generic-library/opencv
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

HAL universal intrinsics tests and documentation

Browse Source
This commit is contained in:
Maksim Shabunin
2015-08-27 16:53:44 +03:00
parent 190d00ea3e
commit a275489f0a
10 changed files with 2226 additions and 113 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
modules/hal/include/opencv2/hal.hpp
View File

@@ -49,10 +49,21 @@
/**
@defgroup hal Hardware Acceleration Layer
@{
@defgroup hal_intrin Universal intrinsics
@{
@defgroup hal_intrin_impl Private implementation helpers
@}
@defgroup hal_utils Platform-dependent utils
@}
*/
namespace cv { namespace hal {
//! @addtogroup hal
//! @{
namespace Error {
enum
@@ -93,6 +104,8 @@ void sqrt(const double* src, double* dst, int len);
void invSqrt(const float* src, float* dst, int len);
void invSqrt(const double* src, double* dst, int len);
//! @}
}} //cv::hal
#endif //__OPENCV_HAL_HPP__

20
modules/hal/include/opencv2/hal/defs.h
View File

@@ -45,6 +45,9 @@
#ifndef __OPENCV_DEF_H__
#define __OPENCV_DEF_H__
//! @addtogroup hal_utils
//! @{
#if !defined _CRT_SECURE_NO_DEPRECATE && defined _MSC_VER && _MSC_VER > 1300
# define _CRT_SECURE_NO_DEPRECATE /* to avoid multiple Visual Studio warnings */
#endif
@@ -335,9 +338,6 @@ Cv64suf;
# include "tegra_round.hpp"
#endif
//! @addtogroup core_utils
//! @{
#if CV_VFP
// 1. general scheme
#define ARM_ROUND(_value, _asm_string) \
@@ -567,15 +567,19 @@ CV_INLINE int cvIsInf( float value )
return (ieee754.u & 0x7fffffff) == 0x7f800000;
}
//! @}
#include <algorithm>
namespace cv
{
//! @addtogroup hal_utils
//! @{
/////////////// saturate_cast (used in image & signal processing) ///////////////////
/**
Template function for accurate conversion from one primitive type to another.
/** @brief Template function for accurate conversion from one primitive type to another.
The functions saturate_cast resemble the standard C++ cast operations, such as static_cast\<T\>()
and others. They perform an efficient and accurate conversion from one primitive type to another
@@ -618,8 +622,6 @@ template<typename _Tp> static inline _Tp saturate_cast(int64 v) { return _Tp(
/** @overload */
template<typename _Tp> static inline _Tp saturate_cast(uint64 v) { return _Tp(v); }
//! @cond IGNORED
template<> inline uchar saturate_cast<uchar>(schar v) { return (uchar)std::max((int)v, 0); }
template<> inline uchar saturate_cast<uchar>(ushort v) { return (uchar)std::min((unsigned)v, (unsigned)UCHAR_MAX); }
template<> inline uchar saturate_cast<uchar>(int v) { return (uchar)((unsigned)v <= UCHAR_MAX ? v : v > 0 ? UCHAR_MAX : 0); }
@@ -664,12 +666,10 @@ template<> inline int saturate_cast<int>(double v) { return cvRound(v)
template<> inline unsigned saturate_cast<unsigned>(float v) { return cvRound(v); }
template<> inline unsigned saturate_cast<unsigned>(double v) { return cvRound(v); }
//! @endcond
//! @}
}
#endif // __cplusplus
//! @} core_utils
#endif //__OPENCV_HAL_H__

28
modules/hal/include/opencv2/hal/intrin.hpp
View File

@@ -48,6 +48,7 @@
#include <cmath>
#include <float.h>
#include <stdlib.h>
#include "opencv2/hal/defs.h"
#define OPENCV_HAL_ADD(a, b) ((a) + (b))
#define OPENCV_HAL_AND(a, b) ((a) & (b))
@@ -59,6 +60,10 @@
// access from within opencv code more accessible
namespace cv {
//! @addtogroup hal_intrin
//! @{
//! @cond IGNORED
template<typename _Tp> struct V_TypeTraits
{
typedef _Tp int_type;
@@ -82,6 +87,7 @@ template<> struct V_TypeTraits<uchar>
typedef int sum_type;
typedef ushort w_type;
typedef unsigned q_type;
enum { delta = 128, shift = 8 };
@@ -99,6 +105,7 @@ template<> struct V_TypeTraits<schar>
typedef int sum_type;
typedef short w_type;
typedef int q_type;
enum { delta = 128, shift = 8 };
@@ -265,8 +272,22 @@ template<> struct V_TypeTraits<double>
}
};
template <typename T> struct V_SIMD128Traits
{
enum { nlanes = 16 / sizeof(T) };
};
//! @endcond
//! @}
}
#ifdef CV_DOXYGEN
# undef CV_SSE2
# undef CV_NEON
#endif
#if CV_SSE2
#include "opencv2/hal/intrin_sse.hpp"
@@ -281,12 +302,19 @@ template<> struct V_TypeTraits<double>
#endif
//! @addtogroup hal_intrin
//! @{
#ifndef CV_SIMD128
//! Set to 1 if current compiler supports vector extensions (NEON or SSE is enabled)
#define CV_SIMD128 0
#endif
#ifndef CV_SIMD128_64F
//! Set to 1 if current intrinsics implementation supports 64-bit float vectors
#define CV_SIMD128_64F 0
#endif
//! @}
#endif

