third_party/opencv
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

AVX and SSE optimizations for resize NN

Browse Source
This commit is contained in:
Vitaly Tuzov 2017-05-31 14:06:45 +03:00
parent f935a16ef2
commit 59373a1ae1
1 changed files with 433 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

435
modules/imgproc/src/imgwarp.cpp
View File

@ -417,6 +417,403 @@ private:
resizeNNInvoker& operator=(const resizeNNInvoker&);
};
#if CV_AVX2
class resizeNNInvokerAVX4 :
public ParallelLoopBody
{
public:
resizeNNInvokerAVX4(const Mat& _src, Mat &_dst, int *_x_ofs, int _pix_size4, double _ify) :
ParallelLoopBody(), src(_src), dst(_dst), x_ofs(_x_ofs), pix_size4(_pix_size4),
ify(_ify)
{
}
#pragma optimization_parameter target_arch=AVX
virtual void operator() (const Range& range) const
{
Size ssize = src.size(), dsize = dst.size();
int y, x, pix_size = (int)src.elemSize();
int width = dsize.width;
int avxWidth = width - (width & 0x7);
#if (defined WIN32 || defined _WIN32 || defined __CYGWIN__)
const __declspec(align(64)) __m256i mask = _mm256_set1_epi32(-1);
#elif defined __GNUC__ && __GNUC__ >= 4
const __m256i mask __attribute__((aligned(64))) = _mm256_set1_epi32(-1);
#else
const __m256i mask = _mm256_set1_epi32(-1);
#endif
if(((int64)(dst.data + dst.step) & 0x1f) == 0)
{
for(y = range.start; y < range.end; y++)
{
uchar* D = dst.data + dst.step*y;
uchar* Dstart = D;
int sy = std::min(cvFloor(y*ify), ssize.height-1);
const uchar* S = src.data + sy*src.step;
#pragma unroll(4)
for(x = 0; x < avxWidth; x += 8)
{
#if (defined WIN32 || defined _WIN32 || defined __CYGWIN__)
__declspec(align(64)) const __m256i *addr = (__m256i*)(x_ofs + x);
__declspec(align(64)) __m256i indices = _mm256_lddqu_si256(addr);
__declspec(align(64)) __m256i pixels = _mm256_i32gather_epi32((const int*)S, indices, 1);
#elif defined __GNUC__ && __GNUC__ >= 4
const __m256i *addr __attribute__((aligned(64))) = (__m256i*)(x_ofs + x);
__m256i indices __attribute__((aligned(64))) = _mm256_lddqu_si256(addr);
__m256i pixels __attribute__((aligned(64))) = _mm256_i32gather_epi32((const int*)S, indices, 1);
#else
const __m256i *addr = (__m256i*)(x_ofs + x);
__m256i indices = _mm256_lddqu_si256(addr);
__m256i pixels = _mm256_i32gather_epi32((const int*)S, indices, 1);
#endif
_mm256_maskstore_epi32((int*)D, mask, pixels);
D += 32;
}
for(; x < width; x++)
{
*(int*)(Dstart + x*4) = *(int*)(S + x_ofs[x]);
}
}
}
else
{
for(y = range.start; y < range.end; y++)
{
uchar* D = dst.data + dst.step*y;
uchar* Dstart = D;
int sy = std::min(cvFloor(y*ify), ssize.height-1);
const uchar* S = src.data + sy*src.step;
#pragma unroll(4)
for(x = 0; x < avxWidth; x += 8)
{
#if (defined WIN32 || defined _WIN32 || defined __CYGWIN__)
__declspec(align(64)) const __m256i *addr = (__m256i*)(x_ofs + x);
__declspec(align(64)) __m256i indices = _mm256_lddqu_si256(addr);
__declspec(align(64)) __m256i pixels = _mm256_i32gather_epi32((const int*)S, indices, 1);
#elif defined __GNUC__ && __GNUC__ >= 4
const __m256i *addr __attribute__((aligned(64))) = (__m256i*)(x_ofs + x);
__m256i indices __attribute__((aligned(64))) = _mm256_lddqu_si256(addr);
__m256i pixels __attribute__((aligned(64))) = _mm256_i32gather_epi32((const int*)S, indices, 1);
#else
