Intel® oneAPI Math Kernel Library (oneMKL) Bug Fixes

ID 标签 765897
已更新 7/2/2024
版本 2024.2
公共

author-image

作者

NOTE: Defects and feature requests described below represent specific issues with specific test cases. It is difficult to succinctly describe an issue and how it impacted the specific test case. Some of the issues listed may impact multiple architectures, operating systems, and/or languages. If you have any questions about the issues discussed in this report, please visit our Get Help page for support options.

2024.2 Release (June 2024)

MKLD-11741 Fixed oneMKL Pardiso failure when solving for partial solution
MKLD-17048 Stopped to expose the MKL_LIBRARIES variable in the MKLConfig.cmake file to users to avoid misusage
MKLD-17054 Fixed the accuracy failures in batched getrf for complex precisions on Intel® Data Center GPU Max Series
MKLD-17109 Fixed the issue that ?trnlspbc fails to find a low-residual solution even if the solution is within the limits
MKLD-17187 Fixed the Conda-forge blas-feedstock tests failure with LAPACK 3.10.1
MKLD-17226 Fixed the incorrect check for ldb parameter in gesv_scratchpad_size and gesv
MKLD-17265 Fixed the issue that Nuget package comes without Fortran libraries
MKLD-17315 Fixed the typos in mkl_lapack.fi to avoid compilation errors
MKLD-17340 Fixed the wrong result of ?LASSQ for large input
MKLD-17439 Fixed the failure of 3D complex descriptor commit for SYCL API on CPU 

2024.1 Release (March 2024)

MKLD-11348 Fixed the problem in ?LANGE and all Lapack norm routines when the input contains NaNs
MKLD-14339 Fixed thread safety issue when using oneMKL PARDISO+METIS
MKLD-15333 Fixed Non-determinism problem when solving multiple linear systems with oneMKL PARDISO+METIS
MKLD-15406
Fixed COPT selection for different languages in MKLConfig.cmake
MKLD-15828
Fixed the issue that oneMKL Redist Nuget package copies runtimes to the wrong directory
MKLD-16067
Improved the performance of SYCL API on CPU for small-sized SDOT routine
MKLD-16656 Removed MKL warning on the processor with AVX support
MKLD-16708
Moved libmkl_sycl.so from the runtime package to the devel package
MKLD-16995  Fixed the issue that 1D FFT fails for some of specific problem sizes

2024.0 Release (November 2023)

MKLD-11567 Improved the accuracy for solving generalized eigenvalue problems. 
MKLD-13797 Function declarations in uppercase with trailing underscore (such as “CGESVDQ_” vs. “cgesvdq”) in mkl_lapack.h without the corresponding symbols in the libraries have been removed. 
MKLD-14216 Improved perfomance of dormqr on Intel® Data Center GPU Max Series. 
MKLD-14317 Cluster Sparse Solver raised a fatal error in PMPI_Gatherv. 
MKLD-14556 cmake reported incorrect MKL_VERSION if both MKLROOT and CMAKE_PREFIX_PATH are set but mismatch. 
MKLD-14896 Fixed an inconsistency between the oneAPI specification and the Intel® oneMKL documentation for sparse::release_matrix_handle 
MKLD-15785 Using the CBLAS macro caused warning messages when mkl_direct_call is defined. 
MKLD-16003 dgemm_batch had extra host overhead for each group in batch when using OpenMP* offloading compared to SYCL*. 
MKLD-16057 The samples in examples_core_c.tgz now use the Intel®oneAPI DPC++/C++ Compiler (icx) by default, as the Intel® C++ Compiler Classic (icc) has been removed from the oneAPI base package. 
MKLD-16096 MKLConfig.cmake detected system-installed instead of user-installed libraries first on Linux*. 
MKLD-16177 Fixed typos and misprints in mp_linpack readme.txt 
MKLD-16220 High-performance LINPACK benchmark did not discover Intel® Data Center GPU Max Series with device id 1450. 

