core: dispatch convert

This commit is contained in:
Alexander Alekhin 2019-02-08 17:37:59 +03:00
parent 33d477e049
commit 39b90ae9fb
3 changed files with 53 additions and 429 deletions

View File

@ -3,6 +3,7 @@ set(the_description "The Core Functionality")
ocv_add_dispatched_file(mathfuncs_core SSE2 AVX AVX2) ocv_add_dispatched_file(mathfuncs_core SSE2 AVX AVX2)
ocv_add_dispatched_file(stat SSE4_2 AVX2) ocv_add_dispatched_file(stat SSE4_2 AVX2)
ocv_add_dispatched_file(arithm SSE2 SSE4_1 AVX2 VSX3) ocv_add_dispatched_file(arithm SSE2 SSE4_1 AVX2 VSX3)
ocv_add_dispatched_file(convert SSE2 AVX2)
# dispatching for accuracy tests # dispatching for accuracy tests
ocv_add_dispatched_file_force_all(test_intrin128 TEST SSE2 SSE3 SSSE3 SSE4_1 SSE4_2 AVX FP16 AVX2) ocv_add_dispatched_file_force_all(test_intrin128 TEST SSE2 SSE3 SSSE3 SSE4_1 SSE4_2 AVX FP16 AVX2)

View File

@ -4,239 +4,12 @@
#include "precomp.hpp" #include "precomp.hpp"
#include "opencl_kernels_core.hpp" #include "opencl_kernels_core.hpp"
#include "convert.hpp"
#include "convert.simd.hpp"
#include "convert.simd_declarations.hpp" // defines CV_CPU_DISPATCH_MODES_ALL=AVX2,...,BASELINE based on CMakeLists.txt content
namespace cv { namespace cv {
/*namespace hal {
void cvt16f32f( const float16_t* src, float* dst, int len )
{
int j = 0;
#if CV_SIMD
const int VECSZ = v_float32::nlanes;
for( ; j < len; j += VECSZ )
{
if( j > len - VECSZ )
{
if( j == 0 )
break;
j = len - VECSZ;
}
v_store(dst + j, vx_load_expand(src + j));
}
#endif
for( ; j < len; j++ )
dst[j] = (float)src[j];
}
void cvt32f16f( const float* src, float16_t* dst, int len )
{
int j = 0;
#if CV_SIMD
const int VECSZ = v_float32::nlanes;
for( ; j < len; j += VECSZ )
{
if( j > len - VECSZ )
{
if( j == 0 )
break;
j = len - VECSZ;
}
v_pack_store(dst + j, vx_load(src + j));
}
#endif
for( ; j < len; j++ )
dst[j] = float16_t(src[j]);
}
/*void addRNGBias32f( float* arr, const float* scaleBiasPairs, int len )
{
// the loop is simple enough, so we let the compiler to vectorize it
for( int i = 0; i < len; i++ )
arr[i] = scaleBiasPairs[i*2 + 1];
}
void addRNGBias64f( double* arr, const double* scaleBiasPairs, int len )
{
// the loop is simple enough, so we let the compiler to vectorize it
for( int i = 0; i < len; i++ )
arr[i] = scaleBiasPairs[i*2 + 1];
}
}*/
template<typename _Ts, typename _Td, typename _Twvec> inline void
cvt_( const _Ts* src, size_t sstep, _Td* dst, size_t dstep, Size size )
{
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
const int VECSZ = _Twvec::nlanes*2;
for( ; j < size.width; j += VECSZ )
{
if( j > size.width - VECSZ )
{
if( j == 0 || src == (_Ts*)dst )
break;
j = size.width - VECSZ;
}
_Twvec v0, v1;
vx_load_pair_as(src + j, v0, v1);
v_store_pair_as(dst + j, v0, v1);
}
#endif
for( ; j < size.width; j++ )
dst[j] = saturate_cast<_Td>(src[j]);
}
}
// in order to reduce the code size, for (16f <-> ...) conversions
