zdoom/gzdoom-gles
0
0
Fork 0
mirror of https://github.com/ZDoom/gzdoom-gles.git synced 2024-11-29 07:22:07 +00:00
Code Issues Projects Releases Packages Wiki Activity

Hook up all php generated wall drawers and enable them per default

Browse source
This commit is contained in:
Magnus Norddahl 2017-02-18 05:34:26 +01:00
parent 89a2e434d4
commit 097dda38a9
4 changed files with 181 additions and 126 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

42
src/swrenderer/drawers/r_draw_rgba.cpp
View file

@ -59,7 +59,7 @@ CVAR(Bool, r_mipmap, true, CVAR_ARCHIVE | CVAR_GLOBALCONFIG);
// Level of detail texture bias // Level of detail texture bias
CVAR(Float, r_lod_bias, -1.5, 0); // To do: add CVAR_ARCHIVE | CVAR_GLOBALCONFIG when a good default has been decided CVAR(Float, r_lod_bias, -1.5, 0); // To do: add CVAR_ARCHIVE | CVAR_GLOBALCONFIG when a good default has been decided
CVAR(Bool, r_phpdrawers, false, 0); CVAR(Bool, r_phpdrawers, true, 0);
namespace swrenderer namespace swrenderer
{ {
@ -71,6 +71,46 @@ namespace swrenderer
Queue->Push<DrawWall1LLVMCommand>(args); Queue->Push<DrawWall1LLVMCommand>(args);
} }
void SWTruecolorDrawers::DrawWallMaskedColumn(const WallDrawerArgs &args)
{
if (r_phpdrawers)
Queue->Push<DrawWallMasked32Command>(args);
else
Queue->Push<DrawWallMasked1LLVMCommand>(args);
}
void SWTruecolorDrawers::DrawWallAddColumn(const WallDrawerArgs &args)
{
if (r_phpdrawers)
Queue->Push<DrawWallAddClamp32Command>(args);
else
Queue->Push<DrawWallAdd1LLVMCommand>(args);
}
void SWTruecolorDrawers::DrawWallAddClampColumn(const WallDrawerArgs &args)
{
if (r_phpdrawers)
Queue->Push<DrawWallAddClamp32Command>(args);
else
Queue->Push<DrawWallAddClamp1LLVMCommand>(args);
}
void SWTruecolorDrawers::DrawWallSubClampColumn(const WallDrawerArgs &args)
{
if (r_phpdrawers)
Queue->Push<DrawWallSubClamp32Command>(args);
else
Queue->Push<DrawWallSubClamp1LLVMCommand>(args);
}
void SWTruecolorDrawers::DrawWallRevSubClampColumn(const WallDrawerArgs &args)
{
if (r_phpdrawers)
Queue->Push<DrawWallRevSubClamp32Command>(args);
else
Queue->Push<DrawWallRevSubClamp1LLVMCommand>(args);
}
DrawSpanLLVMCommand::DrawSpanLLVMCommand(const SpanDrawerArgs &drawerargs) DrawSpanLLVMCommand::DrawSpanLLVMCommand(const SpanDrawerArgs &drawerargs)
{ {
auto shade_constants = drawerargs.ColormapConstants(); auto shade_constants = drawerargs.ColormapConstants();

10
src/swrenderer/drawers/r_draw_rgba.h
View file

@ -360,11 +360,11 @@ namespace swrenderer
using SWPixelFormatDrawers::SWPixelFormatDrawers; using SWPixelFormatDrawers::SWPixelFormatDrawers;
void DrawWallColumn(const WallDrawerArgs &args) override; void DrawWallColumn(const WallDrawerArgs &args) override;
void DrawWallMaskedColumn(const WallDrawerArgs &args) override { Queue->Push<DrawWallMasked1LLVMCommand>(args); } void DrawWallMaskedColumn(const WallDrawerArgs &args) override;
void DrawWallAddColumn(const WallDrawerArgs &args) override { Queue->Push<DrawWallAdd1LLVMCommand>(args); } void DrawWallAddColumn(const WallDrawerArgs &args) override;
void DrawWallAddClampColumn(const WallDrawerArgs &args) override { Queue->Push<DrawWallAddClamp1LLVMCommand>(args); } void DrawWallAddClampColumn(const WallDrawerArgs &args) override;
void DrawWallSubClampColumn(const WallDrawerArgs &args) override { Queue->Push<DrawWallSubClamp1LLVMCommand>(args); } void DrawWallSubClampColumn(const WallDrawerArgs &args) override;
void DrawWallRevSubClampColumn(const WallDrawerArgs &args) override { Queue->Push<DrawWallRevSubClamp1LLVMCommand>(args); } void DrawWallRevSubClampColumn(const WallDrawerArgs &args) override;
void DrawSingleSkyColumn(const SkyDrawerArgs &args) override { Queue->Push<DrawSingleSky1LLVMCommand>(args); } void DrawSingleSkyColumn(const SkyDrawerArgs &args) override { Queue->Push<DrawSingleSky1LLVMCommand>(args); }
void DrawDoubleSkyColumn(const SkyDrawerArgs &args) override { Queue->Push<DrawDoubleSky1LLVMCommand>(args); } void DrawDoubleSkyColumn(const SkyDrawerArgs &args) override { Queue->Push<DrawDoubleSky1LLVMCommand>(args); }
void DrawColumn(const SpriteDrawerArgs &args) override { Queue->Push<DrawColumnLLVMCommand>(args); } void DrawColumn(const SpriteDrawerArgs &args) override { Queue->Push<DrawColumnLLVMCommand>(args); }

230
src/swrenderer/drawers/r_draw_wall32.h
View file

@ -71,7 +71,6 @@ namespace swrenderer
dest = thread->dest_for_thread(dest_y, pitch, dest); dest = thread->dest_for_thread(dest_y, pitch, dest);
fracstep *= thread->num_cores; fracstep *= thread->num_cores;
pitch *= thread->num_cores; pitch *= thread->num_cores;
__m128i srcalpha = _mm_set1_epi16(args.SrcAlpha()); __m128i srcalpha = _mm_set1_epi16(args.SrcAlpha());
__m128i destalpha = _mm_set1_epi16(args.DestAlpha()); __m128i destalpha = _mm_set1_epi16(args.DestAlpha());
@ -120,7 +119,7 @@ namespace swrenderer
dest = thread->dest_for_thread(dest_y, pitch, dest); dest = thread->dest_for_thread(dest_y, pitch, dest);
fracstep *= thread->num_cores; fracstep *= thread->num_cores;
pitch *= thread->num_cores; pitch *= thread->num_cores;
frac -= one / 2;
__m128i srcalpha = _mm_set1_epi16(args.SrcAlpha()); __m128i srcalpha = _mm_set1_epi16(args.SrcAlpha());
__m128i destalpha = _mm_set1_epi16(args.DestAlpha()); __m128i destalpha = _mm_set1_epi16(args.DestAlpha());
@ -141,8 +140,8 @@ namespace swrenderer
unsigned int p11 = source2[y1]; unsigned int p11 = source2[y1];
unsigned int inv_b = texturefracx; unsigned int inv_b = texturefracx;
unsigned int a = (frac_y1 >> (FRACBITS - 4)) & 15; unsigned int inv_a = (frac_y1 >> (FRACBITS - 4)) & 15;
unsigned int inv_a = 16 - a; unsigned int a = 16 - inv_a;
unsigned int b = 16 - inv_b; unsigned int b = 16 - inv_b;
unsigned int sred = (RPART(p00) * (a * b) + RPART(p01) * (inv_a * b) + RPART(p10) * (a * inv_b) + RPART(p11) * (inv_a * inv_b) + 127) >> 8; unsigned int sred = (RPART(p00) * (a * b) + RPART(p01) * (inv_a * b) + RPART(p10) * (a * inv_b) + RPART(p11) * (inv_a * inv_b) + 127) >> 8;
@ -179,8 +178,9 @@ namespace swrenderer
int light = 256 - (args.Light() >> (FRACBITS - 8)); int light = 256 - (args.Light() >> (FRACBITS - 8));
__m128i mlight = _mm_set_epi16(256, light, light, light, 256, light, light, light); __m128i mlight = _mm_set_epi16(256, light, light, light, 256, light, light, light);
__m128i inv_light = _mm_set_epi16(0, 256 - light, 256 - light, 256 - light, 0, 256 - light, 256 - light, 256 - light); __m128i inv_light = _mm_set_epi16(0, 256 - light, 256 - light, 256 - light, 0, 256 - light, 256 - light, 256 - light);
__m128i inv_desaturate = _mm_setr_epi16(0, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 0, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate); __m128i inv_desaturate = _mm_setr_epi16(256, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate);
__m128i shade_fade = _mm_set_epi16(shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue, shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue); __m128i shade_fade = _mm_set_epi16(shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue, shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue);
shade_fade = _mm_mullo_epi16(shade_fade, inv_light);
