BLAS - MKL: fixing HostBlas calls to handle MKL_INT type #2112
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
MKL redefines the BLAS interface based on how MKL_INT is defined we need to wrap that definition with our own Kokkos Kernels INT type to make both compatible with regular BLAS. applying clang-format
|
Status Flag 'Pre-Test Inspection' - Auto Inspected - Inspection is Not Necessary for this Pull Request. |
ndellingwood
approved these changes
Feb 16, 2024
There was a problem hiding this comment.
Changes look reasonable, replace int in the blas function signatures with KK_INT for added flexibility. Thanks @lucbv !
|
Status Flag 'Pull Request AutoTester' - Failure: Timed out waiting for job KokkosKernels_PullRequest_VEGA90A_Tpls_ROCM560 to start: Total Wait = 3603
|
|
Status Flag 'Pull Request AutoTester' - User Requested Retest - Label AT: RETEST will be reset after testing. |
|
Status Flag 'Pull Request AutoTester' - Testing Jenkins Projects: Pull Request Auto Testing STARTING (click to expand)Build InformationTest Name: KokkosKernels_PullRequest_CUDA11_CUDA11_LayoutRight
Jenkins Parameters
Build InformationTest Name: KokkosKernels_PullRequest_GCC930_Light_Tpls_GCC930_Tpls_CLANG13CUDA10
Jenkins Parameters
Build InformationTest Name: KokkosKernels_PullRequest_GNU1021
Jenkins Parameters
Build InformationTest Name: KokkosKernels_PullRequest_GNU1021_Light_LayoutRight
Jenkins Parameters
Build InformationTest Name: KokkosKernels_PullRequest_Tpls_GNU1021
Jenkins Parameters
Build InformationTest Name: KokkosKernels_PullRequest_Tpls_INTEL19_solo
Jenkins Parameters
Build InformationTest Name: KokkosKernels_PullRequest_CLANG1001_solo
Jenkins Parameters
Build InformationTest Name: KokkosKernels_PullRequest_A64FX_Tpls_ARMPL2110
Jenkins Parameters
Build InformationTest Name: KokkosKernels_PullRequest_A64FX_GCC1020
Jenkins Parameters
Build InformationTest Name: KokkosKernels_PullRequest_VEGA90A_ROCM560
Jenkins Parameters
Build InformationTest Name: KokkosKernels_PullRequest_VEGA90A_Tpls_ROCM560
Jenkins Parameters
Using Repos:
Pull Request Author: lucbv |
|
Status Flag 'Pull Request AutoTester' - Jenkins Testing: all Jobs PASSED Pull Request Auto Testing has PASSED (click to expand)Build InformationTest Name: KokkosKernels_PullRequest_CUDA11_CUDA11_LayoutRight
Jenkins Parameters
Build InformationTest Name: KokkosKernels_PullRequest_GCC930_Light_Tpls_GCC930_Tpls_CLANG13CUDA10
Jenkins Parameters
Build InformationTest Name: KokkosKernels_PullRequest_GNU1021
Jenkins Parameters
Build InformationTest Name: KokkosKernels_PullRequest_GNU1021_Light_LayoutRight
Jenkins Parameters
Build InformationTest Name: KokkosKernels_PullRequest_Tpls_GNU1021
Jenkins Parameters
Build InformationTest Name: KokkosKernels_PullRequest_Tpls_INTEL19_solo
Jenkins Parameters
Build InformationTest Name: KokkosKernels_PullRequest_CLANG1001_solo
Jenkins Parameters
Build InformationTest Name: KokkosKernels_PullRequest_A64FX_Tpls_ARMPL2110
Jenkins Parameters
Build InformationTest Name: KokkosKernels_PullRequest_A64FX_GCC1020
Jenkins Parameters
Build InformationTest Name: KokkosKernels_PullRequest_VEGA90A_ROCM560
Jenkins Parameters
Build InformationTest Name: KokkosKernels_PullRequest_VEGA90A_Tpls_ROCM560
Jenkins Parameters
|
|
Status Flag 'Pre-Merge Inspection' - SUCCESS: The last commit to this Pull Request has been INSPECTED AND APPROVED by [ ndellingwood ]! |
|
Status Flag 'Pull Request AutoTester' - Pull Request MUST BE MERGED MANUALLY BY Project Team - This Repo does not support Automerge |
lucbv
added a commit
that referenced
this pull request
Apr 8, 2024
* HIP: since Kokkos has moved it out of experimental we should clean up
Just reflecting the move of HIP and HIPSpace out of experimental
so that we do not get deprecation warning and even failures down
the road. This was really done in Kokkos Core 4.0.0 so it is time
to catch up...
