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

more powerful map()

Browse Source 
Change-Id: Icbae002999a295e3a9d1d2e6046e686784d5f608
Reviewed-on: https://skia-review.googlesource.com/69901
Reviewed-by: Florin Malita <fmalita@chromium.org>
Commit-Queue: Mike Klein <mtklein@chromium.org>
This commit is contained in:
Mike Klein 2017-11-10 13:01:38 -05:00 committed by Skia Commit-Bot
parent 037d1fc453
commit a3dd5ec3a7
1 changed files with 30 additions and 24 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

54
src/jumper/SkJumper_stages_lowp.cpp
View File

@ -183,14 +183,28 @@ SI D join(S lo, S hi) {
memcpy((char*)&v + 1*sizeof(S), &hi, sizeof(S));
return v;
}
template <typename V, typename H>
SI V map(V v, H (*fn)(H)) {
H lo,hi;
split(v, &lo,&hi);
lo = fn(lo);
hi = fn(hi);
return join<V>(lo,hi);
}
#if defined(__AVX2__)
template <typename V, typename Fn>
SI auto map(V v, Fn&& fn) -> decltype(
__builtin_shufflevector(fn(__builtin_shufflevector(v,v, 0,1, 2, 3, 4, 5, 6, 7)),
fn(__builtin_shufflevector(v,v, 8,9,10,11,12,13,14,15)),
0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15)) {
return __builtin_shufflevector(fn(__builtin_shufflevector(v,v, 0,1, 2, 3, 4, 5, 6, 7)),
fn(__builtin_shufflevector(v,v, 8,9,10,11,12,13,14,15)),
0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15);
}
#else
template <typename V, typename Fn>
SI auto map(V v, Fn&& fn) -> decltype(
__builtin_shufflevector(fn(__builtin_shufflevector(v,v, 0,1,2,3)),
fn(__builtin_shufflevector(v,v, 4,5,6,7)),
0,1,2,3,4,5,6,7)) {
return __builtin_shufflevector(fn(__builtin_shufflevector(v,v, 0,1,2,3)),
fn(__builtin_shufflevector(v,v, 4,5,6,7)),
0,1,2,3,4,5,6,7);
}
#endif
// TODO: do we need platform-specific intrinsics for any of these?
SI F if_then_else(I32 c, F t, F e) {
@ -208,7 +222,7 @@ SI F rcp(F x) {
#elif defined(__SSE__)
return map(x, _mm_rcp_ps);
#elif defined(__ARM_NEON)
return map(x, +[](float32x4_t v) {
return map(x, [](float32x4_t v) {
auto est = vrecpeq_f32(v);
return vrecpsq_f32(v,est)*est;
});
@ -224,7 +238,7 @@ SI F sqrt_(F x) {
#elif defined(__aarch64__)
return map(x, vsqrtq_f32);
#elif defined(__ARM_NEON)
return map(x, +[](float32x4_t v) {
return map(x, [](float32x4_t v) {
auto est = vrsqrteq_f32(v); // Estimate and two refinement steps for est = rsqrt(v).
est *= vrsqrtsq_f32(v,est*est);
est *= vrsqrtsq_f32(v,est*est);
@ -242,9 +256,9 @@ SI F floor_(F x) {
#if defined(__aarch64__)
return map(x, vrndmq_f32);
#elif defined(__AVX2__)
return map(x, +[](__m256 v){ return _mm256_floor_ps(v); }); // _mm256_floor_ps is a macro...
return map(x, [](__m256 v){ return _mm256_floor_ps(v); }); // _mm256_floor_ps is a macro...
#elif defined(__SSE4_1__)
return map(x, +[](__m128 v){ return _mm_floor_ps(v); }); // _mm_floor_ps() is a macro too.
return map(x, [](__m128 v){ return _mm_floor_ps(v); }); // _mm_floor_ps() is a macro too.
#else
F roundtrip = cast<F>(cast<I32>(x));
return roundtrip - if_then_else(roundtrip > x, F(1), F(0));
@ -480,21 +494,13 @@ SI void store(T* ptr, size_t tail, V v) {
}
template<>
F gather(const float* p, U32 ix) {
__m256i lo, hi;
split(ix, &lo, &hi);
return join<F>(_mm256_i32gather_ps(p, lo, 4),
_mm256_i32gather_ps(p, hi, 4));
F gather(const float* ptr, U32 ix) {
return map(ix, [&](__m256i ix) { return _mm256_i32gather_ps(ptr, ix, 4); });
}
template<>
U32 gather(const uint32_t* p, U32 ix) {
__m256i lo, hi;
split(ix, &lo, &hi);
return join<U32>(_mm256_i32gather_epi32(p, lo, 4),
_mm256_i32gather_epi32(p, hi, 4));
U32 gather(const uint32_t* ptr, U32 ix) {
return map(ix, [&](__m256i ix) { return _mm256_i32gather_epi32(ptr, ix, 4); });
}
#else
template <typename V, typename T>