__m128i shade_light = _mm_set_epi16(shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue, shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue); __m128i shade_light = _mm_set_epi16(shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue, shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue);
int desaturate = shade_constants.desaturate; int desaturate = shade_constants.desaturate;
@ -198,7 +198,6 @@ namespace swrenderer
dest = thread->dest_for_thread(dest_y, pitch, dest); dest = thread->dest_for_thread(dest_y, pitch, dest);
fracstep *= thread->num_cores; fracstep *= thread->num_cores;
pitch *= thread->num_cores; pitch *= thread->num_cores;
__m128i srcalpha = _mm_set1_epi16(args.SrcAlpha()); __m128i srcalpha = _mm_set1_epi16(args.SrcAlpha());
__m128i destalpha = _mm_set1_epi16(args.DestAlpha()); __m128i destalpha = _mm_set1_epi16(args.DestAlpha());
@ -213,15 +212,16 @@ namespace swrenderer
__m128i fgcolor = _mm_unpacklo_epi8(_mm_cvtsi32_si128(ifgcolor), _mm_setzero_si128()); __m128i fgcolor = _mm_unpacklo_epi8(_mm_cvtsi32_si128(ifgcolor), _mm_setzero_si128());
// Shade // Shade
int blue = BPART(ifgcolor); int blue0 = BPART(ifgcolor);
int green = GPART(ifgcolor); int green0 = GPART(ifgcolor);
int red = RPART(ifgcolor); int red0 = RPART(ifgcolor);
__m128i intensity = _mm_set1_epi16(((red * 77 + green * 143 + blue * 37) >> 8) * desaturate); int intensity0 = ((red0 * 77 + green0 * 143 + blue0 * 37) >> 8) * desaturate;
__m128i intensity = _mm_set_epi16(0, intensity0, intensity0, intensity0, 0, intensity0, intensity0, intensity0);
fgcolor = _mm_srli_epi16(_mm_add_epi16(_mm_mullo_epi16(fgcolor, inv_desaturate), intensity), 8); fgcolor = _mm_srli_epi16(_mm_add_epi16(_mm_mullo_epi16(fgcolor, inv_desaturate), intensity), 8);
fgcolor = _mm_mullo_epi16(fgcolor, _mm_set1_epi16(light)); fgcolor = _mm_mullo_epi16(fgcolor, mlight);
fgcolor = _mm_srli_epi16(_mm_add_epi16(_mm_mullo_epi16(shade_fade, inv_light), fgcolor), 8); fgcolor = _mm_srli_epi16(_mm_add_epi16(shade_fade, fgcolor), 8);
fgcolor = _mm_srli_epi16(_mm_mullo_epi16(fgcolor, shade_light), 8); fgcolor = _mm_srli_epi16(_mm_mullo_epi16(fgcolor, shade_light), 8);
// Blend // Blend
@ -241,8 +241,9 @@ namespace swrenderer
int light = 256 - (args.Light() >> (FRACBITS - 8)); int light = 256 - (args.Light() >> (FRACBITS - 8));
__m128i mlight = _mm_set_epi16(256, light, light, light, 256, light, light, light); __m128i mlight = _mm_set_epi16(256, light, light, light, 256, light, light, light);
__m128i inv_light = _mm_set_epi16(0, 256 - light, 256 - light, 256 - light, 0, 256 - light, 256 - light, 256 - light); __m128i inv_light = _mm_set_epi16(0, 256 - light, 256 - light, 256 - light, 0, 256 - light, 256 - light, 256 - light);
__m128i inv_desaturate = _mm_setr_epi16(0, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 0, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate); __m128i inv_desaturate = _mm_setr_epi16(256, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate);
__m128i shade_fade = _mm_set_epi16(shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue, shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue); __m128i shade_fade = _mm_set_epi16(shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue, shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue);
shade_fade = _mm_mullo_epi16(shade_fade, inv_light);
__m128i shade_light = _mm_set_epi16(shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue, shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue); __m128i shade_light = _mm_set_epi16(shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue, shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue);
int desaturate = shade_constants.desaturate; int desaturate = shade_constants.desaturate;
@ -260,7 +261,7 @@ namespace swrenderer
dest = thread->dest_for_thread(dest_y, pitch, dest); dest = thread->dest_for_thread(dest_y, pitch, dest);
fracstep *= thread->num_cores; fracstep *= thread->num_cores;
pitch *= thread->num_cores; pitch *= thread->num_cores;
frac -= one / 2;
__m128i srcalpha = _mm_set1_epi16(args.SrcAlpha()); __m128i srcalpha = _mm_set1_epi16(args.SrcAlpha());
__m128i destalpha = _mm_set1_epi16(args.DestAlpha()); __m128i destalpha = _mm_set1_epi16(args.DestAlpha());
@ -281,8 +282,8 @@ namespace swrenderer
unsigned int p11 = source2[y1]; unsigned int p11 = source2[y1];
unsigned int inv_b = texturefracx; unsigned int inv_b = texturefracx;
unsigned int a = (frac_y1 >> (FRACBITS - 4)) & 15; unsigned int inv_a = (frac_y1 >> (FRACBITS - 4)) & 15;
unsigned int inv_a = 16 - a; unsigned int a = 16 - inv_a;
unsigned int b = 16 - inv_b; unsigned int b = 16 - inv_b;
unsigned int sred = (RPART(p00) * (a * b) + RPART(p01) * (inv_a * b) + RPART(p10) * (a * inv_b) + RPART(p11) * (inv_a * inv_b) + 127) >> 8; unsigned int sred = (RPART(p00) * (a * b) + RPART(p01) * (inv_a * b) + RPART(p10) * (a * inv_b) + RPART(p11) * (inv_a * inv_b) + 127) >> 8;
@ -294,15 +295,16 @@ namespace swrenderer
__m128i fgcolor = _mm_unpacklo_epi8(_mm_cvtsi32_si128(ifgcolor), _mm_setzero_si128()); __m128i fgcolor = _mm_unpacklo_epi8(_mm_cvtsi32_si128(ifgcolor), _mm_setzero_si128());
// Shade // Shade
int blue = BPART(ifgcolor); int blue0 = BPART(ifgcolor);
int green = GPART(ifgcolor); int green0 = GPART(ifgcolor);
int red = RPART(ifgcolor); int red0 = RPART(ifgcolor);
__m128i intensity = _mm_set1_epi16(((red * 77 + green * 143 + blue * 37) >> 8) * desaturate); int intensity0 = ((red0 * 77 + green0 * 143 + blue0 * 37) >> 8) * desaturate;
__m128i intensity = _mm_set_epi16(0, intensity0, intensity0, intensity0, 0, intensity0, intensity0, intensity0);
fgcolor = _mm_srli_epi16(_mm_add_epi16(_mm_mullo_epi16(fgcolor, inv_desaturate), intensity), 8); fgcolor = _mm_srli_epi16(_mm_add_epi16(_mm_mullo_epi16(fgcolor, inv_desaturate), intensity), 8);
fgcolor = _mm_mullo_epi16(fgcolor, _mm_set1_epi16(light)); fgcolor = _mm_mullo_epi16(fgcolor, mlight);
fgcolor = _mm_srli_epi16(_mm_add_epi16(_mm_mullo_epi16(shade_fade, inv_light), fgcolor), 8); fgcolor = _mm_srli_epi16(_mm_add_epi16(shade_fade, fgcolor), 8);
fgcolor = _mm_srli_epi16(_mm_mullo_epi16(fgcolor, shade_light), 8); fgcolor = _mm_srli_epi16(_mm_mullo_epi16(fgcolor, shade_light), 8);
// Blend // Blend
@ -359,7 +361,6 @@ namespace swrenderer