2023.2 Release (July 2023)

MKLD-13292 MKLD-14768

Fixed the wrong error message in oneMKL PARDISO when using the minimum degree reordering algorithm with rank-deficient, invalid input
MKLD-13622 Fixed the issue that the type of dummy and actual parameters doesn’t match for the Fortran APIs in Vector Math
MKLD-14559 Added note about additional error codes for oneMKL PARDISO in the developer reference for C
MKLD-14659 Fixed the issue that the DPC++ version of LAPACK routine GETRF fails to throw an exception on singular input matrices
MKLD-14856 Fixed errors, inconsistency between description and implementation, and added validation in some Sparse BLAS examples
MKLD-14886 Fixed the DPC++ documentation about LAPACK routines, getr{i,s}, potr{f,s}, trtrs, on the supported precision on Intel GPU
MKLD-14938 Added exit if an internal allocation fails for Cluster Pardiso
MKLD-14955 Fixed the issue that BLAS routines her, her2, hpr and hpr2 generates incorrect result when alpha is set to zero
MKLD-15064 Fixed the issue that dgetrf hangs when using multiple host threads
MKLD-15217 Fixed the wrong output of LAPACK routine dlamc2
MKLD-15229 Fixed the inappropriate use of sycl::abs in poisson_impl.hpp
MKLD-15340 Fixed the issue that vzAbs may generate spurious exception when using AVX2 or older architecture
MKLD-15750 Fixed the issue that oneMKL CMake configure file doesn't work with oneMKL Conda package

2023.1 Release (April 2023)

MKLD-13049 Removed the LAPACK example and applied various other fixes to the introductory material for BLAS and LAPACK in the developer reference 
MKLD-13828 Fixed the issue that the strided API is much faster than the group API for batched getri 
MKLD-13836 Fixed the issue that 32-bit version oneMKL produces different 1D FFT result between single- and multi-threaded 
MKLD-13880 Fixed the issue that Link Line Advisor is missing -lmkl_sycl for OpenMP offloading 
MKLD-13906 Fixed the issue that throws exception when calling mkl_sparse_d_syprd for large arrays 
MKLD-14119 Fixed the performance issue with Philox RNG for the SYCL API 
MKLD-14168 Fixed the memory management issue in the cl_solver_export_c example 
MKLD-14362 Fixed the wrong result of array DL from GTSV 
MKLD-14407 Fixed the misprint in the gemm_usm_multi_stack example 
MKLD-14516 Fixed the missing uppercase/alias function declarations in mkl_pardiso.h
MKLD-14558 Fixed the CNR sample code from oneMKL developer guide 
MKLD-14689 Fixed some wrong declarations in the Fortran files in the $MKLROOT/include directory 

2023.0 Release (December 2022)

MKLD-14199 Fixed issue with uninitialized device number in VML OpenMP offloading examples
MKLD-14060 Fixed issues with incorrect descriptions in ScaLAPACK examples
MKLD-13858 Fixed issue with missing option -MKLD_ILP64 in Cmake
MKLD-13818 Fixed issue with oneMKL examples failing due to recent llvm updates
MKLD-13796 Fixed issue with Different function-signatures between some LAPACK function Aliases
MKLD-13746 Fixed issue with missing symbols in LAPACK functions `slaqz3_64` and `SLAQZ3_64`
MKLD-13742 Fixed issue with Floating invalid exception in solving phase of MKL cluster sparse solver
MKLD-13430 Fixed issue with memory leaking in mkl_sparse_sypr

2022.2 Release (August 2022)

MKLD-13840 Fixed issue with the missing heasder files fftw3_mkl.h and fftw3-mpi_mkl.h
MKLD-13733 Fixed issue with libmkl_vml_mc2.dylib still remaining in the Mac OS package
MKLD-13621 Fixed issue with missing files mkl_mc3.dll and libmkl_mc3.dylib in the MacOS Package
MKLD-13502 Fixed issue in spblas.f90
MKLD-13451 Fixed the security issue in LAPACK's ?larrv
MKLD-13353 Fixed corner cases in MKLConfig.cmake
MKLD-13135 Fixed issue with Batch FFT performance degrading when using non-default scaling
MKLD-13076 Fixed Issue with DSTEDC
MKLD-13048 Fixed issues in potrs.cpp, potrs_usm.cpp
MKLD-12888 Fixed issue with natural cubic spline with not-a-knot boundary conditions
MKLD-12822 Fixed PARDISO issue when using with{{msglvl = 1}}

