 0 ], A[ silk_SMULBB( M, M ) - 1 ] ), SILK_FIX_CONST( FIND_LTP_COND_FAC, 31 ) ), 1 << 9 ); - for( loop_count = 0; loop_count < M && status == 1; loop_count++ ) { - status = 0; - for( j = 0; j < M; j++ ) { - ptr1 = matrix_adr( L_Q16, j, 0, M ); - tmp_32 = 0; - for( i = 0; i < j; i++ ) { - v_Q0[ i ] = silk_SMULWW( D_Q0[ i ], ptr1[ i ] ); /* Q0 */ - tmp_32 = silk_SMLAWW( tmp_32, v_Q0[ i ], ptr1[ i ] ); /* Q0 */ - } - tmp_32 = silk_SUB32( matrix_ptr( A, j, j, M ), tmp_32 ); - - if( tmp_32 < diag_min_value ) { - tmp_32 = silk_SUB32( silk_SMULBB( loop_count + 1, diag_min_value ), tmp_32 ); - /* Matrix not positive semi-definite, or ill conditioned */ - for( i = 0; i < M; i++ ) { - matrix_ptr( A, i, i, M ) = silk_ADD32( matrix_ptr( A, i, i, M ), tmp_32 ); - } - status = 1; - break; - } - D_Q0[ j ] = tmp_32; /* always < max(Correlation) */ - - /* two-step division */ - one_div_diag_Q36 = silk_INVERSE32_varQ( tmp_32, 36 ); /* Q36 */ - one_div_diag_Q40 = silk_LSHIFT( one_div_diag_Q36, 4 ); /* Q40 */ - err = silk_SUB32( (opus_int32)1 << 24, silk_SMULWW( tmp_32, one_div_diag_Q40 ) ); /* Q24 */ - one_div_diag_Q48 = silk_SMULWW( err, one_div_diag_Q40 ); /* Q48 */ - - /* Save 1/Ds */ - inv_D[ j ].Q36_part = one_div_diag_Q36; - inv_D[ j ].Q48_part = one_div_diag_Q48; - - matrix_ptr( L_Q16, j, j, M ) = 65536; /* 1.0 in Q16 */ - ptr1 = matrix_adr( A, j, 0, M ); - ptr2 = matrix_adr( L_Q16, j + 1, 0, M ); - for( i = j + 1; i < M; i++ ) { - tmp_32 = 0; - for( k = 0; k < j; k++ ) { - tmp_32 = silk_SMLAWW( tmp_32, v_Q0[ k ], ptr2[ k ] ); /* Q0 */ - } - tmp_32 = silk_SUB32( ptr1[ i ], tmp_32 ); /* always < max(Correlation) */ - - /* tmp_32 / D_Q0[j] : Divide to Q16 */ - matrix_ptr( L_Q16, i, j, M ) = silk_ADD32( silk_SMMUL( tmp_32, one_div_diag_Q48 ), - silk_RSHIFT( silk_SMULWW( tmp_32, one_div_diag_Q36 ), 4 ) ); - - /* go to next column */ - ptr2 += M; - } - } - } - - silk_assert( status == 0 ); -} - -static OPUS_INLINE void silk_LS_divide_Q16_FIX( - opus_int32 T[], /* I/O Numenator vector */ - inv_D_t *inv_D, /* I 1 / D vector */ - opus_int M /* I dimension */ -) -{ - opus_int i; - opus_int32 tmp_32; - opus_int32 one_div_diag_Q36, one_div_diag_Q48; - - for( i = 0; i < M; i++ ) { - one_div_diag_Q36 = inv_D[ i ].Q36_part; - one_div_diag_Q48 = inv_D[ i ].Q48_part; - - tmp_32 = T[ i ]; - T[ i ] = silk_ADD32( silk_SMMUL( tmp_32, one_div_diag_Q48 ), silk_RSHIFT( silk_SMULWW( tmp_32, one_div_diag_Q36 ), 4 ) ); - } -} - -/* Solve Lx = b, when L is lower triangular and has ones on the diagonal */ -static OPUS_INLINE void silk_LS_SolveFirst_FIX( - const opus_int32 *L_Q16, /* I Pointer to Lower Triangular Matrix */ - opus_int M, /* I Dim of Matrix equation */ - const opus_int32 *b, /* I b Vector */ - opus_int32 *x_Q16 /* O x Vector */ -) -{ - opus_int i, j; - const opus_int32 *ptr32; - opus_int32 tmp_32; - - for( i = 0; i < M; i++ ) { - ptr32 = matrix_adr( L_Q16, i, 0, M ); - tmp_32 = 0; - for( j = 0; j < i; j++ ) { - tmp_32 = silk_SMLAWW( tmp_32, ptr32[ j ], x_Q16[ j ] ); - } - x_Q16[ i ] = silk_SUB32( b[ i ], tmp_32 ); - } -} - -/* Solve L^t*x = b, where L is lower triangular with ones on the diagonal */ -static OPUS_INLINE void silk_LS_SolveLast_FIX( - const opus_int32 *L_Q16, /* I Pointer to Lower Triangular Matrix */ - const opus_int M, /* I Dim of Matrix equation */ - const opus_int32 *b, /* I b Vector */ - opus_int32 *x_Q16 /* O x Vector */ -) -{ - opus_int i, j; - const opus_int32 *ptr32; - opus_int32 tmp_32; - - for( i = M - 1; i >= 0; i-- ) { - ptr32 = matrix_adr( L_Q16, 0, i, M ); - tmp_32 = 0; - for( j = M - 1; j > i; j-- ) { - tmp_32 = silk_SMLAWW( tmp_32, ptr32[ silk_SMULBB( j, M ) ], x_Q16[ j ] ); - } - x_Q16[ i ] = silk_SUB32( b[ i ], tmp_32 ); - } -} diff --git a/code/opus-1.1.4/silk/fixed/structs_FIX.h b/code/opus-1.1.4/silk/fixed/structs_FIX.h index 244b4793..2774a97b 100644 --- a/code/opus-1.1.4/silk/fixed/structs_FIX.h +++ b/code/opus-1.1.4/silk/fixed/structs_FIX.h @@ -47,31 +47,17 @@ typedef struct { opus_int32 Tilt_smth_Q16; } silk_shape_state_FIX; -/********************************/ -/* Prefilter state */ -/********************************/ -typedef struct { - opus_int16 sLTP_shp[ LTP_BUF_LENGTH ]; - opus_int32 sAR_shp[ MAX_SHAPE_LPC_ORDER + 1 ]; - opus_int sLTP_shp_buf_idx; - opus_int32 sLF_AR_shp_Q12; - opus_int32 sLF_MA_shp_Q12; - opus_int32 sHarmHP_Q2; - opus_int32 rand_seed; - opus_int lagPrev; -} silk_prefilter_state_FIX; - /********************************/ /* Encoder state FIX */ /********************************/ typedef struct { silk_encoder_state sCmn; /* Common struct, shared with floating-point code */ silk_shape_state_FIX sShape; /* Shape state */ - silk_prefilter_state_FIX sPrefilt; /* Prefilter State */ /* Buffer for find pitch and noise shape analysis */ silk_DWORD_ALIGN opus_int16 x_buf[ 2 * MAX_FRAME_LENGTH + LA_SHAPE_MAX ];/* Buffer for find pitch and noise shape analysis */ opus_int LTPCorr_Q15; /* Normalized correlation from pitch lag estimator */ + opus_int32 resNrgSmth; } silk_encoder_state_FIX; /************************/ @@ -87,11 +73,8 @@ typedef struct { /* Noise shaping parameters */ /* Testing */ - silk_DWORD_ALIGN opus_int16 AR1_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ]; - silk_DWORD_ALIGN opus_int16 AR2_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ]; + silk_DWORD_ALIGN opus_int16 AR_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ]; opus_int32 LF_shp_Q14[ MAX_NB_SUBFR ]; /* Packs two int16 coefficients per int32 value */ - opus_int GainsPre_Q14[ MAX_NB_SUBFR ]; - opus_int HarmBoost_Q14[ MAX_NB_SUBFR ]; opus_int Tilt_Q14[ MAX_NB_SUBFR ]; opus_int HarmShapeGain_Q14[ MAX_NB_SUBFR ]; opus_int Lambda_Q10; @@ -99,7 +82,6 @@ typedef struct { opus_int coding_quality_Q14; /* measures */ - opus_int sparseness_Q8; opus_int32 predGain_Q16; opus_int LTPredCodGain_Q7; opus_int32 ResNrg[ MAX_NB_SUBFR ]; /* Residual energy per subframe */ @@ -116,6 +98,7 @@ typedef struct { typedef struct { silk_encoder_state_FIX state_Fxx[ ENCODER_NUM_CHANNELS ]; stereo_enc_state sStereo; + opus_int32 nBitsUsedLBRR; opus_int32 nBitsExceeded; opus_int nChannelsAPI; opus_int nChannelsInternal; diff --git a/code/opus-1.1.4/silk/fixed/vector_ops_FIX.c b/code/opus-1.1.4/silk/fixed/vector_ops_FIX.c index 509c8b35..d9498001 100644 --- a/code/opus-1.1.4/silk/fixed/vector_ops_FIX.c +++ b/code/opus-1.1.4/silk/fixed/vector_ops_FIX.c @@ -30,6 +30,7 @@ POSSIBILITY OF SUCH DAMAGE. #endif #include "SigProc_FIX.h" +#include "pitch.h" /* Copy and multiply a vector by a constant */ void silk_scale_copy_vector16( @@ -70,18 +71,23 @@ void silk_scale_vector32_Q26_lshift_18( opus_int32 silk_inner_prod_aligned( const opus_int16 *const inVec1, /* I input vector 1 */ const opus_int16 *const inVec2, /* I input vector 2 */ - const opus_int len /* I vector lengths */ + const opus_int len, /* I vector lengths */ + int arch /* I Run-time architecture */ ) { +#ifdef FIXED_POINT + return celt_inner_prod(inVec1, inVec2, len, arch); +#else opus_int i; opus_int32 sum = 0; for( i = 0; i < len; i++ ) { sum = silk_SMLABB( sum, inVec1[ i ], inVec2[ i ] ); } return sum; +#endif } -opus_int64 silk_inner_prod16_aligned_64( +opus_int64 silk_inner_prod16_aligned_64_c( const opus_int16 *inVec1, /* I input vector 1 */ const opus_int16 *inVec2, /* I input vector 2 */ const opus_int len /* I vector lengths */ diff --git a/code/opus-1.1.4/silk/fixed/warped_autocorrelation_FIX.c b/code/opus-1.1.4/silk/fixed/warped_autocorrelation_FIX.c index a4a579b1..994c299a 100644 --- a/code/opus-1.1.4/silk/fixed/warped_autocorrelation_FIX.c +++ b/code/opus-1.1.4/silk/fixed/warped_autocorrelation_FIX.c @@ -31,11 +31,13 @@ POSSIBILITY OF SUCH DAMAGE. #include "main_FIX.h" -#define QC 10 -#define QS 14 +#if defined(MIPSr1_ASM) +#include "mips/warped_autocorrelation_FIX_mipsr1.h" +#endif + /* Autocorrelations for a warped frequency axis */ -void silk_warped_autocorrelation_FIX( +void silk_warped_autocorrelation_FIX_c( opus_int32 *corr, /* O Result [order + 1] */ opus_int *scale, /* O Scaling of the correlation vector */ const opus_int16 *input, /* I Input data to correlate */ diff --git a/code/opus-1.1.4/silk/float/LPC_inv_pred_gain_FLP.c b/code/opus-1.1.4/silk/float/LPC_inv_pred_gain_FLP.c index 25178bac..2be2122d 100644 --- a/code/opus-1.1.4/silk/float/LPC_inv_pred_gain_FLP.c +++ b/code/opus-1.1.4/silk/float/LPC_inv_pred_gain_FLP.c @@ -31,8 +31,7 @@ POSSIBILITY OF SUCH DAMAGE. #include "SigProc_FIX.h" #include "SigProc_FLP.h" - -#define RC_THRESHOLD 0.9999f +#include "define.h" /* compute inverse of LPC prediction gain, and */ /* test if LPC coefficients are stable (all poles within unit circle) */ @@ -43,34 +42,32 @@ silk_float silk_LPC_inverse_pred_gain_FLP( /* O return inverse prediction ga ) { opus_int k, n; - double invGain, rc, rc_mult1, rc_mult2; - silk_float Atmp[ 2 ][ SILK_MAX_ORDER_LPC ]; - silk_float *Aold, *Anew; + double invGain, rc, rc_mult1, rc_mult2, tmp1, tmp2; + silk_float Atmp[ SILK_MAX_ORDER_LPC ]; - Anew = Atmp[ order & 1 ]; - silk_memcpy( Anew, A, order * sizeof(silk_float) ); + silk_memcpy( Atmp, A, order * sizeof(silk_float) ); invGain = 1.0; for( k = order - 1; k > 0; k-- ) { - rc = -Anew[ k ]; - if( rc > RC_THRESHOLD || rc < -RC_THRESHOLD ) { + rc = -Atmp[ k ]; + rc_mult1 = 1.0f - rc * rc; + invGain *= rc_mult1; + if( invGain * MAX_PREDICTION_POWER_GAIN < 1.0f ) { return 0.0f; } - rc_mult1 = 1.0f - rc * rc; rc_mult2 = 1.0f / rc_mult1; - invGain *= rc_mult1; - /* swap pointers */ - Aold = Anew; - Anew = Atmp[ k & 1 ]; - for( n = 0; n < k; n++ ) { - Anew[ n ] = (silk_float)( ( Aold[ n ] - Aold[ k - n - 1 ] * rc ) * rc_mult2 ); + for( n = 0; n < (k + 1) >> 1; n++ ) { + tmp1 = Atmp[ n ]; + tmp2 = Atmp[ k - n - 1 ]; + Atmp[ n ] = (silk_float)( ( tmp1 - tmp2 * rc ) * rc_mult2 ); + Atmp[ k - n - 1 ] = (silk_float)( ( tmp2 - tmp1 * rc ) * rc_mult2 ); } } - rc = -Anew[ 0 ]; - if( rc > RC_THRESHOLD || rc < -RC_THRESHOLD ) { - return 0.0f; - } + rc = -Atmp[ 0 ]; rc_mult1 = 1.0f - rc * rc; invGain *= rc_mult1; + if( invGain * MAX_PREDICTION_POWER_GAIN < 1.0f ) { + return 0.0f; + } return (silk_float)invGain; } diff --git a/code/opus-1.1.4/silk/float/SigProc_FLP.h b/code/opus-1.1.4/silk/float/SigProc_FLP.h index f0cb3733..953de8b0 100644 --- a/code/opus-1.1.4/silk/float/SigProc_FLP.h +++ b/code/opus-1.1.4/silk/float/SigProc_FLP.h @@ -68,13 +68,6 @@ void silk_k2a_FLP( opus_int32 order /* I prediction order */ ); -/* Solve the normal equations using the Levinson-Durbin recursion */ -silk_float silk_levinsondurbin_FLP( /* O prediction error energy */ - silk_float A[], /* O prediction coefficients [order] */ - const silk_float corr[], /* I input auto-correlations [order + 1] */ - const opus_int order /* I prediction order */ -); - /* compute autocorrelation */ void silk_autocorrelation_FLP( silk_float *results, /* O result (length correlationCount) */ diff --git a/code/opus-1.1.4/silk/float/encode_frame_FLP.c b/code/opus-1.1.4/silk/float/encode_frame_FLP.c index 2092a4d9..c3ad50a9 100644 --- a/code/opus-1.1.4/silk/float/encode_frame_FLP.c +++ b/code/opus-1.1.4/silk/float/encode_frame_FLP.c @@ -29,6 +29,7 @@ POSSIBILITY OF SUCH DAMAGE. #include "config.h" #endif +#include #include "main_FLP.h" #include "tuning_parameters.h" @@ -85,7 +86,6 @@ opus_int silk_encode_frame_FLP( silk_encoder_control_FLP sEncCtrl; opus_int i, iter, maxIter, found_upper, found_lower, ret = 0; silk_float *x_frame, *res_pitch_frame; - silk_float xfw[ MAX_FRAME_LENGTH ]; silk_float res_pitch[ 2 * MAX_FRAME_LENGTH + LA_PITCH_MAX ]; ec_enc sRangeEnc_copy, sRangeEnc_copy2; silk_nsq_state sNSQ_copy, sNSQ_copy2; @@ -97,6 +97,9 @@ opus_int silk_encode_frame_FLP( opus_int8 LastGainIndex_copy2; opus_int32 pGains_Q16[ MAX_NB_SUBFR ]; opus_uint8 ec_buf_copy[ 1275 ]; + opus_int gain_lock[ MAX_NB_SUBFR ] = {0}; + opus_int16 best_gain_mult[ MAX_NB_SUBFR ]; + opus_int best_sum[ MAX_NB_SUBFR ]; /* This is totally unnecessary but many compilers (including gcc) are too dumb to realise it */ LastGainIndex_copy2 = nBits_lower = nBits_upper = gainMult_lower = gainMult_upper = 0; @@ -139,22 +142,17 @@ opus_int silk_encode_frame_FLP( /***************************************************/ /* Find linear prediction coefficients (LPC + LTP) */ /***************************************************/ - silk_find_pred_coefs_FLP( psEnc, &sEncCtrl, res_pitch, x_frame, condCoding ); + silk_find_pred_coefs_FLP( psEnc, &sEncCtrl, res_pitch_frame, x_frame, condCoding ); /****************************************/ /* Process gains */ /****************************************/ silk_process_gains_FLP( psEnc, &sEncCtrl, condCoding ); - /*****************************************/ - /* Prefiltering for noise shaper */ - /*****************************************/ - silk_prefilter_FLP( psEnc, &sEncCtrl, xfw, x_frame ); - /****************************************/ /* Low Bitrate Redundant Encoding */ /****************************************/ - silk_LBRR_encode_FLP( psEnc, &sEncCtrl, xfw, condCoding ); + silk_LBRR_encode_FLP( psEnc, &sEncCtrl, x_frame, condCoding ); /* Loop over quantizer and entroy coding to control bitrate */ maxIter = 6; @@ -188,7 +186,11 @@ opus_int silk_encode_frame_FLP( /*****************************************/ /* Noise shaping quantization */ /*****************************************/ - silk_NSQ_wrapper_FLP( psEnc, &sEncCtrl, &psEnc->sCmn.indices, &psEnc->sCmn.sNSQ, psEnc->sCmn.pulses, xfw ); + silk_NSQ_wrapper_FLP( psEnc, &sEncCtrl, &psEnc->sCmn.indices, &psEnc->sCmn.sNSQ, psEnc->sCmn.pulses, x_frame ); + + if ( iter == maxIter && !found_lower ) { + silk_memcpy( &sRangeEnc_copy2, psRangeEnc, sizeof( ec_enc ) ); + } /****************************************/ /* Encode Parameters */ @@ -203,6 +205,33 @@ opus_int silk_encode_frame_FLP( nBits = ec_tell( psRangeEnc ); + /* If we still bust after the last iteration, do some damage control. */ + if ( iter == maxIter && !found_lower && nBits > maxBits ) { + silk_memcpy( psRangeEnc, &sRangeEnc_copy2, sizeof( ec_enc ) ); + + /* Keep gains the same as the last frame. */ + psEnc->sShape.LastGainIndex = sEncCtrl.lastGainIndexPrev; + for ( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) { + psEnc->sCmn.indices.GainsIndices[ i ] = 4; + } + if (condCoding != CODE_CONDITIONALLY) { + psEnc->sCmn.indices.GainsIndices[ 0 ] = sEncCtrl.lastGainIndexPrev; + } + psEnc->sCmn.ec_prevLagIndex = ec_prevLagIndex_copy; + psEnc->sCmn.ec_prevSignalType = ec_prevSignalType_copy; + /* Clear all pulses. */ + for ( i = 0; i < psEnc->sCmn.frame_length; i++ ) { + psEnc->sCmn.pulses[ i ] = 0; + } + + silk_encode_indices( &psEnc->sCmn, psRangeEnc, psEnc->sCmn.nFramesEncoded, 0, condCoding ); + + silk_encode_pulses( psRangeEnc, psEnc->sCmn.indices.signalType, psEnc->sCmn.indices.quantOffsetType, + psEnc->sCmn.pulses, psEnc->sCmn.frame_length ); + + nBits = ec_tell( psRangeEnc ); + } + if( useCBR == 0 && iter == 0 && nBits <= maxBits ) { break; } @@ -223,7 +252,9 @@ opus_int silk_encode_frame_FLP( if( nBits > maxBits ) { if( found_lower == 0 && iter >= 2 ) { /* Adjust the quantizer's rate/distortion tradeoff and discard previous "upper" results */ - sEncCtrl.Lambda *= 1.5f; + sEncCtrl.Lambda = silk_max_float(sEncCtrl.Lambda*1.5f, 1.5f); + /* Reducing dithering can help us hit the target. */ + psEnc->sCmn.indices.quantOffsetType = 0; found_upper = 0; gainsID_upper = -1; } else { @@ -250,15 +281,34 @@ opus_int silk_encode_frame_FLP( break; } + if ( !found_lower && nBits > maxBits ) { + int j; + for ( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) { + int sum=0; + for ( j = i*psEnc->sCmn.subfr_length; j < (i+1)*psEnc->sCmn.subfr_length; j++ ) { + sum += abs( psEnc->sCmn.pulses[j] ); + } + if ( iter == 0 || (sum < best_sum[i] && !gain_lock[i]) ) { + best_sum[i] = sum; + best_gain_mult[i] = gainMult_Q8; + } else { + gain_lock[i] = 1; + } + } + } if( ( found_lower & found_upper ) == 0 ) { /* Adjust gain according to high-rate rate/distortion curve */ - opus_int32 gain_factor_Q16; - gain_factor_Q16 = silk_log2lin( silk_LSHIFT( nBits - maxBits, 7 ) / psEnc->sCmn.frame_length + SILK_FIX_CONST( 16, 7 ) ); - gain_factor_Q16 = silk_min_32( gain_factor_Q16, SILK_FIX_CONST( 2, 16 ) ); if( nBits > maxBits ) { - gain_factor_Q16 = silk_max_32( gain_factor_Q16, SILK_FIX_CONST( 1.3, 16 ) ); + if (gainMult_Q8 < 16384) { + gainMult_Q8 *= 2; + } else { + gainMult_Q8 = 32767; + } + } else { + opus_int32 gain_factor_Q16; + gain_factor_Q16 = silk_log2lin( silk_LSHIFT( nBits - maxBits, 7 ) / psEnc->sCmn.frame_length + SILK_FIX_CONST( 16, 7 ) ); + gainMult_Q8 = silk_SMULWB( gain_factor_Q16, gainMult_Q8 ); } - gainMult_Q8 = silk_SMULWB( gain_factor_Q16, gainMult_Q8 ); } else { /* Adjust gain by interpolating */ gainMult_Q8 = gainMult_lower + ( ( gainMult_upper - gainMult_lower ) * ( maxBits - nBits_lower ) ) / ( nBits_upper - nBits_lower ); @@ -272,7 +322,13 @@ opus_int silk_encode_frame_FLP( } for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) { - pGains_Q16[ i ] = silk_LSHIFT_SAT32( silk_SMULWB( sEncCtrl.GainsUnq_Q16[ i ], gainMult_Q8 ), 8 ); + opus_int16 tmp; + if ( gain_lock[i] ) { + tmp = best_gain_mult[i]; + } else { + tmp = gainMult_Q8; + } + pGains_Q16[ i ] = silk_LSHIFT_SAT32( silk_SMULWB( sEncCtrl.GainsUnq_Q16[ i ], tmp ), 8 ); } /* Quantize gains */ diff --git a/code/opus-1.1.4/silk/float/energy_FLP.c b/code/opus-1.1.4/silk/float/energy_FLP.c index 24b8179f..7bc7173c 100644 --- a/code/opus-1.1.4/silk/float/energy_FLP.c +++ b/code/opus-1.1.4/silk/float/energy_FLP.c @@ -37,13 +37,12 @@ double silk_energy_FLP( opus_int dataSize ) { - opus_int i, dataSize4; + opus_int i; double result; /* 4x unrolled loop */ result = 0.0; - dataSize4 = dataSize & 0xFFFC; - for( i = 0; i < dataSize4; i += 4 ) { + for( i = 0; i < dataSize - 3; i += 4 ) { result += data[ i + 0 ] * (double)data[ i + 0 ] + data[ i + 1 ] * (double)data[ i + 1 ] + data[ i + 2 ] * (double)data[ i + 2 ] + diff --git a/code/opus-1.1.4/silk/float/find_LPC_FLP.c b/code/opus-1.1.4/silk/float/find_LPC_FLP.c index fcfe1c36..4d63964f 100644 --- a/code/opus-1.1.4/silk/float/find_LPC_FLP.c +++ b/code/opus-1.1.4/silk/float/find_LPC_FLP.c @@ -73,7 +73,7 @@ void silk_find_LPC_FLP( silk_interpolate( NLSF0_Q15, psEncC->prev_NLSFq_Q15, NLSF_Q15, k, psEncC->predictLPCOrder ); /* Convert to LPC for residual energy evaluation */ - silk_NLSF2A_FLP( a_tmp, NLSF0_Q15, psEncC->predictLPCOrder ); + silk_NLSF2A_FLP( a_tmp, NLSF0_Q15, psEncC->predictLPCOrder, psEncC->arch ); /* Calculate residual energy with LSF interpolation */ silk_LPC_analysis_filter_FLP( LPC_res, a_tmp, x, 2 * subfr_length, psEncC->predictLPCOrder ); diff --git a/code/opus-1.1.4/silk/float/find_LTP_FLP.c b/code/opus-1.1.4/silk/float/find_LTP_FLP.c index 72299960..f9706493 100644 --- a/code/opus-1.1.4/silk/float/find_LTP_FLP.c +++ b/code/opus-1.1.4/silk/float/find_LTP_FLP.c @@ -33,100 +33,32 @@ POSSIBILITY OF SUCH DAMAGE. #include "tuning_parameters.h" void silk_find_LTP_FLP( - silk_float b[ MAX_NB_SUBFR * LTP_ORDER ], /* O LTP coefs */ - silk_float WLTP[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ], /* O Weight for LTP quantization */ - silk_float *LTPredCodGain, /* O LTP coding gain */ - const silk_float r_lpc[], /* I LPC residual */ - const opus_int lag[ MAX_NB_SUBFR ], /* I LTP lags */ - const silk_float Wght[ MAX_NB_SUBFR ], /* I Weights */ + silk_float XX[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ], /* O Weight for LTP quantization */ + silk_float xX[ MAX_NB_SUBFR * LTP_ORDER ], /* O Weight for LTP quantization */ + const silk_float r_ptr[], /* I LPC residual */ + const opus_int lag[ MAX_NB_SUBFR ], /* I LTP lags */ const opus_int subfr_length, /* I Subframe length */ - const opus_int nb_subfr, /* I number of subframes */ - const opus_int mem_offset /* I Number of samples in LTP memory */ + const opus_int nb_subfr /* I number of subframes */ ) { - opus_int i, k; - silk_float *b_ptr, temp, *WLTP_ptr; - silk_float LPC_res_nrg, LPC_LTP_res_nrg; - silk_float d[ MAX_NB_SUBFR ], m, g, delta_b[ LTP_ORDER ]; - silk_float w[ MAX_NB_SUBFR ], nrg[ MAX_NB_SUBFR ], regu; - silk_float Rr[ LTP_ORDER ], rr[ MAX_NB_SUBFR ]; - const silk_float *r_ptr, *lag_ptr; + opus_int k; + silk_float *xX_ptr, *XX_ptr; + const silk_float *lag_ptr; + silk_float xx, temp; - b_ptr = b; - WLTP_ptr = WLTP; - r_ptr = &r_lpc[ mem_offset ]; + xX_ptr = xX; + XX_ptr = XX; for( k = 0; k < nb_subfr; k++ ) { lag_ptr = r_ptr - ( lag[ k ] + LTP_ORDER / 2 ); + silk_corrMatrix_FLP( lag_ptr, subfr_length, LTP_ORDER, XX_ptr ); + silk_corrVector_FLP( lag_ptr, r_ptr, subfr_length, LTP_ORDER, xX_ptr ); + xx = ( silk_float )silk_energy_FLP( r_ptr, subfr_length + LTP_ORDER ); + temp = 1.0f / silk_max( xx, LTP_CORR_INV_MAX * 0.5f * ( XX_ptr[ 0 ] + XX_ptr[ 24 ] ) + 1.0f ); + silk_scale_vector_FLP( XX_ptr, temp, LTP_ORDER * LTP_ORDER ); + silk_scale_vector_FLP( xX_ptr, temp, LTP_ORDER ); - silk_corrMatrix_FLP( lag_ptr, subfr_length, LTP_ORDER, WLTP_ptr ); - silk_corrVector_FLP( lag_ptr, r_ptr, subfr_length, LTP_ORDER, Rr ); - - rr[ k ] = ( silk_float )silk_energy_FLP( r_ptr, subfr_length ); - regu = 1.0f + rr[ k ] + - matrix_ptr( WLTP_ptr, 0, 0, LTP_ORDER ) + - matrix_ptr( WLTP_ptr, LTP_ORDER-1, LTP_ORDER-1, LTP_ORDER ); - regu *= LTP_DAMPING / 3; - silk_regularize_correlations_FLP( WLTP_ptr, &rr[ k ], regu, LTP_ORDER ); - silk_solve_LDL_FLP( WLTP_ptr, LTP_ORDER, Rr, b_ptr ); - - /* Calculate residual energy */ - nrg[ k ] = silk_residual_energy_covar_FLP( b_ptr, WLTP_ptr, Rr, rr[ k ], LTP_ORDER ); - - temp = Wght[ k ] / ( nrg[ k ] * Wght[ k ] + 0.01f * subfr_length ); - silk_scale_vector_FLP( WLTP_ptr, temp, LTP_ORDER * LTP_ORDER ); - w[ k ] = matrix_ptr( WLTP_ptr, LTP_ORDER / 2, LTP_ORDER / 2, LTP_ORDER ); - - r_ptr += subfr_length; - b_ptr += LTP_ORDER; - WLTP_ptr += LTP_ORDER * LTP_ORDER; - } - - /* Compute LTP coding gain */ - if( LTPredCodGain != NULL ) { - LPC_LTP_res_nrg = 1e-6f; - LPC_res_nrg = 0.0f; - for( k = 0; k < nb_subfr; k++ ) { - LPC_res_nrg += rr[ k ] * Wght[ k ]; - LPC_LTP_res_nrg += nrg[ k ] * Wght[ k ]; - } - - silk_assert( LPC_LTP_res_nrg > 0 ); - *LTPredCodGain = 3.0f * silk_log2( LPC_res_nrg / LPC_LTP_res_nrg ); - } - - /* Smoothing */ - /* d = sum( B, 1 ); */ - b_ptr = b; - for( k = 0; k < nb_subfr; k++ ) { - d[ k ] = 0; - for( i = 0; i < LTP_ORDER; i++ ) { - d[ k ] += b_ptr[ i ]; - } - b_ptr += LTP_ORDER; - } - /* m = ( w * d' ) / ( sum( w ) + 1e-3 ); */ - temp = 1e-3f; - for( k = 0; k < nb_subfr; k++ ) { - temp += w[ k ]; - } - m = 0; - for( k = 0; k < nb_subfr; k++ ) { - m += d[ k ] * w[ k ]; - } - m = m / temp; - - b_ptr = b; - for( k = 0; k < nb_subfr; k++ ) { - g = LTP_SMOOTHING / ( LTP_SMOOTHING + w[ k ] ) * ( m - d[ k ] ); - temp = 0; - for( i = 0; i < LTP_ORDER; i++ ) { - delta_b[ i ] = silk_max_float( b_ptr[ i ], 0.1f ); - temp += delta_b[ i ]; - } - temp = g / temp; - for( i = 0; i < LTP_ORDER; i++ ) { - b_ptr[ i ] = b_ptr[ i ] + delta_b[ i ] * temp; - } - b_ptr += LTP_ORDER; + r_ptr += subfr_length; + XX_ptr += LTP_ORDER * LTP_ORDER; + xX_ptr += LTP_ORDER; } } diff --git a/code/opus-1.1.4/silk/float/find_pred_coefs_FLP.c b/code/opus-1.1.4/silk/float/find_pred_coefs_FLP.c index 1af4fe5f..cb2e763b 100644 --- a/code/opus-1.1.4/silk/float/find_pred_coefs_FLP.c +++ b/code/opus-1.1.4/silk/float/find_pred_coefs_FLP.c @@ -41,8 +41,9 @@ void silk_find_pred_coefs_FLP( ) { opus_int i; - silk_float WLTP[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ]; - silk_float invGains[ MAX_NB_SUBFR ], Wght[ MAX_NB_SUBFR ]; + silk_float XXLTP[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ]; + silk_float xXLTP[ MAX_NB_SUBFR * LTP_ORDER ]; + silk_float invGains[ MAX_NB_SUBFR ]; opus_int16 NLSF_Q15[ MAX_LPC_ORDER ]; const silk_float *x_ptr; silk_float *x_pre_ptr, LPC_in_pre[ MAX_NB_SUBFR * MAX_LPC_ORDER + MAX_FRAME_LENGTH ]; @@ -52,7 +53,6 @@ void silk_find_pred_coefs_FLP( for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) { silk_assert( psEncCtrl->Gains[ i ] > 0.0f ); invGains[ i ] = 1.0f / psEncCtrl->Gains[ i ]; - Wght[ i ] = invGains[ i ] * invGains[ i ]; } if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) { @@ -62,13 +62,11 @@ void silk_find_pred_coefs_FLP( silk_assert( psEnc->sCmn.ltp_mem_length - psEnc->sCmn.predictLPCOrder >= psEncCtrl->pitchL[ 0 ] + LTP_ORDER / 2 ); /* LTP analysis */ - silk_find_LTP_FLP( psEncCtrl->LTPCoef, WLTP, &psEncCtrl->LTPredCodGain, res_pitch, - psEncCtrl->pitchL, Wght, psEnc->sCmn.subfr_length, psEnc->sCmn.nb_subfr, psEnc->sCmn.ltp_mem_length ); + silk_find_LTP_FLP( XXLTP, xXLTP, res_pitch, psEncCtrl->pitchL, psEnc->sCmn.subfr_length, psEnc->sCmn.nb_subfr ); /* Quantize LTP gain parameters */ silk_quant_LTP_gains_FLP( psEncCtrl->LTPCoef, psEnc->sCmn.indices.LTPIndex, &psEnc->sCmn.indices.PERIndex, - &psEnc->sCmn.sum_log_gain_Q7, WLTP, psEnc->sCmn.mu_LTP_Q9, psEnc->sCmn.LTPQuantLowComplexity, psEnc->sCmn.nb_subfr, - psEnc->sCmn.arch ); + &psEnc->sCmn.sum_log_gain_Q7, &psEncCtrl->LTPredCodGain, XXLTP, xXLTP, psEnc->sCmn.subfr_length, psEnc->sCmn.nb_subfr, psEnc->sCmn.arch ); /* Control LTP scaling */ silk_LTP_scale_ctrl_FLP( psEnc, psEncCtrl, condCoding ); diff --git a/code/opus-1.1.4/silk/float/inner_product_FLP.c b/code/opus-1.1.4/silk/float/inner_product_FLP.c index 029c0129..cdd39d24 100644 --- a/code/opus-1.1.4/silk/float/inner_product_FLP.c +++ b/code/opus-1.1.4/silk/float/inner_product_FLP.c @@ -38,13 +38,12 @@ double silk_inner_product_FLP( opus_int dataSize ) { - opus_int i, dataSize4; + opus_int i; double result; /* 4x unrolled loop */ result = 0.0; - dataSize4 = dataSize & 0xFFFC; - for( i = 0; i < dataSize4; i += 4 ) { + for( i = 0; i < dataSize - 3; i += 4 ) { result += data1[ i + 0 ] * (double)data2[ i + 0 ] + data1[ i + 1 ] * (double)data2[ i + 1 ] + data1[ i + 2 ] * (double)data2[ i + 2 ] + diff --git a/code/opus-1.1.4/silk/float/k2a_FLP.c b/code/opus-1.1.4/silk/float/k2a_FLP.c index 12af4e76..1448008d 100644 --- a/code/opus-1.1.4/silk/float/k2a_FLP.c +++ b/code/opus-1.1.4/silk/float/k2a_FLP.c @@ -39,15 +39,16 @@ void silk_k2a_FLP( ) { opus_int k, n; - silk_float Atmp[ SILK_MAX_ORDER_LPC ]; + silk_float rck, tmp1, tmp2; for( k = 0; k < order; k++ ) { - for( n = 0; n < k; n++ ) { - Atmp[ n ] = A[ n ]; + rck = rc[ k ]; + for( n = 0; n < (k + 1) >> 1; n++ ) { + tmp1 = A[ n ]; + tmp2 = A[ k - n - 1 ]; + A[ n ] = tmp1 + tmp2 * rck; + A[ k - n - 1 ] = tmp2 + tmp1 * rck; } - for( n = 0; n < k; n++ ) { - A[ n ] += Atmp[ k - n - 1 ] * rc[ k ]; - } - A[ k ] = -rc[ k ]; + A[ k ] = -rck; } } diff --git a/code/opus-1.1.4/silk/float/levinsondurbin_FLP.c b/code/opus-1.1.4/silk/float/levinsondurbin_FLP.c deleted file mode 100644 index f0ba6069..00000000 --- a/code/opus-1.1.4/silk/float/levinsondurbin_FLP.c +++ /dev/null @@ -1,81 +0,0 @@ -/*********************************************************************** -Copyright (c) 2006-2011, Skype Limited. All rights reserved. -Redistribution and use in source and binary forms, with or without -modification, are permitted provided that the following conditions -are met: -- Redistributions of source code must retain the above copyright notice, -this list of conditions and the following disclaimer. -- Redistributions in binary form must reproduce the above copyright -notice, this list of conditions and the following disclaimer in the -documentation and/or other materials provided with the distribution. -- Neither the name of Internet Society, IETF or