const __m256i *addr = (__m256i*)(x_ofs + x);
__m256i indices = _mm256_lddqu_si256(addr);
__m256i pixels = _mm256_i32gather_epi32((const int*)S, indices, 1);
#endif
_mm256_storeu_si256((__m256i*)D, pixels);
D += 32;
}
for(; x < width; x++)
{
*(int*)(Dstart + x*4) = *(int*)(S + x_ofs[x]);
}
}
}
}
private:
const Mat src;
Mat dst;
int* x_ofs, pix_size4;
double ify;
resizeNNInvokerAVX4(const resizeNNInvokerAVX4&);
resizeNNInvokerAVX4& operator=(const resizeNNInvokerAVX4&);
};
class resizeNNInvokerAVX2 :
public ParallelLoopBody
{
public:
resizeNNInvokerAVX2(const Mat& _src, Mat &_dst, int *_x_ofs, int _pix_size4, double _ify) :
ParallelLoopBody(), src(_src), dst(_dst), x_ofs(_x_ofs), pix_size4(_pix_size4),
ify(_ify)
{
}
#pragma optimization_parameter target_arch=AVX
virtual void operator() (const Range& range) const
{
Size ssize = src.size(), dsize = dst.size();
int y, x, pix_size = (int)src.elemSize();
int width = dsize.width;
//int avxWidth = (width - 1) - ((width - 1) & 0x7);
int avxWidth = width - (width & 0xf);
#if (defined WIN32 || defined _WIN32 || defined __CYGWIN__)
const __declspec(align(64)) __m256i mask = _mm256_set1_epi32(-1);
const __declspec(align(64)) __m256i shuffle_mask = _mm256_set_epi8(15,14,11,10,13,12,9,8,7,6,3,2,5,4,1,0,
15,14,11,10,13,12,9,8,7,6,3,2,5,4,1,0);
const __declspec(align(64)) __m256i permute_mask = _mm256_set_epi32(7, 5, 3, 1, 6, 4, 2, 0);
const __declspec(align(64)) __m256i shift_shuffle_mask = _mm256_set_epi8(13,12,15,14,9,8,11,10,5,4,7,6,1,0,3,2,
13,12,15,14,9,8,11,10,5,4,7,6,1,0,3,2);
#elif defined __GNUC__ && __GNUC__ >= 4
const __m256i mask __attribute__((aligned(64))) = _mm256_set1_epi32(-1);
const __m256i shuffle_mask __attribute__((aligned(64))) = _mm256_set_epi8(15,14,11,10,13,12,9,8,7,6,3,2,5,4,1,0,
15,14,11,10,13,12,9,8,7,6,3,2,5,4,1,0);
const __m256i permute_mask __attribute__((aligned(64))) = _mm256_set_epi32(7, 5, 3, 1, 6, 4, 2, 0);
const __m256i shift_shuffle_mask __attribute__((aligned(64))) = _mm256_set_epi8(13,12,15,14,9,8,11,10,5,4,7,6,1,0,3,2,
13,12,15,14,9,8,11,10,5,4,7,6,1,0,3,2);
#else
const __m256i mask = _mm256_set1_epi32(-1);
const __m256i shuffle_mask = _mm256_set_epi8(15,14,11,10,13,12,9,8,7,6,3,2,5,4,1,0,
15,14,11,10,13,12,9,8,7,6,3,2,5,4,1,0);
const __m256i permute_mask = _mm256_set_epi32(7, 5, 3, 1, 6, 4, 2, 0);
const __m256i shift_shuffle_mask = _mm256_set_epi8(13,12,15,14,9,8,11,10,5,4,7,6,1,0,3,2,
13,12,15,14,9,8,11,10,5,4,7,6,1,0,3,2);
#endif
if(((int64)(dst.data + dst.step) & 0x1f) == 0)
{
for(y = range.start; y < range.end; y++)
{
uchar* D = dst.data + dst.step*y;
uchar* Dstart = D;
int sy = std::min(cvFloor(y*ify), ssize.height-1);
const uchar* S = src.data + sy*src.step;
const uchar* S2 = S - 2;
#pragma unroll(4)
for(x = 0; x < avxWidth; x += 16)
{
#if (defined WIN32 || defined _WIN32 || defined __CYGWIN__)
