mirror of
https://github.com/opencv/opencv.git
synced 2025-01-18 22:44:02 +08:00
Merge pull request #7999 from alalek:fix_lapack
This commit is contained in:
Alexander Alekhin
commit
7c91700cc6
@ -1,78 +1,159 @@
|
||||
macro(_find_file_in_dirs VAR NAME DIRS)
|
||||
find_path(${VAR} ${NAME} ${DIRS} NO_DEFAULT_PATH)
|
||||
set(${VAR} ${${VAR}}/${NAME})
|
||||
unset(${VAR} CACHE)
|
||||
macro(_find_header_file_in_dirs VAR NAME)
|
||||
unset(${VAR})
|
||||
unset(${VAR} CACHE)
|
||||
if(" ${ARGN}" STREQUAL " ")
|
||||
check_include_file("${NAME}" HAVE_${VAR})
|
||||
if(HAVE_${VAR})
|
||||
set(${VAR} "${NAME}") # fallback
|
||||
else()
|
||||
set(${VAR} "")
|
||||
endif()
|
||||
else()
|
||||
find_path(${VAR} "${NAME}" ${ARGN} NO_DEFAULT_PATH)
|
||||
if(${VAR})
|
||||
set(${VAR} "${${VAR}}/${NAME}")
|
||||
unset(${VAR} CACHE)
|
||||
else()
|
||||
unset(${VAR} CACHE)
|
||||
set(${VAR} "")
|
||||
endif()
|
||||
endif()
|
||||
endmacro()
|
||||
|
||||
macro(ocv_lapack_check)
|
||||
string(REGEX REPLACE "[^a-zA-Z0-9_]" "_" _lapack_impl "${LAPACK_IMPL}")
|
||||
message(STATUS "LAPACK(${LAPACK_IMPL}): LAPACK_LIBRARIES: ${LAPACK_LIBRARIES}")
|
||||
_find_header_file_in_dirs(OPENCV_CBLAS_H_PATH_${_lapack_impl} "${LAPACK_CBLAS_H}" "${LAPACK_INCLUDE_DIR}")
|
||||
_find_header_file_in_dirs(OPENCV_LAPACKE_H_PATH_${_lapack_impl} "${LAPACK_LAPACKE_H}" "${LAPACK_INCLUDE_DIR}")
|
||||
if(NOT OPENCV_CBLAS_H_PATH_${_lapack_impl} OR NOT OPENCV_LAPACKE_H_PATH_${_lapack_impl})
|
||||
message(WARNING "LAPACK(${LAPACK_IMPL}): CBLAS/LAPACK headers are not found in '${LAPACK_INCLUDE_DIR}'")
|
||||
unset(LAPACK_LIBRARIES)
|
||||
else()
|
||||
# adding proxy opencv_lapack.h header
|
||||
set(CBLAS_H_PROXY_PATH ${CMAKE_BINARY_DIR}/opencv_lapack.h)
|
||||
set(_lapack_include_str "\#include \"${OPENCV_CBLAS_H_PATH_${_lapack_impl}}\"")
|
||||
if(NOT "${OPENCV_CBLAS_H_PATH_${_lapack_impl}}" STREQUAL "${OPENCV_LAPACKE_H_PATH_${_lapack_impl}}")
|
||||
set(_lapack_include_str "${_lapack_include_str}\n#include \"${OPENCV_LAPACKE_H_PATH_${_lapack_impl}}\"")
|
||||
endif()
|
||||
# update file contents (if required)
|
||||
set(__content_str "")
|
||||
if(EXISTS "${CBLAS_H_PROXY_PATH}")
|
||||
file(READ "${CBLAS_H_PROXY_PATH}" __content_str)
|
||||
endif()
|
||||
if(NOT " ${__content_str}" STREQUAL " ${_lapack_include_str}")
|
||||
file(WRITE "${CBLAS_H_PROXY_PATH}" "${_lapack_include_str}")
|
||||
endif()
|
||||
|
||||
try_compile(__VALID_LAPACK
|
||||
"${OpenCV_BINARY_DIR}"
|
||||
"${OpenCV_SOURCE_DIR}/cmake/checks/lapack_check.cpp"
|
||||
CMAKE_FLAGS "-DINCLUDE_DIRECTORIES:STRING=${LAPACK_INCLUDE_DIR}\;${CMAKE_BINARY_DIR}"
|
||||
"-DLINK_DIRECTORIES:STRING=${LAPACK_LINK_LIBRARIES}"
|
||||
"-DLINK_LIBRARIES:STRING=${LAPACK_LIBRARIES}"
|
||||
OUTPUT_VARIABLE TRY_OUT
|
||||
)
|
||||
if(NOT __VALID_LAPACK)
|
||||
#message(FATAL_ERROR "LAPACK: check build log:\n${TRY_OUT}")
|
||||
message(STATUS "LAPACK(${LAPACK_IMPL}): Can't build LAPACK check code. This LAPACK version is not supported.")