2022.1 Release (March 2022)

MKLD-13001 Fixed CBLAS - MKL Developer Guide
MKLD-12926 Fixed issue with MKL dpcpp axpy running much slower then the pure dpcpp code
MKLD-12917 Fixed dpcpp fft example( dp_real_1d.cpp) description
MKLD-12792 Fixed issue with LAPACKE_cgesvd_work generating core dump
MKLD-12762 Fixed issue with GPU function names different from those of CPU in VERBOSE mode.
MKLD-12672 Fixed issue with SVD ( jobz = 'O' ) of a tall thin matrix slowing down comparing to those of version 2019.3
MKLD-12609 Fixed issue with DFTI Compute Forward aborting while running with level-zero backend.
MKLD-12503 Fixed issue with MKL library miscounting available cores.
MKLD-11952 Fixed issue with the VML default values
MKLD-11544 Fixed issue with dfeast_scsrgv returns different numbers of EV when the search interval bounds coincide with the eigenvalues

2022.0 Release (December 2021)

MKLD-12475 Added verbose mode for GPU.
MKLD-12463 Fixed the misprint in mkl_types.h
MKLD-12191 Fixed the performance degradation of SVD in version 2021 for tall-skinny matrices comparing to that of version 2019.3.
MKLD-12181 Fixed the issue that the SPARSE_MATRIX_CHECKER routine not recognizing when the input matrix is not sorted in the ascending order.
MKLD-12149 Fixed the wrong comments in mkl_df_functions.h and mkl_vsl_functions.h files.
MKLD-12129 Fixed the warning issue of the integer-size 64 for parameter VSL_QRNG_LEAPFROG_COMPONENTS.
MKLD-12118 Fixed the issue that GEMM routines output NaN when y contains NaN even if beta is zero.
MKLD-12052 Fixed the memory leak issue when using MKL_SPARSE_CONVERT_CSR ROUTINE.
MKLD-12019 Updated the content description of FFT dpcpp examples.
MKLD-11821 Fixed the issue oneMKL Pardiso reporting wrong size of memory allocated in phases 11 ans 22.
MKLD-11358 Fixed the crashing issue with runtime error CL_INVALID_VALUE of GEMM routine in DPC++.

2021.4 Release (September 2021)

MKLD-12148 Fixed the wrong comments in the file mkl_vmlfunctions.h.
MKLD-12052 Fixed the memory leak issue when using MKL_SPARSE_CONVERT_CSR.
MKLD-11763 Fixed the error "Variable MKLROOT not defined" reported by pkg-config.
MKLD-11725 Fixed the hanging issue in the cluster sparse solver when iparm[3]=31.
MKLD-11723 Fixed the crash issue when using P?GEEVX.
MKLD-11665 Fixed the incorrect result produced by MKL PARDISO when a permutation vectore is used.
MKLD-11542 Fixed the misprints in mkl_vml_defines.h.
MKLD-11531 Fixed the issue MKL_SPARSE_QR_REORDER returning the "-6" error code.
MKLD-11515 The /MD option is documented in the oneMKL Developer Guide for Windows.
MKLD-11514 Fixed the TBB static linking issue in Link Line Advisor.
MKLD-11372 Fixed the issue with incorrect parameter 3 in DGEMM when using the compler option -qmkl.
MKLD-11370 Fixed the issue that cause FFT function to behave in ILP64 mode while in the lP64 mode.
MKLD-10903 Fixed issue CLUSTER_SPARSE_SOLVER_EXPORT reporting wrong index mode.
MKLD-10772 Fixed issue with MKL-SPARSE_S_MM in BSR being slower than in CSR.

