potok/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2026-07-30 15:53:03 +04:00
Code Packages Releases Wiki Activity

Merge pull request #14210 from terfendail:wui_512

Browse Source 
AVX512 wide universal intrinsics (#14210)

* Added implementation of 512-bit wide universal intrinsics(WIP)

* Added implementation of 512-bit wide universal intrinsics: implemented WUI vector types(WIP)

* Added implementation of 512-bit wide universal intrinsics(WIP): implemented load/store

* Added implementation of 512-bit wide universal intrinsics(WIP): implemented fp16 load/store

* Added implementation of 512-bit wide universal intrinsics(WIP): implemented recombine and zip, implemented non-saturating and saturating arithmetics

* Added implementation of 512-bit wide universal intrinsics(WIP): implemented bit operations

* Added implementation of 512-bit wide universal intrinsics(WIP): implemented comparisons

* Added implementation of 512-bit wide universal intrinsics(WIP): implemented lane shifts and reduction

* Added implementation of 512-bit wide universal intrinsics(WIP): implemented absolute values

* Added implementation of 512-bit wide universal intrinsics(WIP): implemented rounding and cast to float

* Added implementation of 512-bit wide universal intrinsics(WIP): implemented LUT

* Added implementation of 512-bit wide universal intrinsics(WIP): implemented type extension/narrowing and matrix operations

* Added implementation of 512-bit wide universal intrinsics(WIP): implemented load_deinterleave for 2 and 3 channels images

* Added implementation of 512-bit wide universal intrinsics(WIP): reimplemented load_deinterleave for 2- and implemented for 4-channel images

* Added implementation of 512-bit wide universal intrinsics(WIP): implemented store_interleave

* Added implementation of 512-bit wide universal intrinsics(WIP): implemented signmask and checks

* Added implementation of 512-bit wide universal intrinsics(WIP): build fixes

* Added implementation of 512-bit wide universal intrinsics(WIP): reimplemented popcount in case AVX512_BITALG is unavailable

* Added implementation of 512-bit wide universal intrinsics(WIP): reimplemented zip

* Added implementation of 512-bit wide universal intrinsics(WIP): reimplemented rotate for s8 and s16

* Added implementation of 512-bit wide universal intrinsics(WIP): reimplemented interleave/deinterleave for s8 and s16

* Added implementation of 512-bit wide universal intrinsics(WIP): updated v512_set macros

* Added implementation of 512-bit wide universal intrinsics(WIP): fix for GCC wrong _mm512_abs_pd definition

* Added implementation of 512-bit wide universal intrinsics(WIP): reworked v_zip to avoid AVX512_VBMI intrinsics

* Added implementation of 512-bit wide universal intrinsics(WIP): reworked v_invsqrt to avoid AVX512_ER intrinsics

* Added implementation of 512-bit wide universal intrinsics(WIP): reworked v_rotate, v_popcount and interleave/deinterleave for U8 to avoid AVX512_VBMI intrinsics

* Added implementation of 512-bit wide universal intrinsics(WIP): fixed integral image SIMD part

* Added implementation of 512-bit wide universal intrinsics(WIP): fixed warnings

* Added implementation of 512-bit wide universal intrinsics(WIP): fixed load_deinterleave for u8 and u16

* Added implementation of 512-bit wide universal intrinsics(WIP): fixed v_invsqrt accuracy for f64

* Added implementation of 512-bit wide universal intrinsics(WIP): fixed interleave/deinterleave for u32 and u64

* Added implementation of 512-bit wide universal intrinsics(WIP): fixed interleave_pairs, interleave_quads and pack_triplets

* Added implementation of 512-bit wide universal intrinsics(WIP): fixed rotate_left

* Added implementation of 512-bit wide universal intrinsics(WIP): fixed rotate_left/right, part 2

* Added implementation of 512-bit wide universal intrinsics(WIP): fixed 512-wide universal intrinsics based resize

* Added implementation of 512-bit wide universal intrinsics(WIP): fixed findContours by avoiding use of uint64 dependent 512-wide v_signmask()

* Added implementation of 512-bit wide universal intrinsics(WIP): fixed trailing whitespaces

* Added implementation of 512-bit wide universal intrinsics(WIP): reworked specific intrinsic sets dependent parts to check availability of intrinsics based on CPU feature group defines

* Added implementation of 512-bit wide universal intrinsics(WIP):Updated AVX512 implementation of v_popcount to avoid AVX512VPOPCNTDQ intrinsics if unavailable.

* Added implementation of 512-bit wide universal intrinsics(WIP): Fixed universal intrinsics data initialisation, v_mul_wrap, v_floor, v_ceil and v_signmask.

