Add AOCL-Sparse to CI (#49)

* Add AOCL-Sparse to CI, update CMakeLists.txt.

* Update AOCL-Sparse backend to support mixed CSR and CSC.

Author: BenBrock

.github/workflows/ci.yml

@@ -94,6 +94,27 @@ jobs:
       run: |
         ./build/test/gtest/spblas-tests
 
+  aocl:
+    runs-on: 'cpu_amd'
+    steps:
+    - uses: actions/checkout@v4
+    - name: Add AOCL-Sparse to Environment
+      run: |
+        source /apps/spacks/current/share/spack/setup-env.sh
+        echo "AOCLSPARSE_DIR=$(spack location -i aocl-sparse)" >> $GITHUB_ENV
+        echo "AOCLUTILS_DIR=$(spack location -i aocl-utils)" >> $GITHUB_ENV
+    - name: CMake
+      run: |
+        source /apps/spacks/current/share/spack/setup-env.sh
+        spack load /ia2365b
+        cmake -B build -DENABLE_AOCLSPARSE=ON
+    - name: Build
+      run: |
+        make -C build -j `nproc`
+    - name: Test
+      run: |
+        ./build/test/gtest/spblas-tests
+
   rocsparse:
     runs-on: 'gpu_amd'
     steps:

CMakeLists.txt

@@ -42,7 +42,24 @@ if (ENABLE_AOCLSPARSE)
     message(FATAL_ERROR "Environment variable AOCLUTILS_DIR must be set when the AOCLSPARSE is enabled.")
   endif()
   target_include_directories(spblas INTERFACE $ENV{AOCLSPARSE_DIR}/include $ENV{AOCLUTILS_DIR}/include)
-  target_link_libraries(spblas INTERFACE $ENV{AOCLSPARSE_DIR}/lib/libaoclsparse.a $ENV{AOCLUTILS_DIR}/lib/libaoclutils.a)
+
+  find_library(AOCLSPARSE_LIBRARY NAMES aoclsparse libaoclsparse
+               PATHS $ENV{AOCLSPARSE_DIR}/lib
+               NO_DEFAULT_PATH)
+
+  find_library(AOCLUTILS_LIBRARY NAMES aoclutils libaoclutils
+               PATHS $ENV{AOCLUTILS_DIR}/lib
+               NO_DEFAULT_PATH)
+
+  if (NOT AOCLSPARSE_LIBRARY)
+    message(FATAL_ERROR "Could not find AOCL-Sparse library in $ENV{AOCLSPARSE_DIR}/lib")
+  endif()
+  if (NOT AOCLUTILS_LIBRARY)
+    message(FATAL_ERROR "Could not find AOCL-Utils library in $ENV{AOCLUTILS_DIR}/lib")
+  endif()
+
+  target_link_libraries(spblas INTERFACE ${AOCLSPARSE_LIBRARY} ${AOCLUTILS_LIBRARY})
+
   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DSPBLAS_ENABLE_AOCLSPARSE")
 endif()
 

include/spblas/vendor/aoclsparse/aocl_wrappers.hpp

@@ -119,6 +119,32 @@ inline aoclsparse_status aoclsparse_create_csr<double>(
   return aoclsparse_create_dcsr(mat, base, M, N, nnz, row_ptr, col_idx, val);
 }
 
+template <class T>
+inline aoclsparse_status
+aoclsparse_create_csc(aoclsparse_matrix* mat, aoclsparse_index_base base,
+                      aoclsparse_int M, aoclsparse_int N, aoclsparse_int nnz,
+                      aoclsparse_int* col_ptr, aoclsparse_int* row_idx,
+                      T* val) {
+  log_warning("matrix creating with this data type is currently not supported");
+  return aoclsparse_status_not_implemented;
+}
+
+template <>
+inline aoclsparse_status
+aoclsparse_create_csc<float>(aoclsparse_matrix* mat, aoclsparse_index_base base,
+                             aoclsparse_int M, aoclsparse_int N,
+                             aoclsparse_int nnz, aoclsparse_int* col_ptr,
+                             aoclsparse_int* row_idx, float* val) {
+  return aoclsparse_create_scsc(mat, base, M, N, nnz, col_ptr, row_idx, val);
+}
+template <>
+inline aoclsparse_status aoclsparse_create_csc<double>(
+    aoclsparse_matrix* mat, aoclsparse_index_base base, aoclsparse_int M,
+    aoclsparse_int N, aoclsparse_int nnz, aoclsparse_int* col_ptr,
+    aoclsparse_int* row_idx, double* val) {
+  return aoclsparse_create_dcsc(mat, base, M, N, nnz, col_ptr, row_idx, val);
+}
+
 template <typename T>
 inline aoclsparse_status
 aoclsparse_export_csr(const aoclsparse_matrix mat, aoclsparse_index_base* base,

