ShuriZma
/
dolphin
mirror of https://github.com/dolphin-emu/dolphin.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Refactor all the SSSE3 functions in TextureDecoder so that the cpu_info check isn't looped over. Speeds up most textures dramatically (where it has previously slowed them). 

git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@6784 8ced0084-cf51-0410-be5f-012b33b47a6e
This commit is contained in:
xsacha 2011-01-08 09:08:36 +00:00
parent 394534814b
commit 5c725262ba
1 changed files with 139 additions and 100 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

239
Source/Core/VideoCommon/Src/TextureDecoder.cpp
View File

@ -957,27 +957,21 @@ PC_TexFormat TexDecoder_Decode_RGBA(u32 * dst, const u8 * src, int width, int he
break;
case GX_TF_I4:
{
// JSD optimized with SSE2 intrinsics.
// Produces a ~76% speed increase over reference C implementation.
const __m128i kMask_x0f = _mm_set_epi32(0x0f0f0f0fL, 0x0f0f0f0fL, 0x0f0f0f0fL, 0x0f0f0f0fL);
const __m128i kMask_xf0 = _mm_set_epi32(0xf0f0f0f0L, 0xf0f0f0f0L, 0xf0f0f0f0L, 0xf0f0f0f0L);
for (int y = 0; y < height; y += 8)
for (int x = 0; x < width; x += 8)
for (int iy = 0; iy < 8; iy += 2, src += 8)
{
// Expand [BA] to [BB][BB][BB][BB] [AA][AA][AA][AA], where [BA] is a single byte and A and B are 4-bit values.
// Load 64 bits with upper 64 bits zeroed: (0000 0000 hgfe dcba)
// dcba is row #0 and hgfe is row #1. We process two rows at once with each loop iteration, hence iy += 2.
const __m128i r0 = _mm_loadl_epi64((const __m128i *)src);
__m128i o1, o2, o3, o4;
#if _M_SSE >= 0x301
if (cpu_info.bSSSE3) {
const __m128i mask9180 = _mm_set_epi8(9,9,9,9,1,1,1,1,8,8,8,8,0,0,0,0);
const __m128i maskB3A2 = _mm_set_epi8(11,11,11,11,3,3,3,3,10,10,10,10,2,2,2,2);
const __m128i maskD5C4 = _mm_set_epi8(13,13,13,13,5,5,5,5,12,12,12,12,4,4,4,4);
const __m128i maskF7E6 = _mm_set_epi8(15,15,15,15,7,7,7,7,14,14,14,14,6,6,6,6);
// xsacha optimized with SSSE3 intrinsics
// Produces a ~40% speed improvement over SSE2 implementation
if (cpu_info.bSSSE3) {
const __m128i mask9180 = _mm_set_epi8(9,9,9,9,1,1,1,1,8,8,8,8,0,0,0,0);
const __m128i maskB3A2 = _mm_set_epi8(11,11,11,11,3,3,3,3,10,10,10,10,2,2,2,2);
const __m128i maskD5C4 = _mm_set_epi8(13,13,13,13,5,5,5,5,12,12,12,12,4,4,4,4);
const __m128i maskF7E6 = _mm_set_epi8(15,15,15,15,7,7,7,7,14,14,14,14,6,6,6,6);
for (int y = 0; y < height; y += 8)
for (int x = 0; x < width; x += 8)
for (int iy = 0; iy < 8; iy += 2, src += 8)
{
const __m128i r0 = _mm_loadl_epi64((const __m128i *)src);
// We want the hi 4 bits of each 8-bit word replicated to 32-bit words:
// (00000000 00000000 HhGgFfEe DdCcBbAa) -> (00000000 00000000 HHGGFFEE DDCCBBAA)
const __m128i i1 = _mm_and_si128(r0, kMask_xf0);
@ -994,9 +988,24 @@ PC_TexFormat TexDecoder_Decode_RGBA(u32 * dst, const u8 * src, int width, int he
const __m128i o2 = _mm_shuffle_epi8(base, maskB3A2);
const __m128i o3 = _mm_shuffle_epi8(base, maskD5C4);
const __m128i o4 = _mm_shuffle_epi8(base, maskF7E6);
} else
// Write row 0:
_mm_storeu_si128( (__m128i*)( dst+(y + iy) * width + x ), o1 );
_mm_storeu_si128( (__m128i*)( dst+(y + iy) * width + x + 4 ), o2 );
// Write row 1:
_mm_storeu_si128( (__m128i*)( dst+(y + iy+1) * width + x ), o3 );
_mm_storeu_si128( (__m128i*)( dst+(y + iy+1) * width + x + 4 ), o4 );
}
} else
#endif
// JSD optimized with SSE2 intrinsics.