|
||||
unset(LAPACK_LIBRARIES)
|
||||
else()
|
||||
message(STATUS "LAPACK(${LAPACK_IMPL}): Support is enabled.")
|
||||
ocv_include_directories(${LAPACK_INCLUDE_DIR})
|
||||
set(HAVE_LAPACK 1)
|
||||
endif()
|
||||
endif()
|
||||
endmacro()
|
||||
|
||||
if(WITH_LAPACK)
|
||||
ocv_update(LAPACK_IMPL "Unknown")
|
||||
if(NOT LAPACK_LIBRARIES)
|
||||
include(cmake/OpenCVFindMKL.cmake)
|
||||
if(HAVE_MKL)
|
||||
set(LAPACK_INCLUDE_DIR ${MKL_INCLUDE_DIRS})
|
||||
set(LAPACK_LIBRARIES ${MKL_LIBRARIES} )
|
||||
set(LAPACK_CBLAS_H "mkl_cblas.h" )
|
||||
set(LAPACK_LAPACKE_H "mkl_lapack.h" )
|
||||
set(LAPACK_IMPL "MKL")
|
||||
endif()
|
||||
ocv_update(LAPACK_IMPL "Unknown")
|
||||
if(NOT OPENCV_LAPACK_FIND_PACKAGE_ONLY)
|
||||
if(NOT LAPACK_LIBRARIES AND NOT OPENCV_LAPACK_DISABLE_MKL)
|
||||
include(cmake/OpenCVFindMKL.cmake)
|
||||
if(HAVE_MKL)
|
||||
set(LAPACK_INCLUDE_DIR ${MKL_INCLUDE_DIRS})
|
||||
set(LAPACK_LIBRARIES ${MKL_LIBRARIES})
|
||||
set(LAPACK_CBLAS_H "mkl_cblas.h")
|
||||
set(LAPACK_LAPACKE_H "mkl_lapack.h")
|
||||
set(LAPACK_IMPL "MKL")
|
||||
ocv_lapack_check()
|
||||
endif()
|
||||
endif()
|
||||
if(NOT LAPACK_LIBRARIES)
|
||||
include(cmake/OpenCVFindOpenBLAS.cmake)
|
||||
if(OpenBLAS_FOUND)
|
||||
set(LAPACK_INCLUDE_DIR ${OpenBLAS_INCLUDE_DIR} )
|
||||
set(LAPACK_LIBRARIES ${OpenBLAS_LIB} )
|
||||
set(LAPACK_CBLAS_H "cblas.h" )
|
||||
set(LAPACK_LAPACKE_H "lapacke.h" )
|
||||
set(LAPACK_IMPL "OpenBLAS")
|
||||
endif()
|
||||
include(cmake/OpenCVFindOpenBLAS.cmake)
|
||||
if(OpenBLAS_FOUND)
|
||||
set(LAPACK_INCLUDE_DIR ${OpenBLAS_INCLUDE_DIR})
|
||||
set(LAPACK_LIBRARIES ${OpenBLAS_LIB})
|
||||
set(LAPACK_CBLAS_H "cblas.h")
|
||||
set(LAPACK_LAPACKE_H "lapacke.h")
|
||||
set(LAPACK_IMPL "OpenBLAS")
|
||||
ocv_lapack_check()
|
||||
endif()
|
||||
endif()
|
||||
if(NOT LAPACK_LIBRARIES AND UNIX)
|
||||
include(cmake/OpenCVFindAtlas.cmake)
|
||||
if(ATLAS_FOUND)
|
||||
set(LAPACK_INCLUDE_DIR ${Atlas_INCLUDE_DIR})
|
||||
set(LAPACK_LIBRARIES ${Atlas_LIBRARIES} )
|
||||
set(LAPACK_CBLAS_H "cblas.h" )
|
||||
set(LAPACK_LAPACKE_H "lapacke.h" )
|
||||
set(LAPACK_IMPL "Atlas")
|
||||
endif()
|
||||
include(cmake/OpenCVFindAtlas.cmake)
|
||||
if(ATLAS_FOUND)
|
||||
set(LAPACK_INCLUDE_DIR ${Atlas_INCLUDE_DIR})
|
||||
set(LAPACK_LIBRARIES ${Atlas_LIBRARIES})
|
||||
set(LAPACK_CBLAS_H "cblas.h")
|
||||
set(LAPACK_LAPACKE_H "lapacke.h")
|
||||
set(LAPACK_IMPL "Atlas")
|
||||
ocv_lapack_check()
|
||||
endif()
|
||||
