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

avx2 specialization for gather32

Browse Source 
This wraps up my recent interpreter optimization push.
I've been informally measuring my progress by running

    viewer --slide GM_color_cube_rt --skvm --nojit --stats

The frame cost has dropped from 13-14ms to just under 10ms;
the JIT's right around 3ms.

Change-Id: Ieff977ac7777dbbf15c8c3b7e40a0031fb95fe7d
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/317718
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
This commit is contained in:
Mike Klein 2020-09-18 08:02:46 -05:00 committed by Skia Commit-Bot
parent 6798e5d0d1
commit 2e69a13d83
1 changed files with 17 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
src/opts/SkVM_opts.h
View File

@ -37,6 +37,22 @@ static inline skvx::Vec<N,int16_t> mul_q14(const skvx::Vec<N,int16_t>& x,
skvx::cast<int>(y) + 0x4000)>>15 ) <<1;
}
template <int N>
static inline skvx::Vec<N,int> gather32(const int* ptr, const skvx::Vec<N,int>& ix) {
#if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_AVX2
if constexpr (N == 8) {
return skvx::bit_pun<skvx::Vec<N,int>>(
_mm256_i32gather_epi32(ptr, skvx::bit_pun<__m256i>(ix), 4));
}
#endif
// Try to recurse on specializations, falling back on standard scalar map()-based impl.
if constexpr (N > 8) {
return join(gather32(ptr, ix.lo),
gather32(ptr, ix.hi));
}
return map(ix, [&](int i) { return ptr[i]; });
}
namespace SK_OPTS_NS {
inline void interpret_skvm(const skvm::InterpreterInstruction insts[], const int ninsts,
@ -180,7 +196,7 @@ namespace SK_OPTS_NS {
STRIDE_K(Op::gather32): {
const int* ptr;
memcpy(&ptr, (const uint8_t*)args[immy] + immz, sizeof(ptr));
r[d].i32 = map(r[x].i32, [&](int ix) { return ptr[ix]; });
r[d].i32 = gather32(ptr, r[x].i32);
} break;
#undef STRIDE_1