// Produces a ~76% speed improvement over reference C implementation.
{
for (int y = 0; y < height; y += 8)
for (int x = 0; x < width; x += 8)
for (int iy = 0; iy < 8; iy += 2, src += 8)
{
const __m128i r0 = _mm_loadl_epi64((const __m128i *)src);
// Shuffle low 64-bits with itself to expand from (0000 0000 hgfe dcba) to (hhgg ffee ddcc bbaa)
const __m128i r1 = _mm_unpacklo_epi8(r0, r0);
@ -1040,20 +1049,20 @@ PC_TexFormat TexDecoder_Decode_RGBA(u32 * dst, const u8 * src, int width, int he
// (00000000 BBBBBBBB 00000000 AAAAAAAA) | (bbbbbbbb 00000000 aaaaaaaa 00000000) -> (bbbbbbbb BBBBBBBB aaaaaaaa AAAAAAAA)
const __m128i kMask_x00000000ffffffff = _mm_set_epi32(0x00000000L, 0xffffffffL, 0x00000000L, 0xffffffffL);
const __m128i kMask_xffffffff00000000 = _mm_set_epi32(0xffffffffL, 0x00000000L, 0xffffffffL, 0x00000000L);
o1 = _mm_or_si128(_mm_and_si128(i151, kMask_x00000000ffffffff), _mm_and_si128(i251, kMask_xffffffff00000000));
o2 = _mm_or_si128(_mm_and_si128(i152, kMask_x00000000ffffffff), _mm_and_si128(i252, kMask_xffffffff00000000));
const __m128i o1 = _mm_or_si128(_mm_and_si128(i151, kMask_x00000000ffffffff), _mm_and_si128(i251, kMask_xffffffff00000000));
const __m128i o2 = _mm_or_si128(_mm_and_si128(i152, kMask_x00000000ffffffff), _mm_and_si128(i252, kMask_xffffffff00000000));
// These two are for the next row; same pattern as above. We batched up two rows because our input was 64 bits.
o3 = _mm_or_si128(_mm_and_si128(i161, kMask_x00000000ffffffff), _mm_and_si128(i261, kMask_xffffffff00000000));
o4 = _mm_or_si128(_mm_and_si128(i162, kMask_x00000000ffffffff), _mm_and_si128(i262, kMask_xffffffff00000000));
const __m128i o3 = _mm_or_si128(_mm_and_si128(i161, kMask_x00000000ffffffff), _mm_and_si128(i261, kMask_xffffffff00000000));
const __m128i o4 = _mm_or_si128(_mm_and_si128(i162, kMask_x00000000ffffffff), _mm_and_si128(i262, kMask_xffffffff00000000));
// Write row 0:
_mm_storeu_si128( (__m128i*)( dst+(y + iy) * width + x ), o1 );
_mm_storeu_si128( (__m128i*)( dst+(y + iy) * width + x + 4 ), o2 );
// Write row 1:
_mm_storeu_si128( (__m128i*)( dst+(y + iy+1) * width + x ), o3 );
_mm_storeu_si128( (__m128i*)( dst+(y + iy+1) * width + x + 4 ), o4 );
}
// Write row 0:
_mm_storeu_si128( (__m128i*)( dst+(y + iy) * width + x ), o1 );
_mm_storeu_si128( (__m128i*)( dst+(y + iy) * width + x + 4 ), o2 );
// Write row 1:
_mm_storeu_si128( (__m128i*)( dst+(y + iy+1) * width + x ), o3 );
_mm_storeu_si128( (__m128i*)( dst+(y + iy+1) * width + x + 4 ), o4 );
}
}
#if 0
// Reference C implementation:
for (int y = 0; y < height; y += 8)
@ -1072,13 +1081,14 @@ PC_TexFormat TexDecoder_Decode_RGBA(u32 * dst, const u8 * src, int width, int he
