Skip to content

O
opencv
Commit d2aadabc authored by Vitaly Tuzov's avatar Vitaly Tuzov Committed by Alexander Alekhin
Browse files
Options

Merge pull request #14743 from terfendail:wui512_fixvswarn 

Fix for MSVS2019 build warnings (#14743)

* AVX512 arch support for MSVS

* Fix for MSVS2019 build warnings: updated integral() AVX512 implementation

* Fix for MSVS2019 build warnings: reworked v_rotate_right AVX512 implementation

* fix indentation
parent b3e6c97d
Hide whitespace changes
Inline Side-by-side
Showing with 101 additions and 84 deletions
cmake/OpenCVCompilerOptimizations.cmake
View file @ d2aadabc
...@@ -294,6 +294,7 @@ if(X86 OR X86_64) ...@@ -294,6 +294,7 @@ if(X86 OR X86_64)
else() else()
ocv_update(CPU_SSE_SUPPORTED ON) ocv_update(CPU_SSE_SUPPORTED ON)
ocv_update(CPU_SSE2_SUPPORTED ON) ocv_update(CPU_SSE2_SUPPORTED ON)
ocv_update(CPU_AVX_512F_FLAGS_ON "/arch:AVX512")
endif() endif()
# Other instruction sets are supported by default since MSVC 2008 at least # Other instruction sets are supported by default since MSVC 2008 at least
else() else()
......
modules/core/include/opencv2/core/hal/intrin_avx512.hpp
View file @ d2aadabc
...@@ -893,37 +893,58 @@ OPENCV_HAL_IMPL_AVX512_BIN_FUNC(v_min, v_float64x8, _mm512_min_pd) ...@@ -893,37 +893,58 @@ OPENCV_HAL_IMPL_AVX512_BIN_FUNC(v_min, v_float64x8, _mm512_min_pd)
OPENCV_HAL_IMPL_AVX512_BIN_FUNC(v_max, v_float64x8, _mm512_max_pd) OPENCV_HAL_IMPL_AVX512_BIN_FUNC(v_max, v_float64x8, _mm512_max_pd)
/** Rotate **/ /** Rotate **/
template<int imm> namespace {
inline v_int8x64 v_rotate_right(const v_int8x64& a, const v_int8x64& b) template<bool prec, int imm4, bool part, int imm32>
{ struct _v_rotate_right { static inline v_int8x64 eval(const v_int8x64&, const v_int8x64&) { return v_int8x64(); }};
if (imm == 0) return a; template<int imm4, int imm32>
if (imm == 64) return b; struct _v_rotate_right<true, imm4, false, imm32> { static inline v_int8x64 eval(const v_int8x64& a, const v_int8x64& b)
if (imm >= 128) return v_int8x64(); {
return v_int8x64(_mm512_or_si512(_mm512_srli_epi32(_mm512_alignr_epi32(b.val, a.val, imm32 ), imm4 *8),
_mm512_slli_epi32(_mm512_alignr_epi32(b.val, a.val, imm32 + 1), (4-imm4)*8)));
}};
template<int imm4>
struct _v_rotate_right<true, imm4, false, 15> { static inline v_int8x64 eval(const v_int8x64& a, const v_int8x64& b)
{
return v_int8x64(_mm512_or_si512(_mm512_srli_epi32(_mm512_alignr_epi32(b.val, a.val, 15), imm4 *8),
_mm512_slli_epi32( b.val, (4-imm4)*8)));
}};
template<int imm4, int imm32>
struct _v_rotate_right<true, imm4, true, imm32> { static inline v_int8x64 eval(const v_int8x64&, const v_int8x64& b)
{
return v_int8x64(_mm512_or_si512(_mm512_srli_epi32(_mm512_alignr_epi32(_mm512_setzero_si512(), b.val, imm32 - 16), imm4 *8),
_mm512_slli_epi32(_mm512_alignr_epi32(_mm512_setzero_si512(), b.val, imm32 - 15), (4-imm4)*8)));
}};
template<int imm4>
struct _v_rotate_right<true, imm4, true, 31> { static inline v_int8x64 eval(const v_int8x64&, const v_int8x64& b)