dest = thread->dest_for_thread(dest_y, pitch, dest); dest = thread->dest_for_thread(dest_y, pitch, dest);
fracstep *= thread->num_cores; fracstep *= thread->num_cores;
pitch *= thread->num_cores; pitch *= thread->num_cores;
__m128i srcalpha = _mm_set1_epi16(args.SrcAlpha()); __m128i srcalpha = _mm_set1_epi16(args.SrcAlpha());
__m128i destalpha = _mm_set1_epi16(args.DestAlpha()); __m128i destalpha = _mm_set1_epi16(args.DestAlpha());
@ -415,7 +416,7 @@ namespace swrenderer
dest = thread->dest_for_thread(dest_y, pitch, dest); dest = thread->dest_for_thread(dest_y, pitch, dest);
fracstep *= thread->num_cores; fracstep *= thread->num_cores;
pitch *= thread->num_cores; pitch *= thread->num_cores;
frac -= one / 2;
__m128i srcalpha = _mm_set1_epi16(args.SrcAlpha()); __m128i srcalpha = _mm_set1_epi16(args.SrcAlpha());
__m128i destalpha = _mm_set1_epi16(args.DestAlpha()); __m128i destalpha = _mm_set1_epi16(args.DestAlpha());
@ -436,8 +437,8 @@ namespace swrenderer
unsigned int p11 = source2[y1]; unsigned int p11 = source2[y1];
unsigned int inv_b = texturefracx; unsigned int inv_b = texturefracx;
unsigned int a = (frac_y1 >> (FRACBITS - 4)) & 15; unsigned int inv_a = (frac_y1 >> (FRACBITS - 4)) & 15;
unsigned int inv_a = 16 - a; unsigned int a = 16 - inv_a;
unsigned int b = 16 - inv_b; unsigned int b = 16 - inv_b;
unsigned int sred = (RPART(p00) * (a * b) + RPART(p01) * (inv_a * b) + RPART(p10) * (a * inv_b) + RPART(p11) * (inv_a * inv_b) + 127) >> 8; unsigned int sred = (RPART(p00) * (a * b) + RPART(p01) * (inv_a * b) + RPART(p10) * (a * inv_b) + RPART(p11) * (inv_a * inv_b) + 127) >> 8;
@ -481,8 +482,9 @@ namespace swrenderer
int light = 256 - (args.Light() >> (FRACBITS - 8)); int light = 256 - (args.Light() >> (FRACBITS - 8));
__m128i mlight = _mm_set_epi16(256, light, light, light, 256, light, light, light); __m128i mlight = _mm_set_epi16(256, light, light, light, 256, light, light, light);
__m128i inv_light = _mm_set_epi16(0, 256 - light, 256 - light, 256 - light, 0, 256 - light, 256 - light, 256 - light); __m128i inv_light = _mm_set_epi16(0, 256 - light, 256 - light, 256 - light, 0, 256 - light, 256 - light, 256 - light);
__m128i inv_desaturate = _mm_setr_epi16(0, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 0, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate); __m128i inv_desaturate = _mm_setr_epi16(256, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate);
__m128i shade_fade = _mm_set_epi16(shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue, shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue); __m128i shade_fade = _mm_set_epi16(shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue, shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue);
shade_fade = _mm_mullo_epi16(shade_fade, inv_light);
__m128i shade_light = _mm_set_epi16(shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue, shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue); __m128i shade_light = _mm_set_epi16(shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue, shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue);
int desaturate = shade_constants.desaturate; int desaturate = shade_constants.desaturate;
@ -500,7 +502,6 @@ namespace swrenderer
dest = thread->dest_for_thread(dest_y, pitch, dest); dest = thread->dest_for_thread(dest_y, pitch, dest);
fracstep *= thread->num_cores; fracstep *= thread->num_cores;
pitch *= thread->num_cores; pitch *= thread->num_cores;
__m128i srcalpha = _mm_set1_epi16(args.SrcAlpha()); __m128i srcalpha = _mm_set1_epi16(args.SrcAlpha());
__m128i destalpha = _mm_set1_epi16(args.DestAlpha()); __m128i destalpha = _mm_set1_epi16(args.DestAlpha());
@ -515,15 +516,16 @@ namespace swrenderer
__m128i fgcolor = _mm_unpacklo_epi8(_mm_cvtsi32_si128(ifgcolor), _mm_setzero_si128()); __m128i fgcolor = _mm_unpacklo_epi8(_mm_cvtsi32_si128(ifgcolor), _mm_setzero_si128());
// Shade // Shade
int blue = BPART(ifgcolor); int blue0 = BPART(ifgcolor);
int green = GPART(ifgcolor); int green0 = GPART(ifgcolor);
int red = RPART(ifgcolor); int red0 = RPART(ifgcolor);
__m128i intensity = _mm_set1_epi16(((red * 77 + green * 143 + blue * 37) >> 8) * desaturate); int intensity0 = ((red0 * 77 + green0 * 143 + blue0 * 37) >> 8) * desaturate;
__m128i intensity = _mm_set_epi16(0, intensity0, intensity0, intensity0, 0, intensity0, intensity0, intensity0);
fgcolor = _mm_srli_epi16(_mm_add_epi16(_mm_mullo_epi16(fgcolor, inv_desaturate), intensity), 8); fgcolor = _mm_srli_epi16(_mm_add_epi16(_mm_mullo_epi16(fgcolor, inv_desaturate), intensity), 8);
fgcolor = _mm_mullo_epi16(fgcolor, _mm_set1_epi16(light)); fgcolor = _mm_mullo_epi16(fgcolor, mlight);
fgcolor = _mm_srli_epi16(_mm_add_epi16(_mm_mullo_epi16(shade_fade, inv_light), fgcolor), 8); fgcolor = _mm_srli_epi16(_mm_add_epi16(shade_fade, fgcolor), 8);
fgcolor = _mm_srli_epi16(_mm_mullo_epi16(fgcolor, shade_light), 8); fgcolor = _mm_srli_epi16(_mm_mullo_epi16(fgcolor, shade_light), 8);
// Blend // Blend
@ -550,8 +552,9 @@ namespace swrenderer
int light = 256 - (args.Light() >> (FRACBITS - 8)); int light = 256 - (args.Light() >> (FRACBITS - 8));
__m128i mlight = _mm_set_epi16(256, light, light, light, 256, light, light, light); __m128i mlight = _mm_set_epi16(256, light, light, light, 256, light, light, light);
__m128i inv_light = _mm_set_epi16(0, 256 - light, 256 - light, 256 - light, 0, 256 - light, 256 - light, 256 - light); __m128i inv_light = _mm_set_epi16(0, 256 - light, 256 - light, 256 - light, 0, 256 - light, 256 - light, 256 - light);
__m128i inv_desaturate = _mm_setr_epi16(0, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 0, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate); __m128i inv_desaturate = _mm_setr_epi16(256, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate);
__m128i shade_fade = _mm_set_epi16(shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue, shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue); __m128i shade_fade = _mm_set_epi16(shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue, shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue);
shade_fade = _mm_mullo_epi16(shade_fade, inv_light);
__m128i shade_light = _mm_set_epi16(shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue, shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue); __m128i shade_light = _mm_set_epi16(shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue, shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue);
int desaturate = shade_constants.desaturate; int desaturate = shade_constants.desaturate;
@ -569,7 +572,7 @@ namespace swrenderer
dest = thread->dest_for_thread(dest_y, pitch, dest); dest = thread->dest_for_thread(dest_y, pitch, dest);
