Add matrix_opt for cuSPARSE backend

examples/cusparse/CMakeLists.txt

@@ -4,3 +4,4 @@ function(add_cuda_example example_name)
 endfunction()
 
 add_cuda_example(cusparse_simple_spmv)
+add_cuda_example(cusparse_simple_spmm)

examples/cusparse/cusparse_simple_spmm.cpp

@@ -0,0 +1,119 @@
+#include <iostream>
+#include <spblas/spblas.hpp>
+
+#include <cuda_runtime.h>
+
+#include "util.hpp"
+
+#include <fmt/core.h>
+#include <fmt/ranges.h>
+
+int main(int argc, char** argv) {
+  namespace md = spblas::__mdspan;
+
+  using value_t = float;
+  using index_t = spblas::index_t;
+  using offset_t = spblas::offset_t;
+
+  spblas::index_t m = 100;
+  spblas::index_t n = 10;
+  spblas::index_t k = 100;
+  spblas::index_t nnz_in = 10;
+
+  fmt::print("\n\t###########################################################"
+             "######################");
+  fmt::print("\n\t### Running SpMM Example:");
+  fmt::print("\n\t###");
+  fmt::print("\n\t###   Y = alpha * A * X");
+  fmt::print("\n\t###");
+  fmt::print("\n\t### with ");
+  fmt::print("\n\t### A, in CSR format, of size ({}, {}) with nnz = {}", m, n,
+             nnz_in);
+  fmt::print("\n\t### X, a dense matrix, of size ({}, {})", n, k);
+  fmt::print("\n\t### Y, a dense matrix, of size ({}, {})", m, k);
+  fmt::print("\n\t### using float and spblas::index_t (size = {} bytes)",
+             sizeof(spblas::index_t));
+  fmt::print("\n\t###########################################################"
+             "######################");
+  fmt::print("\n");
+
+  auto&& [values, rowptr, colind, shape, nnz] =
+      spblas::generate_csr<value_t, index_t, offset_t>(m, n, nnz_in);
+
+  value_t* d_values;
+  offset_t* d_rowptr;
+  index_t* d_colind;
+
+  CUDA_CHECK(cudaMalloc(&d_values, values.size() * sizeof(value_t)));
+  CUDA_CHECK(cudaMalloc(&d_rowptr, rowptr.size() * sizeof(offset_t)));
+  CUDA_CHECK(cudaMalloc(&d_colind, colind.size() * sizeof(index_t)));
+
+  CUDA_CHECK(cudaMemcpy(d_values, values.data(),
+                        values.size() * sizeof(value_t), cudaMemcpyDefault));
+  CUDA_CHECK(cudaMemcpy(d_rowptr, rowptr.data(),
+                        rowptr.size() * sizeof(offset_t), cudaMemcpyDefault));
+  CUDA_CHECK(cudaMemcpy(d_colind, colind.data(),
+                        colind.size() * sizeof(index_t), cudaMemcpyDefault));
+
+  spblas::csr_view<value_t, index_t, offset_t> a(d_values, d_rowptr, d_colind,
+                                                 shape, nnz);
+
+  // Scale every value of `a` by 5 in place.
+  // scale(5.f, a);
+
+  std::vector<value_t> x(n * k, 1);
+  std::vector<value_t> y(m * k, 0);
+
+  value_t* d_x;
+  value_t* d_y;
+
+  CUDA_CHECK(cudaMalloc(&d_x, x.size() * sizeof(value_t)));
+  CUDA_CHECK(cudaMalloc(&d_y, y.size() * sizeof(value_t)));
+
+  CUDA_CHECK(
+      cudaMemcpy(d_x, x.data(), x.size() * sizeof(value_t), cudaMemcpyDefault));
+  CUDA_CHECK(
+      cudaMemcpy(d_y, y.data(), y.size() * sizeof(value_t), cudaMemcpyDefault));
+
+  md::mdspan x_span(d_x, n, k);
+  md::mdspan y_span(d_y, m, k);
+
+  // Y = A * X
+  spblas::operation_info_t info;
+  spblas::multiply(info, a, x_span, y_span);
+
+  CUDA_CHECK(
+      cudaMemcpy(y.data(), d_y, y.size() * sizeof(value_t), cudaMemcpyDefault));
+
+  // CPU reference
+  std::vector<value_t> y_ref(m * k, 0);
+  for (index_t i = 0; i < m; i++) {
+    for (offset_t j = rowptr[i]; j < rowptr[i + 1]; j++) {
+      index_t col = colind[j];
+      value_t val = values[j];
+      for (index_t l = 0; l < k; l++) {
+        y_ref[i * k + l] += val * x[col * k + l];
+      }
+    }
+  }
+
+  bool failed = false;
+
+  for (size_t i = 0; i < y.size(); ++i) {
+    if (y[i] != y_ref[i]) {
+        fprintf(stderr, "Value mismatch at index %ld: y_ref[%ld] = %f, y[%ld] = %f\n", i, i, y_ref[i], i, y_ref[i]);
+        failed = true;
+    }
+  }
+
+  if (failed) {
+    fmt::print("\tValidation failed!\n");
+  }
+  else {
+    fmt::print("\tValidation succeeded!\n");
+  }
+
+  fmt::print("\tExample is completed!\n");
+
+  return failed;
+}