// we add a conversion function without loop unrolling
template<typename _Ts, typename _Td, typename _Twvec> inline void
cvt1_( const _Ts* src, size_t sstep, _Td* dst, size_t dstep, Size size )
{
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
const int VECSZ = _Twvec::nlanes;
for( ; j < size.width; j += VECSZ )
{
if( j > size.width - VECSZ )
{
if( j == 0 || src == (_Ts*)dst )
break;
j = size.width - VECSZ;
}
_Twvec v;
vx_load_as(src + j, v);
v_store_as(dst + j, v);
}
vx_cleanup();
#endif
for( ; j < size.width; j++ )
dst[j] = saturate_cast<_Td>(src[j]);
}
}
static void cvtCopy( const uchar* src, size_t sstep,
uchar* dst, size_t dstep, Size size, size_t elemsize)
{
size_t len = size.width*elemsize;
for( int i = 0; i < size.height; i++, src += sstep, dst += dstep )
{
memcpy( dst, src, len );
}
}
#define DEF_CVT_FUNC(suffix, cvtfunc, _Ts, _Td, _Twvec) \
static void cvt##suffix(const _Ts* src, size_t sstep, uchar*, size_t, \
_Td* dst, size_t dstep, Size size, void*) \
{ cvtfunc<_Ts, _Td, _Twvec>(src, sstep, dst, dstep, size); }
////////////////////// 8u -> ... ////////////////////////
DEF_CVT_FUNC(8u8s, cvt_, uchar, schar, v_int16)
DEF_CVT_FUNC(8u16u, cvt_, uchar, ushort, v_uint16)
DEF_CVT_FUNC(8u16s, cvt_, uchar, short, v_int16)
DEF_CVT_FUNC(8u32s, cvt_, uchar, int, v_int32)
DEF_CVT_FUNC(8u32f, cvt_, uchar, float, v_float32)
DEF_CVT_FUNC(8u64f, cvt_, uchar, double, v_int32)
//DEF_CVT_FUNC(8u16f, cvt1_, uchar, float16_t, v_float32)
////////////////////// 8s -> ... ////////////////////////
DEF_CVT_FUNC(8s8u, cvt_, schar, uchar, v_int16)
DEF_CVT_FUNC(8s16u, cvt_, schar, ushort, v_uint16)
DEF_CVT_FUNC(8s16s, cvt_, schar, short, v_int16)
DEF_CVT_FUNC(8s32s, cvt_, schar, int, v_int32)
DEF_CVT_FUNC(8s32f, cvt_, schar, float, v_float32)
DEF_CVT_FUNC(8s64f, cvt_, schar, double, v_int32)
//DEF_CVT_FUNC(8s16f, cvt1_, schar, float16_t, v_float32)
////////////////////// 16u -> ... ////////////////////////
DEF_CVT_FUNC(16u8u, cvt_, ushort, uchar, v_uint16)
DEF_CVT_FUNC(16u8s, cvt_, ushort, schar, v_uint16)
DEF_CVT_FUNC(16u16s, cvt_, ushort, short, v_int32)
DEF_CVT_FUNC(16u32s, cvt_, ushort, int, v_int32)
DEF_CVT_FUNC(16u32f, cvt_, ushort, float, v_float32)
DEF_CVT_FUNC(16u64f, cvt_, ushort, double, v_int32)
//DEF_CVT_FUNC(16u16f, cvt1_,ushort, float16_t, v_float32)
////////////////////// 16s -> ... ////////////////////////
DEF_CVT_FUNC(16s8u, cvt_, short, uchar, v_int16)
DEF_CVT_FUNC(16s8s, cvt_, short, schar, v_int16)
DEF_CVT_FUNC(16s16u, cvt_, short, ushort, v_int32)
DEF_CVT_FUNC(16s32s, cvt_, short, int, v_int32)
DEF_CVT_FUNC(16s32f, cvt_, short, float, v_float32)
DEF_CVT_FUNC(16s64f, cvt_, short, double, v_int32)
//DEF_CVT_FUNC(16s16f, cvt1_,short, float16_t, v_float32)
////////////////////// 32s -> ... ////////////////////////
DEF_CVT_FUNC(32s8u, cvt_, int, uchar, v_int32)
DEF_CVT_FUNC(32s8s, cvt_, int, schar, v_int32)
DEF_CVT_FUNC(32s16u, cvt_, int, ushort, v_int32)
