Attachment #757507 for bug #800539

Lines 51-56 macro(ocv_lapack_check) Link Here

(-)a/cmake/OpenCVFindLAPACK.cmake (-1 / +19 lines)
51	if(NOT "${OPENCV_CBLAS_H_PATH_${_lapack_impl}}" STREQUAL "${OPENCV_LAPACKE_H_PATH_${_lapack_impl}}")	51	if(NOT "${OPENCV_CBLAS_H_PATH_${_lapack_impl}}" STREQUAL "${OPENCV_LAPACKE_H_PATH_${_lapack_impl}}")
52	list(APPEND _lapack_content "#include \"${OPENCV_LAPACKE_H_PATH_${_lapack_impl}}\"")	52	list(APPEND _lapack_content "#include \"${OPENCV_LAPACKE_H_PATH_${_lapack_impl}}\"")
53	endif()	53	endif()
		54	list(APPEND _lapack_content "
		55	#ifdef OPENCV_USE_LAPACK_PREFIX
		56	#define LAPACK_FUNC(f) LAPACK_##f
		57	#else
		58	#define LAPACK_FUNC(f) f##_
		59	#endif
		60	")
54	if(${_lapack_add_extern_c})	61	if(${_lapack_add_extern_c})
55	list(APPEND _lapack_content "}")	62	list(APPEND _lapack_content "}")
56	endif()	63	endif()
Lines 156-163 if(WITH_LAPACK) Link Here
156	if(WIN32 AND NOT OPENCV_LAPACK_SHARED_LIBS)	163	if(WIN32 AND NOT OPENCV_LAPACK_SHARED_LIBS)
157	set(BLA_STATIC 1)	164	set(BLA_STATIC 1)
158	endif()	165	endif()
159	find_package(LAPACK)	166	find_package(LAPACK CONFIG)
160	if(LAPACK_FOUND)	167	if(LAPACK_FOUND)
		168	set(USE_LAPACK_PREFIX_MIN_VERSION "3.9.1")
		169	if(LAPACK_VERSION VERSION_GREATER_EQUAL USE_LAPACK_PREFIX_MIN_VERSION)
		170	message(STATUS "Setting OPENCV_USE_LAPACK_PREFIX to ON as LAPACK_VERSION=${LAPACK_VERSION} >= USE_LAPACK_PREFIX_MIN_VERSION=${USE_LAPACK_PREFIX_MIN_VERSION}")
		171	set(OPENCV_USE_LAPACK_PREFIX ON)
		172	else()
		173	message(STATUS "Setting OPENCV_USE_LAPACK_PREFIX to OFF as LAPACK_VERSION=${LAPACK_VERSION} < USE_LAPACK_PREFIX_MIN_VERSION=${USE_LAPACK_PREFIX_MIN_VERSION}")
		174	set(OPENCV_USE_LAPACK_PREFIX OFF)
		175	endif()
161	if(NOT DEFINED LAPACKE_INCLUDE_DIR)	176	if(NOT DEFINED LAPACKE_INCLUDE_DIR)
162	find_path(LAPACKE_INCLUDE_DIR "lapacke.h")	177	find_path(LAPACKE_INCLUDE_DIR "lapacke.h")
163	endif()	178	endif()
Lines 210-213 if(WITH_LAPACK) Link Here
210	set(LAPACK_LAPACKE_H ${LAPACK_LAPACKE_H} CACHE STRING "Alternative name of lapacke.h" FORCE)	225	set(LAPACK_LAPACKE_H ${LAPACK_LAPACKE_H} CACHE STRING "Alternative name of lapacke.h" FORCE)
211	set(LAPACK_LIBRARIES ${LAPACK_LIBRARIES} CACHE STRING "Names of BLAS & LAPACK binaries (.so, .dll, .a, .lib)" FORCE)	226	set(LAPACK_LIBRARIES ${LAPACK_LIBRARIES} CACHE STRING "Names of BLAS & LAPACK binaries (.so, .dll, .a, .lib)" FORCE)
212	set(LAPACK_IMPL ${LAPACK_IMPL} CACHE STRING "Lapack implementation id" FORCE)	227	set(LAPACK_IMPL ${LAPACK_IMPL} CACHE STRING "Lapack implementation id" FORCE)
		228	if(OPENCV_USE_LAPACK_PREFIX)
		229	add_definitions(-DOPENCV_USE_LAPACK_PREFIX)
		230	endif()
213	endif()	231	endif()




