Sparse matrix addition (related to Trilinos pull request #1957) #122

brian-kelley · 2017-11-19T08:19:31Z

Same interface as spgemm (separate symbolic/numeric)
Implemented in sparse/KokkosSparse_spadd.hpp
Has associated handle type (like spgemm) and unit tests

Implemented in sparse/KokkosSparse_spadd.hpp Has associated handle type (like spgemm) and unit tests

mndevec · 2018-02-14T21:58:37Z

src/sparse/KokkosSparse_spadd.hpp

+      addHandle->set_a_b_pos(a_pos, b_pos);
+    }
+    //provide the number of NNZ in C to user through handle
+    addHandle->set_max_result_nnz(c_rowmap(nrows));


This causes memory access errors if UVM is not enabled on GPUs. Something like below is needed for this.
auto d_c_nnz_size = Kokkos::subview(c_rowmap, nrows);
auto h_c_nnz_size = Kokkos::create_mirror_view (d_c_nnz_size);
Kokkos::deep_copy (h_c_nnz_size, d_c_nnz_size);
execution_space::fence();
addHandle->set_max_result_nnz(h_c_nnz_size());

mndevec · 2018-02-14T22:22:58Z

src/sparse/KokkosSparse_spadd.hpp

+        Kokkos::parallel_scan(range_type(0, nrows + 1), prefix);
+        //compute uncompressed entries of C (just indices, no scalars)
+        execution_space::fence();
+        c_nnz_upperbound = c_rowmap_upperbound(nrows);


auto d_c_nnz_size = Kokkos::subview(c_rowmap_upperbound, nrows); auto h_c_nnz_size = Kokkos::create_mirror_view (d_c_nnz_size); Kokkos::deep_copy (h_c_nnz_size, d_c_nnz_size); execution_space::fence(); c_nnz_upperbound = h_c_nnz_size();

mndevec · 2018-02-14T22:32:48Z

src/sparse/KokkosSparse_spadd.hpp

+      execution_space::fence();
+      clno_nnz_view_t_ a_pos("A entry positions", a_rowmap(nrows));
+      clno_nnz_view_t_ b_pos("B entry positions", b_rowmap(nrows));
+      //merge the entries and compute Apos/Bpos, as well as Crowcounts


same this for lines 477 and 478. The array cannot be accessed by the host.
Simply, we can use a_entries.dimension_0() here, right?

- Changing the name of the function to spadd to be consistent with the file name. - Minor fixes for UVM spaces.

Minor edits for SpADD function.

crtrott · 2018-02-14T23:32:48Z

Somebody needs to rework this to follow our coding guidelines as well as adding ETI and the TPL hooks etc.

mndevec · 2018-02-14T23:34:11Z

@crtrott
I have the issue listed here:
#151

SpAdd handle was originally a copy-paste of the spgemm handle way back in kokkos#122, and included get_sort_option() and set_sort_option() from spgemm. But these try to use the member bool sort_option, which doesn't exist. Somehow these functions never produced compile errors until someone tried to call them.

SpAdd handle was originally a copy-paste of the spgemm handle way back in #122, and included get_sort_option() and set_sort_option() from spgemm. But these try to use the member bool sort_option, which doesn't exist. Somehow these functions never produced compile errors until someone tried to call them.

SpAdd handle was originally a copy-paste of the spgemm handle way back in kokkos#122, and included get_sort_option() and set_sort_option() from spgemm. But these try to use the member bool sort_option, which doesn't exist. Somehow these functions never produced compile errors until someone tried to call them.

* Restore size_t as default offset, in Tribits builds (#2313) If building KokkosKernels standalone, leave int as the default offset (this was the case since #2140). But if building KokkosKernels as a Trilinos/Tribits package, then make size_t the default offset because this is what the Tpetra stack currently uses. Signed-off-by: Brian Kelley <bmkelle@sandia.gov> * Improve crs/bsr sorting performance (#2293) * CRS sorting improvements - Wrote bulk sort/permutation based sorting for CRS graph, matrix, and BSR matrix (bulk = one large sort of all the entries, using row-major dense index as keys) - This is more performant for imbalanced entries per row - If matrix dimensions are too large to do bulk sort, fall back to sorting within each row with a thread. * Add perf test for sort_crs_matrix * sort_crs: improve parallel labels * Work around kokkos issue 7036 * sort_crs: replace radix sort lambda with functor (Lambda segfaults with nvcc+openmp) --------- Signed-off-by: Brian Kelley <bmkelle@sandia.gov> * SpAdd handle: delete sort_option getter/setter (#2296) SpAdd handle was originally a copy-paste of the spgemm handle way back in #122, and included get_sort_option() and set_sort_option() from spgemm. But these try to use the member bool sort_option, which doesn't exist. Somehow these functions never produced compile errors until someone tried to call them. * Improve GH action to produce release artifacts (#2312) * coo2csr: add parens to function calls (#2318) * Update changelog * Update master_history.txt * .github/workflows: Group jobs under common github-AT2 name (#2320) * Update master_history.txt --------- Signed-off-by: Brian Kelley <bmkelle@sandia.gov> Co-authored-by: brian-kelley <bmkelle@sandia.gov> Co-authored-by: Damien L-G <dalg24@gmail.com> Co-authored-by: Carl Pearson <cwpearson@users.noreply.github.com> Co-authored-by: Evan Harvey <57234914+e10harvey@users.noreply.github.com>