endif()
|
||||
endif()
|
||||
|
||||
if(NOT LAPACK_LIBRARIES AND APPLE)
|
||||
set(LAPACK_INCLUDE_DIR "Accelerate")
|
||||
set(LAPACK_LIBRARIES "-framework Accelerate")
|
||||
set(LAPACK_CBLAS_H "cblas.h" )
|
||||
set(LAPACK_LAPACKE_H "lapacke.h" )
|
||||
set(LAPACK_IMPL "Apple")
|
||||
if(NOT LAPACK_LIBRARIES)
|
||||
if(WIN32 AND NOT OPENCV_LAPACK_SHARED_LIBS)
|
||||
set(BLA_STATIC 1)
|
||||
endif()
|
||||
|
||||
set(LAPACK_INCLUDE_DIR ${LAPACK_INCLUDE_DIR} CACHE PATH "Path to BLAS include dir" FORCE)
|
||||
set(LAPACK_CBLAS_H ${LAPACK_CBLAS_H} CACHE STRING "Alternative name of cblas.h" FORCE)
|
||||
set(LAPACK_LAPACKE_H ${LAPACK_LAPACKE_H} CACHE STRING "Alternative name of lapacke.h" FORCE)
|
||||
set(LAPACK_LIBRARIES ${LAPACK_LIBRARIES} CACHE STRING "Names of BLAS & LAPACK binaries (.so, .dll, .a, .lib)" FORCE)
|
||||
set(LAPACK_IMPL ${LAPACK_IMPL} CACHE STRING "Lapack implementation id" FORCE)
|
||||
|
||||
if(LAPACK_LIBRARIES) #adding proxy cblas.h header
|
||||
message(STATUS "LAPACK_IMPL: ${LAPACK_IMPL}, LAPACK_LIBRARIES: ${LAPACK_LIBRARIES}")
|
||||
if("${LAPACK_IMPL}" STREQUAL "Apple")
|
||||
set(CBLAS_H_PATH "Accelerate/Accelerate.h")
|
||||
set(LAPACKE_H_PATH "Accelerate/Accelerate.h")
|
||||
else()
|
||||
_find_file_in_dirs(CBLAS_H_PATH "${LAPACK_CBLAS_H}" "${LAPACK_INCLUDE_DIR}")
|
||||
_find_file_in_dirs(LAPACKE_H_PATH "${LAPACK_LAPACKE_H}" "${LAPACK_INCLUDE_DIR}")
|
||||
endif()
|
||||
if(NOT CBLAS_H_PATH OR NOT LAPACKE_H_PATH)
|
||||
message(WARNING "CBLAS/LAPACK headers are not found in '${LAPACK_INCLUDE_DIR}'")
|
||||
endif()
|
||||
ocv_include_directories(${LAPACK_INCLUDE_DIR})
|
||||
list(APPEND OPENCV_LINKER_LIBS ${LAPACK_LIBRARIES})
|
||||
set(HAVE_LAPACK 1)
|
||||
|
||||
set(CBLAS_H_PROXY_PATH ${CMAKE_BINARY_DIR}/opencv_lapack.h)
|
||||
set(_include_str "\#include \"${CBLAS_H_PATH}\"")
|
||||
if("${CBLAS_H_PATH}" STREQUAL "${LAPACKE_H_PATH}")
|
||||
else()
|
||||
set(_include_str "${_include_str}\n\#include \"${LAPACKE_H_PATH}\"")
|
||||
endif()
|
||||
file(WRITE ${CBLAS_H_PROXY_PATH} ${_include_str})
|
||||
find_package(LAPACK)
|
||||
if(LAPACK_FOUND)
|
||||
if(NOT DEFINED LAPACKE_INCLUDE_DIR)
|
||||
find_path(LAPACKE_INCLUDE_DIR "lapacke.h")
|
||||
endif()
|
||||
if(NOT DEFINED MKL_LAPACKE_INCLUDE_DIR)
|
||||
find_path(MKL_LAPACKE_INCLUDE_DIR "mkl_lapack.h")
|
||||
endif()
|
||||
if(MKL_LAPACKE_INCLUDE_DIR AND NOT OPENCV_LAPACK_DISABLE_MKL)
|
||||
set(LAPACK_INCLUDE_DIR ${MKL_LAPACKE_INCLUDE_DIR})
|