DEF_CVT_FUNC(32s16s, cvt_, int, short, v_int32)
DEF_CVT_FUNC(32s32f, cvt_, int, float, v_float32)
DEF_CVT_FUNC(32s64f, cvt_, int, double, v_int32)
//DEF_CVT_FUNC(32s16f, cvt1_,int, float16_t, v_float32)
////////////////////// 32f -> ... ////////////////////////
DEF_CVT_FUNC(32f8u, cvt_, float, uchar, v_float32)
DEF_CVT_FUNC(32f8s, cvt_, float, schar, v_float32)
DEF_CVT_FUNC(32f16u, cvt_, float, ushort, v_float32)
DEF_CVT_FUNC(32f16s, cvt_, float, short, v_float32)
DEF_CVT_FUNC(32f32s, cvt_, float, int, v_float32)
DEF_CVT_FUNC(32f64f, cvt_, float, double, v_float32)
DEF_CVT_FUNC(32f16f, cvt1_,float, float16_t, v_float32)
////////////////////// 64f -> ... ////////////////////////
DEF_CVT_FUNC(64f8u, cvt_, double, uchar, v_int32)
DEF_CVT_FUNC(64f8s, cvt_, double, schar, v_int32)
DEF_CVT_FUNC(64f16u, cvt_, double, ushort, v_int32)
DEF_CVT_FUNC(64f16s, cvt_, double, short, v_int32)
DEF_CVT_FUNC(64f32s, cvt_, double, int, v_int32)
DEF_CVT_FUNC(64f32f, cvt_, double, float, v_float32)
//DEF_CVT_FUNC(64f16f, cvt1_,double, float16_t, v_float32)
////////////////////// 16f -> ... ////////////////////////
//DEF_CVT_FUNC(16f8u, cvt_, float16_t, uchar, v_float32)
//DEF_CVT_FUNC(16f8s, cvt_, float16_t, schar, v_float32)
//DEF_CVT_FUNC(16f16u, cvt1_, float16_t, ushort, v_float32)
//DEF_CVT_FUNC(16f16s, cvt1_, float16_t, short, v_float32)
//DEF_CVT_FUNC(16f32s, cvt1_, float16_t, int, v_float32)
DEF_CVT_FUNC(16f32f, cvt1_, float16_t, float, v_float32)
//DEF_CVT_FUNC(16f64f, cvt1_, float16_t, double, v_float32)
///////////// "conversion" w/o conversion ///////////////
static void cvt8u(const uchar* src, size_t sstep, uchar*, size_t, uchar* dst, size_t dstep, Size size, void*)
{ cvtCopy(src, sstep, dst, dstep, size, 1); }
static void cvt16u(const ushort* src, size_t sstep, uchar*, size_t, ushort* dst, size_t dstep, Size size, void*)
{ cvtCopy((const uchar*)src, sstep, (uchar*)dst, dstep, size, 2); }
static void cvt32s(const int* src, size_t sstep, uchar*, size_t, int* dst, size_t dstep, Size size, void*)
{ cvtCopy((const uchar*)src, sstep, (uchar*)dst, dstep, size, 4); }
static void cvt64s(const int64* src, size_t sstep, uchar*, size_t, int64* dst, size_t dstep, Size size, void*)
{ cvtCopy((const uchar*)src, sstep, (uchar*)dst, dstep, size, 8); }
/* [TODO] Recover IPP calls /* [TODO] Recover IPP calls
#if defined(HAVE_IPP) #if defined(HAVE_IPP)
#define DEF_CVT_FUNC_F(suffix, stype, dtype, ippFavor) \ #define DEF_CVT_FUNC_F(suffix, stype, dtype, ippFavor) \
@ -334,50 +107,22 @@ DEF_CPY_FUNC(64s, int64)
BinaryFunc getConvertFunc(int sdepth, int ddepth) BinaryFunc getConvertFunc(int sdepth, int ddepth)
{ {
static BinaryFunc cvtTab[][8] = CV_INSTRUMENT_REGION();
{ CV_CPU_DISPATCH(getConvertFunc, (sdepth, ddepth),
{ CV_CPU_DISPATCH_MODES_ALL);
(BinaryFunc)(cvt8u), (BinaryFunc)GET_OPTIMIZED(cvt8s8u), (BinaryFunc)GET_OPTIMIZED(cvt16u8u), }
(BinaryFunc)GET_OPTIMIZED(cvt16s8u), (BinaryFunc)GET_OPTIMIZED(cvt32s8u), (BinaryFunc)GET_OPTIMIZED(cvt32f8u),