break;
case GX_TF_I8: // speed critical
{
for (int y = 0; y < height; y += 4)
for (int x = 0; x < width; x += 8)
{
#if _M_SSE >= 0x301
if (cpu_info.bSSSE3)
if (cpu_info.bSSSE3)
{
for (int y = 0; y < height; y += 4)
for (int x = 0; x < width; x += 8)
{
// SSSE3 intrinsics: About 5-10% faster than SSE2 version
// xsacha optimized with SSSE3 intrinsics
// Produces a ~10% speed improvement over SSE2 implementation
for (int iy = 0; iy < 4; ++iy, src+=8)
{
const __m128i mask3210 = _mm_set_epi8(3, 3, 3, 3, 2, 2, 2, 2, 1, 1, 1, 1, 0, 0, 0, 0);
@ -1095,11 +1105,15 @@ PC_TexFormat TexDecoder_Decode_RGBA(u32 * dst, const u8 * src, int width, int he
_mm_storeu_si128(quaddst, rgba0);
_mm_storeu_si128(quaddst+1, rgba1);
}
} else
}
} else
#endif
// JSD optimized with SSE2 intrinsics.
// Produces an ~86% speed improvement over reference C implementation.
{
for (int y = 0; y < height; y += 4)
for (int x = 0; x < width; x += 8)
{
// JSD optimized with SSE2 intrinsics.
// Produces an ~86% speed increase over reference C implementation.
// Each loop iteration processes 4 rows from 4 64-bit reads.
// TODO: is it more efficient to group the loads together sequentially and also the stores at the end?
@ -1177,7 +1191,7 @@ PC_TexFormat TexDecoder_Decode_RGBA(u32 * dst, const u8 * src, int width, int he
src += 8;
}
}
}
#if 0
// Reference C implementation
for (int y = 0; y < height; y += 4)
@ -1238,7 +1252,8 @@ PC_TexFormat TexDecoder_Decode_RGBA(u32 * dst, const u8 * src, int width, int he
case GX_TF_IA8:
{
#if _M_SSE >= 0x301
// SSSE3 implementation is approximately 50% faster than SSE2 version.
// xsacha optimized with SSSE3 intrinsics.
// Produces an ~50% speed improvement over SSE2 implementation.
if (cpu_info.bSSSE3)
{
for (int y = 0; y < height; y += 4)
@ -1254,9 +1269,9 @@ PC_TexFormat TexDecoder_Decode_RGBA(u32 * dst, const u8 * src, int width, int he
}
} else
#endif
// JSD optimized with SSE2 intrinsics.
// Produces an ~80% speed improvement over reference C implementation.
{
// JSD optimized with SSE2 intrinsics.
// Produces an ~80% speed improvement over reference C implementation.
const __m128i kMask_xf0 = _mm_set_epi32(0x00000000L, 0x00000000L, 0xff00ff00L, 0xff00ff00L);
const __m128i kMask_x0f = _mm_set_epi32(0x00000000L, 0x00000000L, 0x00ff00ffL, 0x00ff00ffL);
const __m128i kMask_xf000 = _mm_set_epi32(0xff000000L, 0xff000000L, 0xff000000L, 0xff000000L);
@ -1431,25 +1446,24 @@ PC_TexFormat TexDecoder_Decode_RGBA(u32 * dst, const u8 * src, int width, int he
break;
case GX_TF_RGB5A3:
{
// JSD optimized with SSE2 intrinsics in 2 out of 4 cases.