2021.3 Release (June 2021)

MKLD-11609 Fixed SSYSV_AA returning the wrong optimal size of the work array.
MKLD-11590 Fixed slowdown in CBLAS_SCAL and CBLAS_DCOPY.
MKLD-11440 Fixed PZHEEVD returning wrong results when running with 16 MPI tasks and the matrix size >= 4000.
MKLD-11382 Fixed CLUSTER_SPARSE_SOLVER_EXPORT returning wrong results in the symmetric matrix.
MKLD-11380 Fixed crashed issue when unloading a plugin DLL linking to oneMKL in  an application dynamically linking to oneMKL.
MKLD-11347 Updated oneMKL Link Line Advisor to show -qmkl option.
MKLD-11317 Fixed incorrect paths in the builder tool.
MKLD-11299 Fixed issue with MKL FEAST reporting internal error when computing big problem sizes.
MKLD-11208 Fixed issue with PDSYRK failing in LP64, not in ILP64 mode even though the problem size is within LP64 size limitations.
MKLD-11109 oneMKL is opening libgomp.so instead of libgomp.so.1.
MKLD-11040 Fixed hanging issue in AVX2 version of batch DGEMM.
MKLD-10944 Fixed a DGEMM regression issue when running in multi-threaded applications.
MKLD-10907 Fixed issue with CLUSTER_SPARSE_SOLVER showing significant nonuniform memory distribution across MPI ranks.
MKLD-10904 Fixed issue with CLUSTER_SPARSE_SOLVER spending 85% of reordering time in memory allocation.
MKLD-10884 Fixed issue with MKL_SPARSE_D_MV flipping when using CSC.
MKLD-10874 Fixed issue with MKL_SPARSE_D_MM producing wrong results.
MKLD-10851 Fixed issue with VDRNGGAUSSIAN raising divide-by-zero exception.
MKLD-10578 Added missing description line in verbose mode when using TBB mode.
MKLD-9898 Fixed memory leak issue in DPOTRF and DPOTRI.
MKLD-9189 Fixed issue with makefile.
MKLD-9026 Added CMake support for oneMKL.
MKLD-8345 Fixed verbose mode for ?DTSVB. 
MKLD-3959 Added support for TBB in custom DLL builder.

2021.2 Release (March 2021)

MKLD-11173 Added missing DLL files at runtime with FORTRAN sample in Visual Studio 19.
MKLD-11027 Fixed oneMKL Link Line Advisor ( v6.13) to check boxes for LAPACK95 and BLAS95.
MKLD-10796 Fixed format of the INTEL_MKL_VERSION in mkl_version.h.
MKLD-10629 Fixed nrm2 output NaN on graphics when the input vector was a zero vector.
MKLD-10532 Implemented fix for FFTW producing the wrong results when Intel® AVX-512 is enabled.
MKLD-10488 Fixed zero error in DFEAST_SCSRGV when divided by float.
MKLD-10448 Fixed MKL_ZOMATADD routine producing wrong results
MKLD-10254 Improved MKL PARDISO scalability behavior at the solving stage with many of the RHS.
MKLD-10247 Improved performane of MKL_SPARSE_D_GV when the problem size exceeds.
MKLD-10228 Fixed crash reported for MKL_SPARSE_D_GV.
MKLD-10217 Fixed regression reported for DSYEVX in earlier versions.
MKLD-9070 Cleaned up MKL PARDISO to not use deprecated omp_get/set_nested routines.
MKLD-10307 Enabled the extended precision iterative refinement in MKL PARDISO.
MKLD-10076 Added MKL_DIRECT_CALL mode for ?GEEV routine.
MKLD-10612 Fixed Random Number Generator (RNG) - Numerical Divergence for vslSkipAheadStream.

2021.1 Initial Release (December 2020)

MKLD-10499  Fixed DGESVD function returning essential different results with different number of threads.
MKLD-10397 Implemented Application Note for oneMKL PARDISOINIT to describe that oneMKL PARDISOINIT works with only 4-Byte integers.
MKLD-10359 Fixed inconsistent returns of oneMKL PARDISO_64 and ILP64 API.
MKLD-10136 Fixed different results produced on different branches of the codepath for RNG - vslloadstreamf and vslsavestreamf.
MKLD-10059 Added -fPIC as default option to build BLAS95 and LAPACK95 interfaces.
MKLD-9831 Fixed unexpected return value for Poisson solver for iparm values.
MKLD-9331 Resolved Linking error LNK2038 message while linking SparseBLAS routine with /MD option.
MKLD-8512 Fixed simultaneous data racing issue occured when several streams try to call DataFitting functions for the first time.
MKLD-10188 Updated documentation for ?FGMRES_CHECK and other RCI check routine.