IETF Trust, nor the -names of specific contributors, may be used to endorse or promote -products derived from this software without specific prior written -permission. -THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" -AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE -IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE -ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE -LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR -CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF -SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS -INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN -CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) -ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE -POSSIBILITY OF SUCH DAMAGE. -***********************************************************************/ - -#ifdef HAVE_CONFIG_H -#include "config.h" -#endif - -#include "SigProc_FLP.h" - -/* Solve the normal equations using the Levinson-Durbin recursion */ -silk_float silk_levinsondurbin_FLP( /* O prediction error energy */ - silk_float A[], /* O prediction coefficients [order] */ - const silk_float corr[], /* I input auto-correlations [order + 1] */ - const opus_int order /* I prediction order */ -) -{ - opus_int i, mHalf, m; - silk_float min_nrg, nrg, t, km, Atmp1, Atmp2; - - min_nrg = 1e-12f * corr[ 0 ] + 1e-9f; - nrg = corr[ 0 ]; - nrg = silk_max_float(min_nrg, nrg); - A[ 0 ] = corr[ 1 ] / nrg; - nrg -= A[ 0 ] * corr[ 1 ]; - nrg = silk_max_float(min_nrg, nrg); - - for( m = 1; m < order; m++ ) - { - t = corr[ m + 1 ]; - for( i = 0; i < m; i++ ) { - t -= A[ i ] * corr[ m - i ]; - } - - /* reflection coefficient */ - km = t / nrg; - - /* residual energy */ - nrg -= km * t; - nrg = silk_max_float(min_nrg, nrg); - - mHalf = m >> 1; - for( i = 0; i < mHalf; i++ ) { - Atmp1 = A[ i ]; - Atmp2 = A[ m - i - 1 ]; - A[ m - i - 1 ] -= km * Atmp1; - A[ i ] -= km * Atmp2; - } - if( m & 1 ) { - A[ mHalf ] -= km * A[ mHalf ]; - } - A[ m ] = km; - } - - /* return the residual energy */ - return nrg; -} - diff --git a/code/opus-1.1.4/silk/float/main_FLP.h b/code/opus-1.1.4/silk/float/main_FLP.h index e5a75972..f47fc93b 100644 --- a/code/opus-1.1.4/silk/float/main_FLP.h +++ b/code/opus-1.1.4/silk/float/main_FLP.h @@ -79,22 +79,11 @@ opus_int silk_init_encoder( opus_int silk_control_encoder( silk_encoder_state_FLP *psEnc, /* I/O Pointer to Silk encoder state FLP */ silk_EncControlStruct *encControl, /* I Control structure */ - const opus_int32 TargetRate_bps, /* I Target max bitrate (bps) */ const opus_int allow_bw_switch, /* I Flag to allow switching audio bandwidth */ const opus_int channelNb, /* I Channel number */ const opus_int force_fs_kHz ); -/****************/ -/* Prefiltering */ -/****************/ -void silk_prefilter_FLP( - silk_encoder_state_FLP *psEnc, /* I/O Encoder state FLP */ - const silk_encoder_control_FLP *psEncCtrl, /* I Encoder control FLP */ - silk_float xw[], /* O Weighted signal */ - const silk_float x[] /* I Speech signal */ -); - /**************************/ /* Noise shaping analysis */ /**************************/ @@ -153,15 +142,12 @@ void silk_find_LPC_FLP( /* LTP analysis */ void silk_find_LTP_FLP( - silk_float b[ MAX_NB_SUBFR * LTP_ORDER ], /* O LTP coefs */ - silk_float WLTP[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ], /* O Weight for LTP quantization */ - silk_float *LTPredCodGain, /* O LTP coding gain */ - const silk_float r_lpc[], /* I LPC residual */ + silk_float XX[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ], /* O Weight for LTP quantization */ + silk_float xX[ MAX_NB_SUBFR * LTP_ORDER ], /* O Weight for LTP quantization */ + const silk_float r_ptr[], /* I LPC residual */ const opus_int lag[ MAX_NB_SUBFR ], /* I LTP lags */ - const silk_float Wght[ MAX_NB_SUBFR ], /* I Weights */ const opus_int subfr_length, /* I Subframe length */ - const opus_int nb_subfr, /* I number of subframes */ - const opus_int mem_offset /* I Number of samples in LTP memory */ + const opus_int nb_subfr /* I number of subframes */ ); void silk_LTP_analysis_filter_FLP( @@ -198,14 +184,15 @@ void silk_LPC_analysis_filter_FLP( /* LTP tap quantizer */ void silk_quant_LTP_gains_FLP( - silk_float B[ MAX_NB_SUBFR * LTP_ORDER ], /* I/O (Un-)quantized LTP gains */ + silk_float B[ MAX_NB_SUBFR * LTP_ORDER ], /* O Quantized LTP gains */ opus_int8 cbk_index[ MAX_NB_SUBFR ], /* O Codebook index */ opus_int8 *periodicity_index, /* O Periodicity index */ opus_int32 *sum_log_gain_Q7, /* I/O Cumulative max prediction gain */ - const silk_float W[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ], /* I Error weights */ - const opus_int mu_Q10, /* I Mu value (R/D tradeoff) */ - const opus_int lowComplexity, /* I Flag for low complexity */ - const opus_int nb_subfr, /* I number of subframes */ + silk_float *pred_gain_dB, /* O LTP prediction gain */ + const silk_float XX[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ], /* I Correlation matrix */ + const silk_float xX[ MAX_NB_SUBFR * LTP_ORDER ], /* I Correlation vector */ + const opus_int subfr_len, /* I Number of samples per subframe */ + const opus_int nb_subfr, /* I Number of subframes */ int arch /* I Run-time architecture */ ); @@ -245,22 +232,6 @@ void silk_corrVector_FLP( silk_float *Xt /* O X'*t correlation vector [order] */ ); -/* Add noise to matrix diagonal */ -void silk_regularize_correlations_FLP( - silk_float *XX, /* I/O Correlation matrices */ - silk_float *xx, /* I/O Correlation values */ - const silk_float noise, /* I Noise energy to add */ - const opus_int D /* I Dimension of XX */ -); - -/* Function to solve linear equation Ax = b, where A is an MxM symmetric matrix */ -void silk_solve_LDL_FLP( - silk_float *A, /* I/O Symmetric square matrix, out: reg. */ - const opus_int M, /* I Size of matrix */ - const silk_float *b, /* I Pointer to b vector */ - silk_float *x /* O Pointer to x solution vector */ -); - /* Apply sine window to signal vector. */ /* Window types: */ /* 1 -> sine window from 0 to pi/2 */ @@ -285,7 +256,8 @@ void silk_A2NLSF_FLP( void silk_NLSF2A_FLP( silk_float *pAR, /* O LPC coefficients [ LPC_order ] */ const opus_int16 *NLSF_Q15, /* I NLSF vector [ LPC_order ] */ - const opus_int LPC_order /* I LPC order */ + const opus_int LPC_order, /* I LPC order */ + int arch /* I Run-time architecture */ ); /* Limit, stabilize, and quantize NLSFs */ diff --git a/code/opus-1.1.4/silk/float/noise_shape_analysis_FLP.c b/code/opus-1.1.4/silk/float/noise_shape_analysis_FLP.c index 65f6ea58..cb3d8a50 100644 --- a/code/opus-1.1.4/silk/float/noise_shape_analysis_FLP.c +++ b/code/opus-1.1.4/silk/float/noise_shape_analysis_FLP.c @@ -55,25 +55,21 @@ static OPUS_INLINE silk_float warped_gain( /* Convert warped filter coefficients to monic pseudo-warped coefficients and limit maximum */ /* amplitude of monic warped coefficients by using bandwidth expansion on the true coefficients */ static OPUS_INLINE void warped_true2monic_coefs( - silk_float *coefs_syn, - silk_float *coefs_ana, + silk_float *coefs, silk_float lambda, silk_float limit, opus_int order ) { opus_int i, iter, ind = 0; - silk_float tmp, maxabs, chirp, gain_syn, gain_ana; + silk_float tmp, maxabs, chirp, gain; /* Convert to monic coefficients */ for( i = order - 1; i > 0; i-- ) { - coefs_syn[ i - 1 ] -= lambda * coefs_syn[ i ]; - coefs_ana[ i - 1 ] -= lambda * coefs_ana[ i ]; + coefs[ i - 1 ] -= lambda * coefs[ i ]; } - gain_syn = ( 1.0f - lambda * lambda ) / ( 1.0f + lambda * coefs_syn[ 0 ] ); - gain_ana = ( 1.0f - lambda * lambda ) / ( 1.0f + lambda * coefs_ana[ 0 ] ); + gain = ( 1.0f - lambda * lambda ) / ( 1.0f + lambda * coefs[ 0 ] ); for( i = 0; i < order; i++ ) { - coefs_syn[ i ] *= gain_syn; - coefs_ana[ i ] *= gain_ana; + coefs[ i ] *= gain; } /* Limit */ @@ -81,7 +77,7 @@ static OPUS_INLINE void warped_true2monic_coefs( /* Find maximum absolute value */ maxabs = -1.0f; for( i = 0; i < order; i++ ) { - tmp = silk_max( silk_abs_float( coefs_syn[ i ] ), silk_abs_float( coefs_ana[ i ] ) ); + tmp = silk_abs_float( coefs[ i ] ); if( tmp > maxabs ) { maxabs = tmp; ind = i; @@ -94,36 +90,59 @@ static OPUS_INLINE void warped_true2monic_coefs( /* Convert back to true warped coefficients */ for( i = 1; i < order; i++ ) { - coefs_syn[ i - 1 ] += lambda * coefs_syn[ i ]; - coefs_ana[ i - 1 ] += lambda * coefs_ana[ i ]; + coefs[ i - 1 ] += lambda * coefs[ i ]; } - gain_syn = 1.0f / gain_syn; - gain_ana = 1.0f / gain_ana; + gain = 1.0f / gain; for( i = 0; i < order; i++ ) { - coefs_syn[ i ] *= gain_syn; - coefs_ana[ i ] *= gain_ana; + coefs[ i ] *= gain; } /* Apply bandwidth expansion */ chirp = 0.99f - ( 0.8f + 0.1f * iter ) * ( maxabs - limit ) / ( maxabs * ( ind + 1 ) ); - silk_bwexpander_FLP( coefs_syn, order, chirp ); - silk_bwexpander_FLP( coefs_ana, order, chirp ); + silk_bwexpander_FLP( coefs, order, chirp ); /* Convert to monic warped coefficients */ for( i = order - 1; i > 0; i-- ) { - coefs_syn[ i - 1 ] -= lambda * coefs_syn[ i ]; - coefs_ana[ i - 1 ] -= lambda * coefs_ana[ i ]; + coefs[ i - 1 ] -= lambda * coefs[ i ]; } - gain_syn = ( 1.0f - lambda * lambda ) / ( 1.0f + lambda * coefs_syn[ 0 ] ); - gain_ana = ( 1.0f - lambda * lambda ) / ( 1.0f + lambda * coefs_ana[ 0 ] ); + gain = ( 1.0f - lambda * lambda ) / ( 1.0f + lambda * coefs[ 0 ] ); for( i = 0; i < order; i++ ) { - coefs_syn[ i ] *= gain_syn; - coefs_ana[ i ] *= gain_ana; + coefs[ i ] *= gain; } } silk_assert( 0 ); } +static OPUS_INLINE void limit_coefs( + silk_float *coefs, + silk_float limit, + opus_int order +) { + opus_int i, iter, ind = 0; + silk_float tmp, maxabs, chirp; + + for( iter = 0; iter < 10; iter++ ) { + /* Find maximum absolute value */ + maxabs = -1.0f; + for( i = 0; i < order; i++ ) { + tmp = silk_abs_float( coefs[ i ] ); + if( tmp > maxabs ) { + maxabs = tmp; + ind = i; + } + } + if( maxabs <= limit ) { + /* Coefficients are within range - done */ + return; + } + + /* Apply bandwidth expansion */ + chirp = 0.99f - ( 0.8f + 0.1f * iter ) * ( maxabs - limit ) / ( maxabs * ( ind + 1 ) ); + silk_bwexpander_FLP( coefs, order, chirp ); + } + silk_assert( 0 ); +} + /* Compute noise shaping coefficients and initial gain values */ void silk_noise_shape_analysis_FLP( silk_encoder_state_FLP *psEnc, /* I/O Encoder state FLP */ @@ -133,12 +152,13 @@ void silk_noise_shape_analysis_FLP( ) { silk_shape_state_FLP *psShapeSt = &psEnc->sShape; - opus_int k, nSamples; - silk_float SNR_adj_dB, HarmBoost, HarmShapeGain, Tilt; - silk_float nrg, pre_nrg, log_energy, log_energy_prev, energy_variation; - silk_float delta, BWExp1, BWExp2, gain_mult, gain_add, strength, b, warping; + opus_int k, nSamples, nSegs; + silk_float SNR_adj_dB, HarmShapeGain, Tilt; + silk_float nrg, log_energy, log_energy_prev, energy_variation; + silk_float BWExp, gain_mult, gain_add, strength, b, warping; silk_float x_windowed[ SHAPE_LPC_WIN_MAX ]; silk_float auto_corr[ MAX_SHAPE_LPC_ORDER + 1 ]; + silk_float rc[ MAX_SHAPE_LPC_ORDER + 1 ]; const silk_float *x_ptr, *pitch_res_ptr; /* Point to start of first LPC analysis block */ @@ -176,14 +196,14 @@ void silk_noise_shape_analysis_FLP( if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) { /* Initially set to 0; may be overruled in process_gains(..) */ psEnc->sCmn.indices.quantOffsetType = 0; - psEncCtrl->sparseness = 0.0f; } else { /* Sparseness measure, based on relative fluctuations of energy per 2 milliseconds */ nSamples = 2 * psEnc->sCmn.fs_kHz; energy_variation = 0.0f; log_energy_prev = 0.0f; pitch_res_ptr = pitch_res; - for( k = 0; k < silk_SMULBB( SUB_FRAME_LENGTH_MS, psEnc->sCmn.nb_subfr ) / 2; k++ ) { + nSegs = silk_SMULBB( SUB_FRAME_LENGTH_MS, psEnc->sCmn.nb_subfr ) / 2; + for( k = 0; k < nSegs; k++ ) { nrg = ( silk_float )nSamples + ( silk_float )silk_energy_FLP( pitch_res_ptr, nSamples ); log_energy = silk_log2( nrg ); if( k > 0 ) { @@ -192,17 +212,13 @@ void silk_noise_shape_analysis_FLP( log_energy_prev = log_energy; pitch_res_ptr += nSamples; } - psEncCtrl->sparseness = silk_sigmoid( 0.4f * ( energy_variation - 5.0f ) ); /* Set quantization offset depending on sparseness measure */ - if( psEncCtrl->sparseness > SPARSENESS_THRESHOLD_QNT_OFFSET ) { + if( energy_variation > ENERGY_VARIATION_THRESHOLD_QNT_OFFSET * (nSegs-1) ) { psEnc->sCmn.indices.quantOffsetType = 0; } else { psEnc->sCmn.indices.quantOffsetType = 1; } - - /* Increase coding SNR for sparse signals */ - SNR_adj_dB += SPARSE_SNR_INCR_dB * ( psEncCtrl->sparseness - 0.5f ); } /*******************************/ @@ -210,19 +226,10 @@ void silk_noise_shape_analysis_FLP( /*******************************/ /* More BWE for signals with high prediction gain */ strength = FIND_PITCH_WHITE_NOISE_FRACTION * psEncCtrl->predGain; /* between 0.0 and 1.0 */ - BWExp1 = BWExp2 = BANDWIDTH_EXPANSION / ( 1.0f + strength * strength ); - delta = LOW_RATE_BANDWIDTH_EXPANSION_DELTA * ( 1.0f - 0.75f * psEncCtrl->coding_quality ); - BWExp1 -= delta; - BWExp2 += delta; - /* BWExp1 will be applied after BWExp2, so make it relative */ - BWExp1 /= BWExp2; + BWExp = BANDWIDTH_EXPANSION / ( 1.0f + strength * strength ); - if( psEnc->sCmn.warping_Q16 > 0 ) { - /* Slightly more warping in analysis will move quantization noise up in frequency, where it's better masked */ - warping = (silk_float)psEnc->sCmn.warping_Q16 / 65536.0f + 0.01f * psEncCtrl->coding_quality; - } else { - warping = 0.0f; - } + /* Slightly more warping in analysis will move quantization noise up in frequency, where it's better masked */ + warping = (silk_float)psEnc->sCmn.warping_Q16 / 65536.0f + 0.01f * psEncCtrl->coding_quality; /********************************************/ /* Compute noise shaping AR coefs and gains */ @@ -252,37 +259,28 @@ void silk_noise_shape_analysis_FLP( } /* Add white noise, as a fraction of energy */ - auto_corr[ 0 ] += auto_corr[ 0 ] * SHAPE_WHITE_NOISE_FRACTION; + auto_corr[ 0 ] += auto_corr[ 0 ] * SHAPE_WHITE_NOISE_FRACTION + 1.0f; /* Convert correlations to prediction coefficients, and compute residual energy */ - nrg = silk_levinsondurbin_FLP( &psEncCtrl->AR2[ k * MAX_SHAPE_LPC_ORDER ], auto_corr, psEnc->sCmn.shapingLPCOrder ); + nrg = silk_schur_FLP( rc, auto_corr, psEnc->sCmn.shapingLPCOrder ); + silk_k2a_FLP( &psEncCtrl->AR[ k * MAX_SHAPE_LPC_ORDER ], rc, psEnc->sCmn.shapingLPCOrder ); psEncCtrl->Gains[ k ] = ( silk_float )sqrt( nrg ); if( psEnc->sCmn.warping_Q16 > 0 ) { /* Adjust gain for warping */ - psEncCtrl->Gains[ k ] *= warped_gain( &psEncCtrl->AR2[ k * MAX_SHAPE_LPC_ORDER ], warping, psEnc->sCmn.shapingLPCOrder ); + psEncCtrl->Gains[ k ] *= warped_gain( &psEncCtrl->AR[ k * MAX_SHAPE_LPC_ORDER ], warping, psEnc->sCmn.shapingLPCOrder ); } /* Bandwidth expansion for synthesis filter shaping */ - silk_bwexpander_FLP( &psEncCtrl->AR2[ k * MAX_SHAPE_LPC_ORDER ], psEnc->sCmn.shapingLPCOrder, BWExp2 ); + silk_bwexpander_FLP( &psEncCtrl->AR[ k * MAX_SHAPE_LPC_ORDER ], psEnc->sCmn.shapingLPCOrder, BWExp ); - /* Compute noise shaping filter coefficients */ - silk_memcpy( - &psEncCtrl->AR1[ k * MAX_SHAPE_LPC_ORDER ], - &psEncCtrl->AR2[ k * MAX_SHAPE_LPC_ORDER ], - psEnc->sCmn.shapingLPCOrder * sizeof( silk_float ) ); - - /* Bandwidth expansion for analysis filter shaping */ - silk_bwexpander_FLP( &psEncCtrl->AR1[ k * MAX_SHAPE_LPC_ORDER ], psEnc->sCmn.shapingLPCOrder, BWExp1 ); - - /* Ratio of prediction gains, in energy domain */ - pre_nrg = silk_LPC_inverse_pred_gain_FLP( &psEncCtrl->AR2[ k * MAX_SHAPE_LPC_ORDER ], psEnc->sCmn.shapingLPCOrder ); - nrg = silk_LPC_inverse_pred_gain_FLP( &psEncCtrl->AR1[ k * MAX_SHAPE_LPC_ORDER ], psEnc->sCmn.shapingLPCOrder ); - psEncCtrl->GainsPre[ k ] = 1.0f - 0.7f * ( 1.0f - pre_nrg / nrg ); - - /* Convert to monic warped prediction coefficients and limit absolute values */ - warped_true2monic_coefs( &psEncCtrl->AR2[ k * MAX_SHAPE_LPC_ORDER ], &psEncCtrl->AR1[ k * MAX_SHAPE_LPC_ORDER ], - warping, 3.999f, psEnc->sCmn.shapingLPCOrder ); + if( psEnc->sCmn.warping_Q16 > 0 ) { + /* Convert to monic warped prediction coefficients and limit absolute values */ + warped_true2monic_coefs( &psEncCtrl->AR[ k * MAX_SHAPE_LPC_ORDER ], warping, 3.999f, psEnc->sCmn.shapingLPCOrder ); + } else { + /* Limit absolute values */ + limit_coefs( &psEncCtrl->AR[ k * MAX_SHAPE_LPC_ORDER ], 3.999f, psEnc->sCmn.shapingLPCOrder ); + } } /*****************/ @@ -296,11 +294,6 @@ void silk_noise_shape_analysis_FLP( psEncCtrl->Gains[ k ] += gain_add; } - gain_mult = 1.0f + INPUT_TILT + psEncCtrl->coding_quality * HIGH_RATE_INPUT_TILT; - for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) { - psEncCtrl->GainsPre[ k ] *= gain_mult; - } - /************************************************/ /* Control low-frequency shaping and noise tilt */ /************************************************/ @@ -331,12 +324,6 @@ void silk_noise_shape_analysis_FLP( /****************************/ /* HARMONIC SHAPING CONTROL */ /****************************/ - /* Control boosting of harmonic frequencies */ - HarmBoost = LOW_RATE_HARMONIC_BOOST * ( 1.0f - psEncCtrl->coding_quality ) * psEnc->LTPCorr; - - /* More harmonic boost for noisy input signals */ - HarmBoost += LOW_INPUT_QUALITY_HARMONIC_BOOST * ( 1.0f - psEncCtrl->input_quality ); - if( USE_HARM_SHAPING && psEnc->sCmn.indices.signalType == TYPE_VOICED ) { /* Harmonic noise shaping */ HarmShapeGain = HARMONIC_SHAPING; @@ -355,8 +342,6 @@ void silk_noise_shape_analysis_FLP( /* Smooth over subframes */ /*************************/ for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) { - psShapeSt->HarmBoost_smth += SUBFR_SMTH_COEF * ( HarmBoost - psShapeSt->HarmBoost_smth ); - psEncCtrl->HarmBoost[ k ] = psShapeSt->HarmBoost_smth; psShapeSt->HarmShapeGain_smth += SUBFR_SMTH_COEF * ( HarmShapeGain - psShapeSt->HarmShapeGain_smth ); psEncCtrl->HarmShapeGain[ k ] = psShapeSt->HarmShapeGain_smth; psShapeSt->Tilt_smth += SUBFR_SMTH_COEF * ( Tilt - psShapeSt->Tilt_smth ); diff --git a/code/opus-1.1.4/silk/float/pitch_analysis_core_FLP.c b/code/opus-1.1.4/silk/float/pitch_analysis_core_FLP.c index d0e637a2..b3716937 100644 --- a/code/opus-1.1.4/silk/float/pitch_analysis_core_FLP.c +++ b/code/opus-1.1.4/silk/float/pitch_analysis_core_FLP.c @@ -159,7 +159,7 @@ opus_int silk_pitch_analysis_core_FLP( /* O Voicing estimate: 0 voiced, /* Low-pass filter */ for( i = frame_length_4kHz - 1; i > 0; i-- ) { - frame_4kHz[ i ] += frame_4kHz[ i - 1 ]; + frame_4kHz[ i ] = silk_ADD_SAT16( frame_4kHz[ i ], frame_4kHz[ i - 1 ] ); } /****************************************************************************** diff --git a/code/opus-1.1.4/silk/float/prefilter_FLP.c b/code/opus-1.1.4/silk/float/prefilter_FLP.c deleted file mode 100644 index 8bc32fb4..00000000 --- a/code/opus-1.1.4/silk/float/prefilter_FLP.c +++ /dev/null @@ -1,206 +0,0 @@ -/*********************************************************************** -Copyright (c) 2006-2011, Skype Limited. All rights reserved. -Redistribution and use in source and binary forms, with or without -modification, are permitted provided that the following conditions -are met: -- Redistributions of source code must retain the above copyright notice, -this list of conditions and the following disclaimer. -- Redistributions in binary form must reproduce the above copyright -notice, this list of conditions and the following disclaimer in the -documentation and/or other materials provided with the distribution. -- Neither the name of Internet Society, IETF or IETF Trust, nor the -names of specific contributors, may be used to endorse or promote -products derived from this software without specific prior written -permission. -THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" -AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE -IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE -ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE -LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR -CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF -SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS -INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN -CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) -ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE -POSSIBILITY OF SUCH DAMAGE. -***********************************************************************/ - -#ifdef HAVE_CONFIG_H -#include "config.h" -#endif - -#include "main_FLP.h" -#include "tuning_parameters.h" - -/* -* Prefilter for finding Quantizer input signal -*/ -static OPUS_INLINE void silk_prefilt_FLP( - silk_prefilter_state_FLP *P, /* I/O state */ - silk_float st_res[], /* I */ - silk_float xw[], /* O */ - silk_float *HarmShapeFIR, /* I */ - silk_float Tilt, /* I */ - silk_float LF_MA_shp, /* I */ - silk_float LF_AR_shp, /* I */ - opus_int lag, /* I */ - opus_int length /* I */ -); - -static void silk_warped_LPC_analysis_filter_FLP( - silk_float state[], /* I/O State [order + 1] */ - silk_float res[], /* O Residual signal [length] */ - const silk_float coef[], /* I Coefficients [order] */ - const silk_float input[], /* I Input signal [length] */ - const silk_float lambda, /* I Warping factor */ - const opus_int length, /* I Length of input signal */ - const opus_int order /* I Filter order (even) */ -) -{ - opus_int n, i; - silk_float acc, tmp1, tmp2; - - /* Order must be even */ - silk_assert( ( order & 1 ) == 0 ); - - for( n = 0; n < length; n++ ) { - /* Output of lowpass section */ - tmp2 = state[ 0 ] + lambda * state[ 1 ]; - state[ 0 ] = input[ n ]; - /* Output of allpass section */ - tmp1 = state[ 1 ] + lambda * ( state[ 2 ] - tmp2 ); - state[ 1 ] = tmp2; - acc = coef[ 0 ] * tmp2; - /* Loop over allpass sections */ - for( i = 2; i < order; i += 2 ) { - /* Output of allpass section */ - tmp2 = state[ i ] + lambda * ( state[ i + 1 ] - tmp1 ); - state[ i ] = tmp1; - acc += coef[ i - 1 ] * tmp1; - /* Output of allpass section */ - tmp1 = state[ i + 1 ] + lambda * ( state[ i + 2 ] - tmp2 ); - state[ i + 1 ] = tmp2; - acc += coef[ i ] * tmp2; - } - state[ order ] = tmp1; - acc += coef[ order - 1 ] * tmp1; - res[ n ] = input[ n ] - acc; - } -} - -/* -* silk_prefilter. Main prefilter function -*/ -void silk_prefilter_FLP( - silk_encoder_state_FLP *psEnc, /* I/O Encoder state FLP */ - const silk_encoder_control_FLP *psEncCtrl, /* I Encoder control FLP */ - silk_float xw[], /* O Weighted signal */ - const silk_float x[] /* I Speech signal */ -) -{ - silk_prefilter_state_FLP *P = &psEnc->sPrefilt; - opus_int j, k, lag; - silk_float HarmShapeGain, Tilt, LF_MA_shp, LF_AR_shp; - silk_float B[ 2 ]; - const silk_float *AR1_shp; - const silk_float *px; - silk_float *pxw; - silk_float HarmShapeFIR[ 3 ]; - silk_float st_res[ MAX_SUB_FRAME_LENGTH + MAX_LPC_ORDER ]; - - /* Set up pointers */ - px = x; - pxw = xw; - lag = P->lagPrev; - for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) { - /* Update Variables that change per sub frame */ - if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) { - lag = psEncCtrl->pitchL[ k ]; - } - - /* Noise shape parameters */ - HarmShapeGain = psEncCtrl->HarmShapeGain[ k ] * ( 1.0f - psEncCtrl->HarmBoost[ k ] ); - HarmShapeFIR[ 0 ] = 0.25f * HarmShapeGain; - HarmShapeFIR[ 1 ] = 32767.0f / 65536.0f * HarmShapeGain; - HarmShapeFIR[ 2 ] = 0.25f * HarmShapeGain; - Tilt = psEncCtrl->Tilt[ k ]; - LF_MA_shp = psEncCtrl->LF_MA_shp[ k ]; - LF_AR_shp = psEncCtrl->LF_AR_shp[ k ]; - AR1_shp = &psEncCtrl->AR1[ k * MAX_SHAPE_LPC_ORDER ]; - - /* Short term FIR filtering */ - silk_warped_LPC_analysis_filter_FLP( P->sAR_shp, st_res, AR1_shp, px, - (silk_float)psEnc->sCmn.warping_Q16 / 65536.0f, psEnc->sCmn.subfr_length, psEnc->sCmn.shapingLPCOrder ); - - /* Reduce (mainly) low frequencies during harmonic emphasis */ - B[ 0 ] = psEncCtrl->GainsPre[ k ]; - B[ 1 ] = -psEncCtrl->GainsPre[ k ] * - ( psEncCtrl->HarmBoost[ k ] * HarmShapeGain + INPUT_TILT + psEncCtrl->coding_quality * HIGH_RATE_INPUT_TILT ); - pxw[ 0 ] = B[ 0 ] * st_res[ 0 ] + B[ 1 ] * P->sHarmHP; - for( j = 1; j < psEnc->sCmn.subfr_length; j++ ) { - pxw[ j ] = B[ 0 ] * st_res[ j ] + B[ 1 ] * st_res[ j - 1 ]; - } - P->sHarmHP = st_res[ psEnc->sCmn.subfr_length - 1 ]; - - silk_prefilt_FLP( P, pxw, pxw, HarmShapeFIR, Tilt, LF_MA_shp, LF_AR_shp, lag, psEnc->sCmn.subfr_length ); - - px += psEnc->sCmn.subfr_length; - pxw += psEnc->sCmn.subfr_length; - } - P->lagPrev = psEncCtrl->pitchL[ psEnc->sCmn.nb_subfr - 1 ]; -} - -/* -* Prefilter for finding Quantizer input signal -*/ -static OPUS_INLINE void silk_prefilt_FLP( - silk_prefilter_state_FLP *P, /* I/O state */ - silk_float st_res[], /* I */ - silk_float xw[], /* O */ - silk_float *HarmShapeFIR, /* I */ - silk_float Tilt, /* I */ - silk_float LF_MA_shp, /* I */ - silk_float LF_AR_shp, /* I */ - opus_int lag, /* I */ - opus_int length /* I */ -) -{ - opus_int i; - opus_int idx, LTP_shp_buf_idx; - silk_float n_Tilt, n_LF, n_LTP; - silk_float sLF_AR_shp, sLF_MA_shp; - silk_float *LTP_shp_buf; - - /* To speed up use temp variables instead of using the struct */ - LTP_shp_buf = P->sLTP_shp; - LTP_shp_buf_idx = P->sLTP_shp_buf_idx; - sLF_AR_shp = P->sLF_AR_shp; - sLF_MA_shp = P->sLF_MA_shp; - - for( i = 0; i < length; i++ ) { - if( lag > 0 ) { - silk_assert( HARM_SHAPE_FIR_TAPS == 3 ); - idx = lag + LTP_shp_buf_idx; - n_LTP = LTP_shp_buf[ ( idx - HARM_SHAPE_FIR_TAPS / 2 - 1) & LTP_MASK ] * HarmShapeFIR[ 0 ]; - n_LTP += LTP_shp_buf[ ( idx - HARM_SHAPE_FIR_TAPS / 2 ) & LTP_MASK ] * HarmShapeFIR[ 1 ]; - n_LTP += LTP_shp_buf[ ( idx - HARM_SHAPE_FIR_TAPS / 2 + 1) & LTP_MASK ] * HarmShapeFIR[ 2 ]; - } else { - n_LTP = 0; - } - - n_Tilt = sLF_AR_shp * Tilt; - n_LF = sLF_AR_shp * LF_AR_shp + sLF_MA_shp * LF_MA_shp; - - sLF_AR_shp = st_res[ i ] - n_Tilt; - sLF_MA_shp = sLF_AR_shp - n_LF; - - LTP_shp_buf_idx = ( LTP_shp_buf_idx - 1 ) & LTP_MASK; - LTP_shp_buf[ LTP_shp_buf_idx ] = sLF_MA_shp; - - xw[ i ] = sLF_MA_shp - n_LTP; - } - /* Copy temp variable back to state */ - P->sLF_AR_shp = sLF_AR_shp; - P->sLF_MA_shp = sLF_MA_shp; - P->sLTP_shp_buf_idx = LTP_shp_buf_idx; -} diff --git a/code/opus-1.1.4/silk/float/schur_FLP.c b/code/opus-1.1.4/silk/float/schur_FLP.c index ee436f83..c1e0bbb5 100644 --- a/code/opus-1.1.4/silk/float/schur_FLP.c +++ b/code/opus-1.1.4/silk/float/schur_FLP.c @@ -38,22 +38,23 @@ silk_float silk_schur_FLP( /* O returns residual energy ) { opus_int k, n; - silk_float C[ SILK_MAX_ORDER_LPC + 1 ][ 2 ]; - silk_float Ctmp1, Ctmp2, rc_tmp; + double C[ SILK_MAX_ORDER_LPC + 1 ][ 2 ]; + double Ctmp1, Ctmp2, rc_tmp; - silk_assert( order==6||order==8||order==10||order==12||order==14||order==16 ); + silk_assert( order >= 0 && order <= SILK_MAX_ORDER_LPC ); /* Copy correlations */ - for( k = 0; k < order+1; k++ ) { + k = 0; + do { C[ k ][ 0 ] = C[ k ][ 1 ] = auto_corr[ k ]; - } + } while( ++k <= order ); for( k = 0; k < order; k++ ) { /* Get reflection coefficient */ rc_tmp = -C[ k + 1 ][ 0 ] / silk_max_float( C[ 0 ][ 1 ], 1e-9f ); /* Save the output */ - refl_coef[ k ] = rc_tmp; + refl_coef[ k ] = (silk_float)rc_tmp; /* Update correlations */ for( n = 0; n < order - k; n++ ) { @@ -65,6 +66,5 @@ silk_float silk_schur_FLP( /* O returns residual energy } /* Return residual energy */ - return C[ 0 ][ 1 ]; + return (silk_float)C[ 0 ][ 1 ]; } - diff --git a/code/opus-1.1.4/silk/float/solve_LS_FLP.c b/code/opus-1.1.4/silk/float/solve_LS_FLP.c deleted file mode 100644 index b35f0a57..00000000 --- a/code/opus-1.1.4/silk/float/solve_LS_FLP.c +++ /dev/null @@ -1,207 +0,0 @@ -/*********************************************************************** -Copyright (c) 2006-2011, Skype Limited. All rights reserved. -Redistribution and use in source and binary forms, with or without -modification, are permitted provided that the following conditions -are met: -- Redistributions of source code must retain the above copyright notice, -this list of conditions and the following disclaimer. -- Redistributions in binary form must reproduce the above copyright -notice, this list of conditions and the following disclaimer in the -documentation and/or other materials provided with the distribution. -- Neither the name of Internet Society, IETF or IETF Trust, nor the -names of specific contributors, may be used to endorse or promote -products derived from this software without specific prior written -permission. -THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" -AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE -IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE -ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE -LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR -CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF -SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS -INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN -CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) -ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE -POSSIBILITY OF SUCH DAMAGE. -***********************************************************************/ - -#ifdef HAVE_CONFIG_H -#include "config.h" -#endif - -#include "main_FLP.h" -#include "tuning_parameters.h" - -/********************************************************************** - * LDL Factorisation. Finds the upper triangular matrix L and the diagonal - * Matrix D (only the diagonal elements returned in a vector)such that - * the symmetric matric A is given by A = L*D*L'. - **********************************************************************/ -static OPUS_INLINE void silk_LDL_FLP( - silk_float *A, /* I/O Pointer to Symetric Square Matrix */ - opus_int M, /* I Size of Matrix */ - silk_float *L, /* I/O Pointer to Square Upper triangular Matrix */ - silk_float *Dinv /* I/O Pointer to vector holding the inverse diagonal elements of D */ -); - -/********************************************************************** - * Function to solve linear equation Ax = b, when A is a MxM lower - * triangular matrix, with ones on the diagonal. - **********************************************************************/ -static OPUS_INLINE void silk_SolveWithLowerTriangularWdiagOnes_FLP( - const silk_float *L, /* I Pointer to Lower Triangular Matrix */ - opus_int M, /* I Dim of Matrix equation */ - const silk_float *b, /* I b Vector */ - silk_float *x /* O x Vector */ -); - -/********************************************************************** - * Function to solve linear equation (A^T)x = b, when A is a MxM lower - * triangular, with ones on the diagonal. (ie then A^T is upper triangular) - **********************************************************************/ -static OPUS_INLINE void silk_SolveWithUpperTriangularFromLowerWdiagOnes_FLP( - const silk_float *L, /* I Pointer to Lower Triangular Matrix */ - opus_int M, /* I Dim of Matrix equation */ - const silk_float *b, /* I b Vector */ - silk_float *x /* O x Vector */ -); - -/********************************************************************** - * Function to solve linear equation Ax = b, when A is a MxM - * symmetric square matrix - using LDL factorisation - **********************************************************************/ -void silk_solve_LDL_FLP( - silk_float *A, /* I/O Symmetric square matrix, out: reg. */ - const opus_int M, /* I Size of matrix */ - const silk_float *b, /* I Pointer to b vector */ - silk_float *x /* O Pointer to x solution vector */ -) -{ - opus_int i; - silk_float L[ MAX_MATRIX_SIZE ][ MAX_MATRIX_SIZE ]; - silk_float T[ MAX_MATRIX_SIZE ]; - silk_float Dinv[ MAX_MATRIX_SIZE ]; /* inverse diagonal elements of D*/ - - silk_assert( M <= MAX_MATRIX_SIZE ); - - /*************************************************** - Factorize A by LDL such that A = L*D*(L^T), - where L is lower triangular with ones on diagonal - ****************************************************/ - silk_LDL_FLP( A, M, &L[ 0 ][ 0 ], Dinv ); - - /**************************************************** - * substitute D*(L^T) = T. ie: - L*D*(L^T)*x = b => L*T = b <=> T = inv(L)*b - ******************************************************/ - silk_SolveWithLowerTriangularWdiagOnes_FLP( &L[ 0 ][ 0 ], M, b, T ); - - /**************************************************** - D*(L^T)*x = T <=> (L^T)*x = inv(D)*T, because D is - diagonal just multiply with 1/d_i - ****************************************************/ - for( i = 0; i < M; i++ ) { - T[ i ] = T[ i ] * Dinv[ i ]; - } - /**************************************************** - x = inv(L') * inv(D) * T - *****************************************************/ - silk_SolveWithUpperTriangularFromLowerWdiagOnes_FLP( &L[ 0 ][ 0 ], M, T, x ); -} - -static OPUS_INLINE void silk_SolveWithUpperTriangularFromLowerWdiagOnes_FLP( - const silk_float *L, /* I Pointer to Lower Triangular Matrix */ - opus_int M, /* I Dim of Matrix equation */ - const silk_float *b, /* I b Vector */ - silk_float *x /* O x Vector */ -) -{ - opus_int i, j; - silk_float temp; - const silk_float *ptr1; - - for( i = M - 1; i >= 0; i-- ) { - ptr1 = matrix_adr( L, 0, i, M ); - temp = 0; - for( j = M - 1; j > i ; j-- ) { - temp += ptr1[ j * M ] * x[ j ]; - } - temp = b[ i ] - temp; - x[ i ] = temp; - } -} - -static OPUS_INLINE void silk_SolveWithLowerTriangularWdiagOnes_FLP( - const silk_float *L, /* I Pointer to Lower Triangular Matrix */ - opus_int M, /* I Dim of Matrix equation */ - const silk_float *b, /* I b Vector */ - silk_float *x /* O x Vector */ -) -{ - opus_int i, j; - silk_float temp; - const silk_float *ptr1; - - for( i = 0; i < M; i++ ) { - ptr1 = matrix_adr( L, i, 0, M ); - temp = 0; - for( j = 0; j < i; j++ ) { - temp += ptr1[ j ] * x[ j ]; - } - temp = b[ i ] - temp; - x[ i ] = temp; - } -} - -static OPUS_INLINE void silk_LDL_FLP( - silk_float *A, /* I/O Pointer to Symetric Square Matrix */ - opus_int M, /* I Size of Matrix */ - silk_float *L, /* I/O Pointer to Square Upper triangular Matrix */ - silk_float *Dinv /* I/O Pointer to vector holding the inverse diagonal elements of D */ -) -{ - opus_int i, j, k, loop_count, err = 1; - silk_float *ptr1, *ptr2; - double temp, diag_min_value; - silk_float v[ MAX_MATRIX_SIZE ] = { 0 }, D[ MAX_MATRIX_SIZE ]; /* temp arrays*/ - - silk_assert( M <= MAX_MATRIX_SIZE ); - - diag_min_value = FIND_LTP_COND_FAC * 0.5f * ( A[ 0 ] + A[ M * M - 1 ] ); - for( loop_count = 0; loop_count < M && err == 1; loop_count++ ) { - err = 0; - for( j = 0; j < M; j++ ) { - ptr1 = matrix_adr( L, j, 0, M ); - temp = matrix_ptr( A, j, j, M ); /* element in row j column j*/ - for( i = 0; i < j; i++ ) { - v[ i ] = ptr1[ i ] * D[ i ]; - temp -= ptr1[ i ] * v[ i ]; - } - if( temp < diag_min_value ) { - /* Badly conditioned matrix: add white noise and run again */ - temp = ( loop_count + 1 ) * diag_min_value - temp; - for( i = 0; i < M; i++ ) { - matrix_ptr( A, i, i, M ) += ( silk_float )temp; - } - err = 1; - break; - } - D[ j ] = ( silk_float )temp; - Dinv[ j ] = ( silk_float )( 1.0f / temp ); - matrix_ptr( L, j, j, M ) = 1.0f; - - ptr1 = matrix_adr( A, j, 0, M ); - ptr2 = matrix_adr( L, j + 1, 0, M); - for( i = j + 1; i < M; i++ ) { - temp = 0.0; - for( k = 0; k < j; k++ ) { - temp += ptr2[ k ] * v[ k ]; - } - matrix_ptr( L, i, j, M ) = ( silk_float )( ( ptr1[ i ] - temp ) * Dinv[ j ] ); - ptr2 += M; /* go to next column*/ - } - } - } - silk_assert( err == 0 ); -} - diff --git a/code/opus-1.1.4/silk/float/structs_FLP.h b/code/opus-1.1.4/silk/float/structs_FLP.h index 14d647ce..3150b386 100644 --- a/code/opus-1.1.4/silk/float/structs_FLP.h +++ b/code/opus-1.1.4/silk/float/structs_FLP.h @@ -42,32 +42,16 @@ extern "C" /********************************/ typedef struct { opus_int8 LastGainIndex; - silk_float HarmBoost_smth; silk_float HarmShapeGain_smth; silk_float Tilt_smth; } silk_shape_state_FLP; -/********************************/ -/* Prefilter state */ -/********************************/ -typedef struct { - silk_float sLTP_shp[ LTP_BUF_LENGTH ]; - silk_float sAR_shp[ MAX_SHAPE_LPC_ORDER + 1 ]; - opus_int sLTP_shp_buf_idx; - silk_float sLF_AR_shp; - silk_float sLF_MA_shp; - silk_float sHarmHP; - opus_int32 rand_seed; - opus_int lagPrev; -} silk_prefilter_state_FLP; - /********************************/ /* Encoder state FLP */ /********************************/ typedef struct { silk_encoder_state sCmn; /* Common struct, shared with fixed-point code */ silk_shape_state_FLP sShape; /* Noise shaping state */ - silk_prefilter_state_FLP sPrefilt; /* Prefilter State */ /* Buffer for find pitch and noise shape analysis */ silk_float x_buf[ 2 * MAX_FRAME_LENGTH + LA_SHAPE_MAX ];/* Buffer for find pitch and noise shape analysis */ @@ -86,12 +70,9 @@ typedef struct { opus_int pitchL[ MAX_NB_SUBFR ]; /* Noise shaping parameters */ - silk_float AR1[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ]; - silk_float AR2[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ]; + silk_float AR[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ]; silk_float LF_MA_shp[ MAX_NB_SUBFR ]; silk_float LF_AR_shp[ MAX_NB_SUBFR ]; - silk_float GainsPre[ MAX_NB_SUBFR ]; - silk_float HarmBoost[ MAX_NB_SUBFR ]; silk_float Tilt[ MAX_NB_SUBFR ]; silk_float HarmShapeGain[ MAX_NB_SUBFR ]; silk_float Lambda; @@ -99,7 +80,6 @@ typedef struct { silk_float coding_quality; /* Measures */ - silk_float sparseness; silk_float predGain; silk_float LTPredCodGain; silk_float ResNrg[ MAX_NB_SUBFR ]; /* Residual energy per subframe */ diff --git a/code/opus-1.1.4/silk/float/wrappers_FLP.c b/code/opus-1.1.4/silk/float/wrappers_FLP.c index 6666b8ef..ad90b874 100644 --- a/code/opus-1.1.4/silk/float/wrappers_FLP.c +++ b/code/opus-1.1.4/silk/float/wrappers_FLP.c @@ -54,13 +54,14 @@ void silk_A2NLSF_FLP( void silk_NLSF2A_FLP( silk_float *pAR, /* O LPC coefficients [ LPC_order ] */ const opus_int16 *NLSF_Q15, /* I NLSF vector [ LPC_order ] */ - const opus_int LPC_order /* I LPC order */ + const opus_int LPC_order, /* I LPC order */ + int arch /* I Run-time architecture */ ) { opus_int i; opus_int16 a_fix_Q12[ MAX_LPC_ORDER ]; - silk_NLSF2A( a_fix_Q12, NLSF_Q15, LPC_order ); + silk_NLSF2A( a_fix_Q12, NLSF_Q15, LPC_order, arch ); for( i = 0; i < LPC_order; i++ ) { pAR[ i ] = ( silk_float )a_fix_Q12[ i ] * ( 1.0f / 4096.0f ); @@ -102,14 +103,14 @@ void silk_NSQ_wrapper_FLP( ) { opus_int i, j; - opus_int32 x_Q3[ MAX_FRAME_LENGTH ]; + opus_int16 x16[ MAX_FRAME_LENGTH ]; opus_int32 Gains_Q16[ MAX_NB_SUBFR ]; silk_DWORD_ALIGN opus_int16 PredCoef_Q12[ 2 ][ MAX_LPC_ORDER ]; opus_int16 LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ]; opus_int LTP_scale_Q14; /* Noise shaping parameters */ - opus_int16 AR2_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ]; + opus_int16 AR_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ]; opus_int32 LF_shp_Q14[ MAX_NB_SUBFR ]; /* Packs two int16 coefficients per int32 value */ opus_int Lambda_Q10; opus_int Tilt_Q14[ MAX_NB_SUBFR ]; @@ -119,7 +120,7 @@ void silk_NSQ_wrapper_FLP( /* Noise shape parameters */ for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) { for( j = 0; j < psEnc->sCmn.shapingLPCOrder; j++ ) { - AR2_Q13[ i * MAX_SHAPE_LPC_ORDER + j ] = silk_float2int( psEncCtrl->AR2[ i * MAX_SHAPE_LPC_ORDER + j ] * 8192.0f ); + AR_Q13[ i * MAX_SHAPE_LPC_ORDER + j ] = silk_float2int( psEncCtrl->AR[ i * MAX_SHAPE_LPC_ORDER + j ] * 8192.0f ); } } @@ -155,16 +156,16 @@ void silk_NSQ_wrapper_FLP( /* Convert input to fix */ for( i = 0; i < psEnc->sCmn.frame_length; i++ ) { - x_Q3[ i ] = silk_float2int( 8.0f * x[ i ] ); + x16[ i ] = silk_float2int( x[ i ] ); } /* Call NSQ */ if( psEnc->sCmn.nStatesDelayedDecision > 1 || psEnc->sCmn.warping_Q16 > 0 ) { - silk_NSQ_del_dec( &psEnc->sCmn, psNSQ, psIndices, x_Q3, pulses, PredCoef_Q12[ 0 ], LTPCoef_Q14, - AR2_Q13, HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, psEncCtrl->pitchL, Lambda_Q10, LTP_scale_Q14, psEnc->sCmn.arch ); + silk_NSQ_del_dec( &psEnc->sCmn, psNSQ, psIndices, x16, pulses, PredCoef_Q12[ 0 ], LTPCoef_Q14, + AR_Q13, HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, psEncCtrl->pitchL, Lambda_Q10, LTP_scale_Q14, psEnc->sCmn.arch ); } else { - silk_NSQ( &psEnc->sCmn, psNSQ, psIndices, x_Q3, pulses, PredCoef_Q12[ 0 ], LTPCoef_Q14, - AR2_Q13, HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, psEncCtrl->pitchL, Lambda_Q10, LTP_scale_Q14, psEnc->sCmn.arch ); + silk_NSQ( &psEnc->sCmn, psNSQ, psIndices, x16, pulses, PredCoef_Q12[ 0 ], LTPCoef_Q14, + AR_Q13, HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, psEncCtrl->pitchL, Lambda_Q10, LTP_scale_Q14, psEnc->sCmn.arch ); } } @@ -172,31 +173,35 @@ void silk_NSQ_wrapper_FLP( /* Floating-point Silk LTP quantiation wrapper */ /***********************************************/ void silk_quant_LTP_gains_FLP( - silk_float B[ MAX_NB_SUBFR * LTP_ORDER ], /* I/O (Un-)quantized LTP gains */ + silk_float B[ MAX_NB_SUBFR * LTP_ORDER ], /* O Quantized LTP gains */ opus_int8 cbk_index[ MAX_NB_SUBFR ], /* O Codebook index */ opus_int8 *periodicity_index, /* O Periodicity index */ opus_int32 *sum_log_gain_Q7, /* I/O Cumulative max prediction gain */ - const silk_float W[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ], /* I Error weights */ - const opus_int mu_Q10, /* I Mu value (R/D tradeoff) */ - const opus_int lowComplexity, /* I Flag for low complexity */ - const opus_int nb_subfr, /* I number of subframes */ + silk_float *pred_gain_dB, /* O LTP prediction gain */ + const silk_float XX[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ], /* I Correlation matrix */ + const silk_float xX[ MAX_NB_SUBFR * LTP_ORDER ], /* I Correlation vector */ + const opus_int subfr_len, /* I Number of samples per subframe */ + const opus_int nb_subfr, /* I Number of subframes */ int arch /* I Run-time architecture */ ) { - opus_int i; + opus_int i, pred_gain_dB_Q7; opus_int16 B_Q14[ MAX_NB_SUBFR * LTP_ORDER ]; - opus_int32 W_Q18[ MAX_NB_SUBFR*LTP_ORDER*LTP_ORDER ]; + opus_int32 XX_Q17[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ]; + opus_int32 xX_Q17[ MAX_NB_SUBFR * LTP_ORDER ]; - for( i = 0; i < nb_subfr * LTP_ORDER; i++ ) { - B_Q14[ i ] = (opus_int16)silk_float2int( B[ i ] * 16384.0f ); - } for( i = 0; i < nb_subfr * LTP_ORDER * LTP_ORDER; i++ ) { - W_Q18[ i ] = (opus_int32)silk_float2int( W[ i ] * 262144.0f ); + XX_Q17[ i ] = (opus_int32)silk_float2int( XX[ i ] * 131072.0f ); + } + for( i = 0; i < nb_subfr * LTP_ORDER; i++ ) { + xX_Q17[ i ] = (opus_int32)silk_float2int( xX[ i ] * 131072.0f ); } - silk_quant_LTP_gains( B_Q14, cbk_index, periodicity_index, sum_log_gain_Q7, W_Q18, mu_Q10, lowComplexity, nb_subfr, arch ); + silk_quant_LTP_gains( B_Q14, cbk_index, periodicity_index, sum_log_gain_Q7, &pred_gain_dB_Q7, XX_Q17, xX_Q17, subfr_len, nb_subfr, arch ); for( i = 0; i < nb_subfr * LTP_ORDER; i++ ) { B[ i ] = (silk_float)B_Q14[ i ] * ( 1.0f / 16384.0f ); } + + *pred_gain_dB = (silk_float)pred_gain_dB_Q7 * ( 1.0f / 128.0f ); } diff --git a/code/opus-1.1.4/silk/gain_quant.c b/code/opus-1.1.4/silk/gain_quant.c index 64ccd061..ee65245a 100644 --- a/code/opus-1.1.4/silk/gain_quant.c +++ b/code/opus-1.1.4/silk/gain_quant.c @@ -76,6 +76,7 @@ void silk_gains_quant( /* Accumulate deltas */ if( ind[ k ] > double_step_size_threshold ) { *prev_ind += silk_LSHIFT( ind[ k ], 1 ) - double_step_size_threshold; + *prev_ind = silk_min_int( *prev_ind, N_LEVELS_QGAIN - 1 ); } else { *prev_ind += ind[ k ]; } diff --git a/code/opus-1.1.4/silk/init_decoder.c b/code/opus-1.1.4/silk/init_decoder.c index f887c678..16c03dcd 100644 --- a/code/opus-1.1.4/silk/init_decoder.c +++ b/code/opus-1.1.4/silk/init_decoder.c @@ -44,6 +44,7 @@ opus_int silk_init_decoder( /* Used to deactivate LSF interpolation */ psDec->first_frame_after_reset = 1; psDec->prev_gain_Q16 = 65536; + psDec->arch = opus_select_arch(); /* Reset CNG state */ silk_CNG_Reset( psDec ); diff --git a/code/opus-1.1.4/silk/lin2log.c b/code/opus-1.1.4/silk/lin2log.c index d4fe5153..0d5155aa 100644 --- a/code/opus-1.1.4/silk/lin2log.c +++ b/code/opus-1.1.4/silk/lin2log.c @@ -41,6 +41,6 @@ opus_int32 silk_lin2log( silk_CLZ_FRAC( inLin, &lz, &frac_Q7 ); /* Piece-wise parabolic approximation */ - return silk_LSHIFT( 31 - lz, 7 ) + silk_SMLAWB( frac_Q7, silk_MUL( frac_Q7, 128 - frac_Q7 ), 179 ); + return silk_ADD_LSHIFT32( silk_SMLAWB( frac_Q7, silk_MUL( frac_Q7, 128 - frac_Q7 ), 179 ), 31 - lz, 7 ); } diff --git a/code/opus-1.1.4/silk/macros.h b/code/opus-1.1.4/silk/macros.h index d3ca3475..3c67b6e5 100644 --- a/code/opus-1.1.4/silk/macros.h +++ b/code/opus-1.1.4/silk/macros.h @@ -36,14 +36,6 @@ POSSIBILITY OF SUCH DAMAGE. #include "opus_defines.h" #include "arch.h" -#if OPUS_GNUC_PREREQ(3, 0) -#define opus_likely(x) (__builtin_expect(!!(x), 1)) -#define opus_unlikely(x) (__builtin_expect(!!(x), 0)) -#else -#define opus_likely(x) (!!(x)) -#define opus_unlikely(x) (!!(x)) -#endif - /* This is an OPUS_INLINE header file for general platform. */ /* (a32 * (opus_int32)((opus_int16)(b32))) >> 16 output have to be 32bit int */ diff --git a/code/opus-1.1.4/silk/main.h b/code/opus-1.1.4/silk/main.h index 2f90d68f..1a33eed5 100644 --- a/code/opus-1.1.4/silk/main.h +++ b/code/opus-1.1.4/silk/main.h @@ -42,6 +42,10 @@ POSSIBILITY OF SUCH DAMAGE. #include "x86/main_sse.h" #endif +#if (defined(OPUS_ARM_ASM) || defined(OPUS_ARM_MAY_HAVE_NEON_INTR)) +#include "arm/NSQ_del_dec_arm.h" +#endif + /* Convert Left/Right stereo signal to adaptive Mid/Side representation */ void silk_stereo_LR_to_MS( stereo_enc_state *state, /* I/O State */ @@ -109,22 +113,22 @@ void silk_stereo_decode_mid_only( /* Encodes signs of excitation */ void silk_encode_signs( - ec_enc *psRangeEnc, /* I/O Compressor data structure */ - const opus_int8 pulses[], /* I pulse signal */ - opus_int length, /* I length of input */ - const opus_int signalType, /* I Signal type */ - const opus_int quantOffsetType, /* I Quantization offset type */ - const opus_int sum_pulses[ MAX_NB_SHELL_BLOCKS ] /* I Sum of absolute pulses per block */ + ec_enc *psRangeEnc, /* I/O Compressor data structure */ + const opus_int8 pulses[], /* I pulse signal */ + opus_int length, /* I length of input */ + const opus_int signalType, /* I Signal type */ + const opus_int quantOffsetType, /* I Quantization offset type */ + const opus_int sum_pulses[ MAX_NB_SHELL_BLOCKS ] /* I Sum of absolute pulses per block */ ); /* Decodes signs of excitation */ void silk_decode_signs( - ec_dec *psRangeDec, /* I/O Compressor data structure */ - opus_int16 pulses[], /* I/O pulse signal */ - opus_int length, /* I length of input */ - const opus_int signalType, /* I Signal type */ - const opus_int quantOffsetType, /* I Quantization offset type */ - const opus_int sum_pulses[ MAX_NB_SHELL_BLOCKS ] /* I Sum of absolute pulses per block */ + ec_dec *psRangeDec, /* I/O Compressor data structure */ + opus_int16 pulses[], /* I/O pulse signal */ + opus_int length, /* I length of input */ + const opus_int signalType, /* I Signal type */ + const opus_int quantOffsetType, /* I Quantization offset type */ + const opus_int sum_pulses[ MAX_NB_SHELL_BLOCKS ] /* I Sum of absolute pulses per block */ ); /* Check encoder control struct */ @@ -205,37 +209,37 @@ void silk_interpolate( /* LTP tap quantizer */ void silk_quant_LTP_gains( - opus_int16 B_Q14[ MAX_NB_SUBFR * LTP_ORDER ], /* I/O (un)quantized LTP gains */ + opus_int16 B_Q14[ MAX_NB_SUBFR * LTP_ORDER ], /* O Quantized LTP gains */ opus_int8 cbk_index[ MAX_NB_SUBFR ], /* O Codebook Index */ opus_int8 *periodicity_index, /* O Periodicity Index */ opus_int32 *sum_gain_dB_Q7, /* I/O Cumulative max prediction gain */ - const opus_int32 W_Q18[ MAX_NB_SUBFR*LTP_ORDER*LTP_ORDER ], /* I Error Weights in Q18 */ - opus_int mu_Q9, /* I Mu value (R/D tradeoff) */ - opus_int lowComplexity, /* I Flag for low complexity */ - const opus_int nb_subfr, /* I number of subframes */ + opus_int *pred_gain_dB_Q7, /* O LTP prediction gain */ + const opus_int32 XX_Q17[ MAX_NB_SUBFR*LTP_ORDER*LTP_ORDER ], /* I Correlation matrix in Q18 */ + const opus_int32 xX_Q17[ MAX_NB_SUBFR*LTP_ORDER ], /* I Correlation vector in Q18 */ + const opus_int subfr_len, /* I Number of samples per subframe */ + const opus_int nb_subfr, /* I Number of subframes */ int arch /* I Run-time architecture */ ); /* Entropy constrained matrix-weighted VQ, for a single input data vector */ void silk_VQ_WMat_EC_c( opus_int8 *ind, /* O index of best codebook vector */ - opus_int32 *rate_dist_Q14, /* O best weighted quant error + mu * rate */ + opus_int32 *res_nrg_Q15, /* O best residual energy */ + opus_int32 *rate_dist_Q8, /* O best total bitrate */ opus_int *gain_Q7, /* O sum of absolute LTP coefficients */ - const opus_int16 *in_Q14, /* I input vector to be quantized */ - const opus_int32 *W_Q18, /* I weighting matrix */ + const opus_int32 *XX_Q17, /* I correlation matrix */ + const opus_int32 *xX_Q17, /* I correlation vector */ const opus_int8 *cb_Q7, /* I codebook */ const opus_uint8 *cb_gain_Q7, /* I codebook effective gain */ const opus_uint8 *cl_Q5, /* I code length for each codebook vector */ - const opus_int mu_Q9, /* I tradeoff betw. weighted error and rate */ + const opus_int subfr_len, /* I number of samples per subframe */ const opus_int32 max_gain_Q7, /* I maximum sum of absolute LTP coefficients */ - opus_int L /* I number of vectors in codebook */ + const opus_int L /* I number of vectors in codebook */ ); #if !defined(OVERRIDE_silk_VQ_WMat_EC) -#define silk_VQ_WMat_EC(ind, rate_dist_Q14, gain_Q7, in_Q14, W_Q18, cb_Q7, cb_gain_Q7, cl_Q5, \ - mu_Q9, max_gain_Q7, L, arch) \ - ((void)(arch),silk_VQ_WMat_EC_c(ind, rate_dist_Q14, gain_Q7, in_Q14, W_Q18, cb_Q7, cb_gain_Q7, cl_Q5, \ - mu_Q9, max_gain_Q7, L)) +#define silk_VQ_WMat_EC(ind, res_nrg_Q15, rate_dist_Q8, gain_Q7, XX_Q17, xX_Q17, cb_Q7, cb_gain_Q7, cl_Q5, subfr_len, max_gain_Q7, L, arch) \ + ((void)(arch),silk_VQ_WMat_EC_c(ind, res_nrg_Q15, rate_dist_Q8, gain_Q7, XX_Q17, xX_Q17, cb_Q7, cb_gain_Q7, cl_Q5, subfr_len, max_gain_Q7, L)) #endif /************************************/ @@ -243,14 +247,14 @@ void silk_VQ_WMat_EC_c( /************************************/ void silk_NSQ_c( - const silk_encoder_state *psEncC, /* I/O Encoder State */ + const silk_encoder_state *psEncC, /* I Encoder State */ silk_nsq_state *NSQ, /* I/O NSQ state */ SideInfoIndices *psIndices, /* I/O Quantization Indices */ - const opus_int32 x_Q3[], /* I Prefiltered input signal */ + const opus_int16 x16[], /* I Input */ opus_int8 pulses[], /* O Quantized pulse signal */ const opus_int16 PredCoef_Q12[ 2 * MAX_LPC_ORDER ], /* I Short term prediction coefs */ const opus_int16 LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ], /* I Long term prediction coefs */ - const opus_int16 AR2_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs */ + const opus_int16 AR_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs */ const opus_int HarmShapeGain_Q14[ MAX_NB_SUBFR ], /* I Long term shaping coefs */ const opus_int Tilt_Q14[ MAX_NB_SUBFR ], /* I Spectral tilt */ const opus_int32 LF_shp_Q14[ MAX_NB_SUBFR ], /* I Low frequency shaping coefs */ @@ -261,22 +265,22 @@ void silk_NSQ_c( ); #if !defined(OVERRIDE_silk_NSQ) -#define silk_NSQ(psEncC, NSQ, psIndices, x_Q3, pulses, PredCoef_Q12, LTPCoef_Q14, AR2_Q13, \ +#define silk_NSQ(psEncC, NSQ, psIndices, x16, pulses, PredCoef_Q12, LTPCoef_Q14, AR_Q13, \ HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, LTP_scale_Q14, arch) \ - ((void)(arch),silk_NSQ_c(psEncC, NSQ, psIndices, x_Q3, pulses, PredCoef_Q12, LTPCoef_Q14, AR2_Q13, \ + ((void)(arch),silk_NSQ_c(psEncC, NSQ, psIndices, x16, pulses, PredCoef_Q12, LTPCoef_Q14, AR_Q13, \ HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, LTP_scale_Q14)) #endif /* Noise shaping using delayed decision */ void silk_NSQ_del_dec_c( - const silk_encoder_state *psEncC, /* I/O Encoder State */ + const silk_encoder_state *psEncC, /* I Encoder State */ silk_nsq_state *NSQ, /* I/O NSQ state */ SideInfoIndices *psIndices, /* I/O Quantization Indices */ - const opus_int32 x_Q3[], /* I Prefiltered input signal */ + const opus_int16 x16[], /* I Input */ opus_int8 pulses[], /* O Quantized pulse signal */ const opus_int16 PredCoef_Q12[ 2 * MAX_LPC_ORDER ], /* I Short term prediction coefs */ const opus_int16 LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ], /* I Long term prediction coefs */ - const opus_int16 AR2_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs */ + const opus_int16 AR_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs */ const opus_int HarmShapeGain_Q14[ MAX_NB_SUBFR ], /* I Long term shaping coefs */ const opus_int Tilt_Q14[ MAX_NB_SUBFR ], /* I Spectral tilt */ const opus_int32 LF_shp_Q14[ MAX_NB_SUBFR ], /* I Low frequency shaping coefs */ @@ -287,9 +291,9 @@ void silk_NSQ_del_dec_c( ); #if !defined(OVERRIDE_silk_NSQ_del_dec) -#define silk_NSQ_del_dec(psEncC, NSQ, psIndices, x_Q3, pulses, PredCoef_Q12, LTPCoef_Q14, AR2_Q13, \ +#define silk_NSQ_del_dec(psEncC, NSQ, psIndices, x16, pulses, PredCoef_Q12, LTPCoef_Q14, AR_Q13, \ HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, LTP_scale_Q14, arch) \ - ((void)(arch),silk_NSQ_del_dec_c(psEncC, NSQ, psIndices, x_Q3, pulses, PredCoef_Q12, LTPCoef_Q14, AR2_Q13, \ + ((void)(arch),silk_NSQ_del_dec_c(psEncC, NSQ, psIndices, x16, pulses, PredCoef_Q12, LTPCoef_Q14, AR_Q13, \ HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, LTP_scale_Q14)) #endif @@ -346,6 +350,7 @@ void silk_NLSF_VQ( opus_int32 err_Q26[], /* O Quantization errors [K] */ const opus_int16 in_Q15[], /* I Input vectors to be quantized [LPC_order] */ const opus_uint8 pCB_Q8[], /* I Codebook vectors [K*LPC_order] */ + const opus_int16 pWght_Q9[], /* I Codebook weights [K*LPC_order] */ const opus_int K, /* I Number of codebook vectors */ const opus_int LPC_order /* I Number of LPCs */ ); diff --git a/code/opus-1.1.4/silk/mips/NSQ_del_dec_mipsr1.h b/code/opus-1.1.4/silk/mips/NSQ_del_dec_mipsr1.h index ad1cfe2a..cd70713a 100644 --- a/code/opus-1.1.4/silk/mips/NSQ_del_dec_mipsr1.h +++ b/code/opus-1.1.4/silk/mips/NSQ_del_dec_mipsr1.h @@ -61,7 +61,7 @@ static inline void silk_noise_shape_quantizer_del_dec( opus_int predictLPCOrder, /* I Prediction filter order */ opus_int warping_Q16, /* I */ opus_int nStatesDelayedDecision, /* I Number of states in decision tree */ - opus_int *smpl_buf_idx, /* I Index to newest samples in buffers */ + opus_int *smpl_buf_idx, /* I/O Index to newest samples in buffers */ opus_int decisionDelay, /* I */ int arch /* I */ ) @@ -323,8 +323,9 @@ static inline void silk_noise_shape_quantizer_del_dec( psSS[ 1 ].xq_Q14 = xq_Q14; } - *smpl_buf_idx = ( *smpl_buf_idx - 1 ) & DECISION_DELAY_MASK; /* Index to newest samples */ - last_smple_idx = ( *smpl_buf_idx + decisionDelay ) & DECISION_DELAY_MASK; /* Index to decisionDelay old samples */ + *smpl_buf_idx = ( *smpl_buf_idx - 1 ) % DECISION_DELAY; + if( *smpl_buf_idx < 0 ) *smpl_buf_idx += DECISION_DELAY; + last_smple_idx = ( *smpl_buf_idx + decisionDelay ) % DECISION_DELAY; /* Find winner */ RDmin_Q10 = psSampleState[ 0 ][ 0 ].RD_Q10; diff --git a/code/opus-1.1.4/silk/mips/sigproc_fix_mipsr1.h b/code/opus-1.1.4/silk/mips/sigproc_fix_mipsr1.h index 3b0a6953..51520c0a 100644 --- a/code/opus-1.1.4/silk/mips/sigproc_fix_mipsr1.h +++ b/code/opus-1.1.4/silk/mips/sigproc_fix_mipsr1.h @@ -28,11 +28,6 @@ POSSIBILITY OF SUCH DAMAGE. #ifndef SILK_SIGPROC_FIX_MIPSR1_H #define SILK_SIGPROC_FIX_MIPSR1_H -#ifdef __cplusplus -extern "C" -{ -#endif - #undef silk_SAT16 static inline short int silk_SAT16(int a) { diff --git a/code/opus-1.1.4/silk/process_NLSFs.c b/code/opus-1.1.4/silk/process_NLSFs.c index 0ab71f01..2f10f8df 100644 --- a/code/opus-1.1.4/silk/process_NLSFs.c +++ b/code/opus-1.1.4/silk/process_NLSFs.c @@ -89,7 +89,7 @@ void silk_process_NLSFs( NLSF_mu_Q20, psEncC->NLSF_MSVQ_Survivors, psEncC->indices.signalType ); /* Convert quantized NLSFs back to LPC coefficients */ - silk_NLSF2A( PredCoef_Q12[ 1 ], pNLSF_Q15, psEncC->predictLPCOrder ); + silk_NLSF2A( PredCoef_Q12[ 1 ], pNLSF_Q15, psEncC->predictLPCOrder, psEncC->arch ); if( doInterpolate ) { /* Calculate the interpolated, quantized LSF vector for the first half */ @@ -97,7 +97,7 @@ void silk_process_NLSFs( psEncC->indices.NLSFInterpCoef_Q2, psEncC->predictLPCOrder ); /* Convert back to LPC coefficients */ - silk_NLSF2A( PredCoef_Q12[ 0 ], pNLSF0_temp_Q15, psEncC->predictLPCOrder ); + silk_NLSF2A( PredCoef_Q12[ 0 ], pNLSF0_temp_Q15, psEncC->predictLPCOrder, psEncC->arch ); } else { /* Copy LPC coefficients for first half from second half */ diff --git a/code/opus-1.1.4/silk/quant_LTP_gains.c b/code/opus-1.1.4/silk/quant_LTP_gains.c index 513a8c44..d6b8eff8 100644 --- a/code/opus-1.1.4/silk/quant_LTP_gains.c +++ b/code/opus-1.1.4/silk/quant_LTP_gains.c @@ -33,14 +33,15 @@ POSSIBILITY OF SUCH DAMAGE. #include "tuning_parameters.h" void silk_quant_LTP_gains( - opus_int16 B_Q14[ MAX_NB_SUBFR * LTP_ORDER ], /* I/O (un)quantized LTP gains */ + opus_int16 B_Q14[ MAX_NB_SUBFR * LTP_ORDER ], /* O Quantized LTP gains */ opus_int8 cbk_index[ MAX_NB_SUBFR ], /* O Codebook Index */ opus_int8 *periodicity_index, /* O Periodicity Index */ opus_int32 *sum_log_gain_Q7, /* I/O Cumulative max prediction gain */ - const opus_int32 W_Q18[ MAX_NB_SUBFR*LTP_ORDER*LTP_ORDER ], /* I Error Weights in Q18 */ - opus_int mu_Q9, /* I Mu value (R/D tradeoff) */ - opus_int lowComplexity, /* I Flag for low complexity */ - const opus_int nb_subfr, /* I number of subframes */ + opus_int *pred_gain_dB_Q7, /* O LTP prediction gain */ + const opus_int32 XX_Q17[ MAX_NB_SUBFR*LTP_ORDER*LTP_ORDER ], /* I Correlation matrix in Q18 */ + const opus_int32 xX_Q17[ MAX_NB_SUBFR*LTP_ORDER ], /* I Correlation vector in Q18 */ + const opus_int subfr_len, /* I Number of samples per subframe */ + const opus_int nb_subfr, /* I Number of subframes */ int arch /* I Run-time architecture */ ) { @@ -49,16 +50,16 @@ void silk_quant_LTP_gains( const opus_uint8 *cl_ptr_Q5; const opus_int8 *cbk_ptr_Q7; const opus_uint8 *cbk_gain_ptr_Q7; - const opus_int16 *b_Q14_ptr; - const opus_int32 *W_Q18_ptr; - opus_int32 rate_dist_Q14_subfr, rate_dist_Q14, min_rate_dist_Q14; - opus_int32 sum_log_gain_tmp_Q7, best_sum_log_gain_Q7, max_gain_Q7, gain_Q7; + const opus_int32 *XX_Q17_ptr, *xX_Q17_ptr; + opus_int32 res_nrg_Q15_subfr, res_nrg_Q15, rate_dist_Q7_subfr, rate_dist_Q7, min_rate_dist_Q7; + opus_int32 sum_log_gain_tmp_Q7, best_sum_log_gain_Q7, max_gain_Q7; + opus_int gain_Q7; /***************************************************/ /* iterate over different codebooks with different */ /* rates/distortions, and choose best */ /***************************************************/ - min_rate_dist_Q14 = silk_int32_MAX; + min_rate_dist_Q7 = silk_int32_MAX; best_sum_log_gain_Q7 = 0; for( k = 0; k < 3; k++ ) { /* Safety margin for pitch gain control, to take into account factors @@ -70,53 +71,47 @@ void silk_quant_LTP_gains( cbk_gain_ptr_Q7 = silk_LTP_vq_gain_ptrs_Q7[ k ]; cbk_size = silk_LTP_vq_sizes[ k ]; - /* Set up pointer to first subframe */ - W_Q18_ptr = W_Q18; - b_Q14_ptr = B_Q14; + /* Set up pointers to first subframe */ + XX_Q17_ptr = XX_Q17; + xX_Q17_ptr = xX_Q17; - rate_dist_Q14 = 0; + res_nrg_Q15 = 0; + rate_dist_Q7 = 0; sum_log_gain_tmp_Q7 = *sum_log_gain_Q7; for( j = 0; j < nb_subfr; j++ ) { max_gain_Q7 = silk_log2lin( ( SILK_FIX_CONST( MAX_SUM_LOG_GAIN_DB / 6.0, 7 ) - sum_log_gain_tmp_Q7 ) + SILK_FIX_CONST( 7, 7 ) ) - gain_safety; - silk_VQ_WMat_EC( &temp_idx[ j ], /* O index of best codebook vector */ - &rate_dist_Q14_subfr, /* O best weighted quantization error + mu * rate */ + &res_nrg_Q15_subfr, /* O residual energy */ + &rate_dist_Q7_subfr, /* O best weighted quantization error + mu * rate */ &gain_Q7, /* O sum of absolute LTP coefficients */ - b_Q14_ptr, /* I input vector to be quantized */ - W_Q18_ptr, /* I weighting matrix */ + XX_Q17_ptr, /* I correlation matrix */ + xX_Q17_ptr, /* I correlation vector */ cbk_ptr_Q7, /* I codebook */ cbk_gain_ptr_Q7, /* I codebook effective gains */ cl_ptr_Q5, /* I code length for each codebook vector */ - mu_Q9, /* I tradeoff between weighted error and rate */ + subfr_len, /* I number of samples per subframe */ max_gain_Q7, /* I maximum sum of absolute LTP coefficients */ cbk_size, /* I number of vectors in codebook */ arch /* I Run-time architecture */ ); - rate_dist_Q14 = silk_ADD_POS_SAT32( rate_dist_Q14, rate_dist_Q14_subfr ); + res_nrg_Q15 = silk_ADD_POS_SAT32( res_nrg_Q15, res_nrg_Q15_subfr ); + rate_dist_Q7 = silk_ADD_POS_SAT32( rate_dist_Q7, rate_dist_Q7_subfr ); sum_log_gain_tmp_Q7 = silk_max(0, sum_log_gain_tmp_Q7 + silk_lin2log( gain_safety + gain_Q7 ) - SILK_FIX_CONST( 7, 7 )); - b_Q14_ptr += LTP_ORDER; - W_Q18_ptr += LTP_ORDER * LTP_ORDER; + XX_Q17_ptr += LTP_ORDER * LTP_ORDER; + xX_Q17_ptr += LTP_ORDER; } - /* Avoid never finding a codebook */ - rate_dist_Q14 = silk_min( silk_int32_MAX - 1, rate_dist_Q14 ); - - if( rate_dist_Q14 < min_rate_dist_Q14 ) { - min_rate_dist_Q14 = rate_dist_Q14; + if( rate_dist_Q7 <= min_rate_dist_Q7 ) { + min_rate_dist_Q7 = rate_dist_Q7; *periodicity_index = (opus_int8)k; silk_memcpy( cbk_index, temp_idx, nb_subfr * sizeof( opus_int8 ) ); best_sum_log_gain_Q7 = sum_log_gain_tmp_Q7; } - - /* Break early in low-complexity mode if rate distortion is below threshold */ - if( lowComplexity && ( rate_dist_Q14 < silk_LTP_gain_middle_avg_RD_Q14 ) ) { - break; - } } cbk_ptr_Q7 = silk_LTP_vq_ptrs_Q7[ *periodicity_index ]; @@ -125,5 +120,13 @@ void silk_quant_LTP_gains( B_Q14[ j * LTP_ORDER + k ] = silk_LSHIFT( cbk_ptr_Q7[ cbk_index[ j ] * LTP_ORDER + k ], 7 ); } } + + if( nb_subfr == 2 ) { + res_nrg_Q15 = silk_RSHIFT32( res_nrg_Q15, 1 ); + } else { + res_nrg_Q15 = silk_RSHIFT32( res_nrg_Q15, 2 ); + } + *sum_log_gain_Q7 = best_sum_log_gain_Q7; + *pred_gain_dB_Q7 = (opus_int)silk_SMULBB( -3, silk_lin2log( res_nrg_Q15 ) - ( 15 << 7 ) ); } diff --git a/code/opus-1.1.4/silk/structs.h b/code/opus-1.1.4/silk/structs.h index 827829dc..4ff590b0 100644 --- a/code/opus-1.1.4/silk/structs.h +++ b/code/opus-1.1.4/silk/structs.h @@ -48,6 +48,7 @@ typedef struct { opus_int32 sLPC_Q14[ MAX_SUB_FRAME_LENGTH + NSQ_LPC_BUF_LENGTH ]; opus_int32 sAR2_Q14[ MAX_SHAPE_LPC_ORDER ]; opus_int32 sLF_AR_shp_Q14; + opus_int32 sDiff_shp_Q14; opus_int lagPrev; opus_int sLTP_buf_idx; opus_int sLTP_shp_buf_idx; @@ -86,6 +87,7 @@ typedef struct { const opus_int16 quantStepSize_Q16; const opus_int16 invQuantStepSize_Q6; const opus_uint8 *CB1_NLSF_Q8; + const opus_int16 *CB1_Wght_Q9; const opus_uint8 *CB1_iCDF; const opus_uint8 *pred_Q8; const opus_uint8 *ec_sel; @@ -169,8 +171,6 @@ typedef struct { opus_int pitchEstimationComplexity; /* Complexity level for pitch estimator */ opus_int pitchEstimationLPCOrder; /* Whitening filter order for pitch estimator */ opus_int32 pitchEstimationThreshold_Q16; /* Threshold for pitch estimator */ - opus_int LTPQuantLowComplexity; /* Flag for low complexity LTP quantization */ - opus_int mu_LTP_Q9; /* Rate-distortion tradeoff in LTP quantization */ opus_int32 sum_log_gain_Q7; /* Cumulative max prediction gain */ opus_int NLSF_MSVQ_Survivors; /* Number of survivors in NLSF MSVQ */ opus_int first_frame_after_reset; /* Flag for deactivating NLSF interpolation, pitch prediction */ @@ -301,6 +301,7 @@ typedef struct { /* Stuff used for PLC */ opus_int lossCnt; opus_int prevSignalType; + int arch; silk_PLC_struct sPLC; diff --git a/code/opus-1.1.4/silk/sum_sqr_shift.c b/code/opus-1.1.4/silk/sum_sqr_shift.c index 129df191..4fd0c3d7 100644 --- a/code/opus-1.1.4/silk/sum_sqr_shift.c +++ b/code/opus-1.1.4/silk/sum_sqr_shift.c @@ -41,43 +41,40 @@ void silk_sum_sqr_shift( ) { opus_int i, shft; - opus_int32 nrg_tmp, nrg; + opus_uint32 nrg_tmp; + opus_int32 nrg; - nrg = 0; - shft = 0; - len--; - for( i = 0; i < len; i += 2 ) { - nrg = silk_SMLABB_ovflw( nrg, x[ i ], x[ i ] ); - nrg = silk_SMLABB_ovflw( nrg, x[ i + 1 ], x[ i + 1 ] ); - if( nrg < 0 ) { - /* Scale down */ - nrg = (opus_int32)silk_RSHIFT_uint( (opus_uint32)nrg, 2 ); - shft = 2; - i+=2; - break; - } - } - for( ; i < len; i += 2 ) { + /* Do a first run with the maximum shift we could have. */ + shft = 31-silk_CLZ32(len); + /* Let's be conservative with rounding and start with nrg=len. */ + nrg = len; + for( i = 0; i < len - 1; i += 2 ) { nrg_tmp = silk_SMULBB( x[ i ], x[ i ] ); nrg_tmp = silk_SMLABB_ovflw( nrg_tmp, x[ i + 1 ], x[ i + 1 ] ); - nrg = (opus_int32)silk_ADD_RSHIFT_uint( nrg, (opus_uint32)nrg_tmp, shft ); - if( nrg < 0 ) { - /* Scale down */ - nrg = (opus_int32)silk_RSHIFT_uint( (opus_uint32)nrg, 2 ); - shft += 2; - } + nrg = (opus_int32)silk_ADD_RSHIFT_uint( nrg, nrg_tmp, shft ); } - if( i == len ) { + if( i < len ) { + /* One sample left to process */ + nrg_tmp = silk_SMULBB( x[ i ], x[ i ] ); + nrg = (opus_int32)silk_ADD_RSHIFT_uint( nrg, nrg_tmp, shft ); + } + silk_assert( nrg >= 0 ); + /* Make sure the result will fit in a 32-bit signed integer with two bits + of headroom. */ + shft = silk_max_32(0, shft+3 - silk_CLZ32(nrg)); + nrg = 0; + for( i = 0 ; i < len - 1; i += 2 ) { + nrg_tmp = silk_SMULBB( x[ i ], x[ i ] ); + nrg_tmp = silk_SMLABB_ovflw( nrg_tmp, x[ i + 1 ], x[ i + 1 ] ); + nrg = (opus_int32)silk_ADD_RSHIFT_uint( nrg, nrg_tmp, shft ); + } + if( i < len ) { /* One sample left to process */ nrg_tmp = silk_SMULBB( x[ i ], x[ i ] ); nrg = (opus_int32)silk_ADD_RSHIFT_uint( nrg, nrg_tmp, shft ); } - /* Make sure to have at least one extra leading zero (two leading zeros in total) */ - if( nrg & 0xC0000000 ) { - nrg = silk_RSHIFT_uint( (opus_uint32)nrg, 2 ); - shft += 2; - } + silk_assert( nrg >= 0 ); /* Output arguments */ *shift = shft; diff --git a/code/opus-1.1.4/silk/tables.h b/code/opus-1.1.4/silk/tables.h index 7fea6fda..8b0380ee 100644 --- a/code/opus-1.1.4/silk/tables.h +++ b/code/opus-1.1.4/silk/tables.h @@ -76,10 +76,8 @@ extern const opus_uint8 silk_NLSF_EXT_iCDF[ 7 ]; extern const opus_uint8 silk_LTP_per_index_iCDF[ 3 ]; /* 3 */ extern const opus_uint8 * const silk_LTP_gain_iCDF_ptrs[ NB_LTP_CBKS ]; /* 3 */ extern const opus_uint8 * const silk_LTP_gain_BITS_Q5_ptrs[ NB_LTP_CBKS ]; /* 3 */ -extern const opus_int16 silk_LTP_gain_middle_avg_RD_Q14; extern const opus_int8 * const silk_LTP_vq_ptrs_Q7[ NB_LTP_CBKS ]; /* 168 */ extern const opus_uint8 * const silk_LTP_vq_gain_ptrs_Q7[NB_LTP_CBKS]; - extern const opus_int8 silk_LTP_vq_sizes[ NB_LTP_CBKS ]; /* 3 */ extern const opus_uint8 silk_LTPscale_iCDF[ 3 ]; /* 4 */ diff --git a/code/opus-1.1.4/silk/tables_LTP.c b/code/opus-1.1.4/silk/tables_LTP.c index 0e6a0254..5e12c864 100644 --- a/code/opus-1.1.4/silk/tables_LTP.c +++ b/code/opus-1.1.4/silk/tables_LTP.c @@ -51,8 +51,6 @@ static const opus_uint8 silk_LTP_gain_iCDF_2[32] = { 24, 20, 16, 12, 9, 5, 2, 0 }; -const opus_int16 silk_LTP_gain_middle_avg_RD_Q14 = 12304; - static const opus_uint8 silk_LTP_gain_BITS_Q5_0[8] = { 15, 131, 138, 138, 155, 155, 173, 173 }; diff --git a/code/opus-1.1.4/silk/tables_NLSF_CB_NB_MB.c b/code/opus-1.1.4/silk/tables_NLSF_CB_NB_MB.c index 8c59d207..195d5b95 100644 --- a/code/opus-1.1.4/silk/tables_NLSF_CB_NB_MB.c +++ b/code/opus-1.1.4/silk/tables_NLSF_CB_NB_MB.c @@ -74,6 +74,41 @@ static const opus_uint8 silk_NLSF_CB1_NB_MB_Q8[ 320 ] = { 64, 84, 104, 118, 156, 177, 201, 230 }; +static const opus_int16 silk_NLSF_CB1_Wght_Q9[ 320 ] = { + 2897, 2314, 2314, 2314, 2287, 2287, 2314, 2300, 2327, 2287, + 2888, 2580, 2394, 2367, 2314, 2274, 2274, 2274, 2274, 2194, + 2487, 2340, 2340, 2314, 2314, 2314, 2340, 2340, 2367, 2354, + 3216, 2766, 2340, 2340, 2314, 2274, 2221, 2207, 2261, 2194, + 2460, 2474, 2367, 2394, 2394, 2394, 2394, 2367, 2407, 2314, + 3479, 3056, 2127, 2207, 2274, 2274, 2274, 2287, 2314, 2261, + 3282, 3141, 2580, 2394, 2247, 2221, 2207, 2194, 2194, 2114, + 4096, 3845, 2221, 2620, 2620, 2407, 2314, 2394, 2367, 2074, + 3178, 3244, 2367, 2221, 2553, 2434, 2340, 2314, 2167, 2221, + 3338, 3488, 2726, 2194, 2261, 2460, 2354, 2367, 2207, 2101, + 2354, 2420, 2327, 2367, 2394, 2420, 2420, 2420, 2460, 2367, + 3779, 3629, 2434, 2527, 2367, 2274, 2274, 2300, 2207, 2048, + 3254, 3225, 2713, 2846, 2447, 2327, 2300, 2300, 2274, 2127, + 3263, 3300, 2753, 2806, 2447, 2261, 2261, 2247, 2127, 2101, + 2873, 2981, 2633, 2367, 2407, 2354, 2194, 2247, 2247, 2114, + 3225, 3197, 2633, 2580, 2274, 2181, 2247, 2221, 2221, 2141, + 3178, 3310, 2740, 2407, 2274, 2274, 2274, 2287, 2194, 2114, + 3141, 3272, 2460, 2061, 2287, 2500, 2367, 2487, 2434, 2181, + 3507, 3282, 2314, 2700, 2647, 2474, 2367, 2394, 2340, 2127, + 3423, 3535, 3038, 3056, 2300, 1950, 2221, 2274, 2274, 2274, + 3404, 3366, 2087, 2687, 2873, 2354, 2420, 2274, 2474, 2540, + 3760, 3488, 1950, 2660, 2897, 2527, 2394, 2367, 2460, 2261, + 3028, 3272, 2740, 2888, 2740, 2154, 2127, 2287, 2234, 2247, + 3695, 3657, 2025, 1969, 2660, 2700, 2580, 2500, 2327, 2367, + 3207, 3413, 2354, 2074, 2888, 2888, 2340, 2487, 2247, 2167, + 3338, 3366, 2846, 2780, 2327, 2154, 2274, 2287, 2114, 2061, + 2327, 2300, 2181, 2167, 2181, 2367, 2633, 2700, 2700, 2553, + 2407, 2434, 2221, 2261, 2221, 2221, 2340, 2420, 2607, 2700, + 3038, 3244, 2806, 2888, 2474, 2074, 2300, 2314, 2354, 2380, + 2221, 2154, 2127, 2287, 2500, 2793, 2793, 2620, 2580, 2367, + 3676, 3713, 2234, 1838, 2181, 2753, 2726, 2673, 2513, 2207, + 2793, 3160, 2726, 2553, 2846, 2513, 2181, 2394, 2221, 2181 +}; + static const opus_uint8 silk_NLSF_CB1_iCDF_NB_MB[ 64 ] = { 212, 178, 148, 129, 108, 96, 85, 82, 79, 77, 61, 59, 57, 56, 51, 49, @@ -150,6 +185,7 @@ const silk_NLSF_CB_struct silk_NLSF_CB_NB_MB = SILK_FIX_CONST( 0.18, 16 ), SILK_FIX_CONST( 1.0 / 0.18, 6 ), silk_NLSF_CB1_NB_MB_Q8, + silk_NLSF_CB1_Wght_Q9, silk_NLSF_CB1_iCDF_NB_MB, silk_NLSF_PRED_NB_MB_Q8, silk_NLSF_CB2_SELECT_NB_MB, diff --git a/code/opus-1.1.4/silk/tables_NLSF_CB_WB.c b/code/opus-1.1.4/silk/tables_NLSF_CB_WB.c index 50af87eb..5cc9f57b 100644 --- a/code/opus-1.1.4/silk/tables_NLSF_CB_WB.c +++ b/code/opus-1.1.4/silk/tables_NLSF_CB_WB.c @@ -98,6 +98,41 @@ static const opus_uint8 silk_NLSF_CB1_WB_Q8[ 512 ] = { 110, 119, 129, 141, 175, 198, 218, 237 }; +static const opus_int16 silk_NLSF_CB1_WB_Wght_Q9[ 512 ] = { + 3657, 2925, 2925, 2925, 2925, 2925, 2925, 2925, 2925, 2925, 2925, 2925, 2963, 2963, 2925, 2846, + 3216, 3085, 2972, 3056, 3056, 3010, 3010, 3010, 2963, 2963, 3010, 2972, 2888, 2846, 2846, 2726, + 3920, 4014, 2981, 3207, 3207, 2934, 3056, 2846, 3122, 3244, 2925, 2846, 2620, 2553, 2780, 2925, + 3516, 3197, 3010, 3103, 3019, 2888, 2925, 2925, 2925, 2925, 2888, 2888, 2888, 2888, 2888, 2753, + 5054, 5054, 2934, 3573, 3385, 3056, 3085, 2793, 3160, 3160, 2972, 2846, 2513, 2540, 2753, 2888, + 4428, 4149, 2700, 2753, 2972, 3010, 2925, 2846, 2981, 3019, 2925, 2925, 2925, 2925, 2888, 2726, + 3620, 3019, 2972, 3056, 3056, 2873, 2806, 3056, 3216, 3047, 2981, 3291, 3291, 2981, 3310, 2991, + 5227, 5014, 2540, 3338, 3526, 3385, 3197, 3094, 3376, 2981, 2700, 2647, 2687, 2793, 2846, 2673, + 5081, 5174, 4615, 4428, 2460, 2897, 3047, 3207, 3169, 2687, 2740, 2888, 2846, 2793, 2846, 2700, + 3122, 2888, 2963, 2925, 2925, 2925, 2925, 2963, 2963, 2963, 2963, 2925, 2925, 2963, 2963, 2963, + 4202, 3207, 2981, 3103, 3010, 2888, 2888, 2925, 2972, 2873, 2916, 3019, 2972, 3010, 3197, 2873, + 3760, 3760, 3244, 3103, 2981, 2888, 2925, 2888, 2972, 2934, 2793, 2793, 2846, 2888, 2888, 2660, + 3854, 4014, 3207, 3122, 3244, 2934, 3047, 2963, 2963, 3085, 2846, 2793, 2793, 2793, 2793, 2580, + 3845, 4080, 3357, 3516, 3094, 2740, 3010, 2934, 3122, 3085, 2846, 2846, 2647, 2647, 2846, 2806, + 5147, 4894, 3225, 3845, 3441, 3169, 2897, 3413, 3451, 2700, 2580, 2673, 2740, 2846, 2806, 2753, + 4109, 3789, 3291, 3160, 2925, 2888, 2888, 2925, 2793, 2740, 2793, 2740, 2793, 2846, 2888, 2806, + 5081, 5054, 3047, 3545, 3244, 3056, 3085, 2944, 3103, 2897, 2740, 2740, 2740, 2846, 2793, 2620, + 4309, 4309, 2860, 2527, 3207, 3376, 3376, 3075, 3075, 3376, 3056, 2846, 2647, 2580, 2726, 2753, + 3056, 2916, 2806, 2888, 2740, 2687, 2897, 3103, 3150, 3150, 3216, 3169, 3056, 3010, 2963, 2846, + 4375, 3882, 2925, 2888, 2846, 2888, 2846, 2846, 2888, 2888, 2888, 2846, 2888, 2925, 2888, 2846, + 2981, 2916, 2916, 2981, 2981, 3056, 3122, 3216, 3150, 3056, 3010, 2972, 2972, 2972, 2925, 2740, + 4229, 4149, 3310, 3347, 2925, 2963, 2888, 2981, 2981, 2846, 2793, 2740, 2846, 2846, 2846, 2793, + 4080, 4014, 3103, 3010, 2925, 2925, 2925, 2888, 2925, 2925, 2846, 2846, 2846, 2793, 2888, 2780, + 4615, 4575, 3169, 3441, 3207, 2981, 2897, 3038, 3122, 2740, 2687, 2687, 2687, 2740, 2793, 2700, + 4149, 4269, 3789, 3657, 2726, 2780, 2888, 2888, 3010, 2972, 2925, 2846, 2687, 2687, 2793, 2888, + 4215, 3554, 2753, 2846, 2846, 2888, 2888, 2888, 2925, 2925, 2888, 2925, 2925, 2925, 2963, 2888, + 5174, 4921, 2261, 3432, 3789, 3479, 3347, 2846, 3310, 3479, 3150, 2897, 2460, 2487, 2753, 2925, + 3451, 3685, 3122, 3197, 3357, 3047, 3207, 3207, 2981, 3216, 3085, 2925, 2925, 2687, 2540, 2434, + 2981, 3010, 2793, 2793, 2740, 2793, 2846, 2972, 3056, 3103, 3150, 3150, 3150, 3103, 3010, 3010, + 2944, 2873, 2687, 2726, 2780, 3010, 3432, 3545, 3357, 3244, 3056, 3010, 2963, 2925, 2888, 2846, + 3019, 2944, 2897, 3010, 3010, 2972, 3019, 3103, 3056, 3056, 3010, 2888, 2846, 2925, 2925, 2888, + 3920, 3967, 3010, 3197, 3357, 3216, 3291, 3291, 3479, 3704, 3441, 2726, 2181, 2460, 2580, 2607 +}; + static const opus_uint8 silk_NLSF_CB1_iCDF_WB[ 64 ] = { 225, 204, 201, 184, 183, 175, 158, 154, 153, 135, 119, 115, 113, 110, 109, 99, @@ -188,6 +223,7 @@ const silk_NLSF_CB_struct silk_NLSF_CB_WB = SILK_FIX_CONST( 0.15, 16 ), SILK_FIX_CONST( 1.0 / 0.15, 6 ), silk_NLSF_CB1_WB_Q8, + silk_NLSF_CB1_WB_Wght_Q9, silk_NLSF_CB1_iCDF_WB, silk_NLSF_PRED_WB_Q8, silk_NLSF_CB2_SELECT_WB, diff --git a/code/opus-1.1.4/silk/tuning_parameters.h b/code/opus-1.1.4/silk/tuning_parameters.h index 5b8f4042..d70275fd 100644 --- a/code/opus-1.1.4/silk/tuning_parameters.h +++ b/code/opus-1.1.4/silk/tuning_parameters.h @@ -53,19 +53,12 @@ extern "C" /* LPC analysis regularization */ #define FIND_LPC_COND_FAC 1e-5f -/* LTP analysis defines */ -#define FIND_LTP_COND_FAC 1e-5f -#define LTP_DAMPING 0.05f -#define LTP_SMOOTHING 0.1f - -/* LTP quantization settings */ -#define MU_LTP_QUANT_NB 0.03f -#define MU_LTP_QUANT_MB 0.025f -#define MU_LTP_QUANT_WB 0.02f - /* Max cumulative LTP gain */ #define MAX_SUM_LOG_GAIN_DB 250.0f +/* LTP analysis defines */ +#define LTP_CORR_INV_MAX 0.03f + /***********************/ /* High pass filtering */ /***********************/ @@ -103,25 +96,16 @@ extern "C" #define SPARSE_SNR_INCR_dB 2.0f /* threshold for sparseness measure above which to use lower quantization offset during unvoiced */ -#define SPARSENESS_THRESHOLD_QNT_OFFSET 0.75f +#define ENERGY_VARIATION_THRESHOLD_QNT_OFFSET 0.6f /* warping control */ #define WARPING_MULTIPLIER 0.015f /* fraction added to first autocorrelation value */ -#define SHAPE_WHITE_NOISE_FRACTION 5e-5f +#define SHAPE_WHITE_NOISE_FRACTION 3e-5f /* noise shaping filter chirp factor */ -#define BANDWIDTH_EXPANSION 0.95f - -/* difference between chirp factors for analysis and synthesis noise shaping filters at low bitrates */ -#define LOW_RATE_BANDWIDTH_EXPANSION_DELTA 0.01f - -/* extra harmonic boosting (signal shaping) at low bitrates */ -#define LOW_RATE_HARMONIC_BOOST 0.1f - -/* extra harmonic boosting (signal shaping) for noisy input signals */ -#define LOW_INPUT_QUALITY_HARMONIC_BOOST 0.1f +#define BANDWIDTH_EXPANSION 0.94f /* harmonic noise shaping */ #define HARMONIC_SHAPING 0.3f diff --git a/code/opus-1.1.4/silk/x86/NSQ_del_dec_sse.c b/code/opus-1.1.4/silk/x86/NSQ_del_dec_sse.c index 21d4a8bc..c5212bee 100644 --- a/code/opus-1.1.4/silk/x86/NSQ_del_dec_sse.c +++ b/code/opus-1.1.4/silk/x86/NSQ_del_dec_sse.c @@ -107,12 +107,12 @@ static OPUS_INLINE void silk_noise_shape_quantizer_del_dec_sse4_1( opus_int predictLPCOrder, /* I Prediction filter order */ opus_int warping_Q16, /* I */ opus_int nStatesDelayedDecision, /* I Number of states in decision tree */ - opus_int *smpl_buf_idx, /* I Index to newest samples in buffers */ + opus_int *smpl_buf_idx, /* I/O Index to newest samples in buffers */ opus_int decisionDelay /* I */ ); void silk_NSQ_del_dec_sse4_1( - const silk_encoder_state *psEncC, /* I/O Encoder State */ + const silk_encoder_state *psEncC, /* I Encoder State */ silk_nsq_state *NSQ, /* I/O NSQ state */ SideInfoIndices *psIndices, /* I/O Quantization Indices */ const opus_int32 x_Q3[], /* I Prefiltered input signal */ @@ -234,7 +234,8 @@ void silk_NSQ_del_dec_sse4_1( psDD = &psDelDec[ Winner_ind ]; last_smple_idx = smpl_buf_idx + decisionDelay; for( i = 0; i < decisionDelay; i++ ) { - last_smple_idx = ( last_smple_idx - 1 ) & DECISION_DELAY_MASK; + last_smple_idx = ( last_smple_idx - 1 ) % DECISION_DELAY; + if( last_smple_idx < 0 ) last_smple_idx += DECISION_DELAY; pulses[ i - decisionDelay ] = (opus_int8)silk_RSHIFT_ROUND( psDD->Q_Q10[ last_smple_idx ], 10 ); pxq[ i - decisionDelay ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( psDD->Xq_Q14[ last_smple_idx ], Gains_Q16[ 1 ] ), 14 ) ); @@ -285,7 +286,8 @@ void silk_NSQ_del_dec_sse4_1( last_smple_idx = smpl_buf_idx + decisionDelay; Gain_Q10 = silk_RSHIFT32( Gains_Q16[ psEncC->nb_subfr - 1 ], 6 ); for( i = 0; i < decisionDelay; i++ ) { - last_smple_idx = ( last_smple_idx - 1 ) & DECISION_DELAY_MASK; + last_smple_idx = ( last_smple_idx - 1 ) % DECISION_DELAY; + if( last_smple_idx < 0 ) last_smple_idx += DECISION_DELAY; pulses[ i - decisionDelay ] = (opus_int8)silk_RSHIFT_ROUND( psDD->Q_Q10[ last_smple_idx ], 10 ); pxq[ i - decisionDelay ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( psDD->Xq_Q14[ last_smple_idx ], Gain_Q10 ), 8 ) ); @@ -299,7 +301,6 @@ void silk_NSQ_del_dec_sse4_1( NSQ->lagPrev = pitchL[ psEncC->nb_subfr - 1 ]; /* Save quantized speech signal */ - /* DEBUG_STORE_DATA( enc.pcm, &NSQ->xq[psEncC->ltp_mem_length], psEncC->frame_length * sizeof( opus_int16 ) ) */ silk_memmove( NSQ->xq, &NSQ->xq[ psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int16 ) ); silk_memmove( NSQ->sLTP_shp_Q14, &NSQ->sLTP_shp_Q14[ psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int32 ) ); RESTORE_STACK; @@ -333,7 +334,7 @@ static OPUS_INLINE void silk_noise_shape_quantizer_del_dec_sse4_1( opus_int predictLPCOrder, /* I Prediction filter order */ opus_int warping_Q16, /* I */ opus_int nStatesDelayedDecision, /* I Number of states in decision tree */ - opus_int *smpl_buf_idx, /* I Index to newest samples in buffers */ + opus_int *smpl_buf_idx, /* I/O Index to newest samples in buffers */ opus_int decisionDelay /* I */ ) { @@ -638,8 +639,9 @@ static OPUS_INLINE void silk_noise_shape_quantizer_del_dec_sse4_1( psSS[ 1 ].xq_Q14 = xq_Q14; } } - *smpl_buf_idx = ( *smpl_buf_idx - 1 ) & DECISION_DELAY_MASK; /* Index to newest samples */ - last_smple_idx = ( *smpl_buf_idx + decisionDelay ) & DECISION_DELAY_MASK; /* Index to decisionDelay old samples */ + *smpl_buf_idx = ( *smpl_buf_idx - 1 ) % DECISION_DELAY; + if( *smpl_buf_idx < 0 ) *smpl_buf_idx += DECISION_DELAY; + last_smple_idx = ( *smpl_buf_idx + decisionDelay ) % DECISION_DELAY; /* Find winner */ RDmin_Q10 = psSampleState[ 0 ][ 0 ].RD_Q10; diff --git a/code/opus-1.1.4/silk/x86/NSQ_sse.c b/code/opus-1.1.4/silk/x86/NSQ_sse.c index bb3c5f19..6a9e6e96 100644 --- a/code/opus-1.1.4/silk/x86/NSQ_sse.c +++ b/code/opus-1.1.4/silk/x86/NSQ_sse.c @@ -71,7 +71,7 @@ static OPUS_INLINE void silk_noise_shape_quantizer_10_16_sse4_1( ); void silk_NSQ_sse4_1( - const silk_encoder_state *psEncC, /* I/O Encoder State */ + const silk_encoder_state *psEncC, /* I Encoder State */ silk_nsq_state *NSQ, /* I/O NSQ state */ SideInfoIndices *psIndices, /* I/O Quantization Indices */ const opus_int32 x_Q3[], /* I Prefiltered input signal */ @@ -233,7 +233,6 @@ void silk_NSQ_sse4_1( NSQ->lagPrev = pitchL[ psEncC->nb_subfr - 1 ]; /* Save quantized speech and noise shaping signals */ - /* DEBUG_STORE_DATA( enc.pcm, &NSQ->xq[ psEncC->ltp_mem_length ], psEncC->frame_length * sizeof( opus_int16 ) ) */ silk_memmove( NSQ->xq, &NSQ->xq[ psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int16 ) ); silk_memmove( NSQ->sLTP_shp_Q14, &NSQ->sLTP_shp_Q14[ psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int32 ) ); RESTORE_STACK; diff --git a/code/opus-1.1.4/silk/x86/main_sse.h b/code/opus-1.1.4/silk/x86/main_sse.h index d8d61310..2f15d448 100644 --- a/code/opus-1.1.4/silk/x86/main_sse.h +++ b/code/opus-1.1.4/silk/x86/main_sse.h @@ -34,6 +34,7 @@ # if defined(OPUS_X86_MAY_HAVE_SSE4_1) +#if 0 /* FIXME: SSE disabled until silk_VQ_WMat_EC_sse4_1() gets updated. */ # define OVERRIDE_silk_VQ_WMat_EC void silk_VQ_WMat_EC_sse4_1( @@ -79,11 +80,13 @@ extern void (*const SILK_VQ_WMAT_EC_IMPL[OPUS_ARCHMASK + 1])( mu_Q9, max_gain_Q7, L)) #endif +#endif +#if 0 /* FIXME: SSE disabled until the NSQ code gets updated. */ # define OVERRIDE_silk_NSQ void silk_NSQ_sse4_1( - const silk_encoder_state *psEncC, /* I/O Encoder State */ + const silk_encoder_state *psEncC, /* I Encoder State */ silk_nsq_state *NSQ, /* I/O NSQ state */ SideInfoIndices *psIndices, /* I/O Quantization Indices */ const opus_int32 x_Q3[], /* I Prefiltered input signal */ @@ -110,7 +113,7 @@ void silk_NSQ_sse4_1( #else extern void (*const SILK_NSQ_IMPL[OPUS_ARCHMASK + 1])( - const silk_encoder_state *psEncC, /* I/O Encoder State */ + const silk_encoder_state *psEncC, /* I Encoder State */ silk_nsq_state *NSQ, /* I/O NSQ state */ SideInfoIndices *psIndices, /* I/O Quantization Indices */ const opus_int32 x_Q3[], /* I Prefiltered input signal */ @@ -137,7 +140,7 @@ extern void (*const SILK_NSQ_IMPL[OPUS_ARCHMASK + 1])( # define OVERRIDE_silk_NSQ_del_dec void silk_NSQ_del_dec_sse4_1( - const silk_encoder_state *psEncC, /* I/O Encoder State */ + const silk_encoder_state *psEncC, /* I Encoder State */ silk_nsq_state *NSQ, /* I/O NSQ state */ SideInfoIndices *psIndices, /* I/O Quantization Indices */ const opus_int32 x_Q3[], /* I Prefiltered input signal */ @@ -164,7 +167,7 @@ void silk_NSQ_del_dec_sse4_1( #else extern void (*const SILK_NSQ_DEL_DEC_IMPL[OPUS_ARCHMASK + 1])( - const silk_encoder_state *psEncC, /* I/O Encoder State */ + const silk_encoder_state *psEncC, /* I Encoder State */ silk_nsq_state *NSQ, /* I/O NSQ state */ SideInfoIndices *psIndices, /* I/O Quantization Indices */ const opus_int32 x_Q3[], /* I Prefiltered input signal */ @@ -187,6 +190,7 @@ extern void (*const SILK_NSQ_DEL_DEC_IMPL[OPUS_ARCHMASK + 1])( HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, LTP_scale_Q14)) #endif +#endif void silk_noise_shape_quantizer( silk_nsq_state *NSQ, /* I/O NSQ state */ @@ -238,39 +242,6 @@ extern opus_int (*const SILK_VAD_GETSA_Q8_IMPL[OPUS_ARCHMASK + 1])( silk_encoder_state *psEnC, const opus_int16 pIn[]); -# define OVERRIDE_silk_warped_LPC_analysis_filter_FIX - -#endif - -void silk_warped_LPC_analysis_filter_FIX_sse4_1( - opus_int32 state[], /* I/O State [order + 1] */ - opus_int32 res_Q2[], /* O Residual signal [length] */ - const opus_int16 coef_Q13[], /* I Coefficients [order] */ - const opus_int16 input[], /* I Input signal [length] */ - const opus_int16 lambda_Q16, /* I Warping factor */ - const opus_int length, /* I Length of input signal */ - const opus_int order /* I Filter order (even) */ -); - -#if defined(OPUS_X86_PRESUME_SSE4_1) -#define silk_warped_LPC_analysis_filter_FIX(state, res_Q2, coef_Q13, input, lambda_Q16, length, order, arch) \ - ((void)(arch),silk_warped_LPC_analysis_filter_FIX_c(state, res_Q2, coef_Q13, input, lambda_Q16, length, order)) - -#else - -extern void (*const SILK_WARPED_LPC_ANALYSIS_FILTER_FIX_IMPL[OPUS_ARCHMASK + 1])( - opus_int32 state[], /* I/O State [order + 1] */ - opus_int32 res_Q2[], /* O Residual signal [length] */ - const opus_int16 coef_Q13[], /* I Coefficients [order] */ - const opus_int16 input[], /* I Input signal [length] */ - const opus_int16 lambda_Q16, /* I Warping factor */ - const opus_int length, /* I Length of input signal */ - const opus_int order /* I Filter order (even) */ -); - -# define silk_warped_LPC_analysis_filter_FIX(state, res_Q2, coef_Q13, input, lambda_Q16, length, order, arch) \ - ((*SILK_WARPED_LPC_ANALYSIS_FILTER_FIX_IMPL[(arch) & OPUS_ARCHMASK])(state, res_Q2, coef_Q13, input, lambda_Q16, length, order)) - #endif # endif diff --git a/code/opus-1.1.4/silk/x86/x86_silk_map.c b/code/opus-1.1.4/silk/x86/x86_silk_map.c index 818841f2..32dcc3ca 100644 --- a/code/opus-1.1.4/silk/x86/x86_silk_map.c +++ b/code/opus-1.1.4/silk/x86/x86_silk_map.c @@ -66,8 +66,9 @@ opus_int (*const SILK_VAD_GETSA_Q8_IMPL[ OPUS_ARCHMASK + 1 ] )( MAY_HAVE_SSE4_1( silk_VAD_GetSA_Q8 ) /* avx */ }; +#if 0 /* FIXME: SSE disabled until the NSQ code gets updated. */ void (*const SILK_NSQ_IMPL[ OPUS_ARCHMASK + 1 ] )( - const silk_encoder_state *psEncC, /* I/O Encoder State */ + const silk_encoder_state *psEncC, /* I Encoder State */ silk_nsq_state *NSQ, /* I/O NSQ state */ SideInfoIndices *psIndices, /* I/O Quantization Indices */ const opus_int32 x_Q3[], /* I Prefiltered input signal */ @@ -89,7 +90,9 @@ void (*const SILK_NSQ_IMPL[ OPUS_ARCHMASK + 1 ] )( MAY_HAVE_SSE4_1( silk_NSQ ), /* sse4.1 */ MAY_HAVE_SSE4_1( silk_NSQ ) /* avx */ }; +#endif +#if 0 /* FIXME: SSE disabled until silk_VQ_WMat_EC_sse4_1() gets updated. */ void (*const SILK_VQ_WMAT_EC_IMPL[ OPUS_ARCHMASK + 1 ] )( opus_int8 *ind, /* O index of best codebook vector */ opus_int32 *rate_dist_Q14, /* O best weighted quant error + mu * rate */ @@ -109,9 +112,11 @@ void (*const SILK_VQ_WMAT_EC_IMPL[ OPUS_ARCHMASK + 1 ] )( MAY_HAVE_SSE4_1( silk_VQ_WMat_EC ), /* sse4.1 */ MAY_HAVE_SSE4_1( silk_VQ_WMat_EC ) /* avx */ }; +#endif +#if 0 /* FIXME: SSE disabled until the NSQ code gets updated. */ void (*const SILK_NSQ_DEL_DEC_IMPL[ OPUS_ARCHMASK + 1 ] )( - const silk_encoder_state *psEncC, /* I/O Encoder State */ + const silk_encoder_state *psEncC, /* I Encoder State */ silk_nsq_state *NSQ, /* I/O NSQ state */ SideInfoIndices *psIndices, /* I/O Quantization Indices */ const opus_int32 x_Q3[], /* I Prefiltered input signal */ @@ -133,25 +138,10 @@ void (*const SILK_NSQ_DEL_DEC_IMPL[ OPUS_ARCHMASK + 1 ] )( MAY_HAVE_SSE4_1( silk_NSQ_del_dec ), /* sse4.1 */ MAY_HAVE_SSE4_1( silk_NSQ_del_dec ) /* avx */ }; +#endif #if defined(FIXED_POINT) -void (*const SILK_WARPED_LPC_ANALYSIS_FILTER_FIX_IMPL[ OPUS_ARCHMASK + 1 ] )( - opus_int32 state[], /* I/O State [order + 1] */ - opus_int32 res_Q2[], /* O Residual signal [length] */ - const opus_int16 coef_Q13[], /* I Coefficients [order] */ - const opus_int16 input[], /* I Input signal [length] */ - const opus_int16 lambda_Q16, /* I Warping factor */ - const opus_int length, /* I Length of input signal */ - const opus_int order /* I Filter order (even) */ -) = { - silk_warped_LPC_analysis_filter_FIX_c, /* non-sse */ - silk_warped_LPC_analysis_filter_FIX_c, - silk_warped_LPC_analysis_filter_FIX_c, - MAY_HAVE_SSE4_1( silk_warped_LPC_analysis_filter_FIX ), /* sse4.1 */ - MAY_HAVE_SSE4_1( silk_warped_LPC_analysis_filter_FIX ) /* avx */ -}; - void (*const SILK_BURG_MODIFIED_IMPL[ OPUS_ARCHMASK + 1 ] )( opus_int32 *res_nrg, /* O Residual energy */ opus_int *res_nrg_Q, /* O Residual energy Q value */ diff --git a/code/opus-1.1.4/src/analysis.c b/code/opus-1.1.4/src/analysis.c index 663431a4..f4160e4b 100644 --- a/code/opus-1.1.4/src/analysis.c +++ b/code/opus-1.1.4/src/analysis.c @@ -29,20 +29,27 @@ #include "config.h" #endif +#define ANALYSIS_C + +#include + +#include "mathops.h" #include "kiss_fft.h" #include "celt.h" #include "modes.h" #include "arch.h" #include "quant_bands.h" -#include #include "analysis.h" #include "mlp.h" #include "stack_alloc.h" +#include "float_cast.h" #ifndef M_PI #define M_PI 3.141592653 #endif +#ifndef DISABLE_FLOAT_API + static const float dct_table[128] = { 0.250000f, 0.250000f, 0.250000f, 0.250000f, 0.250000f, 0.250000f, 0.250000f, 0.250000f, 0.250000f, 0.250000f, 0.250000f, 0.250000f, 0.250000f, 0.250000f, 0.250000f, 0.250000f, @@ -96,52 +103,118 @@ static const float analysis_window[240] = { }; static const int tbands[NB_TBANDS+1] = { - 2, 4, 6, 8, 10, 12, 14, 16, 20, 24, 28, 32, 40, 48, 56, 68, 80, 96, 120 + 4, 8, 12, 16, 20, 24, 28, 32, 40, 48, 56, 64, 80, 96, 112, 136, 160, 192, 240 }; -static const int extra_bands[NB_TOT_BANDS+1] = { - 1, 2, 4, 6, 8, 10, 12, 14, 16, 20, 24, 28, 32, 40, 48, 56, 68, 80, 96, 120, 160, 200 -}; - -/*static const float tweight[NB_TBANDS+1] = { - .3, .4, .5, .6, .7, .8, .9, 1., 1., 1., 1., 1., 1., 1., .8, .7, .6, .5 -};*/ - #define NB_TONAL_SKIP_BANDS 9 -#define cA 0.43157974f -#define cB 0.67848403f -#define cC 0.08595542f -#define cE ((float)M_PI/2) -static OPUS_INLINE float fast_atan2f(float y, float x) { - float x2, y2; - /* Should avoid underflow on the values we'll get */ - if (ABS16(x)+ABS16(y)<1e-9f) - { - x*=1e12f; - y*=1e12f; - } - x2 = x*x; - y2 = y*y; - if(x2> (2*SIG_SHIFT + 8); +#endif + return (opus_val32)hp_ener; } -void tonality_analysis_init(TonalityAnalysisState *tonal) +static opus_val32 downmix_and_resample(downmix_func downmix, const void *_x, opus_val32 *y, opus_val32 S[3], int subframe, int offset, int c1, int c2, int C, int Fs) +{ + VARDECL(opus_val32, tmp); + opus_val32 scale; + int j; + opus_val32 ret = 0; + SAVE_STACK; + + if (subframe==0) return 0; + if (Fs == 48000) + { + subframe *= 2; + offset *= 2; + } else if (Fs == 16000) { + subframe = subframe*2/3; + offset = offset*2/3; + } + ALLOC(tmp, subframe, opus_val32); + + downmix(_x, tmp, subframe, offset, c1, c2, C); +#ifdef FIXED_POINT + scale = (1<-1) + scale /= 2; + for (j=0;jarch = opus_select_arch(); + tonal->Fs = Fs; /* Clear remaining fields. */ tonality_analysis_reset(tonal); } @@ -151,6 +224,8 @@ void tonality_analysis_reset(TonalityAnalysisState *tonal) /* Clear non-reusable fields. */ char *start = (char*)&tonal->TONALITY_ANALYSIS_RESET_START; OPUS_CLEAR(start, sizeof(TonalityAnalysisState) - (start - (char*)tonal)); + tonal->music_confidence = .9f; + tonal->speech_confidence = .1f; } void tonality_get_info(TonalityAnalysisState *tonal, AnalysisInfo *info_out, int len) @@ -158,6 +233,9 @@ void tonality_get_info(TonalityAnalysisState *tonal, AnalysisInfo *info_out, int int pos; int curr_lookahead; float psum; + float tonality_max; + float tonality_avg; + int tonality_count; int i; pos = tonal->read_pos; @@ -165,7 +243,8 @@ void tonality_get_info(TonalityAnalysisState *tonal, AnalysisInfo *info_out, int if (curr_lookahead<0) curr_lookahead += DETECT_SIZE; - if (len > 480 && pos != tonal->write_pos) + /* On long frames, look at the second analysis window rather than the first. */ + if (len > tonal->Fs/50 && pos != tonal->write_pos) { pos++; if (pos==DETECT_SIZE) @@ -176,18 +255,32 @@ void tonality_get_info(TonalityAnalysisState *tonal, AnalysisInfo *info_out, int if (pos<0) pos = DETECT_SIZE-1; OPUS_COPY(info_out, &tonal->info[pos], 1); - tonal->read_subframe += len/120; - while (tonal->read_subframe>=4) + tonality_max = tonality_avg = info_out->tonality; + tonality_count = 1; + /* If possible, look ahead for a tone to compensate for the delay in the tone detector. */ + for (i=0;i<3;i++) { - tonal->read_subframe -= 4; + pos++; + if (pos==DETECT_SIZE) + pos = 0; + if (pos == tonal->write_pos) + break; + tonality_max = MAX32(tonality_max, tonal->info[pos].tonality); + tonality_avg += tonal->info[pos].tonality; + tonality_count++; + } + info_out->tonality = MAX32(tonality_avg/tonality_count, tonality_max-.2f); + tonal->read_subframe += len/(tonal->Fs/400); + while (tonal->read_subframe>=8) + { + tonal->read_subframe -= 8; tonal->read_pos++; } if (tonal->read_pos>=DETECT_SIZE) tonal->read_pos-=DETECT_SIZE; - /* Compensate for the delay in the features themselves. - FIXME: Need a better estimate the 10 I just made up */ - curr_lookahead = IMAX(curr_lookahead-10, 0); + /* The -1 is to compensate for the delay in the features themselves. */ + curr_lookahead = IMAX(curr_lookahead-1, 0); psum=0; /* Summing the probability of transition patterns that involve music at @@ -197,11 +290,28 @@ void tonality_get_info(TonalityAnalysisState *tonal, AnalysisInfo *info_out, int for (;ipspeech[i]; psum = psum*tonal->music_confidence + (1-psum)*tonal->speech_confidence; - /*printf("%f %f %f\n", psum, info_out->music_prob, info_out->tonality);*/ + /*printf("%f %f %f %f %f\n", psum, info_out->music_prob, info_out->vad_prob, info_out->activity_probability, info_out->tonality);*/ info_out->music_prob = psum; } +static const float std_feature_bias[9] = { + 5.684947f, 3.475288f, 1.770634f, 1.599784f, 3.773215f, + 2.163313f, 1.260756f, 1.116868f, 1.918795f +}; + +#define LEAKAGE_OFFSET 2.5f +#define LEAKAGE_SLOPE 2.f + +#ifdef FIXED_POINT +/* For fixed-point, the input is +/-2^15 shifted up by SIG_SHIFT, so we need to + compensate for that in the energy. */ +#define SCALE_COMPENS (1.f/((opus_int32)1<<(15+SIG_SHIFT))) +#define SCALE_ENER(e) ((SCALE_COMPENS*SCALE_COMPENS)*(e)) +#else +#define SCALE_ENER(e) (e) +#endif + static void tonality_analysis(TonalityAnalysisState *tonal, const CELTMode *celt_mode, const void *x, int len, int offset, int c1, int c2, int C, int lsb_depth, downmix_func downmix) { int i, b; @@ -235,19 +345,41 @@ static void tonality_analysis(TonalityAnalysisState *tonal, const CELTMode *celt float noise_floor; int remaining; AnalysisInfo *info; + float hp_ener; + float tonality2[240]; + float midE[8]; + float spec_variability=0; + float band_log2[NB_TBANDS+1]; + float leakage_from[NB_TBANDS+1]; + float leakage_to[NB_TBANDS+1]; SAVE_STACK; - tonal->last_transition++; - alpha = 1.f/IMIN(20, 1+tonal->count); - alphaE = 1.f/IMIN(50, 1+tonal->count); - alphaE2 = 1.f/IMIN(1000, 1+tonal->count); + alpha = 1.f/IMIN(10, 1+tonal->count); + alphaE = 1.f/IMIN(25, 1+tonal->count); + alphaE2 = 1.f/IMIN(500, 1+tonal->count); - if (tonal->count<4) - tonal->music_prob = .5; + if (tonal->Fs == 48000) + { + /* len and offset are now at 24 kHz. */ + len/= 2; + offset /= 2; + } else if (tonal->Fs == 16000) { + len = 3*len/2; + offset = 3*offset/2; + } + + if (tonal->count<4) { + if (tonal->application == OPUS_APPLICATION_VOIP) + tonal->music_prob = .1f; + else + tonal->music_prob = .625f; + } kfft = celt_mode->mdct.kfft[0]; if (tonal->count==0) tonal->mem_fill = 240; - downmix(x, &tonal->inmem[tonal->mem_fill], IMIN(len, ANALYSIS_BUF_SIZE-tonal->mem_fill), offset, c1, c2, C); + tonal->hp_ener_accum += (float)downmix_and_resample(downmix, x, + &tonal->inmem[tonal->mem_fill], tonal->downmix_state, + IMIN(len, ANALYSIS_BUF_SIZE-tonal->mem_fill), offset, c1, c2, C, tonal->Fs); if (tonal->mem_fill+len < ANALYSIS_BUF_SIZE) { tonal->mem_fill += len; @@ -255,6 +387,7 @@ static void tonality_analysis(TonalityAnalysisState *tonal, const CELTMode *celt RESTORE_STACK; return; } + hp_ener = tonal->hp_ener_accum; info = &tonal->info[tonal->write_pos++]; if (tonal->write_pos>=DETECT_SIZE) tonal->write_pos-=DETECT_SIZE; @@ -273,7 +406,9 @@ static void tonality_analysis(TonalityAnalysisState *tonal, const CELTMode *celt } OPUS_MOVE(tonal->inmem, tonal->inmem+ANALYSIS_BUF_SIZE-240, 240); remaining = len - (ANALYSIS_BUF_SIZE-tonal->mem_fill); - downmix(x, &tonal->inmem[240], remaining, offset+ANALYSIS_BUF_SIZE-tonal->mem_fill, c1, c2, C); + tonal->hp_ener_accum = (float)downmix_and_resample(downmix, x, + &tonal->inmem[240], tonal->downmix_state, remaining, + offset+ANALYSIS_BUF_SIZE-tonal->mem_fill, c1, c2, C, tonal->Fs); tonal->mem_fill = 240 + remaining; opus_fft(kfft, in, out, tonal->arch); #ifndef FIXED_POINT @@ -305,24 +440,31 @@ static void tonality_analysis(TonalityAnalysisState *tonal, const CELTMode *celt d_angle2 = angle2 - angle; d2_angle2 = d_angle2 - d_angle; - mod1 = d2_angle - (float)floor(.5+d2_angle); + mod1 = d2_angle - (float)float2int(d2_angle); noisiness[i] = ABS16(mod1); mod1 *= mod1; mod1 *= mod1; - mod2 = d2_angle2 - (float)floor(.5+d2_angle2); + mod2 = d2_angle2 - (float)float2int(d2_angle2); noisiness[i] += ABS16(mod2); mod2 *= mod2; mod2 *= mod2; - avg_mod = .25f*(d2A[i]+2.f*mod1+mod2); + avg_mod = .25f*(d2A[i]+mod1+2*mod2); + /* This introduces an extra delay of 2 frames in the detection. */ tonality[i] = 1.f/(1.f+40.f*16.f*pi4*avg_mod)-.015f; + /* No delay on this detection, but it's less reliable. */ + tonality2[i] = 1.f/(1.f+40.f*16.f*pi4*mod2)-.015f; A[i] = angle2; dA[i] = d_angle2; d2A[i] = mod2; } - + for (i=2;ilowE[b] = MIN32(logE[b], tonal->lowE[b]+.01f); - tonal->highE[b] = MAX32(logE[b], tonal->highE[b]-.1f); - if (tonal->highE[b] < tonal->lowE[b]+1.f) + band_log2[b+1] = .5f*1.442695f*(float)log(E+1e-10f); + tonal->logE[tonal->E_count][b] = logE[b]; + if (tonal->count==0) + tonal->highE[b] = tonal->lowE[b] = logE[b]; + if (tonal->highE[b] > tonal->lowE[b] + 7.5) { - tonal->highE[b]+=.5f; - tonal->lowE[b]-=.5f; + if (tonal->highE[b] - logE[b] > logE[b] - tonal->lowE[b]) + tonal->highE[b] -= .01f; + else + tonal->lowE[b] += .01f; } - relativeE += (logE[b]-tonal->lowE[b])/(1e-15f+tonal->highE[b]-tonal->lowE[b]); + if (logE[b] > tonal->highE[b]) + { + tonal->highE[b] = logE[b]; + tonal->lowE[b] = MAX32(tonal->highE[b]-15, tonal->lowE[b]); + } else if (logE[b] < tonal->lowE[b]) + { + tonal->lowE[b] = logE[b]; + tonal->highE[b] = MIN32(tonal->lowE[b]+15, tonal->highE[b]); + } + relativeE += (logE[b]-tonal->lowE[b])/(1e-15f + (tonal->highE[b]-tonal->lowE[b])); L1=L2=0; for (i=0;iprev_band_tonality[b] = band_tonality[b]; } + leakage_from[0] = band_log2[0]; + leakage_to[0] = band_log2[0] - LEAKAGE_OFFSET; + for (b=1;b=0;b--) + { + float leak_slope = LEAKAGE_SLOPE*(tbands[b+1]-tbands[b])/4; + leakage_from[b] = MIN16(leakage_from[b+1]+leak_slope, leakage_from[b]); + leakage_to[b] = MAX16(leakage_to[b+1]-leak_slope, leakage_to[b]); + } + celt_assert(NB_TBANDS+1 <= LEAK_BANDS); + for (b=0;bleak_boost[b] = IMIN(255, (int)floor(.5 + 64.f*boost)); + } + for (;bleak_boost[b] = 0; + + for (i=0;ilogE[i][k] - tonal->logE[j][k]; + dist += tmp*tmp; + } + if (j!=i) + mindist = MIN32(mindist, dist); + } + spec_variability += mindist; + } + spec_variability = (float)sqrt(spec_variability/NB_FRAMES/NB_TBANDS); bandwidth_mask = 0; bandwidth = 0; maxE = 0; noise_floor = 5.7e-4f/(1<<(IMAX(0,lsb_depth-8))); -#ifdef FIXED_POINT - noise_floor *= 1<<(15+SIG_SHIFT); -#endif noise_floor *= noise_floor; - for (b=0;bmeanE[b] = MAX32((1-alphaE2)*tonal->meanE[b], E); E = MAX32(E, tonal->meanE[b]); @@ -441,14 +656,36 @@ static void tonality_analysis(TonalityAnalysisState *tonal, const CELTMode *celt 2) less than 90 dB below the peak band (maximal masking possible considering both the ATH and the loudness-dependent slope of the spreading function) 3) above the PCM quantization noise floor + We use b+1 because the first CELT band isn't included in tbands[] */ if (E>.1*bandwidth_mask && E*1e9f > maxE && E > noise_floor*(band_end-band_start)) - bandwidth = b; + bandwidth = b+1; + } + /* Special case for the last two bands, for which we don't have spectrum but only + the energy above 12 kHz. */ + if (tonal->Fs == 48000) { + float ratio; + float E = hp_ener*(1.f/(240*240)); + ratio = tonal->prev_bandwidth==20 ? 0.03f : 0.07f; +#ifdef FIXED_POINT + /* silk_resampler_down2_hp() shifted right by an extra 8 bits. */ + E *= 256.f*(1.f/Q15ONE)*(1.f/Q15ONE); +#endif + maxE = MAX32(maxE, E); + tonal->meanE[b] = MAX32((1-alphaE2)*tonal->meanE[b], E); + E = MAX32(E, tonal->meanE[b]); + /* Use a simple follower with 13 dB/Bark slope for spreading function */ + bandwidth_mask = MAX32(.05f*bandwidth_mask, E); + if (E>ratio*bandwidth_mask && E*1e9f > maxE && E > noise_floor*160) + bandwidth = 20; + /* This detector is unreliable, so if the bandwidth is close to SWB, assume it's FB. */ + if (bandwidth >= 17) + bandwidth = 20; } if (tonal->count<=2) bandwidth = 20; frame_loudness = 20*(float)log10(frame_loudness); - tonal->Etracker = MAX32(tonal->Etracker-.03f, frame_loudness); + tonal->Etracker = MAX32(tonal->Etracker-.003f, frame_loudness); tonal->lowECount *= (1-alphaE); if (frame_loudness < tonal->Etracker-30) tonal->lowECount += alphaE; @@ -460,11 +697,18 @@ static void tonality_analysis(TonalityAnalysisState *tonal, const CELTMode *celt sum += dct_table[i*16+b]*logE[b]; BFCC[i] = sum; } + for (i=0;i<8;i++) + { + float sum=0; + for (b=0;b<16;b++) + sum += dct_table[i*16+b]*.5f*(tonal->highE[b]+tonal->lowE[b]); + midE[i] = sum; + } frame_stationarity /= NB_TBANDS; relativeE /= NB_TBANDS; if (tonal->count<10) - relativeE = .5; + relativeE = .5f; frame_noisiness /= NB_TBANDS; #if 1 info->activity = frame_noisiness + (1-frame_noisiness)*relativeE; @@ -479,7 +723,7 @@ static void tonality_analysis(TonalityAnalysisState *tonal, const CELTMode *celt info->tonality_slope = slope; tonal->E_count = (tonal->E_count+1)%NB_FRAMES; - tonal->count++; + tonal->count = IMIN(tonal->count+1, ANALYSIS_COUNT_MAX); info->tonality = frame_tonality; for (i=0;i<4;i++) @@ -498,6 +742,8 @@ static void tonality_analysis(TonalityAnalysisState *tonal, const CELTMode *celt for (i=0;i<9;i++) tonal->std[i] = (1-alpha)*tonal->std[i] + alpha*features[i]*features[i]; } + for (i=0;i<4;i++) + features[i] = BFCC[i]-midE[i]; for (i=0;i<8;i++) { @@ -507,24 +753,26 @@ static void tonality_analysis(TonalityAnalysisState *tonal, const CELTMode *celt tonal->mem[i] = BFCC[i]; } for (i=0;i<9;i++) - features[11+i] = (float)sqrt(tonal->std[i]); - features[20] = info->tonality; - features[21] = info->activity; - features[22] = frame_stationarity; - features[23] = info->tonality_slope; - features[24] = tonal->lowECount; + features[11+i] = (float)sqrt(tonal->std[i]) - std_feature_bias[i]; + features[18] = spec_variability - 0.78f; + features[20] = info->tonality - 0.154723f; + features[21] = info->activity - 0.724643f; + features[22] = frame_stationarity - 0.743717f; + features[23] = info->tonality_slope + 0.069216f; + features[24] = tonal->lowECount - 0.067930f; -#ifndef DISABLE_FLOAT_API mlp_process(&net, features, frame_probs); frame_probs[0] = .5f*(frame_probs[0]+1); /* Curve fitting between the MLP probability and the actual probability */ - frame_probs[0] = .01f + 1.21f*frame_probs[0]*frame_probs[0] - .23f*(float)pow(frame_probs[0], 10); + /*frame_probs[0] = .01f + 1.21f*frame_probs[0]*frame_probs[0] - .23f*(float)pow(frame_probs[0], 10);*/ /* Probability of active audio (as opposed to silence) */ frame_probs[1] = .5f*frame_probs[1]+.5f; - /* Consider that silence has a 50-50 probability. */ - frame_probs[0] = frame_probs[1]*frame_probs[0] + (1-frame_probs[1])*.5f; + frame_probs[1] *= frame_probs[1]; - /*printf("%f %f ", frame_probs[0], frame_probs[1]);*/ + /* Probability of speech or music vs noise */ + info->activity_probability = frame_probs[1]; + + /*printf("%f %f\n", frame_probs[0], frame_probs[1]);*/ { /* Probability of state transition */ float tau; @@ -542,12 +790,32 @@ static void tonality_analysis(TonalityAnalysisState *tonal, const CELTMode *celt float music0; float p, q; + /* More silence transitions for speech than for music. */ + tau = .001f*tonal->music_prob + .01f*(1-tonal->music_prob); + p = MAX16(.05f,MIN16(.95f,frame_probs[1])); + q = MAX16(.05f,MIN16(.95f,tonal->vad_prob)); + beta = .02f+.05f*ABS16(p-q)/(p*(1-q)+q*(1-p)); + /* p0 and p1 are the probabilities of speech and music at this frame + using only information from previous frame and applying the + state transition model */ + p0 = (1-tonal->vad_prob)*(1-tau) + tonal->vad_prob *tau; + p1 = tonal->vad_prob *(1-tau) + (1-tonal->vad_prob)*tau; + /* We apply the current probability with exponent beta to work around + the fact that the probability estimates aren't independent. */ + p0 *= (float)pow(1-frame_probs[1], beta); + p1 *= (float)pow(frame_probs[1], beta); + /* Normalise the probabilities to get the Marokv probability of music. */ + tonal->vad_prob = p1/(p0+p1); + info->vad_prob = tonal->vad_prob; + /* Consider that silence has a 50-50 probability of being speech or music. */ + frame_probs[0] = tonal->vad_prob*frame_probs[0] + (1-tonal->vad_prob)*.5f; + /* One transition every 3 minutes of active audio */ - tau = .00005f*frame_probs[1]; + tau = .0001f; /* Adapt beta based on how "unexpected" the new prob is */ p = MAX16(.05f,MIN16(.95f,frame_probs[0])); q = MAX16(.05f,MIN16(.95f,tonal->music_prob)); - beta = .01f+.05f*ABS16(p-q)/(p*(1-q)+q*(1-p)); + beta = .02f+.05f*ABS16(p-q)/(p*(1-q)+q*(1-p)); /* p0 and p1 are the probabilities of speech and music at this frame using only information from previous frame and applying the state transition model */ @@ -561,6 +829,7 @@ static void tonality_analysis(TonalityAnalysisState *tonal, const CELTMode *celt tonal->music_prob = p1/(p0+p1); info->music_prob = tonal->music_prob; + /*printf("%f %f %f %f\n", frame_probs[0], frame_probs[1], tonal->music_prob, tonal->vad_prob);*/ /* This chunk of code deals with delayed decision. */ psum=1e-20f; /* Instantaneous probability of speech and music, with beta pre-applied. */ @@ -568,8 +837,11 @@ static void tonality_analysis(TonalityAnalysisState *tonal, const CELTMode *celt music0 = (float)pow(frame_probs[0], beta); if (tonal->count==1) { - tonal->pspeech[0]=.5; - tonal->pmusic [0]=.5; + if (tonal->application == OPUS_APPLICATION_VOIP) + tonal->pmusic[0] = .1f; + else + tonal->pmusic[0] = .625f; + tonal->pspeech[0] = 1-tonal->pmusic[0]; } /* Updated probability of having only speech (s0) or only music (m0), before considering the new observation. */ @@ -619,24 +891,17 @@ static void tonality_analysis(TonalityAnalysisState *tonal, const CELTMode *celt tonal->speech_confidence_count = IMIN(tonal->speech_confidence_count, 500); tonal->speech_confidence += adapt*MIN16(.2f,frame_probs[0]-tonal->speech_confidence); } - } else { - if (tonal->music_confidence_count==0) - tonal->music_confidence = .9f; - if (tonal->speech_confidence_count==0) - tonal->speech_confidence = .1f; } } - if (tonal->last_music != (tonal->music_prob>.5f)) - tonal->last_transition=0; tonal->last_music = tonal->music_prob>.5f; -#else - info->music_prob = 0; -#endif - /*for (i=0;i<25;i++) +#ifdef MLP_TRAINING + for (i=0;i<25;i++) printf("%f ", features[i]); - printf("\n");*/ + printf("\n"); +#endif info->bandwidth = bandwidth; + tonal->prev_bandwidth = bandwidth; /*printf("%d %d\n", info->bandwidth, info->opus_bandwidth);*/ info->noisiness = frame_noisiness; info->valid = 1; @@ -650,18 +915,19 @@ void run_analysis(TonalityAnalysisState *analysis, const CELTMode *celt_mode, co int offset; int pcm_len; + analysis_frame_size -= analysis_frame_size&1; if (analysis_pcm != NULL) { /* Avoid overflow/wrap-around of the analysis buffer */ - analysis_frame_size = IMIN((DETECT_SIZE-5)*Fs/100, analysis_frame_size); + analysis_frame_size = IMIN((DETECT_SIZE-5)*Fs/50, analysis_frame_size); pcm_len = analysis_frame_size - analysis->analysis_offset; offset = analysis->analysis_offset; - do { - tonality_analysis(analysis, celt_mode, analysis_pcm, IMIN(480, pcm_len), offset, c1, c2, C, lsb_depth, downmix); - offset += 480; - pcm_len -= 480; - } while (pcm_len>0); + while (pcm_len>0) { + tonality_analysis(analysis, celt_mode, analysis_pcm, IMIN(Fs/50, pcm_len), offset, c1, c2, C, lsb_depth, downmix); + offset += Fs/50; + pcm_len -= Fs/50; + } analysis->analysis_offset = analysis_frame_size; analysis->analysis_offset -= frame_size; @@ -670,3 +936,5 @@ void run_analysis(TonalityAnalysisState *analysis, const CELTMode *celt_mode, co analysis_info->valid = 0; tonality_get_info(analysis, analysis_info, frame_size); } + +#endif /* DISABLE_FLOAT_API */ diff --git a/code/opus-1.1.4/src/analysis.h b/code/opus-1.1.4/src/analysis.h index 9eae56a5..cac51dfa 100644 --- a/code/opus-1.1.4/src/analysis.h +++ b/code/opus-1.1.4/src/analysis.h @@ -33,13 +33,20 @@ #define NB_FRAMES 8 #define NB_TBANDS 18 -#define NB_TOT_BANDS 21 -#define ANALYSIS_BUF_SIZE 720 /* 15 ms at 48 kHz */ +#define ANALYSIS_BUF_SIZE 720 /* 30 ms at 24 kHz */ -#define DETECT_SIZE 200 +/* At that point we can stop counting frames because it no longer matters. */ +#define ANALYSIS_COUNT_MAX 10000 + +#define DETECT_SIZE 100 + +/* Uncomment this to print the MLP features on stdout. */ +/*#define MLP_TRAINING*/ typedef struct { int arch; + int application; + opus_int32 Fs; #define TONALITY_ANALYSIS_RESET_START angle float angle[240]; float d_angle[240]; @@ -48,21 +55,22 @@ typedef struct { int mem_fill; /* number of usable samples in the buffer */ float prev_band_tonality[NB_TBANDS]; float prev_tonality; + int prev_bandwidth; float E[NB_FRAMES][NB_TBANDS]; + float logE[NB_FRAMES][NB_TBANDS]; float lowE[NB_TBANDS]; float highE[NB_TBANDS]; - float meanE[NB_TOT_BANDS]; + float meanE[NB_TBANDS+1]; float mem[32]; float cmean[8]; float std[9]; float music_prob; + float vad_prob; float Etracker; float lowECount; int E_count; int last_music; - int last_transition; int count; - float subframe_mem[3]; int analysis_offset; /** Probability of having speech for time i to DETECT_SIZE-1 (and music before). pspeech[0] is the probability that all frames in the window are speech. */ @@ -77,6 +85,8 @@ typedef struct { int write_pos; int read_pos; int read_subframe; + float hp_ener_accum; + opus_val32 downmix_state[3]; AnalysisInfo info[DETECT_SIZE]; } TonalityAnalysisState; @@ -86,7 +96,7 @@ typedef struct { * not be repeated every analysis step. No allocated memory is retained * by the state struct, so no cleanup call is required. */ -void tonality_analysis_init(TonalityAnalysisState *analysis); +void tonality_analysis_init(TonalityAnalysisState *analysis, opus_int32 Fs); /** Reset a TonalityAnalysisState stuct. * diff --git a/code/opus-1.1.4/src/mlp_data.c b/code/opus-1.1.4/src/mlp_data.c index c2fda4e2..a819880b 100644 --- a/code/opus-1.1.4/src/mlp_data.c +++ b/code/opus-1.1.4/src/mlp_data.c @@ -1,106 +1,109 @@ -/* The contents of this file was automatically generated by mlp_train.c - It contains multi-layer perceptron (MLP) weights. */ - #ifdef