Skip to content

Commit

Permalink
*improve AVX2 optimizations of class ResizerFloatBilinear (part 5: ca…
Browse files Browse the repository at this point in the history 
…se of large scale, channels = 1).
  • Loading branch information
@ermig1979
ermig1979 committed Jan 16, 2025
1 parent 1fc4efe commit 3bafc89
Show file tree
Hide file tree
Showing 3 changed files with 70 additions and 3 deletions.
69 changes: 68 additions & 1 deletion src/Simd/SimdAvx2ResizerBilinear.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -809,6 +809,7 @@ namespace Simd
size_t cn = _param.channels,
cnH = AlignLo(cn, HF), cnTH = cn - cnH, cnLH = cnTH - HF,
cnF = AlignLo(cn, F), cnTF = cn - cnF, cnLF = cnTF - F;
size_t dw = _param.dstW, dw2 = AlignLo(dw, 2), dw4 = AlignLo(dw, 4), dw8 = AlignLo(dw, 8), dw1 = dw - 1;
__m256 _1 = _mm256_set1_ps(1.0f);
if (_rowBuf)
{
Expand Down Expand Up @@ -1005,7 +1006,73 @@ namespace Simd
}
else
{
Sse41::ResizerFloatBilinear::Run(src, srcStride, dst, dstStride);
if (cn > 1)
{
Sse41::ResizerFloatBilinear::Run(src, srcStride, dst, dstStride);
return;
}
for (size_t dy = 0; dy < _param.dstH; dy++, dst += dstStride)
{
__m128 fy1 = _mm_set1_ps(_ay[dy]);
__m128 fy0 = _mm_sub_ps(_mm256_castps256_ps128(_1), fy1);
const float* src0 = src + _iy[dy] * srcStride, * src1 = src0 + srcStride;
if (cn == 1)
{
size_t dx = 0;
if (Avx2::SlowGather)
{
for (; dx < dw8; dx += 8)
{
__m256 fx1 = _mm256_load_ps(_ax.data + dx);
__m256 fx0 = _mm256_sub_ps(_1, fx1);
__m256 s00 = Load(src0 + _ix[dx + 0], src0 + _ix[dx + 1], src0 + _ix[dx + 4], src0 + _ix[dx + 5]);
__m256 s01 = Load(src0 + _ix[dx + 2], src0 + _ix[dx + 3], src0 + _ix[dx + 6], src0 + _ix[dx + 7]);
__m256 r0 = _mm256_fmadd_ps(_mm256_shuffle_ps(s00, s01, 0x88), fx0, _mm256_mul_ps(_mm256_shuffle_ps(s00, s01, 0xDD), fx1));
__m256 s10 = Load(src1 + _ix[dx + 0], src1 + _ix[dx + 1], src1 + _ix[dx + 4], src1 + _ix[dx + 5]);
__m256 s11 = Load(src1 + _ix[dx + 2], src1 + _ix[dx + 3], src1 + _ix[dx + 6], src1 + _ix[dx + 7]);
__m256 r1 = _mm256_fmadd_ps(_mm256_shuffle_ps(s10, s11, 0x88), fx0, _mm256_mul_ps(_mm256_shuffle_ps(s10, s11, 0xDD), fx1));
_mm256_store_ps(dst + dx, _mm256_fmadd_ps(r0, fx0, _mm256_mul_ps(r1, fx1)));
}
}
else
{
for (; dx < dw8; dx += 8)
{
__m256 fx1 = _mm256_load_ps(_ax.data + dx);
__m256 fx0 = _mm256_sub_ps(_1, fx1);
__m256i idx = Avx2::LoadPermuted<true>((__m256i*)(_ix.data + dx));
__m256 s00 = _mm256_castpd_ps(_mm256_i32gather_pd((double*)src0, _mm256_extracti128_si256(idx, 0), 4));
__m256 s01 = _mm256_castpd_ps(_mm256_i32gather_pd((double*)src0, _mm256_extracti128_si256(idx, 1), 4));
__m256 r0 = _mm256_fmadd_ps(_mm256_shuffle_ps(s00, s01, 