* Applying clang-format
* Sparse: fix cusparse spgemm hang properly
The issue is fixed by disabling the TPL in spec_avail
when a problematic version of CUDA/cuSPARSE is being used.
* Sparse: fix logic for bad cursparse spgemm version.
Just inverted the logic statement to enable the TPL when it is
known to work correctly.
* Improvements on the unification attempt logic for axpby(), including new tests
* Addressing feedbacks from Luc, plus some small changes here and there:
In KokkosBlas1_axpby_unification_attempt.hpp:
- Removed unnecessary variables, routines, and checks
- Imposed terminology consistency: variable names begin with lower case letters, type names begin with upper case letters
- Using static_assert as much as possible
- Using 'public' and 'private' keywords accordingly
- Improved some explanations and error messages
In KokkosBlas1_axpby_spec.hpp:
- Replace 'a' and 'b' by 'scalar_x' and 'scalar_y' where appropriate, to keep consistency with the terminology used in the 'impl' and 'mv_impl' files of the axpby operation.
- Not using the 'KOKKOSBLAS_OPTIMIZATION_LEVEL_AXPBY' define anymore. Code is now consistent with the 'old' value 3 for such define.
In KokkosBlas1_axpby_impl.hpp and KokkosBlas1_axpby_mv_impl.hpp:
- Not using the 'KOKKOSBLAS_OPTIMIZATION_LEVEL_AXPBY' define anymore. Code is now consistent with the 'old' value 3 for such define.
- Using 'if constexpr' whenever possible
- Checking that -1 <= scalar_x <= 2 and that -1 <= scalar_y <= 2
- Replaced '} else {' by '} else if (scalar_x == 2)' or by '} else if (scalar_y == 2)', whenever possible
- Improved error messages
- Improved explanation headers a bit
In KokkosBlas1_axpby.hpp:
- Renamed some variables to more meaningful names
* Formatting
* Using 'ifdef HAVE_KOKKOSKERNELS_DEBUG', per Luc's suggestion
* Addressing feedbacks from Luc
* Correcting compilation errors in my Mac
* Backup
* SYR2: fix unit-test type issue
On KokkosEco_Trilinos_Weaver_CUDA112_opt-uvm the SYR2 test
enerates a compile time error probably due to a mixed use of host
and device views when comparing implemented vs. reference results.
* CUDA 11.0.1 / cuSPARSE 11.0.0 changed SpMM enums
* SYR2: applying clang-format
* CUDA 11.2.1 / cuSPARSE 11.4.0 changed SpMV
* KokkosBlas1_axpby: include <iostream> for debug builds
Resolve compilation errors in debug mode:
"error: no member named 'cout' in namespace 'std';"
* Backup
* Backup
* Backup
* Backup
* Backup
* Backup
* Backup
* Backup
* Backup
* Backup
* Backup
* Backup
* Address CI build errors
* Some cleanup on current pull request, making it more related to 'just' the creation of the lapack subdirectory and the moving of some files to there
* More cleanup
* Re-enabling gesv unit tests under the lapack subdirectory
* Adding BLAS routines back, for backwards compatibility
* Formatting
* Small cleaning
* Correcting error in Jenkins
* Fixing compilation error on Jenkins when dealing with HIP
* Add required rtd conf file
* README.md: Use correct project slug
* docs/requirements.txt: Add sphinx-rtd-theme
* Addressing latest feedbacks from Luc.