// Produces a ~25% speed improvement over reference C implementation.
const __m128i kMask_x1f = _mm_set_epi32(0x0000001fL, 0x0000001fL, 0x0000001fL, 0x0000001fL);
const __m128i kMask_x0f = _mm_set_epi32(0x0000000fL, 0x0000000fL, 0x0000000fL, 0x0000000fL);
const __m128i kMask_x07 = _mm_set_epi32(0x00000007L, 0x00000007L, 0x00000007L, 0x00000007L);
// This is the hard-coded 0xFF alpha constant that is ORed in place after the RGB are calculated
// for the RGB555 case when (s[x] & 0x8000) is true for all pixels.
const __m128i aVxff00 = _mm_set_epi32(0xFF000000L, 0xFF000000L, 0xFF000000L, 0xFF000000L);
for (int y = 0; y < height; y += 4)
// xsacha optimized with SSSE3 intrinsics (2 in 4 cases)
// Produces a ~18% speed improvement over SSE2 implementation
#if _M_SSE >= 0x301
if (cpu_info.bSSSE3)
{
for (int y = 0; y < height; y += 4)
for (int x = 0; x < width; x += 4)
for (int iy = 0; iy < 4; iy++, src += 8)
{
u32 *newdst = dst+(y+iy)*width+x;
#if _M_SSE >= 0x301
// Produces a ~40% speed improvement over reference C implementation
if (cpu_info.bSSSE3)
{
const __m128i mask = _mm_set_epi8(128,128,6,7,128,128,4,5,128,128,2,3,128,128,0,1);
const __m128i mask = _mm_set_epi8(128,128,6,7,128,128,4,5,128,128,2,3,128,128,0,1);
const __m128i valV = _mm_shuffle_epi8(_mm_loadl_epi64((const __m128i*)src),mask);
int cmp = _mm_movemask_epi8(valV); //MSB: 0x2 = val0; 0x20=val1; 0x200 = val2; 0x2000=val3
if ((cmp&0x2222)==0x2222) // SSSE3 case #1: all 4 pixels are in RGB555 and alpha = 0xFF.
@ -1529,9 +1543,17 @@ PC_TexFormat TexDecoder_Decode_RGBA(u32 * dst, const u8 * src, int width, int he
newdst[i] = r | (g << 8) | (b << 16) | (a << 24);
}
}
} else
}
} else
#endif
// JSD optimized with SSE2 intrinsics (2 in 4 cases)
// Produces a ~25% speed improvement over reference C implementation.
{
for (int y = 0; y < height; y += 4)
for (int x = 0; x < width; x += 4)
for (int iy = 0; iy < 4; iy++, src += 8)
{
u32 *newdst = dst+(y+iy)*width+x;
const u16 *newsrc = (const u16*)src;
// TODO: weak point
@ -1686,7 +1708,7 @@ PC_TexFormat TexDecoder_Decode_RGBA(u32 * dst, const u8 * src, int width, int he
newdst[3] = r3 | (g3 << 8) | (b3 << 16) | (a3 << 24);
}
}
}
}
#if 0
// Reference C implementation:
for (int y = 0; y < height; y += 4)
@ -1698,43 +1720,19 @@ PC_TexFormat TexDecoder_Decode_RGBA(u32 * dst, const u8 * src, int width, int he
break;
case GX_TF_RGBA8: // speed critical
{
// JSD optimized with SSE2 intrinsics.
// Produces a ~68% improvement in speed over reference C implementation.
for (int y = 0; y < height; y += 4)
for (int x = 0; x < width; x += 4, src += 64)
{
// Input is divided up into 16-bit words. The texels are split up into AR and GB components where all
// AR components come grouped up first in 32 bytes followed by the GB components in 32 bytes. We are
// processing 16 texels per each loop iteration, numbered from 0-f.
