Merge pull request #13780 from alalek:core_dispatch_convertTo

modules/core/CMakeLists.txt

@@ -3,6 +3,8 @@ set(the_description "The Core Functionality")
 ocv_add_dispatched_file(mathfuncs_core SSE2 AVX AVX2)
 ocv_add_dispatched_file(stat SSE4_2 AVX2)
 ocv_add_dispatched_file(arithm SSE2 SSE4_1 AVX2 VSX3)
+ocv_add_dispatched_file(convert SSE2 AVX2)
+ocv_add_dispatched_file(convert_scale SSE2 AVX2)
 
 # dispatching for accuracy tests
 ocv_add_dispatched_file_force_all(test_intrin128 TEST SSE2 SSE3 SSSE3 SSE4_1 SSE4_2 AVX FP16 AVX2)

modules/core/src/convert.dispatch.cpp

@@ -0,0 +1,272 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html
+
+#include "precomp.hpp"
+#include "opencl_kernels_core.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 {
+
+/* [TODO] Recover IPP calls
+#if defined(HAVE_IPP)
+#define DEF_CVT_FUNC_F(suffix, stype, dtype, ippFavor) \
+static void cvt##suffix( const stype* src, size_t sstep, const uchar*, size_t, \
+                         dtype* dst, size_t dstep, Size size, double*) \
+{ \
+    CV_IPP_RUN(src && dst, CV_INSTRUMENT_FUN_IPP(ippiConvert_##ippFavor, src, (int)sstep, dst, (int)dstep, ippiSize(size.width, size.height)) >= 0) \
+    cvt_(src, sstep, dst, dstep, size); \
+}
+
+#define DEF_CVT_FUNC_F2(suffix, stype, dtype, ippFavor) \
+static void cvt##suffix( const stype* src, size_t sstep, const uchar*, size_t, \
+                         dtype* dst, size_t dstep, Size size, double*) \
+{ \
+    CV_IPP_RUN(src && dst, CV_INSTRUMENT_FUN_IPP(ippiConvert_##ippFavor, src, (int)sstep, dst, (int)dstep, ippiSize(size.width, size.height), ippRndFinancial, 0) >= 0) \
+    cvt_(src, sstep, dst, dstep, size); \
+}
+#else
+#define DEF_CVT_FUNC_F(suffix, stype, dtype, ippFavor) \
+static void cvt##suffix( const stype* src, size_t sstep, const uchar*, size_t, \
+                         dtype* dst, size_t dstep, Size size, double*) \
+{ \
+    cvt_(src, sstep, dst, dstep, size); \
+}
+#define DEF_CVT_FUNC_F2 DEF_CVT_FUNC_F
+#endif
+
+#define DEF_CVT_FUNC(suffix, stype, dtype) \
+static void cvt##suffix( const stype* src, size_t sstep, const uchar*, size_t, \
+                         dtype* dst, size_t dstep, Size size, double*) \
+{ \
+    cvt_(src, sstep, dst, dstep, size); \
+}
+
+#define DEF_CPY_FUNC(suffix, stype) \
+static void cvt##suffix( const stype* src, size_t sstep, const uchar*, size_t, \
+                         stype* dst, size_t dstep, Size size, double*) \
+{ \
+    cpy_(src, sstep, dst, dstep, size); \
+}
+
+DEF_CPY_FUNC(8u,     uchar)
+DEF_CVT_FUNC_F(8s8u,   schar, uchar, 8s8u_C1Rs)
+DEF_CVT_FUNC_F(16u8u,  ushort, uchar, 16u8u_C1R)
+DEF_CVT_FUNC_F(16s8u,  short, uchar, 16s8u_C1R)
+DEF_CVT_FUNC_F(32s8u,  int, uchar, 32s8u_C1R)
+DEF_CVT_FUNC_F2(32f8u,  float, uchar, 32f8u_C1RSfs)
+DEF_CVT_FUNC(64f8u,  double, uchar)
+
+DEF_CVT_FUNC_F2(8u8s,   uchar, schar, 8u8s_C1RSfs)
+DEF_CVT_FUNC_F2(16u8s,  ushort, schar, 16u8s_C1RSfs)