HAVE_CONFIG_H #include "config.h" #endif #include "mlp.h" -/* RMS error was 0.138320, seed was 1361535663 */ +/* RMS error was 0.280492, seed was 1480478173 */ +/* 0.005976 0.031821 (0.280494 0.280492) done */ -static const float weights[422] = { +static const float weights[450] = { /* hidden layer */ --0.0941125f, -0.302976f, -0.603555f, -0.19393f, -0.185983f, --0.601617f, -0.0465317f, -0.114563f, -0.103599f, -0.618938f, --0.317859f, -0.169949f, -0.0702885f, 0.148065f, 0.409524f, -0.548432f, 0.367649f, -0.494393f, 0.764306f, -1.83957f, -0.170849f, 12.786f, -1.08848f, -1.27284f, -16.2606f, -24.1773f, -5.57454f, -0.17276f, -0.163388f, -0.224421f, --0.0948944f, -0.0728695f, -0.26557f, -0.100283f, -0.0515459f, --0.146142f, -0.120674f, -0.180655f, 0.12857f, 0.442138f, --0.493735f, 0.167767f, 0.206699f, -0.197567f, 0.417999f, -1.50364f, -0.773341f, -10.0401f, 0.401872f, 2.97966f, -15.2165f, -1.88905f, -1.19254f, 0.0285397f, -0.00405139f, -0.0707565f, 0.00825699f, -0.0927269f, -0.010393f, -0.00428882f, --0.00489743f, -0.0709731f, -0.00255992f, 0.0395619f, 0.226424f, -0.0325231f, 0.162175f, -0.100118f, 0.485789f, 0.12697f, -0.285937f, 0.0155637f, 0.10546f, 3.05558f, 1.15059f, --1.00904f, -1.83088f, 3.31766f, -3.42516f, -0.119135f, --0.0405654f, 0.00690068f, 0.0179877f, -0.0382487f, 0.00597941f, --0.0183611f, 0.00190395f, -0.144322f, -0.0435671f, 0.000990594f, -0.221087f, 0.142405f, 0.484066f, 0.404395f, 0.511955f, --0.237255f, 0.241742f, 0.35045f, -0.699428f, 10.3993f, -2.6507f, -2.43459f, -4.18838f, 1.05928f, 1.71067f, -0.00667811f, -0.0721335f, -0.0397346f, 0.0362704f, -0.11496f, --0.0235776f, 0.0082161f, -0.0141741f, -0.0329699f, -0.0354253f, -0.00277404f, -0.290654f, -1.14767f, -0.319157f, -0.686544f, -0.36897f, 0.478899f, 0.182579f, -0.411069f, 0.881104f, --4.60683f, 1.4697f, 0.335845f, -1.81905f, -30.1699f, -5.55225f, 0.0019508f, -0.123576f, -0.0727332f, -0.0641597f, --0.0534458f, -0.108166f, -0.0937368f, -0.0697883f, -0.0275475f, --0.192309f, -0.110074f, 0.285375f, -0.405597f, 0.0926724f, --0.287881f, -0.851193f, -0.099493f, -0.233764f, -1.2852f, -1.13611f, 3.12168f, -0.0699f, -1.86216f, 2.65292f, --7.31036f, 2.44776f, -0.00111802f, -0.0632786f, -0.0376296f, --0.149851f, 0.142963f, 0.184368f, 0.123433f, 0.0756158f, -0.117312f, 0.0933395f, 0.0692163f, 0.0842592f, 0.0704683f, -0.0589963f, 0.0942205f, -0.448862f, 0.0262677f, 0.270352f, --0.262317f, 0.172586f, 2.00227f, -0.159216f, 0.038422f, -10.2073f, 4.15536f, -2.3407f, -0.0550265f, 0.00964792f, --0.141336f, 0.0274501f, 0.0343921f, -0.0487428f, 0.0950172f, --0.00775017f, -0.0372492f, -0.00548121f, -0.0663695f, 0.0960506f, --0.200008f, -0.0412827f, 0.58728f, 0.0515787f, 0.337254f, -0.855024f, 0.668371f, -0.114904f, -3.62962f, -0.467477f, --0.215472f, 2.61537f, 0.406117f, -1.36373f, 0.0425394f, -0.12208f, 0.0934502f, 0.123055f, 0.0340935f, -0.142466f, -0.035037f, -0.0490666f, 0.0733208f, 0.0576672f, 0.123984f, --0.0517194f, -0.253018f, 0.590565f, 0.145849f, 0.315185f, -0.221534f, -0.149081f, 0.216161f, -0.349575f, 24.5664f, --0.994196f, 0.614289f, -18.7905f, -2.83277f, -0.716801f, --0.347201f, 0.479515f, -0.246027f, 0.0758683f, 0.137293f, --0.17781f, 0.118751f, -0.00108329f, -0.237334f, 0.355732f, --0.12991f, -0.0547627f, -0.318576f, -0.325524f, 0.180494f, --0.0625604f, 0.141219f, 0.344064f, 0.37658f, -0.591772f, -5.8427f, -0.38075f, 0.221894f, -1.41934f, -1.87943e+06f, -1.34114f, 0.0283355f, -0.0447856f, -0.0211466f, -0.0256927f, -0.0139618f, 0.0207934f, -0.0107666f, 0.0110969f, 0.0586069f, --0.0253545f, -0.0328433f, 0.11872f, -0.216943f, 0.145748f, -0.119808f, -0.0915211f, -0.120647f, -0.0787719f, -0.143644f, --0.595116f, -1.152f, -1.25335f, -1.17092f, 4.34023f, --975268.f, -1.37033f, -0.0401123f, 0.210602f, -0.136656f, -0.135962f, -0.0523293f, 0.0444604f, 0.0143928f, 0.00412666f, --0.0193003f, 0.218452f, -0.110204f, -2.02563f, 0.918238f, --2.45362f, 1.19542f, -0.061362f, -1.92243f, 0.308111f, -0.49764f, 0.912356f, 0.209272f, -2.34525f, 2.19326f, --6.47121f, 1.69771f, -0.725123f, 0.0118929f, 0.0377944f, -0.0554003f, 0.0226452f, -0.0704421f, -0.0300309f, 0.0122978f, --0.0041782f, -0.0686612f, 0.0313115f, 0.039111f, 0.364111f, --0.0945548f, 0.0229876f, -0.17414f, 0.329795f, 0.114714f, -0.30022f, 0.106997f, 0.132355f, 5.79932f, 0.908058f, --0.905324f, -3.3561f, 0.190647f, 0.184211f, -0.673648f, -0.231807f, -0.0586222f, 0.230752f, -0.438277f, 0.245857f, --0.17215f, 0.0876383f, -0.720512f, 0.162515f, 0.0170571f, -0.101781f, 0.388477f, 1.32931f, 1.08548f, -0.936301f, --2.36958f, -6.71988f, -3.44376f, 2.13818f, 14.2318f, -4.91459f, -3.09052f, -9.69191f, -0.768234f, 1.79604f, -0.0549653f, 0.163399f, 0.0797025f, 0.0343933f, -0.0555876f, --0.00505673f, 0.0187258f, 0.0326628f, 0.0231486f, 0.15573f, -0.0476223f, -0.254824f, 1.60155f, -0.801221f, 2.55496f, -0.737629f, -1.36249f, -0.695463f, -2.44301f, -1.73188f, -3.95279f, 1.89068f, 0.486087f, -11.3343f, 3.9416e+06f, - +-0.514624f, 0.0234227f, -0.14329f, -0.0878216f, -0.00187827f, +-0.0257443f, 0.108524f, 0.00333881f, 0.00585017f, -0.0246132f, +0.142723f, -0.00436494f, 0.0101354f, -0.11124f, -0.0809367f, +-0.0750772f, 0.0295524f, 0.00823944f, 0.150392f, 0.0320876f, +-0.0710564f, -1.43818f, 0.652076f, 0.0650744f, -1.54821f, +0.168949f, -1.92724f, 0.0517976f, -0.0670737f, -0.0690121f, +0.00247528f, -0.0522024f, 0.0631368f, 0.0532776f, 0.047751f, +-0.011715f, 0.142374f, -0.0290885f, -0.279263f, -0.433499f, +-0.0795174f, -0.380458f, -0.051263f, 0.218537f, -0.322478f, +1.06667f, -0.104607f, -4.70108f, 0.312037f, 0.277397f, +-2.71859f, 1.70037f, -0.141845f, 0.0115618f, 0.0629883f, +0.0403871f, 0.0139428f, -0.00430733f, -0.0429038f, -0.0590318f, +-0.0501526f, -0.0284802f, -0.0415686f, -0.0438999f, 0.0822666f, +0.197194f, 0.0363275f, -0.0584307f, 0.0752364f, -0.0799796f, +-0.146275f, 0.161661f, -0.184585f, 0.145568f, 0.442823f, +1.61221f, 1.11162f, 2.62177f, -2.482f, -0.112599f, +-0.110366f, -0.140794f, -0.181694f, 0.0648674f, 0.0842248f, +0.0933993f, 0.150122f, 0.129171f, 0.176848f, 0.141758f, +-0.271822f, 0.235113f, 0.0668579f, -0.433957f, 0.113633f, +-0.169348f, -1.40091f, 0.62861f, -0.134236f, 0.402173f, +1.86373f, 1.53998f, -4.32084f, 0.735343f, 0.800214f, +-0.00968415f, 0.0425904f, 0.0196811f, -0.018426f, -0.000343953f, +-0.00416389f, 0.00111558f, 0.0173069f, -0.00998596f, -0.025898f, +0.00123764f, -0.00520373f, -0.0565033f, 0.0637394f, 0.0051213f, +0.0221361f, 0.00819962f, -0.0467061f, -0.0548258f, -0.00314063f, +-1.18332f, 1.88091f, -0.41148f, -2.95727f, -0.521449f, +-0.271641f, 0.124946f, -0.0532936f, 0.101515f, 0.000208564f, +-0.0488748f, 0.0642388f, -0.0383848f, 0.0135046f, -0.0413592f, +-0.0326402f, -0.0137421f, -0.0225219f, -0.0917294f, -0.277759f, +-0.185418f, 0.0471128f, -0.125879f, 0.262467f, -0.212794f, +-0.112931f, -1.99885f, -0.404787f, 0.224402f, 0.637962f, +-0.27808f, -0.0723953f, -0.0537655f, -0.0336359f, -0.0906601f, +-0.0641309f, -0.0713542f, 0.0524317f, 0.00608819f, 0.0754101f, +-0.0488401f, -0.00671865f, 0.0418239f, 0.0536284f, -0.132639f, +0.0267648f, -0.248432f, -0.0104153f, 0.035544f, -0.212753f, +-0.302895f, -0.0357854f, 0.376838f, 0.597025f, -0.664647f, +0.268422f, -0.376772f, -1.05472f, 0.0144178f, 0.179122f, +0.0360155f, 0.220262f, -0.0056381f, 0.0317197f, 0.0621066f, +-0.00779298f, 0.00789378f, 0.00350605f, 0.0104809f, 0.0362871f, +-0.157708f, -0.0659779f, -0.0926278f, 0.00770791f, 0.0631621f, +0.0817343f, -0.424295f, -0.0437727f, -0.24251f, 0.711217f, +-0.736455f, -2.194f, -0.107612f, -0.175156f, -0.0366573f, +-0.0123156f, -0.0628516f, -0.0218977f, -0.00693699f, 0.00695185f, +0.00507362f, 0.00359334f, 0.0052661f, 0.035561f, 0.0382701f, +0.0342179f, -0.00790271f, -0.0170925f, 0.047029f, 0.0197362f, +-0.0153435f, 0.0644152f, -0.36862f, -0.0674876f, -2.82672f, +1.34122f, -0.0788029f, -3.47792f, 0.507246f, -0.816378f, +-0.0142383f, -0.127349f, -0.106926f, -0.0359524f, 0.105045f, +0.291554f, 0.195413f, 0.0866214f, -0.066577f, -0.102188f, +0.0979466f, -0.12982f, 0.400181f, -0.409336f, -0.0593326f, +-0.0656203f, -0.204474f, 0.179802f, 0.000509084f, 0.0995954f, +-2.377f, -0.686359f, 0.934861f, 1.10261f, 1.3901f, +-4.33616f, -0.00264017f, 0.00713045f, 0.106264f, 0.143726f, +-0.0685305f, -0.054656f, -0.0176725f, -0.0772669f, -0.0264526f, +-0.0103824f, -0.0269872f, -0.00687f, 0.225804f, 0.407751f, +-0.0612611f, -0.0576863f, -0.180131f, -0.222772f, -0.461742f, +0.335236f, 1.03399f, 4.24112f, -0.345796f, -0.594549f, +-76.1407f, -0.265276f, 0.0507719f, 0.0643044f, 0.0384832f, +0.0424459f, -0.0387817f, -0.0235996f, -0.0740556f, -0.0270029f, +0.00882177f, -0.0552371f, -0.00485851f, 0.314295f, 0.360431f, +-0.0787085f, 0.110355f, -0.415958f, -0.385088f, -0.272224f, +-1.55108f, -0.141848f, 0.448877f, -0.563447f, -2.31403f, +-0.120077f, -1.49918f, -0.817726f, -0.0495854f, -0.0230782f, +-0.0224014f, 0.117076f, 0.0393216f, 0.051997f, 0.0330763f, +-0.110796f, 0.0211117f, -0.0197258f, 0.0187461f, 0.0125183f, +0.14876f, 0.0920565f, -0.342475f, 0.135272f, -0.168155f, +-0.033423f, -0.0604611f, -0.128835f, 0.664947f, -0.144997f, +2.27649f, 1.28663f, 0.841217f, -2.42807f, 0.0230471f, +0.226709f, -0.0374803f, 0.155436f, 0.0400342f, -0.184686f, +0.128488f, -0.0939518f, -0.0578559f, 0.0265967f, -0.0999322f, +-0.0322768f, -0.322994f, -0.189371f, -0.738069f, -0.0754914f, +0.214717f, -0.093728f, -0.695741f, 0.0899298f, -2.06188f, +-0.273719f, -0.896977f, 0.130553f, 0.134638f, 1.29355f, +0.00520749f, -0.0324224f, 0.00530451f, 0.0192385f, 0.00328708f, +0.0250838f, 0.0053365f, -0.0177321f, 0.00618789f, 0.00525364f, +0.00104596f, -0.0360459f, 0.0402403f, -0.0406351f, 0.0136883f, +0.0880722f, -0.0197449f, 0.089938f, 0.0100456f, -0.0475638f, +-0.73267f, 0.037433f, -0.146551f, -0.230221f, -3.06489f, +-1.40194f, 0.0198483f, 0.0397953f, -0.0190239f, 0.0470715f, +-0.131363f, -0.191721f, -0.0176224f, -0.0480352f, -0.221799f, +-0.26794f, -0.0292615f, 0.0612127f, -0.129877f, 0.00628332f, +-0.085918f, 0.0175379f, 0.0541011f, -0.0810874f, -0.380809f, +-0.222056f, -0.508859f, -0.473369f, 0.484958f, -2.28411f, +0.0139516f, /* output layer */ --0.381439f, 0.12115f, -0.906927f, 2.93878f, 1.6388f, -0.882811f, 0.874344f, 1.21726f, -0.874545f, 0.321706f, -0.785055f, 0.946558f, -0.575066f, -3.46553f, 0.884905f, -0.0924047f, -9.90712f, 0.391338f, 0.160103f, -2.04954f, -4.1455f, 0.0684029f, -0.144761f, -0.285282f, 0.379244f, --1.1584f, -0.0277241f, -9.85f, -4.82386f, 3.71333f, -3.87308f, 3.52558f}; +3.90017f, 1.71789f, -1.43372f, -2.70839f, 1.77107f, +5.48006f, 1.44661f, 2.01134f, -1.88383f, -3.64958f, +-1.26351f, 0.779421f, 2.11357f, 3.10409f, 1.68846f, +-4.46197f, -1.61455f, 3.59832f, 2.43531f, -1.26458f, +0.417941f, 1.47437f, 2.16635f, -1.909f, -0.828869f, +1.38805f, -2.67975f, -0.110044f, 1.95596f, 0.697931f, +-0.313226f, -0.889315f, 0.283236f, 0.946102f, }; -static const int topo[3] = {25, 15, 2}; +static const int topo[3] = {25, 16, 2}; const MLP net = { 3, diff --git a/code/opus-1.1.4/src/opus.c b/code/opus-1.1.4/src/opus.c index f76f125c..cdbd13a1 100644 --- a/code/opus-1.1.4/src/opus.c +++ b/code/opus-1.1.4/src/opus.c @@ -107,7 +107,7 @@ OPUS_EXPORT void opus_pcm_soft_clip(float *_x, int N, int C, float *declip_mem) /* Slightly boost "a" by 2^-22. This is just enough to ensure -ffast-math does not cause output values larger than +/-1, but small enough not to matter even for 24-bit output. */ - a += a*2.4e-7; + a += a*2.4e-7f; if (x[i*C]>0) a = -a; /* Apply soft clipping */ diff --git a/code/opus-1.1.4/src/opus_decoder.c b/code/opus-1.1.4/src/opus_decoder.c index 080bec50..149ae7f2 100644 --- a/code/opus-1.1.4/src/opus_decoder.c +++ b/code/opus-1.1.4/src/opus_decoder.c @@ -891,7 +891,7 @@ int opus_decoder_ctl(OpusDecoder *st, int request, ...) break; case OPUS_GET_LAST_PACKET_DURATION_REQUEST: { - opus_uint32 *value = va_arg(ap, opus_uint32*); + opus_int32 *value = va_arg(ap, opus_int32*); if (!value) { goto bad_arg; @@ -899,6 +899,26 @@ int opus_decoder_ctl(OpusDecoder *st, int request, ...) *value = st->last_packet_duration; } break; + case OPUS_SET_PHASE_INVERSION_DISABLED_REQUEST: + { + opus_int32 value = va_arg(ap, opus_int32); + if(value<0 || value>1) + { + goto bad_arg; + } + celt_decoder_ctl(celt_dec, OPUS_SET_PHASE_INVERSION_DISABLED(value)); + } + break; + case OPUS_GET_PHASE_INVERSION_DISABLED_REQUEST: + { + opus_int32 *value = va_arg(ap, opus_int32*); + if (!value) + { + goto bad_arg; + } + celt_decoder_ctl(celt_dec, OPUS_GET_PHASE_INVERSION_DISABLED(value)); + } + break; default: /*fprintf(stderr, "unknown opus_decoder_ctl() request: %d", request);*/ ret = OPUS_UNIMPLEMENTED; diff --git a/code/opus-1.1.4/src/opus_encoder.c b/code/opus-1.1.4/src/opus_encoder.c index 9a516a88..474e6c24 100644 --- a/code/opus-1.1.4/src/opus_encoder.c +++ b/code/opus-1.1.4/src/opus_encoder.c @@ -53,6 +53,10 @@ #define MAX_ENCODER_BUFFER 480 +#ifndef DISABLE_FLOAT_API +#define PSEUDO_SNR_THRESHOLD 316.23f /* 10^(25/10) */ +#endif + typedef struct { opus_val32 XX, XY, YY; opus_val16 smoothed_width; @@ -82,6 +86,7 @@ struct OpusEncoder { int encoder_buffer; int lfe; int arch; + int use_dtx; /* general DTX for both SILK and CELT */ #ifndef DISABLE_FLOAT_API TonalityAnalysisState analysis; #endif @@ -97,6 +102,8 @@ struct OpusEncoder { int prev_channels; int prev_framesize; int bandwidth; + /* Bandwidth determined automatically from the rate (before any other adjustment) */ + int auto_bandwidth; int silk_bw_switch; /* Sampling rate (at the API level) */ int first; @@ -105,7 +112,10 @@ struct OpusEncoder { opus_val16 delay_buffer[MAX_ENCODER_BUFFER*2]; #ifndef DISABLE_FLOAT_API int detected_bandwidth; + int nb_no_activity_frames; + opus_val32 peak_signal_energy; #endif + int nonfinal_frame; /* current frame is not the final in a packet */ opus_uint32 rangeFinal; }; @@ -113,32 +123,32 @@ struct OpusEncoder { middle (memoriless) threshold. The second column is the hysteresis (difference with the middle) */ static const opus_int32 mono_voice_bandwidth_thresholds[8] = { - 11000, 1000, /* NB<->MB */ - 14000, 1000, /* MB<->WB */ - 17000, 1000, /* WB<->SWB */ - 21000, 2000, /* SWB<->FB */ + 10000, 1000, /* NB<->MB */ + 11000, 1000, /* MB<->WB */ + 13500, 1000, /* WB<->SWB */ + 14000, 2000, /* SWB<->FB */ }; static const opus_int32 mono_music_bandwidth_thresholds[8] = { - 12000, 1000, /* NB<->MB */ - 15000, 1000, /* MB<->WB */ - 18000, 2000, /* WB<->SWB */ - 22000, 2000, /* SWB<->FB */ + 10000, 1000, /* NB<->MB */ + 11000, 1000, /* MB<->WB */ + 13500, 1000, /* WB<->SWB */ + 14000, 2000, /* SWB<->FB */ }; static const opus_int32 stereo_voice_bandwidth_thresholds[8] = { - 11000, 1000, /* NB<->MB */ - 14000, 1000, /* MB<->WB */ - 21000, 2000, /* WB<->SWB */ - 28000, 2000, /* SWB<->FB */ + 10000, 1000, /* NB<->MB */ + 11000, 1000, /* MB<->WB */ + 13500, 1000, /* WB<->SWB */ + 14000, 2000, /* SWB<->FB */ }; static const opus_int32 stereo_music_bandwidth_thresholds[8] = { - 12000, 1000, /* NB<->MB */ - 18000, 2000, /* MB<->WB */ - 21000, 2000, /* WB<->SWB */ - 30000, 2000, /* SWB<->FB */ + 10000, 1000, /* NB<->MB */ + 11000, 1000, /* MB<->WB */ + 13500, 1000, /* WB<->SWB */ + 14000, 2000, /* SWB<->FB */ }; /* Threshold bit-rates for switching between mono and stereo */ -static const opus_int32 stereo_voice_threshold = 30000; -static const opus_int32 stereo_music_threshold = 30000; +static const opus_int32 stereo_voice_threshold = 24000; +static const opus_int32 stereo_music_threshold = 24000; /* Threshold bit-rate for switching between SILK/hybrid and CELT-only */ static const opus_int32 mode_thresholds[2][2] = { @@ -147,6 +157,14 @@ static const opus_int32 mode_thresholds[2][2] = { { 36000, 16000}, /* stereo */ }; +static const opus_int32 fec_thresholds[] = { + 12000, 1000, /* NB */ + 14000, 1000, /* MB */ + 16000, 1000, /* WB */ + 20000, 1000, /* SWB */ + 22000, 1000, /* FB */ +}; + int opus_encoder_get_size(int channels) { int silkEncSizeBytes, celtEncSizeBytes; @@ -245,7 +263,8 @@ int opus_encoder_init(OpusEncoder* st, opus_int32 Fs, int channels, int applicat st->bandwidth = OPUS_BANDWIDTH_FULLBAND; #ifndef DISABLE_FLOAT_API - tonality_analysis_init(&st->analysis); + tonality_analysis_init(&st->analysis, st->Fs); + st->analysis.application = st->application; #endif return OPUS_OK; @@ -323,10 +342,11 @@ static void silk_biquad_float( } #endif -static void hp_cutoff(const opus_val16 *in, opus_int32 cutoff_Hz, opus_val16 *out, opus_val32 *hp_mem, int len, int channels, opus_int32 Fs) +static void hp_cutoff(const opus_val16 *in, opus_int32 cutoff_Hz, opus_val16 *out, opus_val32 *hp_mem, int len, int channels, opus_int32 Fs, int arch) { opus_int32 B_Q28[ 3 ], A_Q28[ 2 ]; opus_int32 Fc_Q19, r_Q28, r_Q22; + (void)arch; silk_assert( cutoff_Hz <= silk_int32_MAX / SILK_FIX_CONST( 1.5 * 3.14159 / 1000, 19 ) ); Fc_Q19 = silk_DIV32_16( silk_SMULBB( SILK_FIX_CONST( 1.5 * 3.14159 / 1000, 19 ), cutoff_Hz ), Fs/1000 ); @@ -346,9 +366,10 @@ static void hp_cutoff(const opus_val16 *in, opus_int32 cutoff_Hz, opus_val16 *ou A_Q28[ 1 ] = silk_SMULWW( r_Q22, r_Q22 ); #ifdef FIXED_POINT - silk_biquad_alt( in, B_Q28, A_Q28, hp_mem, out, len, channels ); - if( channels == 2 ) { - silk_biquad_alt( in+1, B_Q28, A_Q28, hp_mem+2, out+1, len, channels ); + if( channels == 1 ) { + silk_biquad_alt_stride1( in, B_Q28, A_Q28, hp_mem, out, len ); + } else { + silk_biquad_alt_stride2( in, B_Q28, A_Q28, hp_mem, out, len, arch ); } #else silk_biquad_float( in, B_Q28, A_Q28, hp_mem, out, len, channels ); @@ -371,14 +392,14 @@ static void dc_reject(const opus_val16 *in, opus_int32 cutoff_Hz, opus_val16 *ou for (i=0;iuser_bitrate_bps; } -#ifndef DISABLE_FLOAT_API -/* Don't use more than 60 ms for the frame size analysis */ -#define MAX_DYNAMIC_FRAMESIZE 24 -/* Estimates how much the bitrate will be boosted based on the sub-frame energy */ -static float transient_boost(const float *E, const float *E_1, int LM, int maxM) -{ - int i; - int M; - float sumE=0, sumE_1=0; - float metric; - - M = IMIN(maxM, (1<10 ? 1 : 0;*/ - /*return MAX16(0,1-exp(-.25*(metric-2.)));*/ - return MIN16(1,(float)sqrt(MAX16(0,.05f*(metric-2)))); -} - -/* Viterbi decoding trying to find the best frame size combination using look-ahead - - State numbering: - 0: unused - 1: 2.5 ms - 2: 5 ms (#1) - 3: 5 ms (#2) - 4: 10 ms (#1) - 5: 10 ms (#2) - 6: 10 ms (#3) - 7: 10 ms (#4) - 8: 20 ms (#1) - 9: 20 ms (#2) - 10: 20 ms (#3) - 11: 20 ms (#4) - 12: 20 ms (#5) - 13: 20 ms (#6) - 14: 20 ms (#7) - 15: 20 ms (#8) -*/ -static int transient_viterbi(const float *E, const float *E_1, int N, int frame_cost, int rate) -{ - int i; - float cost[MAX_DYNAMIC_FRAMESIZE][16]; - int states[MAX_DYNAMIC_FRAMESIZE][16]; - float best_cost; - int best_state; - float factor; - /* Take into account that we damp VBR in the 32 kb/s to 64 kb/s range. */ - if (rate<80) - factor=0; - else if (rate>160) - factor=1; - else - factor = (rate-80.f)/80.f; - /* Makes variable framesize less aggressive at lower bitrates, but I can't - find any valid theoretical justification for this (other than it seems - to help) */ - for (i=0;i<16;i++) - { - /* Impossible state */ - states[0][i] = -1; - cost[0][i] = 1e10; - } - for (i=0;i<4;i++) - { - cost[0][1<=0;i--) - { - /*printf("%d ", best_state);*/ - best_state = states[i][best_state]; - } - /*printf("%d\n", best_state);*/ - return best_state; -} - -static int optimize_framesize(const void *x, int len, int C, opus_int32 Fs, - int bitrate, opus_val16 tonality, float *mem, int buffering, - downmix_func downmix) -{ - int N; - int i; - float e[MAX_DYNAMIC_FRAMESIZE+4]; - float e_1[MAX_DYNAMIC_FRAMESIZE+3]; - opus_val32 memx; - int bestLM=0; - int subframe; - int pos; - int offset; - VARDECL(opus_val32, sub); - - subframe = Fs/400; - ALLOC(sub, subframe, opus_val32); - e[0]=mem[0]; - e_1[0]=1.f/(EPSILON+mem[0]); - if (buffering) - { - /* Consider the CELT delay when not in restricted-lowdelay */ - /* We assume the buffering is between 2.5 and 5 ms */ - offset = 2*subframe - buffering; - celt_assert(offset>=0 && offset <= subframe); - len -= offset; - e[1]=mem[1]; - e_1[1]=1.f/(EPSILON+mem[1]); - e[2]=mem[2]; - e_1[2]=1.f/(EPSILON+mem[2]); - pos = 3; - } else { - pos=1; - offset=0; - } - N=IMIN(len/subframe, MAX_DYNAMIC_FRAMESIZE); - /* Just silencing a warning, it's really initialized later */ - memx = 0; - for (i=0;i-1) { for (j=0;j-1) { for (j=0;j= OPUS_FRAMESIZE_2_5_MS && variable_duration <= OPUS_FRAMESIZE_60_MS) - new_size = IMIN(3*Fs/50, (Fs/400)<<(variable_duration-OPUS_FRAMESIZE_2_5_MS)); + else if (variable_duration >= OPUS_FRAMESIZE_2_5_MS && variable_duration <= OPUS_FRAMESIZE_120_MS) + { + if (variable_duration <= OPUS_FRAMESIZE_40_MS) + new_size = (Fs/400)<<(variable_duration-OPUS_FRAMESIZE_2_5_MS); + else + new_size = (variable_duration-OPUS_FRAMESIZE_2_5_MS-2)*Fs/50; + } else return -1; if (new_size>frame_size) return -1; - if (400*new_size!=Fs && 200*new_size!=Fs && 100*new_size!=Fs && - 50*new_size!=Fs && 25*new_size!=Fs && 50*new_size!=3*Fs) + if (400*new_size!=Fs && 200*new_size!=Fs && 100*new_size!=Fs && + 50*new_size!=Fs && 25*new_size!=Fs && 50*new_size!=3*Fs && + 50*new_size!=4*Fs && 50*new_size!=5*Fs && 50*new_size!=6*Fs) return -1; return new_size; } -opus_int32 compute_frame_size(const void *analysis_pcm, int frame_size, - int variable_duration, int C, opus_int32 Fs, int bitrate_bps, - int delay_compensation, downmix_func downmix -#ifndef DISABLE_FLOAT_API - , float *subframe_mem -#endif - ) -{ -#ifndef DISABLE_FLOAT_API - if (variable_duration == OPUS_FRAMESIZE_VARIABLE && frame_size >= Fs/200) - { - int LM = 3; - LM = optimize_framesize(analysis_pcm, frame_size, C, Fs, bitrate_bps, - 0, subframe_mem, delay_compensation, downmix); - while ((Fs/400<frame_size) - LM--; - frame_size = (Fs/400<max_follower))); } +static int decide_fec(int useInBandFEC, int PacketLoss_perc, int last_fec, int mode, int *bandwidth, opus_int32 rate) +{ + int orig_bandwidth; + if (!useInBandFEC || PacketLoss_perc == 0 || mode == MODE_CELT_ONLY) + return 0; + orig_bandwidth = *bandwidth; + for (;;) + { + opus_int32 hysteresis; + opus_int32 LBRR_rate_thres_bps; + /* Compute threshold for using FEC at the current bandwidth setting */ + LBRR_rate_thres_bps = fec_thresholds[2*(*bandwidth - OPUS_BANDWIDTH_NARROWBAND)]; + hysteresis = fec_thresholds[2*(*bandwidth - OPUS_BANDWIDTH_NARROWBAND) + 1]; + if (last_fec == 1) LBRR_rate_thres_bps -= hysteresis; + if (last_fec == 0) LBRR_rate_thres_bps += hysteresis; + LBRR_rate_thres_bps = silk_SMULWB( silk_MUL( LBRR_rate_thres_bps, + 125 - silk_min( PacketLoss_perc, 25 ) ), SILK_FIX_CONST( 0.01, 16 ) ); + /* If loss <= 5%, we look at whether we have enough rate to enable FEC. + If loss > 5%, we decrease the bandwidth until we can enable FEC. */ + if (rate > LBRR_rate_thres_bps) + return 1; + else if (PacketLoss_perc <= 5) + return 0; + else if (*bandwidth > OPUS_BANDWIDTH_NARROWBAND) + (*bandwidth)--; + else + break; + } + /* Couldn't find any bandwidth to enable FEC, keep original bandwidth. */ + *bandwidth = orig_bandwidth; + return 0; +} + +static int compute_silk_rate_for_hybrid(int rate, int bandwidth, int frame20ms, int vbr, int fec) { + int entry; + int i; + int N; + int silk_rate; + static int rate_table[][5] = { + /* |total| |-------- SILK------------| + |-- No FEC -| |--- FEC ---| + 10ms 20ms 10ms 20ms */ + { 0, 0, 0, 0, 0}, + {12000, 10000, 10000, 11000, 11000}, + {16000, 13500, 13500, 15000, 15000}, + {20000, 16000, 16000, 18000, 18000}, + {24000, 18000, 18000, 21000, 21000}, + {32000, 22000, 22000, 28000, 28000}, + {64000, 38000, 38000, 50000, 50000} + }; + entry = 1 + frame20ms + 2*fec; + N = sizeof(rate_table)/sizeof(rate_table[0]); + for (i=1;i rate) break; + } + if (i == N) + { + silk_rate = rate_table[i-1][entry]; + /* For now, just give 50% of the extra bits to SILK. */ + silk_rate += (rate-rate_table[i-1][0])/2; + } else { + opus_int32 lo, hi, x0, x1; + lo = rate_table[i-1][entry]; + hi = rate_table[i][entry]; + x0 = rate_table[i-1][0]; + x1 = rate_table[i][0]; + silk_rate = (lo*(x1-rate) + hi*(rate-x0))/(x1-x0); + } + if (!vbr) + { + /* Tiny boost to SILK for CBR. We should probably tune this better. */ + silk_rate += 100; + } + if (bandwidth==OPUS_BANDWIDTH_SUPERWIDEBAND) + silk_rate += 300; + return silk_rate; +} + +/* Returns the equivalent bitrate corresponding to 20 ms frames, + complexity 10 VBR operation. */ +static opus_int32 compute_equiv_rate(opus_int32 bitrate, int channels, + int frame_rate, int vbr, int mode, int complexity, int loss) +{ + opus_int32 equiv; + equiv = bitrate; + /* Take into account overhead from smaller frames. */ + equiv -= (40*channels+20)*(frame_rate - 50); + /* CBR is about a 8% penalty for both SILK and CELT. */ + if (!vbr) + equiv -= equiv/12; + /* Complexity makes about 10% difference (from 0 to 10) in general. */ + equiv = equiv * (90+complexity)/100; + if (mode == MODE_SILK_ONLY || mode == MODE_HYBRID) + { + /* SILK complexity 0-1 uses the non-delayed-decision NSQ, which + costs about 20%. */ + if (complexity<2) + equiv = equiv*4/5; + equiv -= equiv*loss/(6*loss + 10); + } else if (mode == MODE_CELT_ONLY) { + /* CELT complexity 0-4 doesn't have the pitch filter, which costs + about 10%. */ + if (complexity<5) + equiv = equiv*9/10; + } else { + /* Mode not known yet */ + /* Half the SILK loss*/ + equiv -= equiv*loss/(12*loss + 20); + } + return equiv; +} + +#ifndef DISABLE_FLOAT_API + +static int is_digital_silence(const opus_val16* pcm, int frame_size, int channels, int lsb_depth) +{ + int silence = 0; + opus_val32 sample_max = 0; +#ifdef MLP_TRAINING + return 0; +#endif + sample_max = celt_maxabs16(pcm, frame_size*channels); + +#ifdef FIXED_POINT + silence = (sample_max == 0); + (void)lsb_depth; +#else + silence = (sample_max <= (opus_val16) 1 / (1 << lsb_depth)); +#endif + + return silence; +} + +#ifdef FIXED_POINT +static opus_val32 compute_frame_energy(const opus_val16 *pcm, int frame_size, int channels, int arch) +{ + int i; + opus_val32 sample_max; + int max_shift; + int shift; + opus_val32 energy = 0; + int len = frame_size*channels; + (void)arch; + /* Max amplitude in the signal */ + sample_max = celt_maxabs16(pcm, len); + + /* Compute the right shift required in the MAC to avoid an overflow */ + max_shift = celt_ilog2(len); + shift = IMAX(0, (celt_ilog2(sample_max) << 1) + max_shift - 28); + + /* Compute the energy */ + for (i=0; i= (PSEUDO_SNR_THRESHOLD * noise_energy); + } + } + + if (is_silence || (is_noise && is_sufficiently_quiet)) + { + /* The number of consecutive DTX frames should be within the allowed bounds */ + (*nb_no_activity_frames)++; + + if (*nb_no_activity_frames > NB_SPEECH_FRAMES_BEFORE_DTX) + { + if (*nb_no_activity_frames <= (NB_SPEECH_FRAMES_BEFORE_DTX + MAX_CONSECUTIVE_DTX)) + /* Valid frame for DTX! */ + return 1; + else + (*nb_no_activity_frames) = NB_SPEECH_FRAMES_BEFORE_DTX; + } + } else + (*nb_no_activity_frames) = 0; + + return 0; +} + +#endif + +static opus_int32 encode_multiframe_packet(OpusEncoder *st, + const opus_val16 *pcm, + int nb_frames, + int frame_size, + unsigned char *data, + opus_int32 out_data_bytes, + int to_celt, + int lsb_depth, + int float_api) +{ + int i; + int ret = 0; + VARDECL(unsigned char, tmp_data); + int bak_mode, bak_bandwidth, bak_channels, bak_to_mono; + VARDECL(OpusRepacketizer, rp); + int max_header_bytes; + opus_int32 bytes_per_frame; + opus_int32 cbr_bytes; + opus_int32 repacketize_len; + int tmp_len; + ALLOC_STACK; + + /* Worst cases: + * 2 frames: Code 2 with different compressed sizes + * >2 frames: Code 3 VBR */ + max_header_bytes = nb_frames == 2 ? 