From b628115092dabafc1c02a066a8c513cb4a7d6050 Mon Sep 17 00:00:00 2001
From: derselbst <tom.mbrt@googlemail.com>
Date: Wed, 31 Oct 2018 09:23:39 +0100
Subject: [PATCH] optimize rounding and clipping to int16 samples

---
 src/synth/fluid_synth.c | 90 +++++++++++------------------------------
 1 file changed, 24 insertions(+), 66 deletions(-)

diff --git a/src/synth/fluid_synth.c b/src/synth/fluid_synth.c
index 50028c5f..172abdaf 100644
--- a/src/synth/fluid_synth.c
+++ b/src/synth/fluid_synth.c
@@ -88,7 +88,7 @@ static void fluid_synth_update_presets(fluid_synth_t *synth);
 static void fluid_synth_update_gain_LOCAL(fluid_synth_t *synth);
 static int fluid_synth_update_polyphony_LOCAL(fluid_synth_t *synth, int new_polyphony);
 static void init_dither(void);
-static FLUID_INLINE int roundi(float x);
+static FLUID_INLINE int16_t round_clip_to_i16(float x);
 static int fluid_synth_render_blocks(fluid_synth_t *synth, int blockcount);
 
 static fluid_voice_t *fluid_synth_free_voice_by_kill_LOCAL(fluid_synth_t *synth);
@@ -3808,17 +3808,28 @@ init_dither(void)
 }
 
 /* A portable replacement for roundf(), seems it may actually be faster too! */
-static FLUID_INLINE int
-roundi(float x)
+static FLUID_INLINE int16_t
+round_clip_to_i16(float x)
 {
+    long i;
     if(x >= 0.0f)
     {
-        return (int)(x + 0.5f);
+        i = (long)(x + 0.5f);
+        if (FLUID_UNLIKELY(i > 32767))
+        {
+            i = 32767;
+        }
     }
     else
     {
-        return (int)(x - 0.5f);
+        i = (long)(x - 0.5f);
+        if (FLUID_UNLIKELY(i < -32768))
+        {
+            i = -32768;
+        }
     }
+    
+    return (int16_t)i;
 }
 
 /**
@@ -3847,12 +3858,10 @@ fluid_synth_write_s16(fluid_synth_t *synth, int len,
                       void *rout, int roff, int rincr)
 {
     int i, j, k, cur;
-    signed short *left_out = (signed short *) lout;
-    signed short *right_out = (signed short *) rout;
+    int16_t *left_out = lout;
+    int16_t *right_out = rout;
     fluid_real_t *left_in;
     fluid_real_t *right_in;
-    int left_sample;
-    int right_sample;
     double time = fluid_utime();
     int di;
     float cpu_load;
@@ -3877,37 +3886,13 @@ fluid_synth_write_s16(fluid_synth_t *synth, int len,
             cur = 0;
         }
 
-        left_sample = roundi(left_in[0 * FLUID_BUFSIZE * FLUID_MIXER_MAX_BUFFERS_DEFAULT + cur] * 32766.0f + rand_table[0][di]);
-        right_sample = roundi(right_in[0 * FLUID_BUFSIZE * FLUID_MIXER_MAX_BUFFERS_DEFAULT + cur] * 32766.0f + rand_table[1][di]);
+        left_out[j] = round_clip_to_i16(left_in[0 * FLUID_BUFSIZE * FLUID_MIXER_MAX_BUFFERS_DEFAULT + cur] * 32766.0f + rand_table[0][di]);
+        right_out[k] = round_clip_to_i16(right_in[0 * FLUID_BUFSIZE * FLUID_MIXER_MAX_BUFFERS_DEFAULT + cur] * 32766.0f + rand_table[1][di]);
 
         if(++di >= DITHER_SIZE)
         {
             di = 0;
         }
-
-        /* digital clipping */
-        if(left_sample > 32767)
-        {
-            left_sample = 32767;
-        }
-
-        if(left_sample < -32768)
-        {
-            left_sample = -32768;
-        }
-
-        if(right_sample > 32767)
-        {
-            right_sample = 32767;
-        }
-
-        if(right_sample < -32768)
-        {
-            right_sample = -32768;
-        }
-
-        left_out[j] = (signed short) left_sample;
-        right_out[k] = (signed short) right_sample;
     }
 
     synth->cur = cur;
@@ -3946,47 +3931,20 @@ fluid_synth_dither_s16(int *dither_index, int len, float *lin, float *rin,
                        void *rout, int roff, int rincr)
 {
     int i, j, k;
-    signed short *left_out = (signed short *) lout;
-    signed short *right_out = (signed short *) rout;
-    int left_sample;
-    int right_sample;
+    int16_t *left_out = lout;
+    int16_t *right_out = rout;
     int di = *dither_index;
     fluid_profile_ref_var(prof_ref);
 
     for(i = 0, j = loff, k = roff; i < len; i++, j += lincr, k += rincr)
     {
-
-        left_sample = roundi(lin[i] * 32766.0f + rand_table[0][di]);
-        right_sample = roundi(rin[i] * 32766.0f + rand_table[1][di]);
+        left_out[j] = round_clip_to_i16(lin[i] * 32766.0f + rand_table[0][di]);
+        right_out[k] = round_clip_to_i16(rin[i] * 32766.0f + rand_table[1][di]);
 
         if(++di >= DITHER_SIZE)
         {
             di = 0;
         }
-
-        /* digital clipping */
-        if(left_sample > 32767)
-        {
-            left_sample = 32767;
-        }
-
-        if(left_sample < -32768)
-        {
-            left_sample = -32768;
-        }
-
-        if(right_sample > 32767)
-        {
-            right_sample = 32767;
-        }
-
-        if(right_sample < -32768)
-        {
-            right_sample = -32768;
-        }
-
-        left_out[j] = (signed short) left_sample;
-        right_out[k] = (signed short) right_sample;
     }
 
     *dither_index = di;	/* keep dither buffer continous */