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

GPU: 

- Further optimize the SSE2 versions of ConvertColor555To6665Opaque() and ConvertColor555To8888Opaque().
This commit is contained in:
rogerman 2016-07-01 18:32:10 +00:00
parent a05ddab710
commit e6dac5ec96
2 changed files with 61 additions and 118 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

100
desmume/src/GPU.h
View File

@ -1783,77 +1783,61 @@ FORCEINLINE u16 ConvertColor6665To5551(u32 srcColor)
#ifdef ENABLE_SSE2 #ifdef ENABLE_SSE2
template <bool SWAP_RB> template <bool SWAP_RB>
FORCEINLINE void ConvertColor555To8888Opaque(const __m128i &src, __m128i &dstLo, __m128i &dstHi) FORCEINLINE void ConvertColor555To8888(const __m128i &srcColor, const __m128i &srcAlphaBits32Lo, const __m128i &srcAlphaBits32Hi, __m128i &dstLo, __m128i &dstHi)
{ {
__m128i src32; __m128i src32;
// Conversion algorithm: // Conversion algorithm:
// RGB 5-bit to 8-bit formula: dstRGB8 = (srcRGB5 << 3) | ((srcRGB5 >> 2) & 0x07) // RGB 5-bit to 8-bit formula: dstRGB8 = (srcRGB5 << 3) | ((srcRGB5 >> 2) & 0x07)
if (SWAP_RB) src32 = _mm_unpacklo_epi16(srcColor, _mm_setzero_si128());
{ dstLo = (SWAP_RB) ? _mm_or_si128(_mm_slli_epi32(src32, 19), _mm_srli_epi32(src32, 7)) : _mm_or_si128(_mm_slli_epi32(src32, 3), _mm_slli_epi32(src32, 9));
src32 = _mm_unpacklo_epi16(src, _mm_setzero_si128()); dstLo = _mm_and_si128( dstLo, _mm_set1_epi32(0x00F800F8) );
dstLo = _mm_or_si128(_mm_and_si128(_mm_slli_epi32(src32, 19), _mm_set1_epi32(0x00F80000)), _mm_and_si128(_mm_slli_epi32(src32, 14), _mm_set1_epi32(0x00070000))); dstLo = _mm_or_si128( dstLo, _mm_and_si128(_mm_slli_epi32(src32, 6), _mm_set1_epi32(0x0000F800)) );
dstLo = _mm_or_si128( dstLo, _mm_or_si128(_mm_and_si128(_mm_slli_epi32(src32, 6), _mm_set1_epi32(0x0000F800)), _mm_and_si128(_mm_slli_epi32(src32, 1), _mm_set1_epi32(0x00000700))) ); dstLo = _mm_or_si128( dstLo, _mm_and_si128(_mm_srli_epi32(dstLo, 5), _mm_set1_epi32(0x00070707)) );
dstLo = _mm_or_si128( dstLo, _mm_or_si128(_mm_and_si128(_mm_srli_epi32(src32, 7), _mm_set1_epi32(0x000000F8)), _mm_and_si128(_mm_srli_epi32(src32, 12), _mm_set1_epi32(0x00000007))) ); dstLo = _mm_or_si128( dstLo, srcAlphaBits32Lo );
dstLo = _mm_or_si128( dstLo, _mm_set1_epi32(0xFF000000) );
src32 = _mm_unpackhi_epi16(srcColor, _mm_setzero_si128());
src32 = _mm_unpackhi_epi16(src, _mm_setzero_si128()); dstHi = (SWAP_RB) ? _mm_or_si128(_mm_slli_epi32(src32, 19), _mm_srli_epi32(src32, 7)) : _mm_or_si128(_mm_slli_epi32(src32, 3), _mm_slli_epi32(src32, 9));
dstHi = _mm_or_si128(_mm_and_si128(_mm_slli_epi32(src32, 19), _mm_set1_epi32(0x00F80000)), _mm_and_si128(_mm_slli_epi32(src32, 14), _mm_set1_epi32(0x00070000))); dstHi = _mm_and_si128( dstHi, _mm_set1_epi32(0x00F800F8) );