{ return v_int8x64(_mm512_srli_epi32(_mm512_alignr_epi32(_mm512_setzero_si512(), b.val, 15), imm4*8)); }};
template<int imm32>
struct _v_rotate_right<false, 0, false, imm32> { static inline v_int8x64 eval(const v_int8x64& a, const v_int8x64& b)
{ return v_int8x64(_mm512_alignr_epi32(b.val, a.val, imm32)); }};
template<>
struct _v_rotate_right<false, 0, false, 0> { static inline v_int8x64 eval(const v_int8x64& a, const v_int8x64&) { return a; }};
template<int imm32>
struct _v_rotate_right<false, 0, true, imm32> { static inline v_int8x64 eval(const v_int8x64&, const v_int8x64& b)
{ return v_int8x64(_mm512_alignr_epi32(_mm512_setzero_si512(), b.val, imm32 - 16)); }};
template<>
struct _v_rotate_right<false, 0, true, 16> { static inline v_int8x64 eval(const v_int8x64&, const v_int8x64& b) { return b; }};
template<>
struct _v_rotate_right<false, 0, true, 32> { static inline v_int8x64 eval(const v_int8x64&, const v_int8x64&) { return v_int8x64(); }};
}
template<int imm> inline v_int8x64 v_rotate_right(const v_int8x64& a, const v_int8x64& b)
{
return imm >= 128 ? v_int8x64() :
#if CV_AVX_512VBMI #if CV_AVX_512VBMI
return v_int8x64(_mm512_permutex2var_epi8(a.val, v_int8x64(_mm512_permutex2var_epi8(a.val,
_v512_set_epu8(0x3f + imm,0x3e + imm,0x3d + imm,0x3c + imm,0x3b + imm,0x3a + imm,0x39 + imm,0x38 + imm, _v512_set_epu8(0x3f + imm, 0x3e + imm, 0x3d + imm, 0x3c + imm, 0x3b + imm, 0x3a + imm, 0x39 + imm, 0x38 + imm,
0x37 + imm,0x36 + imm,0x35 + imm,0x34 + imm,0x33 + imm,0x32 + imm,0x31 + imm,0x30 + imm, 0x37 + imm, 0x36 + imm, 0x35 + imm, 0x34 + imm, 0x33 + imm, 0x32 + imm, 0x31 + imm, 0x30 + imm,
0x2f + imm,0x2e + imm,0x2d + imm,0x2c + imm,0x2b + imm,0x2a + imm,0x29 + imm,0x28 + imm, 0x2f + imm, 0x2e + imm, 0x2d + imm, 0x2c + imm, 0x2b + imm, 0x2a + imm, 0x29 + imm, 0x28 + imm,
0x27 + imm,0x26 + imm,0x25 + imm,0x24 + imm,0x23 + imm,0x22 + imm,0x21 + imm,0x20 + imm, 0x27 + imm, 0x26 + imm, 0x25 + imm, 0x24 + imm, 0x23 + imm, 0x22 + imm, 0x21 + imm, 0x20 + imm,
0x1f + imm,0x1e + imm,0x1d + imm,0x1c + imm,0x1b + imm,0x1a + imm,0x19 + imm,0x18 + imm, 0x1f + imm, 0x1e + imm, 0x1d + imm, 0x1c + imm, 0x1b + imm, 0x1a + imm, 0x19 + imm, 0x18 + imm,
0x17 + imm,0x16 + imm,0x15 + imm,0x14 + imm,0x13 + imm,0x12 + imm,0x11 + imm,0x10 + imm, 0x17 + imm, 0x16 + imm, 0x15 + imm, 0x14 + imm, 0x13 + imm, 0x12 + imm, 0x11 + imm, 0x10 + imm,
0x0f + imm,0x0e + imm,0x0d + imm,0x0c + imm,0x0b + imm,0x0a + imm,0x09 + imm,0x08 + imm, 0x0f + imm, 0x0e + imm, 0x0d + imm, 0x0c + imm, 0x0b + imm, 0x0a + imm, 0x09 + imm, 0x08 + imm,
0x07 + imm,0x06 + imm,0x05 + imm,0x04 + imm,0x03 + imm,0x02 + imm,0x01 + imm,0x00 + imm), b.val)); 0x07 + imm, 0x06 + imm, 0x05 + imm, 0x04 + imm, 0x03 + imm, 0x02 + imm, 0x01 + imm, 0x00 + imm), b.val));
#else #else
__m512i pre = _mm512_alignr_epi32(b.val, a.val, imm/4); _v_rotate_right<imm%4!=0, imm%4, (imm/4 > 15), imm/4>::eval(a, b);
if (imm % 4)