        if(n == 1 && b_step == sizeof(fptype))
        {
            if(typeid(fptype) == typeid(float))
                LAPACK_FUNC(sposv)(L, &m, &n, (float*)a, &lda, (float*)b, &m, &lapackStatus);
            else if(typeid(fptype) == typeid(double))
                LAPACK_FUNC(dposv)(L, &m, &n, (double*)a, &lda, (double*)b, &m, &lapackStatus);
        }
        else
        {

            transpose(b, ldb, tmpB, m, m, n);

            if(typeid(fptype) == typeid(float))
                LAPACK_FUNC(sposv)(L, &m, &n, (float*)a, &lda, (float*)tmpB, &m, &lapackStatus);
            else if(typeid(fptype) == typeid(double))
                LAPACK_FUNC(dposv)(L, &m, &n, (double*)a, &lda, (double*)tmpB, &m, &lapackStatus);

            transpose(tmpB, m, b, ldb, n, m);
            delete[] tmpB;

    else
    {
        if(typeid(fptype) == typeid(float))
            LAPACK_FUNC(spotrf)(L, &m, (float*)a, &lda, &lapackStatus);
        else if(typeid(fptype) == typeid(double))
            LAPACK_FUNC(dpotrf)(L, &m, (double*)a, &lda, &lapackStatus);
    }

    if(lapackStatus == 0) *info = true;

    }

    if(typeid(fptype) == typeid(float))
        LAPACK_FUNC(sgesdd)(mode, &m, &n, (float*)a, &lda, (float*)w, (float*)u, &ldu, (float*)vt, &ldv, (float*)&work1, &lwork, iworkBuf, info);
    else if(typeid(fptype) == typeid(double))
        LAPACK_FUNC(dgesdd)(mode, &m, &n, (double*)a, &lda, (double*)w, (double*)u, &ldu, (double*)vt, &ldv, (double*)&work1, &lwork, iworkBuf, info);

    lwork = (int)round(work1); //optimal buffer size
    fptype* buffer = new fptype[lwork + 1];

    if(typeid(fptype) == typeid(float))
        LAPACK_FUNC(sgesdd)(mode, &m, &n, (float*)a, &lda, (float*)w, (float*)u, &ldu, (float*)vt, &ldv, (float*)buffer, &lwork, iworkBuf, info);
    else if(typeid(fptype) == typeid(double))
        LAPACK_FUNC(dgesdd)(mode, &m, &n, (double*)a, &lda, (double*)w, (double*)u, &ldu, (double*)vt, &ldv, (double*)buffer, &lwork, iworkBuf, info);

    if(!(flags & CV_HAL_SVD_NO_UV))
        transpose_square_inplace(vt, ldv, n);

        if (k == 1 && b_step == sizeof(fptype))
        {
            if (typeid(fptype) == typeid(float))
                LAPACK_FUNC(sgels)(mode, &m, &n, &k, (float*)tmpA, &ldtmpA, (float*)b, &m, (float*)&work1, &lwork, info);
            else if (typeid(fptype) == typeid(double))
                LAPACK_FUNC(dgels)(mode, &m, &n, &k, (double*)tmpA, &ldtmpA, (double*)b, &m, (double*)&work1, &lwork, info);

            lwork = cvRound(work1); //optimal buffer size
            std::vector<fptype> workBufMemHolder(lwork + 1);
            fptype* buffer = &workBufMemHolder.front();

            if (typeid(fptype) == typeid(float))
                LAPACK_FUNC(sgels)(mode, &m, &n, &k, (float*)tmpA, &ldtmpA, (float*)b, &m, (float*)buffer, &lwork, info);
            else if (typeid(fptype) == typeid(double))
                LAPACK_FUNC(dgels)(mode, &m, &n, &k, (double*)tmpA, &ldtmpA, (double*)b, &m, (double*)buffer, &lwork, info);
        }
        else
        {

