Merge naive lower triangular solver + updated cuSPARSE solvers

This PR adds a naive lower triangular solver implementation for CUDA to Ginkgo, and updates the cuSPARSE solver wrappers to use the new cuSPARSE 11.* triangular solver. Related PR: #764
ginkgo-project · Dec 14, 2021 · 00a4f67 · 00a4f67
2 parents e780783 + 3916a4b
commit 00a4f67
Show file tree

Hide file tree

Showing 25 changed files with 803 additions and 320 deletions.
diff --git a/core/device_hooks/common_kernels.inc.cpp b/core/device_hooks/common_kernels.inc.cpp
@@ -318,7 +318,6 @@ namespace lower_trs {
 
 
 GKO_STUB(GKO_DECLARE_LOWER_TRS_SHOULD_PERFORM_TRANSPOSE_KERNEL);
-GKO_STUB(GKO_DECLARE_LOWER_TRS_INIT_STRUCT_KERNEL);
 GKO_STUB_VALUE_AND_INDEX_TYPE(GKO_DECLARE_LOWER_TRS_GENERATE_KERNEL);
 GKO_STUB_VALUE_AND_INDEX_TYPE(GKO_DECLARE_LOWER_TRS_SOLVE_KERNEL);
 
@@ -330,7 +329,6 @@ namespace upper_trs {
 
 
 GKO_STUB(GKO_DECLARE_UPPER_TRS_SHOULD_PERFORM_TRANSPOSE_KERNEL);
-GKO_STUB(GKO_DECLARE_UPPER_TRS_INIT_STRUCT_KERNEL);
 GKO_STUB_VALUE_AND_INDEX_TYPE(GKO_DECLARE_UPPER_TRS_GENERATE_KERNEL);
 GKO_STUB_VALUE_AND_INDEX_TYPE(GKO_DECLARE_UPPER_TRS_SOLVE_KERNEL);
 

diff --git a/core/solver/lower_trs.cpp b/core/solver/lower_trs.cpp
@@ -55,7 +55,6 @@ namespace {
 
 
 GKO_REGISTER_OPERATION(generate, lower_trs::generate);
-GKO_REGISTER_OPERATION(init_struct, lower_trs::init_struct);
 GKO_REGISTER_OPERATION(should_perform_transpose,
                        lower_trs::should_perform_transpose);
 GKO_REGISTER_OPERATION(solve, lower_trs::solve);
@@ -85,19 +84,11 @@ std::unique_ptr<LinOp> LowerTrs<ValueType, IndexType>::conj_transpose() const
 }
 
 
-template <typename ValueType, typename IndexType>
-void LowerTrs<ValueType, IndexType>::init_trs_solve_struct()
-{
-    this->get_executor()->run(lower_trs::make_init_struct(this->solve_struct_));
-}
-
-
 template <typename ValueType, typename IndexType>
 void LowerTrs<ValueType, IndexType>::generate()
 {
     this->get_executor()->run(lower_trs::make_generate(
-        gko::lend(system_matrix_), gko::lend(this->solve_struct_),
-        parameters_.num_rhs));
+        gko::lend(system_matrix_), this->solve_struct_, parameters_.num_rhs));
 }
 
 

diff --git a/core/solver/lower_trs_kernels.hpp b/core/solver/lower_trs_kernels.hpp
@@ -58,15 +58,10 @@ namespace lower_trs {
                                   bool& do_transpose)
 
 
-#define GKO_DECLARE_LOWER_TRS_INIT_STRUCT_KERNEL                  \
-    void init_struct(std::shared_ptr<const DefaultExecutor> exec, \
-                     std::shared_ptr<solver::SolveStruct>& solve_struct)
-
-
-#define GKO_DECLARE_LOWER_TRS_GENERATE_KERNEL(_vtype, _itype)  \
-    void generate(std::shared_ptr<const DefaultExecutor> exec, \
-                  const matrix::Csr<_vtype, _itype>* matrix,   \
-                  solver::SolveStruct* solve_struct,           \
+#define GKO_DECLARE_LOWER_TRS_GENERATE_KERNEL(_vtype, _itype)         \
+    void generate(std::shared_ptr<const DefaultExecutor> exec,        \
+                  const matrix::Csr<_vtype, _itype>* matrix,          \
+                  std::shared_ptr<solver::SolveStruct>& solve_struct, \
                   const gko::size_type num_rhs)
 