1069
modules/hal/include/opencv2/hal/intrin_cpp.hpp
View File

File diff suppressed because it is too large Load Diff

31
modules/hal/include/opencv2/hal/intrin_neon.hpp
View File

@@ -48,6 +48,8 @@
namespace cv
{
//! @cond IGNORED
#define CV_SIMD128 1
struct v_uint8x16
@@ -278,14 +280,15 @@ void v_rshr_##pack##_store(_Tp* ptr, const _Tpwvec& a) \
}
OPENCV_HAL_IMPL_NEON_PACK(v_uint8x16, uchar, uint8x8_t, u8, v_uint16x8, u16, pack, n)
OPENCV_HAL_IMPL_NEON_PACK(v_uint8x16, uchar, uint8x8_t, u8, v_int16x8, s16, pack_u, un)
OPENCV_HAL_IMPL_NEON_PACK(v_int8x16, schar, int8x8_t, s8, v_int16x8, s16, pack, n)
OPENCV_HAL_IMPL_NEON_PACK(v_uint16x8, ushort, uint16x4_t, u16, v_uint32x4, u32, pack, n)
OPENCV_HAL_IMPL_NEON_PACK(v_uint16x8, ushort, uint16x4_t, u16, v_int32x4, s32, pack_u, un)
OPENCV_HAL_IMPL_NEON_PACK(v_int16x8, short, int16x4_t, s16, v_int32x4, s32, pack, n)
OPENCV_HAL_IMPL_NEON_PACK(v_uint32x4, unsigned, uint32x2_t, u32, v_uint64x2, u64, pack, n)
OPENCV_HAL_IMPL_NEON_PACK(v_int32x4, int, int32x2_t, s32, v_int64x2, s64, pack, n)
OPENCV_HAL_IMPL_NEON_PACK(v_uint8x16, uchar, uint8x8_t, u8, v_int16x8, s16, pack_u, un)
OPENCV_HAL_IMPL_NEON_PACK(v_uint16x8, ushort, uint16x4_t, u16, v_int32x4, s32, pack_u, un)
inline v_float32x4 v_matmul(const v_float32x4& v, const v_float32x4& m0,
const v_float32x4& m1, const v_float32x4& m2,
const v_float32x4& m3)
@@ -374,7 +377,7 @@ inline v_int32x4 v_dotprod(const v_int16x8& a, const v_int16x8& b)
{
int32x4_t c = vmull_s16(vget_low_s16(a.val), vget_low_s16(b.val));
int32x4_t d = vmull_s16(vget_high_s16(a.val), vget_high_s16(b.val));
int32x4x2_t cd = vtrnq_s32(c, d);
int32x4x2_t cd = vuzpq_s32(c, d);
return v_int32x4(vaddq_s32(cd.val[0], cd.val[1]));
}
@@ -497,6 +500,16 @@ OPENCV_HAL_IMPL_NEON_BIN_FUNC(v_uint16x8, v_absdiff, vabdq_u16)
OPENCV_HAL_IMPL_NEON_BIN_FUNC(v_uint32x4, v_absdiff, vabdq_u32)
OPENCV_HAL_IMPL_NEON_BIN_FUNC(v_float32x4, v_absdiff, vabdq_f32)
#define OPENCV_HAL_IMPL_NEON_BIN_FUNC2(_Tpvec, _Tpvec2, cast, func, intrin) \
inline _Tpvec2 func(const _Tpvec& a, const _Tpvec& b) \
{ \
return _Tpvec2(cast(intrin(a.val, b.val))); \
}
OPENCV_HAL_IMPL_NEON_BIN_FUNC2(v_int8x16, v_uint8x16, vreinterpretq_u8_s8, v_absdiff, vabdq_s8)
OPENCV_HAL_IMPL_NEON_BIN_FUNC2(v_int16x8, v_uint16x8, vreinterpretq_u16_s16, v_absdiff, vabdq_s16)
OPENCV_HAL_IMPL_NEON_BIN_FUNC2(v_int32x4, v_uint32x4, vreinterpretq_u32_s32, v_absdiff, vabdq_s32)
inline v_float32x4 v_magnitude(const v_float32x4& a, const v_float32x4& b)
{
v_float32x4 x(vmlaq_f32(vmulq_f32(a.val, a.val), b.val, b.val));
@@ -641,13 +654,13 @@ inline bool v_check_all(const v_float32x4& a)
{ return v_check_all(v_reinterpret_as_u32(a)); }
inline bool v_check_any(const v_int8x16& a)
{ return v_check_all(v_reinterpret_as_u8(a)); }
{ return v_check_any(v_reinterpret_as_u8(a)); }
inline bool v_check_any(const v_int16x8& a)
{ return v_check_all(v_reinterpret_as_u16(a)); }
{ return v_check_any(v_reinterpret_as_u16(a)); }
inline bool v_check_any(const v_int32x4& a)
{ return v_check_all(v_reinterpret_as_u32(a)); }
{ return v_check_any(v_reinterpret_as_u32(a)); }
inline bool v_check_any(const v_float32x4& a)
{ return v_check_all(v_reinterpret_as_u32(a)); }
{ return v_check_any(v_reinterpret_as_u32(a)); }
#define OPENCV_HAL_IMPL_NEON_SELECT(_Tpvec, suffix, usuffix) \
inline _Tpvec v_select(const _Tpvec& mask, const _Tpvec& a, const _Tpvec& b) \
@@ -678,6 +691,8 @@ OPENCV_HAL_IMPL_NEON_EXPAND(v_uint8x16, v_uint16x8, uchar, u8)
OPENCV_HAL_IMPL_NEON_EXPAND(v_int8x16, v_int16x8, schar, s8)
OPENCV_HAL_IMPL_NEON_EXPAND(v_uint16x8, v_uint32x4, ushort, u16)
OPENCV_HAL_IMPL_NEON_EXPAND(v_int16x8, v_int32x4, short, s16)
OPENCV_HAL_IMPL_NEON_EXPAND(v_uint32x4, v_uint64x2, uint, u32)
OPENCV_HAL_IMPL_NEON_EXPAND(v_int32x4, v_int64x2, int, s32)
inline v_uint32x4 v_load_expand_q(const uchar* ptr)
{
@@ -840,6 +855,8 @@ inline v_float32x4 v_cvt_f32(const v_int32x4& a)
return v_float32x4(vcvtq_f32_s32(a.val));
}
//! @endcond
}
#endif