+DEF_CVT_FUNC_F2(16s8s,  short, schar, 16s8s_C1RSfs)
+DEF_CVT_FUNC_F(32s8s,  int, schar, 32s8s_C1R)
+DEF_CVT_FUNC_F2(32f8s,  float, schar, 32f8s_C1RSfs)
+DEF_CVT_FUNC(64f8s,  double, schar)
+
+DEF_CVT_FUNC_F(8u16u,  uchar, ushort, 8u16u_C1R)
+DEF_CVT_FUNC_F(8s16u,  schar, ushort, 8s16u_C1Rs)
+DEF_CPY_FUNC(16u,    ushort)
+DEF_CVT_FUNC_F(16s16u, short, ushort, 16s16u_C1Rs)
+DEF_CVT_FUNC_F2(32s16u, int, ushort, 32s16u_C1RSfs)
+DEF_CVT_FUNC_F2(32f16u, float, ushort, 32f16u_C1RSfs)
+DEF_CVT_FUNC(64f16u, double, ushort)
+
+DEF_CVT_FUNC_F(8u16s,  uchar, short, 8u16s_C1R)
+DEF_CVT_FUNC_F(8s16s,  schar, short, 8s16s_C1R)
+DEF_CVT_FUNC_F2(16u16s, ushort, short, 16u16s_C1RSfs)
+DEF_CVT_FUNC_F2(32s16s, int, short, 32s16s_C1RSfs)
+DEF_CVT_FUNC(32f16s, float, short)
+DEF_CVT_FUNC(64f16s, double, short)
+
+DEF_CVT_FUNC_F(8u32s,  uchar, int, 8u32s_C1R)
+DEF_CVT_FUNC_F(8s32s,  schar, int, 8s32s_C1R)
+DEF_CVT_FUNC_F(16u32s, ushort, int, 16u32s_C1R)
+DEF_CVT_FUNC_F(16s32s, short, int, 16s32s_C1R)
+DEF_CPY_FUNC(32s,    int)
+DEF_CVT_FUNC_F2(32f32s, float, int, 32f32s_C1RSfs)
+DEF_CVT_FUNC(64f32s, double, int)
+
+DEF_CVT_FUNC_F(8u32f,  uchar, float, 8u32f_C1R)
+DEF_CVT_FUNC_F(8s32f,  schar, float, 8s32f_C1R)
+DEF_CVT_FUNC_F(16u32f, ushort, float, 16u32f_C1R)
+DEF_CVT_FUNC_F(16s32f, short, float, 16s32f_C1R)
+DEF_CVT_FUNC_F(32s32f, int, float, 32s32f_C1R)
+DEF_CVT_FUNC(64f32f, double, float)
+
+DEF_CVT_FUNC(8u64f,  uchar, double)
+DEF_CVT_FUNC(8s64f,  schar, double)
+DEF_CVT_FUNC(16u64f, ushort, double)
+DEF_CVT_FUNC(16s64f, short, double)
+DEF_CVT_FUNC(32s64f, int, double)
+DEF_CVT_FUNC(32f64f, float, double)
+DEF_CPY_FUNC(64s,    int64)
+*/
+
+BinaryFunc getConvertFunc(int sdepth, int ddepth)
+{
+    CV_INSTRUMENT_REGION();
+    CV_CPU_DISPATCH(getConvertFunc, (sdepth, ddepth),
+        CV_CPU_DISPATCH_MODES_ALL);
+}
+
+static BinaryFunc get_cvt32f16f()
+{
+    CV_INSTRUMENT_REGION();
+    CV_CPU_DISPATCH(get_cvt32f16f, (),
+        CV_CPU_DISPATCH_MODES_ALL);
+}
+static BinaryFunc get_cvt16f32f()
+{
+    CV_INSTRUMENT_REGION();
+    CV_CPU_DISPATCH(get_cvt16f32f, (),
+        CV_CPU_DISPATCH_MODES_ALL);
+}
+
+#ifdef HAVE_OPENCL
+static bool ocl_convertFp16( InputArray _src, OutputArray _dst, int sdepth, int ddepth )
+{
+    int type = _src.type(), cn = CV_MAT_CN(type);
+
+    _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);
+}
+#endif
+
+void 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 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)get_cvt32f16f();
+        break;
+    case CV_16S:
+    //case CV_16F:
+        ddepth = CV_32F;
+        func = (BinaryFunc)get_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);
+    }
+}
+
+} // namespace cv

modules/core/src/convert.simd.hpp

@@ -3,10 +3,16 @@
 // of this distribution and at http://opencv.org/license.html
 
 #include "precomp.hpp"
-#include "opencl_kernels_core.hpp"
 #include "convert.hpp"
 