dstHi = _mm_or_si128( dstHi, _mm_or_si128(_mm_and_si128(_mm_slli_epi32(src32, 6), _mm_set1_epi32(0x0000F800)), _mm_and_si128(_mm_slli_epi32(src32, 1), _mm_set1_epi32(0x00000700))) ); dstHi = _mm_or_si128( dstHi, _mm_and_si128(_mm_slli_epi32(src32, 6), _mm_set1_epi32(0x0000F800)) );
dstHi = _mm_or_si128( dstHi, _mm_or_si128(_mm_and_si128(_mm_srli_epi32(src32, 7), _mm_set1_epi32(0x000000F8)), _mm_and_si128(_mm_srli_epi32(src32, 12), _mm_set1_epi32(0x00000007))) ); dstHi = _mm_or_si128( dstHi, _mm_and_si128(_mm_srli_epi32(dstHi, 5), _mm_set1_epi32(0x00070707)) );
dstHi = _mm_or_si128( dstHi, _mm_set1_epi32(0xFF000000) ); dstHi = _mm_or_si128( dstHi, srcAlphaBits32Hi );
}
else
{
src32 = _mm_unpacklo_epi16(src, _mm_setzero_si128());
dstLo = _mm_or_si128(_mm_and_si128(_mm_slli_epi32(src32, 3), _mm_set1_epi32(0x000000F8)), _mm_and_si128(_mm_srli_epi32(src32, 2), _mm_set1_epi32(0x00000007)));
dstLo = _mm_or_si128( dstLo, _mm_or_si128(_mm_and_si128(_mm_slli_epi32(src32, 6), _mm_set1_epi32(0x0000F800)), _mm_and_si128(_mm_slli_epi32(src32, 1), _mm_set1_epi32(0x00000700))) );
dstLo = _mm_or_si128( dstLo, _mm_or_si128(_mm_and_si128(_mm_slli_epi32(src32, 9), _mm_set1_epi32(0x00F80000)), _mm_and_si128(_mm_slli_epi32(src32, 4), _mm_set1_epi32(0x00070000))) );
dstLo = _mm_or_si128( dstLo, _mm_set1_epi32(0xFF000000) );
src32 = _mm_unpackhi_epi16(src, _mm_setzero_si128());
dstHi = _mm_or_si128(_mm_and_si128(_mm_slli_epi32(src32, 3), _mm_set1_epi32(0x000000F8)), _mm_and_si128(_mm_srli_epi32(src32, 2), _mm_set1_epi32(0x00000007)));
dstHi = _mm_or_si128( dstHi, _mm_or_si128(_mm_and_si128(_mm_slli_epi32(src32, 6), _mm_set1_epi32(0x0000F800)), _mm_and_si128(_mm_slli_epi32(src32, 1), _mm_set1_epi32(0x00000700))) );
dstHi = _mm_or_si128( dstHi, _mm_or_si128(_mm_and_si128(_mm_slli_epi32(src32, 9), _mm_set1_epi32(0x00F80000)), _mm_and_si128(_mm_slli_epi32(src32, 4), _mm_set1_epi32(0x00070000))) );
dstHi = _mm_or_si128( dstHi, _mm_set1_epi32(0xFF000000) );
}
} }
template <bool SWAP_RB> template <bool SWAP_RB>
FORCEINLINE void ConvertColor555To6665Opaque(const __m128i &src, __m128i &dstLo, __m128i &dstHi) FORCEINLINE void ConvertColor555To6665(const __m128i &srcColor, const __m128i &srcAlphaBits32Lo, const __m128i &srcAlphaBits32Hi, __m128i &dstLo, __m128i &dstHi)
{ {
__m128i src32; __m128i src32;
// Conversion algorithm: // Conversion algorithm:
// RGB 5-bit to 6-bit formula: dstRGB6 = (srcRGB5 << 1) | ((srcRGB5 >> 4) & 0x01) // RGB 5-bit to 6-bit formula: dstRGB6 = (srcRGB5 << 1) | ((srcRGB5 >> 4) & 0x01)
