AuroraMiddleware/skia2
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

reland: We can mask load and store with just AVX

Browse Source 
Originally reviewed here: https://skia-review.googlesource.com/c/17452/

CQ_INCLUDE_TRYBOTS=skia.primary:Test-Ubuntu-GCC-ShuttleA-GPU-GTX550Ti-x86_64-Release-Valgrind

Change-Id: I2e593e897ce93147ec593c2a5de143217274ba2a
Reviewed-on: https://skia-review.googlesource.com/18267
Reviewed-by: Mike Klein <mtklein@chromium.org>
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Mike Klein <mtklein@chromium.org>
This commit is contained in:
Mike Klein 2017-05-31 16:59:00 -04:00 committed by Skia Commit-Bot
parent 8bf1f9ffcf
commit 823103384c
3 changed files with 1418 additions and 1582 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1491
src/jumper/SkJumper_generated.S
View File

File diff suppressed because it is too large Load Diff

1497
src/jumper/SkJumper_generated_win.S
View File

File diff suppressed because it is too large Load Diff

12
src/jumper/SkJumper_stages.cpp
View File

@ -176,8 +176,8 @@ SI void store(T* dst, V v, size_t tail) {
}
#endif
// AVX2 adds some mask loads and stores that make for shorter, faster code.
#if defined(JUMPER) && defined(__AVX2__)
// AVX adds some mask loads and stores that make for shorter, faster code.
#if defined(JUMPER) && defined(__AVX__)
SI U32 mask(size_t tail) {
// We go a little out of our way to avoid needing large constant values here.
@ -186,14 +186,16 @@ SI void store(T* dst, V v, size_t tail) {
uint64_t mask = 0xffffffffffffffff >> 8*(kStride-tail);
// Sign-extend each mask lane to its full width, 0x00000000 or 0xffffffff.
return _mm256_cvtepi8_epi32(_mm_cvtsi64_si128((int64_t)mask));
using S8 = int8_t __attribute__((ext_vector_type(8)));
using S32 = int32_t __attribute__((ext_vector_type(8)));
return (U32)__builtin_convertvector(unaligned_load<S8>(&mask), S32);
}
template <>
inline U32 load(const uint32_t* src, size_t tail) {
__builtin_assume(tail < kStride);
if (__builtin_expect(tail, 0)) {
return _mm256_maskload_epi32((const int*)src, mask(tail));
return (U32)_mm256_maskload_ps((const float*)src, mask(tail));
}
return unaligned_load<U32>(src);
}
@ -202,7 +204,7 @@ SI void store(T* dst, V v, size_t tail) {
inline void store(uint32_t* dst, U32 v, size_t tail) {
__builtin_assume(tail < kStride);
if (__builtin_expect(tail, 0)) {
return _mm256_maskstore_epi32((int*)dst, mask(tail), v);
return _mm256_maskstore_ps((float*)dst, mask(tail), (F)v);
}
unaligned_store(dst, v);
}