Add new SpGEMM/SpGEAM interface with reuse capabilities #1934
Conversation
ginkgo-bot
added
reg:testing
- add documentation to multiply_reuse_info - more strict input dimension checks - fix test comments Co-authored-by: Marcel Koch <marcel.koch@kit.edu>
|
@MarcelKoch I added the functionality to a benchmark, a very brief check on a 1Mx1M stencil matrix on an A2 GPU shows a roughly 1.4x speedup over normal SpGEMM |
upsj
changed the title
|
I added advanced_spgemm (multiply_add) and spgeam (add_scale) interfaces, as well as reusable versions thereof with extensive tests of all the size checks and cross-executor functionality. |
|
|
||
| multiply_reuse_info(const multiply_reuse_info&) = delete; | ||
|
|
||
| multiply_reuse_info(multiply_reuse_info&&); |
There was a problem hiding this comment.
This (and all other move constructors/assignments) can be made noexcept
| * add_reuse to recompute a sparse matrix-matrix sum | ||
| * with updated values. | ||
| */ | ||
| class add_scale_reuse_info { |
There was a problem hiding this comment.
nit: the other reuse_info structs are defined before the corresponding non-reuse function.
|
|
||
|
|
||
| template <typename ValueType, typename IndexType> | ||
| Csr<ValueType, IndexType>::multiply_reuse_info::multiply_reuse_info() |
|
|
||
|
|
||
| template <typename ValueType, typename IndexType> | ||
| Csr<ValueType, IndexType>::multiply_add_reuse_info::multiply_add_reuse_info() |
| size_type nnz1; | ||
| size_type nnz2; | ||
| size_type nnz_out; | ||
| // no other data yet |
There was a problem hiding this comment.
nit:
| // no other data yet |
| } | ||
|
|
||
|
|
||
| TEST_F(Csr, MultiplyAddReuseCrossExecutor) |
| #endif | ||
|
|
||
|
|
||
| TEST_F(Csr, MultiplyReuseCrossExecutor) |
| set_up_apply_data<Mtx::classical>(); | ||
| auto trans = gko::as<Mtx>(mtx->transpose()); | ||
|
|
||
| auto [dresult, reuse] = dmtx->multiply_reuse(trans); |
There was a problem hiding this comment.
nit:
| auto [dresult, reuse] = dmtx->multiply_reuse(trans); | |
| auto [dresult, dreuse] = dmtx->multiply_reuse(trans); |
I think this can help to make it clearer in the act section that the device reuse is applied to the host matrix.
| mtx2 = gen_mtx<Mtx>(mtx_size[0], mtx_size[1], 0); | ||
| dmtx = gko::clone(exec, mtx); | ||
|
|
||
| auto [dresult, reuse] = dmtx->add_scale_reuse(alpha, beta, mtx2); |
| } | ||
|
|
||
|
|
||
| TEST_F(Csr, AddScaleReuseCrossExecutor) |
3 participants
This adds a new SpGEMM/SpGEAM interface that works without abusing the
LinOp::applyinterface, and a new SpGEMM kernel that operates on an existing output structure (maybe it also makes sense to expose this as a masked SpGEMM operation? The difference is similar to LU vs. ILU kernels)