||||
set(LAPACK_CBLAS_H "mkl_cblas.h")
|
||||
set(LAPACK_LAPACKE_H "mkl_lapack.h")
|
||||
set(LAPACK_IMPL "LAPACK/MKL")
|
||||
ocv_lapack_check()
|
||||
endif()
|
||||
if(LAPACKE_INCLUDE_DIR AND NOT HAVE_LAPACK)
|
||||
set(LAPACK_INCLUDE_DIR ${LAPACKE_INCLUDE_DIR})
|
||||
set(LAPACK_CBLAS_H "cblas.h")
|
||||
set(LAPACK_LAPACKE_H "lapacke.h")
|
||||
set(LAPACK_IMPL "LAPACK/Generic")
|
||||
ocv_lapack_check()
|
||||
elseif(APPLE)
|
||||
set(LAPACK_CBLAS_H "Accelerate/Accelerate.h")
|
||||
set(LAPACK_LAPACKE_H "Accelerate/Accelerate.h")
|
||||
set(LAPACK_IMPL "LAPACK/Apple")
|
||||
ocv_lapack_check()
|
||||
else()
|
||||
unset(LAPACK_LIBRARIES CACHE)
|
||||
endif()
|
||||
endif()
|
||||
endif()
|
||||
|
||||
if(NOT LAPACK_LIBRARIES AND APPLE AND NOT OPENCV_LAPACK_FIND_PACKAGE_ONLY)
|
||||
set(LAPACK_INCLUDE_DIR "")
|
||||
set(LAPACK_LIBRARIES "-framework Accelerate")
|
||||
set(LAPACK_CBLAS_H "Accelerate/Accelerate.h")
|
||||
set(LAPACK_LAPACKE_H "Accelerate/Accelerate.h")
|
||||
set(LAPACK_IMPL "Apple")
|
||||
ocv_lapack_check()
|
||||
endif()
|
||||
|
||||
if(NOT HAVE_LAPACK AND LAPACK_LIBRARIES)
|
||||
ocv_lapack_check()
|
||||
endif()
|
||||
|
||||
set(LAPACK_INCLUDE_DIR ${LAPACK_INCLUDE_DIR} CACHE PATH "Path to BLAS include dir" FORCE)
|
||||
set(LAPACK_CBLAS_H ${LAPACK_CBLAS_H} CACHE STRING "Alternative name of cblas.h" FORCE)
|
||||
set(LAPACK_LAPACKE_H ${LAPACK_LAPACKE_H} CACHE STRING "Alternative name of lapacke.h" FORCE)
|
||||
set(LAPACK_LIBRARIES ${LAPACK_LIBRARIES} CACHE STRING "Names of BLAS & LAPACK binaries (.so, .dll, .a, .lib)" FORCE)
|
||||
set(LAPACK_IMPL ${LAPACK_IMPL} CACHE STRING "Lapack implementation id" FORCE)
|
||||
endif()
|
||||
|
||||
14
cmake/checks/lapack_check.cpp
Normal file
14
cmake/checks/lapack_check.cpp
Normal file
@ -0,0 +1,14 @@
|
||||
#include "opencv_lapack.h"
|
||||
|
||||
static char* check_fn1 = (char*)sgesv_;
|
||||
static char* check_fn2 = (char*)sposv_;
|
||||
static char* check_fn3 = (char*)spotrf_;
|
||||
static char* check_fn4 = (char*)sgesdd_;
|
||||
|
||||
int main(int argc, char* argv[])
|
||||
{
|
||||
(void)argv;
|
||||
if(argc > 1000)
|
||||
return check_fn1[0] + check_fn2[0] + check_fn3[0] + check_fn4[0];
|
||||
return 0;
|
||||
}
|
||||
@ -1,7 +1,7 @@
|
||||
set(the_description "The Core Functionality")
|
||||
ocv_add_module(core
|
||||
"${OPENCV_HAL_LINKER_LIBS}"
|
||||
PRIVATE_REQUIRED ${ZLIB_LIBRARIES} "${OPENCL_LIBRARIES}" "${VA_LIBRARIES}"
|
||||