 
@@ -80,7 +75,6 @@ namespace lower_trs {
 
 #define GKO_DECLARE_ALL_AS_TEMPLATES                          \
     GKO_DECLARE_LOWER_TRS_SHOULD_PERFORM_TRANSPOSE_KERNEL;    \
-    GKO_DECLARE_LOWER_TRS_INIT_STRUCT_KERNEL;                 \
     template <typename ValueType, typename IndexType>         \
     GKO_DECLARE_LOWER_TRS_SOLVE_KERNEL(ValueType, IndexType); \
     template <typename ValueType, typename IndexType>         \

diff --git a/core/solver/upper_trs.cpp b/core/solver/upper_trs.cpp
@@ -55,7 +55,6 @@ namespace {
 
 
 GKO_REGISTER_OPERATION(generate, upper_trs::generate);
-GKO_REGISTER_OPERATION(init_struct, upper_trs::init_struct);
 GKO_REGISTER_OPERATION(should_perform_transpose,
                        upper_trs::should_perform_transpose);
 GKO_REGISTER_OPERATION(solve, upper_trs::solve);
@@ -85,19 +84,11 @@ std::unique_ptr<LinOp> UpperTrs<ValueType, IndexType>::conj_transpose() const
 }
 
 
-template <typename ValueType, typename IndexType>
-void UpperTrs<ValueType, IndexType>::init_trs_solve_struct()
-{
-    this->get_executor()->run(upper_trs::make_init_struct(this->solve_struct_));
-}
-
-
 template <typename ValueType, typename IndexType>
 void UpperTrs<ValueType, IndexType>::generate()
 {
     this->get_executor()->run(upper_trs::make_generate(
-        gko::lend(system_matrix_), gko::lend(this->solve_struct_),
-        parameters_.num_rhs));
+        gko::lend(system_matrix_), this->solve_struct_, parameters_.num_rhs));
 }
 
 

diff --git a/core/solver/upper_trs_kernels.hpp b/core/solver/upper_trs_kernels.hpp
@@ -58,15 +58,10 @@ namespace upper_trs {
                                   bool& do_transpose)
 
 
-#define GKO_DECLARE_UPPER_TRS_INIT_STRUCT_KERNEL                  \
-    void init_struct(std::shared_ptr<const DefaultExecutor> exec, \
-                     std::shared_ptr<gko::solver::SolveStruct>& solve_struct)
-
-
-#define GKO_DECLARE_UPPER_TRS_GENERATE_KERNEL(_vtype, _itype)  \
-    void generate(std::shared_ptr<const DefaultExecutor> exec, \
-                  const matrix::Csr<_vtype, _itype>* matrix,   \
-                  solver::SolveStruct* solve_struct,           \
+#define GKO_DECLARE_UPPER_TRS_GENERATE_KERNEL(_vtype, _itype)              \
+    void generate(std::shared_ptr<const DefaultExecutor> exec,             \
+                  const matrix::Csr<_vtype, _itype>* matrix,               \
+                  std::shared_ptr<gko::solver::SolveStruct>& solve_struct, \
                   const gko::size_type num_rhs)
 
 
@@ -80,7 +75,6 @@ namespace upper_trs {
 
 #define GKO_DECLARE_ALL_AS_TEMPLATES                          \
     GKO_DECLARE_UPPER_TRS_SHOULD_PERFORM_TRANSPOSE_KERNEL;    \
-    GKO_DECLARE_UPPER_TRS_INIT_STRUCT_KERNEL;                 \
     template <typename ValueType, typename IndexType>         \
     GKO_DECLARE_UPPER_TRS_SOLVE_KERNEL(ValueType, IndexType); \
     template <typename ValueType, typename IndexType>         \