* Set version number to 4.4.01 * [4.4.01] Patches to 4.4.01 (#2327) * Restore size_t as default offset, in Tribits builds (#2313) If building KokkosKernels standalone, leave int as the default offset (this was the case since #2140). But if building KokkosKernels as a Trilinos/Tribits package, then make size_t the default offset because this is what the Tpetra stack currently uses. Signed-off-by: Brian Kelley <bmkelle@sandia.gov> * Improve crs/bsr sorting performance (#2293) * CRS sorting improvements - Wrote bulk sort/permutation based sorting for CRS graph, matrix, and BSR matrix (bulk = one large sort of all the entries, using row-major dense index as keys) - This is more performant for imbalanced entries per row - If matrix dimensions are too large to do bulk sort, fall back to sorting within each row with a thread. * Add perf test for sort_crs_matrix * sort_crs: improve parallel labels * Work around kokkos issue 7036 * sort_crs: replace radix sort lambda with functor (Lambda segfaults with nvcc+openmp) --------- Signed-off-by: Brian Kelley <bmkelle@sandia.gov> * SpAdd handle: delete sort_option getter/setter (#2296) SpAdd handle was originally a copy-paste of the spgemm handle way back in #122, and included get_sort_option() and set_sort_option() from spgemm. But these try to use the member bool sort_option, which doesn't exist. Somehow these functions never produced compile errors until someone tried to call them. * Improve GH action to produce release artifacts (#2312) * coo2csr: add parens to function calls (#2318) * Update changelog * Update master_history.txt * .github/workflows: Group jobs under common github-AT2 name (#2320) * Update master_history.txt --------- Signed-off-by: Brian Kelley <bmkelle@sandia.gov> Co-authored-by: brian-kelley <bmkelle@sandia.gov> Co-authored-by: Damien L-G <dalg24@gmail.com> Co-authored-by: Carl Pearson <cwpearson@users.noreply.github.com> Co-authored-by: Evan Harvey <57234914+e10harvey@users.noreply.github.com> * Update master_history.txt for 4.4.01 --------- Signed-off-by: Brian Kelley <bmkelle@sandia.gov> Co-authored-by: brian-kelley <bmkelle@sandia.gov> Co-authored-by: Damien L-G <dalg24@gmail.com> Co-authored-by: Carl Pearson <cwpearson@users.noreply.github.com> Co-authored-by: Evan Harvey <57234914+e10harvey@users.noreply.github.com>

Sparse matrix addition

a682651

Implemented in sparse/KokkosSparse_spadd.hpp Has associated handle type (like spgemm) and unit tests

brian-kelley mentioned this pull request Nov 19, 2017

[WIP] KokkosKernels: migrating sparse m-m addition from Tpetra trilinos/Trilinos#1957

Closed

crtrott assigned mndevec Jan 2, 2018

mndevec self-requested a review February 14, 2018 21:08

mndevec reviewed Feb 14, 2018

View reviewed changes

mndevec and others added 2 commits February 14, 2018 16:02

Minor edits for SpADD function.

e87110f

- Changing the name of the function to spadd to be consistent with the file name. - Minor fixes for UVM spaces.

Merge pull request #1 from mndevec/bmkelley-develop

d805831

Minor edits for SpADD function.

mndevec requested a review from srajama1 February 14, 2018 23:23

srajama1 approved these changes Feb 14, 2018

View reviewed changes

mndevec merged commit c8d99c8 into kokkos:develop Feb 14, 2018

ndellingwood mentioned this pull request Apr 10, 2018

Specilization layer for sparse matrix addition function #151

Closed

brian-kelley mentioned this pull request Aug 5, 2024

SpAdd handle: delete sort_option getter/setter #2296

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Sparse matrix addition (related to Trilinos pull request #1957) #122

Sparse matrix addition (related to Trilinos pull request #1957) #122

brian-kelley commented Nov 19, 2017

mndevec Feb 14, 2018 •

edited

Loading

mndevec Feb 14, 2018 •

edited

Loading

mndevec Feb 14, 2018 •

edited

Loading

crtrott commented Feb 14, 2018

mndevec commented Feb 14, 2018

Sparse matrix addition (related to Trilinos pull request #1957) #122

Sparse matrix addition (related to Trilinos pull request #1957) #122

Conversation

brian-kelley commented Nov 19, 2017

mndevec Feb 14, 2018 • edited Loading

Choose a reason for hiding this comment

mndevec Feb 14, 2018 • edited Loading

Choose a reason for hiding this comment

mndevec Feb 14, 2018 • edited Loading

Choose a reason for hiding this comment

crtrott commented Feb 14, 2018

mndevec commented Feb 14, 2018

mndevec Feb 14, 2018 •

edited

Loading

mndevec Feb 14, 2018 •

edited

Loading

mndevec Feb 14, 2018 •

edited

Loading