include/spblas/vendor/aoclsparse/detail/create_matrix_handle.hpp

@@ -0,0 +1,57 @@
+#pragma once
+
+#include <spblas/vendor/aoclsparse/aocl_wrappers.hpp>
+
+#include <spblas/detail/view_inspectors.hpp>
+
+namespace spblas {
+
+namespace __aoclsparse {
+
+template <matrix M>
+  requires __detail::is_csr_view_v<M>
+aoclsparse_matrix create_matrix_handle(M&& m) {
+  aoclsparse_matrix handle = nullptr;
+  aoclsparse_status status = __aoclsparse::aoclsparse_create_csr(
+      &handle, aoclsparse_index_base_zero, m.shape()[0], m.shape()[1], m.size(),
+      m.rowptr().data(), m.colind().data(), m.values().data());
+
+  if (status != aoclsparse_status_success) {
+    throw std::runtime_error("create_matrix_handle: AOCL-Sparse failed while "
+                             "creating matrix handle.");
+  }
+
+  return handle;
+}
+
+template <matrix M>
+  requires __detail::is_csc_view_v<M>
+aoclsparse_matrix create_matrix_handle(M&& m) {
+  aoclsparse_matrix handle = nullptr;
+  aoclsparse_status status = __aoclsparse::aoclsparse_create_csr(
+      &handle, aoclsparse_index_base_zero, m.shape()[1], m.shape()[0], m.size(),
+      m.colptr().data(), m.rowind().data(), m.values().data());
+
+  if (status != aoclsparse_status_success) {
+    throw std::runtime_error("create_matrix_handle: AOCL-Sparse failed while "
+                             "creating matrix handle.");
+  }
+
+  return handle;
+}
+
+template <matrix M>
+aoclsparse_operation get_transpose(M&& m) {
+  static_assert(__detail::has_csr_base<M> || __detail::has_csc_base<M>);
+  if constexpr (__detail::has_base<M>) {
+    return get_transpose(m.base());
+  } else if constexpr (__detail::is_csr_view_v<M>) {
+    return aoclsparse_operation_none;
+  } else if constexpr (__detail::is_csc_view_v<M>) {
+    return aoclsparse_operation_transpose;
+  }
+}
+
+} // namespace __aoclsparse
+
+} // namespace spblas

include/spblas/vendor/aoclsparse/detail/detail.hpp

@@ -0,0 +1,3 @@
+#pragma once
+
+#include "create_matrix_handle.hpp"

include/spblas/vendor/aoclsparse/spgemm_impl.hpp

@@ -9,11 +9,13 @@
 
 #pragma once
 
-#include "aoclsparse.h"
+#include <aoclsparse.h>
 #include <cstdint>
 
 #include "aocl_wrappers.hpp"
+#include "detail/detail.hpp"
 #include <fmt/core.h>
+#include <spblas/algorithms/transposed.hpp>
 #include <spblas/detail/log.hpp>
 #include <spblas/detail/operation_info_t.hpp>
 #include <spblas/detail/ranges.hpp>
@@ -30,8 +32,9 @@
 namespace spblas {
 
 template <matrix A, matrix B, matrix C>
-  requires(__detail::has_csr_base<A>) &&
-          (__detail::has_csr_base<B>) && __detail::is_csr_view_v<C>
+  requires(__detail::has_csr_base<A> || __detail::has_csc_base<A>) &&
+          (__detail::has_csr_base<B> || __detail::has_csc_base<B>) &&
+          __detail::is_csr_view_v<C>
 operation_info_t multiply_compute(A&& a, B&& b, C&& c) {
   log_trace("");
   auto a_base = __detail::get_ultimate_base(a);
@@ -41,32 +44,27 @@ operation_info_t multiply_compute(A&& a, B&& b, C&& c) {
   using I = tensor_index_t<C>;
   using O = tensor_offset_t<C>;
 