fracstep *= thread->num_cores; fracstep *= thread->num_cores;
pitch *= thread->num_cores; pitch *= thread->num_cores;
frac -= one / 2;
__m128i srcalpha = _mm_set1_epi16(args.SrcAlpha()); __m128i srcalpha = _mm_set1_epi16(args.SrcAlpha());
__m128i destalpha = _mm_set1_epi16(args.DestAlpha()); __m128i destalpha = _mm_set1_epi16(args.DestAlpha());
@ -590,8 +593,8 @@ namespace swrenderer
unsigned int p11 = source2[y1]; unsigned int p11 = source2[y1];
unsigned int inv_b = texturefracx; unsigned int inv_b = texturefracx;
unsigned int a = (frac_y1 >> (FRACBITS - 4)) & 15; unsigned int inv_a = (frac_y1 >> (FRACBITS - 4)) & 15;
unsigned int inv_a = 16 - a; unsigned int a = 16 - inv_a;
unsigned int b = 16 - inv_b; unsigned int b = 16 - inv_b;
unsigned int sred = (RPART(p00) * (a * b) + RPART(p01) * (inv_a * b) + RPART(p10) * (a * inv_b) + RPART(p11) * (inv_a * inv_b) + 127) >> 8; unsigned int sred = (RPART(p00) * (a * b) + RPART(p01) * (inv_a * b) + RPART(p10) * (a * inv_b) + RPART(p11) * (inv_a * inv_b) + 127) >> 8;
@ -603,15 +606,16 @@ namespace swrenderer
__m128i fgcolor = _mm_unpacklo_epi8(_mm_cvtsi32_si128(ifgcolor), _mm_setzero_si128()); __m128i fgcolor = _mm_unpacklo_epi8(_mm_cvtsi32_si128(ifgcolor), _mm_setzero_si128());
// Shade // Shade
int blue = BPART(ifgcolor); int blue0 = BPART(ifgcolor);
int green = GPART(ifgcolor); int green0 = GPART(ifgcolor);
int red = RPART(ifgcolor); int red0 = RPART(ifgcolor);
__m128i intensity = _mm_set1_epi16(((red * 77 + green * 143 + blue * 37) >> 8) * desaturate); int intensity0 = ((red0 * 77 + green0 * 143 + blue0 * 37) >> 8) * desaturate;
__m128i intensity = _mm_set_epi16(0, intensity0, intensity0, intensity0, 0, intensity0, intensity0, intensity0);
fgcolor = _mm_srli_epi16(_mm_add_epi16(_mm_mullo_epi16(fgcolor, inv_desaturate), intensity), 8); fgcolor = _mm_srli_epi16(_mm_add_epi16(_mm_mullo_epi16(fgcolor, inv_desaturate), intensity), 8);
fgcolor = _mm_mullo_epi16(fgcolor, _mm_set1_epi16(light)); fgcolor = _mm_mullo_epi16(fgcolor, mlight);
fgcolor = _mm_srli_epi16(_mm_add_epi16(_mm_mullo_epi16(shade_fade, inv_light), fgcolor), 8); fgcolor = _mm_srli_epi16(_mm_add_epi16(shade_fade, fgcolor), 8);
fgcolor = _mm_srli_epi16(_mm_mullo_epi16(fgcolor, shade_light), 8); fgcolor = _mm_srli_epi16(_mm_mullo_epi16(fgcolor, shade_light), 8);
// Blend // Blend
@ -675,7 +679,6 @@ namespace swrenderer
dest = thread->dest_for_thread(dest_y, pitch, dest); dest = thread->dest_for_thread(dest_y, pitch, dest);
fracstep *= thread->num_cores; fracstep *= thread->num_cores;
pitch *= thread->num_cores; pitch *= thread->num_cores;
__m128i srcalpha = _mm_set1_epi16(args.SrcAlpha()); __m128i srcalpha = _mm_set1_epi16(args.SrcAlpha());
__m128i destalpha = _mm_set1_epi16(args.DestAlpha()); __m128i destalpha = _mm_set1_epi16(args.DestAlpha());
@ -746,7 +749,7 @@ namespace swrenderer
dest = thread->dest_for_thread(dest_y, pitch, dest); dest = thread->dest_for_thread(dest_y, pitch, dest);
fracstep *= thread->num_cores; fracstep *= thread->num_cores;
pitch *= thread->num_cores; pitch *= thread->num_cores;
frac -= one / 2;
__m128i srcalpha = _mm_set1_epi16(args.SrcAlpha()); __m128i srcalpha = _mm_set1_epi16(args.SrcAlpha());
__m128i destalpha = _mm_set1_epi16(args.DestAlpha()); __m128i destalpha = _mm_set1_epi16(args.DestAlpha());
@ -767,8 +770,8 @@ namespace swrenderer
unsigned int p11 = source2[y1]; unsigned int p11 = source2[y1];
unsigned int inv_b = texturefracx; unsigned int inv_b = texturefracx;
unsigned int a = (frac_y1 >> (FRACBITS - 4)) & 15; unsigned int inv_a = (frac_y1 >> (FRACBITS - 4)) & 15;
unsigned int inv_a = 16 - a; unsigned int a = 16 - inv_a;
unsigned int b = 16 - inv_b; unsigned int b = 16 - inv_b;
unsigned int sred = (RPART(p00) * (a * b) + RPART(p01) * (inv_a * b) + RPART(p10) * (a * inv_b) + RPART(p11) * (inv_a * inv_b) + 127) >> 8; unsigned int sred = (RPART(p00) * (a * b) + RPART(p01) * (inv_a * b) + RPART(p10) * (a * inv_b) + RPART(p11) * (inv_a * inv_b) + 127) >> 8;
@ -827,8 +830,9 @@ namespace swrenderer
int light = 256 - (args.Light() >> (FRACBITS - 8)); int light = 256 - (args.Light() >> (FRACBITS - 8));
__m128i mlight = _mm_set_epi16(256, light, light, light, 256, light, light, light); __m128i mlight = _mm_set_epi16(256, light, light, light, 256, light, light, light);
__m128i inv_light = _mm_set_epi16(0, 256 - light, 256 - light, 256 - light, 0, 256 - light, 256 - light, 256 - light); __m128i inv_light = _mm_set_epi16(0, 256 - light, 256 - light, 256 - light, 0, 256 - light, 256 - light, 256 - light);
__m128i inv_desaturate = _mm_setr_epi16(0, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 0, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate); __m128i inv_desaturate = _mm_setr_epi16(256, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate);
__m128i shade_fade = _mm_set_epi16(shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue, shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue); __m128i shade_fade = _mm_set_epi16(shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue, shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue);
shade_fade = _mm_mullo_epi16(shade_fade, inv_light);
__m128i shade_light = _mm_set_epi16(shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue, shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue); __m128i shade_light = _mm_set_epi16(shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue, shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue);
int desaturate = shade_constants.desaturate; int desaturate = shade_constants.desaturate;
@ -846,7 +850,6 @@ namespace swrenderer
dest = thread->dest_for_thread(dest_y, pitch, dest); dest = thread->dest_for_thread(dest_y, pitch, dest);
fracstep *= thread->num_cores; fracstep *= thread->num_cores;
pitch *= thread->num_cores; pitch *= thread->num_cores;
__m128i srcalpha = _mm_set1_epi16(args.SrcAlpha()); __m128i srcalpha = _mm_set1_epi16(args.SrcAlpha());
__m128i destalpha = _mm_set1_epi16(args.DestAlpha()); __m128i destalpha = _mm_set1_epi16(args.DestAlpha());
@ -861,15 +864,16 @@ namespace swrenderer
__m128i fgcolor = _mm_unpacklo_epi8(_mm_cvtsi32_si128(ifgcolor), _mm_setzero_si128()); __m128i fgcolor = _mm_unpacklo_epi8(_mm_cvtsi32_si128(ifgcolor), _mm_setzero_si128());
// Shade // Shade
int blue = BPART(ifgcolor); int blue0 = BPART(ifgcolor);
int green = GPART(ifgcolor); int green0 = GPART(ifgcolor);
int red = RPART(ifgcolor); int red0 = RPART(ifgcolor);
__m128i intensity = _mm_set1_epi16(((red * 77 + green * 143 + blue * 37) >> 8) * desaturate); int intensity0 = ((red0 * 77 + green0 * 143 + blue0 * 37) >> 8) * desaturate;
__m128i intensity = _mm_set_epi16(0, intensity0, intensity0, intensity0, 0, intensity0, intensity0, intensity0);
fgcolor = _mm_srli_epi16(_mm_add_epi16(_mm_mullo_epi16(fgcolor, inv_desaturate), intensity), 8); fgcolor = _mm_srli_epi16(_mm_add_epi16(_mm_mullo_epi16(fgcolor, inv_desaturate), intensity), 8);