3 : (2+(nb_frames-1)*2); + + if (st->use_vbr || st->user_bitrate_bps==OPUS_BITRATE_MAX) + repacketize_len = out_data_bytes; + else { + cbr_bytes = 3*st->bitrate_bps/(3*8*st->Fs/(frame_size*nb_frames)); + repacketize_len = IMIN(cbr_bytes, out_data_bytes); + } + bytes_per_frame = IMIN(1276, 1+(repacketize_len-max_header_bytes)/nb_frames); + + ALLOC(tmp_data, nb_frames*bytes_per_frame, unsigned char); + ALLOC(rp, 1, OpusRepacketizer); + opus_repacketizer_init(rp); + + bak_mode = st->user_forced_mode; + bak_bandwidth = st->user_bandwidth; + bak_channels = st->force_channels; + + st->user_forced_mode = st->mode; + st->user_bandwidth = st->bandwidth; + st->force_channels = st->stream_channels; + + bak_to_mono = st->silk_mode.toMono; + if (bak_to_mono) + st->force_channels = 1; + else + st->prev_channels = st->stream_channels; + + for (i=0;isilk_mode.toMono = 0; + st->nonfinal_frame = i<(nb_frames-1); + + /* When switching from SILK/Hybrid to CELT, only ask for a switch at the last frame */ + if (to_celt && i==nb_frames-1) + st->user_forced_mode = MODE_CELT_ONLY; + + tmp_len = opus_encode_native(st, pcm+i*(st->channels*frame_size), frame_size, + tmp_data+i*bytes_per_frame, bytes_per_frame, lsb_depth, NULL, 0, 0, 0, 0, + NULL, float_api); + + if (tmp_len<0) + { + RESTORE_STACK; + return OPUS_INTERNAL_ERROR; + } + + ret = opus_repacketizer_cat(rp, tmp_data+i*bytes_per_frame, tmp_len); + + if (ret<0) + { + RESTORE_STACK; + return OPUS_INTERNAL_ERROR; + } + } + + ret = opus_repacketizer_out_range_impl(rp, 0, nb_frames, data, repacketize_len, 0, !st->use_vbr); + + if (ret<0) + { + RESTORE_STACK; + return OPUS_INTERNAL_ERROR; + } + + /* Discard configs that were forced locally for the purpose of repacketization */ + st->user_forced_mode = bak_mode; + st->user_bandwidth = bak_bandwidth; + st->force_channels = bak_channels; + st->silk_mode.toMono = bak_to_mono; + + RESTORE_STACK; + return ret; +} + +static int compute_redundancy_bytes(opus_int32 max_data_bytes, opus_int32 bitrate_bps, int frame_rate, int channels) +{ + int redundancy_bytes_cap; + int redundancy_bytes; + opus_int32 redundancy_rate; + int base_bits; + opus_int32 available_bits; + base_bits = (40*channels+20); + + /* Equivalent rate for 5 ms frames. */ + redundancy_rate = bitrate_bps + base_bits*(200 - frame_rate); + /* For VBR, further increase the bitrate if we can afford it. It's pretty short + and we'll avoid artefacts. */ + redundancy_rate = 3*redundancy_rate/2; + redundancy_bytes = redundancy_rate/1600; + + /* Compute the max rate we can use given CBR or VBR with cap. */ + available_bits = max_data_bytes*8 - 2*base_bits; + redundancy_bytes_cap = (available_bits*240/(240+48000/frame_rate) + base_bits)/8; + redundancy_bytes = IMIN(redundancy_bytes, redundancy_bytes_cap); + /* It we can't get enough bits for redundancy to be worth it, rely on the decoder PLC. */ + if (redundancy_bytes > 4 + 8*channels) + redundancy_bytes = IMIN(257, redundancy_bytes); + else + redundancy_bytes = 0; + return redundancy_bytes; +} + opus_int32 opus_encode_native(OpusEncoder *st, const opus_val16 *pcm, int frame_size, unsigned char *data, opus_int32 out_data_bytes, int lsb_depth, const void *analysis_pcm, opus_int32 analysis_size, int c1, int c2, @@ -971,6 +1108,7 @@ opus_int32 opus_encode_native(OpusEncoder *st, const opus_val16 *pcm, int frame_ AnalysisInfo analysis_info; int analysis_read_pos_bak=-1; int analysis_read_subframe_bak=-1; + int is_silence = 0; #endif VARDECL(opus_val16, tmp_prefill); @@ -979,15 +1117,19 @@ opus_int32 opus_encode_native(OpusEncoder *st, const opus_val16 *pcm, int frame_ max_data_bytes = IMIN(1276, out_data_bytes); st->rangeFinal = 0; - if ((!st->variable_duration && 400*frame_size != st->Fs && 200*frame_size != st->Fs && 100*frame_size != st->Fs && - 50*frame_size != st->Fs && 25*frame_size != st->Fs && 50*frame_size != 3*st->Fs) - || (400*frame_size < st->Fs) - || max_data_bytes<=0 - ) + if (frame_size <= 0 || max_data_bytes <= 0) { RESTORE_STACK; return OPUS_BAD_ARG; } + + /* Cannot encode 100 ms in 1 byte */ + if (max_data_bytes==1 && st->Fs==(frame_size*10)) + { + RESTORE_STACK; + return OPUS_BUFFER_TOO_SMALL; + } + silk_enc = (char*)st+st->silk_enc_offset; celt_enc = (CELTEncoder*)((char*)st+st->celt_enc_offset); if (st->application == OPUS_APPLICATION_RESTRICTED_LOWDELAY) @@ -1001,25 +1143,42 @@ opus_int32 opus_encode_native(OpusEncoder *st, const opus_val16 *pcm, int frame_ #ifndef DISABLE_FLOAT_API analysis_info.valid = 0; #ifdef FIXED_POINT - if (st->silk_mode.complexity >= 10 && st->Fs==48000) + if (st->silk_mode.complexity >= 10 && st->Fs>=16000) #else - if (st->silk_mode.complexity >= 7 && st->Fs==48000) + if (st->silk_mode.complexity >= 7 && st->Fs>=16000) #endif { - analysis_read_pos_bak = st->analysis.read_pos; - analysis_read_subframe_bak = st->analysis.read_subframe; - run_analysis(&st->analysis, celt_mode, analysis_pcm, analysis_size, frame_size, - c1, c2, analysis_channels, st->Fs, - lsb_depth, downmix, &analysis_info); + if (is_digital_silence(pcm, frame_size, st->channels, lsb_depth)) + { + is_silence = 1; + } else { + analysis_read_pos_bak = st->analysis.read_pos; + analysis_read_subframe_bak = st->analysis.read_subframe; + run_analysis(&st->analysis, celt_mode, analysis_pcm, analysis_size, frame_size, + c1, c2, analysis_channels, st->Fs, + lsb_depth, downmix, &analysis_info); + } + + /* Track the peak signal energy */ + if (!is_silence && analysis_info.activity_probability > DTX_ACTIVITY_THRESHOLD) + st->peak_signal_energy = MAX32(MULT16_32_Q15(QCONST16(0.999f, 15), st->peak_signal_energy), + compute_frame_energy(pcm, frame_size, st->channels, st->arch)); } #else (void)analysis_pcm; (void)analysis_size; + (void)c1; + (void)c2; + (void)analysis_channels; + (void)downmix; #endif - st->voice_ratio = -1; - #ifndef DISABLE_FLOAT_API + /* Reset voice_ratio if this frame is not silent or if analysis is disabled. + * Otherwise, preserve voice_ratio from the last non-silent frame */ + if (!is_silence) + st->voice_ratio = -1; + st->detected_bandwidth = 0; if (analysis_info.valid) { @@ -1039,6 +1198,8 @@ opus_int32 opus_encode_native(OpusEncoder *st, const opus_val16 *pcm, int frame_ else st->detected_bandwidth = OPUS_BANDWIDTH_FULLBAND; } +#else + st->voice_ratio = -1; #endif if (st->channels==2 && st->force_channels!=1) @@ -1052,12 +1213,13 @@ opus_int32 opus_encode_native(OpusEncoder *st, const opus_val16 *pcm, int frame_ if (!st->use_vbr) { int cbrBytes; - /* Multiply by 3 to make sure the division is exact. */ - int frame_rate3 = 3*st->Fs/frame_size; + /* Multiply by 12 to make sure the division is exact. */ + int frame_rate12 = 12*st->Fs/frame_size; /* We need to make sure that "int" values always fit in 16 bits. */ - cbrBytes = IMIN( (3*st->bitrate_bps/8 + frame_rate3/2)/frame_rate3, max_data_bytes); - st->bitrate_bps = cbrBytes*(opus_int32)frame_rate3*8/3; - max_data_bytes = cbrBytes; + cbrBytes = IMIN( (12*st->bitrate_bps/8 + frame_rate12/2)/frame_rate12, max_data_bytes); + st->bitrate_bps = cbrBytes*(opus_int32)frame_rate12*8/12; + /* Make sure we provide at least one byte to avoid failing. */ + max_data_bytes = IMAX(1, cbrBytes); } if (max_data_bytes<3 || st->bitrate_bps < 3*frame_rate*8 || (frame_rate<50 && (max_data_bytes*frame_rate<300 || st->bitrate_bps < 2400))) @@ -1065,25 +1227,63 @@ opus_int32 opus_encode_native(OpusEncoder *st, const opus_val16 *pcm, int frame_ /*If the space is too low to do something useful, emit 'PLC' frames.*/ int tocmode = st->mode; int bw = st->bandwidth == 0 ? OPUS_BANDWIDTH_NARROWBAND : st->bandwidth; + int packet_code = 0; + int num_multiframes = 0; + if (tocmode==0) tocmode = MODE_SILK_ONLY; if (frame_rate>100) tocmode = MODE_CELT_ONLY; - if (frame_rate < 50) - tocmode = MODE_SILK_ONLY; + /* 40 ms -> 2 x 20 ms if in CELT_ONLY or HYBRID mode */ + if (frame_rate==25 && tocmode!=MODE_SILK_ONLY) + { + frame_rate = 50; + packet_code = 1; + } + + /* >= 60 ms frames */ + if (frame_rate<=16) + { + /* 1 x 60 ms, 2 x 40 ms, 2 x 60 ms */ + if (out_data_bytes==1 || (tocmode==MODE_SILK_ONLY && frame_rate!=10)) + { + tocmode = MODE_SILK_ONLY; + + packet_code = frame_rate <= 12; + frame_rate = frame_rate == 12 ? 25 : 16; + } + else + { + num_multiframes = 50/frame_rate; + frame_rate = 50; + packet_code = 3; + } + } + if(tocmode==MODE_SILK_ONLY&&bw>OPUS_BANDWIDTH_WIDEBAND) bw=OPUS_BANDWIDTH_WIDEBAND; else if (tocmode==MODE_CELT_ONLY&&bw==OPUS_BANDWIDTH_MEDIUMBAND) bw=OPUS_BANDWIDTH_NARROWBAND; else if (tocmode==MODE_HYBRID&&bw<=OPUS_BANDWIDTH_SUPERWIDEBAND) bw=OPUS_BANDWIDTH_SUPERWIDEBAND; + data[0] = gen_toc(tocmode, frame_rate, bw, st->stream_channels); - ret = 1; + data[0] |= packet_code; + + ret = packet_code <= 1 ? 1 : 2; + + max_data_bytes = IMAX(max_data_bytes, ret); + + if (packet_code==3) + data[1] = num_multiframes; + if (!st->use_vbr) { ret = opus_packet_pad(data, ret, max_data_bytes); if (ret == OPUS_OK) ret = max_data_bytes; + else + ret = OPUS_INTERNAL_ERROR; } RESTORE_STACK; return ret; @@ -1091,7 +1291,8 @@ opus_int32 opus_encode_native(OpusEncoder *st, const opus_val16 *pcm, int frame_ max_rate = frame_rate*max_data_bytes*8; /* Equivalent 20-ms rate for mode/channel/bandwidth decisions */ - equiv_rate = st->bitrate_bps - (40*st->channels+20)*(st->Fs/frame_size - 50); + equiv_rate = compute_equiv_rate(st->bitrate_bps, st->channels, st->Fs/frame_size, + st->use_vbr, 0, st->silk_mode.complexity, st->silk_mode.packetLossPercentage); if (st->signal_type == OPUS_SIGNAL_VOICE) voice_est = 127; @@ -1132,7 +1333,9 @@ opus_int32 opus_encode_native(OpusEncoder *st, const opus_val16 *pcm, int frame_ } #endif } - equiv_rate = st->bitrate_bps - (40*st->stream_channels+20)*(st->Fs/frame_size - 50); + /* Update equivalent rate for channels decision. */ + equiv_rate = compute_equiv_rate(st->bitrate_bps, st->stream_channels, st->Fs/frame_size, + st->use_vbr, 0, st->silk_mode.complexity, st->silk_mode.packetLossPercentage); /* Mode selection depending on application and signal type */ if (st->application == OPUS_APPLICATION_RESTRICTED_LOWDELAY) @@ -1181,10 +1384,21 @@ opus_int32 opus_encode_native(OpusEncoder *st, const opus_val16 *pcm, int frame_ /* When FEC is enabled and there's enough packet loss, use SILK */ if (st->silk_mode.useInBandFEC && st->silk_mode.packetLossPercentage > (128-voice_est)>>4) st->mode = MODE_SILK_ONLY; - /* When encoding voice and DTX is enabled, set the encoder to SILK mode (at least for now) */ + /* When encoding voice and DTX is enabled but the generalized DTX cannot be used, + because of complexity and sampling frequency settings, switch to SILK DTX and + set the encoder to SILK mode */ +#ifndef DISABLE_FLOAT_API + st->silk_mode.useDTX = st->use_dtx && !(analysis_info.valid || is_silence); +#else + st->silk_mode.useDTX = st->use_dtx; +#endif if (st->silk_mode.useDTX && voice_est > 100) st->mode = MODE_SILK_ONLY; #endif + + /* If max_data_bytes represents less than 6 kb/s, switch to CELT-only mode */ + if (max_data_bytes < (frame_rate > 50 ? 9000 : 6000)*frame_size / (st->Fs * 8)) + st->mode = MODE_CELT_ONLY; } else { st->mode = st->user_forced_mode; } @@ -1194,19 +1408,6 @@ opus_int32 opus_encode_native(OpusEncoder *st, const opus_val16 *pcm, int frame_ st->mode = MODE_CELT_ONLY; if (st->lfe) st->mode = MODE_CELT_ONLY; - /* If max_data_bytes represents less than 8 kb/s, switch to CELT-only mode */ - if (max_data_bytes < (frame_rate > 50 ? 12000 : 8000)*frame_size / (st->Fs * 8)) - st->mode = MODE_CELT_ONLY; - - if (st->stream_channels == 1 && st->prev_channels ==2 && st->silk_mode.toMono==0 - && st->mode != MODE_CELT_ONLY && st->prev_mode != MODE_CELT_ONLY) - { - /* Delay stereo->mono transition by two frames so that SILK can do a smooth downmix */ - st->silk_mode.toMono = 1; - st->stream_channels = 2; - } else { - st->silk_mode.toMono = 0; - } if (st->prev_mode > 0 && ((st->mode != MODE_CELT_ONLY && st->prev_mode == MODE_CELT_ONLY) || @@ -1226,23 +1427,22 @@ opus_int32 opus_encode_native(OpusEncoder *st, const opus_val16 *pcm, int frame_ } } } - /* For the first frame at a new SILK bandwidth */ - if (st->silk_bw_switch) + + /* When encoding multiframes, we can ask for a switch to CELT only in the last frame. This switch + * is processed above as the requested mode shouldn't interrupt stereo->mono transition. */ + if (st->stream_channels == 1 && st->prev_channels ==2 && st->silk_mode.toMono==0 + && st->mode != MODE_CELT_ONLY && st->prev_mode != MODE_CELT_ONLY) { - redundancy = 1; - celt_to_silk = 1; - st->silk_bw_switch = 0; - prefill=1; + /* Delay stereo->mono transition by two frames so that SILK can do a smooth downmix */ + st->silk_mode.toMono = 1; + st->stream_channels = 2; + } else { + st->silk_mode.toMono = 0; } - if (redundancy) - { - /* Fair share of the max size allowed */ - redundancy_bytes = IMIN(257, max_data_bytes*(opus_int32)(st->Fs/200)/(frame_size+st->Fs/200)); - /* For VBR, target the actual bitrate (subject to the limit above) */ - if (st->use_vbr) - redundancy_bytes = IMIN(redundancy_bytes, st->bitrate_bps/1600); - } + /* Update equivalent rate with mode decision. */ + equiv_rate = compute_equiv_rate(st->bitrate_bps, st->stream_channels, st->Fs/frame_size, + st->use_vbr, st->mode, st->silk_mode.complexity, st->silk_mode.packetLossPercentage); if (st->mode != MODE_CELT_ONLY && st->prev_mode == MODE_CELT_ONLY) { @@ -1257,17 +1457,7 @@ opus_int32 opus_encode_native(OpusEncoder *st, const opus_val16 *pcm, int frame_ const opus_int32 *voice_bandwidth_thresholds, *music_bandwidth_thresholds; opus_int32 bandwidth_thresholds[8]; int bandwidth = OPUS_BANDWIDTH_FULLBAND; - opus_int32 equiv_rate2; - equiv_rate2 = equiv_rate; - if (st->mode != MODE_CELT_ONLY) - { - /* Adjust the threshold +/- 10% depending on complexity */ - equiv_rate2 = equiv_rate2 * (45+st->silk_mode.complexity)/50; - /* CBR is less efficient by ~1 kb/s */ - if (!st->use_vbr) - equiv_rate2 -= 1000; - } if (st->channels==2 && st->force_channels!=1) { voice_bandwidth_thresholds = stereo_voice_bandwidth_thresholds; @@ -1288,15 +1478,15 @@ opus_int32 opus_encode_native(OpusEncoder *st, const opus_val16 *pcm, int frame_ hysteresis = bandwidth_thresholds[2*(bandwidth-OPUS_BANDWIDTH_MEDIUMBAND)+1]; if (!st->first) { - if (st->bandwidth >= bandwidth) + if (st->auto_bandwidth >= bandwidth) threshold -= hysteresis; else threshold += hysteresis; } - if (equiv_rate2 >= threshold) + if (equiv_rate >= threshold) break; } while (--bandwidth>OPUS_BANDWIDTH_NARROWBAND); - st->bandwidth = bandwidth; + st->bandwidth = st->auto_bandwidth = bandwidth; /* Prevents any transition to SWB/FB until the SILK layer has fully switched to WB mode and turned the variable LP filter off */ if (!st->first && st->mode != MODE_CELT_ONLY && !st->silk_mode.inWBmodeWithoutVariableLP && st->bandwidth > OPUS_BANDWIDTH_WIDEBAND) @@ -1349,6 +1539,8 @@ opus_int32 opus_encode_native(OpusEncoder *st, const opus_val16 *pcm, int frame_ st->bandwidth = IMIN(st->bandwidth, st->detected_bandwidth); } #endif + st->silk_mode.LBRR_coded = decide_fec(st->silk_mode.useInBandFEC, st->silk_mode.packetLossPercentage, + st->silk_mode.LBRR_coded, st->mode, &st->bandwidth, equiv_rate); celt_encoder_ctl(celt_enc, OPUS_SET_LSB_DEPTH(lsb_depth)); /* CELT mode doesn't support mediumband, use wideband instead */ @@ -1357,15 +1549,34 @@ opus_int32 opus_encode_native(OpusEncoder *st, const opus_val16 *pcm, int frame_ if (st->lfe) st->bandwidth = OPUS_BANDWIDTH_NARROWBAND; - /* Can't support higher than wideband for >20 ms frames */ - if (frame_size > st->Fs/50 && (st->mode == MODE_CELT_ONLY || st->bandwidth > OPUS_BANDWIDTH_WIDEBAND)) + curr_bandwidth = st->bandwidth; + + /* Chooses the appropriate mode for speech + *NEVER* switch to/from CELT-only mode here as this will invalidate some assumptions */ + if (st->mode == MODE_SILK_ONLY && curr_bandwidth > OPUS_BANDWIDTH_WIDEBAND) + st->mode = MODE_HYBRID; + if (st->mode == MODE_HYBRID && curr_bandwidth <= OPUS_BANDWIDTH_WIDEBAND) + st->mode = MODE_SILK_ONLY; + + /* Can't support higher than >60 ms frames, and >20 ms when in Hybrid or CELT-only modes */ + if ((frame_size > st->Fs/50 && (st->mode != MODE_SILK_ONLY)) || frame_size > 3*st->Fs/50) { - VARDECL(unsigned char, tmp_data); + int enc_frame_size; int nb_frames; - int bak_mode, bak_bandwidth, bak_channels, bak_to_mono; - VARDECL(OpusRepacketizer, rp); - opus_int32 bytes_per_frame; - opus_int32 repacketize_len; + + if (st->mode == MODE_SILK_ONLY) + { + if (frame_size == 2*st->Fs/25) /* 80 ms -> 2x 40 ms */ + enc_frame_size = st->Fs/25; + else if (frame_size == 3*st->Fs/25) /* 120 ms -> 2x 60 ms */ + enc_frame_size = 3*st->Fs/50; + else /* 100 ms -> 5x 20 ms */ + enc_frame_size = st->Fs/50; + } + else + enc_frame_size = st->Fs/50; + + nb_frames = frame_size/enc_frame_size; #ifndef DISABLE_FLOAT_API if (analysis_read_pos_bak!= -1) @@ -1375,74 +1586,33 @@ opus_int32 opus_encode_native(OpusEncoder *st, const opus_val16 *pcm, int frame_ } #endif - nb_frames = frame_size > st->Fs/25 ? 3 : 2; - bytes_per_frame = IMIN(1276,(out_data_bytes-3)/nb_frames); + ret = encode_multiframe_packet(st, pcm, nb_frames, enc_frame_size, data, + out_data_bytes, to_celt, lsb_depth, float_api); - ALLOC(tmp_data, nb_frames*bytes_per_frame, unsigned char); - - ALLOC(rp, 1, OpusRepacketizer); - opus_repacketizer_init(rp); - - bak_mode = st->user_forced_mode; - bak_bandwidth = st->user_bandwidth; - bak_channels = st->force_channels; - - st->user_forced_mode = st->mode; - st->user_bandwidth = st->bandwidth; - st->force_channels = st->stream_channels; - bak_to_mono = st->silk_mode.toMono; - - if (bak_to_mono) - st->force_channels = 1; - else - st->prev_channels = st->stream_channels; - for (i=0;isilk_mode.toMono = 0; - /* When switching from SILK/Hybrid to CELT, only ask for a switch at the last frame */ - if (to_celt && i==nb_frames-1) - st->user_forced_mode = MODE_CELT_ONLY; - tmp_len = opus_encode_native(st, pcm+i*(st->channels*st->Fs/50), st->Fs/50, - tmp_data+i*bytes_per_frame, bytes_per_frame, lsb_depth, - NULL, 0, c1, c2, analysis_channels, downmix, float_api); - if (tmp_len<0) - { - RESTORE_STACK; - return OPUS_INTERNAL_ERROR; - } - ret = opus_repacketizer_cat(rp, tmp_data+i*bytes_per_frame, tmp_len); - if (ret<0) - { - RESTORE_STACK; - return OPUS_INTERNAL_ERROR; - } - } - if (st->use_vbr) - repacketize_len = out_data_bytes; - else - repacketize_len = IMIN(3*st->bitrate_bps/(3*8*50/nb_frames), out_data_bytes); - ret = opus_repacketizer_out_range_impl(rp, 0, nb_frames, data, repacketize_len, 0, !st->use_vbr); - if (ret<0) - { - RESTORE_STACK; - return OPUS_INTERNAL_ERROR; - } - st->user_forced_mode = bak_mode; - st->user_bandwidth = bak_bandwidth; - st->force_channels = bak_channels; - st->silk_mode.toMono = bak_to_mono; RESTORE_STACK; return ret; } - curr_bandwidth = st->bandwidth; - /* Chooses the appropriate mode for speech - *NEVER* switch to/from CELT-only mode here as this will invalidate some assumptions */ - if (st->mode == MODE_SILK_ONLY && curr_bandwidth > OPUS_BANDWIDTH_WIDEBAND) - st->mode = MODE_HYBRID; - if (st->mode == MODE_HYBRID && curr_bandwidth <= OPUS_BANDWIDTH_WIDEBAND) - st->mode = MODE_SILK_ONLY; + /* For the first frame at a new SILK bandwidth */ + if (st->silk_bw_switch) + { + redundancy = 1; + celt_to_silk = 1; + st->silk_bw_switch = 0; + prefill=1; + } + + /* If we decided to go with CELT, make sure redundancy is off, no matter what + we decided earlier. */ + if (st->mode == MODE_CELT_ONLY) + redundancy = 0; + + if (redundancy) + { + redundancy_bytes = compute_redundancy_bytes(max_data_bytes, st->bitrate_bps, frame_rate, st->stream_channels); + if (redundancy_bytes == 0) + redundancy = 0; + } /* printf("%d %d %d %d\n", st->bitrate_bps, st->stream_channels, st->mode, curr_bandwidth); */ bytes_target = IMIN(max_data_bytes-redundancy_bytes, st->bitrate_bps * frame_size / (st->Fs * 8)) - 1; @@ -1467,7 +1637,7 @@ opus_int32 opus_encode_native(OpusEncoder *st, const opus_val16 *pcm, int frame_ if (st->application == OPUS_APPLICATION_VOIP) { - hp_cutoff(pcm, cutoff_Hz, &pcm_buf[total_buffer*st->channels], st->hp_mem, frame_size, st->channels, st->Fs); + hp_cutoff(pcm, cutoff_Hz, &pcm_buf[total_buffer*st->channels], st->hp_mem, frame_size, st->channels, st->Fs, st->arch); } else { dc_reject(pcm, 3, &pcm_buf[total_buffer*st->channels], st->hp_mem, frame_size, st->channels, st->Fs); } @@ -1502,27 +1672,14 @@ opus_int32 opus_encode_native(OpusEncoder *st, const opus_val16 *pcm, int frame_ /* Distribute bits between SILK and CELT */ total_bitRate = 8 * bytes_target * frame_rate; if( st->mode == MODE_HYBRID ) { - int HB_gain_ref; /* Base rate for SILK */ - st->silk_mode.bitRate = st->stream_channels * ( 5000 + 1000 * ( st->Fs == 100 * frame_size ) ); - if( curr_bandwidth == OPUS_BANDWIDTH_SUPERWIDEBAND ) { - /* SILK gets 2/3 of the remaining bits */ - st->silk_mode.bitRate += ( total_bitRate - st->silk_mode.bitRate ) * 2 / 3; - } else { /* FULLBAND */ - /* SILK gets 3/5 of the remaining bits */ - st->silk_mode.bitRate += ( total_bitRate - st->silk_mode.bitRate ) * 3 / 5; - } - /* Don't let SILK use more than 80% */ - if( st->silk_mode.bitRate > total_bitRate * 4/5 ) { - st->silk_mode.bitRate = total_bitRate * 4/5; - } + st->silk_mode.bitRate = compute_silk_rate_for_hybrid(total_bitRate, + curr_bandwidth, st->Fs == 50 * frame_size, st->use_vbr, st->silk_mode.LBRR_coded); if (!st->energy_masking) { /* Increasingly attenuate high band when it gets allocated fewer bits */ celt_rate = total_bitRate - st->silk_mode.bitRate; - HB_gain_ref = (curr_bandwidth == OPUS_BANDWIDTH_SUPERWIDEBAND) ? 3000 : 3600; - HB_gain = SHL32((opus_val32)celt_rate, 9) / SHR32((opus_val32)celt_rate + st->stream_channels * HB_gain_ref, 6); - HB_gain = HB_gain < (opus_val32)Q15ONE*6/7 ? HB_gain + Q15ONE/7 : Q15ONE; + HB_gain = Q15ONE - SHR32(celt_exp2(-celt_rate * QCONST16(1.f/1024, 10)), 1); } } else { /* SILK gets all bits */ @@ -1569,7 +1726,6 @@ opus_int32 opus_encode_native(OpusEncoder *st, const opus_val16 *pcm, int frame_ st->silk_mode.bitRate += 3*rate_offset/5; else st->silk_mode.bitRate += rate_offset; - bytes_target += rate_offset * frame_size / (8 * st->Fs); } st->silk_mode.payloadSize_ms = 1000 * frame_size / st->Fs; @@ -1590,40 +1746,52 @@ opus_int32 opus_encode_native(OpusEncoder *st, const opus_val16 *pcm, int frame_ st->silk_mode.minInternalSampleRate = 8000; } + st->silk_mode.maxInternalSampleRate = 16000; if (st->mode == MODE_SILK_ONLY) { opus_int32 effective_max_rate = max_rate; - st->silk_mode.maxInternalSampleRate = 16000; if (frame_rate > 50) effective_max_rate = effective_max_rate*2/3; - if (effective_max_rate < 13000) + if (effective_max_rate < 8000) { st->silk_mode.maxInternalSampleRate = 12000; st->silk_mode.desiredInternalSampleRate = IMIN(12000, st->silk_mode.desiredInternalSampleRate); } - if (effective_max_rate < 9600) + if (effective_max_rate < 7000) { st->silk_mode.maxInternalSampleRate = 8000; st->silk_mode.desiredInternalSampleRate = IMIN(8000, st->silk_mode.desiredInternalSampleRate); } - } else { - st->silk_mode.maxInternalSampleRate = 16000; } st->silk_mode.useCBR = !st->use_vbr; /* Call SILK encoder for the low band */ - nBytes = IMIN(1275, max_data_bytes-1-redundancy_bytes); - st->silk_mode.maxBits = nBytes*8; - /* Only allow up to 90% of the bits for hybrid mode*/ - if (st->mode == MODE_HYBRID) - st->silk_mode.maxBits = (opus_int32)st->silk_mode.maxBits*9/10; + /* Max bits for SILK, counting ToC, redundancy bytes, and optionally redundancy. */ + st->silk_mode.maxBits = (max_data_bytes-1)*8; + if (redundancy && redundancy_bytes >= 2) + { + /* Counting 1 bit for redundancy position and 20 bits for flag+size (only for hybrid). */ + st->silk_mode.maxBits -= redundancy_bytes*8 + 1; + if (st->mode == MODE_HYBRID) + st->silk_mode.maxBits -= 20; + } if (st->silk_mode.useCBR) { - st->silk_mode.maxBits = (st->silk_mode.bitRate * frame_size / (st->Fs * 8))*8; - /* Reduce the initial target to make it easier to reach the CBR rate */ - st->silk_mode.bitRate = IMAX(1, st->silk_mode.bitRate-2000); + if (st->mode == MODE_HYBRID) + { + st->silk_mode.maxBits = IMIN(st->silk_mode.maxBits, st->silk_mode.bitRate * frame_size / st->Fs); + } + } else { + /* Constrained VBR. */ + if (st->mode == MODE_HYBRID) + { + /* Compute SILK bitrate corresponding to the max total bits available */ + opus_int32 maxBitRate = compute_silk_rate_for_hybrid(st->silk_mode.maxBits*st->Fs / frame_size, + curr_bandwidth, st->Fs == 50 * frame_size, st->use_vbr, st->silk_mode.LBRR_coded); + st->silk_mode.maxBits = maxBitRate * frame_size / st->Fs; + } } if (prefill) @@ -1662,13 +1830,7 @@ opus_int32 opus_encode_native(OpusEncoder *st, const opus_val16 *pcm, int frame_ RESTORE_STACK; return OPUS_INTERNAL_ERROR; } - if (nBytes==0) - { - st->rangeFinal = 0; - data[-1] = gen_toc(st->mode, st->Fs/frame_size, curr_bandwidth, st->stream_channels); - RESTORE_STACK; - return 1; - } + /* Extract SILK internal bandwidth for signaling in first byte */ if( st->mode == MODE_SILK_ONLY ) { if( st->silk_mode.internalSampleRate == 8000 ) { @@ -1682,11 +1844,21 @@ opus_int32 opus_encode_native(OpusEncoder *st, const opus_val16 *pcm, int frame_ silk_assert( st->silk_mode.internalSampleRate == 16000 ); } - st->silk_mode.opusCanSwitch = st->silk_mode.switchReady; + st->silk_mode.opusCanSwitch = st->silk_mode.switchReady && !st->nonfinal_frame; + + if (nBytes==0) + { + st->rangeFinal = 0; + data[-1] = gen_toc(st->mode, st->Fs/frame_size, curr_bandwidth, st->stream_channels); + RESTORE_STACK; + return 1; + } + /* FIXME: How do we allocate the redundancy for CBR? */ if (st->silk_mode.opusCanSwitch) { - redundancy = 1; + redundancy_bytes = compute_redundancy_bytes(max_data_bytes, st->bitrate_bps, frame_rate, st->stream_channels); + redundancy = (redundancy_bytes != 0); celt_to_silk = 0; st->silk_bw_switch = 1; } @@ -1727,40 +1899,18 @@ opus_int32 opus_encode_native(OpusEncoder *st, const opus_val16 *pcm, int frame_ if (st->mode == MODE_HYBRID) { - int len; - - len = (ec_tell(&enc)+7)>>3; - if (redundancy) - len += st->mode == MODE_HYBRID ? 