(BinaryFunc)GET_OPTIMIZED(cvt64f8u), 0 //(BinaryFunc)(cvt16f8u) static BinaryFunc get_cvt32f16f()
}, {
{ CV_INSTRUMENT_REGION();
(BinaryFunc)GET_OPTIMIZED(cvt8u8s), (BinaryFunc)cvt8u, (BinaryFunc)GET_OPTIMIZED(cvt16u8s), CV_CPU_DISPATCH(get_cvt32f16f, (),
(BinaryFunc)GET_OPTIMIZED(cvt16s8s), (BinaryFunc)GET_OPTIMIZED(cvt32s8s), (BinaryFunc)GET_OPTIMIZED(cvt32f8s), CV_CPU_DISPATCH_MODES_ALL);
(BinaryFunc)GET_OPTIMIZED(cvt64f8s), 0 //(BinaryFunc)(cvt16f8s) }
}, static BinaryFunc get_cvt16f32f()
{ {
(BinaryFunc)GET_OPTIMIZED(cvt8u16u), (BinaryFunc)GET_OPTIMIZED(cvt8s16u), (BinaryFunc)cvt16u, CV_INSTRUMENT_REGION();
(BinaryFunc)GET_OPTIMIZED(cvt16s16u), (BinaryFunc)GET_OPTIMIZED(cvt32s16u), (BinaryFunc)GET_OPTIMIZED(cvt32f16u), CV_CPU_DISPATCH(get_cvt16f32f, (),
(BinaryFunc)GET_OPTIMIZED(cvt64f16u), 0 //(BinaryFunc)(cvt16f16u) CV_CPU_DISPATCH_MODES_ALL);
},
{
(BinaryFunc)GET_OPTIMIZED(cvt8u16s), (BinaryFunc)GET_OPTIMIZED(cvt8s16s), (BinaryFunc)GET_OPTIMIZED(cvt16u16s),
(BinaryFunc)cvt16u, (BinaryFunc)GET_OPTIMIZED(cvt32s16s), (BinaryFunc)GET_OPTIMIZED(cvt32f16s),
(BinaryFunc)GET_OPTIMIZED(cvt64f16s), 0 //(BinaryFunc)(cvt16f16s)
},
{
(BinaryFunc)GET_OPTIMIZED(cvt8u32s), (BinaryFunc)GET_OPTIMIZED(cvt8s32s), (BinaryFunc)GET_OPTIMIZED(cvt16u32s),
(BinaryFunc)GET_OPTIMIZED(cvt16s32s), (BinaryFunc)cvt32s, (BinaryFunc)GET_OPTIMIZED(cvt32f32s),
(BinaryFunc)GET_OPTIMIZED(cvt64f32s), 0 //(BinaryFunc)(cvt16f32s)
},
{
(BinaryFunc)GET_OPTIMIZED(cvt8u32f), (BinaryFunc)GET_OPTIMIZED(cvt8s32f), (BinaryFunc)GET_OPTIMIZED(cvt16u32f),
(BinaryFunc)GET_OPTIMIZED(cvt16s32f), (BinaryFunc)GET_OPTIMIZED(cvt32s32f), (BinaryFunc)cvt32s,
(BinaryFunc)GET_OPTIMIZED(cvt64f32f), 0 //(BinaryFunc)(cvt16f32f)
},
{
(BinaryFunc)GET_OPTIMIZED(cvt8u64f), (BinaryFunc)GET_OPTIMIZED(cvt8s64f), (BinaryFunc)GET_OPTIMIZED(cvt16u64f),
(BinaryFunc)GET_OPTIMIZED(cvt16s64f), (BinaryFunc)GET_OPTIMIZED(cvt32s64f), (BinaryFunc)GET_OPTIMIZED(cvt32f64f),
(BinaryFunc)(cvt64s), 0 //(BinaryFunc)(cvt16f64f)
},
{
0, 0, 0, 0, 0, 0, 0, 0
//(BinaryFunc)(cvt8u16f), (BinaryFunc)(cvt8s16f), (BinaryFunc)(cvt16u16f), (BinaryFunc)(cvt16s16f),
//(BinaryFunc)(cvt32s16f), (BinaryFunc)(cvt32f16f), (BinaryFunc)(cvt64f16f), (BinaryFunc)(cvt16u)
}
};
return cvtTab[CV_MAT_DEPTH(ddepth)][CV_MAT_DEPTH(sdepth)];
} }
#ifdef HAVE_OPENCL #ifdef HAVE_OPENCL
@ -410,9 +155,7 @@ static bool ocl_convertFp16( InputArray _src, OutputArray _dst, int sdepth, int
} }
#endif #endif
} // cv:: void Mat::convertTo(OutputArray _dst, int _type, double alpha, double beta) const
void cv::Mat::convertTo(OutputArray _dst, int _type, double alpha, double beta) const
{ {
CV_INSTRUMENT_REGION(); CV_INSTRUMENT_REGION();
@ -467,7 +210,7 @@ void cv::Mat::convertTo(OutputArray _dst, int _type, double alpha, double beta)
//================================================================================================== //==================================================================================================