 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 {
 
@@ -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 )
 {
     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
 // 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 )
 {
     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) \
 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); }
+{ \
+    CV_INSTRUMENT_REGION(); \
+    cvtfunc<_Ts, _Td, _Twvec>(src, sstep, dst, dstep, size); \
+}
 
 ////////////////////// 8u -> ... ////////////////////////
 
@@ -225,16 +234,16 @@ DEF_CVT_FUNC(16f32f, cvt1_, float16_t, float,  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); }
+{ 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*)
-{ 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*)
-{ 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*)
-{ 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
@@ -380,148 +389,17 @@ BinaryFunc getConvertFunc(int sdepth, int ddepth)
     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
 
-} // 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);
-    }
-}

modules/core/src/convert_scale.dispatch.cpp

@@ -0,0 +1,259 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html
+
+
+#include "precomp.hpp"
+#include "opencl_kernels_core.hpp"
+
+#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
+{
+
+static BinaryFunc getCvtScaleAbsFunc(int depth)
+{
+    CV_INSTRUMENT_REGION();
+    CV_CPU_DISPATCH(getCvtScaleAbsFunc, (depth),
+        CV_CPU_DISPATCH_MODES_ALL);
+}
+
+BinaryFunc getConvertScaleFunc(int sdepth, int ddepth)
+{
+    CV_INSTRUMENT_REGION();
+    CV_CPU_DISPATCH(getConvertScaleFunc, (sdepth, ddepth),
+        CV_CPU_DISPATCH_MODES_ALL);
+}
+
+#ifdef HAVE_OPENCL
+
+static bool ocl_convertScaleAbs( InputArray _src, OutputArray _dst, double alpha, double beta )
+{
+    const ocl::Device & d = ocl::Device::getDefault();
+
+    int type = _src.type(), depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type);
+    bool doubleSupport = d.doubleFPConfig() > 0;
+    if (!doubleSupport && depth == CV_64F)
+        return false;
+
+    _dst.create(_src.size(), CV_8UC(cn));
+    int kercn = 1;
+    if (d.isIntel())
+    {
+        static const int vectorWidths[] = {4, 4, 4, 4, 4, 4, 4, -1};
+        kercn = ocl::checkOptimalVectorWidth( vectorWidths, _src, _dst,
+                                              noArray(), noArray(), noArray(),
+                                              noArray(), noArray(), noArray(),
+                                              noArray(), ocl::OCL_VECTOR_MAX);
+    }
+    else
+        kercn = ocl::predictOptimalVectorWidthMax(_src, _dst);
+
+    int rowsPerWI = d.isIntel() ? 4 : 1;
+    char cvt[2][50];
+    int wdepth = std::max(depth, CV_32F);
+    String build_opt = format("-D OP_CONVERT_SCALE_ABS -D UNARY_OP -D dstT=%s -D DEPTH_dst=%d -D srcT1=%s"
+                         " -D workT=%s -D wdepth=%d -D convertToWT1=%s -D convertToDT=%s"
+                         " -D workT1=%s -D rowsPerWI=%d%s",
+                         ocl::typeToStr(CV_8UC(kercn)), CV_8U,
+                         ocl::typeToStr(CV_MAKE_TYPE(depth, kercn)),
+                         ocl::typeToStr(CV_MAKE_TYPE(wdepth, kercn)), wdepth,
+                         ocl::convertTypeStr(depth, wdepth, kercn, cvt[0]),
+                         ocl::convertTypeStr(wdepth, CV_8U, kercn, cvt[1]),
+                         ocl::typeToStr(wdepth), rowsPerWI,
+                         doubleSupport ? " -D DOUBLE_SUPPORT" : "");
+    ocl::Kernel k("KF", ocl::core::arithm_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);
+
+    if (wdepth == CV_32F)
+        k.args(srcarg, dstarg, (float)alpha, (float)beta);
+    else if (wdepth == CV_64F)
+        k.args(srcarg, dstarg, alpha, beta);
+
+    size_t globalsize[2] = { (size_t)src.cols * cn / kercn, ((size_t)src.rows + rowsPerWI - 1) / rowsPerWI };
+    return k.run(2, globalsize, NULL, false);
+}
+
+#endif
+
+void convertScaleAbs(InputArray _src, OutputArray _dst, double alpha, double beta)
+{
+    CV_INSTRUMENT_REGION();
+
+    CV_OCL_RUN(_src.dims() <= 2 && _dst.isUMat(),
+               ocl_convertScaleAbs(_src, _dst, alpha, beta))
+
+    Mat src = _src.getMat();
+    int cn = src.channels();
+    double scale[] = {alpha, beta};
+    _dst.create( src.dims, src.size, CV_8UC(cn) );
+    Mat dst = _dst.getMat();
+    BinaryFunc func = getCvtScaleAbsFunc(src.depth());
+    CV_Assert( func != 0 );
+
+    if( src.dims <= 2 )
+    {
+        Size sz = getContinuousSize2D(src, dst, cn);
+        func( src.ptr(), src.step, 0, 0, dst.ptr(), 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], 0, 0, 0, ptrs[1], 0, sz, scale );
+    }
+}
+
+//==================================================================================================
