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

An SSE4 implementation for I8 texture decode. Slightly faster than SSE2 version on my computer (SSE4: 60%, SSE2: 55% vs reference C on Core2 Duo). 

git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@6768 8ced0084-cf51-0410-be5f-012b33b47a6e
This commit is contained in:
xsacha 2011-01-07 13:40:32 +00:00
parent c02f56fca3
commit 53474403e2
1 changed files with 22 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

24
Source/Core/VideoCommon/Src/TextureDecoder.cpp
View File

@ -1065,11 +1065,30 @@ PC_TexFormat TexDecoder_Decode_RGBA(u32 * dst, const u8 * src, int width, int he
break; break;
case GX_TF_I8: // speed critical case GX_TF_I8: // speed critical
{ {
// JSD optimized with SSE2 intrinsics.
// Produces an ~86% speed increase over reference C implementation.
for (int y = 0; y < height; y += 4) for (int y = 0; y < height; y += 4)
for (int x = 0; x < width; x += 8) for (int x = 0; x < width; x += 8)
{ {
#if _M_SSE >= 0x401
// SSE4 intrinsics: About 5-10% faster than SSE2 version
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);
const __m128i mask7654 = _mm_set_epi8(7, 7, 7, 7, 6, 6, 6, 6, 5, 5, 5, 5, 4, 4, 4, 4);
__m128i *quaddst, r, rgba0, rgba1;
// Load 64 bits from `src` into an __m128i with upper 64 bits zeroed: (0000 0000 hgfe dcba)
r = _mm_loadl_epi64((const __m128i *)src);
// Shuffle select bytes to expand from (0000 0000 hgfe dcba) to:
rgba0 = _mm_shuffle_epi8(r, mask3210); // (dddd cccc bbbb aaaa)
rgba1 = _mm_shuffle_epi8(r, mask7654); // (hhhh gggg ffff eeee)
quaddst = (__m128i *)(dst + (y + iy)*width + x);
_mm_storeu_si128(quaddst, rgba0);
_mm_storeu_si128(quaddst+1, rgba1);
}
#else
// 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. // 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? // TODO: is it more efficient to group the loads together sequentially and also the stores at the end?
@ -1146,6 +1165,7 @@ PC_TexFormat TexDecoder_Decode_RGBA(u32 * dst, const u8 * src, int width, int he
_mm_storeu_si128(quaddst+1, rgba7); _mm_storeu_si128(quaddst+1, rgba7);
src += 8; src += 8;
#endif
} }
#if 0 #if 0
// Reference C implementation // Reference C implementation