fgcolor = _mm_mullo_epi16(fgcolor, _mm_set1_epi16(light)); fgcolor = _mm_mullo_epi16(fgcolor, mlight);
fgcolor = _mm_srli_epi16(_mm_add_epi16(_mm_mullo_epi16(shade_fade, inv_light), fgcolor), 8); fgcolor = _mm_srli_epi16(_mm_add_epi16(shade_fade, fgcolor), 8);
fgcolor = _mm_srli_epi16(_mm_mullo_epi16(fgcolor, shade_light), 8); fgcolor = _mm_srli_epi16(_mm_mullo_epi16(fgcolor, shade_light), 8);
// Blend // Blend
@ -911,8 +915,9 @@ namespace swrenderer
int light = 256 - (args.Light() >> (FRACBITS - 8)); int light = 256 - (args.Light() >> (FRACBITS - 8));
__m128i mlight = _mm_set_epi16(256, light, light, light, 256, light, light, light); __m128i mlight = _mm_set_epi16(256, light, light, light, 256, light, light, light);
__m128i inv_light = _mm_set_epi16(0, 256 - light, 256 - light, 256 - light, 0, 256 - light, 256 - light, 256 - light); __m128i inv_light = _mm_set_epi16(0, 256 - light, 256 - light, 256 - light, 0, 256 - light, 256 - light, 256 - light);
__m128i inv_desaturate = _mm_setr_epi16(0, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 0, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate); __m128i inv_desaturate = _mm_setr_epi16(256, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate);
__m128i shade_fade = _mm_set_epi16(shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue, shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue); __m128i shade_fade = _mm_set_epi16(shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue, shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue);
shade_fade = _mm_mullo_epi16(shade_fade, inv_light);
__m128i shade_light = _mm_set_epi16(shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue, shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue); __m128i shade_light = _mm_set_epi16(shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue, shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue);
int desaturate = shade_constants.desaturate; int desaturate = shade_constants.desaturate;
@ -930,7 +935,7 @@ namespace swrenderer
dest = thread->dest_for_thread(dest_y, pitch, dest); dest = thread->dest_for_thread(dest_y, pitch, dest);
fracstep *= thread->num_cores; fracstep *= thread->num_cores;
pitch *= thread->num_cores; pitch *= thread->num_cores;
frac -= one / 2;
__m128i srcalpha = _mm_set1_epi16(args.SrcAlpha()); __m128i srcalpha = _mm_set1_epi16(args.SrcAlpha());
__m128i destalpha = _mm_set1_epi16(args.DestAlpha()); __m128i destalpha = _mm_set1_epi16(args.DestAlpha());
@ -951,8 +956,8 @@ namespace swrenderer
unsigned int p11 = source2[y1]; unsigned int p11 = source2[y1];
unsigned int inv_b = texturefracx; unsigned int inv_b = texturefracx;
unsigned int a = (frac_y1 >> (FRACBITS - 4)) & 15; unsigned int inv_a = (frac_y1 >> (FRACBITS - 4)) & 15;
unsigned int inv_a = 16 - a; unsigned int a = 16 - inv_a;
unsigned int b = 16 - inv_b; unsigned int b = 16 - inv_b;
unsigned int sred = (RPART(p00) * (a * b) + RPART(p01) * (inv_a * b) + RPART(p10) * (a * inv_b) + RPART(p11) * (inv_a * inv_b) + 127) >> 8; unsigned int sred = (RPART(p00) * (a * b) + RPART(p01) * (inv_a * b) + RPART(p10) * (a * inv_b) + RPART(p11) * (inv_a * inv_b) + 127) >> 8;
@ -964,15 +969,16 @@ namespace swrenderer
__m128i fgcolor = _mm_unpacklo_epi8(_mm_cvtsi32_si128(ifgcolor), _mm_setzero_si128()); __m128i fgcolor = _mm_unpacklo_epi8(_mm_cvtsi32_si128(ifgcolor), _mm_setzero_si128());
// Shade // Shade
int blue = BPART(ifgcolor); int blue0 = BPART(ifgcolor);
int green = GPART(ifgcolor); int green0 = GPART(ifgcolor);
int red = RPART(ifgcolor); int red0 = RPART(ifgcolor);
__m128i intensity = _mm_set1_epi16(((red * 77 + green * 143 + blue * 37) >> 8) * desaturate); int intensity0 = ((red0 * 77 + green0 * 143 + blue0 * 37) >> 8) * desaturate;
__m128i intensity = _mm_set_epi16(0, intensity0, intensity0, intensity0, 0, intensity0, intensity0, intensity0);
fgcolor = _mm_srli_epi16(_mm_add_epi16(_mm_mullo_epi16(fgcolor, inv_desaturate), intensity), 8); fgcolor = _mm_srli_epi16(_mm_add_epi16(_mm_mullo_epi16(fgcolor, inv_desaturate), intensity), 8);
fgcolor = _mm_mullo_epi16(fgcolor, _mm_set1_epi16(light)); fgcolor = _mm_mullo_epi16(fgcolor, mlight);
fgcolor = _mm_srli_epi16(_mm_add_epi16(_mm_mullo_epi16(shade_fade, inv_light), fgcolor), 8); fgcolor = _mm_srli_epi16(_mm_add_epi16(shade_fade, fgcolor), 8);
fgcolor = _mm_srli_epi16(_mm_mullo_epi16(fgcolor, shade_light), 8); fgcolor = _mm_srli_epi16(_mm_mullo_epi16(fgcolor, shade_light), 8);
// Blend // Blend
@ -1051,7 +1057,6 @@ namespace swrenderer
dest = thread->dest_for_thread(dest_y, pitch, dest); dest = thread->dest_for_thread(dest_y, pitch, dest);
fracstep *= thread->num_cores; fracstep *= thread->num_cores;
pitch *= thread->num_cores; pitch *= thread->num_cores;
__m128i srcalpha = _mm_set1_epi16(args.SrcAlpha()); __m128i srcalpha = _mm_set1_epi16(args.SrcAlpha());
__m128i destalpha = _mm_set1_epi16(args.DestAlpha()); __m128i destalpha = _mm_set1_epi16(args.DestAlpha());
@ -1122,7 +1127,7 @@ namespace swrenderer
dest = thread->dest_for_thread(dest_y, pitch, dest); dest = thread->dest_for_thread(dest_y, pitch, dest);
fracstep *= thread->num_cores; fracstep *= thread->num_cores;
pitch *= thread->num_cores; pitch *= thread->num_cores;
frac -= one / 2;
__m128i srcalpha = _mm_set1_epi16(args.SrcAlpha()); __m128i srcalpha = _mm_set1_epi16(args.SrcAlpha());
__m128i destalpha = _mm_set1_epi16(args.DestAlpha()); __m128i destalpha = _mm_set1_epi16(args.DestAlpha());
@ -1143,8 +1148,8 @@ namespace swrenderer
unsigned int p11 = source2[y1]; unsigned int p11 = source2[y1];
unsigned int inv_b = texturefracx; unsigned int inv_b = texturefracx;
unsigned int a = (frac_y1 >> (FRACBITS - 4)) & 15; unsigned int inv_a = (frac_y1 >> (FRACBITS - 4)) & 15;
unsigned int inv_a = 16 - a; unsigned int a = 16 - inv_a;
unsigned int b = 16 - inv_b; unsigned int b = 16 - inv_b;
unsigned int sred = (RPART(p00) * (a * b) + RPART(p01) * (inv_a * b) + RPART(p10) * (a * inv_b) + RPART(p11) * (inv_a * inv_b) + 127) >> 8; unsigned int sred = (RPART(p00) * (a * b) + RPART(p01) * (inv_a * b) + RPART(p10) * (a * inv_b) + RPART(p11) * (inv_a * inv_b) + 127) >> 8;
@ -1203,8 +1208,9 @@ namespace swrenderer
int light = 256 - (args.Light() >> (FRACBITS - 8)); int light = 256 - (args.Light() >> (FRACBITS - 8));