{
__m512i post;
if (imm/4 < 15)
post = _mm512_alignr_epi32(b.val, a.val, imm/4 + 1);
else if (imm/4 == 15)
post = b.val;
else
post = _mm512_alignr_epi32(_mm512_setzero_si512(), b.val, imm/4 - 15);
return v_int8x64(_mm512_or_si512(_mm512_srli_epi32(pre, (imm % 4)*8), _mm512_slli_epi32(post, (4 - imm % 4)*8)));
}
else
return v_int8x64(pre);
#endif #endif
} }
template<int imm> template<int imm>
...@@ -943,8 +964,7 @@ inline v_int8x64 v_rotate_left(const v_int8x64& a, const v_int8x64& b) ...@@ -943,8 +964,7 @@ inline v_int8x64 v_rotate_left(const v_int8x64& a, const v_int8x64& b)
0x4f - imm,0x4e - imm,0x4d - imm,0x4c - imm,0x4b - imm,0x4a - imm,0x49 - imm,0x48 - imm, 0x4f - imm,0x4e - imm,0x4d - imm,0x4c - imm,0x4b - imm,0x4a - imm,0x49 - imm,0x48 - imm,
0x47 - imm,0x46 - imm,0x45 - imm,0x44 - imm,0x43 - imm,0x42 - imm,0x41 - imm,0x40 - imm), a.val)); 0x47 - imm,0x46 - imm,0x45 - imm,0x44 - imm,0x43 - imm,0x42 - imm,0x41 - imm,0x40 - imm), a.val));
#else #else
if (imm < 64) return v_rotate_right<64 - imm>(b, a); return imm < 64 ? v_rotate_right<64 - imm>(b, a) : v_rotate_right<128 - imm>(v512_setzero_s8(), b);
else return v_rotate_right<128 - imm>(v512_setzero_s8(), b);
#endif #endif
} }
template<int imm> template<int imm>
...@@ -986,54 +1006,50 @@ inline v_int8x64 v_rotate_left(const v_int8x64& a) ...@@ -986,54 +1006,50 @@ inline v_int8x64 v_rotate_left(const v_int8x64& a)
#endif #endif
} }
#define OPENCV_HAL_IMPL_AVX512_ROTATE_PM(_Tpvec, suffix) \ #define OPENCV_HAL_IMPL_AVX512_ROTATE_PM(_Tpvec, suffix) \
template<int imm> inline _Tpvec v_rotate_left(const _Tpvec& a, const _Tpvec& b) \ template<int imm> inline _Tpvec v_rotate_left(const _Tpvec& a, const _Tpvec& b) \
{ return v_reinterpret_as_##suffix(v_rotate_left<imm * sizeof(_Tpvec::lane_type)>(v_reinterpret_as_s8(a), v_reinterpret_as_s8(b))); } \ { return v_reinterpret_as_##suffix(v_rotate_left<imm * sizeof(_Tpvec::lane_type)>(v_reinterpret_as_s8(a), v_reinterpret_as_s8(b))); } \
template<int imm> inline _Tpvec v_rotate_right(const _Tpvec& a, const _Tpvec& b) \ template<int imm> inline _Tpvec v_rotate_right(const _Tpvec& a, const _Tpvec& b) \
{ return v_reinterpret_as_##suffix(v_rotate_right<imm * sizeof(_Tpvec::lane_type)>(v_reinterpret_as_s8(a), v_reinterpret_as_s8(b))); } \ { return v_reinterpret_as_##suffix(v_rotate_right<imm * sizeof(_Tpvec::lane_type)>(v_reinterpret_as_s8(a), v_reinterpret_as_s8(b))); } \
template<int imm> inline _Tpvec v_rotate_left(const _Tpvec& a) \ template<int imm> inline _Tpvec v_rotate_left(const _Tpvec& a) \
{ return v_reinterpret_as_##suffix(v_rotate_left<imm * sizeof(_Tpvec::lane_type)>(v_reinterpret_as_s8(a))); } \ { return v_reinterpret_as_##suffix(v_rotate_left<imm * sizeof(_Tpvec::lane_type)>(v_reinterpret_as_s8(a))); } \
template<int imm> inline _Tpvec v_rotate_right(const _Tpvec& a) \ template<int imm> inline _Tpvec v_rotate_right(const _Tpvec& a) \