__declspec(align(64)) const __m256i *addr = (__m256i*)(x_ofs + x);
__declspec(align(64)) __m256i indices = _mm256_lddqu_si256(addr);
__declspec(align(64)) __m256i pixels1 = _mm256_i32gather_epi32((const int*)S, indices, 1);
__declspec(align(64)) const __m256i *addr2 = (__m256i*)(x_ofs + x + 8);
__declspec(align(64)) __m256i indices2 = _mm256_lddqu_si256(addr2);
__declspec(align(64)) __m256i pixels2 = _mm256_i32gather_epi32((const int*)S2, indices2, 1);
__declspec(align(64)) __m256i unpacked = _mm256_blend_epi16(pixels1, pixels2, 0xaa);
__declspec(align(64)) __m256i bytes_shuffled = _mm256_shuffle_epi8(unpacked, shuffle_mask);
__declspec(align(64)) __m256i ints_permuted = _mm256_permutevar8x32_epi32(bytes_shuffled, permute_mask);
#elif defined __GNUC__ && __GNUC__ >= 4
const __m256i *addr __attribute__((aligned(64))) = (__m256i*)(x_ofs + x);
__m256i indices __attribute__((aligned(64))) = _mm256_lddqu_si256(addr);
__m256i pixels1 __attribute__((aligned(64))) = _mm256_i32gather_epi32((const int*)S, indices, 1);
const __m256i *addr2 __attribute__((aligned(64))) = (__m256i*)(x_ofs + x + 8);
__m256i indices2 __attribute__((aligned(64))) = _mm256_lddqu_si256(addr2);
__m256i pixels2 __attribute__((aligned(64))) = _mm256_i32gather_epi32((const int*)S2, indices2, 1);
__m256i unpacked __attribute__((aligned(64))) = _mm256_blend_epi16(pixels1, pixels2, 0xaa);
__m256i bytes_shuffled __attribute__((aligned(64))) = _mm256_shuffle_epi8(unpacked, shuffle_mask);
__m256i ints_permuted __attribute__((aligned(64))) = _mm256_permutevar8x32_epi32(bytes_shuffled, permute_mask);
#else
const __m256i *addr = (__m256i*)(x_ofs + x);
__m256i indices = _mm256_lddqu_si256(addr);
__m256i pixels1 = _mm256_i32gather_epi32((const int*)S, indices, 1);
const __m256i *addr2 = (__m256i*)(x_ofs + x + 8);
__m256i indices2 = _mm256_lddqu_si256(addr2);
__m256i pixels2 = _mm256_i32gather_epi32((const int*)S2, indices2, 1);
__m256i unpacked = _mm256_blend_epi16(pixels1, pixels2, 0xaa);
__m256i bytes_shuffled = _mm256_shuffle_epi8(unpacked, shuffle_mask);
__m256i ints_permuted = _mm256_permutevar8x32_epi32(bytes_shuffled, permute_mask);
#endif
_mm256_maskstore_epi32((int*)D, mask, ints_permuted);
D += 32;
}
for(; x < width; x++)
{
*(ushort*)(Dstart + x*2) = *(ushort*)(S + x_ofs[x]);
}
}
}
else
{
for(y = range.start; y < range.end; y++)
{
uchar* D = dst.data + dst.step*y;
uchar* Dstart = D;
int sy = std::min(cvFloor(y*ify), ssize.height-1);
const uchar* S = src.data + sy*src.step;
const uchar* S2 = S - 2;
#pragma unroll(4)
for(x = 0; x < avxWidth; x += 16)
{
#if (defined WIN32 || defined _WIN32 || defined __CYGWIN__)
__declspec(align(64)) const __m256i *addr = (__m256i*)(x_ofs + x);
__declspec(align(64)) __m256i indices = _mm256_lddqu_si256(addr);
__declspec(align(64)) __m256i pixels1 = _mm256_i32gather_epi32((const int*)S, indices, 1);
__declspec(align(64)) const __m256i *addr2 = (__m256i*)(x_ofs + x + 8);
__declspec(align(64)) __m256i indices2 = _mm256_lddqu_si256(addr2);
__declspec(align(64)) __m256i pixels2 = _mm256_i32gather_epi32((const int*)S2, indices2, 1);
__declspec(align(64)) __m256i unpacked = _mm256_blend_epi16(pixels1, pixels2, 0xaa);
__declspec(align(64)) __m256i bytes_shuffled = _mm256_shuffle_epi8(unpacked, shuffle_mask);
__declspec(align(64)) __m256i ints_permuted = _mm256_permutevar8x32_epi32(bytes_shuffled, permute_mask);
#elif defined __GNUC__ && __GNUC__ >= 4
const __m256i *addr __attribute__((aligned(64))) = (__m256i*)(x_ofs + x);
