Merge pull request #7819 from alalek:fix_mulSpectrums_inplace_2.4

This commit is contained in:
Alexander Alekhin 2016-12-14 11:00:53 +00:00
commit af0ee306cf
4 changed files with 179 additions and 165 deletions

View File

@ -341,8 +341,13 @@ if(MSVC)
string(REPLACE "/W3" "/W4" CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE}")
string(REPLACE "/W3" "/W4" CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG}")
if(NOT ENABLE_NOISY_WARNINGS AND MSVC_VERSION EQUAL 1400)
ocv_warnings_disable(CMAKE_CXX_FLAGS /wd4510 /wd4610 /wd4312 /wd4201 /wd4244 /wd4328 /wd4267)
if(NOT ENABLE_NOISY_WARNINGS)
if(MSVC_VERSION EQUAL 1400)
ocv_warnings_disable(CMAKE_CXX_FLAGS /wd4510 /wd4610 /wd4312 /wd4201 /wd4244 /wd4328 /wd4267)
endif()
if(MSVC_VERSION LESS 1900) # MSVS2015
ocv_warnings_disable(CMAKE_CXX_FLAGS /wd4127) # warning C4127: conditional expression is constant
endif()
endif()
# allow extern "C" functions throw exceptions

View File

@ -1891,13 +1891,131 @@ void cv::idft( InputArray src, OutputArray dst, int flags, int nonzero_rows )
dft( src, dst, flags | DFT_INVERSE, nonzero_rows );
}
namespace {
#define VAL(buf, elem) (((T*)((char*)data ## buf + (step ## buf * (elem))))[0])
#define MUL_SPECTRUMS_COL(A, B, C) \
VAL(C, 0) = VAL(A, 0) * VAL(B, 0); \
for (size_t j = 1; j <= rows - 2; j += 2) \
{ \
double a_re = VAL(A, j), a_im = VAL(A, j + 1); \
double b_re = VAL(B, j), b_im = VAL(B, j + 1); \
if (conjB) b_im = -b_im; \
double c_re = a_re * b_re - a_im * b_im; \
double c_im = a_re * b_im + a_im * b_re; \
VAL(C, j) = (T)c_re; VAL(C, j + 1) = (T)c_im; \
} \
if ((rows & 1) == 0) \
VAL(C, rows-1) = VAL(A, rows-1) * VAL(B, rows-1)
template <typename T, bool conjB> static inline
void mulSpectrums_processCol_noinplace(const T* dataA, const T* dataB, T* dataC, size_t stepA, size_t stepB, size_t stepC, size_t rows)
{
MUL_SPECTRUMS_COL(A, B, C);
}
template <typename T, bool conjB> static inline
void mulSpectrums_processCol_inplaceA(const T* dataB, T* dataAC, size_t stepB, size_t stepAC, size_t rows)
{
MUL_SPECTRUMS_COL(AC, B, AC);
}
template <typename T, bool conjB, bool inplaceA> static inline
void mulSpectrums_processCol(const T* dataA, const T* dataB, T* dataC, size_t stepA, size_t stepB, size_t stepC, size_t rows)
{
if (inplaceA)
mulSpectrums_processCol_inplaceA<T, conjB>(dataB, dataC, stepB, stepC, rows);
else
mulSpectrums_processCol_noinplace<T, conjB>(dataA, dataB, dataC, stepA, stepB, stepC, rows);
}
#undef MUL_SPECTRUMS_COL
#undef VAL
template <typename T, bool conjB, bool inplaceA> static inline
void mulSpectrums_processCols(const T* dataA, const T* dataB, T* dataC, size_t stepA, size_t stepB, size_t stepC, size_t rows, size_t cols)
{
mulSpectrums_processCol<T, conjB, inplaceA>(dataA, dataB, dataC, stepA, stepB, stepC, rows);