//
// Convention is:
// one byte is [component-name texel-number]
// __m128i is (4-bytes 4-bytes 4-bytes 4-bytes)
//
// Input is ([A 7][R 7][A 6][R 6] [A 5][R 5][A 4][R 4] [A 3][R 3][A 2][R 2] [A 1][R 1][A 0][R 0])
// ([A f][R f][A e][R e] [A d][R d][A c][R c] [A b][R b][A a][R a] [A 9][R 9][A 8][R 8])
// ([G 7][B 7][G 6][B 6] [G 5][B 5][G 4][B 4] [G 3][B 3][G 2][B 2] [G 1][B 1][G 0][B 0])
// ([G f][B f][G e][B e] [G d][B d][G c][B c] [G b][B b][G a][B a] [G 9][B 9][G 8][B 8])
//
// Output is (RGBA3 RGBA2 RGBA1 RGBA0)
// (RGBA7 RGBA6 RGBA5 RGBA4)
// (RGBAb RGBAa RGBA9 RGBA8)
// (RGBAf RGBAe RGBAd RGBAc)
// Loads the 1st half of AR components ([A 7][R 7][A 6][R 6] [A 5][R 5][A 4][R 4] [A 3][R 3][A 2][R 2] [A 1][R 1][A 0][R 0])
const __m128i ar0 = _mm_loadu_si128((__m128i*)src);
// Loads the 2nd half of AR components ([A f][R f][A e][R e] [A d][R d][A c][R c] [A b][R b][A a][R a] [A 9][R 9][A 8][R 8])
const __m128i ar1 = _mm_loadu_si128((__m128i*)src+1);
// Loads the 1st half of GB components ([G 7][B 7][G 6][B 6] [G 5][B 5][G 4][B 4] [G 3][B 3][G 2][B 2] [G 1][B 1][G 0][B 0])
const __m128i gb0 = _mm_loadu_si128((__m128i*)src+2);
// Loads the 2nd half of GB components ([G f][B f][G e][B e] [G d][B d][G c][B c] [G b][B b][G a][B a] [G 9][B 9][G 8][B 8])
const __m128i gb1 = _mm_loadu_si128((__m128i*)src+3);
__m128i rgba00, rgba01, rgba10, rgba11;
#if _M_SSE >= 0x301
// SSSE3 Implementation is about 25% faster than SSE2 version
if (cpu_info.bSSSE3)
// xsacha optimized with SSSE3 instrinsics
// Produces a ~25% speed improvement over SSE2 implementation
if (cpu_info.bSSSE3)
{
__m128i rgba00, rgba01, rgba10, rgba11;
for (int y = 0; y < height; y += 4)
for (int x = 0; x < width; x += 4, src += 64)
{
const __m128i ar0 = _mm_loadu_si128((__m128i*)src);
const __m128i ar1 = _mm_loadu_si128((__m128i*)src+1);
const __m128i gb0 = _mm_loadu_si128((__m128i*)src+2);
const __m128i gb1 = _mm_loadu_si128((__m128i*)src+3);
const __m128i mask6xx7 = _mm_set_epi8(6, 128, 128, 7, 4, 128, 128, 5, 2, 128, 128, 3, 0, 128, 128, 1);
const __m128i maskExxF = _mm_set_epi8(14, 128, 128, 15, 12, 128, 128, 13, 10, 128, 128, 11, 8, 128, 128, 9);
const __m128i maskx76x = _mm_set_epi8(128, 7, 6, 128, 128, 5, 4, 128, 128, 3, 2, 128, 128, 1, 0, 128);
@ -1744,9 +1742,51 @@ PC_TexFormat TexDecoder_Decode_RGBA(u32 * dst, const u8 * src, int width, int he
rgba01 = _mm_or_si128(_mm_shuffle_epi8(ar0, maskExxF), _mm_shuffle_epi8(gb0, maskxFEx));
rgba10 = _mm_or_si128(_mm_shuffle_epi8(ar1, mask6xx7), _mm_shuffle_epi8(gb1, maskx76x));
rgba11 = _mm_or_si128(_mm_shuffle_epi8(ar1, maskExxF), _mm_shuffle_epi8(gb1, maskxFEx));
} else
// Write em out!
__m128i *dst128 = (__m128i*)( dst + (y + 0) * width + x );
_mm_storeu_si128(dst128, rgba00);
dst128 = (__m128i*)( dst + (y + 1) * width + x );
_mm_storeu_si128(dst128, rgba01);
dst128 = (__m128i*)( dst + (y + 2) * width + x );
_mm_storeu_si128(dst128, rgba10);
dst128 = (__m128i*)( dst + (y + 3) * width + x );
_mm_storeu_si128(dst128, rgba11);
}
} else
#endif
// JSD optimized with SSE2 intrinsics
// Produces a ~68% speed improvement over reference C implementation.
{
for (int y = 0; y < height; y += 4)
for (int x = 0; x < width; x += 4, src += 64)
{
// Input is divided up into 16-bit words. The texels are split up into AR and GB components where all
// AR components come grouped up first in 32 bytes followed by the GB components in 32 bytes. We are
// processing 16 texels per each loop iteration, numbered from 0-f.
