mirror/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2024-11-24 03:00:14 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

Merge remote-tracking branch 'upstream/3.4' into merge-3.4

Browse Source
This commit is contained in:
Alexander Alekhin 2021-12-22 15:54:58 +00:00
commit 9777fbacf6
4 changed files with 50 additions and 19 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

17
cmake/OpenCVFindLAPACK.cmake
View File

@ -51,6 +51,23 @@ macro(ocv_lapack_check)
if(NOT "${OPENCV_CBLAS_H_PATH_${_lapack_impl}}" STREQUAL "${OPENCV_LAPACKE_H_PATH_${_lapack_impl}}")
list(APPEND _lapack_content "#include \"${OPENCV_LAPACKE_H_PATH_${_lapack_impl}}\"")
endif()
list(APPEND _lapack_content "
#if defined(LAPACK_GLOBAL) || defined(LAPACK_NAME)
/*
* Using netlib's reference LAPACK implementation version >= 3.4.0 (first with C interface).
* Use LAPACK_xxxx to transparently (via predefined lapack macros) deal with pre and post 3.9.1 versions.
* LAPACK 3.9.1 introduces LAPACK_FORTRAN_STRLEN_END and modifies (through preprocessing) the declarations of the following functions used in opencv
* sposv_, dposv_, spotrf_, dpotrf_, sgesdd_, dgesdd_, sgels_, dgels_
* which end up with an extra parameter.
* So we also need to preprocess the function calls in opencv coding by prefixing them with LAPACK_.
* The good news is the preprocessing works fine whatever netlib's LAPACK version.
*/
#define OCV_LAPACK_FUNC(f) LAPACK_##f
#else
/* Using other LAPACK implementations so fall back to opencv's assumption until now */
#define OCV_LAPACK_FUNC(f) f##_
#endif
")
if(${_lapack_add_extern_c})
list(APPEND _lapack_content "}")
endif()