__m128i mlight = _mm_set_epi16(256, light, light, light, 256, light, light, light); __m128i mlight = _mm_set_epi16(256, light, light, light, 256, light, light, light);
__m128i inv_light = _mm_set_epi16(0, 256 - light, 256 - light, 256 - light, 0, 256 - light, 256 - light, 256 - light); __m128i inv_light = _mm_set_epi16(0, 256 - light, 256 - light, 256 - light, 0, 256 - light, 256 - light, 256 - light);
__m128i inv_desaturate = _mm_setr_epi16(0, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 0, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate); __m128i inv_desaturate = _mm_setr_epi16(256, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate);
__m128i shade_fade = _mm_set_epi16(shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue, shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue); __m128i shade_fade = _mm_set_epi16(shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue, shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue);
shade_fade = _mm_mullo_epi16(shade_fade, inv_light);
__m128i shade_light = _mm_set_epi16(shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue, shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue); __m128i shade_light = _mm_set_epi16(shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue, shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue);
int desaturate = shade_constants.desaturate; int desaturate = shade_constants.desaturate;
@ -1222,7 +1228,6 @@ namespace swrenderer
dest = thread->dest_for_thread(dest_y, pitch, dest); dest = thread->dest_for_thread(dest_y, pitch, dest);
fracstep *= thread->num_cores; fracstep *= thread->num_cores;
pitch *= thread->num_cores; pitch *= thread->num_cores;
__m128i srcalpha = _mm_set1_epi16(args.SrcAlpha()); __m128i srcalpha = _mm_set1_epi16(args.SrcAlpha());
__m128i destalpha = _mm_set1_epi16(args.DestAlpha()); __m128i destalpha = _mm_set1_epi16(args.DestAlpha());
@ -1237,15 +1242,16 @@ namespace swrenderer
__m128i fgcolor = _mm_unpacklo_epi8(_mm_cvtsi32_si128(ifgcolor), _mm_setzero_si128()); __m128i fgcolor = _mm_unpacklo_epi8(_mm_cvtsi32_si128(ifgcolor), _mm_setzero_si128());
// Shade // Shade
int blue = BPART(ifgcolor); int blue0 = BPART(ifgcolor);
int green = GPART(ifgcolor); int green0 = GPART(ifgcolor);
int red = RPART(ifgcolor); int red0 = RPART(ifgcolor);
__m128i intensity = _mm_set1_epi16(((red * 77 + green * 143 + blue * 37) >> 8) * desaturate); int intensity0 = ((red0 * 77 + green0 * 143 + blue0 * 37) >> 8) * desaturate;
__m128i intensity = _mm_set_epi16(0, intensity0, intensity0, intensity0, 0, intensity0, intensity0, intensity0);
fgcolor = _mm_srli_epi16(_mm_add_epi16(_mm_mullo_epi16(fgcolor, inv_desaturate), intensity), 8); fgcolor = _mm_srli_epi16(_mm_add_epi16(_mm_mullo_epi16(fgcolor, inv_desaturate), intensity), 8);
fgcolor = _mm_mullo_epi16(fgcolor, _mm_set1_epi16(light)); fgcolor = _mm_mullo_epi16(fgcolor, mlight);
fgcolor = _mm_srli_epi16(_mm_add_epi16(_mm_mullo_epi16(shade_fade, inv_light), fgcolor), 8); fgcolor = _mm_srli_epi16(_mm_add_epi16(shade_fade, fgcolor), 8);
fgcolor = _mm_srli_epi16(_mm_mullo_epi16(fgcolor, shade_light), 8); fgcolor = _mm_srli_epi16(_mm_mullo_epi16(fgcolor, shade_light), 8);
// Blend // Blend
@ -1287,8 +1293,9 @@ namespace swrenderer
int light = 256 - (args.Light() >> (FRACBITS - 8)); int light = 256 - (args.Light() >> (FRACBITS - 8));
__m128i mlight = _mm_set_epi16(256, light, light, light, 256, light, light, light); __m128i mlight = _mm_set_epi16(256, light, light, light, 256, light, light, light);
__m128i inv_light = _mm_set_epi16(0, 256 - light, 256 - light, 256 - light, 0, 256 - light, 256 - light, 256 - light); __m128i inv_light = _mm_set_epi16(0, 256 - light, 256 - light, 256 - light, 0, 256 - light, 256 - light, 256 - light);
__m128i inv_desaturate = _mm_setr_epi16(0, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 0, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate); __m128i inv_desaturate = _mm_setr_epi16(256, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate);
__m128i shade_fade = _mm_set_epi16(shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue, shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue); __m128i shade_fade = _mm_set_epi16(shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue, shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue);
shade_fade = _mm_mullo_epi16(shade_fade, inv_light);
__m128i shade_light = _mm_set_epi16(shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue, shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue); __m128i shade_light = _mm_set_epi16(shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue, shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue);
int desaturate = shade_constants.desaturate; int desaturate = shade_constants.desaturate;
@ -1306,7 +1313,7 @@ namespace swrenderer
dest = thread->dest_for_thread(dest_y, pitch, dest); dest = thread->dest_for_thread(dest_y, pitch, dest);
fracstep *= thread->num_cores; fracstep *= thread->num_cores;
pitch *= thread->num_cores; pitch *= thread->num_cores;
frac -= one / 2;
__m128i srcalpha = _mm_set1_epi16(args.SrcAlpha()); __m128i srcalpha = _mm_set1_epi16(args.SrcAlpha());
__m128i destalpha = _mm_set1_epi16(args.DestAlpha()); __m128i destalpha = _mm_set1_epi16(args.DestAlpha());
@ -1327,8 +1334,8 @@ namespace swrenderer
unsigned int p11 = source2[y1]; unsigned int p11 = source2[y1];
unsigned int inv_b = texturefracx; unsigned int inv_b = texturefracx;
unsigned int a = (frac_y1 >> (FRACBITS - 4)) & 15; unsigned int inv_a = (frac_y1 >> (FRACBITS - 4)) & 15;
unsigned int inv_a = 16 - a; unsigned int a = 16 - inv_a;
unsigned int b = 16 - inv_b; unsigned int b = 16 - inv_b;
unsigned int sred = (RPART(p00) * (a * b) + RPART(p01) * (inv_a * b) + RPART(p10) * (a * inv_b) + RPART(p11) * (inv_a * inv_b) + 127) >> 8; unsigned int sred = (RPART(p00) * (a * b) + RPART(p01) * (inv_a * b) + RPART(p10) * (a * inv_b) + RPART(p11) * (inv_a * inv_b) + 127) >> 8;
@ -1340,15 +1347,16 @@ namespace swrenderer
__m128i fgcolor = _mm_unpacklo_epi8(_mm_cvtsi32_si128(ifgcolor), _mm_setzero_si128()); __m128i fgcolor = _mm_unpacklo_epi8(_mm_cvtsi32_si128(ifgcolor), _mm_setzero_si128());
// Shade // Shade
int blue = BPART(ifgcolor); int blue0 = BPART(ifgcolor);
int green = GPART(ifgcolor); int green0 = GPART(ifgcolor);
int red = RPART(ifgcolor); int red0 = RPART(ifgcolor);
__m128i intensity = _mm_set1_epi16(((red * 77 + green * 143 + blue * 37) >> 8) * desaturate); int intensity0 = ((red0 * 77 + green0 * 143 + blue0 * 37) >> 8) * desaturate;
__m128i intensity = _mm_set_epi16(0, intensity0, intensity0, intensity0, 0, intensity0, intensity0, intensity0);
fgcolor = _mm_srli_epi16(_mm_add_epi16(_mm_mullo_epi16(fgcolor, inv_desaturate), intensity), 8); fgcolor = _mm_srli_epi16(_mm_add_epi16(_mm_mullo_epi16(fgcolor, inv_desaturate), intensity), 8);