+
+#ifdef HAVE_OPENCL
+
+static bool ocl_normalize( InputArray _src, InputOutputArray _dst, InputArray _mask, int dtype,
+                           double scale, double delta )
+{
+    UMat src = _src.getUMat();
+
+    if( _mask.empty() )
+        src.convertTo( _dst, dtype, scale, delta );
+    else if (src.channels() <= 4)
+    {
+        const ocl::Device & dev = ocl::Device::getDefault();
+
+        int stype = _src.type(), sdepth = CV_MAT_DEPTH(stype), cn = CV_MAT_CN(stype),
+                ddepth = CV_MAT_DEPTH(dtype), wdepth = std::max(CV_32F, std::max(sdepth, ddepth)),
+                rowsPerWI = dev.isIntel() ? 4 : 1;
+
+        float fscale = static_cast<float>(scale), fdelta = static_cast<float>(delta);
+        bool haveScale = std::fabs(scale - 1) > DBL_EPSILON,
+                haveZeroScale = !(std::fabs(scale) > DBL_EPSILON),
+                haveDelta = std::fabs(delta) > DBL_EPSILON,
+                doubleSupport = dev.doubleFPConfig() > 0;
+
+        if (!haveScale && !haveDelta && stype == dtype)
+        {
+            _src.copyTo(_dst, _mask);
+            return true;
+        }
+        if (haveZeroScale)
+        {
+            _dst.setTo(Scalar(delta), _mask);
+            return true;
+        }
+
+        if ((sdepth == CV_64F || ddepth == CV_64F) && !doubleSupport)
+            return false;
+
+        char cvt[2][40];
+        String opts = format("-D srcT=%s -D dstT=%s -D convertToWT=%s -D cn=%d -D rowsPerWI=%d"
+                             " -D convertToDT=%s -D workT=%s%s%s%s -D srcT1=%s -D dstT1=%s",
+                             ocl::typeToStr(stype), ocl::typeToStr(dtype),
+                             ocl::convertTypeStr(sdepth, wdepth, cn, cvt[0]), cn,
+                             rowsPerWI, ocl::convertTypeStr(wdepth, ddepth, cn, cvt[1]),
+                             ocl::typeToStr(CV_MAKE_TYPE(wdepth, cn)),
+                             doubleSupport ? " -D DOUBLE_SUPPORT" : "",
+                             haveScale ? " -D HAVE_SCALE" : "",
+                             haveDelta ? " -D HAVE_DELTA" : "",
+                             ocl::typeToStr(sdepth), ocl::typeToStr(ddepth));
+
+        ocl::Kernel k("normalizek", ocl::core::normalize_oclsrc, opts);
+        if (k.empty())
+            return false;
+
+        UMat mask = _mask.getUMat(), dst = _dst.getUMat();
+
+        ocl::KernelArg srcarg = ocl::KernelArg::ReadOnlyNoSize(src),
+                maskarg = ocl::KernelArg::ReadOnlyNoSize(mask),
+                dstarg = ocl::KernelArg::ReadWrite(dst);
+
+        if (haveScale)
+        {
+            if (haveDelta)
+                k.args(srcarg, maskarg, dstarg, fscale, fdelta);
+            else
+                k.args(srcarg, maskarg, dstarg, fscale);
+        }
+        else
+        {
+            if (haveDelta)
+                k.args(srcarg, maskarg, dstarg, fdelta);
+            else
+                k.args(srcarg, maskarg, dstarg);
+        }
+
+        size_t globalsize[2] = { (size_t)src.cols, ((size_t)src.rows + rowsPerWI - 1) / rowsPerWI };
+        return k.run(2, globalsize, NULL, false);
+    }
+    else
+    {
+        UMat temp;
+        src.convertTo( temp, dtype, scale, delta );
+        temp.copyTo( _dst, _mask );
+    }
+
+    return true;
+}
+
+#endif
+
+void normalize(InputArray _src, InputOutputArray _dst, double a, double b,
+               int norm_type, int rtype, InputArray _mask)
+{
+    CV_INSTRUMENT_REGION();
+
+    double scale = 1, shift = 0;
+    int type = _src.type(), depth = CV_MAT_DEPTH(type);
+
+    if( rtype < 0 )
+        rtype = _dst.fixedType() ? _dst.depth() : depth;
+
+    if( norm_type == CV_MINMAX )
+    {
+        double smin = 0, smax = 0;
+        double dmin = MIN( a, b ), dmax = MAX( a, b );
+        minMaxIdx( _src, &smin, &smax, 0, 0, _mask );
+        scale = (dmax - dmin)*(smax - smin > DBL_EPSILON ? 1./(smax - smin) : 0);
+        if( rtype == CV_32F )
+        {
+            scale = (float)scale;
+            shift = (float)dmin - (float)(smin*scale);
+        }
+        else
+            shift = dmin - smin*scale;
+    }
+    else if( norm_type == CV_L2 || norm_type == CV_L1 || norm_type == CV_C )
+    {
+        scale = norm( _src, norm_type, _mask );
+        scale = scale > DBL_EPSILON ? a/scale : 0.;
+        shift = 0;
+    }
+    else
+        CV_Error( CV_StsBadArg, "Unknown/unsupported norm type" );
+
+    CV_OCL_RUN(_dst.isUMat(),
+               ocl_normalize(_src, _dst, _mask, rtype, scale, shift))
+
+    Mat src = _src.getMat();
+    if( _mask.empty() )
+        src.convertTo( _dst, rtype, scale, shift );
+    else
+    {
+        Mat temp;
+        src.convertTo( temp, rtype, scale, shift );
+        temp.copyTo( _dst, _mask );
+    }
+}
+
+} // namespace