//
// Convention is:
// one byte is [component-name texel-number]
// __m128i is (4-bytes 4-bytes 4-bytes 4-bytes)
//
// Input is ([A 7][R 7][A 6][R 6] [A 5][R 5][A 4][R 4] [A 3][R 3][A 2][R 2] [A 1][R 1][A 0][R 0])
// ([A f][R f][A e][R e] [A d][R d][A c][R c] [A b][R b][A a][R a] [A 9][R 9][A 8][R 8])
// ([G 7][B 7][G 6][B 6] [G 5][B 5][G 4][B 4] [G 3][B 3][G 2][B 2] [G 1][B 1][G 0][B 0])
// ([G f][B f][G e][B e] [G d][B d][G c][B c] [G b][B b][G a][B a] [G 9][B 9][G 8][B 8])
//
// Output is (RGBA3 RGBA2 RGBA1 RGBA0)
// (RGBA7 RGBA6 RGBA5 RGBA4)
// (RGBAb RGBAa RGBA9 RGBA8)
// (RGBAf RGBAe RGBAd RGBAc)
// Loads the 1st half of AR components ([A 7][R 7][A 6][R 6] [A 5][R 5][A 4][R 4] [A 3][R 3][A 2][R 2] [A 1][R 1][A 0][R 0])
const __m128i ar0 = _mm_loadu_si128((__m128i*)src);
// Loads the 2nd half of AR components ([A f][R f][A e][R e] [A d][R d][A c][R c] [A b][R b][A a][R a] [A 9][R 9][A 8][R 8])
const __m128i ar1 = _mm_loadu_si128((__m128i*)src+1);
// Loads the 1st half of GB components ([G 7][B 7][G 6][B 6] [G 5][B 5][G 4][B 4] [G 3][B 3][G 2][B 2] [G 1][B 1][G 0][B 0])
const __m128i gb0 = _mm_loadu_si128((__m128i*)src+2);
// Loads the 2nd half of GB components ([G f][B f][G e][B e] [G d][B d][G c][B c] [G b][B b][G a][B a] [G 9][B 9][G 8][B 8])
const __m128i gb1 = _mm_loadu_si128((__m128i*)src+3);
__m128i rgba00, rgba01, rgba10, rgba11;
const __m128i kMask_x000f = _mm_set_epi32(0x000000FFL, 0x000000FFL, 0x000000FFL, 0x000000FFL);
const __m128i kMask_xf000 = _mm_set_epi32(0xFF000000L, 0xFF000000L, 0xFF000000L, 0xFF000000L);
const __m128i kMask_x0ff0 = _mm_set_epi32(0x00FFFF00L, 0x00FFFF00L, 0x00FFFF00L, 0x00FFFF00L);
@ -1800,18 +1840,17 @@ PC_TexFormat TexDecoder_Decode_RGBA(u32 * dst, const u8 * src, int width, int he
rgba01 = _mm_or_si128(r__a01, _gb_01);
rgba10 = _mm_or_si128(r__a10, _gb_10);
rgba11 = _mm_or_si128(r__a11, _gb_11);
// Write em out!
__m128i *dst128 = (__m128i*)( dst + (y + 0) * width + x );
_mm_storeu_si128(dst128, rgba00);
dst128 = (__m128i*)( dst + (y + 1) * width + x );
_mm_storeu_si128(dst128, rgba01);
dst128 = (__m128i*)( dst + (y + 2) * width + x );
_mm_storeu_si128(dst128, rgba10);
dst128 = (__m128i*)( dst + (y + 3) * width + x );
_mm_storeu_si128(dst128, rgba11);
}
// Write em out!
__m128i *dst128 = (__m128i*)( dst + (y + 0) * width + x );
_mm_storeu_si128(dst128, rgba00);
dst128 = (__m128i*)( dst + (y + 1) * width + x );
_mm_storeu_si128(dst128, rgba01);
dst128 = (__m128i*)( dst + (y + 2) * width + x );
_mm_storeu_si128(dst128, rgba10);
dst128 = (__m128i*)( dst + (y + 3) * width + x );
_mm_storeu_si128(dst128, rgba11);
}
}
#if 0
// Reference C implementation.
for (int y = 0; y < height; y += 4)