* Formatting
* KokkosKernelsConfig.cmake: add all_libs target and necessary aliases
* Intent of these changes is to allow for building Trilinos with
KokkosKernels as an external TPL
* hide native merge-path SpMV behind "native-merge"
* test native-merge algorithm
* Quick fix for night compilation with Trilinos
* SPTRSV: check if cusparse is available before calling TPL path
Since SpTRSV does not implement the TPL layer the usual way we need
to be extra careful before calling the TPL implementation path. If
cusparse is not available then we definitely want to revert back to
calling the native implementation. Similarly, if the execution space
is not Kokkos::CUDA, let's use the native implementation.
* SpTRSV: more strickly check prerequisites in SptrsvHandle
Check that CUSPARSE is enabled and that HandleExecSpace is
Kokkos::CUDA before allowing users to set the implementation to use
the CUSPARSE TPL.
* SpTRSV: fix some type definition and variable usaged for cuSPARSE
Since we are guarding the cusparse path a bit better we need to be
careful when some types are defined and to mark some variables
(void) when they do not get used by an implementation...
* SpTRSV: applying clang-format
* SpTRSV: more fixes
* SpTRSV: apply clang-format
* SYCL: fix for Trilinos build with MKL
* Apply clang-format to non-cmake files
* SYR2: fix issue with bad type in test function
After comparing various function signatures and view types, the change
allows tests to pass correctly and seem correct based on input params.
* Update Test_Blas2_syr2.hpp
Fix mistake in host/device view argument
* LAPACK: adding rocsolver TPL
Adding the necessary CMake logic and TPL layer to support rocsolver
for LAPACK. Enabling the TPL in gesv and updating gesv test to run
by default the more common configurations and only run specific ones
when the associated TPL (MAGMA) is enabled.
* Lapack: change according to Brian's review
The SpaceAccessibility of IPIVV needs to be modified for MAGMA.
The value_type of IPIVV needs to be rocblas_int when running with
rocSOLVER.
The types used for gesv_tpl_spec_avail and the actual TPL
instantiation where mismatched leading to linker error.
* cmake/Dependencies.cmake: remove ROCSOLVER
Removing ROCSOLVER to prevent configuration errors with Trilinos
Will bring back when support is added in Trilinos for ROCSOLVER as TPL
* Lapack: cusolver TPL logic and support for gesv
Adding CMake logic to support cusolver and implementing gesv using
cusolver getrf and getrs. Unit-test is passing without problems!
* Lapack: updating logic in cm_generate_makefile for cusolver
There is some specific TPL logic in cm_generate_makefile and it
cannot be found for cusolver, changing that might to the trick!
* Backup
* Backup
* Backup
* Formatting
* mv_unification tests with double are failing by very small amounts, e.g. 5.9e-14 vs. 3.6e-14
* Trying one more increment on tolerance
* Putting pragma's and unrolls properly right before for loops (compilation warning at weaver)
* Giving it another try to larger tolarance, after fixing the warning on pragma and unroll
* Lapack: gesv, implementing review commments
* Adding Changelog for Release 4.2.0 (#2031)
* Adding Changelog for Release 4.2.0
Part of Kokkos C++ Performance Portability Programming EcoSystem 4.2
* Formatting the changelog a bit more
Mentioning more clearly LAPACK vs BLAS, grouping PRs by logical work unit, etc...
* Remove minor revisions, improve text descriptions
* Changelog: add spmv perftest detail
---------
Co-authored-by: Luc Berger <lberge@sandia.gov>
Co-authored-by: Carl Pearson <cwpearson@users.noreply.github.com>
Co-authored-by: brian-kelley <brian.honda11@gmail.com>
* NRM1: refactoring TPL layer a bit with c++17 if constexpr
Hopefully this leads to simpler code, less duplication, less
macro and easier maintenance!
Adding support for oneapi MKL while making tpl layer changes.
* BLAS: Nrm1 implementing Brian's feedback
* Blas: nrm1, fix in tpl spec decl
* BLAS: nrm1 problems with ExecSpace template and lack of Kokkos::Threads
Fix issue with Kokkos::Threads and Kokkos::HIP
* Another attempt while waiting to get access to the solo cluster
* Formatting
* Correction error from the last commit
* Fixing the error that was happening only at the solo cluster
* Increase tolerance a bit more
* ncreasing tolerances in all 4 locations
* Backup
* Backup
* Formatting
* Forgot to add ClusteringAlgorithm:: at some spots
* Formatting
* Lapack: fixing issue with Magma TPL in gesv, trtri, etc...