if (SWAP_RB) src32 = _mm_unpacklo_epi16(srcColor, _mm_setzero_si128());
{ dstLo = (SWAP_RB) ? _mm_or_si128(_mm_slli_epi32(src32, 17), _mm_srli_epi32(src32, 9)) : _mm_or_si128(_mm_slli_epi32(src32, 1), _mm_slli_epi32(src32, 7));
src32 = _mm_unpacklo_epi16(src, _mm_setzero_si128()); dstLo = _mm_and_si128( dstLo, _mm_set1_epi32(0x003E003E) );
dstLo = _mm_or_si128(_mm_and_si128(_mm_slli_epi32(src32, 17), _mm_set1_epi32(0x003E0000)), _mm_and_si128(_mm_slli_epi32(src32, 12), _mm_set1_epi32(0x00010000))); dstLo = _mm_or_si128( dstLo, _mm_and_si128(_mm_slli_epi32(src32, 4), _mm_set1_epi32(0x00003E00)) );
dstLo = _mm_or_si128( dstLo, _mm_or_si128(_mm_and_si128(_mm_slli_epi32(src32, 4), _mm_set1_epi32(0x00003E00)), _mm_and_si128(_mm_srli_epi32(src32, 1), _mm_set1_epi32(0x00000100))) ); dstLo = _mm_or_si128( dstLo, _mm_and_si128(_mm_srli_epi32(dstLo, 5), _mm_set1_epi32(0x00010101)) );
dstLo = _mm_or_si128( dstLo, _mm_or_si128(_mm_and_si128(_mm_srli_epi32(src32, 9), _mm_set1_epi32(0x0000003E)), _mm_and_si128(_mm_srli_epi32(src32, 14), _mm_set1_epi32(0x00000001))) ); dstLo = _mm_or_si128( dstLo, srcAlphaBits32Lo );
dstLo = _mm_or_si128( dstLo, _mm_set1_epi32(0x1F000000) );
src32 = _mm_unpackhi_epi16(srcColor, _mm_setzero_si128());
src32 = _mm_unpackhi_epi16(src, _mm_setzero_si128()); dstHi = (SWAP_RB) ? _mm_or_si128(_mm_slli_epi32(src32, 17), _mm_srli_epi32(src32, 9)) : _mm_or_si128(_mm_slli_epi32(src32, 1), _mm_slli_epi32(src32, 7));
dstHi = _mm_or_si128(_mm_and_si128(_mm_slli_epi32(src32, 17), _mm_set1_epi32(0x003E0000)), _mm_and_si128(_mm_slli_epi32(src32, 12), _mm_set1_epi32(0x00010000))); dstHi = _mm_and_si128( dstHi, _mm_set1_epi32(0x003E003E) );
dstHi = _mm_or_si128( dstHi, _mm_or_si128(_mm_and_si128(_mm_slli_epi32(src32, 4), _mm_set1_epi32(0x00003E00)), _mm_and_si128(_mm_srli_epi32(src32, 1), _mm_set1_epi32(0x00000100))) ); dstHi = _mm_or_si128( dstHi, _mm_and_si128(_mm_slli_epi32(src32, 4), _mm_set1_epi32(0x00003E00)) );
dstHi = _mm_or_si128( dstHi, _mm_or_si128(_mm_and_si128(_mm_srli_epi32(src32, 9), _mm_set1_epi32(0x0000003E)), _mm_and_si128(_mm_srli_epi32(src32, 14), _mm_set1_epi32(0x00000001))) ); dstHi = _mm_or_si128( dstHi, _mm_and_si128(_mm_srli_epi32(dstHi, 5), _mm_set1_epi32(0x00010101)) );
dstHi = _mm_or_si128( dstHi, _mm_set1_epi32(0x1F000000) ); dstHi = _mm_or_si128( dstHi, srcAlphaBits32Hi );
} }
else
{ template <bool SWAP_RB>
src32 = _mm_unpacklo_epi16(src, _mm_setzero_si128()); FORCEINLINE void ConvertColor555To8888Opaque(const __m128i &srcColor, __m128i &dstLo, __m128i &dstHi)
dstLo = _mm_or_si128(_mm_and_si128(_mm_slli_epi32(src32, 1), _mm_set1_epi32(0x0000003E)), _mm_and_si128(_mm_srli_epi32(src32, 4), _mm_set1_epi32(0x00000001))); {