modules/core/src/convert_scale.simd.hpp

@@ -4,16 +4,20 @@
 
 
 #include "precomp.hpp"
-#include "opencl_kernels_core.hpp"
 #include "convert.hpp"
 
+namespace cv {
+CV_CPU_OPTIMIZATION_NAMESPACE_BEGIN
+
+BinaryFunc getCvtScaleAbsFunc(int depth);
+BinaryFunc getConvertScaleFunc(int sdepth, int ddepth);
+
+#ifndef CV_CPU_OPTIMIZATION_DECLARATIONS_ONLY
+
 /****************************************************************************************\
 *                                convertScale[Abs]                                       *
 \****************************************************************************************/
 
-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 )
@@ -287,7 +291,7 @@ 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)
+BinaryFunc getCvtScaleAbsFunc(int depth)
 {
     static BinaryFunc cvtScaleAbsTab[] =
     {
@@ -349,238 +353,7 @@ BinaryFunc getConvertScaleFunc(int sdepth, int ddepth)
     return cvtScaleTab[CV_MAT_DEPTH(ddepth)][CV_MAT_DEPTH(sdepth)];
 }
 
-#ifdef HAVE_OPENCL
-
-static bool ocl_convertScaleAbs( InputArray _src, OutputArray _dst, double alpha, double beta )
-{
-    const ocl::Device & d = ocl::Device::getDefault();
-
-    int type = _src.type(), depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type);
-    bool doubleSupport = d.doubleFPConfig() > 0;
-    if (!doubleSupport && depth == CV_64F)
-        return false;
-
-    _dst.create(_src.size(), CV_8UC(cn));
-    int kercn = 1;
-    if (d.isIntel())
-    {
-        static const int vectorWidths[] = {4, 4, 4, 4, 4, 4, 4, -1};
-        kercn = ocl::checkOptimalVectorWidth( vectorWidths, _src, _dst,
-                                              noArray(), noArray(), noArray(),
-                                              noArray(), noArray(), noArray(),
-                                              noArray(), ocl::OCL_VECTOR_MAX);
-    }
-    else
-        kercn = ocl::predictOptimalVectorWidthMax(_src, _dst);
-
-    int rowsPerWI = d.isIntel() ? 4 : 1;
-    char cvt[2][50];
-    int wdepth = std::max(depth, CV_32F);
-    String build_opt = format("-D OP_CONVERT_SCALE_ABS -D UNARY_OP -D dstT=%s -D DEPTH_dst=%d -D srcT1=%s"
-                         " -D workT=%s -D wdepth=%d -D convertToWT1=%s -D convertToDT=%s"
-                         " -D workT1=%s -D rowsPerWI=%d%s",
-                         ocl::typeToStr(CV_8UC(kercn)), CV_8U,
-                         ocl::typeToStr(CV_MAKE_TYPE(depth, kercn)),
-                         ocl::typeToStr(CV_MAKE_TYPE(wdepth, kercn)), wdepth,
-                         ocl::convertTypeStr(depth, wdepth, kercn, cvt[0]),
-                         ocl::convertTypeStr(wdepth, CV_8U, kercn, cvt[1]),
-                         ocl::typeToStr(wdepth), rowsPerWI,
-                         doubleSupport ? " -D DOUBLE_SUPPORT" : "");
-    ocl::Kernel k("KF", ocl::core::arithm_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);
-
-    if (wdepth == CV_32F)
-        k.args(srcarg, dstarg, (float)alpha, (float)beta);
-    else if (wdepth == CV_64F)
-        k.args(srcarg, dstarg, alpha, beta);
-
-    size_t globalsize[2] = { (size_t)src.cols * cn / kercn, ((size_t)src.rows + rowsPerWI - 1) / rowsPerWI };
-    return k.run(2, globalsize, NULL, false);
-}
-
 #endif
 