__m256i indices __attribute__((aligned(64))) = _mm256_lddqu_si256(addr);
__m256i pixels1 __attribute__((aligned(64))) = _mm256_i32gather_epi32((const int*)S, indices, 1);
const __m256i *addr2 __attribute__((aligned(64))) = (__m256i*)(x_ofs + x + 8);
__m256i indices2 __attribute__((aligned(64))) = _mm256_lddqu_si256(addr2);
__m256i pixels2 __attribute__((aligned(64))) = _mm256_i32gather_epi32((const int*)S2, indices2, 1);
__m256i unpacked __attribute__((aligned(64))) = _mm256_blend_epi16(pixels1, pixels2, 0xaa);
__m256i bytes_shuffled __attribute__((aligned(64))) = _mm256_shuffle_epi8(unpacked, shuffle_mask);
__m256i ints_permuted __attribute__((aligned(64))) = _mm256_permutevar8x32_epi32(bytes_shuffled, permute_mask);
#else
const __m256i *addr = (__m256i*)(x_ofs + x);
__m256i indices = _mm256_lddqu_si256(addr);
__m256i pixels1 = _mm256_i32gather_epi32((const int*)S, indices, 1);
const __m256i *addr2 = (__m256i*)(x_ofs + x + 8);
__m256i indices2 = _mm256_lddqu_si256(addr2);
__m256i pixels2 = _mm256_i32gather_epi32((const int*)S2, indices2, 1);
__m256i unpacked = _mm256_blend_epi16(pixels1, pixels2, 0xaa);
__m256i bytes_shuffled = _mm256_shuffle_epi8(unpacked, shuffle_mask);
__m256i ints_permuted = _mm256_permutevar8x32_epi32(bytes_shuffled, permute_mask);
#endif
_mm256_storeu_si256((__m256i*)D, ints_permuted);
D += 32;
}
for(; x < width; x++)
{
*(ushort*)(Dstart + x*2) = *(ushort*)(S + x_ofs[x]);
}
}
}
}
private:
const Mat src;
Mat dst;
int* x_ofs, pix_size4;
double ify;
resizeNNInvokerAVX2(const resizeNNInvokerAVX2&);
resizeNNInvokerAVX2& operator=(const resizeNNInvokerAVX2&);
};
#endif
#if CV_SSE4_1
class resizeNNInvokerSSE2 :
public ParallelLoopBody
{
public:
resizeNNInvokerSSE2(const Mat& _src, Mat &_dst, int *_x_ofs, int _pix_size4, double _ify) :
ParallelLoopBody(), src(_src), dst(_dst), x_ofs(_x_ofs), pix_size4(_pix_size4),
ify(_ify)
{
}
#pragma optimization_parameter target_arch=SSE4.2
virtual void operator() (const Range& range) const
{
Size ssize = src.size(), dsize = dst.size();
int y, x, pix_size = (int)src.elemSize();
int width = dsize.width;
int sseWidth = width - (width & 0x7);
for(y = range.start; y < range.end; y++)
{
uchar* D = dst.data + dst.step*y;
uchar* Dstart = D;
int sy = std::min(cvFloor(y*ify), ssize.height-1);
const uchar* S = src.data + sy*src.step;
#if (defined WIN32 || defined _WIN32 || defined __CYGWIN__)
__declspec(align(64)) __m128i pixels = _mm_set1_epi16(0);
#elif defined __GNUC__ && __GNUC__ >= 4
__m128i pixels __attribute__((aligned(64))) = _mm_set1_epi16(0);
#else
__m128i pixels = _mm_set1_epi16(0);
#endif
for(x = 0; x < sseWidth; x += 8)
{
ushort imm = *(ushort*)(S + x_ofs[x + 0]);
pixels = _mm_insert_epi16(pixels, imm, 0);
imm = *(ushort*)(S + x_ofs[x + 1]);
pixels = _mm_insert_epi16(pixels, imm, 1);
imm = *(ushort*)(S + x_ofs[x + 2]);
pixels = _mm_insert_epi16(pixels, imm, 2);
imm = *(ushort*)(S + x_ofs[x + 3]);
pixels = _mm_insert_epi16(pixels, imm, 3);
imm = *(ushort*)(S + x_ofs[x + 4]);
pixels = _mm_insert_epi16(pixels, imm, 4);
imm = *(ushort*)(S + x_ofs[x + 5]);
pixels = _mm_insert_epi16(pixels, imm, 5);
imm = *(ushort*)(S + x_ofs[x + 6]);
pixels = _mm_insert_epi16(pixels, imm, 6);
imm = *(ushort*)(S + x_ofs[x + 7]);