dstLo = _mm_or_si128( dstLo, _mm_or_si128(_mm_and_si128(_mm_slli_epi32(src32, 4), _mm_set1_epi32(0x00003E00)), _mm_and_si128(_mm_srli_epi32(src32, 1), _mm_set1_epi32(0x00000100))) ); const __m128i srcAlphaBits32 = _mm_set1_epi32(0xFF000000);
dstLo = _mm_or_si128( dstLo, _mm_or_si128(_mm_and_si128(_mm_slli_epi32(src32, 7), _mm_set1_epi32(0x003E0000)), _mm_and_si128(_mm_slli_epi32(src32, 2), _mm_set1_epi32(0x00010000))) ); ConvertColor555To8888<SWAP_RB>(srcColor, srcAlphaBits32, srcAlphaBits32, dstLo, dstHi);
dstLo = _mm_or_si128( dstLo, _mm_set1_epi32(0x1F000000) ); }
src32 = _mm_unpackhi_epi16(src, _mm_setzero_si128()); template <bool SWAP_RB>
dstHi = _mm_or_si128(_mm_and_si128(_mm_slli_epi32(src32, 1), _mm_set1_epi32(0x0000003E)), _mm_and_si128(_mm_srli_epi32(src32, 4), _mm_set1_epi32(0x00000001))); FORCEINLINE void ConvertColor555To6665Opaque(const __m128i &srcColor, __m128i &dstLo, __m128i &dstHi)
dstHi = _mm_or_si128( dstHi, _mm_or_si128(_mm_and_si128(_mm_slli_epi32(src32, 4), _mm_set1_epi32(0x00003E00)), _mm_and_si128(_mm_srli_epi32(src32, 1), _mm_set1_epi32(0x00000100))) ); {
dstHi = _mm_or_si128( dstHi, _mm_or_si128(_mm_and_si128(_mm_slli_epi32(src32, 7), _mm_set1_epi32(0x003E0000)), _mm_and_si128(_mm_slli_epi32(src32, 2), _mm_set1_epi32(0x00010000))) ); const __m128i srcAlphaBits32 = _mm_set1_epi32(0x1F000000);
dstHi = _mm_or_si128( dstHi, _mm_set1_epi32(0x1F000000) ); ConvertColor555To6665<SWAP_RB>(srcColor, srcAlphaBits32, srcAlphaBits32, dstLo, dstHi);
}
} }
template <bool SWAP_RB> template <bool SWAP_RB>

79
desmume/src/texcache.cpp
View File

@ -870,77 +870,36 @@ public:
const __m128i palColor0 = _mm_shuffle_epi8(pal_vec128, idx0); const __m128i palColor0 = _mm_shuffle_epi8(pal_vec128, idx0);
const __m128i palColor1 = _mm_shuffle_epi8(pal_vec128, idx1); const __m128i palColor1 = _mm_shuffle_epi8(pal_vec128, idx1);
__m128i tmpColor; __m128i tmpAlpha[2];
__m128i tmpAlpha;
__m128i convertedColor[4]; __m128i convertedColor[4];
if (TEXFORMAT == TexFormat_15bpp) if (TEXFORMAT == TexFormat_15bpp)
{ {
__m128i alpha = _mm_srli_epi16( _mm_and_si128(bits, _mm_set1_epi8(0xF8)), 3 ); const __m128i alpha = _mm_srli_epi16( _mm_and_si128(bits, _mm_set1_epi8(0xF8)), 3 );
__m128i alphaLo = _mm_unpacklo_epi8(_mm_setzero_si128(), alpha); const __m128i alphaLo = _mm_unpacklo_epi8(_mm_setzero_si128(), alpha);
__m128i alphaHi = _mm_unpackhi_epi8(_mm_setzero_si128(), alpha); const __m128i alphaHi = _mm_unpackhi_epi8(_mm_setzero_si128(), alpha);
tmpColor = _mm_unpacklo_epi16(palColor0, _mm_setzero_si128()); tmpAlpha[0] = _mm_unpacklo_epi16(_mm_setzero_si128(), alphaLo);