PRIVATE_REQUIRED ${ZLIB_LIBRARIES} "${OPENCL_LIBRARIES}" "${VA_LIBRARIES}" "${LAPACK_LIBRARIES}"
|
||||
OPTIONAL opencv_cudev
|
||||
WRAP java python)
|
||||
|
||||
|
||||
@ -98,7 +98,7 @@ set_value(fptype *dst, size_t dst_ld, fptype value, size_t m, size_t n)
|
||||
template <typename fptype> static inline int
|
||||
lapack_LU(fptype* a, size_t a_step, int m, fptype* b, size_t b_step, int n, int* info)
|
||||
{
|
||||
int lda = a_step / sizeof(fptype), sign = 0;
|
||||
int lda = (int)(a_step / sizeof(fptype)), sign = 0;
|
||||
int* piv = new int[m];
|
||||
|
||||
transpose_square_inplace(a, lda, m);
|
||||
@ -114,7 +114,7 @@ lapack_LU(fptype* a, size_t a_step, int m, fptype* b, size_t b_step, int n, int*
|
||||
}
|
||||
else
|
||||
{
|
||||
int ldb = b_step / sizeof(fptype);
|
||||
int ldb = (int)(b_step / sizeof(fptype));
|
||||
fptype* tmpB = new fptype[m*n];
|
||||
|
||||
transpose(b, ldb, tmpB, m, m, n);
|
||||
@ -153,7 +153,7 @@ template <typename fptype> static inline int
|
||||
lapack_Cholesky(fptype* a, size_t a_step, int m, fptype* b, size_t b_step, int n, bool* info)
|
||||
{
|
||||
int lapackStatus = 0;
|
||||
int lda = a_step / sizeof(fptype);
|
||||
int lda = (int)(a_step / sizeof(fptype));
|
||||
char L[] = {'L', '\0'};
|
||||
|
||||
if(b)
|
||||
@ -167,7 +167,7 @@ lapack_Cholesky(fptype* a, size_t a_step, int m, fptype* b, size_t b_step, int n
|
||||
}
|
||||
else
|
||||
{
|
||||
int ldb = b_step / sizeof(fptype);
|
||||
int ldb = (int)(b_step / sizeof(fptype));
|
||||
fptype* tmpB = new fptype[m*n];
|
||||
transpose(b, ldb, tmpB, m, m, n);
|
||||
|
||||
@ -197,9 +197,9 @@ lapack_Cholesky(fptype* a, size_t a_step, int m, fptype* b, size_t b_step, int n
|
||||
template <typename fptype> static inline int
|
||||
lapack_SVD(fptype* a, size_t a_step, fptype *w, fptype* u, size_t u_step, fptype* vt, size_t v_step, int m, int n, int flags, int* info)
|
||||
{
|
||||
int lda = a_step / sizeof(fptype);
|
||||
int ldv = v_step / sizeof(fptype);
|
||||
int ldu = u_step / sizeof(fptype);
|
||||
int lda = (int)(a_step / sizeof(fptype));
|
||||
int ldv = (int)(v_step / sizeof(fptype));
|
||||
int ldu = (int)(u_step / sizeof(fptype));
|
||||
int lwork = -1;
|
||||
int* iworkBuf = new int[8*std::min(m, n)];
|
||||
fptype work1 = 0;
|
||||
@ -256,7 +256,7 @@ lapack_SVD(fptype* a, size_t a_step, fptype *w, fptype* u, size_t u_step, fptype
|
||||
template <typename fptype> static inline int
|
||||