            transpose(b, ldb, tmpB, m, m, k);

            if (typeid(fptype) == typeid(float))
                LAPACK_FUNC(sgels)(mode, &m, &n, &k, (float*)tmpA, &ldtmpA, (float*)tmpB, &m, (float*)&work1, &lwork, info);
            else if (typeid(fptype) == typeid(double))
                LAPACK_FUNC(dgels)(mode, &m, &n, &k, (double*)tmpA, &ldtmpA, (double*)tmpB, &m, (double*)&work1, &lwork, info);

            lwork = cvRound(work1); //optimal buffer size
            std::vector<fptype> workBufMemHolder(lwork + 1);
            fptype* buffer = &workBufMemHolder.front();

            if (typeid(fptype) == typeid(float))
                LAPACK_FUNC(sgels)(mode, &m, &n, &k, (float*)tmpA, &ldtmpA, (float*)tmpB, &m, (float*)buffer, &lwork, info);
            else if (typeid(fptype) == typeid(double))
                LAPACK_FUNC(dgels)(mode, &m, &n, &k, (double*)tmpA, &ldtmpA, (double*)tmpB, &m, (double*)buffer, &lwork, info);

            transpose(tmpB, m, b, ldb, k, m);
        }

Return to bug 800539

Lines 163-171 lapack_Cholesky(fptype* a, size_t a_step, int m, fptype* b, size_t b_step, int n Link Here