-} //cv::
+CV_CPU_OPTIMIZATION_NAMESPACE_END
-
+} // namespace
-
-void cv::convertScaleAbs( InputArray _src, OutputArray _dst, double alpha, double beta )
-{
-    CV_INSTRUMENT_REGION();
-
-    CV_OCL_RUN(_src.dims() <= 2 && _dst.isUMat(),
-               ocl_convertScaleAbs(_src, _dst, alpha, beta))
-
-    Mat src = _src.getMat();
-    int cn = src.channels();
-    double scale[] = {alpha, beta};
-    _dst.create( src.dims, src.size, CV_8UC(cn) );
-    Mat dst = _dst.getMat();
-    BinaryFunc func = getCvtScaleAbsFunc(src.depth());
-    CV_Assert( func != 0 );
-
-    if( src.dims <= 2 )
-    {
-        Size sz = getContinuousSize2D(src, dst, cn);
-        func( src.ptr(), src.step, 0, 0, dst.ptr(), 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], 0, 0, 0, ptrs[1], 0, sz, scale );
-    }
-}
-
-//==================================================================================================
-
-namespace cv {
-
-#ifdef HAVE_OPENCL
-
-static bool ocl_normalize( InputArray _src, InputOutputArray _dst, InputArray _mask, int dtype,
-                           double scale, double delta )
-{
-    UMat src = _src.getUMat();
-
-    if( _mask.empty() )
-        src.convertTo( _dst, dtype, scale, delta );
-    else if (src.channels() <= 4)
-    {
-        const ocl::Device & dev = ocl::Device::getDefault();
-
-        int stype = _src.type(), sdepth = CV_MAT_DEPTH(stype), cn = CV_MAT_CN(stype),
-                ddepth = CV_MAT_DEPTH(dtype), wdepth = std::max(CV_32F, std::max(sdepth, ddepth)),
-                rowsPerWI = dev.isIntel() ? 4 : 1;
-
-        float fscale = static_cast<float>(scale), fdelta = static_cast<float>(delta);
-        bool haveScale = std::fabs(scale - 1) > DBL_EPSILON,
-                haveZeroScale = !(std::fabs(scale) > DBL_EPSILON),
-                haveDelta = std::fabs(delta) > DBL_EPSILON,
-                doubleSupport = dev.doubleFPConfig() > 0;
-
-        if (!haveScale && !haveDelta && stype == dtype)
-        {
-            _src.copyTo(_dst, _mask);
-            return true;
-        }
-        if (haveZeroScale)
-        {
-            _dst.setTo(Scalar(delta), _mask);
-            return true;
-        }
-
-        if ((sdepth == CV_64F || ddepth == CV_64F) && !doubleSupport)
-            return false;
-
-        char cvt[2][40];
-        String opts = format("-D srcT=%s -D dstT=%s -D convertToWT=%s -D cn=%d -D rowsPerWI=%d"
-                             " -D convertToDT=%s -D workT=%s%s%s%s -D srcT1=%s -D dstT1=%s",
-                             ocl::typeToStr(stype), ocl::typeToStr(dtype),
-                             ocl::convertTypeStr(sdepth, wdepth, cn, cvt[0]), cn,
-                             rowsPerWI, ocl::convertTypeStr(wdepth, ddepth, cn, cvt[1]),
-                             ocl::typeToStr(CV_MAKE_TYPE(wdepth, cn)),