36
modules/core/src/hal_internal.cpp
View File

@ -163,9 +163,9 @@ lapack_Cholesky(fptype* a, size_t a_step, int m, fptype* b, size_t b_step, int n
if(n == 1 && b_step == sizeof(fptype))
{
if(typeid(fptype) == typeid(float))
sposv_(L, &m, &n, (float*)a, &lda, (float*)b, &m, &lapackStatus);
OCV_LAPACK_FUNC(sposv)(L, &m, &n, (float*)a, &lda, (float*)b, &m, &lapackStatus);
else if(typeid(fptype) == typeid(double))
dposv_(L, &m, &n, (double*)a, &lda, (double*)b, &m, &lapackStatus);
OCV_LAPACK_FUNC(dposv)(L, &m, &n, (double*)a, &lda, (double*)b, &m, &lapackStatus);
}
else
{
@ -174,9 +174,9 @@ lapack_Cholesky(fptype* a, size_t a_step, int m, fptype* b, size_t b_step, int n
transpose(b, ldb, tmpB, m, m, n);
if(typeid(fptype) == typeid(float))
sposv_(L, &m, &n, (float*)a, &lda, (float*)tmpB, &m, &lapackStatus);
OCV_LAPACK_FUNC(sposv)(L, &m, &n, (float*)a, &lda, (float*)tmpB, &m, &lapackStatus);
else if(typeid(fptype) == typeid(double))
dposv_(L, &m, &n, (double*)a, &lda, (double*)tmpB, &m, &lapackStatus);
OCV_LAPACK_FUNC(dposv)(L, &m, &n, (double*)a, &lda, (double*)tmpB, &m, &lapackStatus);
transpose(tmpB, m, b, ldb, n, m);
delete[] tmpB;
@ -185,9 +185,9 @@ lapack_Cholesky(fptype* a, size_t a_step, int m, fptype* b, size_t b_step, int n
else
{
if(typeid(fptype) == typeid(float))
spotrf_(L, &m, (float*)a, &lda, &lapackStatus);
OCV_LAPACK_FUNC(spotrf)(L, &m, (float*)a, &lda, &lapackStatus);
else if(typeid(fptype) == typeid(double))
dpotrf_(L, &m, (double*)a, &lda, &lapackStatus);
OCV_LAPACK_FUNC(dpotrf)(L, &m, (double*)a, &lda, &lapackStatus);
}
if(lapackStatus == 0) *info = true;
@ -227,17 +227,17 @@ lapack_SVD(fptype* a, size_t a_step, fptype *w, fptype* u, size_t u_step, fptype
}
if(typeid(fptype) == typeid(float))
sgesdd_(mode, &m, &n, (float*)a, &lda, (float*)w, (float*)u, &ldu, (float*)vt, &ldv, (float*)&work1, &lwork, iworkBuf, info);
OCV_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))
dgesdd_(mode, &m, &n, (double*)a, &lda, (double*)w, (double*)u, &ldu, (double*)vt, &ldv, (double*)&work1, &lwork, iworkBuf, info);
OCV_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))
sgesdd_(mode, &m, &n, (float*)a, &lda, (float*)w, (float*)u, &ldu, (float*)vt, &ldv, (float*)buffer, &lwork, iworkBuf, info);
OCV_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))
dgesdd_(mode, &m, &n, (double*)a, &lda, (double*)w, (double*)u, &ldu, (double*)vt, &ldv, (double*)buffer, &lwork, iworkBuf, info);
OCV_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);
@ -288,18 +288,18 @@ lapack_QR(fptype* a, size_t a_step, int m, int n, int k, fptype* b, size_t b_ste
if (k == 1 && b_step == sizeof(fptype))
{
if (typeid(fptype) == typeid(float))
sgels_(mode, &m, &n, &k, (float*)tmpA, &ldtmpA, (float*)b, &m, (float*)&work1, &lwork, info);
OCV_LAPACK_FUNC(sgels)(mode, &m, &n, &k, (float*)tmpA, &ldtmpA, (float*)b, &m, (float*)&work1, &lwork, info);
else if (typeid(fptype) == typeid(double))
dgels_(mode, &m, &n, &k, (double*)tmpA, &ldtmpA, (double*)b, &m, (double*)&work1, &lwork, info);
OCV_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))
sgels_(mode, &m, &n, &k, (float*)tmpA, &ldtmpA, (float*)b, &m, (float*)buffer, &lwork, info);
OCV_LAPACK_FUNC(sgels)(mode, &m, &n, &k, (float*)tmpA, &ldtmpA, (float*)b, &m, (float*)buffer, &lwork, info);
else if (typeid(fptype) == typeid(double))
dgels_(mode, &m, &n, &k, (double*)tmpA, &ldtmpA, (double*)b, &m, (double*)buffer, &lwork, info);
OCV_LAPACK_FUNC(dgels)(mode, &m, &n, &k, (double*)tmpA, &ldtmpA, (double*)b, &m, (double*)buffer, &lwork, info);
}
else
{
@ -309,18 +309,18 @@ lapack_QR(fptype* a, size_t a_step, int m, int n, int k, fptype* b, size_t b_ste
transpose(b, ldb, tmpB, m, m, k);
if (typeid(fptype) == typeid(float))
sgels_(mode, &m, &n, &k, (float*)tmpA, &ldtmpA, (float*)tmpB, &m, (float*)&work1, &lwork, info);
OCV_LAPACK_FUNC(sgels)(mode, &m, &n, &k, (float*)tmpA, &ldtmpA, (float*)tmpB, &m, (float*)&work1, &lwork, info);
else if (typeid(fptype) == typeid(double))
dgels_(mode, &m, &n, &k, (double*)tmpA, &ldtmpA, (double*)tmpB, &m, (double*)&work1, &lwork, info);
OCV_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))
sgels_(mode, &m, &n, &k, (float*)tmpA, &ldtmpA, (float*)tmpB, &m, (float*)buffer, &lwork, info);
OCV_LAPACK_FUNC(sgels)(mode, &m, &n, &k, (float*)tmpA, &ldtmpA, (float*)tmpB, &m, (float*)buffer, &lwork, info);
else if (typeid(fptype) == typeid(double))
dgels_(mode, &m, &n, &k, (double*)tmpA, &ldtmpA, (double*)tmpB, &m, (double*)buffer, &lwork, info);
OCV_LAPACK_FUNC(dgels)(mode, &m, &n, &k, (double*)tmpA, &ldtmpA, (double*)tmpB, &m, (double*)buffer, &lwork, info);
transpose(tmpB, m, b, ldb, k, m);
}

11
modules/dnn/src/onnx/onnx_importer.cpp
View File

@ -1711,7 +1711,16 @@ void ONNXImporter::parseMul(LayerParams& layerParams, const opencv_onnx::NodePro
blob = blob.reshape(1, 1);
if (blob.total() == 1) {
float blob_value = blob.ptr<float>()[0];
float coeff = isDiv ? 1.0 / blob_value : blob_value;
float coeff = blob_value;
if (isDiv)
{
coeff = 1.f / blob_value;
if (constId == 0)
{
// Power layer calculates (x*scale + shift)^power, so const/x -> (x * (1/const) + 0)^(-1)
layerParams.set("power", -1.f);
}
}
layerParams.set("scale", coeff);
layerParams.type = "Power";
}

5
modules/dnn/test/test_onnx_importer.cpp
View File

@ -1338,6 +1338,11 @@ TEST_P(Test_ONNX_layers, SubFromConst)
testONNXModels("sub_from_const_broadcast");
}
TEST_P(Test_ONNX_layers, DivConst)
{
testONNXModels("div_const");
}
TEST_P(Test_ONNX_layers, Quantized_Convolution)
{