* Added implementation of 512-bit wide universal intrinsics(WIP): Removed hasSIMD512()

* Added implementation of 512-bit wide universal intrinsics(WIP): Fixes for gcc build

* Added implementation of 512-bit wide universal intrinsics(WIP): Reworked v_signmask, v_check_any() and v_check_all() implementation.
This commit is contained in:
Vitaly Tuzov
2019-06-03 18:05:35 +03:00
committed by Alexander Alekhin
parent 3289a0aff9
commit 3b015dfc7d
10 changed files with 2902 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/core/include/opencv2/core/hal/intrin.hpp
+41 -1
View File
@@ -180,6 +180,18 @@ using namespace CV_CPU_OPTIMIZATION_HAL_NAMESPACE;
#endif
// AVX512 can be used together with SSE2 and AVX2, so
// we define those sets of intrinsics at once.
// For some of AVX512 intrinsics get v512_ prefix instead of v_, e.g. v512_load() vs v_load().
// Wide intrinsics will be mapped to v512_ counterparts in this case(e.g. vx_load() => v512_load())
#if CV_AVX512_SKX
#define CV__SIMD_FORWARD 512
#include "opencv2/core/hal/intrin_forward.hpp"
#include "opencv2/core/hal/intrin_avx512.hpp"
#endif
//! @cond IGNORED
namespace cv {
@@ -321,13 +333,41 @@ template<typename _Tp> struct V_RegTraits
CV_DEF_REG_TRAITS(v256, v_float64x4, double, f64, v_float64x4, void, void, v_int64x4, v_int32x8);
#endif
#if CV_SIMD512
CV_DEF_REG_TRAITS(v512, v_uint8x64, uchar, u8, v_uint8x64, v_uint16x32, v_uint32x16, v_int8x64, void);
CV_DEF_REG_TRAITS(v512, v_int8x64, schar, s8, v_uint8x64, v_int16x32, v_int32x16, v_int8x64, void);
CV_DEF_REG_TRAITS(v512, v_uint16x32, ushort, u16, v_uint16x32, v_uint32x16, v_uint64x8, v_int16x32, void);
CV_DEF_REG_TRAITS(v512, v_int16x32, short, s16, v_uint16x32, v_int32x16, v_int64x8, v_int16x32, void);
CV_DEF_REG_TRAITS(v512, v_uint32x16, unsigned, u32, v_uint32x16, v_uint64x8, void, v_int32x16, void);
CV_DEF_REG_TRAITS(v512, v_int32x16, int, s32, v_uint32x16, v_int64x8, void, v_int32x16, void);
CV_DEF_REG_TRAITS(v512, v_float32x16, float, f32, v_float32x16, v_float64x8, void, v_int32x16, v_int32x16);
CV_DEF_REG_TRAITS(v512, v_uint64x8, uint64, u64, v_uint64x8, void, void, v_int64x8, void);
CV_DEF_REG_TRAITS(v512, v_int64x8, int64, s64, v_uint64x8, void, void, v_int64x8, void);
CV_DEF_REG_TRAITS(v512, v_float64x8, double, f64, v_float64x8, void, void, v_int64x8, v_int32x16);
#endif
#if CV_SIMD512 && (!defined(CV__SIMD_FORCE_WIDTH) || CV__SIMD_FORCE_WIDTH == 512)
#define CV__SIMD_NAMESPACE simd512
namespace CV__SIMD_NAMESPACE {
#define CV_SIMD 1
#define CV_SIMD_64F CV_SIMD512_64F
#define CV_SIMD_FP16 CV_SIMD512_FP16
#define CV_SIMD_WIDTH 64
// TODO typedef v_uint8 / v_int32 / etc types here
typedef v_uint8x64 v_uint8;
typedef v_int8x64 v_int8;
typedef v_uint16x32 v_uint16;
typedef v_int16x32 v_int16;
typedef v_uint32x16 v_uint32;
typedef v_int32x16 v_int32;
typedef v_uint64x8 v_uint64;
typedef v_int64x8 v_int64;
typedef v_float32x16 v_float32;
CV_INTRIN_DEFINE_WIDE_INTRIN_ALL_TYPES(v512)
#if CV_SIMD512_64F
typedef v_float64x8 v_float64;
CV_INTRIN_DEFINE_WIDE_INTRIN(double, v_float64, f64, v512, load)
#endif
inline void vx_cleanup() { v512_cleanup(); }
} // namespace
using namespace CV__SIMD_NAMESPACE;
#elif CV_SIMD256 && (!defined(CV__SIMD_FORCE_WIDTH) || CV__SIMD_FORCE_WIDTH == 256)
modules/core/include/opencv2/core/hal/intrin_avx512.hpp
+2743
View File
File diff suppressed because it is too large Load Diff
modules/core/include/opencv2/core/hal/intrin_forward.hpp
+26 -3
View File
@@ -14,9 +14,32 @@ namespace cv
CV_CPU_OPTIMIZATION_HAL_NAMESPACE_BEGIN
/** Types **/
#if CV__SIMD_FORWARD == 512
// [todo] 512
#error "AVX512 Not implemented yet"
#if CV__SIMD_FORWARD == 1024
// [todo] 1024
#error "1024-long ops not implemented yet"
#elif CV__SIMD_FORWARD == 512
// 512
#define __CV_VX(fun) v512_##fun
#define __CV_V_UINT8 v_uint8x64
#define __CV_V_INT8 v_int8x64
#define __CV_V_UINT16 v_uint16x32
#define __CV_V_INT16 v_int16x32
#define __CV_V_UINT32 v_uint32x16
#define __CV_V_INT32 v_int32x16
#define __CV_V_UINT64 v_uint64x8
#define __CV_V_INT64 v_int64x8
#define __CV_V_FLOAT32 v_float32x16
#define __CV_V_FLOAT64 v_float64x8
struct v_uint8x64;
struct v_int8x64;
struct v_uint16x32;
struct v_int16x32;
struct v_uint32x16;
struct v_int32x16;
struct v_uint64x8;
struct v_int64x8;
struct v_float32x16;
struct v_float64x8;
#elif CV__SIMD_FORWARD == 256
// 256
#define __CV_VX(fun) v256_##fun