Fix ParILUT threshold select for larger matrices with GPU backends

common/cuda_hip/factorization/par_ilut_select_kernels.cpp

@@ -119,7 +119,7 @@ void threshold_select(std::shared_ptr<const DefaultExecutor> exec,
                            partial_counts, total_counts);
         auto new_bucket = sampleselect_find_bucket(exec, total_counts, rank);
         sampleselect_filter(exec, tmp_in, bucket.size, oracles, partial_counts,
-                            bucket.idx, tmp_out);
+                            new_bucket.idx, tmp_out);
 
         rank -= new_bucket.begin;
         bucket.size = new_bucket.size;

contributors.txt

@@ -3,6 +3,7 @@ license, as specified in the repository's LICENSE file.
 
 Aliaga José I. <aliaga@uji.es> Universitat Jaume I
 Anzt Hartwig <hartwig.anzt@kit.edu> Karlsruhe Institute of Technology, The University of Tennessee Knoxville
+Beams Natalie <nbeams@icl.utk.edu> University of Tennessee, Knoxville
 Boman Erik <egboman@sandia.gov> Sandia National Laboratories
 Castelli Fabian <fabian.castelli@kit.edu> Karlsruhe Institute of Technology
 Chen Yenchen <yanjen224@gmail.com> National Taiwan University

dpcpp/factorization/par_ilut_select_kernel.dp.cpp

@@ -1,4 +1,4 @@
-// SPDX-FileCopyrightText: 2017 - 2024 The Ginkgo authors
+// SPDX-FileCopyrightText: 2017 - 2025 The Ginkgo authors
 //
 // SPDX-License-Identifier: BSD-3-Clause
 
@@ -122,7 +122,7 @@ void threshold_select(std::shared_ptr<const DefaultExecutor> exec,
                            partial_counts, total_counts);
         auto new_bucket = sampleselect_find_bucket(exec, total_counts, rank);
         sampleselect_filter(exec, tmp_in, bucket.size, oracles, partial_counts,
-                            bucket.idx, tmp_out);
+                            new_bucket.idx, tmp_out);
 
         rank -= new_bucket.begin;
         bucket.size = new_bucket.size;

test/factorization/par_ilut_kernels.cpp

@@ -65,6 +65,9 @@ class ParIlut : public CommonTestFixture {
             mtx_size[0], true,
             std::uniform_int_distribution<index_type>(1, mtx_size[0]),
             std::normal_distribution<>(0.0, 1.0), rand_engine, ref);
+        mtx_l3 = gko::test::generate_random_lower_triangular_matrix<Csr>(
+            800, false, std::uniform_int_distribution<index_type>(256, 256),
+            std::normal_distribution<>(-0.667, 3.333), rand_engine, ref);
         mtx_u = gko::test::generate_random_upper_triangular_matrix<Csr>(
             mtx_size[0], false,
             std::uniform_int_distribution<index_type>(10, mtx_size[0]),
@@ -79,6 +82,7 @@ class ParIlut : public CommonTestFixture {
         dmtx_ut_ani = Csr::create(exec);
         dmtx_l = gko::clone(exec, mtx_l);
         dmtx_l2 = gko::clone(exec, mtx_l2);
+        dmtx_l3 = gko::clone(exec, mtx_l3);
         dmtx_u = gko::clone(exec, mtx_u);
 
         std::string file_name(gko::matrices::location_ani4_mtx);
@@ -219,6 +223,7 @@ class ParIlut : public CommonTestFixture {
     std::unique_ptr<Csr> mtx_ut_ani;
     std::unique_ptr<Csr> mtx_l;
     std::unique_ptr<Csr> mtx_l2;
+    std::unique_ptr<Csr> mtx_l3;
     std::unique_ptr<Csr> mtx_u;
 
     std::unique_ptr<Csr> dmtx1;
@@ -230,6 +235,7 @@ class ParIlut : public CommonTestFixture {
     std::unique_ptr<Csr> dmtx_ut_ani;
     std::unique_ptr<Csr> dmtx_l;
     std::unique_ptr<Csr> dmtx_l2;
+    std::unique_ptr<Csr> dmtx_l3;
     std::unique_ptr<Csr> dmtx_u;
 };
 
@@ -246,8 +252,8 @@ TYPED_TEST(ParIlut, KernelThresholdSelectIsEquivalentToRef)
     SKIP_IF_BFLOAT16(value_type);
 #endif
 
-    this->test_select(this->mtx_l, this->dmtx_l,
-                      this->mtx_l->get_num_stored_elements() / 3);
+    this->test_select(this->mtx_l3, this->dmtx_l3,
+                      this->mtx_l3->get_num_stored_elements() / 3);
 }