void cv::convertFp16( InputArray _src, OutputArray _dst ) void convertFp16(InputArray _src, OutputArray _dst)
{ {
CV_INSTRUMENT_REGION(); CV_INSTRUMENT_REGION();
@ -485,12 +228,12 @@ void cv::convertFp16( InputArray _src, OutputArray _dst )
} }
else else
ddepth = CV_16S; ddepth = CV_16S;
func = (BinaryFunc)cvt32f16f; func = (BinaryFunc)get_cvt32f16f();
break; break;
case CV_16S: case CV_16S:
//case CV_16F: //case CV_16F:
ddepth = CV_32F; ddepth = CV_32F;
func = (BinaryFunc)cvt16f32f; func = (BinaryFunc)get_cvt16f32f();
break; break;
default: default:
CV_Error(Error::StsUnsupportedFormat, "Unsupported input depth"); CV_Error(Error::StsUnsupportedFormat, "Unsupported input depth");
@ -525,3 +268,5 @@ void cv::convertFp16( InputArray _src, OutputArray _dst )
func(ptrs[0], 0, 0, 0, ptrs[1], 0, sz, 0); func(ptrs[0], 0, 0, 0, ptrs[1], 0, sz, 0);
} }
} }
} // namespace cv

View File

@ -3,10 +3,16 @@
// of this distribution and at http://opencv.org/license.html // of this distribution and at http://opencv.org/license.html
#include "precomp.hpp" #include "precomp.hpp"
#include "opencl_kernels_core.hpp"
#include "convert.hpp" #include "convert.hpp"
namespace cv { namespace cv {
CV_CPU_OPTIMIZATION_NAMESPACE_BEGIN
BinaryFunc getConvertFunc(int sdepth, int ddepth);
BinaryFunc get_cvt32f16f();
BinaryFunc get_cvt16f32f();
#ifndef CV_CPU_OPTIMIZATION_DECLARATIONS_ONLY
/*namespace hal { /*namespace hal {
@ -66,7 +72,7 @@ void addRNGBias64f( double* arr, const double* scaleBiasPairs, int len )
}*/ }*/
template<typename _Ts, typename _Td, typename _Twvec> inline void template<typename _Ts, typename _Td, typename _Twvec> static inline void
cvt_( const _Ts* src, size_t sstep, _Td* dst, size_t dstep, Size size ) cvt_( const _Ts* src, size_t sstep, _Td* dst, size_t dstep, Size size )
{ {
sstep /= sizeof(src[0]); sstep /= sizeof(src[0]);
@ -97,7 +103,7 @@ cvt_( const _Ts* src, size_t sstep, _Td* dst, size_t dstep, Size size )
// in order to reduce the code size, for (16f <-> ...) conversions // in order to reduce the code size, for (16f <-> ...) conversions
// we add a conversion function without loop unrolling // we add a conversion function without loop unrolling
template<typename _Ts, typename _Td, typename _Twvec> inline void template<typename _Ts, typename _Td, typename _Twvec> static inline void
cvt1_( const _Ts* src, size_t sstep, _Td* dst, size_t dstep, Size size ) cvt1_( const _Ts* src, size_t sstep, _Td* dst, size_t dstep, Size size )
{ {
sstep /= sizeof(src[0]); sstep /= sizeof(src[0]);
@ -140,7 +146,10 @@ static void cvtCopy( const uchar* src, size_t sstep,
#define DEF_CVT_FUNC(suffix, cvtfunc, _Ts, _Td, _Twvec) \ #define DEF_CVT_FUNC(suffix, cvtfunc, _Ts, _Td, _Twvec) \
static void cvt##suffix(const _Ts* src, size_t sstep, uchar*, size_t, \ static void cvt##suffix(const _Ts* src, size_t sstep, uchar*, size_t, \