Adding proper support for MAGMA after having it moved to the Lapack
directory and checking it does not create issues with cuSOLVER.
* Update blas/unit_test/Test_Blas1_swap.hpp
Co-authored-by: brian-kelley <brian.honda11@gmail.com>
* cmake: Add workaround check for CUSOLVER support with Trilinos
TPL_ENABLE_CUDA default enables CUBLAS and CUSOLVER in Trilinos, but not CUSPARSE
This PR modifies the TPL requirement checks to maintain compatibility with existing configration options of Trilinos
Attempt to resolve/workaround issue #2047
* Addressing Brian Kelley's feedbacks
* Formatting
* Removing 'ClusteringAlgorithm::'
* Lapack: gesv, incorporate Brian's feedback
* Applying clang-format
* Fixing some deprecation warnings/errors for ROCm 6
* BLAS: fix bug in TPL layer of KokkosBlas::swap
The cuBLAS Kokkos::complex<float> specialization had a small bug
where the rank of the view was not specified correctly!
* CMake: fix bugs in deciding KOKKOSKERNELS_TPL_BLAS_RETURN_COMPLEX
* TPL: revise BLAS1 dot implementation
* Fix compile errors for C-linkage dot functions returning std::complex
* Use a C struct for complex numbers
to avoid error: '_Complex' is a C99 extension [-Werror,-Wc99-extensions].
* Add a workaround by disabling host MKL dot with complex numbers
* Allow KokkosKernels_ENABLE_PERFTESTS=ON to build perf_tests without KokkosKernels_ENABLE_TESTS=ON
* format sparse/tpls/KokkosSparse_spmv_tpl_spec_decl.hpp
* cmake: fix tpl check so cusolver can be disabled when needed
* Link std::filesystem for IntelLLVM in perf_test/sparse
* gemm3 perf test: user CUDA, SYCL, or HIP device for kokkos:initialize
* Fix for rocm_verison header inclusion
* fence Kokkos before timed interations
* Deprecate KOKKOSLINALG_OPT_LEVEL
* Add CMake warning message if KokkosKernels_LINALG_OPT_LEVEL is used
* Async matrix release for MKL >= 2023.2
* Support CUBLAS_{LIBRARIES,LIBRARY_DIRS,INCLUDE_DIRS,ROOT} and KokkosKernels_CUBLAS_ROOT
* KokkosSparse_spmv_impl_merge.hpp: use capture by reference
Resolve warnings in builds with c++20 support enabled:
"kokkos-kernels/sparse/impl/KokkosSparse_spmv_impl_merge.hpp:166:81: warning: implicit capture of 'this' via '[=]' is deprecated in C++20 [-Wdeprecated]"
* KokkosSparse_par_ilut_numeric_impl.hpp: use capture by reference
Resolve warnings in builds with c++20 support enabled:
"kokkos-kernels/sparse/impl/KokkosSparse_par_ilut_numeric_impl.hpp(591):
warning #2908-D: the implicit by-copy capture of "this" is deprecated"
* Backup
* Backup
* Backup
* Backup
* Formatting
* Correcting compilation error
* Typo
* Changes for syr and syr2, to be tested at weaver
* Formatting
* Changes for axpby
* Backup
* Formatting
* Just to force new checking tests in github
* Addressing feedback from Luc.
* Don't call optimize_gemv for one-shot spmv
* Add HIPManagedSpace support
- CMake option for ETI
- Run unit tests with a Kokkos::Device, not just Kokkos::HIP
- Like we do for Cuda
- Still use HIPSpace unless Managed is the only enabled memspace
- Couple of minor fixes
- Allow querying free HIPManagedSpace memory for SpGEMM
- Disable VBD coloring (not a huge deal, had to do same on CUDA)
- Use correct memory space in SpTRSV solve
* Backup
* Backup
* Backup
* Minor typo
* Add block support to all SPILUK algorithms (#2064)
* Interface for block iluk
* Progress. Test hooked up
* Progress on test refactoring
* More test reorg
* Fix test
* Refactor spiluk numeric a bit with a struct wrapper
* Add good logging
* progress
* Fix block test
* Progress but potential dead end
* Giving up on this approach for now
* progress
* Make verbose
* Progress
* Progress
* RP working?
* Progress on TP alg
* Bug fix
* Progress on template stuff
* Progress on block TP
* Progress
* Get rid of all the static_casts
* More cleanup. Steams now support blocks
* Tests not passing
* Serail tests all working, both algs, blocked
* Remove output coming from spiluk test
* Final fixes for CPU
* Cuda req full template specification for SerialGemm::invoke
* Don't use scratch for now
* Formatting
* Fix warnings
* Formatting
* Add tolerance to view checks. Use macro and remove redundant test util
* Fix for HIP
* formatting
* Another test reorg to fix weirdness on solo
* formatting
* Remove unused var
* Github feedback
* Remove test cout
* formatting
* Zero-size arrays can cause problems
* Fix unused var warning
* Add CUDA/HIP TPL support for KokkosSparse::spadd (#1962)
* spadd: change arguments to ctor of SPADDHandle
add a default value to input_sorted;
add a second argument input_merged to indicate unqiue entries;
So that we can easily know whether we can use TPLs on the input matrices
* spadd: add cuda/rocm TPL support for spadd_symbolic/numeric
* Make spiluk_handle::reset backwards compatible (#2087)
* Make spiluk_handle::reset backwards compatible
By making block_size default to -1, which means don't change
block size.
* Switch default val for block_size for reset_handle
* formatting
* Fix comment
* spadd: add APIs without an execution space argument (#2090)
* Lapack - SVD: adding initial files that do not implement anything (#2092)
Adding SVD feature to Lapack component, the interface is similar
to classic Lapack and the implementation relies on the TPL layer
to provide initial capabilities. The TPL supported are LAPACK,
MKL, cuSOLVER and rocSOLVER.
Testing three analytical cases 2x2, 2x3 and 3x2 and then some
randomly generated matrices.
* Hands off namespace `Kokkos::Impl` - cleanup couple violations that snuck in (#2094)
* Do not use things from namespace Kokkos::Impl (Kokkos::{Impl:: -> }ALL_t)
* Do not use things from namespace Kokkos::Impl (Kokkos::Impl::DeepCopy)
Can achieve the same with Kokkos::deep_copy
* Fix warning `declaration of ‘std::size_t n’ shadows a parameter`
* Change name of yaml-cpp to yamlcpp
* Fix macro setting in CMakeLists
* GMRES: Add support for BSR matrices
Also, add a test for this.
* Remove all mentions of HBWSpace
* Reintroduce EXECSPACE_(SERIAL,OPENMP,THREADS}_VALID_MEM_SPACES
Drop HBWSPACE as an option
* Lapack: adding svd benchmark
Fixing unit-test for CUSOLVER and adding benchmark to check the
algorithm performance on various platforms.
* Fix Cuda TPL finding (#2098)
- Allow finding cusparse, cusolver based on manually provided paths
- This is necessary when using an nvhpc toolchain instead of a
standard cuda toolchain
- Set header paths correctly (this is redundant in a cuda installation,
in which $CUDA_ROOT/include is already a system include dir, but
needed in other cases)
* Add support for BSR matrices to some trsv routines (#2104)
* Add support for BSR matrices to some trsv routines
* Change trsv to gesv
* Lapack - SVD: adding quick return when cuSOLVER is skipped (#2107)
Currently we still run the tests on U, S and Vt which does not
make sense since we actively skip this test because cuSOLVER does
not support more columns than rows...