lapack_QR(fptype* a, size_t a_step, int m, int n, int k, fptype* b, size_t b_step, fptype* dst, int* info)
|
||||
{
|
||||
int lda = a_step / sizeof(fptype);
|
||||
int lda = (int)(a_step / sizeof(fptype));
|
||||
char mode[] = { 'N', '\0' };
|
||||
if(m < n)
|
||||
return CV_HAL_ERROR_NOT_IMPLEMENTED;
|
||||
@ -303,7 +303,7 @@ lapack_QR(fptype* a, size_t a_step, int m, int n, int k, fptype* b, size_t b_ste
|
||||
{
|
||||
std::vector<fptype> tmpBMemHolder(m*k);
|
||||
fptype* tmpB = &tmpBMemHolder.front();
|
||||
int ldb = b_step / sizeof(fptype);
|
||||
int ldb = (int)(b_step / sizeof(fptype));
|
||||
transpose(b, ldb, tmpB, m, m, k);
|
||||
|
||||
if (typeid(fptype) == typeid(float))
|
||||
@ -357,10 +357,10 @@ template <typename fptype> static inline int
|
||||
lapack_gemm(const fptype *src1, size_t src1_step, const fptype *src2, size_t src2_step, fptype alpha,
|
||||
const fptype *src3, size_t src3_step, fptype beta, fptype *dst, size_t dst_step, int a_m, int a_n, int d_n, int flags)
|
||||
{
|
||||
int ldsrc1 = src1_step / sizeof(fptype);
|
||||
int ldsrc2 = src2_step / sizeof(fptype);
|
||||
int ldsrc3 = src3_step / sizeof(fptype);
|
||||
int lddst = dst_step / sizeof(fptype);
|
||||
int ldsrc1 = (int)(src1_step / sizeof(fptype));
|
||||
int ldsrc2 = (int)(src2_step / sizeof(fptype));
|
||||
int ldsrc3 = (int)(src3_step / sizeof(fptype));
|
||||
int lddst = (int)(dst_step / sizeof(fptype));
|
||||
int c_m, c_n, d_m;
|
||||
CBLAS_TRANSPOSE transA, transB;
|
||||
|
||||
@ -434,10 +434,10 @@ template <typename fptype> static inline int
|
||||
lapack_gemm_c(const fptype *src1, size_t src1_step, const fptype *src2, size_t src2_step, fptype alpha,
|
||||
const fptype *src3, size_t src3_step, fptype beta, fptype *dst, size_t dst_step, int a_m, int a_n, int d_n, int flags)
|
||||
{
|
||||
int ldsrc1 = src1_step / sizeof(std::complex<fptype>);
|
||||
int ldsrc2 = src2_step / sizeof(std::complex<fptype>);
|
||||
int ldsrc3 = src3_step / sizeof(std::complex<fptype>);
|
||||
int lddst = dst_step / sizeof(std::complex<fptype>);
|
||||
int ldsrc1 = (int)(src1_step / sizeof(std::complex<fptype>));
|
||||
int ldsrc2 = (int)(src2_step / sizeof(std::complex<fptype>));
|
||||
int ldsrc3 = (int)(src3_step / sizeof(std::complex<fptype>));
|
||||
int lddst = (int)(dst_step / sizeof(std::complex<fptype>));
|
||||
int c_m, c_n, d_m;
|
||||
CBLAS_TRANSPOSE transA, transB;
|
||||
std::complex<fptype> cAlpha(alpha, 0.0);
|
||||
|
||||
Loading…
Reference in New Issue
Block a user