if ((cols & 1) == 0)
{
mulSpectrums_processCol<T, conjB, inplaceA>(dataA + cols - 1, dataB + cols - 1, dataC + cols - 1, stepA, stepB, stepC, rows);
}
}
#define VAL(buf, elem) (data ## buf[(elem)])
#define MUL_SPECTRUMS_ROW(A, B, C) \
for (size_t j = j0; j < j1; j += 2) \
{ \
double a_re = VAL(A, j), a_im = VAL(A, j + 1); \
double b_re = VAL(B, j), b_im = VAL(B, j + 1); \
if (conjB) b_im = -b_im; \
double c_re = a_re * b_re - a_im * b_im; \
double c_im = a_re * b_im + a_im * b_re; \
VAL(C, j) = (T)c_re; VAL(C, j + 1) = (T)c_im; \
}
template <typename T, bool conjB> static inline
void mulSpectrums_processRow_noinplace(const T* dataA, const T* dataB, T* dataC, size_t j0, size_t j1)
{
MUL_SPECTRUMS_ROW(A, B, C);
}
template <typename T, bool conjB> static inline
void mulSpectrums_processRow_inplaceA(const T* dataB, T* dataAC, size_t j0, size_t j1)
{
MUL_SPECTRUMS_ROW(AC, B, AC);
}
template <typename T, bool conjB, bool inplaceA> static inline
void mulSpectrums_processRow(const T* dataA, const T* dataB, T* dataC, size_t j0, size_t j1)
{
if (inplaceA)
mulSpectrums_processRow_inplaceA<T, conjB>(dataB, dataC, j0, j1);
else
mulSpectrums_processRow_noinplace<T, conjB>(dataA, dataB, dataC, j0, j1);
}
#undef MUL_SPECTRUMS_ROW
#undef VAL
template <typename T, bool conjB, bool inplaceA> static inline
void mulSpectrums_processRows(const T* dataA, const T* dataB, T* dataC, size_t stepA, size_t stepB, size_t stepC, size_t rows, size_t cols, size_t j0, size_t j1, bool is_1d_CN1)
{
while (rows-- > 0)
{
if (is_1d_CN1)
dataC[0] = dataA[0]*dataB[0];
mulSpectrums_processRow<T, conjB, inplaceA>(dataA, dataB, dataC, j0, j1);
if (is_1d_CN1 && (cols & 1) == 0)
dataC[j1] = dataA[j1]*dataB[j1];
dataA = (const T*)(((char*)dataA) + stepA);
dataB = (const T*)(((char*)dataB) + stepB);
dataC = (T*)(((char*)dataC) + stepC);
}
}
template <typename T, bool conjB, bool inplaceA> static inline
void mulSpectrums_Impl_(const T* dataA, const T* dataB, T* dataC, size_t stepA, size_t stepB, size_t stepC, size_t rows, size_t cols, size_t j0, size_t j1, bool is_1d, bool isCN1)
{
if (!is_1d && isCN1)
{
mulSpectrums_processCols<T, conjB, inplaceA>(dataA, dataB, dataC, stepA, stepB, stepC, rows, cols);
}
mulSpectrums_processRows<T, conjB, inplaceA>(dataA, dataB, dataC, stepA, stepB, stepC, rows, cols, j0, j1, is_1d && isCN1);
}
template <typename T, bool conjB> static inline
void mulSpectrums_Impl(const T* dataA, const T* dataB, T* dataC, size_t stepA, size_t stepB, size_t stepC, size_t rows, size_t cols, size_t j0, size_t j1, bool is_1d, bool isCN1)
{
if (dataA == dataC)
mulSpectrums_Impl_<T, conjB, true>(dataA, dataB, dataC, stepA, stepB, stepC, rows, cols, j0, j1, is_1d, isCN1);
else