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

@@ -0,0 +1,49 @@
+#pragma once
+
+#include <cusparse.h>
+
+#include <stdexcept>
+
+#include <spblas/detail/view_inspectors.hpp>
+
+namespace spblas {
+
+namespace __cusparse {
+
+template <matrix M>
+  requires __detail::is_csr_view_v<M>
+cusparseSpMatDescr_t create_matrix_descriptor(M&& m) {
+  cusparseSpMatDescr_t mat_descr;
+  __cusparse::throw_if_error(cusparseCreateCsr(
+      &mat_descr, __backend::shape(m)[0], __backend::shape(m)[1],
+      m.values().size(), m.rowptr().data(), m.colind().data(),
+      m.values().data(), detail::cusparse_index_type_v<tensor_offset_t<M>>,
+      detail::cusparse_index_type_v<tensor_index_t<M>>,
+      CUSPARSE_INDEX_BASE_ZERO, detail::cuda_data_type_v<tensor_scalar_t<M>>));
+
+  return mat_descr;
+}
+
+template <matrix M>
+  requires __detail::is_csc_view_v<M>
+cusparseSpMatDescr_t create_matrix_descriptor(M&& m) {
+  cusparseSpMatDescr_t mat_descr;
+  __cusparse::throw_if_error(cusparseCreateCsc(
+      &mat_descr, __backend::shape(m)[0], __backend::shape(m)[1],
+      m.values().size(), m.rowptr().data(), m.colind().data(),
+      m.values().data(), detail::cusparse_index_type_v<tensor_offset_t<M>>,
+      detail::cusparse_index_type_v<tensor_index_t<M>>,
+      CUSPARSE_INDEX_BASE_ZERO, detail::cuda_data_type_v<tensor_scalar_t<M>>));
+
+  return mat_descr;
+}
+
+template <matrix M>
+  requires __detail::has_base<M>
+cusparseSpMatDescr_t create_matrix_descriptor(M&& m) {
+  return create_matrix_descriptor(m.base());
+}
+
+} // namespace __cusparse
+
+} // namespace spblas

include/spblas/vendor/cusparse/detail/detail.hpp

@@ -0,0 +1,4 @@
+#pragma once
+
+#include "create_matrix_handle.hpp"
+#include "get_matrix_handle.hpp"