tmpAlpha = _mm_unpacklo_epi16(_mm_setzero_si128(), alphaLo); tmpAlpha[1] = _mm_unpackhi_epi16(_mm_setzero_si128(), alphaLo);
convertedColor[0] = _mm_or_si128(_mm_and_si128(_mm_slli_epi32(tmpColor, 1), _mm_set1_epi32(0x0000003E)), _mm_and_si128(_mm_srli_epi32(tmpColor, 4), _mm_set1_epi32(0x00000001))); ConvertColor555To6665<false>(palColor0, tmpAlpha[0], tmpAlpha[1], convertedColor[0], convertedColor[1]);
convertedColor[0] = _mm_or_si128( convertedColor[0], _mm_or_si128(_mm_and_si128(_mm_slli_epi32(tmpColor, 4), _mm_set1_epi32(0x00003E00)), _mm_and_si128(_mm_srli_epi32(tmpColor, 1), _mm_set1_epi32(0x00000100))) );
convertedColor[0] = _mm_or_si128( convertedColor[0], _mm_or_si128(_mm_and_si128(_mm_slli_epi32(tmpColor, 7), _mm_set1_epi32(0x003E0000)), _mm_and_si128(_mm_slli_epi32(tmpColor, 2), _mm_set1_epi32(0x00010000))) );
convertedColor[0] = _mm_or_si128( convertedColor[0], tmpAlpha);
tmpColor = _mm_unpackhi_epi16(palColor0, _mm_setzero_si128()); tmpAlpha[0] = _mm_unpacklo_epi16(_mm_setzero_si128(), alphaHi);
tmpAlpha = _mm_unpackhi_epi16(_mm_setzero_si128(), alphaLo); tmpAlpha[1] = _mm_unpackhi_epi16(_mm_setzero_si128(), alphaHi);
convertedColor[1] = _mm_or_si128(_mm_and_si128(_mm_slli_epi32(tmpColor, 1), _mm_set1_epi32(0x0000003E)), _mm_and_si128(_mm_srli_epi32(tmpColor, 4), _mm_set1_epi32(0x00000001))); ConvertColor555To6665<false>(palColor1, tmpAlpha[0], tmpAlpha[1], convertedColor[2], convertedColor[3]);
convertedColor[1] = _mm_or_si128( convertedColor[1], _mm_or_si128(_mm_and_si128(_mm_slli_epi32(tmpColor, 4), _mm_set1_epi32(0x00003E00)), _mm_and_si128(_mm_srli_epi32(tmpColor, 1), _mm_set1_epi32(0x00000100))) );
convertedColor[1] = _mm_or_si128( convertedColor[1], _mm_or_si128(_mm_and_si128(_mm_slli_epi32(tmpColor, 7), _mm_set1_epi32(0x003E0000)), _mm_and_si128(_mm_slli_epi32(tmpColor, 2), _mm_set1_epi32(0x00010000))) );
convertedColor[1] = _mm_or_si128( convertedColor[1], tmpAlpha);
tmpColor = _mm_unpacklo_epi16(palColor1, _mm_setzero_si128());
tmpAlpha = _mm_unpacklo_epi16(_mm_setzero_si128(), alphaHi);
convertedColor[2] = _mm_or_si128(_mm_and_si128(_mm_slli_epi32(tmpColor, 1), _mm_set1_epi32(0x0000003E)), _mm_and_si128(_mm_srli_epi32(tmpColor, 4), _mm_set1_epi32(0x00000001)));
convertedColor[2] = _mm_or_si128( convertedColor[2], _mm_or_si128(_mm_and_si128(_mm_slli_epi32(tmpColor, 4), _mm_set1_epi32(0x00003E00)), _mm_and_si128(_mm_srli_epi32(tmpColor, 1), _mm_set1_epi32(0x00000100))) );
convertedColor[2] = _mm_or_si128( convertedColor[2], _mm_or_si128(_mm_and_si128(_mm_slli_epi32(tmpColor, 7), _mm_set1_epi32(0x003E0000)), _mm_and_si128(_mm_slli_epi32(tmpColor, 2), _mm_set1_epi32(0x00010000))) );