pixels = _mm_insert_epi16(pixels, imm, 7);
_mm_storeu_si128((__m128i*)D, pixels);
D += 16;
}
for(; x < width; x++)
{
*(ushort*)(Dstart + x*2) = *(ushort*)(S + x_ofs[x]);
}
}
}
private:
const Mat src;
Mat dst;
int* x_ofs, pix_size4;
double ify;
resizeNNInvokerSSE2(const resizeNNInvokerSSE2&);
resizeNNInvokerSSE2& operator=(const resizeNNInvokerSSE2&);
};
class resizeNNInvokerSSE4 :
public ParallelLoopBody
{
public:
resizeNNInvokerSSE4(const Mat& _src, Mat &_dst, int *_x_ofs, int _pix_size4, double _ify) :
ParallelLoopBody(), src(_src), dst(_dst), x_ofs(_x_ofs), pix_size4(_pix_size4),
ify(_ify)
{
}
#pragma optimization_parameter target_arch=SSE4.2
virtual void operator() (const Range& range) const
{
Size ssize = src.size(), dsize = dst.size();
int y, x, pix_size = (int)src.elemSize();
int width = dsize.width;
int sseWidth = width - (width & 0x3);
for(y = range.start; y < range.end; y++)
{
uchar* D = dst.data + dst.step*y;
uchar* Dstart = D;
int sy = std::min(cvFloor(y*ify), ssize.height-1);
const uchar* S = src.data + sy*src.step;
#if (defined WIN32 || defined _WIN32 || defined __CYGWIN__)
__declspec(align(64)) __m128i pixels = _mm_set1_epi16(0);
#elif defined __GNUC__ && __GNUC__ >= 4
__m128i pixels __attribute__((aligned(64))) = _mm_set1_epi16(0);
#else
__m128i pixels = _mm_set1_epi16(0);
#endif
for(x = 0; x < sseWidth; x += 4)
{
int imm = *(int*)(S + x_ofs[x + 0]);
pixels = _mm_insert_epi32(pixels, imm, 0);
imm = *(int*)(S + x_ofs[x + 1]);
pixels = _mm_insert_epi32(pixels, imm, 1);
imm = *(int*)(S + x_ofs[x + 2]);
pixels = _mm_insert_epi32(pixels, imm, 2);
imm = *(int*)(S + x_ofs[x + 3]);
pixels = _mm_insert_epi32(pixels, imm, 3);
_mm_storeu_si128((__m128i*)D, pixels);
D += 16;
}
for(; x < width; x++)
{
*(int*)(Dstart + x*4) = *(int*)(S + x_ofs[x]);
}
}
}
private:
const Mat src;
Mat dst;
int* x_ofs, pix_size4;
double ify;
resizeNNInvokerSSE4(const resizeNNInvokerSSE4&);
resizeNNInvokerSSE4& operator=(const resizeNNInvokerSSE4&);
};
#endif
static void
resizeNN( const Mat& src, Mat& dst, double fx, double fy )
{
@ -435,8 +832,42 @@ resizeNN( const Mat& src, Mat& dst, double fx, double fy )
}
Range range(0, dsize.height);
resizeNNInvoker invoker(src, dst, x_ofs, pix_size4, ify);
parallel_for_(range, invoker, dst.total()/(double)(1<<16));
#if CV_AVX2
if(checkHardwareSupport(CV_CPU_AVX2) && ((pix_size == 2) || (pix_size == 4)))
{
if(pix_size == 2)
{
resizeNNInvokerAVX2 invoker(src, dst, x_ofs, pix_size4, ify);
parallel_for_(range, invoker, dst.total()/(double)(1<<16));
}
else if (pix_size == 4)
{
resizeNNInvokerAVX4 invoker(src, dst, x_ofs, pix_size4, ify);
parallel_for_(range, invoker, dst.total()/(double)(1<<16));
}
}
else
#endif
#if CV_SSE4_1
if(checkHardwareSupport(CV_CPU_SSE4_1) && ((pix_size == 2) || (pix_size == 4)))
{
if(pix_size == 2)
{
resizeNNInvokerSSE2 invoker(src, dst, x_ofs, pix_size4, ify);
parallel_for_(range, invoker, dst.total()/(double)(1<<16));
}
else if(pix_size == 4)
{
resizeNNInvokerSSE4 invoker(src, dst, x_ofs, pix_size4, ify);
parallel_for_(range, invoker, dst.total()/(double)(1<<16));
}
}
else
#endif
{
resizeNNInvoker invoker(src, dst, x_ofs, pix_size4, ify);
parallel_for_(range, invoker, dst.total()/(double)(1<<16));
}
}