(-)a/modules/core/src/hal_internal.cpp (-18 / +18 lines)
163	if(n == 1 && b_step == sizeof(fptype))	163	if(n == 1 && b_step == sizeof(fptype))
164	{	164	{
165	if(typeid(fptype) == typeid(float))	165	if(typeid(fptype) == typeid(float))
166	sposv_(L, &m, &n, (float)a, &lda, (float)b, &m, &lapackStatus);	166	LAPACK_FUNC(sposv)(L, &m, &n, (float)a, &lda, (float)b, &m, &lapackStatus);
167	else if(typeid(fptype) == typeid(double))	167	else if(typeid(fptype) == typeid(double))
168	dposv_(L, &m, &n, (double)a, &lda, (double)b, &m, &lapackStatus);	168	LAPACK_FUNC(dposv)(L, &m, &n, (double)a, &lda, (double)b, &m, &lapackStatus);
169	}	169	}
170	else	170	else
171	{	171	{
Lines 174-182 lapack_Cholesky(fptype* a, size_t a_step, int m, fptype* b, size_t b_step, int n Link Here
174	transpose(b, ldb, tmpB, m, m, n);	174	transpose(b, ldb, tmpB, m, m, n);
175		175
176	if(typeid(fptype) == typeid(float))	176	if(typeid(fptype) == typeid(float))
177	sposv_(L, &m, &n, (float)a, &lda, (float)tmpB, &m, &lapackStatus);	177	LAPACK_FUNC(sposv)(L, &m, &n, (float)a, &lda, (float)tmpB, &m, &lapackStatus);
178	else if(typeid(fptype) == typeid(double))	178	else if(typeid(fptype) == typeid(double))
179	dposv_(L, &m, &n, (double)a, &lda, (double)tmpB, &m, &lapackStatus);	179	LAPACK_FUNC(dposv)(L, &m, &n, (double)a, &lda, (double)tmpB, &m, &lapackStatus);
180		180
181	transpose(tmpB, m, b, ldb, n, m);	181	transpose(tmpB, m, b, ldb, n, m);
182	delete[] tmpB;	182	delete[] tmpB;
Lines 185-193 lapack_Cholesky(fptype* a, size_t a_step, int m, fptype* b, size_t b_step, int n Link Here
185	else	185	else
186	{	186	{
187	if(typeid(fptype) == typeid(float))	187	if(typeid(fptype) == typeid(float))
188	spotrf_(L, &m, (float*)a, &lda, &lapackStatus);	188	LAPACK_FUNC(spotrf)(L, &m, (float*)a, &lda, &lapackStatus);
189	else if(typeid(fptype) == typeid(double))	189	else if(typeid(fptype) == typeid(double))
190	dpotrf_(L, &m, (double*)a, &lda, &lapackStatus);	190	LAPACK_FUNC(dpotrf)(L, &m, (double*)a, &lda, &lapackStatus);
191	}	191	}
192		192
193	if(lapackStatus == 0) *info = true;	193	if(lapackStatus == 0) *info = true;
Lines 227-243 lapack_SVD(fptype* a, size_t a_step, fptype w, fptype u, size_t u_step, fptype Link Here
227	}	227	}
228		228
229	if(typeid(fptype) == typeid(float))	229	if(typeid(fptype) == typeid(float))
230	sgesdd_(mode, &m, &n, (float)a, &lda, (float)w, (float)u, &ldu, (float)vt, &ldv, (float*)&work1, &lwork, iworkBuf, info);	230	LAPACK_FUNC(sgesdd)(mode, &m, &n, (float)a, &lda, (float)w, (float)u, &ldu, (float)vt, &ldv, (float*)&work1, &lwork, iworkBuf, info);
231	else if(typeid(fptype) == typeid(double))	231	else if(typeid(fptype) == typeid(double))
232	dgesdd_(mode, &m, &n, (double)a, &lda, (double)w, (double)u, &ldu, (double)vt, &ldv, (double*)&work1, &lwork, iworkBuf, info);	232	LAPACK_FUNC(dgesdd)(mode, &m, &n, (double)a, &lda, (double)w, (double)u, &ldu, (double)vt, &ldv, (double*)&work1, &lwork, iworkBuf, info);
233		233
234	lwork = (int)round(work1); //optimal buffer size	234	lwork = (int)round(work1); //optimal buffer size
235	fptype* buffer = new fptype[lwork + 1];	235	fptype* buffer = new fptype[lwork + 1];
236		236
237	if(typeid(fptype) == typeid(float))	237	if(typeid(fptype) == typeid(float))
238	sgesdd_(mode, &m, &n, (float)a, &lda, (float)w, (float)u, &ldu, (float)vt, &ldv, (float*)buffer, &lwork, iworkBuf, info);	238	LAPACK_FUNC(sgesdd)(mode, &m, &n, (float)a, &lda, (float)w, (float)u, &ldu, (float)vt, &ldv, (float*)buffer, &lwork, iworkBuf, info);
239	else if(typeid(fptype) == typeid(double))	239	else if(typeid(fptype) == typeid(double))
240	dgesdd_(mode, &m, &n, (double)a, &lda, (double)w, (double)u, &ldu, (double)vt, &ldv, (double*)buffer, &lwork, iworkBuf, info);	240	LAPACK_FUNC(dgesdd)(mode, &m, &n, (double)a, &lda, (double)w, (double)u, &ldu, (double)vt, &ldv, (double*)buffer, &lwork, iworkBuf, info);