mulSpectrums_Impl_<T, conjB, false>(dataA, dataB, dataC, stepA, stepB, stepC, rows, cols, j0, j1, is_1d, isCN1);
}
} // namespace
void cv::mulSpectrums( InputArray _srcA, InputArray _srcB,
OutputArray _dst, int flags, bool conjB )
{
Mat srcA = _srcA.getMat(), srcB = _srcB.getMat();
int depth = srcA.depth(), cn = srcA.channels(), type = srcA.type();
int rows = srcA.rows, cols = srcA.cols;
int j, k;
size_t rows = srcA.rows, cols = srcA.cols;
CV_Assert( type == srcB.type() && srcA.size() == srcB.size() );
CV_Assert( type == CV_32FC1 || type == CV_32FC2 || type == CV_64FC1 || type == CV_64FC2 );
@ -1905,149 +2023,42 @@ void cv::mulSpectrums( InputArray _srcA, InputArray _srcB,
_dst.create( srcA.rows, srcA.cols, type );
Mat dst = _dst.getMat();
bool is_1d = (flags & DFT_ROWS) || (rows == 1 || (cols == 1 &&
srcA.isContinuous() && srcB.isContinuous() && dst.isContinuous()));
// correct inplace support
// Case 'dst.data == srcA.data' is handled by implementation,
// because it is used frequently (filter2D, matchTemplate)
if (dst.data == srcB.data)
srcB = srcB.clone(); // workaround for B only
bool is_1d = (flags & DFT_ROWS)
|| (rows == 1)
|| (cols == 1 && srcA.isContinuous() && srcB.isContinuous() && dst.isContinuous());
if( is_1d && !(flags & DFT_ROWS) )
cols = cols + rows - 1, rows = 1;
int ncols = cols*cn;
int j0 = cn == 1;
int j1 = ncols - (cols % 2 == 0 && cn == 1);
bool isCN1 = cn == 1;
size_t j0 = isCN1 ? 1 : 0;
size_t j1 = cols*cn - (((cols & 1) == 0 && cn == 1) ? 1 : 0);
if( depth == CV_32F )
if (depth == CV_32F)
{
const float* dataA = (const float*)srcA.data;
const float* dataB = (const float*)srcB.data;
float* dataC = (float*)dst.data;
size_t stepA = srcA.step/sizeof(dataA[0]);
size_t stepB = srcB.step/sizeof(dataB[0]);
size_t stepC = dst.step/sizeof(dataC[0]);
if( !is_1d && cn == 1 )
{
for( k = 0; k < (cols % 2 ? 1 : 2); k++ )
{
if( k == 1 )
dataA += cols - 1, dataB += cols - 1, dataC += cols - 1;
dataC[0] = dataA[0]*dataB[0];
if( rows % 2 == 0 )
dataC[(rows-1)*stepC] = dataA[(rows-1)*stepA]*dataB[(rows-1)*stepB];
if( !conjB )
for( j = 1; j <= rows - 2; j += 2 )
{
double re = (double)dataA[j*stepA]*dataB[j*stepB] -
(double)dataA[(j+1)*stepA]*dataB[(j+1)*stepB];
double im = (double)dataA[j*stepA]*dataB[(j+1)*stepB] +
(double)dataA[(j+1)*stepA]*dataB[j*stepB];
dataC[j*stepC] = (float)re; dataC[(j+1)*stepC] = (float)im;
}
else
for( j = 1; j <= rows - 2; j += 2 )
{
double re = (double)dataA[j*stepA]*dataB[j*stepB] +
(double)dataA[(j+1)*stepA]*dataB[(j+1)*stepB];
double im = (double)dataA[(j+1)*stepA]*dataB[j*stepB] -
(double)dataA[j*stepA]*dataB[(j+1)*stepB];
dataC[j*stepC] = (float)re; dataC[(j+1)*stepC] = (float)im;
}
if( k == 1 )
dataA -= cols - 1, dataB -= cols - 1, dataC -= cols - 1;
}
}
for( ; rows--; dataA += stepA, dataB += stepB, dataC += stepC )
{
