piwalker/dolphin
1
0
Fork 0
mirror of https://github.com/dolphin-emu/dolphin.git synced 2025-07-21 05:09:34 -06:00
Code Issues Packages Projects Releases Wiki Activity

VideoCommon: remove XFReg copy optimization

Browse Source 
This code is just ugly and I doubt there is a way that copying twice is faster.
This commit is contained in:
degasus
2014-09-03 22:39:26 +02:00
parent e1248599eb
commit ef6f6a7fa9
4 changed files with 15 additions and 101 deletions
Download Patch File Download Diff File Expand all files Collapse all files

40
Source/Core/VideoCommon/DataReader.h
View File

@ -92,46 +92,6 @@ __forceinline u32 DataReadU32()
return DataRead<u32>();
}
#if _M_SSE >= 0x301
const __m128i bs_mask = _mm_set_epi32(0x0C0D0E0FL, 0x08090A0BL, 0x04050607L, 0x00010203L);
template<unsigned int N>
void DataReadU32xN_SSSE3(u32 *bufx16)
{
memcpy(bufx16, g_pVideoData, sizeof(u32) * N);
__m128i* buf = (__m128i *)bufx16;
if (N>12) { _mm_store_si128(buf, _mm_shuffle_epi8(_mm_load_si128(buf), bs_mask)); buf++; }
if (N>8) { _mm_store_si128(buf, _mm_shuffle_epi8(_mm_load_si128(buf), bs_mask)); buf++; }
if (N>4) { _mm_store_si128(buf, _mm_shuffle_epi8(_mm_load_si128(buf), bs_mask)); buf++; }
_mm_store_si128(buf, _mm_shuffle_epi8(_mm_load_si128(buf), bs_mask));
g_pVideoData += (sizeof(u32) * N);
}
#endif
template<unsigned int N>
void DataReadU32xN(u32 *bufx16)
{
memcpy(bufx16, g_pVideoData, sizeof(u32) * N);
if (N >= 1) bufx16[0] = Common::swap32(bufx16[0]);
if (N >= 2) bufx16[1] = Common::swap32(bufx16[1]);
if (N >= 3) bufx16[2] = Common::swap32(bufx16[2]);
if (N >= 4) bufx16[3] = Common::swap32(bufx16[3]);
if (N >= 5) bufx16[4] = Common::swap32(bufx16[4]);
if (N >= 6) bufx16[5] = Common::swap32(bufx16[5]);
if (N >= 7) bufx16[6] = Common::swap32(bufx16[6]);
if (N >= 8) bufx16[7] = Common::swap32(bufx16[7]);
if (N >= 9) bufx16[8] = Common::swap32(bufx16[8]);
if (N >= 10) bufx16[9] = Common::swap32(bufx16[9]);
if (N >= 11) bufx16[10] = Common::swap32(bufx16[10]);
if (N >= 12) bufx16[11] = Common::swap32(bufx16[11]);
if (N >= 13) bufx16[12] = Common::swap32(bufx16[12]);
if (N >= 14) bufx16[13] = Common::swap32(bufx16[13]);
if (N >= 15) bufx16[14] = Common::swap32(bufx16[14]);
if (N >= 16) bufx16[15] = Common::swap32(bufx16[15]);
g_pVideoData += (sizeof(u32) * N);
}
__forceinline u32 DataReadU32Unswapped()
{
u32 tmp = *(u32*)g_pVideoData;