convertedColor[2] = _mm_or_si128( convertedColor[2], tmpAlpha);
tmpColor = _mm_unpackhi_epi16(palColor1, _mm_setzero_si128());
tmpAlpha = _mm_unpackhi_epi16(_mm_setzero_si128(), alphaHi);
convertedColor[3] = _mm_or_si128(_mm_and_si128(_mm_slli_epi32(tmpColor, 1), _mm_set1_epi32(0x0000003E)), _mm_and_si128(_mm_srli_epi32(tmpColor, 4), _mm_set1_epi32(0x00000001)));
convertedColor[3] = _mm_or_si128( convertedColor[3], _mm_or_si128(_mm_and_si128(_mm_slli_epi32(tmpColor, 4), _mm_set1_epi32(0x00003E00)), _mm_and_si128(_mm_srli_epi32(tmpColor, 1), _mm_set1_epi32(0x00000100))) );
convertedColor[3] = _mm_or_si128( convertedColor[3], _mm_or_si128(_mm_and_si128(_mm_slli_epi32(tmpColor, 7), _mm_set1_epi32(0x003E0000)), _mm_and_si128(_mm_slli_epi32(tmpColor, 2), _mm_set1_epi32(0x00010000))) );
convertedColor[3] = _mm_or_si128( convertedColor[3], tmpAlpha);
} }
else else
{ {
__m128i alpha = _mm_or_si128( _mm_and_si128(bits, _mm_set1_epi8(0xF8)), _mm_srli_epi16(_mm_and_si128(bits, _mm_set1_epi8(0xE0)), 5) ); const __m128i alpha = _mm_or_si128( _mm_and_si128(bits, _mm_set1_epi8(0xF8)), _mm_srli_epi16(_mm_and_si128(bits, _mm_set1_epi8(0xE0)), 5) );
__m128i alphaLo = _mm_unpacklo_epi8(_mm_setzero_si128(), alpha); const __m128i alphaLo = _mm_unpacklo_epi8(_mm_setzero_si128(), alpha);
__m128i alphaHi = _mm_unpackhi_epi8(_mm_setzero_si128(), alpha); const __m128i alphaHi = _mm_unpackhi_epi8(_mm_setzero_si128(), alpha);
tmpColor = _mm_unpacklo_epi16(palColor0, _mm_setzero_si128()); tmpAlpha[0] = _mm_unpacklo_epi16(_mm_setzero_si128(), alphaLo);
tmpAlpha = _mm_unpacklo_epi16(_mm_setzero_si128(), alphaLo); tmpAlpha[1] = _mm_unpackhi_epi16(_mm_setzero_si128(), alphaLo);
convertedColor[0] = _mm_or_si128(_mm_and_si128(_mm_slli_epi32(tmpColor, 3), _mm_set1_epi32(0x000000F8)), _mm_and_si128(_mm_srli_epi32(tmpColor, 2), _mm_set1_epi32(0x00000007))); ConvertColor555To8888<false>(palColor0, tmpAlpha[0], tmpAlpha[1], convertedColor[0], convertedColor[1]);
convertedColor[0] = _mm_or_si128( convertedColor[0], _mm_or_si128(_mm_and_si128(_mm_slli_epi32(tmpColor, 6), _mm_set1_epi32(0x0000F800)), _mm_and_si128(_mm_slli_epi32(tmpColor, 1), _mm_set1_epi32(0x00000700))) );
convertedColor[0] = _mm_or_si128( convertedColor[0], _mm_or_si128(_mm_and_si128(_mm_slli_epi32(tmpColor, 9), _mm_set1_epi32(0x00F80000)), _mm_and_si128(_mm_slli_epi32(tmpColor, 4), _mm_set1_epi32(0x00070000))) );
convertedColor[0] = _mm_or_si128( convertedColor[0], tmpAlpha);
tmpColor = _mm_unpackhi_epi16(palColor0, _mm_setzero_si128()); tmpAlpha[0] = _mm_unpacklo_epi16(_mm_setzero_si128(), alphaHi);