_Td* dst, size_t dstep, Size size, void*) \ _Td* dst, size_t dstep, Size size, void*) \
{ cvtfunc<_Ts, _Td, _Twvec>(src, sstep, dst, dstep, size); } { \
CV_INSTRUMENT_REGION(); \
cvtfunc<_Ts, _Td, _Twvec>(src, sstep, dst, dstep, size); \
}
////////////////////// 8u -> ... //////////////////////// ////////////////////// 8u -> ... ////////////////////////
@ -225,16 +234,16 @@ DEF_CVT_FUNC(16f32f, cvt1_, float16_t, float, v_float32)
///////////// "conversion" w/o conversion /////////////// ///////////// "conversion" w/o conversion ///////////////
static void cvt8u(const uchar* src, size_t sstep, uchar*, size_t, uchar* dst, size_t dstep, Size size, void*) static void cvt8u(const uchar* src, size_t sstep, uchar*, size_t, uchar* dst, size_t dstep, Size size, void*)
{ cvtCopy(src, sstep, dst, dstep, size, 1); } { CV_INSTRUMENT_REGION(); cvtCopy(src, sstep, dst, dstep, size, 1); }
static void cvt16u(const ushort* src, size_t sstep, uchar*, size_t, ushort* dst, size_t dstep, Size size, void*) static void cvt16u(const ushort* src, size_t sstep, uchar*, size_t, ushort* dst, size_t dstep, Size size, void*)
{ cvtCopy((const uchar*)src, sstep, (uchar*)dst, dstep, size, 2); } { CV_INSTRUMENT_REGION(); cvtCopy((const uchar*)src, sstep, (uchar*)dst, dstep, size, 2); }
static void cvt32s(const int* src, size_t sstep, uchar*, size_t, int* dst, size_t dstep, Size size, void*) static void cvt32s(const int* src, size_t sstep, uchar*, size_t, int* dst, size_t dstep, Size size, void*)
{ cvtCopy((const uchar*)src, sstep, (uchar*)dst, dstep, size, 4); } { CV_INSTRUMENT_REGION(); cvtCopy((const uchar*)src, sstep, (uchar*)dst, dstep, size, 4); }
static void cvt64s(const int64* src, size_t sstep, uchar*, size_t, int64* dst, size_t dstep, Size size, void*) static void cvt64s(const int64* src, size_t sstep, uchar*, size_t, int64* dst, size_t dstep, Size size, void*)
{ cvtCopy((const uchar*)src, sstep, (uchar*)dst, dstep, size, 8); } { CV_INSTRUMENT_REGION(); cvtCopy((const uchar*)src, sstep, (uchar*)dst, dstep, size, 8); }
/* [TODO] Recover IPP calls /* [TODO] Recover IPP calls
@ -380,148 +389,17 @@ BinaryFunc getConvertFunc(int sdepth, int ddepth)
return cvtTab[CV_MAT_DEPTH(ddepth)][CV_MAT_DEPTH(sdepth)]; return cvtTab[CV_MAT_DEPTH(ddepth)][CV_MAT_DEPTH(sdepth)];
} }
#ifdef HAVE_OPENCL BinaryFunc get_cvt32f16f()
static bool ocl_convertFp16( InputArray _src, OutputArray _dst, int sdepth, int ddepth )
{ {
int type = _src.type(), cn = CV_MAT_CN(type); return (BinaryFunc)cvt32f16f;
_dst.createSameSize( _src, CV_MAKETYPE(ddepth, cn) );
int kercn = 1;
int rowsPerWI = 1;
String build_opt = format("-D HALF_SUPPORT -D srcT=%s -D dstT=%s -D rowsPerWI=%d%s",
sdepth == CV_32F ? "float" : "half",
sdepth == CV_32F ? "half" : "float",
rowsPerWI,
sdepth == CV_32F ? " -D FLOAT_TO_HALF " : "");
ocl::Kernel k("convertFp16", ocl::core::halfconvert_oclsrc, build_opt);