{ return v_reinterpret_as_##suffix(v_rotate_right<imm * sizeof(_Tpvec::lane_type)>(v_reinterpret_as_s8(a))); } { return v_reinterpret_as_##suffix(v_rotate_right<imm * sizeof(_Tpvec::lane_type)>(v_reinterpret_as_s8(a))); }
#define OPENCV_HAL_IMPL_AVX512_ROTATE_EC(_Tpvec, suffix) \ #define OPENCV_HAL_IMPL_AVX512_ROTATE_EC(_Tpvec, suffix) \
template<int imm> \ template<int imm> \
inline _Tpvec v_rotate_left(const _Tpvec& a, const _Tpvec& b) \ inline _Tpvec v_rotate_left(const _Tpvec& a, const _Tpvec& b) \
{ \ { \
enum { SHIFT2 = _Tpvec::nlanes - imm }; \ enum { SHIFT2 = (_Tpvec::nlanes - imm) }; \
enum { MASK = (1 << _Tpvec::nlanes) - 1 }; \ enum { MASK = ((1 << _Tpvec::nlanes) - 1) }; \
if (imm == 0) return a; \ if (imm == 0) return a; \
if (imm == _Tpvec::nlanes) return b; \ if (imm == _Tpvec::nlanes) return b; \
if (imm >= 2*_Tpvec::nlanes) return _Tpvec(); \ if (imm >= 2*_Tpvec::nlanes) return _Tpvec(); \
return _Tpvec(_mm512_mask_expand_##suffix(_mm512_maskz_compress_##suffix((MASK << SHIFT2)&MASK, b.val), (MASK << imm)&MASK, a.val)); \ return _Tpvec(_mm512_mask_expand_##suffix(_mm512_maskz_compress_##suffix((MASK << SHIFT2)&MASK, b.val), (MASK << (imm))&MASK, a.val)); \
} \ } \
template<int imm> \ template<int imm> \
inline _Tpvec v_rotate_right(const _Tpvec& a, const _Tpvec& b) \ inline _Tpvec v_rotate_right(const _Tpvec& a, const _Tpvec& b) \
{ \ { \
enum { SHIFT2 = _Tpvec::nlanes - imm }; \ enum { SHIFT2 = (_Tpvec::nlanes - imm) }; \
enum { MASK = (1 << _Tpvec::nlanes) - 1 }; \ enum { MASK = ((1 << _Tpvec::nlanes) - 1) }; \
if (imm == 0) return a; \ if (imm == 0) return a; \
if (imm == _Tpvec::nlanes) return b; \ if (imm == _Tpvec::nlanes) return b; \
if (imm >= 2*_Tpvec::nlanes) return _Tpvec(); \ if (imm >= 2*_Tpvec::nlanes) return _Tpvec(); \
return _Tpvec(_mm512_mask_expand_##suffix(_mm512_maskz_compress_##suffix((MASK << imm)&MASK, a.val), (MASK << SHIFT2)&MASK, b.val)); \ return _Tpvec(_mm512_mask_expand_##suffix(_mm512_maskz_compress_##suffix((MASK << (imm))&MASK, a.val), (MASK << SHIFT2)&MASK, b.val)); \
} \ } \
template<int imm> \ template<int imm> \
inline _Tpvec v_rotate_left(const _Tpvec& a) \ inline _Tpvec v_rotate_left(const _Tpvec& a) \
{ \ { \
enum { SHIFT2 = _Tpvec::nlanes - imm }; \ if (imm == 0) return a; \
enum { MASK = (1 << _Tpvec::nlanes) - 1 }; \ if (imm >= _Tpvec::nlanes) return _Tpvec(); \
if (imm == 0) return a; \ return _Tpvec(_mm512_maskz_expand_##suffix((1 << _Tpvec::nlanes) - (1 << (imm)), a.val)); \
if (imm >= _Tpvec::nlanes) return _Tpvec(); \ } \
return _Tpvec(_mm512_maskz_expand_##suffix((MASK << imm)&MASK, a.val)); \ template<int imm> \
} \ inline _Tpvec v_rotate_right(const _Tpvec& a) \
template<int imm> \ { \
inline _Tpvec v_rotate_right(const _Tpvec& a) \ if (imm == 0) return a; \
{ \ if (imm >= _Tpvec::nlanes) return _Tpvec(); \
enum { SHIFT2 = _Tpvec::nlanes - imm }; \ return _Tpvec(_mm512_maskz_compress_##suffix((1 << _Tpvec::nlanes) - (1 << (imm)), a.val)); \