if( is_1d && cn == 1 )
{
dataC[0] = dataA[0]*dataB[0];
if( cols % 2 == 0 )
dataC[j1] = dataA[j1]*dataB[j1];
}
if( !conjB )
for( j = j0; j < j1; j += 2 )
{
double re = (double)dataA[j]*dataB[j] - (double)dataA[j+1]*dataB[j+1];
double im = (double)dataA[j+1]*dataB[j] + (double)dataA[j]*dataB[j+1];
dataC[j] = (float)re; dataC[j+1] = (float)im;
}
else
for( j = j0; j < j1; j += 2 )
{
double re = (double)dataA[j]*dataB[j] + (double)dataA[j+1]*dataB[j+1];
double im = (double)dataA[j+1]*dataB[j] - (double)dataA[j]*dataB[j+1];
dataC[j] = (float)re; dataC[j+1] = (float)im;
}
}
if (!conjB)
mulSpectrums_Impl<float, false>(dataA, dataB, dataC, srcA.step, srcB.step, dst.step, rows, cols, j0, j1, is_1d, isCN1);
else
mulSpectrums_Impl<float, true>(dataA, dataB, dataC, srcA.step, srcB.step, dst.step, rows, cols, j0, j1, is_1d, isCN1);
}
else
{
const double* dataA = (const double*)srcA.data;
const double* dataB = (const double*)srcB.data;
double* dataC = (double*)dst.data;
size_t stepA = srcA.step/sizeof(dataA[0]);
size_t stepB = srcB.step/sizeof(dataB[0]);
size_t stepC = dst.step/sizeof(dataC[0]);
if( !is_1d && cn == 1 )
{
for( k = 0; k < (cols % 2 ? 1 : 2); k++ )
{
if( k == 1 )
dataA += cols - 1, dataB += cols - 1, dataC += cols - 1;
dataC[0] = dataA[0]*dataB[0];
if( rows % 2 == 0 )
dataC[(rows-1)*stepC] = dataA[(rows-1)*stepA]*dataB[(rows-1)*stepB];
if( !conjB )
for( j = 1; j <= rows - 2; j += 2 )
{
double re = dataA[j*stepA]*dataB[j*stepB] -
dataA[(j+1)*stepA]*dataB[(j+1)*stepB];
double im = dataA[j*stepA]*dataB[(j+1)*stepB] +
dataA[(j+1)*stepA]*dataB[j*stepB];
dataC[j*stepC] = re; dataC[(j+1)*stepC] = im;
}
else
for( j = 1; j <= rows - 2; j += 2 )
{
double re = dataA[j*stepA]*dataB[j*stepB] +
dataA[(j+1)*stepA]*dataB[(j+1)*stepB];
double im = dataA[(j+1)*stepA]*dataB[j*stepB] -
dataA[j*stepA]*dataB[(j+1)*stepB];
dataC[j*stepC] = re; dataC[(j+1)*stepC] = im;
}
if( k == 1 )
dataA -= cols - 1, dataB -= cols - 1, dataC -= cols - 1;
}
}
for( ; rows--; dataA += stepA, dataB += stepB, dataC += stepC )
{
if( is_1d && cn == 1 )
{
dataC[0] = dataA[0]*dataB[0];
if( cols % 2 == 0 )
dataC[j1] = dataA[j1]*dataB[j1];
}
if( !conjB )
for( j = j0; j < j1; j += 2 )
{
double re = dataA[j]*dataB[j] - dataA[j+1]*dataB[j+1];
double im = dataA[j+1]*dataB[j] + dataA[j]*dataB[j+1];
dataC[j] = re; dataC[j+1] = im;
}
else
for( j = j0; j < j1; j += 2 )
{
double re = dataA[j]*dataB[j] + dataA[j+1]*dataB[j+1];
double im = dataA[j+1]*dataB[j] - dataA[j]*dataB[j+1];
dataC[j] = re; dataC[j+1] = im;
}
}
if (!conjB)
mulSpectrums_Impl<double, false>(dataA, dataB, dataC, srcA.step, srcB.step, dst.step, rows, cols, j0, j1, is_1d, isCN1);
else
mulSpectrums_Impl<double, true>(dataA, dataB, dataC, srcA.step, srcB.step, dst.step, rows, cols, j0, j1, is_1d, isCN1);
}
}