fgcolor = _mm_mullo_epi16(fgcolor, _mm_set1_epi16(light)); fgcolor = _mm_mullo_epi16(fgcolor, mlight);
fgcolor = _mm_srli_epi16(_mm_add_epi16(_mm_mullo_epi16(shade_fade, inv_light), fgcolor), 8); fgcolor = _mm_srli_epi16(_mm_add_epi16(shade_fade, fgcolor), 8);
fgcolor = _mm_srli_epi16(_mm_mullo_epi16(fgcolor, shade_light), 8); fgcolor = _mm_srli_epi16(_mm_mullo_epi16(fgcolor, shade_light), 8);
// Blend // Blend
@ -1427,7 +1435,6 @@ namespace swrenderer
dest = thread->dest_for_thread(dest_y, pitch, dest); dest = thread->dest_for_thread(dest_y, pitch, dest);
fracstep *= thread->num_cores; fracstep *= thread->num_cores;
pitch *= thread->num_cores; pitch *= thread->num_cores;
__m128i srcalpha = _mm_set1_epi16(args.SrcAlpha()); __m128i srcalpha = _mm_set1_epi16(args.SrcAlpha());
__m128i destalpha = _mm_set1_epi16(args.DestAlpha()); __m128i destalpha = _mm_set1_epi16(args.DestAlpha());
@ -1498,7 +1505,7 @@ namespace swrenderer
dest = thread->dest_for_thread(dest_y, pitch, dest); dest = thread->dest_for_thread(dest_y, pitch, dest);
fracstep *= thread->num_cores; fracstep *= thread->num_cores;
pitch *= thread->num_cores; pitch *= thread->num_cores;
frac -= one / 2;
__m128i srcalpha = _mm_set1_epi16(args.SrcAlpha()); __m128i srcalpha = _mm_set1_epi16(args.SrcAlpha());
__m128i destalpha = _mm_set1_epi16(args.DestAlpha()); __m128i destalpha = _mm_set1_epi16(args.DestAlpha());
@ -1519,8 +1526,8 @@ namespace swrenderer
unsigned int p11 = source2[y1]; unsigned int p11 = source2[y1];
unsigned int inv_b = texturefracx; unsigned int inv_b = texturefracx;
unsigned int a = (frac_y1 >> (FRACBITS - 4)) & 15; unsigned int inv_a = (frac_y1 >> (FRACBITS - 4)) & 15;
unsigned int inv_a = 16 - a; unsigned int a = 16 - inv_a;
unsigned int b = 16 - inv_b; unsigned int b = 16 - inv_b;
unsigned int sred = (RPART(p00) * (a * b) + RPART(p01) * (inv_a * b) + RPART(p10) * (a * inv_b) + RPART(p11) * (inv_a * inv_b) + 127) >> 8; unsigned int sred = (RPART(p00) * (a * b) + RPART(p01) * (inv_a * b) + RPART(p10) * (a * inv_b) + RPART(p11) * (inv_a * inv_b) + 127) >> 8;
@ -1579,8 +1586,9 @@ namespace swrenderer
int light = 256 - (args.Light() >> (FRACBITS - 8)); int light = 256 - (args.Light() >> (FRACBITS - 8));
__m128i mlight = _mm_set_epi16(256, light, light, light, 256, light, light, light); __m128i mlight = _mm_set_epi16(256, light, light, light, 256, light, light, light);
__m128i inv_light = _mm_set_epi16(0, 256 - light, 256 - light, 256 - light, 0, 256 - light, 256 - light, 256 - light); __m128i inv_light = _mm_set_epi16(0, 256 - light, 256 - light, 256 - light, 0, 256 - light, 256 - light, 256 - light);
__m128i inv_desaturate = _mm_setr_epi16(0, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 0, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate); __m128i inv_desaturate = _mm_setr_epi16(256, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate);
__m128i shade_fade = _mm_set_epi16(shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue, shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue); __m128i shade_fade = _mm_set_epi16(shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue, shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue);
shade_fade = _mm_mullo_epi16(shade_fade, inv_light);
__m128i shade_light = _mm_set_epi16(shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue, shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue); __m128i shade_light = _mm_set_epi16(shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue, shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue);
int desaturate = shade_constants.desaturate; int desaturate = shade_constants.desaturate;
@ -1598,7 +1606,6 @@ namespace swrenderer
dest = thread->dest_for_thread(dest_y, pitch, dest); dest = thread->dest_for_thread(dest_y, pitch, dest);
fracstep *= thread->num_cores; fracstep *= thread->num_cores;
pitch *= thread->num_cores; pitch *= thread->num_cores;
__m128i srcalpha = _mm_set1_epi16(args.SrcAlpha()); __m128i srcalpha = _mm_set1_epi16(args.SrcAlpha());
__m128i destalpha = _mm_set1_epi16(args.DestAlpha()); __m128i destalpha = _mm_set1_epi16(args.DestAlpha());
@ -1613,15 +1620,16 @@ namespace swrenderer
__m128i fgcolor = _mm_unpacklo_epi8(_mm_cvtsi32_si128(ifgcolor), _mm_setzero_si128()); __m128i fgcolor = _mm_unpacklo_epi8(_mm_cvtsi32_si128(ifgcolor), _mm_setzero_si128());
// Shade // Shade
int blue = BPART(ifgcolor); int blue0 = BPART(ifgcolor);
int green = GPART(ifgcolor); int green0 = GPART(ifgcolor);
int red = RPART(ifgcolor); int red0 = RPART(ifgcolor);
__m128i intensity = _mm_set1_epi16(((red * 77 + green * 143 + blue * 37) >> 8) * desaturate); int intensity0 = ((red0 * 77 + green0 * 143 + blue0 * 37) >> 8) * desaturate;
__m128i intensity = _mm_set_epi16(0, intensity0, intensity0, intensity0, 0, intensity0, intensity0, intensity0);
fgcolor = _mm_srli_epi16(_mm_add_epi16(_mm_mullo_epi16(fgcolor, inv_desaturate), intensity), 8); fgcolor = _mm_srli_epi16(_mm_add_epi16(_mm_mullo_epi16(fgcolor, inv_desaturate), intensity), 8);
fgcolor = _mm_mullo_epi16(fgcolor, _mm_set1_epi16(light)); fgcolor = _mm_mullo_epi16(fgcolor, mlight);
fgcolor = _mm_srli_epi16(_mm_add_epi16(_mm_mullo_epi16(shade_fade, inv_light), fgcolor), 8); fgcolor = _mm_srli_epi16(_mm_add_epi16(shade_fade, fgcolor), 8);
fgcolor = _mm_srli_epi16(_mm_mullo_epi16(fgcolor, shade_light), 8); fgcolor = _mm_srli_epi16(_mm_mullo_epi16(fgcolor, shade_light), 8);
// Blend // Blend
@ -1663,8 +1671,9 @@ namespace swrenderer
int light = 256 - (args.Light() >> (FRACBITS - 8)); int light = 256 - (args.Light() >> (FRACBITS - 8));
__m128i mlight = _mm_set_epi16(256, light, light, light, 256, light, light, light); __m128i mlight = _mm_set_epi16(256, light, light, light, 256, light, light, light);
__m128i inv_light = _mm_set_epi16(0, 256 - light, 256 - light, 256 - light, 0, 256 - light, 256 - light, 256 - light); __m128i inv_light = _mm_set_epi16(0, 256 - light, 256 - light, 256 - light, 0, 256 - light, 256 - light, 256 - light);
__m128i inv_desaturate = _mm_setr_epi16(0, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 0, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate); __m128i inv_desaturate = _mm_setr_epi16(256, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate);
__m128i shade_fade = _mm_set_epi16(shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue, shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue); __m128i shade_fade = _mm_set_epi16(shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue, shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue);
shade_fade = _mm_mullo_epi16(shade_fade, inv_light);