-  aoclsparse_matrix csrA = nullptr;
-  aoclsparse_mat_descr descrA;
+  aoclsparse_matrix csrA = __aoclsparse::create_matrix_handle(a_base);
+  aoclsparse_matrix csrB = __aoclsparse::create_matrix_handle(b_base);
+
+  aoclsparse_operation opA = __aoclsparse::get_transpose(a);
+  aoclsparse_operation opB = __aoclsparse::get_transpose(b);
+
+  aoclsparse_mat_descr descrA = NULL;
   aoclsparse_status status = aoclsparse_create_mat_descr(&descrA);
   if (status != aoclsparse_status_success) {
     fmt::print("\t descr creation failed\n");
   }
   aoclsparse_set_mat_type(descrA, aoclsparse_matrix_type_general);
-  aoclsparse_index_base indexingA = aoclsparse_index_base_zero;
-  aoclsparse_operation opA = aoclsparse_operation_none;
 
-  const index_t a_nrows = __backend::shape(a_base)[0];
-  const index_t a_ncols = __backend::shape(a_base)[1];
-
-  aoclsparse_matrix csrB = nullptr;
-  aoclsparse_mat_descr descrB;
+  aoclsparse_mat_descr descrB = NULL;
   status = aoclsparse_create_mat_descr(&descrB);
   if (status != aoclsparse_status_success) {
     fmt::print("\t descr creation failed\n");
   }
 
   aoclsparse_set_mat_type(descrB, aoclsparse_matrix_type_general);
   aoclsparse_index_base indexingB = aoclsparse_index_base_zero;
-  aoclsparse_operation opB = aoclsparse_operation_none;
-
-  const index_t b_nrows = __backend::shape(b_base)[0];
-  const index_t b_ncols = __backend::shape(b_base)[1];
 
   aoclsparse_matrix csrC = nullptr;
   aoclsparse_index_base indexingC = aoclsparse_index_base_zero;
@@ -76,22 +74,6 @@ operation_info_t multiply_compute(A&& a, B&& b, C&& c) {
   index_t* c_colind = nullptr;
   T* c_values = nullptr;
 
-  const aoclsparse_int nnzA = a_base.rowptr().data()[a_nrows] - indexingA;
-  const aoclsparse_int nnzB = b_base.rowptr().data()[b_nrows] - indexingB;
-
-  status = __aoclsparse::aoclsparse_create_csr(
-      &csrA, indexingA, a_nrows, a_ncols, nnzA, a_base.rowptr().data(),
-      a_base.colind().data(), a_base.values().data());
-  if (status != aoclsparse_status_success) {
-    fmt::print("\t csr matrix A creation failed\n");
-  }
-  status = __aoclsparse::aoclsparse_create_csr(
-      &csrB, indexingB, b_nrows, b_ncols, nnzB, b_base.rowptr().data(),
-      b_base.colind().data(), b_base.values().data());
-  if (status != aoclsparse_status_success) {
-    fmt::print("\t csr matrix B creation failed\n");
-  }
-
   aoclsparse_request request = aoclsparse_stage_nnz_count;
   status =
       aoclsparse_sp2m(opA, descrA, csrA, opB, descrB, csrB, request, &csrC);
@@ -108,20 +90,22 @@ operation_info_t multiply_compute(A&& a, B&& b, C&& c) {
   // Check: csrA and csrB destroyed when the operation_info destructor is
   // called?
 