View File

@ -419,9 +419,6 @@ static void fixCCS( Mat& mat, int cols, int flags )
}
}
#if defined _MSC_VER && _MSC_VER >= 1700
#pragma optimize("", off)
#endif
static void mulComplex( const Mat& src1, const Mat& src2, Mat& dst, int flags )
{
dst.create(src1.rows, src1.cols, src1.type());
@ -430,12 +427,27 @@ static void mulComplex( const Mat& src1, const Mat& src2, Mat& dst, int flags )
CV_Assert( src1.size == src2.size && src1.type() == src2.type() &&
(src1.type() == CV_32FC2 || src1.type() == CV_64FC2) );
const Mat* src1_ = &src1;
Mat src1_tmp;
if (dst.data == src1.data)
{
src1_tmp = src1.clone();
src1_ = &src1_tmp;
}
const Mat* src2_ = &src2;
Mat src2_tmp;
if (dst.data == src2.data)
{
src2_tmp = src2.clone();
src2_ = &src2_tmp;
}
for( i = 0; i < dst.rows; i++ )
{
if( depth == CV_32F )
{
const float* a = src1.ptr<float>(i);
const float* b = src2.ptr<float>(i);
const float* a = src1_->ptr<float>(i);
const float* b = src2_->ptr<float>(i);
float* c = dst.ptr<float>(i);
if( !(flags & CV_DXT_MUL_CONJ) )
@ -459,8 +471,8 @@ static void mulComplex( const Mat& src1, const Mat& src2, Mat& dst, int flags )
}
else
{
const double* a = src1.ptr<double>(i);
const double* b = src2.ptr<double>(i);
const double* a = src1_->ptr<double>(i);
const double* b = src2_->ptr<double>(i);
double* c = dst.ptr<double>(i);
if( !(flags & CV_DXT_MUL_CONJ) )
@ -484,9 +496,6 @@ static void mulComplex( const Mat& src1, const Mat& src2, Mat& dst, int flags )
}
}
}
#if defined _MSC_VER && _MSC_VER >= 1700
#pragma optimize("", on)
#endif
}
@ -778,9 +787,7 @@ public:
protected:
void run_func();
void prepare_to_validation( int test_case_idx );
#if defined(__aarch64__) && defined(NDEBUG)
double get_success_error_level( int test_case_idx, int i, int j );
#endif
};
@ -788,31 +795,19 @@ CxCore_MulSpectrumsTest::CxCore_MulSpectrumsTest() : CxCore_DXTBaseTest( true, t
{
}
#if defined(__aarch64__) && defined(NDEBUG)
double CxCore_MulSpectrumsTest::get_success_error_level( int test_case_idx, int i, int j )
{
(void)test_case_idx;
CV_Assert(i == OUTPUT);
CV_Assert(j == 0);
int elem_depth = CV_MAT_DEPTH(cvGetElemType(test_array[i][j]));
if( elem_depth <= CV_32F )
{
return ArrayTest::get_success_error_level( test_case_idx, i, j );
}
switch( test_case_idx )
{
// Usual threshold is too strict for these test cases due to the difference of fmsub and fsub
case 399:
case 420:
return DBL_EPSILON * 20000;
case 65:
case 161:
case 287:
case 351:
case 458:
return DBL_EPSILON * 10000;
default:
return ArrayTest::get_success_error_level( test_case_idx, i, j );
}
CV_Assert(elem_depth == CV_32F || elem_depth == CV_64F);
element_wise_relative_error = false;
double maxInputValue = 1000; // ArrayTest::get_minmax_bounds
double err = 8 * maxInputValue; // result = A*B + C*D
return (elem_depth == CV_32F ? FLT_EPSILON : DBL_EPSILON) * err;
}
#endif
void CxCore_MulSpectrumsTest::run_func()
{

View File

@ -167,6 +167,9 @@ static void divSpectrums( InputArray _srcA, InputArray _srcB, OutputArray _dst,
_dst.create( srcA.rows, srcA.cols, type );
Mat dst = _dst.getMat();
CV_Assert(dst.data != srcA.data); // non-inplace check
CV_Assert(dst.data != srcB.data); // non-inplace check
bool is_1d = (flags & DFT_ROWS) || (rows == 1 || (cols == 1 &&
srcA.isContinuous() && srcB.isContinuous() && dst.isContinuous()));