241		241
242	if(!(flags & CV_HAL_SVD_NO_UV))	242	if(!(flags & CV_HAL_SVD_NO_UV))
243	transpose_square_inplace(vt, ldv, n);	243	transpose_square_inplace(vt, ldv, n);
Lines 288-305 lapack_QR(fptype* a, size_t a_step, int m, int n, int k, fptype* b, size_t b_ste Link Here
288	if (k == 1 && b_step == sizeof(fptype))	288	if (k == 1 && b_step == sizeof(fptype))
289	{	289	{
290	if (typeid(fptype) == typeid(float))	290	if (typeid(fptype) == typeid(float))
291	sgels_(mode, &m, &n, &k, (float)tmpA, &ldtmpA, (float)b, &m, (float*)&work1, &lwork, info);	291	LAPACK_FUNC(sgels)(mode, &m, &n, &k, (float)tmpA, &ldtmpA, (float)b, &m, (float*)&work1, &lwork, info);
292	else if (typeid(fptype) == typeid(double))	292	else if (typeid(fptype) == typeid(double))
293	dgels_(mode, &m, &n, &k, (double)tmpA, &ldtmpA, (double)b, &m, (double*)&work1, &lwork, info);	293	LAPACK_FUNC(dgels)(mode, &m, &n, &k, (double)tmpA, &ldtmpA, (double)b, &m, (double*)&work1, &lwork, info);
294		294
295	lwork = cvRound(work1); //optimal buffer size	295	lwork = cvRound(work1); //optimal buffer size
296	std::vector<fptype> workBufMemHolder(lwork + 1);	296	std::vector<fptype> workBufMemHolder(lwork + 1);
297	fptype* buffer = &workBufMemHolder.front();	297	fptype* buffer = &workBufMemHolder.front();
298		298
299	if (typeid(fptype) == typeid(float))	299	if (typeid(fptype) == typeid(float))
300	sgels_(mode, &m, &n, &k, (float)tmpA, &ldtmpA, (float)b, &m, (float*)buffer, &lwork, info);	300	LAPACK_FUNC(sgels)(mode, &m, &n, &k, (float)tmpA, &ldtmpA, (float)b, &m, (float*)buffer, &lwork, info);
301	else if (typeid(fptype) == typeid(double))	301	else if (typeid(fptype) == typeid(double))
302	dgels_(mode, &m, &n, &k, (double)tmpA, &ldtmpA, (double)b, &m, (double*)buffer, &lwork, info);	302	LAPACK_FUNC(dgels)(mode, &m, &n, &k, (double)tmpA, &ldtmpA, (double)b, &m, (double*)buffer, &lwork, info);
303	}	303	}
304	else	304	else
305	{	305	{
Lines 309-326 lapack_QR(fptype* a, size_t a_step, int m, int n, int k, fptype* b, size_t b_ste Link Here
309	transpose(b, ldb, tmpB, m, m, k);	309	transpose(b, ldb, tmpB, m, m, k);
310		310
311	if (typeid(fptype) == typeid(float))	311	if (typeid(fptype) == typeid(float))
312	sgels_(mode, &m, &n, &k, (float)tmpA, &ldtmpA, (float)tmpB, &m, (float*)&work1, &lwork, info);	312	LAPACK_FUNC(sgels)(mode, &m, &n, &k, (float)tmpA, &ldtmpA, (float)tmpB, &m, (float*)&work1, &lwork, info);
313	else if (typeid(fptype) == typeid(double))	313	else if (typeid(fptype) == typeid(double))
314	dgels_(mode, &m, &n, &k, (double)tmpA, &ldtmpA, (double)tmpB, &m, (double*)&work1, &lwork, info);	314	LAPACK_FUNC(dgels)(mode, &m, &n, &k, (double)tmpA, &ldtmpA, (double)tmpB, &m, (double*)&work1, &lwork, info);
315		315
316	lwork = cvRound(work1); //optimal buffer size	316	lwork = cvRound(work1); //optimal buffer size
317	std::vector<fptype> workBufMemHolder(lwork + 1);	317	std::vector<fptype> workBufMemHolder(lwork + 1);
318	fptype* buffer = &workBufMemHolder.front();	318	fptype* buffer = &workBufMemHolder.front();
319		319
320	if (typeid(fptype) == typeid(float))	320	if (typeid(fptype) == typeid(float))
321	sgels_(mode, &m, &n, &k, (float)tmpA, &ldtmpA, (float)tmpB, &m, (float*)buffer, &lwork, info);	321	LAPACK_FUNC(sgels)(mode, &m, &n, &k, (float)tmpA, &ldtmpA, (float)tmpB, &m, (float*)buffer, &lwork, info);
322	else if (typeid(fptype) == typeid(double))	322	else if (typeid(fptype) == typeid(double))
323	dgels_(mode, &m, &n, &k, (double)tmpA, &ldtmpA, (double)tmpB, &m, (double*)buffer, &lwork, info);	323	LAPACK_FUNC(dgels)(mode, &m, &n, &k, (double)tmpA, &ldtmpA, (double)tmpB, &m, (double*)buffer, &lwork, info);
324		324
325	transpose(tmpB, m, b, ldb, k, m);	325	transpose(tmpB, m, b, ldb, k, m);
326	}	326	}