piwalker/dolphin
1
0
Fork 0
mirror of https://github.com/dolphin-emu/dolphin.git synced 2025-07-23 14:19:46 -06:00
Code Issues Packages Projects Releases Wiki Activity

Use templates for 8+16 bit indexed versions of SSSE3/SSE4 TextCoord reading.

Browse Source
This commit is contained in:
Jordan Woyak
2013-02-21 12:42:09 -06:00
parent c7f4d6b9ac
commit 7987d6babb
1 changed files with 22 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

34
Source/Core/VideoCommon/Src/VertexLoader_TextCoord.cpp
View File

@ -93,11 +93,14 @@ void LOADERDECL TexCoord_ReadIndex()
#if _M_SSE >= 0x401 #if _M_SSE >= 0x401
static const __m128i kMaskSwap16_2 = _mm_set_epi32(0xFFFFFFFFL, 0xFFFFFFFFL, 0xFFFFFFFFL, 0x02030001L); static const __m128i kMaskSwap16_2 = _mm_set_epi32(0xFFFFFFFFL, 0xFFFFFFFFL, 0xFFFFFFFFL, 0x02030001L);
void LOADERDECL TexCoord_ReadIndex16_Short2_SSE4() template <typename I>
void LOADERDECL TexCoord_ReadIndex_Short2_SSE4()
{ {
static_assert(!std::numeric_limits<I>::is_signed, "Only unsigned I is sane!");
// Heavy in ZWW // Heavy in ZWW
u16 Index = DataReadU16(); auto const index = DataRead<I>();
const s32 *pData = (const s32*)(cached_arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex])); const s32 *pData = (const s32*)(cached_arraybases[ARRAY_TEXCOORD0+tcIndex] + (index * arraystrides[ARRAY_TEXCOORD0+tcIndex]));
const __m128i a = _mm_cvtsi32_si128(*pData); const __m128i a = _mm_cvtsi32_si128(*pData);
const __m128i b = _mm_shuffle_epi8(a, kMaskSwap16_2); const __m128i b = _mm_shuffle_epi8(a, kMaskSwap16_2);
const __m128i c = _mm_cvtepi16_epi32(b); const __m128i c = _mm_cvtepi16_epi32(b);
@ -105,7 +108,7 @@ void LOADERDECL TexCoord_ReadIndex16_Short2_SSE4()
const __m128 e = _mm_load1_ps(&tcScale[tcIndex]); const __m128 e = _mm_load1_ps(&tcScale[tcIndex]);
const __m128 f = _mm_mul_ps(d, e); const __m128 f = _mm_mul_ps(d, e);
_mm_storeu_ps((float*)VertexManager::s_pCurBufferPointer, f); _mm_storeu_ps((float*)VertexManager::s_pCurBufferPointer, f);
VertexManager::s_pCurBufferPointer += 8; VertexManager::s_pCurBufferPointer += sizeof(float) * 2;
LOG_TEX<2>(); LOG_TEX<2>();
tcIndex++; tcIndex++;
} }
@ -114,14 +117,17 @@ void LOADERDECL TexCoord_ReadIndex16_Short2_SSE4()
#if _M_SSE >= 0x301 #if _M_SSE >= 0x301
static const __m128i kMaskSwap32 = _mm_set_epi32(0xFFFFFFFFL, 0xFFFFFFFFL, 0x04050607L, 0x00010203L); static const __m128i kMaskSwap32 = _mm_set_epi32(0xFFFFFFFFL, 0xFFFFFFFFL, 0x04050607L, 0x00010203L);
void LOADERDECL TexCoord_ReadIndex16_Float2_SSSE3() template <typename I>
void LOADERDECL TexCoord_ReadIndex_Float2_SSSE3()
{ {
u16 Index = DataReadU16(); static_assert(!std::numeric_limits<I>::is_signed, "Only unsigned I is sane!");
const u32 *pData = (const u32 *)(cached_arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]));
auto const index = DataRead<I>();
const u32 *pData = (const u32 *)(cached_arraybases[ARRAY_TEXCOORD0+tcIndex] + (index * arraystrides[ARRAY_TEXCOORD0+tcIndex]));
GC_ALIGNED128(const __m128i a = _mm_loadl_epi64((__m128i*)pData)); GC_ALIGNED128(const __m128i a = _mm_loadl_epi64((__m128i*)pData));
GC_ALIGNED128(const __m128i b = _mm_shuffle_epi8(a, kMaskSwap32)); GC_ALIGNED128(const __m128i b = _mm_shuffle_epi8(a, kMaskSwap32));
_mm_storel_epi64((__m128i*)VertexManager::s_pCurBufferPointer, b); _mm_storel_epi64((__m128i*)VertexManager::s_pCurBufferPointer, b);
VertexManager::s_pCurBufferPointer += 8; VertexManager::s_pCurBufferPointer += sizeof(float) * 2;
LOG_TEX<2>(); LOG_TEX<2>();
tcIndex++; tcIndex++;
} }
@ -177,16 +183,20 @@ void VertexLoader_TextCoord::Init(void) {
#if _M_SSE >= 0x301 #if _M_SSE >= 0x301
if (cpu_info.bSSSE3) { if (cpu_info.bSSSE3)
tableReadTexCoord[3][4][1] = TexCoord_ReadIndex16_Float2_SSSE3; {
tableReadTexCoord[2][4][1] = TexCoord_ReadIndex_Float2_SSSE3<u8>;
tableReadTexCoord[3][4][1] = TexCoord_ReadIndex_Float2_SSSE3<u16>;
} }
#endif #endif
#if _M_SSE >= 0x401 #if _M_SSE >= 0x401
if (cpu_info.bSSE4_1) { if (cpu_info.bSSE4_1)
tableReadTexCoord[3][3][1] = TexCoord_ReadIndex16_Short2_SSE4; {
tableReadTexCoord[2][3][1] = TexCoord_ReadIndex_Short2_SSE4<u8>;
tableReadTexCoord[3][3][1] = TexCoord_ReadIndex_Short2_SSE4<u16>;
} }
#endif #endif