Draft: Attempt to specialize matrix_vector_product for parallel_policy #255
Conversation
Changes to the namespace (std::experimental to std) broke the existing unit tests and examples. This commit should fix the problem.
The include guards were failing for certain versions of GCC (such as 11.2.0) without TBB. This commit adds TBB as an optional dependency and disables anything involving execution if the build would fail.
LINALG_HAS_TBB should have been LINALG_ENABLE_TBB. This change causes the only test using executors to fail.
|
FYI, TBB can be built and installed from scratch using the following repo: https://github.com/oneapi-src/oneTBB . @amklinv-nnl tested with GCC 13 and it still requires TBB for parallel algorithms to compile, alas. |
|
@amklinv-nnl Christian Trott explained offline how specializations for different policies work.
I don't have time to try this at the moment, but it sounds like this should fix the recursion issue (because users wouldn't ever pass in one of the internal execution policies). |
| @@ -102,8 +102,6 @@ jobs: | |||
|
|
|||
| test-stdBLAS: | |||
| runs-on: ubuntu-latest | |||
| container: | |||
There was a problem hiding this comment.
This change has already been merged into the main branch.
| @@ -121,6 +121,14 @@ option(LINALG_ENABLE_BLAS | |||
| "Assume that we are linking with a BLAS library." | |||
| ${BLAS_FOUND}) | |||
|
|
|||
| find_package(TBB) | |||
There was a problem hiding this comment.
PR #256 incorporates the changes in this file.
| @@ -28,6 +28,8 @@ Other compilers, including MSVC 2019, have been tested in the past. | |||
| - If you want to build examples, set LINALG_ENABLE_EXAMPLES=ON | |||
| - If you have a BLAS installation, set LINALG_ENABLE_BLAS=ON. | |||
| BLAS support is currently experimental. | |||
| - If you have a TBB installation, set LINALG_ENABLE_TBB=ON. | |||
There was a problem hiding this comment.
PR #256 incorporates the changes in this file.
| using std::experimental::mdspan; | ||
| using std::experimental::extents; | ||
| using std::experimental::dynamic_extent; | ||
| using MDSPAN_IMPL_STANDARD_NAMESPACE::mdspan; |
There was a problem hiding this comment.
An equivalent fix is already in the main branch.
| @@ -30,7 +22,7 @@ int main(int argc, char* argv[]) { | |||
|
|
|||
| // Call linalg::scale x = 2.0*x; | |||
| std::experimental::linalg::scale(2.0, x); | |||
| #ifdef MDSPAN_EXAMPLES_USE_EXECUTION_POLICIES | |||
| #ifdef LINALG_HAS_EXECUTION | |||
@crtrott @dalg24
@amklinv-nnl has been working on parallel specialization of stdblas algorithms. The two of us tried to specialize
matrix_vector_productforstd::execution::parallel_policy, but keep getting run-time recursion. We're guessing that the compiler thinks the genericExecutionPolicy&&overload is "more specialized."The test lives in
tests/native/gemv_no_ambig.cpp. We're building using the following CMake options:Please don't merge this branch, btw; it will almost certainly conflict with other PRs.