0x88), fx0, _mm256_mul_ps(_mm256_shuffle_ps(s00, s01, 0xDD), fx1));
__m256 s10 = _mm256_castpd_ps(_mm256_i32gather_pd((double*)src1, _mm256_extracti128_si256(idx, 0), 4));
__m256 s11 = _mm256_castpd_ps(_mm256_i32gather_pd((double*)src1, _mm256_extracti128_si256(idx, 1), 4));
__m256 r1 = _mm256_fmadd_ps(_mm256_shuffle_ps(s10, s11, 0x88), fx0, _mm256_mul_ps(_mm256_shuffle_ps(s10, s11, 0xDD), fx1));
_mm256_store_ps(dst + dx, _mm256_fmadd_ps(r0, fx0, _mm256_mul_ps(r1, fx1)));
}
}
for (; dx < dw4; dx += 4)
{
__m128 fx1 = _mm_loadu_ps(_ax.data + dx);
__m128 fx0 = _mm_sub_ps(_mm256_castps256_ps128(_1), fx1);
__m128 s00 = Sse41::Load(src0 + _ix[dx + 0], src0 + _ix[dx + 1]);
__m128 s01 = Sse41::Load(src0 + _ix[dx + 2], src0 + _ix[dx + 3]);
__m128 r0 = _mm_add_ps(_mm_mul_ps(_mm_shuffle_ps(s00, s01, 0x88), fx0), _mm_mul_ps(_mm_shuffle_ps(s00, s01, 0xDD), fx1));
__m128 s10 = Sse41::Load(src1 + _ix[dx + 0], src1 + _ix[dx + 1]);
__m128 s11 = Sse41::Load(src1 + _ix[dx + 2], src1 + _ix[dx + 3]);
__m128 r1 = _mm_add_ps(_mm_mul_ps(_mm_shuffle_ps(s10, s11, 0x88), fx0), _mm_mul_ps(_mm_shuffle_ps(s10, s11, 0xDD), fx1));
_mm_storeu_ps(dst + dx, _mm_add_ps(_mm_mul_ps(r0, fy0), _mm_mul_ps(r1, fy1)));
}
for (; dx < dw; dx++)
{
size_t os = _ix[dx];
__m128 fx1 = _mm_set1_ps(_ax[dx]);
__m128 fx0 = _mm_sub_ps(_mm256_castps256_ps128(_1), fx1);
__m128 r0 = _mm_add_ps(_mm_mul_ps(_mm_load_ss(src0 + os), fx0), _mm_mul_ps(_mm_load_ss(src0 + os + 1), fx1));
__m128 r1 = _mm_add_ps(_mm_mul_ps(_mm_load_ss(src1 + os), fx0), _mm_mul_ps(_mm_load_ss(src1 + os + 1), fx1));
_mm_store_ss(dst + dx, _mm_add_ps(_mm_mul_ps(r0, fy0), _mm_mul_ps(r1, fy1)));
}
}
}
}
}

Expand Down
2 changes: 1 addition & 1 deletion src/Simd/SimdBaseResizerBilinear.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -334,7 +334,7 @@ namespace Simd
ResizerFloatBilinear::ResizerFloatBilinear(const ResParam & param)
: Resizer(param)
{
_rowBuf = _param.align < 16 || (_param.channels < 4 && _param.align > 16) || _param.dstH >= _param.srcH;
_rowBuf = _param.align < 16 || (_param.channels < 4 && (_param.align > 32 || _param.channels < 2) ) || _param.dstH >= _param.srcH;
#if defined(SIMD_ARM_ENABLE) || defined(SIMD_ARM64_ENABLE)
_rowBuf = true;
#endif
Expand Down
2 changes: 1 addition & 1 deletion src/Test/TestResize.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -244,7 +244,7 @@ namespace Test

bool ResizerAutoTest(const FuncRS & f1, const FuncRS & f2)
{
//return ResizerAutoTest(SimdResizeMethodBilinear, SimdResizeChannelFloat, 10, f1, f2);
//return ResizerAutoTest(SimdResizeMethodBilinear, SimdResizeChannelFloat, 1, f1, f2);
bool result = true;

#if 1
Expand Down

0 comments on commit 3bafc89

Please sign in to comment.