tmpAlpha = _mm_unpackhi_epi16(_mm_setzero_si128(), alphaLo); tmpAlpha[1] = _mm_unpackhi_epi16(_mm_setzero_si128(), alphaHi);
convertedColor[1] = _mm_or_si128(_mm_and_si128(_mm_slli_epi32(tmpColor, 3), _mm_set1_epi32(0x000000F8)), _mm_and_si128(_mm_srli_epi32(tmpColor, 2), _mm_set1_epi32(0x00000007))); ConvertColor555To8888<false>(palColor1, tmpAlpha[0], tmpAlpha[1], convertedColor[2], convertedColor[3]);
convertedColor[1] = _mm_or_si128( convertedColor[1], _mm_or_si128(_mm_and_si128(_mm_slli_epi32(tmpColor, 6), _mm_set1_epi32(0x0000F800)), _mm_and_si128(_mm_slli_epi32(tmpColor, 1), _mm_set1_epi32(0x00000700))) );
convertedColor[1] = _mm_or_si128( convertedColor[1], _mm_or_si128(_mm_and_si128(_mm_slli_epi32(tmpColor, 9), _mm_set1_epi32(0x00F80000)), _mm_and_si128(_mm_slli_epi32(tmpColor, 4), _mm_set1_epi32(0x00070000))) );
convertedColor[1] = _mm_or_si128( convertedColor[1], tmpAlpha);
tmpColor = _mm_unpacklo_epi16(palColor1, _mm_setzero_si128());
tmpAlpha = _mm_unpacklo_epi16(_mm_setzero_si128(), alphaHi);
convertedColor[2] = _mm_or_si128(_mm_and_si128(_mm_slli_epi32(tmpColor, 3), _mm_set1_epi32(0x000000F8)), _mm_and_si128(_mm_srli_epi32(tmpColor, 2), _mm_set1_epi32(0x00000007)));
convertedColor[2] = _mm_or_si128( convertedColor[2], _mm_or_si128(_mm_and_si128(_mm_slli_epi32(tmpColor, 6), _mm_set1_epi32(0x0000F800)), _mm_and_si128(_mm_slli_epi32(tmpColor, 1), _mm_set1_epi32(0x00000700))) );
convertedColor[2] = _mm_or_si128( convertedColor[2], _mm_or_si128(_mm_and_si128(_mm_slli_epi32(tmpColor, 9), _mm_set1_epi32(0x00F80000)), _mm_and_si128(_mm_slli_epi32(tmpColor, 4), _mm_set1_epi32(0x00070000))) );
convertedColor[2] = _mm_or_si128( convertedColor[2], tmpAlpha);
tmpColor = _mm_unpackhi_epi16(palColor1, _mm_setzero_si128());
tmpAlpha = _mm_unpackhi_epi16(_mm_setzero_si128(), alphaHi);
convertedColor[3] = _mm_or_si128(_mm_and_si128(_mm_slli_epi32(tmpColor, 3), _mm_set1_epi32(0x000000F8)), _mm_and_si128(_mm_srli_epi32(tmpColor, 2), _mm_set1_epi32(0x00000007)));
convertedColor[3] = _mm_or_si128( convertedColor[3], _mm_or_si128(_mm_and_si128(_mm_slli_epi32(tmpColor, 6), _mm_set1_epi32(0x0000F800)), _mm_and_si128(_mm_slli_epi32(tmpColor, 1), _mm_set1_epi32(0x00000700))) );
convertedColor[3] = _mm_or_si128( convertedColor[3], _mm_or_si128(_mm_and_si128(_mm_slli_epi32(tmpColor, 9), _mm_set1_epi32(0x00F80000)), _mm_and_si128(_mm_slli_epi32(tmpColor, 4), _mm_set1_epi32(0x00070000))) );
convertedColor[3] = _mm_or_si128( convertedColor[3], tmpAlpha);
} }
_mm_store_si128((__m128i *)(dwdst + 0), convertedColor[0]); _mm_store_si128((__m128i *)(dwdst + 0), convertedColor[0]);