-                             doubleSupport ? " -D DOUBLE_SUPPORT" : "",
-                             haveScale ? " -D HAVE_SCALE" : "",
-                             haveDelta ? " -D HAVE_DELTA" : "",
-                             ocl::typeToStr(sdepth), ocl::typeToStr(ddepth));
-
-        ocl::Kernel k("normalizek", ocl::core::normalize_oclsrc, opts);
-        if (k.empty())
-            return false;
-
-        UMat mask = _mask.getUMat(), dst = _dst.getUMat();
-
-        ocl::KernelArg srcarg = ocl::KernelArg::ReadOnlyNoSize(src),
-                maskarg = ocl::KernelArg::ReadOnlyNoSize(mask),
-                dstarg = ocl::KernelArg::ReadWrite(dst);
-
-        if (haveScale)
-        {
-            if (haveDelta)
-                k.args(srcarg, maskarg, dstarg, fscale, fdelta);
-            else
-                k.args(srcarg, maskarg, dstarg, fscale);
-        }
-        else
-        {
-            if (haveDelta)
-                k.args(srcarg, maskarg, dstarg, fdelta);
-            else
-                k.args(srcarg, maskarg, dstarg);
-        }
-
-        size_t globalsize[2] = { (size_t)src.cols, ((size_t)src.rows + rowsPerWI - 1) / rowsPerWI };
-        return k.run(2, globalsize, NULL, false);
-    }
-    else
-    {
-        UMat temp;
-        src.convertTo( temp, dtype, scale, delta );
-        temp.copyTo( _dst, _mask );
-    }
-
-    return true;
-}
-
-#endif
-
-} // cv::
-
-void cv::normalize( InputArray _src, InputOutputArray _dst, double a, double b,
-                    int norm_type, int rtype, InputArray _mask )
-{
-    CV_INSTRUMENT_REGION();
-
-    double scale = 1, shift = 0;
-    int type = _src.type(), depth = CV_MAT_DEPTH(type);
-
-    if( rtype < 0 )
-        rtype = _dst.fixedType() ? _dst.depth() : depth;
-
-    if( norm_type == CV_MINMAX )
-    {
-        double smin = 0, smax = 0;
-        double dmin = MIN( a, b ), dmax = MAX( a, b );
-        minMaxIdx( _src, &smin, &smax, 0, 0, _mask );
-        scale = (dmax - dmin)*(smax - smin > DBL_EPSILON ? 1./(smax - smin) : 0);
-        if( rtype == CV_32F )
-        {
-            scale = (float)scale;
-            shift = (float)dmin - (float)(smin*scale);
-        }
-        else
-            shift = dmin - smin*scale;
-    }
-    else if( norm_type == CV_L2 || norm_type == CV_L1 || norm_type == CV_C )
-    {
-        scale = norm( _src, norm_type, _mask );
-        scale = scale > DBL_EPSILON ? a/scale : 0.;
-        shift = 0;
-    }
-    else
-        CV_Error( CV_StsBadArg, "Unknown/unsupported norm type" );
-
-    CV_OCL_RUN(_dst.isUMat(),
-               ocl_normalize(_src, _dst, _mask, rtype, scale, shift))
-
-    Mat src = _src.getMat();
-    if( _mask.empty() )
-        src.convertTo( _dst, rtype, scale, shift );
-    else
-    {
-        Mat temp;
-        src.convertTo( temp, rtype, scale, shift );
-        temp.copyTo( _dst, _mask );
-    }
-}