* Fix build error in trsv on gcc8
* Add a workaround for compilation errors with cuda-12.2.0 + gcc-12.3 (#2108)
On Perlmutter@NERSC, I met this error
/usr/lib64/gcc/x86_64-suse-linux/12/include/avx512fp16intrin.h(38): error: vector_size attribute requires an arithmetic or enum type
typedef __half __v8hf __attribute__ ((__vector_size__ (16)));
The workaround was mentioned at https://forums.developer.nvidia.com/t/including-cub-header-breakes-compilation-with-gcc-12-and-sse2-or-better/255018
* Lapack - SVD: fix for unit-test when MKL is enabled (#2110)
This is really a problem with our implementation of the BLAS
interface when MKL is enabled since MKL redefines the function
signatures of blas functions using MKL_INT instead if int...
* Revert "Merge pull request #2037 from ndellingwood/remove-rocsolver-optional-dependency" (#2106)
This reverts commit 5a36d57, reversing
changes made to 2c66d29.
* Fixing missing inclusion in source file
* BLAS - MKL: fixing HostBlas calls to handle MKL_INT type (#2112)
MKL redefines the BLAS interface based on how MKL_INT is defined
we need to wrap that definition with our own Kokkos Kernels INT
type to make both compatible with regular BLAS.
applying clang-format
* Fix weird Trilinos compiler error
It seemed to have a problem with these deep_copies, so just do
the copy by hand like it was being done before my recent trsv
PR.
* Update changelog
* Update changelog
* Block spiluk follow up (#2085)
* Fix for gemm
* Remove unused divide method
* Enhancements to spiluk test
* Progress. Block spiluk now checks out against analytical results
* LUPrec test with spiluk woring
* Disable spiluk LU test on non-host
* Enhancements to spiluk test
* Clean up a few issues uncovered by gh review
* github workflows: update to v4 (use Node 20)
* Refactor Test_Sparse_sptrsv (#2102)
* Refactor Test_Sparse_sptrsv
* More cleanups
* Remove old commented-out code
* CMake: error out in certain case (#2115)
Graph unit tests are unique in that they use default_scalar for the
KokkosKernelsHandle. So if test-eti-only is ON, but neither float nor
double is instatiated, then error out for the graph unit tests.
Users can still build without float or double if they want, but only if
they turn off tests or the graph component.
* Wiki examples for BLAS2 functions are added (#2122)
Some small additional change the the function headers themselves
to add some missing header file inclusions.
Applying clang-format
Removing constexpr since it won't happen before some work in Core.
* Increase tolerance on gesv test (Fix #2123) (#2124)
And uncomment the verbose output for when tolerance is exceeded,
since that helps debug this sort of issue.
This is only printed at most once so it won't spam the output if
the entire vector is wrong.
* Spmv handle (#2126)
* spmv handle, TPL reuse
* using handle in unification layer and hooking up new algorithm
enums with old Controls options
* Update spmv_merge perf test
Compare KK merge vs. default and KK native
* Small changes to help text of spmv_merge perf test
* Complete backwards compatibility with Controls interface
- copy over spmv algorithm selection correctly
- copy expert tuning parameters
* Controls spmv: accept other name for bsr algo
* bsr spmv test: disable tensor core
It was not actually being run before due to a different name
actually enabling it (experimental_bsr_tc rather than experimental_tc)
* Disable OneMKL spmv for complex types
oneapi 2023.2 throws error saying complex isn't supported
* OneMKL: call optimize_gemv during setup
* Option to apply RCM reordering to extracted CRS diagonal blocks (#2125)
* Add rcm option when extracting diagonal blocks
* Update kk_extract_diagonal_blocks_crsmatrix_sequential
* Add test for extracting diagonal blocks with rcm
* Update RCM checking
* cm_test_all_sandia: various updates
- updates for blake
* cm_test_all_sandia: drop decommissioned/unavailable machines
- remove voltrino, mayer
* Fix2130 (#2132)
* Fix #2130
- Do not call BsrMatrix spmv impl if block size is 1
- Instead, convert it to unmanaged CrsMatrix and call spmv again
- cuSPARSE returned an error code in this case
- Better performance
* Formatting
* Remove redundant remove_pointer_t
Handle is already a non-pointer type
* Benchmark: modifying spmv benchmark to run range of spmv tests (#2135)
This could be further automated to run on matrix from suite sparse
* Kokkos Kernels: update version guards to drop old version of Kokkos (#2133)
Since we are now in the 4.2 series we only support up to 4.1.00.