enum { MASK = (1 << _Tpvec::nlanes) - 1 }; \
if (imm == 0) return a; \
if (imm >= _Tpvec::nlanes) return _Tpvec(); \
return _Tpvec(_mm512_maskz_compress_##suffix((MASK << imm)&MASK, a.val)); \
} }
OPENCV_HAL_IMPL_AVX512_ROTATE_PM(v_uint8x64, u8) OPENCV_HAL_IMPL_AVX512_ROTATE_PM(v_uint8x64, u8)
...@@ -1175,7 +1191,7 @@ inline unsigned v_reduce_sad(const v_uint8x64& a, const v_uint8x64& b) ...@@ -1175,7 +1191,7 @@ inline unsigned v_reduce_sad(const v_uint8x64& a, const v_uint8x64& b)
} }
inline unsigned v_reduce_sad(const v_int8x64& a, const v_int8x64& b) inline unsigned v_reduce_sad(const v_int8x64& a, const v_int8x64& b)
{ {
__m512i val = _mm512_set1_epi8(0x80); __m512i val = _mm512_set1_epi8(-128);
val = _mm512_sad_epu8(_mm512_add_epi8(a.val, val), _mm512_add_epi8(b.val, val)); val = _mm512_sad_epu8(_mm512_add_epi8(a.val, val), _mm512_add_epi8(b.val, val));
__m256i half = _mm256_add_epi32(_v512_extract_low(val), _v512_extract_high(val)); __m256i half = _mm256_add_epi32(_v512_extract_low(val), _v512_extract_high(val));
__m128i quarter = _mm_add_epi32(_mm256_castsi256_si128(half), _mm256_extracti128_si256(half, 1)); __m128i quarter = _mm_add_epi32(_mm256_castsi256_si128(half), _mm256_extracti128_si256(half, 1));
......
modules/imgproc/src/sumpixels.avx512_skx.cpp
View file @ d2aadabc
...@@ -79,9 +79,9 @@ public: ...@@ -79,9 +79,9 @@ public:
{ {
// Note the negative index is because the sums/sqsums pointers point to the first real pixel // Note the negative index is because the sums/sqsums pointers point to the first real pixel
// after the border pixel so we have to look backwards // after the border pixel so we have to look backwards
_mm512_mask_storeu_epi64(&sums[-num_channels], (1<<num_channels)-1, _mm512_setzero_si512()); _mm512_mask_storeu_epi64(&sums[-(ptrdiff_t)num_channels], (1<<num_channels)-1, _mm512_setzero_si512());
if (sqsums) if (sqsums)
_mm512_mask_storeu_epi64(&sqsums[-num_channels], (1<<num_channels)-1, _mm512_setzero_si512()); _mm512_mask_storeu_epi64(&sqsums[-(ptrdiff_t)num_channels], (1<<num_channels)-1, _mm512_setzero_si512());
} }
...@@ -180,11 +180,11 @@ public: ...@@ -180,11 +180,11 @@ public:
// //
_mm512_mask_storeu_pd( _mm512_mask_storeu_pd(
results_ptr, // Store the result here results_ptr, // Store the result here
data_mask, // Using the data mask to avoid overrunning the line (__mmask8)data_mask, // Using the data mask to avoid overrunning the line
calculate_integral( // Writing the value of the integral derived from: calculate_integral( // Writing the value of the integral derived from:
src_longs, // input data src_longs, // input data
_mm512_maskz_loadu_pd(data_mask, above_values_ptr), // and the results from line above _mm512_maskz_loadu_pd((__mmask8)data_mask, above_values_ptr), // and the results from line above
accumulator // keeping track of the accumulator accumulator // keeping track of the accumulator
) )
); );
} }
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment