potok/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2026-07-29 15:23:05 +04:00
Code Packages Releases Wiki Activity

core: dispatch convert_scale

Browse Source
This commit is contained in:
Alexander Alekhin
2019-02-08 17:56:25 +03:00
parent 39b90ae9fb
commit d32d576d6d
3 changed files with 26 additions and 579 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/core/src/convert_scale.dispatch.cpp
+14 -341
View File
@@ -5,348 +5,26 @@
#include "precomp.hpp"
#include "opencl_kernels_core.hpp"
#include "convert.hpp"
/****************************************************************************************\
* convertScale[Abs] *
\****************************************************************************************/
#include "convert_scale.simd.hpp"
#include "convert_scale.simd_declarations.hpp" // defines CV_CPU_DISPATCH_MODES_ALL=AVX2,...,BASELINE based on CMakeLists.txt content
namespace cv
{
template<typename _Ts, typename _Td> inline void
cvtabs_32f( const _Ts* src, size_t sstep, _Td* dst, size_t dstep,
Size size, float a, float b )
{
#if CV_SIMD
v_float32 va = vx_setall_f32(a), vb = vx_setall_f32(b);
const int VECSZ = v_float32::nlanes*2;
#endif
sstep /= sizeof(src[0]);
dstep /= sizeof(dst[0]);
for( int i = 0; i < size.height; i++, src += sstep, dst += dstep )
{
int j = 0;
#if CV_SIMD
for( ; j < size.width; j += VECSZ )
{
if( j > size.width - VECSZ )
{
if( j == 0 || src == (_Ts*)dst )
break;
j = size.width - VECSZ;
}
v_float32 v0, v1;
vx_load_pair_as(src + j, v0, v1);
v0 = v_fma(v0, va, vb);
v1 = v_fma(v1, va, vb);
v_store_pair_as(dst + j, v_abs(v0), v_abs(v1));
}
#endif
for( ; j < size.width; j++ )
dst[j] = saturate_cast<_Td>(std::abs(src[j]*a + b));
}
}
// variant for convrsions 16f <-> ... w/o unrolling
template<typename _Ts, typename _Td> inline void
cvtabs1_32f( const _Ts* src, size_t sstep, _Td* dst, size_t dstep,
Size size, float a, float b )
{
#if CV_SIMD
v_float32 va = vx_setall_f32(a), vb = vx_setall_f32(b);
const int VECSZ = v_float32::nlanes*2;
#endif
sstep /= sizeof(src[0]);
dstep /= sizeof(dst[0]);
for( int i = 0; i < size.height; i++, src += sstep, dst += dstep )
{
int j = 0;
#if CV_SIMD
for( ; j < size.width; j += VECSZ )
{
if( j > size.width - VECSZ )
{
if( j == 0 || src == (_Ts*)dst )
break;
j = size.width - VECSZ;
}
v_float32 v0;
vx_load_as(src + j, v0);
v0 = v_fma(v0, va, vb);
v_store_as(dst + j, v_abs(v0));
}
#endif
for( ; j < size.width; j++ )
dst[j] = saturate_cast<_Td>(src[j]*a + b);
}
}
template<typename _Ts, typename _Td> inline void
cvt_32f( const _Ts* src, size_t sstep, _Td* dst, size_t dstep,
Size size, float a, float b )
{
#if CV_SIMD
v_float32 va = vx_setall_f32(a), vb = vx_setall_f32(b);
const int VECSZ = v_float32::nlanes*2;
#endif
sstep /= sizeof(src[0]);
dstep /= sizeof(dst[0]);
for( int i = 0; i < size.height; i++, src += sstep, dst += dstep )
{
int j = 0;
#if CV_SIMD
for( ; j < size.width; j += VECSZ )
{
if( j > size.width - VECSZ )
{
if( j == 0 || src == (_Ts*)dst )
break;
j = size.width - VECSZ;
}
v_float32 v0, v1;
vx_load_pair_as(src + j, v0, v1);
v0 = v_fma(v0, va, vb);
v1 = v_fma(v1, va, vb);
v_store_pair_as(dst + j, v0, v1);
}
#endif
for( ; j < size.width; j++ )
dst[j] = saturate_cast<_Td>(src[j]*a + b);
}
}
// variant for convrsions 16f <-> ... w/o unrolling
template<typename _Ts, typename _Td> inline void
cvt1_32f( const _Ts* src, size_t sstep, _Td* dst, size_t dstep,
Size size, float a, float b )
{
#if CV_SIMD
v_float32 va = vx_setall_f32(a), vb = vx_setall_f32(b);
const int VECSZ = v_float32::nlanes;
#endif
sstep /= sizeof(src[0]);
dstep /= sizeof(dst[0]);
for( int i = 0; i < size.height; i++, src += sstep, dst += dstep )
{
int j = 0;
#if CV_SIMD
for( ; j < size.width; j += VECSZ )
{
if( j > size.width - VECSZ )
{
if( j == 0 || src == (_Ts*)dst )
break;
j = size.width - VECSZ;
}
v_float32 v0;
vx_load_as(src + j, v0);
v0 = v_fma(v0, va, vb);
v_store_as(dst + j, v0);
}
#endif
for( ; j < size.width; j++ )
dst[j] = saturate_cast<_Td>(src[j]*a + b);
}
}
template<typename _Ts, typename _Td> inline void
cvt_64f( const _Ts* src, size_t sstep, _Td* dst, size_t dstep,
Size size, double a, double b )
{
#if CV_SIMD_64F
v_float64 va = vx_setall_f64(a), vb = vx_setall_f64(b);
const int VECSZ = v_float64::nlanes*2;
#endif
sstep /= sizeof(src[0]);
dstep /= sizeof(dst[0]);
for( int i = 0; i < size.height; i++, src += sstep, dst += dstep )
{
int j = 0;
#if CV_SIMD_64F
for( ; j < size.width; j += VECSZ )
{
if( j > size.width - VECSZ )
{
if( j == 0 || src == (_Ts*)dst )
break;
j = size.width - VECSZ;
}
v_float64 v0, v1;
vx_load_pair_as(src + j, v0, v1);
v0 = v_fma(v0, va, vb);
v1 = v_fma(v1, va, vb);
v_store_pair_as(dst + j, v0, v1);
}
#endif
for( ; j < size.width; j++ )
dst[j] = saturate_cast<_Td>(src[j]*a + b);
}
}
//==================================================================================================
#define DEF_CVT_SCALE_ABS_FUNC(suffix, cvt, stype, dtype, wtype) \
static void cvtScaleAbs##suffix( const stype* src, size_t sstep, const uchar*, size_t, \
dtype* dst, size_t dstep, Size size, double* scale) \
{ \
cvt(src, sstep, dst, dstep, size, (wtype)scale[0], (wtype)scale[1]); \
}
#define DEF_CVT_SCALE_FUNC(suffix, cvt, stype, dtype, wtype) \
static void cvtScale##suffix( const stype* src, size_t sstep, const uchar*, size_t, \
dtype* dst, size_t dstep, Size size, double* scale) \
{ \
cvt(src, sstep, dst, dstep, size, (wtype)scale[0], (wtype)scale[1]); \
}
DEF_CVT_SCALE_ABS_FUNC(8u, cvtabs_32f, uchar, uchar, float)
DEF_CVT_SCALE_ABS_FUNC(8s8u, cvtabs_32f, schar, uchar, float)
DEF_CVT_SCALE_ABS_FUNC(16u8u, cvtabs_32f, ushort, uchar, float)
DEF_CVT_SCALE_ABS_FUNC(16s8u, cvtabs_32f, short, uchar, float)
DEF_CVT_SCALE_ABS_FUNC(32s8u, cvtabs_32f, int, uchar, float)
DEF_CVT_SCALE_ABS_FUNC(32f8u, cvtabs_32f, float, uchar, float)
DEF_CVT_SCALE_ABS_FUNC(64f8u, cvtabs_32f, double, uchar, float)
DEF_CVT_SCALE_FUNC(8u, cvt_32f, uchar, uchar, float)
DEF_CVT_SCALE_FUNC(8s8u, cvt_32f, schar, uchar, float)
DEF_CVT_SCALE_FUNC(16u8u, cvt_32f, ushort, uchar, float)
DEF_CVT_SCALE_FUNC(16s8u, cvt_32f, short, uchar, float)
DEF_CVT_SCALE_FUNC(32s8u, cvt_32f, int, uchar, float)
DEF_CVT_SCALE_FUNC(32f8u, cvt_32f, float, uchar, float)
DEF_CVT_SCALE_FUNC(64f8u, cvt_32f, double, uchar, float)
//DEF_CVT_SCALE_FUNC(16f8u, cvt_32f, float16_t, uchar, float)
DEF_CVT_SCALE_FUNC(8u8s, cvt_32f, uchar, schar, float)
DEF_CVT_SCALE_FUNC(8s, cvt_32f, schar, schar, float)
DEF_CVT_SCALE_FUNC(16u8s, cvt_32f, ushort, schar, float)
DEF_CVT_SCALE_FUNC(16s8s, cvt_32f, short, schar, float)
DEF_CVT_SCALE_FUNC(32s8s, cvt_32f, int, schar, float)
DEF_CVT_SCALE_FUNC(32f8s, cvt_32f, float, schar, float)
DEF_CVT_SCALE_FUNC(64f8s, cvt_32f, double, schar, float)
//DEF_CVT_SCALE_FUNC(16f8s, cvt_32f, float16_t, schar, float)
DEF_CVT_SCALE_FUNC(8u16u, cvt_32f, uchar, ushort, float)
DEF_CVT_SCALE_FUNC(8s16u, cvt_32f, schar, ushort, float)
DEF_CVT_SCALE_FUNC(16u, cvt_32f, ushort, ushort, float)
DEF_CVT_SCALE_FUNC(16s16u, cvt_32f, short, ushort, float)
DEF_CVT_SCALE_FUNC(32s16u, cvt_32f, int, ushort, float)
DEF_CVT_SCALE_FUNC(32f16u, cvt_32f, float, ushort, float)
DEF_CVT_SCALE_FUNC(64f16u, cvt_32f, double, ushort, float)
//DEF_CVT_SCALE_FUNC(16f16u, cvt1_32f, float16_t, ushort, float)
DEF_CVT_SCALE_FUNC(8u16s, cvt_32f, uchar, short, float)
DEF_CVT_SCALE_FUNC(8s16s, cvt_32f, schar, short, float)
DEF_CVT_SCALE_FUNC(16u16s, cvt_32f, ushort, short, float)
DEF_CVT_SCALE_FUNC(16s, cvt_32f, short, short, float)
DEF_CVT_SCALE_FUNC(32s16s, cvt_32f, int, short, float)
DEF_CVT_SCALE_FUNC(32f16s, cvt_32f, float, short, float)
DEF_CVT_SCALE_FUNC(64f16s, cvt_32f, double, short, float)
//DEF_CVT_SCALE_FUNC(16f16s, cvt1_32f, float16_t, short, float)
DEF_CVT_SCALE_FUNC(8u32s, cvt_32f, uchar, int, float)
DEF_CVT_SCALE_FUNC(8s32s, cvt_32f, schar, int, float)
DEF_CVT_SCALE_FUNC(16u32s, cvt_32f, ushort, int, float)
DEF_CVT_SCALE_FUNC(16s32s, cvt_32f, short, int, float)
DEF_CVT_SCALE_FUNC(32s, cvt_64f, int, int, double)
DEF_CVT_SCALE_FUNC(32f32s, cvt_32f, float, int, float)
DEF_CVT_SCALE_FUNC(64f32s, cvt_64f, double, int, double)
//DEF_CVT_SCALE_FUNC(16f32s, cvt1_32f, float16_t, int, float)
DEF_CVT_SCALE_FUNC(8u32f, cvt_32f, uchar, float, float)
DEF_CVT_SCALE_FUNC(8s32f, cvt_32f, schar, float, float)
DEF_CVT_SCALE_FUNC(16u32f, cvt_32f, ushort, float, float)
DEF_CVT_SCALE_FUNC(16s32f, cvt_32f, short, float, float)
DEF_CVT_SCALE_FUNC(32s32f, cvt_32f, int, float, float)
DEF_CVT_SCALE_FUNC(32f, cvt_32f, float, float, float)
DEF_CVT_SCALE_FUNC(64f32f, cvt_64f, double, float, double)
//DEF_CVT_SCALE_FUNC(16f32f, cvt1_32f, float16_t, float, float)
DEF_CVT_SCALE_FUNC(8u64f, cvt_64f, uchar, double, double)
DEF_CVT_SCALE_FUNC(8s64f, cvt_64f, schar, double, double)
DEF_CVT_SCALE_FUNC(16u64f, cvt_64f, ushort, double, double)
DEF_CVT_SCALE_FUNC(16s64f, cvt_64f, short, double, double)
DEF_CVT_SCALE_FUNC(32s64f, cvt_64f, int, double, double)
DEF_CVT_SCALE_FUNC(32f64f, cvt_64f, float, double, double)
DEF_CVT_SCALE_FUNC(64f, cvt_64f, double, double, double)
//DEF_CVT_SCALE_FUNC(16f64f, cvt_64f, float16_t, double, double)
/*DEF_CVT_SCALE_FUNC(8u16f, cvt1_32f, uchar, float16_t, float)
DEF_CVT_SCALE_FUNC(8s16f, cvt1_32f, schar, float16_t, float)
DEF_CVT_SCALE_FUNC(16u16f, cvt1_32f, ushort, float16_t, float)
DEF_CVT_SCALE_FUNC(16s16f, cvt1_32f, short, float16_t, float)
DEF_CVT_SCALE_FUNC(32s16f, cvt1_32f, int, float16_t, float)
DEF_CVT_SCALE_FUNC(32f16f, cvt1_32f, float, float16_t, float)
DEF_CVT_SCALE_FUNC(64f16f, cvt_64f, double, float16_t, double)
DEF_CVT_SCALE_FUNC(16f, cvt1_32f, float16_t, float16_t, float)*/
static BinaryFunc getCvtScaleAbsFunc(int depth)
{
static BinaryFunc cvtScaleAbsTab[] =
{
(BinaryFunc)cvtScaleAbs8u, (BinaryFunc)cvtScaleAbs8s8u, (BinaryFunc)cvtScaleAbs16u8u,
(BinaryFunc)cvtScaleAbs16s8u, (BinaryFunc)cvtScaleAbs32s8u, (BinaryFunc)cvtScaleAbs32f8u,
(BinaryFunc)cvtScaleAbs64f8u, 0
};
return cvtScaleAbsTab[depth];
CV_INSTRUMENT_REGION();
CV_CPU_DISPATCH(getCvtScaleAbsFunc, (depth),
CV_CPU_DISPATCH_MODES_ALL);
}
BinaryFunc getConvertScaleFunc(int sdepth, int ddepth)
{
static BinaryFunc cvtScaleTab[][8] =
{
{
(BinaryFunc)GET_OPTIMIZED(cvtScale8u), (BinaryFunc)GET_OPTIMIZED(cvtScale8s8u), (BinaryFunc)GET_OPTIMIZED(cvtScale16u8u),
(BinaryFunc)GET_OPTIMIZED(cvtScale16s8u), (BinaryFunc)GET_OPTIMIZED(cvtScale32s8u), (BinaryFunc)GET_OPTIMIZED(cvtScale32f8u),
(BinaryFunc)cvtScale64f8u, 0 //(BinaryFunc)cvtScale16f8u
},
{
(BinaryFunc)GET_OPTIMIZED(cvtScale8u8s), (BinaryFunc)GET_OPTIMIZED(cvtScale8s), (BinaryFunc)GET_OPTIMIZED(cvtScale16u8s),
(BinaryFunc)GET_OPTIMIZED(cvtScale16s8s), (BinaryFunc)GET_OPTIMIZED(cvtScale32s8s), (BinaryFunc)GET_OPTIMIZED(cvtScale32f8s),
(BinaryFunc)cvtScale64f8s, 0 //(BinaryFunc)cvtScale16f8s
},
{
(BinaryFunc)GET_OPTIMIZED(cvtScale8u16u), (BinaryFunc)GET_OPTIMIZED(cvtScale8s16u), (BinaryFunc)GET_OPTIMIZED(cvtScale16u),
(BinaryFunc)GET_OPTIMIZED(cvtScale16s16u), (BinaryFunc)GET_OPTIMIZED(cvtScale32s16u), (BinaryFunc)GET_OPTIMIZED(cvtScale32f16u),
(BinaryFunc)cvtScale64f16u, 0 //(BinaryFunc)cvtScale16f16u
},
{
(BinaryFunc)GET_OPTIMIZED(cvtScale8u16s), (BinaryFunc)GET_OPTIMIZED(cvtScale8s16s), (BinaryFunc)GET_OPTIMIZED(cvtScale16u16s),
(BinaryFunc)GET_OPTIMIZED(cvtScale16s), (BinaryFunc)GET_OPTIMIZED(cvtScale32s16s), (BinaryFunc)GET_OPTIMIZED(cvtScale32f16s),
(BinaryFunc)cvtScale64f16s, 0 //(BinaryFunc)cvtScale16f16s
},
{
(BinaryFunc)GET_OPTIMIZED(cvtScale8u32s), (BinaryFunc)GET_OPTIMIZED(cvtScale8s32s), (BinaryFunc)GET_OPTIMIZED(cvtScale16u32s),
(BinaryFunc)GET_OPTIMIZED(cvtScale16s32s), (BinaryFunc)GET_OPTIMIZED(cvtScale32s), (BinaryFunc)GET_OPTIMIZED(cvtScale32f32s),
(BinaryFunc)cvtScale64f32s, 0 //(BinaryFunc)cvtScale16f32s
},
{
(BinaryFunc)GET_OPTIMIZED(cvtScale8u32f), (BinaryFunc)GET_OPTIMIZED(cvtScale8s32f), (BinaryFunc)GET_OPTIMIZED(cvtScale16u32f),
(BinaryFunc)GET_OPTIMIZED(cvtScale16s32f), (BinaryFunc)GET_OPTIMIZED(cvtScale32s32f), (BinaryFunc)GET_OPTIMIZED(cvtScale32f),
(BinaryFunc)cvtScale64f32f, 0 //(BinaryFunc)cvtScale16f32f
},
{
(BinaryFunc)cvtScale8u64f, (BinaryFunc)cvtScale8s64f, (BinaryFunc)cvtScale16u64f,
(BinaryFunc)cvtScale16s64f, (BinaryFunc)cvtScale32s64f, (BinaryFunc)cvtScale32f64f,
(BinaryFunc)cvtScale64f, 0 //(BinaryFunc)cvtScale16f64f
},
{
0, 0, 0, 0, 0, 0, 0, 0
/*(BinaryFunc)cvtScale8u16f, (BinaryFunc)cvtScale8s16f, (BinaryFunc)cvtScale16u16f,
(BinaryFunc)cvtScale16s16f, (BinaryFunc)cvtScale32s16f, (BinaryFunc)cvtScale32f16f,
(BinaryFunc)cvtScale64f16f, (BinaryFunc)cvtScale16f*/
},
};
return cvtScaleTab[CV_MAT_DEPTH(ddepth)][CV_MAT_DEPTH(sdepth)];
CV_INSTRUMENT_REGION();
CV_CPU_DISPATCH(getConvertScaleFunc, (sdepth, ddepth),
CV_CPU_DISPATCH_MODES_ALL);
}
#ifdef HAVE_OPENCL
@@ -407,10 +85,7 @@ static bool ocl_convertScaleAbs( InputArray _src, OutputArray _dst, double alpha
#endif
} //cv::
void cv::convertScaleAbs( InputArray _src, OutputArray _dst, double alpha, double beta )
void convertScaleAbs(InputArray _src, OutputArray _dst, double alpha, double beta)
{
CV_INSTRUMENT_REGION();
@@ -444,8 +119,6 @@ void cv::convertScaleAbs( InputArray _src, OutputArray _dst, double alpha, doubl
//==================================================================================================
namespace cv {
#ifdef HAVE_OPENCL
static bool ocl_normalize( InputArray _src, InputOutputArray _dst, InputArray _mask, int dtype,
@@ -535,10 +208,8 @@ static bool ocl_normalize( InputArray _src, InputOutputArray _dst, InputArray _m
#endif
} // cv::
void cv::normalize( InputArray _src, InputOutputArray _dst, double a, double b,
int norm_type, int rtype, InputArray _mask )
void normalize(InputArray _src, InputOutputArray _dst, double a, double b,
int norm_type, int rtype, InputArray _mask)
{
CV_INSTRUMENT_REGION();
@@ -584,3 +255,5 @@ void cv::normalize( InputArray _src, InputOutputArray _dst, double a, double b,
temp.copyTo( _dst, _mask );
}
}
} // namespace