Older version of Kokkos Core will require older version of Kokkos
Kernels for compatibility. Once 4.3.00 is out we will move to
drop support for the 4.1 series and only keep 4.2 and 4.3 series.
* ODE: BDF methods (#1930)
* ODE: adding BDF algorithms
Implementing BDF formula for stiff ODEs.
Orders 1 to 5 are available and tested.
The integrators can be called on GPU to
solve multiple systems in parallel.
* ODE: fixing storage handling for start-up RK stack
* ODE: clang-format
* ODE: first adaptive version of BDF
The current implementation only allows for adaptivity in time,
at this point the BDF Step actually converges as expected with
first order integration!
* ODE: fixing issues with adaptive BDF
The unit-test BDF_adaptive now shows the integration
of the logistic equation using adaptive time steps and
increasing integration order from 1 to 5.
* ODE: running BDF on StiffChemistry problem
The problem runs fine and is solved but there are oscillations
while the behavior of the solution is smooth. More investigation
is needed...
* BDF: fixing types and template parameters in batched calls
Bascially we need template parameters to be more versatile
and cannot assume that all rank1 views will have the exact
same underlying type, for instance layouts can be different.
* More fixes for GPUs only in tests this time.
* ODE: BDF adaptive, fix small bug
After adding rhs and update vectors to temp the subviews taken for
other variables need to be offset appropriately...
* Revert "More fixes for GPUs only in tests this time."
This reverts commit 2f70432.
* Revert "Revert "More fixes for GPUs only in tests this time.""
This reverts commit 836012b.
* ODE: BDF small change to temporarily avoid compile time issue
True fix involving a KOKKOS_VERSION check is upcoming after more
tests on GPU side...
* ODE: BDF fix for some printf statements that will go away soon...
* ODE: adding benchmark for BDF
The benchmark helps us monitor the performance of the BDF
implementaiton across multiple platforms as well as impact of
changes over time.
* ODE: improve benchmark interface...
* ODE: BDF changes to use RMS norm and change some default values
Small changes to compare more closely with reference implementation.
Some of these might be reverted eventually but that's fine for now.
* ODE: BDF convergence more stable and results look pretty good now!
Changing the Newton solver convergence criteria as well as changing
a few default input parameters leads to a more stable algorithms
which can now integrate the stiff Henderson autocatalytic example
well in 66 time steps instead of 200k for fixed order integration...
* ODE: BDF fix bug in initial time step calculation
The initial step routine was overwriting the initial right hand side
which led to obvious issues further down the road... now things should
work fine. Need to figure out if I can re-initialize the variables in
the perf test while excluding that time from each iteration.
* ODE: BDF removing bad print statement...
std::cout in device code
* ODE - BDF: improving perf test
Basically adding new untimed setup within the main loop of the
benchmark to reset the intial conditions, buffers and vectors
ahead of each iteration.
* Modifying unit-test to catch proper return type
* Applying clang-format
* cm_test_all_sandia: update caraway compilers
add rocm/5.6.1 and rocm/6.0.0, and openblas/0.3.23 as tpl
* Sparse MKL: changing the location of the MKL_SAFE_CALL macro (#2134)
* Sparse MKL: changing the location of the MKL_SAFE_CALL macro
Moving the macro outside of namespaces to ensure that it will be
interpreted correctly when called from any other location in the
library.
It does not make much sense to guard Impl code in the Experimental
namespace and in this case it cleans up a problem with namespace
disambiguation for the compiler...
* Sparse BsrSpMV: removing Experimental namespace from Impl namespace
* Applying clang-format
* Sparse SpMV: fixing more namespace issues!
* Fixing missing descriptor for bsr spmv
* Kokkos Kernels: change the default offset ETI from size_t to int (#2140)
This change makes it easier for customer to leverage TPL support
which almost always requires offset=int, ordinal=int to be enabled
meaning that no TPL support is available with our default ETI...