66
modules/hal/include/opencv2/hal/intrin_sse.hpp
View File

@@ -51,6 +51,8 @@
namespace cv
{
//! @cond IGNORED
struct v_uint8x16
{
typedef uchar lane_type;
@@ -296,6 +298,11 @@ OPENCV_HAL_IMPL_SSE_INIT_FROM_FLT(v_int32x4, s32)
OPENCV_HAL_IMPL_SSE_INIT_FROM_FLT(v_uint64x2, u64)
OPENCV_HAL_IMPL_SSE_INIT_FROM_FLT(v_int64x2, s64)
inline v_float32x4 v_reinterpret_as_f32(const v_float32x4& a) {return a; }
inline v_float64x2 v_reinterpret_as_f64(const v_float64x2& a) {return a; }
inline v_float32x4 v_reinterpret_as_f32(const v_float64x2& a) {return v_float32x4(_mm_castpd_ps(a.val)); }
inline v_float64x2 v_reinterpret_as_f64(const v_float32x4& a) {return v_float64x2(_mm_castps_pd(a.val)); }
//////////////// PACK ///////////////
inline v_uint8x16 v_pack(const v_uint16x8& a, const v_uint16x8& b)
{
@@ -430,6 +437,17 @@ inline void v_pack_u_store(ushort* ptr, const v_int32x4& a)
_mm_storel_epi64((__m128i*)ptr, r);
}
template<int n> inline
v_uint16x8 v_rshr_pack_u(const v_int32x4& a, const v_int32x4& b)
{
__m128i delta = _mm_set1_epi32(1 << (n-1)), delta32 = _mm_set1_epi32(32768);
__m128i a1 = _mm_sub_epi32(_mm_srai_epi32(_mm_add_epi32(a.val, delta), n), delta32);
__m128i a2 = _mm_sub_epi16(_mm_packs_epi32(a1, a1), _mm_set1_epi16(-32768));
__m128i b1 = _mm_sub_epi32(_mm_srai_epi32(_mm_add_epi32(b.val, delta), n), delta32);
__m128i b2 = _mm_sub_epi16(_mm_packs_epi32(b1, b1), _mm_set1_epi16(-32768));
return v_uint16x8(_mm_unpacklo_epi64(a2, b2));
}
template<int n> inline
void v_rshr_pack_u_store(ushort* ptr, const v_int32x4& a)
{
@@ -460,7 +478,7 @@ void v_rshr_pack_store(short* ptr, const v_int32x4& a)
{
__m128i delta = _mm_set1_epi32(1 << (n-1));
__m128i a1 = _mm_srai_epi32(_mm_add_epi32(a.val, delta), n);
_mm_storel_epi64((__m128i*)ptr, a1);
_mm_storel_epi64((__m128i*)ptr, _mm_packs_epi32(a1, a1));
}
@@ -469,7 +487,7 @@ inline v_uint32x4 v_pack(const v_uint64x2& a, const v_uint64x2& b)
{
__m128i v0 = _mm_unpacklo_epi32(a.val, b.val); // a0 a1 0 0
__m128i v1 = _mm_unpackhi_epi32(a.val, b.val); // b0 b1 0 0
return v_uint32x4(_mm_unpacklo_epi64(v0, v1));
return v_uint32x4(_mm_unpacklo_epi32(v0, v1));
}
inline void v_pack_store(unsigned* ptr, const v_uint64x2& a)
@@ -483,7 +501,7 @@ inline v_int32x4 v_pack(const v_int64x2& a, const v_int64x2& b)
{
__m128i v0 = _mm_unpacklo_epi32(a.val, b.val); // a0 a1 0 0
__m128i v1 = _mm_unpackhi_epi32(a.val, b.val); // b0 b1 0 0
return v_int32x4(_mm_unpacklo_epi64(v0, v1));
return v_int32x4(_mm_unpacklo_epi32(v0, v1));
}
inline void v_pack_store(int* ptr, const v_int64x2& a)
@@ -501,7 +519,7 @@ v_uint32x4 v_rshr_pack(const v_uint64x2& a, const v_uint64x2& b)
__m128i b1 = _mm_srli_epi64(_mm_add_epi64(b.val, delta2.val), n);
__m128i v0 = _mm_unpacklo_epi32(a1, b1); // a0 a1 0 0
__m128i v1 = _mm_unpackhi_epi32(a1, b1); // b0 b1 0 0
return v_uint32x4(_mm_unpacklo_epi64(v0, v1));
return v_uint32x4(_mm_unpacklo_epi32(v0, v1));
}
template<int n> inline
@@ -534,7 +552,7 @@ v_int32x4 v_rshr_pack(const v_int64x2& a, const v_int64x2& b)
__m128i b1 = v_srai_epi64(_mm_add_epi64(b.val, delta2.val), n);
__m128i v0 = _mm_unpacklo_epi32(a1, b1); // a0 a1 0 0
__m128i v1 = _mm_unpackhi_epi32(a1, b1); // b0 b1 0 0
return v_int32x4(_mm_unpacklo_epi64(v0, v1));
return v_int32x4(_mm_unpacklo_epi32(v0, v1));
}
template<int n> inline