if (k.empty())
return false;
UMat src = _src.getUMat();
UMat dst = _dst.getUMat();
ocl::KernelArg srcarg = ocl::KernelArg::ReadOnlyNoSize(src),
dstarg = ocl::KernelArg::WriteOnly(dst, cn, kercn);
k.args(srcarg, dstarg);
size_t globalsize[2] = { (size_t)src.cols * cn / kercn, ((size_t)src.rows + rowsPerWI - 1) / rowsPerWI };
return k.run(2, globalsize, NULL, false);
} }
BinaryFunc get_cvt16f32f()
{
return (BinaryFunc)cvt16f32f;
}
#endif #endif
} // cv:: CV_CPU_OPTIMIZATION_NAMESPACE_END
} // namespace
void cv::Mat::convertTo(OutputArray _dst, int _type, double alpha, double beta) const
{
CV_INSTRUMENT_REGION();
if( empty() )
{
_dst.release();
return;
}
bool noScale = fabs(alpha-1) < DBL_EPSILON && fabs(beta) < DBL_EPSILON;
if( _type < 0 )
_type = _dst.fixedType() ? _dst.type() : type();
else
_type = CV_MAKETYPE(CV_MAT_DEPTH(_type), channels());
int sdepth = depth(), ddepth = CV_MAT_DEPTH(_type);
if( sdepth == ddepth && noScale )
{
copyTo(_dst);
return;
}
Mat src = *this;
if( dims <= 2 )
_dst.create( size(), _type );
else
_dst.create( dims, size, _type );
Mat dst = _dst.getMat();
BinaryFunc func = noScale ? getConvertFunc(sdepth, ddepth) : getConvertScaleFunc(sdepth, ddepth);
double scale[] = {alpha, beta};
int cn = channels();
CV_Assert( func != 0 );
if( dims <= 2 )
{
Size sz = getContinuousSize2D(src, dst, cn);
func( src.data, src.step, 0, 0, dst.data, dst.step, sz, scale );
}
else
{
const Mat* arrays[] = {&src, &dst, 0};
uchar* ptrs[2] = {};
NAryMatIterator it(arrays, ptrs);
Size sz((int)(it.size*cn), 1);
for( size_t i = 0; i < it.nplanes; i++, ++it )
func(ptrs[0], 1, 0, 0, ptrs[1], 1, sz, scale);
}
}
//==================================================================================================
void cv::convertFp16( InputArray _src, OutputArray _dst )
{
CV_INSTRUMENT_REGION();
int sdepth = _src.depth(), ddepth = 0;
BinaryFunc func = 0;
switch( sdepth )
{
case CV_32F:
if(_dst.fixedType())
{
ddepth = _dst.depth();
CV_Assert(ddepth == CV_16S /*|| ddepth == CV_16F*/);
CV_Assert(_dst.channels() == _src.channels());
}
else
ddepth = CV_16S;
func = (BinaryFunc)cvt32f16f;
break;
case CV_16S:
//case CV_16F:
ddepth = CV_32F;
func = (BinaryFunc)cvt16f32f;
break;
default:
CV_Error(Error::StsUnsupportedFormat, "Unsupported input depth");
return;
}
CV_OCL_RUN(_src.dims() <= 2 && _dst.isUMat(),
ocl_convertFp16(_src, _dst, sdepth, ddepth))
Mat src = _src.getMat();
int type = CV_MAKETYPE(ddepth, src.channels());
_dst.create( src.dims, src.size, type );
Mat dst = _dst.getMat();
int cn = src.channels();
CV_Assert( func != 0 );
if( src.dims <= 2 )
{
Size sz = getContinuousSize2D(src, dst, cn);
func( src.data, src.step, 0, 0, dst.data, dst.step, sz, 0);
}
else
{
const Mat* arrays[] = {&src, &dst, 0};
uchar* ptrs[2] = {};
NAryMatIterator it(arrays, ptrs);
Size sz((int)(it.size*cn), 1);
for( size_t i = 0; i < it.nplanes; i++, ++it )
func(ptrs[0], 0, 0, 0, ptrs[1], 0, sz, 0);
}
}