__m128i shade_light = _mm_set_epi16(shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue, shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue); __m128i shade_light = _mm_set_epi16(shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue, shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue);
int desaturate = shade_constants.desaturate; int desaturate = shade_constants.desaturate;
@ -1682,7 +1691,7 @@ namespace swrenderer
dest = thread->dest_for_thread(dest_y, pitch, dest); dest = thread->dest_for_thread(dest_y, pitch, dest);
fracstep *= thread->num_cores; fracstep *= thread->num_cores;
pitch *= thread->num_cores; pitch *= thread->num_cores;
frac -= one / 2;
__m128i srcalpha = _mm_set1_epi16(args.SrcAlpha()); __m128i srcalpha = _mm_set1_epi16(args.SrcAlpha());
__m128i destalpha = _mm_set1_epi16(args.DestAlpha()); __m128i destalpha = _mm_set1_epi16(args.DestAlpha());
@ -1703,8 +1712,8 @@ namespace swrenderer
unsigned int p11 = source2[y1]; unsigned int p11 = source2[y1];
unsigned int inv_b = texturefracx; unsigned int inv_b = texturefracx;
unsigned int a = (frac_y1 >> (FRACBITS - 4)) & 15; unsigned int inv_a = (frac_y1 >> (FRACBITS - 4)) & 15;
unsigned int inv_a = 16 - a; unsigned int a = 16 - inv_a;
unsigned int b = 16 - inv_b; unsigned int b = 16 - inv_b;
unsigned int sred = (RPART(p00) * (a * b) + RPART(p01) * (inv_a * b) + RPART(p10) * (a * inv_b) + RPART(p11) * (inv_a * inv_b) + 127) >> 8; unsigned int sred = (RPART(p00) * (a * b) + RPART(p01) * (inv_a * b) + RPART(p10) * (a * inv_b) + RPART(p11) * (inv_a * inv_b) + 127) >> 8;
@ -1716,15 +1725,16 @@ namespace swrenderer
__m128i fgcolor = _mm_unpacklo_epi8(_mm_cvtsi32_si128(ifgcolor), _mm_setzero_si128()); __m128i fgcolor = _mm_unpacklo_epi8(_mm_cvtsi32_si128(ifgcolor), _mm_setzero_si128());
// Shade // Shade
int blue = BPART(ifgcolor); int blue0 = BPART(ifgcolor);
int green = GPART(ifgcolor); int green0 = GPART(ifgcolor);
int red = RPART(ifgcolor); int red0 = RPART(ifgcolor);
__m128i intensity = _mm_set1_epi16(((red * 77 + green * 143 + blue * 37) >> 8) * desaturate); int intensity0 = ((red0 * 77 + green0 * 143 + blue0 * 37) >> 8) * desaturate;
__m128i intensity = _mm_set_epi16(0, intensity0, intensity0, intensity0, 0, intensity0, intensity0, intensity0);
fgcolor = _mm_srli_epi16(_mm_add_epi16(_mm_mullo_epi16(fgcolor, inv_desaturate), intensity), 8); fgcolor = _mm_srli_epi16(_mm_add_epi16(_mm_mullo_epi16(fgcolor, inv_desaturate), intensity), 8);
fgcolor = _mm_mullo_epi16(fgcolor, _mm_set1_epi16(light)); fgcolor = _mm_mullo_epi16(fgcolor, mlight);
fgcolor = _mm_srli_epi16(_mm_add_epi16(_mm_mullo_epi16(shade_fade, inv_light), fgcolor), 8); fgcolor = _mm_srli_epi16(_mm_add_epi16(shade_fade, fgcolor), 8);
fgcolor = _mm_srli_epi16(_mm_mullo_epi16(fgcolor, shade_light), 8); fgcolor = _mm_srli_epi16(_mm_mullo_epi16(fgcolor, shade_light), 8);
// Blend // Blend

25
src/swrenderer/drawers/r_draw_wall32.php
View file

@ -95,8 +95,9 @@ namespace swrenderer
__m128i inv_light = _mm_set_epi16(0, 256 - light, 256 - light, 256 - light, 0, 256 - light, 256 - light, 256 - light); __m128i inv_light = _mm_set_epi16(0, 256 - light, 256 - light, 256 - light, 0, 256 - light, 256 - light, 256 - light);
<? if ($isSimpleShade == false) <? if ($isSimpleShade == false)
{ ?> { ?>
__m128i inv_desaturate = _mm_setr_epi16(0, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 0, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate); __m128i inv_desaturate = _mm_setr_epi16(256, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate, 256 - shade_constants.desaturate);
__m128i shade_fade = _mm_set_epi16(shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue, shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue); __m128i shade_fade = _mm_set_epi16(shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue, shade_constants.fade_alpha, shade_constants.fade_red, shade_constants.fade_green, shade_constants.fade_blue);
shade_fade = _mm_mullo_epi16(shade_fade, inv_light);
__m128i shade_light = _mm_set_epi16(shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue, shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue); __m128i shade_light = _mm_set_epi16(shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue, shade_constants.light_alpha, shade_constants.light_red, shade_constants.light_green, shade_constants.light_blue);
int desaturate = shade_constants.desaturate; int desaturate = shade_constants.desaturate;
<? } ?> <? } ?>
@ -115,7 +116,10 @@ namespace swrenderer
dest = thread->dest_for_thread(dest_y, pitch, dest); dest = thread->dest_for_thread(dest_y, pitch, dest);
fracstep *= thread->num_cores; fracstep *= thread->num_cores;
pitch *= thread->num_cores; pitch *= thread->num_cores;
<? if ($isNearestFilter == false)
{ ?>
frac -= one / 2;
<? } ?>
__m128i srcalpha = _mm_set1_epi16(args.SrcAlpha()); __m128i srcalpha = _mm_set1_epi16(args.SrcAlpha());
__m128i destalpha = _mm_set1_epi16(args.DestAlpha()); __m128i destalpha = _mm_set1_epi16(args.DestAlpha());
@ -159,8 +163,8 @@ namespace swrenderer
unsigned int p11 = source2[y1]; unsigned int p11 = source2[y1];
unsigned int inv_b = texturefracx; unsigned int inv_b = texturefracx;
unsigned int a = (frac_y1 >> (FRACBITS - 4)) & 15; unsigned int inv_a = (frac_y1 >> (FRACBITS - 4)) & 15;
unsigned int inv_a = 16 - a; unsigned int a = 16 - inv_a;
unsigned int b = 16 - inv_b; unsigned int b = 16 - inv_b;
unsigned int sred = (RPART(p00) * (a * b) + RPART(p01) * (inv_a * b) + RPART(p10) * (a * inv_b) + RPART(p11) * (inv_a * inv_b) + 127) >> 8; unsigned int sred = (RPART(p00) * (a * b) + RPART(p01) * (inv_a * b) + RPART(p10) * (a * inv_b) + RPART(p11) * (inv_a * inv_b) + 127) >> 8;
@ -181,15 +185,16 @@ namespace swrenderer
<? } <? }
else else
{ ?> { ?>
int blue = BPART(ifgcolor); int blue0 = BPART(ifgcolor);
int green = GPART(ifgcolor); int green0 = GPART(ifgcolor);
int red = RPART(ifgcolor); int red0 = RPART(ifgcolor);
__m128i intensity = _mm_set1_epi16(((red * 77 + green * 143 + blue * 37) >> 8) * desaturate); int intensity0 = ((red0 * 77 + green0 * 143 + blue0 * 37) >> 8) * desaturate;
__m128i intensity = _mm_set_epi16(0, intensity0, intensity0, intensity0, 0, intensity0, intensity0, intensity0);
fgcolor = _mm_srli_epi16(_mm_add_epi16(_mm_mullo_epi16(fgcolor, inv_desaturate), intensity), 8); fgcolor = _mm_srli_epi16(_mm_add_epi16(_mm_mullo_epi16(fgcolor, inv_desaturate), intensity), 8);
fgcolor = _mm_mullo_epi16(fgcolor, _mm_set1_epi16(light)); fgcolor = _mm_mullo_epi16(fgcolor, mlight);
fgcolor = _mm_srli_epi16(_mm_add_epi16(_mm_mullo_epi16(shade_fade, inv_light), fgcolor), 8); fgcolor = _mm_srli_epi16(_mm_add_epi16(shade_fade, fgcolor), 8);
fgcolor = _mm_srli_epi16(_mm_mullo_epi16(fgcolor, shade_light), 8); fgcolor = _mm_srli_epi16(_mm_mullo_epi16(fgcolor, shade_light), 8);
<? } <? }
} }