+  aoclsparse_destroy_mat_descr(descrA);
+  aoclsparse_destroy_mat_descr(descrB);
   return operation_info_t{
       index<>{__backend::shape(c)[0], __backend::shape(c)[1]}, c_nnz,
       __aoclsparse::operation_state_t{csrA, csrB, csrC}};
 }
 
 template <matrix A, matrix B, matrix C>
-  requires(__detail::has_csr_base<A>) &&
-          (__detail::has_csr_base<B>) && __detail::is_csr_view_v<C>
+  requires(__detail::has_csr_base<A> || __detail::has_csc_base<A>) &&
+          (__detail::has_csr_base<B> || __detail::has_csc_base<B>) &&
+          __detail::is_csr_view_v<C>
 void multiply_fill(operation_info_t& info, A&& a, B&& b, C&& c) {
   log_trace("");
 
   auto a_base = __detail::get_ultimate_base(a);
   auto b_base = __detail::get_ultimate_base(b);
-  auto c_base = __detail::get_ultimate_base(c);
 
   using T = tensor_scalar_t<C>;
   using I = tensor_index_t<C>;
@@ -135,35 +119,30 @@ void multiply_fill(operation_info_t& info, A&& a, B&& b, C&& c) {
   aoclsparse_matrix csrC = info.state_.c_handle;
   offset_t c_nnz = info.result_nnz();
 
-  aoclsparse_mat_descr descrA;
+  aoclsparse_operation opA = __aoclsparse::get_transpose(a);
+  aoclsparse_operation opB = __aoclsparse::get_transpose(b);
+
+  aoclsparse_mat_descr descrA = NULL;
   aoclsparse_status status = aoclsparse_create_mat_descr(&descrA);
   if (status != aoclsparse_status_success) {
     fmt::print("\t descr creation failed\n");
   }
 
   aoclsparse_set_mat_type(descrA, aoclsparse_matrix_type_general);
   aoclsparse_index_base indexingA = aoclsparse_index_base_zero;
-  aoclsparse_operation opA = aoclsparse_operation_none;
-
-  const index_t a_nrows = __backend::shape(a_base)[0];
-  const index_t a_ncols = __backend::shape(a_base)[1];
 
-  aoclsparse_mat_descr descrB;
+  aoclsparse_mat_descr descrB = NULL;
   status = aoclsparse_create_mat_descr(&descrB);
   if (status != aoclsparse_status_success) {
     fmt::print("\t descr creation failed\n");
   }
 
   aoclsparse_set_mat_type(descrB, aoclsparse_matrix_type_general);
   aoclsparse_index_base indexingB = aoclsparse_index_base_zero;
-  aoclsparse_operation opB = aoclsparse_operation_none;
-
-  const index_t b_nrows = __backend::shape(b_base)[0];
-  const index_t b_ncols = __backend::shape(b_base)[1];
 
   aoclsparse_index_base indexingC = aoclsparse_index_base_zero;
-  index_t c_nrows = __backend::shape(c_base)[0];
-  index_t c_ncols = __backend::shape(c_base)[1];
+  index_t c_nrows = __backend::shape(c)[0];
+  index_t c_ncols = __backend::shape(c)[1];
   offset_t* c_rowptr = nullptr;
   index_t* c_colind = nullptr;
   T* c_values = nullptr;
@@ -194,6 +173,24 @@ void multiply_fill(operation_info_t& info, A&& a, B&& b, C&& c) {
   if (alpha_optional.has_value()) {
     scale(alpha, c);
   }
+  aoclsparse_destroy_mat_descr(descrA);
+  aoclsparse_destroy_mat_descr(descrB);
+}
+
+template <matrix A, matrix B, matrix C>
+  requires(__detail::has_csr_base<A> || __detail::has_csc_base<A>) &&
+          (__detail::has_csr_base<B> || __detail::has_csc_base<B>) &&
+          __detail::is_csc_view_v<C>
+operation_info_t multiply_compute(A&& a, B&& b, C&& c) {
+  return multiply_compute(transposed(b), transposed(a), transposed(c));
+}
+
+template <matrix A, matrix B, matrix C>
+  requires((__detail::has_csr_base<A> || __detail::has_csc_base<A>) &&
+           (__detail::has_csr_base<B> || __detail::has_csc_base<B>) &&
+           __detail::is_csc_view_v<C>)
+void multiply_fill(operation_info_t& info, A&& a, B&& b, C&& c) {
+  multiply_fill(info, transposed(b), transposed(a), transposed(c));
 }
 
 } // namespace spblas