@@ -630,8 +648,8 @@ inline void v_mul_expand(const v_int16x8& a, const v_int16x8& b,
{
__m128i v0 = _mm_mullo_epi16(a.val, b.val);
__m128i v1 = _mm_mulhi_epi16(a.val, b.val);
c.val = _mm_unpacklo_epi32(v0, v1);
d.val = _mm_unpackhi_epi32(v0, v1);
c.val = _mm_unpacklo_epi16(v0, v1);
d.val = _mm_unpackhi_epi16(v0, v1);
}
inline void v_mul_expand(const v_uint16x8& a, const v_uint16x8& b,
@@ -639,8 +657,8 @@ inline void v_mul_expand(const v_uint16x8& a, const v_uint16x8& b,
{
__m128i v0 = _mm_mullo_epi16(a.val, b.val);
__m128i v1 = _mm_mulhi_epu16(a.val, b.val);
c.val = _mm_unpacklo_epi32(v0, v1);
d.val = _mm_unpackhi_epi32(v0, v1);
c.val = _mm_unpacklo_epi16(v0, v1);
d.val = _mm_unpackhi_epi16(v0, v1);
}
inline void v_mul_expand(const v_uint32x4& a, const v_uint32x4& b,
@@ -869,6 +887,18 @@ inline _Tpuvec v_absdiff(const _Tpsvec& a, const _Tpsvec& b) \
OPENCV_HAL_IMPL_SSE_ABSDIFF_8_16(v_uint8x16, v_int8x16, 8, (int)0x80808080)
OPENCV_HAL_IMPL_SSE_ABSDIFF_8_16(v_uint16x8, v_int16x8, 16, (int)0x80008000)
inline v_uint32x4 v_absdiff(const v_uint32x4& a, const v_uint32x4& b)
{
return v_max(a, b) - v_min(a, b);
}
inline v_uint32x4 v_absdiff(const v_int32x4& a, const v_int32x4& b)
{
__m128i d = _mm_sub_epi32(a.val, b.val);
__m128i m = _mm_cmpgt_epi32(b.val, a.val);
return v_uint32x4(_mm_sub_epi32(_mm_xor_si128(d, m), m));
}
#define OPENCV_HAL_IMPL_SSE_MISC_FLT_OP(_Tpvec, _Tp, _Tpreg, suffix, absmask_vec) \
inline _Tpvec v_absdiff(const _Tpvec& a, const _Tpvec& b) \
{ \
@@ -1047,8 +1077,8 @@ OPENCV_HAL_IMPL_SSE_SELECT(v_uint16x8, si128)
OPENCV_HAL_IMPL_SSE_SELECT(v_int16x8, si128)
OPENCV_HAL_IMPL_SSE_SELECT(v_uint32x4, si128)
OPENCV_HAL_IMPL_SSE_SELECT(v_int32x4, si128)
OPENCV_HAL_IMPL_SSE_SELECT(v_uint64x2, si128)
OPENCV_HAL_IMPL_SSE_SELECT(v_int64x2, si128)
// OPENCV_HAL_IMPL_SSE_SELECT(v_uint64x2, si128)
// OPENCV_HAL_IMPL_SSE_SELECT(v_int64x2, si128)
OPENCV_HAL_IMPL_SSE_SELECT(v_float32x4, ps)
OPENCV_HAL_IMPL_SSE_SELECT(v_float64x2, pd)
@@ -1257,7 +1287,7 @@ inline void v_load_deinterleave(const uchar* ptr, v_uint8x16& a, v_uint8x16& b,
__m128i v0 = _mm_unpacklo_epi8(u0, u2); // a0 a8 b0 b8 ...
__m128i v1 = _mm_unpackhi_epi8(u0, u2); // a2 a10 b2 b10 ...
__m128i v2 = _mm_unpacklo_epi8(u1, u3); // a4 a12 b4 b12 ...
__m128i v3 = _mm_unpackhi_epi8(u1, u3); // a6 a14 b4 b14 ...
__m128i v3 = _mm_unpackhi_epi8(u1, u3); // a6 a14 b6 b14 ...
u0 = _mm_unpacklo_epi8(v0, v2); // a0 a4 a8 a12 ...
u1 = _mm_unpacklo_epi8(v1, v3); // a2 a6 a10 a14 ...
@@ -1266,13 +1296,13 @@ inline void v_load_deinterleave(const uchar* ptr, v_uint8x16& a, v_uint8x16& b,
v0 = _mm_unpacklo_epi8(u0, u1); // a0 a2 a4 a6 ...
v1 = _mm_unpacklo_epi8(u2, u3); // a1 a3 a5 a7 ...
v2 = _mm_unpackhi_epi8(u0, u1); // b0 b2 b4 b6 ...
v3 = _mm_unpackhi_epi8(u2, u3); // b1 b3 b5 b7 ...
v2 = _mm_unpackhi_epi8(u0, u1); // c0 c2 c4 c6 ...
v3 = _mm_unpackhi_epi8(u2, u3); // c1 c3 c5 c7 ...
a.val = _mm_unpacklo_epi8(v0, v1);
b.val = _mm_unpacklo_epi8(v2, v3);
c.val = _mm_unpackhi_epi8(v0, v1);
d.val = _mm_unpacklo_epi8(v2, v3);
b.val = _mm_unpackhi_epi8(v0, v1);
c.val = _mm_unpacklo_epi8(v2, v3);
d.val = _mm_unpackhi_epi8(v2, v3);
}
inline void v_load_deinterleave(const ushort* ptr, v_uint16x8& a, v_uint16x8& b, v_uint16x8& c)
@@ -1560,6 +1590,8 @@ inline v_float64x2 v_cvt_f64(const v_float32x4& a)
return v_float64x2(_mm_cvtps_pd(a.val));
}
//! @endcond
}
#endif