3 : 1; if( st->use_vbr ) { - nb_compr_bytes = len + bytes_target - (st->silk_mode.bitRate * frame_size) / (8 * st->Fs); - } else { - /* check if SILK used up too much */ - nb_compr_bytes = len > bytes_target ? len : bytes_target; + celt_encoder_ctl(celt_enc, OPUS_SET_BITRATE(st->bitrate_bps-st->silk_mode.bitRate)); + celt_encoder_ctl(celt_enc, OPUS_SET_VBR_CONSTRAINT(0)); } } else { if (st->use_vbr) { - opus_int32 bonus=0; -#ifndef DISABLE_FLOAT_API - if (st->variable_duration==OPUS_FRAMESIZE_VARIABLE && frame_size != st->Fs/50) - { - bonus = (60*st->stream_channels+40)*(st->Fs/frame_size-50); - if (analysis_info.valid) - bonus = (opus_int32)(bonus*(1.f+.5f*analysis_info.tonality)); - } -#endif celt_encoder_ctl(celt_enc, OPUS_SET_VBR(1)); celt_encoder_ctl(celt_enc, OPUS_SET_VBR_CONSTRAINT(st->vbr_constraint)); - celt_encoder_ctl(celt_enc, OPUS_SET_BITRATE(st->bitrate_bps+bonus)); - nb_compr_bytes = max_data_bytes-1-redundancy_bytes; - } else { - nb_compr_bytes = bytes_target; + celt_encoder_ctl(celt_enc, OPUS_SET_BITRATE(st->bitrate_bps)); } } - - } else { - nb_compr_bytes = 0; } ALLOC(tmp_prefill, st->channels*st->Fs/400, opus_val16); @@ -1786,7 +1936,7 @@ opus_int32 opus_encode_native(OpusEncoder *st, const opus_val16 *pcm, int frame_ } st->prev_HB_gain = HB_gain; if (st->mode != MODE_HYBRID || st->stream_channels==1) - st->silk_mode.stereoWidth_Q14 = IMIN((1<<14),2*IMAX(0,equiv_rate-30000)); + st->silk_mode.stereoWidth_Q14 = IMIN((1<<14),2*IMAX(0,equiv_rate-24000)); if( !st->energy_masking && st->channels == 2 ) { /* Apply stereo width reduction (at low bitrates) */ if( st->hybrid_stereo_width_Q14 < (1 << 14) || st->silk_mode.stereoWidth_Q14 < (1 << 14) ) { @@ -1809,19 +1959,23 @@ opus_int32 opus_encode_native(OpusEncoder *st, const opus_val16 *pcm, int frame_ if ( st->mode != MODE_CELT_ONLY && ec_tell(&enc)+17+20*(st->mode == MODE_HYBRID) <= 8*(max_data_bytes-1)) { /* For SILK mode, the redundancy is inferred from the length */ - if (st->mode == MODE_HYBRID && (redundancy || ec_tell(&enc)+37 <= 8*nb_compr_bytes)) + if (st->mode == MODE_HYBRID) ec_enc_bit_logp(&enc, redundancy, 12); if (redundancy) { int max_redundancy; ec_enc_bit_logp(&enc, celt_to_silk, 1); if (st->mode == MODE_HYBRID) - max_redundancy = (max_data_bytes-1)-nb_compr_bytes; + { + /* Reserve the 8 bits needed for the redundancy length, + and at least a few bits for CELT if possible */ + max_redundancy = (max_data_bytes-1)-((ec_tell(&enc)+8+3+7)>>3); + } else max_redundancy = (max_data_bytes-1)-((ec_tell(&enc)+7)>>3); /* Target the same bit-rate for redundancy as for the rest, up to a max of 257 bytes */ - redundancy_bytes = IMIN(max_redundancy, st->bitrate_bps/1600); + redundancy_bytes = IMIN(max_redundancy, redundancy_bytes); redundancy_bytes = IMIN(257, IMAX(2, redundancy_bytes)); if (st->mode == MODE_HYBRID) ec_enc_uint(&enc, redundancy_bytes-2, 256); @@ -1843,7 +1997,7 @@ opus_int32 opus_encode_native(OpusEncoder *st, const opus_val16 *pcm, int frame_ ec_enc_done(&enc); nb_compr_bytes = ret; } else { - nb_compr_bytes = IMIN((max_data_bytes-1)-redundancy_bytes, nb_compr_bytes); + nb_compr_bytes = (max_data_bytes-1)-redundancy_bytes; ec_enc_shrink(&enc, nb_compr_bytes); } @@ -1851,6 +2005,14 @@ opus_int32 opus_encode_native(OpusEncoder *st, const opus_val16 *pcm, int frame_ if (redundancy || st->mode != MODE_SILK_ONLY) celt_encoder_ctl(celt_enc, CELT_SET_ANALYSIS(&analysis_info)); #endif + if (st->mode == MODE_HYBRID) { + SILKInfo info; + info.signalType = st->silk_mode.signalType; + info.offset = st->silk_mode.offset; + celt_encoder_ctl(celt_enc, CELT_SET_SILK_INFO(&info)); + } else { + celt_encoder_ctl(celt_enc, CELT_SET_SILK_INFO((SILKInfo*)NULL)); + } /* 5 ms redundant frame for CELT->SILK */ if (redundancy && celt_to_silk) @@ -1858,6 +2020,7 @@ opus_int32 opus_encode_native(OpusEncoder *st, const opus_val16 *pcm, int frame_ int err; celt_encoder_ctl(celt_enc, CELT_SET_START_BAND(0)); celt_encoder_ctl(celt_enc, OPUS_SET_VBR(0)); + celt_encoder_ctl(celt_enc, OPUS_SET_BITRATE(OPUS_BITRATE_MAX)); err = celt_encode_with_ec(celt_enc, pcm_buf, st->Fs/200, data+nb_compr_bytes, redundancy_bytes, NULL); if (err < 0) { @@ -1881,15 +2044,25 @@ opus_int32 opus_encode_native(OpusEncoder *st, const opus_val16 *pcm, int frame_ celt_encode_with_ec(celt_enc, tmp_prefill, st->Fs/400, dummy, 2, NULL); celt_encoder_ctl(celt_enc, CELT_SET_PREDICTION(0)); } - /* If false, we already busted the budget and we'll end up with a "PLC packet" */ + /* If false, we already busted the budget and we'll end up with a "PLC frame" */ if (ec_tell(&enc) <= 8*nb_compr_bytes) { + /* Set the bitrate again if it was overridden in the redundancy code above*/ + if (redundancy && celt_to_silk && st->mode==MODE_HYBRID && st->use_vbr) + celt_encoder_ctl(celt_enc, OPUS_SET_BITRATE(st->bitrate_bps-st->silk_mode.bitRate)); + celt_encoder_ctl(celt_enc, OPUS_SET_VBR(st->use_vbr)); ret = celt_encode_with_ec(celt_enc, pcm_buf, frame_size, NULL, nb_compr_bytes, &enc); if (ret < 0) { RESTORE_STACK; return OPUS_INTERNAL_ERROR; } + /* Put CELT->SILK redundancy data in the right place. */ + if (redundancy && celt_to_silk && st->mode==MODE_HYBRID && st->use_vbr) + { + OPUS_MOVE(data+ret, data+nb_compr_bytes, redundancy_bytes); + nb_compr_bytes = nb_compr_bytes+redundancy_bytes; + } } } @@ -1905,7 +2078,15 @@ opus_int32 opus_encode_native(OpusEncoder *st, const opus_val16 *pcm, int frame_ celt_encoder_ctl(celt_enc, OPUS_RESET_STATE); celt_encoder_ctl(celt_enc, CELT_SET_START_BAND(0)); celt_encoder_ctl(celt_enc, CELT_SET_PREDICTION(0)); + celt_encoder_ctl(celt_enc, OPUS_SET_VBR(0)); + celt_encoder_ctl(celt_enc, OPUS_SET_BITRATE(OPUS_BITRATE_MAX)); + if (st->mode == MODE_HYBRID) + { + /* Shrink packet to what the encoder actually used. */ + nb_compr_bytes = ret; + ec_enc_shrink(&enc, nb_compr_bytes); + } /* NOTE: We could speed this up slightly (at the expense of code size) by just adding a function that prefills the buffer */ celt_encode_with_ec(celt_enc, pcm_buf+st->channels*(frame_size-N2-N4), N4, dummy, 2, NULL); @@ -1935,6 +2116,21 @@ opus_int32 opus_encode_native(OpusEncoder *st, const opus_val16 *pcm, int frame_ st->first = 0; + /* DTX decision */ +#ifndef DISABLE_FLOAT_API + if (st->use_dtx && (analysis_info.valid || is_silence)) + { + if (decide_dtx_mode(analysis_info.activity_probability, &st->nb_no_activity_frames, + st->peak_signal_energy, pcm, frame_size, st->channels, is_silence, st->arch)) + { + st->rangeFinal = 0; + data[0] = gen_toc(st->mode, st->Fs/frame_size, curr_bandwidth, st->stream_channels); + RESTORE_STACK; + return 1; + } + } +#endif + /* In the unlikely case that the SILK encoder busted its target, tell the decoder to call the PLC */ if (ec_tell(&enc) > (max_data_bytes-1)*8) @@ -1962,7 +2158,6 @@ opus_int32 opus_encode_native(OpusEncoder *st, const opus_val16 *pcm, int frame_ if (!st->use_vbr) { if (opus_packet_pad(data, ret, max_data_bytes) != OPUS_OK) - { RESTORE_STACK; return OPUS_INTERNAL_ERROR; @@ -1981,18 +2176,15 @@ opus_int32 opus_encode_float(OpusEncoder *st, const float *pcm, int analysis_fra { int i, ret; int frame_size; - int delay_compensation; VARDECL(opus_int16, in); ALLOC_STACK; - if (st->application == OPUS_APPLICATION_RESTRICTED_LOWDELAY) - delay_compensation = 0; - else - delay_compensation = st->delay_compensation; - frame_size = compute_frame_size(pcm, analysis_frame_size, - st->variable_duration, st->channels, st->Fs, st->bitrate_bps, - delay_compensation, downmix_float, st->analysis.subframe_mem); - + frame_size = frame_size_select(analysis_frame_size, st->variable_duration, st->Fs); + if (frame_size <= 0) + { + RESTORE_STACK; + return OPUS_BAD_ARG; + } ALLOC(in, frame_size*st->channels, opus_int16); for (i=0;ichannels;i++) @@ -2008,18 +2200,7 @@ opus_int32 opus_encode(OpusEncoder *st, const opus_int16 *pcm, int analysis_fram unsigned char *data, opus_int32 out_data_bytes) { int frame_size; - int delay_compensation; - if (st->application == OPUS_APPLICATION_RESTRICTED_LOWDELAY) - delay_compensation = 0; - else - delay_compensation = st->delay_compensation; - frame_size = compute_frame_size(pcm, analysis_frame_size, - st->variable_duration, st->channels, st->Fs, st->bitrate_bps, - delay_compensation, downmix_int -#ifndef DISABLE_FLOAT_API - , st->analysis.subframe_mem -#endif - ); + frame_size = frame_size_select(analysis_frame_size, st->variable_duration, st->Fs); return opus_encode_native(st, pcm, frame_size, data, out_data_bytes, 16, pcm, analysis_frame_size, 0, -2, st->channels, downmix_int, 0); } @@ -2030,18 +2211,15 @@ opus_int32 opus_encode(OpusEncoder *st, const opus_int16 *pcm, int analysis_fram { int i, ret; int frame_size; - int delay_compensation; VARDECL(float, in); ALLOC_STACK; - if (st->application == OPUS_APPLICATION_RESTRICTED_LOWDELAY) - delay_compensation = 0; - else - delay_compensation = st->delay_compensation; - frame_size = compute_frame_size(pcm, analysis_frame_size, - st->variable_duration, st->channels, st->Fs, st->bitrate_bps, - delay_compensation, downmix_int, st->analysis.subframe_mem); - + frame_size = frame_size_select(analysis_frame_size, st->variable_duration, st->Fs); + if (frame_size <= 0) + { + RESTORE_STACK; + return OPUS_BAD_ARG; + } ALLOC(in, frame_size*st->channels, float); for (i=0;ichannels;i++) @@ -2055,14 +2233,7 @@ opus_int32 opus_encode_float(OpusEncoder *st, const float *pcm, int analysis_fra unsigned char *data, opus_int32 out_data_bytes) { int frame_size; - int delay_compensation; - if (st->application == OPUS_APPLICATION_RESTRICTED_LOWDELAY) - delay_compensation = 0; - else - delay_compensation = st->delay_compensation; - frame_size = compute_frame_size(pcm, analysis_frame_size, - st->variable_duration, st->channels, st->Fs, st->bitrate_bps, - delay_compensation, downmix_float, st->analysis.subframe_mem); + frame_size = frame_size_select(analysis_frame_size, st->variable_duration, st->Fs); return opus_encode_native(st, pcm, frame_size, data, out_data_bytes, 24, pcm, analysis_frame_size, 0, -2, st->channels, downmix_float, 1); } @@ -2093,6 +2264,9 @@ int opus_encoder_ctl(OpusEncoder *st, int request, ...) break; } st->application = value; +#ifndef DISABLE_FLOAT_API + st->analysis.application = value; +#endif } break; case OPUS_GET_APPLICATION_REQUEST: @@ -2211,7 +2385,7 @@ int opus_encoder_ctl(OpusEncoder *st, int request, ...) { goto bad_arg; } - st->silk_mode.useDTX = value; + st->use_dtx = value; } break; case OPUS_GET_DTX_REQUEST: @@ -2221,7 +2395,7 @@ int opus_encoder_ctl(OpusEncoder *st, int request, ...) { goto bad_arg; } - *value = st->silk_mode.useDTX; + *value = st->use_dtx; } break; case OPUS_SET_COMPLEXITY_REQUEST: @@ -2422,10 +2596,11 @@ int opus_encoder_ctl(OpusEncoder *st, int request, ...) case OPUS_SET_EXPERT_FRAME_DURATION_REQUEST: { opus_int32 value = va_arg(ap, opus_int32); - if (value != OPUS_FRAMESIZE_ARG && value != OPUS_FRAMESIZE_2_5_MS && - value != OPUS_FRAMESIZE_5_MS && value != OPUS_FRAMESIZE_10_MS && - value != OPUS_FRAMESIZE_20_MS && value != OPUS_FRAMESIZE_40_MS && - value != OPUS_FRAMESIZE_60_MS && value != OPUS_FRAMESIZE_VARIABLE) + if (value != OPUS_FRAMESIZE_ARG && value != OPUS_FRAMESIZE_2_5_MS && + value != OPUS_FRAMESIZE_5_MS && value != OPUS_FRAMESIZE_10_MS && + value != OPUS_FRAMESIZE_20_MS && value != OPUS_FRAMESIZE_40_MS && + value != OPUS_FRAMESIZE_60_MS && value != OPUS_FRAMESIZE_80_MS && + value != OPUS_FRAMESIZE_100_MS && value != OPUS_FRAMESIZE_120_MS) { goto bad_arg; } @@ -2459,6 +2634,26 @@ int opus_encoder_ctl(OpusEncoder *st, int request, ...) *value = st->silk_mode.reducedDependency; } break; + case OPUS_SET_PHASE_INVERSION_DISABLED_REQUEST: + { + opus_int32 value = va_arg(ap, opus_int32); + if(value<0 || value>1) + { + goto bad_arg; + } + celt_encoder_ctl(celt_enc, OPUS_SET_PHASE_INVERSION_DISABLED(value)); + } + break; + case OPUS_GET_PHASE_INVERSION_DISABLED_REQUEST: + { + opus_int32 *value = va_arg(ap, opus_int32*); + if (!value) + { + goto bad_arg; + } + celt_encoder_ctl(celt_enc, OPUS_GET_PHASE_INVERSION_DISABLED(value)); + } + break; case OPUS_RESET_STATE: { void *silk_enc; diff --git a/code/opus-1.1.4/src/opus_multistream_decoder.c b/code/opus-1.1.4/src/opus_multistream_decoder.c index b95eaa6e..e421726f 100644 --- a/code/opus-1.1.4/src/opus_multistream_decoder.c +++ b/code/opus-1.1.4/src/opus_multistream_decoder.c @@ -237,7 +237,8 @@ static int opus_multistream_decode_native( for (s=0;slayout.nb_streams;s++) { OpusDecoder *dec; - int packet_offset, ret; + opus_int32 packet_offset; + int ret; dec = (OpusDecoder*)ptr; ptr += (s < st->layout.nb_coupled_streams) ? align(coupled_size) : align(mono_size); @@ -425,6 +426,7 @@ int opus_multistream_decoder_ctl(OpusMSDecoder *st, int request, ...) case OPUS_GET_SAMPLE_RATE_REQUEST: case OPUS_GET_GAIN_REQUEST: case OPUS_GET_LAST_PACKET_DURATION_REQUEST: + case OPUS_GET_PHASE_INVERSION_DISABLED_REQUEST: { OpusDecoder *dec; /* For int32* GET params, just query the first stream */ @@ -499,6 +501,7 @@ int opus_multistream_decoder_ctl(OpusMSDecoder *st, int request, ...) } break; case OPUS_SET_GAIN_REQUEST: + case OPUS_SET_PHASE_INVERSION_DISABLED_REQUEST: { int s; /* This works for int32 params */ diff --git a/code/opus-1.1.4/src/opus_multistream_encoder.c b/code/opus-1.1.4/src/opus_multistream_encoder.c index e722e31a..032fc003 100644 --- a/code/opus-1.1.4/src/opus_multistream_encoder.c +++ b/code/opus-1.1.4/src/opus_multistream_encoder.c @@ -87,7 +87,6 @@ struct OpusMSEncoder { int variable_duration; MappingType mapping_type; opus_int32 bitrate_bps; - float subframe_mem[3]; /* Encoder states go here */ /* then opus_val32 window_mem[channels*120]; */ /* then opus_val32 preemph_mem[channels]; */ @@ -133,6 +132,29 @@ static opus_val32 *ms_get_window_mem(OpusMSEncoder *st) return (opus_val32*)(void*)ptr; } +#ifdef ENABLE_EXPERIMENTAL_AMBISONICS +static int validate_ambisonics(int nb_channels, int *nb_streams, int *nb_coupled_streams) +{ + int order_plus_one; + int acn_channels; + int nondiegetic_channels; + + order_plus_one = isqrt32(nb_channels); + acn_channels = order_plus_one * order_plus_one; + nondiegetic_channels = nb_channels - acn_channels; + + if (order_plus_one < 1 || order_plus_one > 15 || + (nondiegetic_channels != 0 && nondiegetic_channels != 2)) + return 0; + + if (nb_streams) + *nb_streams = acn_channels + (nondiegetic_channels != 0); + if (nb_coupled_streams) + *nb_coupled_streams = nondiegetic_channels != 0; + return 1; +} +#endif + static int validate_encoder_layout(const ChannelLayout *layout) { int s; @@ -240,6 +262,7 @@ void surround_analysis(const CELTMode *celt_mode, const void *pcm, opus_val16 *b int pos[8] = {0}; int upsample; int frame_size; + int freq_size; opus_val16 channel_offset; opus_val32 bandE[21]; opus_val16 maskLogE[3][21]; @@ -250,6 +273,7 @@ void surround_analysis(const CELTMode *celt_mode, const void *pcm, opus_val16 *b upsample = resampling_factor(rate); frame_size = len*upsample; + freq_size = IMIN(960, frame_size); /* LM = log2(frame_size / 120) */ for (LM=0;LMmaxLM;LM++) @@ -258,7 +282,7 @@ void surround_analysis(const CELTMode *celt_mode, const void *pcm, opus_val16 *b ALLOC(in, frame_size+overlap, opus_val32); ALLOC(x, len, opus_val16); - ALLOC(freq, frame_size, opus_val32); + ALLOC(freq, freq_size, opus_val32); channel_pos(channels, pos); @@ -268,6 +292,9 @@ void surround_analysis(const CELTMode *celt_mode, const void *pcm, opus_val16 *b for (c=0;cpreemph, preemph_mem+c, 0); @@ -277,25 +304,33 @@ void surround_analysis(const CELTMode *celt_mode, const void *pcm, opus_val16 *b sum = celt_inner_prod(in, in, frame_size+overlap, 0); /* This should filter out both NaNs and ridiculous signals that could cause NaNs further down. */ - if (!(sum < 1e9f) || celt_isnan(sum)) + if (!(sum < 1e18f) || celt_isnan(sum)) { OPUS_CLEAR(in, frame_size+overlap); preemph_mem[c] = 0; } } #endif - clt_mdct_forward(&celt_mode->mdct, in, freq, celt_mode->window, - overlap, celt_mode->maxLM-LM, 1, arch); - if (upsample != 1) + OPUS_CLEAR(bandE, 21); + for (frame=0;framemdct, in+960*frame, freq, celt_mode->window, + overlap, celt_mode->maxLM-LM, 1, arch); + if (upsample != 1) + { + int bound = freq_size/upsample; + for (i=0;ilayout.nb_channels = channels; st->layout.nb_streams = streams; st->layout.nb_coupled_streams = coupled_streams; - st->subframe_mem[0]=st->subframe_mem[1]=st->subframe_mem[2]=0; if (mapping_type != MAPPING_TYPE_SURROUND) st->lfe_stream = -1; st->bitrate_bps = OPUS_AUTO; @@ -456,8 +490,16 @@ static int opus_multistream_encoder_init_impl( st->variable_duration = OPUS_FRAMESIZE_ARG; for (i=0;ilayout.nb_channels;i++) st->layout.mapping[i] = mapping[i]; - if (!validate_layout(&st->layout) || !validate_encoder_layout(&st->layout)) + if (!validate_layout(&st->layout)) return OPUS_BAD_ARG; + if (mapping_type == MAPPING_TYPE_SURROUND && + !validate_encoder_layout(&st->layout)) + return OPUS_BAD_ARG; +#ifdef ENABLE_EXPERIMENTAL_AMBISONICS + if (mapping_type == MAPPING_TYPE_AMBISONICS && + !validate_ambisonics(st->layout.nb_channels, NULL, NULL)) + return OPUS_BAD_ARG; +#endif ptr = (char*)st + align(sizeof(OpusMSEncoder)); coupled_size = opus_encoder_get_size(2); mono_size = opus_encoder_get_size(1); @@ -553,10 +595,12 @@ int opus_multistream_surround_encoder_init( } else if (mapping_family==254) { int i; - *streams=channels; - *coupled_streams=0; - for(i=0;ibitrate_bps > st->layout.nb_channels*40000) - stream_offset = 20000; - else - stream_offset = st->bitrate_bps/st->layout.nb_channels/2; - stream_offset += 60*(Fs/frame_size-50); - /* We start by giving each stream (coupled or uncoupled) the same bitrate. + nb_lfe = (st->lfe_stream!=-1); + nb_coupled = st->layout.nb_coupled_streams; + nb_uncoupled = st->layout.nb_streams-nb_coupled-nb_lfe; + nb_normal = 2*nb_coupled + nb_uncoupled; + + /* Give each non-LFE channel enough bits per channel for coding band energy. */ + channel_offset = 40*IMAX(50, Fs/frame_size); + + if (st->bitrate_bps==OPUS_AUTO) + { + bitrate = nb_normal*(channel_offset + Fs + 10000) + 8000*nb_lfe; + } else if (st->bitrate_bps==OPUS_BITRATE_MAX) + { + bitrate = nb_normal*300000 + nb_lfe*128000; + } else { + bitrate = st->bitrate_bps; + } + + /* Give LFE some basic stream_channel allocation but never exceed 1/20 of the + total rate for the non-energy part to avoid problems at really low rate. */ + lfe_offset = IMIN(bitrate/20, 3000) + 15*IMAX(50, Fs/frame_size); + + /* We give each stream (coupled or uncoupled) a starting bitrate. This models the main saving of coupled channels over uncoupled. */ - /* The LFE stream is an exception to the above and gets fewer bits. */ - lfe_offset = 3500 + 60*(Fs/frame_size-50); - /* Coupled streams get twice the mono rate after the first 20 kb/s. */ + stream_offset = (bitrate - channel_offset*nb_normal - lfe_offset*nb_lfe)/nb_normal/2; + stream_offset = IMAX(0, IMIN(20000, stream_offset)); + + /* Coupled streams get twice the mono rate after the offset is allocated. */ coupled_ratio = 512; /* Should depend on the bitrate, for now we assume LFE gets 1/8 the bits of mono */ lfe_ratio = 32; - /* Compute bitrate allocation between streams */ - if (st->bitrate_bps==OPUS_AUTO) - { - channel_rate = Fs+60*Fs/frame_size; - } else if (st->bitrate_bps==OPUS_BITRATE_MAX) - { - channel_rate = 300000; - } else { - int nb_lfe; - int nb_uncoupled; - int nb_coupled; - int total; - nb_lfe = (st->lfe_stream!=-1); - nb_coupled = st->layout.nb_coupled_streams; - nb_uncoupled = st->layout.nb_streams-nb_coupled-nb_lfe; - total = (nb_uncoupled<<8) /* mono */ - + coupled_ratio*nb_coupled /* stereo */ - + nb_lfe*lfe_ratio; - channel_rate = 256*(st->bitrate_bps-lfe_offset*nb_lfe-stream_offset*(nb_coupled+nb_uncoupled))/total; - } -#ifndef FIXED_POINT - if (st->variable_duration==OPUS_FRAMESIZE_VARIABLE && frame_size != Fs/50) - { - opus_int32 bonus; - bonus = 60*(Fs/frame_size-50); - channel_rate += bonus; - } -#endif + total = (nb_uncoupled<<8) /* mono */ + + coupled_ratio*nb_coupled /* stereo */ + + nb_lfe*lfe_ratio; + channel_rate = 256*(opus_int64)(bitrate - lfe_offset*nb_lfe - stream_offset*(nb_coupled+nb_uncoupled) - channel_offset*nb_normal)/total; for (i=0;ilayout.nb_streams;i++) { if (ilayout.nb_coupled_streams) - rate[i] = stream_offset+(channel_rate*coupled_ratio>>8); + rate[i] = 2*channel_offset + IMAX(0, stream_offset+(channel_rate*coupled_ratio>>8)); else if (i!=st->lfe_stream) - rate[i] = stream_offset+channel_rate; + rate[i] = channel_offset + IMAX(0, stream_offset + channel_rate); else - rate[i] = lfe_offset+(channel_rate*lfe_ratio>>8); + rate[i] = IMAX(0, lfe_offset+(channel_rate*lfe_ratio>>8)); } } @@ -736,47 +781,72 @@ static void ambisonics_rate_allocation( ) { int i; - int non_mono_rate; int total_rate; + int directional_rate; + int nondirectional_rate; + int leftover_bits; - /* The mono channel gets (rate_ratio_num / rate_ratio_den) times as many bits - * as all other channels */ + /* Each nondirectional channel gets (rate_ratio_num / rate_ratio_den) times + * as many bits as all other ambisonics channels. + */ const int rate_ratio_num = 4; const int rate_ratio_den = 3; - const int num_channels = st->layout.nb_streams; + const int nb_channels = st->layout.nb_streams + st->layout.nb_coupled_streams; + const int nb_nondirectional_channels = st->layout.nb_coupled_streams * 2 + 1; + const int nb_directional_channels = st->layout.nb_streams - 1; if (st->bitrate_bps==OPUS_AUTO) { - total_rate = num_channels * (20000 + st->layout.nb_streams*(Fs+60*Fs/frame_size)); + total_rate = (st->layout.nb_coupled_streams + st->layout.nb_streams) * + (Fs+60*Fs/frame_size) + st->layout.nb_streams * 15000; } else if (st->bitrate_bps==OPUS_BITRATE_MAX) { - total_rate = num_channels * 320000; - } else { + total_rate = nb_channels * 320000; + } else + { total_rate = st->bitrate_bps; } - /* Let y be the non-mono rate and let p, q be integers such that the mono - * channel rate is (p/q) * y. + /* Let y be the directional rate, m be the num of nondirectional channels + * m = (s + 1) + * and let p, q be integers such that the nondirectional rate is + * m_rate = (p / q) * y * Also let T be the total bitrate to allocate. Then - * (n - 1) y + (p/q) y = T - * y = (T q) / (qn - q + p) + * T = (n - m) * y + m * m_rate + * Solving for y, + * y = (q * T) / (m * (p - q) + n * q) */ - non_mono_rate = - total_rate * rate_ratio_den - / (rate_ratio_den*num_channels + rate_ratio_num - rate_ratio_den); + directional_rate = + total_rate * rate_ratio_den + / (nb_nondirectional_channels * (rate_ratio_num - rate_ratio_den) + + nb_channels * rate_ratio_den); -#ifndef FIXED_POINT - if (st->variable_duration==OPUS_FRAMESIZE_VARIABLE && frame_size != Fs/50) - { - opus_int32 bonus = 60*(Fs/frame_size-50); - non_mono_rate += bonus; - } -#endif + /* Calculate the nondirectional rate. + * m_rate = y * (p / q) + */ + nondirectional_rate = directional_rate * rate_ratio_num / rate_ratio_den; - rate[0] = total_rate - (num_channels - 1) * non_mono_rate; - for (i=1;ilayout.nb_streams;i++) + /* Calculate the leftover from truncation error. + * leftover = T - y * (n - m) - m_rate * m + * Place leftover bits in omnidirectional channel. + */ + leftover_bits = total_rate + - directional_rate * nb_directional_channels + - nondirectional_rate * nb_nondirectional_channels; + + /* Calculate rates for each channel */ + for (i = 0; i < st->layout.nb_streams; i++) { - rate[i] = non_mono_rate; + if (i < st->layout.nb_coupled_streams) + { + rate[i] = nondirectional_rate * 2; + } else if (i == st->layout.nb_coupled_streams) + { + rate[i] = nondirectional_rate + leftover_bits; + } else + { + rate[i] = directional_rate; + } } } #endif /* ENABLE_EXPERIMENTAL_AMBISONICS */ @@ -812,8 +882,8 @@ static opus_int32 rate_allocation( return rate_sum; } -/* Max size in case the encoder decides to return three frames */ -#define MS_FRAME_TMP (3*1275+7) +/* Max size in case the encoder decides to return six frames (6 x 20 ms = 120 ms) */ +#define MS_FRAME_TMP (6*1275+12) static int opus_multistream_encode_native ( OpusMSEncoder *st, @@ -859,32 +929,8 @@ static int opus_multistream_encode_native opus_encoder_ctl((OpusEncoder*)ptr, OPUS_GET_VBR(&vbr)); opus_encoder_ctl((OpusEncoder*)ptr, CELT_GET_MODE(&celt_mode)); - { - opus_int32 delay_compensation; - int channels; - - channels = st->layout.nb_streams + st->layout.nb_coupled_streams; - opus_encoder_ctl((OpusEncoder*)ptr, OPUS_GET_LOOKAHEAD(&delay_compensation)); - delay_compensation -= Fs/400; - frame_size = compute_frame_size(pcm, analysis_frame_size, - st->variable_duration, channels, Fs, st->bitrate_bps, - delay_compensation, downmix -#ifndef DISABLE_FLOAT_API - , st->subframe_mem -#endif - ); - } - - if (400*frame_size < Fs) - { - RESTORE_STACK; - return OPUS_BAD_ARG; - } - /* Validate frame_size before using it to allocate stack space. - This mirrors the checks in opus_encode[_float](). */ - if (400*frame_size != Fs && 200*frame_size != Fs && - 100*frame_size != Fs && 50*frame_size != Fs && - 25*frame_size != Fs && 50*frame_size != 3*Fs) + frame_size = frame_size_select(analysis_frame_size, st->variable_duration, Fs); + if (frame_size <= 0) { RESTORE_STACK; return OPUS_BAD_ARG; @@ -892,6 +938,9 @@ static int opus_multistream_encode_native /* Smallest packet the encoder can produce. */ smallest_packet = st->layout.nb_streams*2-1; + /* 100 ms needs an extra byte per stream for the ToC. */ + if (Fs/frame_size == 10) + smallest_packet += st->layout.nb_streams; if (max_data_bytes < smallest_packet) { RESTORE_STACK; @@ -1013,6 +1062,9 @@ static int opus_multistream_encode_native curr_max = max_data_bytes - tot_size; /* Reserve one byte for the last stream and two for the others */ curr_max -= IMAX(0,2*(st->layout.nb_streams-s-1)-1); + /* For 100 ms, reserve an extra byte per stream for the ToC */ + if (Fs/frame_size == 10) + curr_max -= st->layout.nb_streams-s-1; curr_max = IMIN(curr_max,MS_FRAME_TMP); /* Repacketizer will add one or two bytes for self-delimited frames */ if (s != st->layout.nb_streams-1) curr_max -= curr_max>253 ? 2 : 1; @@ -1161,9 +1213,11 @@ int opus_multistream_encoder_ctl(OpusMSEncoder *st, int request, ...) case OPUS_SET_BITRATE_REQUEST: { opus_int32 value = va_arg(ap, opus_int32); - if (value<0 && value!=OPUS_AUTO && value!=OPUS_BITRATE_MAX) + if (value != OPUS_AUTO && value != OPUS_BITRATE_MAX) { - goto bad_arg; + if (value <= 0) + goto bad_arg; + value = IMIN(300000*st->layout.nb_channels, IMAX(500*st->layout.nb_channels, value)); } st->bitrate_bps = value; } @@ -1206,6 +1260,7 @@ int opus_multistream_encoder_ctl(OpusMSEncoder *st, int request, ...) case OPUS_GET_INBAND_FEC_REQUEST: case OPUS_GET_FORCE_CHANNELS_REQUEST: case OPUS_GET_PREDICTION_DISABLED_REQUEST: + case OPUS_GET_PHASE_INVERSION_DISABLED_REQUEST: { OpusEncoder *enc; /* For int32* GET params, just query the first stream */ @@ -1252,6 +1307,7 @@ int opus_multistream_encoder_ctl(OpusMSEncoder *st, int request, ...) case OPUS_SET_FORCE_MODE_REQUEST: case OPUS_SET_FORCE_CHANNELS_REQUEST: case OPUS_SET_PREDICTION_DISABLED_REQUEST: + case OPUS_SET_PHASE_INVERSION_DISABLED_REQUEST: { int s; /* This works for int32 params */ @@ -1313,7 +1369,6 @@ int opus_multistream_encoder_ctl(OpusMSEncoder *st, int request, ...) case OPUS_RESET_STATE: { int s; - st->subframe_mem[0] = st->subframe_mem[1] = st->subframe_mem[2] = 0; if (st->mapping_type == MAPPING_TYPE_SURROUND) { OPUS_CLEAR(ms_get_preemph_mem(st), st->layout.nb_channels); diff --git a/code/opus-1.1.4/src/opus_private.h b/code/opus-1.1.4/src/opus_private.h index 3b62eed0..a731cc55 100644 --- a/code/opus-1.1.4/src/opus_private.h +++ b/code/opus-1.1.4/src/opus_private.h @@ -92,14 +92,6 @@ int encode_size(int size, unsigned char *data); opus_int32 frame_size_select(opus_int32 frame_size, int variable_duration, opus_int32 Fs); -opus_int32 compute_frame_size(const void *analysis_pcm, int frame_size, - int variable_duration, int C, opus_int32 Fs, int bitrate_bps, - int delay_compensation, downmix_func downmix -#ifndef DISABLE_FLOAT_API - , float *subframe_mem -#endif - ); - opus_int32 opus_encode_native(OpusEncoder *st, const opus_val16 *pcm, int frame_size, unsigned char *data, opus_int32 out_data_bytes, int lsb_depth, const void *analysis_pcm, opus_int32 analysis_size, int c1, int c2, diff --git a/misc/msvc12/quake3.vcxproj b/misc/msvc12/quake3.vcxproj index 992fc5ea..4321ecc8 100644 --- a/misc/msvc12/quake3.vcxproj +++ b/misc/msvc12/quake3.vcxproj @@ -523,21 +523,18 @@ - - - @@ -550,6 +547,7 @@ +