* KokkosSparse_spmv_bsrmatrix_spec: fix Bsr_TC_Precision namespacing
Resolve compilation errors in nightly cuda/12.2 A100 build
* Drop comment for cleaner clang-format fix
* Fix usage of RAII to set cusparse/rocsparse stream (#2141)
Temporary objects like "A()" get destructed immediately.
For the object to have scope lifetime, it needs a name like "A a();".
This was causing cusparse/rocsparse spmv to always execute on the default stream,
causing incorrect timing in the spmv perf test.
* Use execution space operator== (#2136)
It actually is part of the public interface
* cm_test_all_sandia: more caraway module updates and cleanup (#2145)
* Spmv perftest improvements (#2146)
* Spmv perf test improvements
- Add option to flush caches by filling a dummy buffer between
iterations
- Add option to call the non-reuse interface instead of handle/reuse
interface
- Fix modes T, H in nonsquare case (make x,y the correct length)
* Fix mode help text
* Update version to 4.3.0
* Revert "Kokkos Kernels: change the default offset ETI from size_t to int (#2140)"
This reverts commit 3a5498d.
* Fix signed/unsigned comparison warnings (#2150)
This is only hit when spmv is called with integer scalars,
which doesn't happen in our CI but does often in Tpetra.
* SPMV tpl fixes, cusparse workaround (#2152)
* SPMV tpl fixes, workaround
* Avoid possible integer conversion warnings
* Document cusparseSpMM algos that were tested
* Merge pull request #2147 from lucbv/KK_Utils_cleanup
KokkosKernels Utils: cleaning the zero_vector interface
(cherry picked from commit 363868e)
* KokkosBlas1_axpby.hpp: change debug macro guard for printInformation (#2157)
* KokkosBlas1_axpby.hpp: change debug macro guard for printInformation
- resolves test failures in Trilinos (MueLu) that rely on gold file diff
comparisons by removing extra output in debug builds
* fix compilation error
* Update changelog for 4.3.00 (#2148)
* Update changelog for 4.3.00
* Update CHANGELOG.md
---------
Co-authored-by: Luc Berger <lberge@sandia.gov>
* FIx changelog typo
* Fix merge artifacts
* CMakeLists.txt: fix Kokkos_VERSION check
* Merge pull request #2165 from ndellingwood/test-updates
Updates from feedback runnig Trilinos testing
(cherry picked from commit cacba80)
* Update master_history.txt for 4.3.0
* KokkosLapack_svd_tpl_spec_decl: defer to MKL spec when LAPACK also enabled
Resolves redefintion of struct SVD compilation errors with both MKL and LAPACK are enabled
Reported by @maartenarnst in trilinos/Trilinos#12891
Co-authored-by: brian-kelley <brian.honda11@gmail.com>
(cherry picked from commit 5bf5474)
---------
Co-authored-by: Luc Berger-Vergiat <lberge@sandia.gov>
Co-authored-by: Ernesto Prudencio <eeprude@sandia.gov>
Co-authored-by: Carl Pearson <cwpears@sandia.gov>
Co-authored-by: Evan Harvey <eharvey@sandia.gov>
Co-authored-by: Carl Pearson <cwpearson@users.noreply.github.com>
Co-authored-by: brian-kelley <brian.honda11@gmail.com>
Co-authored-by: Sean Miller <sean.miller@amd.com>
Co-authored-by: Junchao Zhang <jczhang@mcs.anl.gov>
Co-authored-by: Junchao Zhang <jczhang@anl.gov>
Co-authored-by: Brian Kelley <bmkelle@sandia.gov>
Co-authored-by: James Foucar <jgfouca@sandia.gov>
Co-authored-by: Damien L-G <dalg24+github@gmail.com>
Co-authored-by: Caleb Schilly <cwschilly@gmail.com>
Co-authored-by: Damien L-G <dalg24@gmail.com>
Co-authored-by: Vinh Dang <vqdang@sandia.gov>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
MKL redefines the BLAS interface based on how MKL_INT is defined we need to wrap that definition with our own Kokkos Kernels INT type to make both compatible with regular BLAS.
This should clean-up significantly some of our nightly oneAPI builds, hopefully the serial+openmp should actually pass now!