include/spblas/vendor/cusparse/detail/get_matrix_handle.hpp

@@ -0,0 +1,44 @@
+#pragma once
+
+#include <cusparse.h>
+
+#include <spblas/detail/log.hpp>
+#include <spblas/detail/operation_info_t.hpp>
+#include <spblas/detail/ranges.hpp>
+#include <spblas/detail/view_inspectors.hpp>
+#include <spblas/views/matrix_opt.hpp>
+
+#include <spblas/vendor/cusparse/detail/create_matrix_handle.hpp>
+
+namespace spblas {
+
+namespace __cusparse {
+
+template <matrix M>
+cusparseSpMatDescr_t
+get_matrix_descriptor(M&& m,
+                      cusparseSpMatDescr_t handle = nullptr) {
+  if constexpr (__detail::is_matrix_opt_v<decltype(m)>) {
+    log_trace("using A as matrix_opt");
+
+    if (m.matrix_descriptor_ == nullptr) {
+      m.matrix_descriptor_ = create_matrix_descriptor(m.base());
+    }
+
+    return m.matrix_descriptor_;
+  } else if constexpr (__detail::has_base<M>) {
+    return get_matrix_descriptor(m.base(), handle);
+  } else if (handle != nullptr) {
+    log_trace("using A from operation_info_t");
+
+    return handle;
+  } else {
+    log_trace("using A as csr_base");
+
+    return create_matrix_descriptor(m);
+  }
+}
+
+} // namespace __cusparse
+
+} // namespace spblas

include/spblas/vendor/cusparse/detail/spmm_state_t.hpp

@@ -0,0 +1,55 @@
+#pragma once
+
+#include <cusparse.h>
+#include <memory>
+
+#include "abstract_operation_state.hpp"
+
+namespace spblas {
+namespace __cusparse {
+
+class spmm_state_t : public abstract_operation_state_t {
+public:
+  spmm_state_t() = default;
+  ~spmm_state_t() {
+    if (a_descr_) {
+      cusparseDestroySpMat(a_descr_);
+    }
+    if (x_descr_) {
+      cusparseDestroyDnMat(x_descr_);
+    }
+    if (y_descr_) {
+      cusparseDestroyDnMat(y_descr_);
+    }
+  }
+
+  // Accessors for the descriptors
+  cusparseSpMatDescr_t a_descriptor() const {
+    return a_descr_;
+  }
+  cusparseDnMatDescr_t x_descriptor() const {
+    return x_descr_;
+  }
+  cusparseDnMatDescr_t y_descriptor() const {
+    return y_descr_;
+  }
+
+  // Setters for the descriptors
+  void set_a_descriptor(cusparseSpMatDescr_t descr) {
+    a_descr_ = descr;
+  }
+  void set_x_descriptor(cusparseDnMatDescr_t descr) {
+    x_descr_ = descr;
+  }
+  void set_y_descriptor(cusparseDnMatDescr_t descr) {
+    y_descr_ = descr;
+  }
+
+private:
+  cusparseSpMatDescr_t a_descr_ = nullptr;
+  cusparseDnMatDescr_t x_descr_ = nullptr;
+  cusparseDnMatDescr_t y_descr_ = nullptr;
+};
+
+} // namespace __cusparse
+} // namespace spblas

include/spblas/vendor/cusparse/detail/spmv_state_t.hpp

@@ -15,40 +15,18 @@ class spmv_state_t : public abstract_operation_state_t {
     if (a_descr_) {
       cusparseDestroySpMat(a_descr_);
     }
-    if (b_descr_) {
-      cusparseDestroyDnVec(b_descr_);
-    }
-    if (c_descr_) {
-      cusparseDestroyDnVec(c_descr_);
-    }
   }
 
-  // Accessors for the descriptors
   cusparseSpMatDescr_t a_descriptor() const {
     return a_descr_;
   }
-  cusparseDnVecDescr_t b_descriptor() const {
-    return b_descr_;
-  }
-  cusparseDnVecDescr_t c_descriptor() const {
-    return c_descr_;
-  }
 
-  // Setters for the descriptors
   void set_a_descriptor(cusparseSpMatDescr_t descr) {
     a_descr_ = descr;
   }
-  void set_b_descriptor(cusparseDnVecDescr_t descr) {
-    b_descr_ = descr;
-  }
-  void set_c_descriptor(cusparseDnVecDescr_t descr) {
-    c_descr_ = descr;
-  }
 
 private:
   cusparseSpMatDescr_t a_descr_ = nullptr;
-  cusparseDnVecDescr_t b_descr_ = nullptr;
-  cusparseDnVecDescr_t c_descr_ = nullptr;
 };
 
 } // namespace __cusparse

include/spblas/vendor/cusparse/multiply.hpp

@@ -1,3 +1,4 @@
 #pragma once
 
 #include "spmv_impl.hpp"
+#include "spmm_impl.hpp"