Cleanup before merging into QMCkl's GPU branch

org/qmckl_blas.org

@@ -2288,6 +2288,98 @@ qmckl_transpose (qmckl_context context,
 
      #+end_src
 
+* cuBLAS interface (optional)
+We propose a cuBLAS version of some QMCkl kernels. However, because cuBLAS is written in C, we need to define a Fortran interface for it. We start by defining functions to manage the cuBLAS handle structure from Fortran, before writing interfaces for the specific cuBLAS functions we are interested in.
+
+TODO These are the C functions that are supposed to be called from Fortran. We still need to write the interfaces themselves.
+
+#+begin_src c :tangle (eval h_private_func) :comments org
+#ifdef HAVE_CUBLAS_OFFLOAD
+#include <cublas_v2.h>
+#endif
+#+end_src
+
+#+begin_src c :tangle (eval h_private_func) :comments org
+#ifdef HAVE_CUBLAS_OFFLOAD
+cublasHandle_t* get_cublas_handle_interfaced();
+#endif
+#+end_src
+
+#+begin_src c :comments org :tangle (eval c) :exports none
+#ifdef HAVE_CUBLAS_OFFLOAD
+cublasHandle_t* get_cublas_handle_interfaced() {
+  cublasHandle_t* handle = malloc(sizeof(cublasHandle_t));
+
+  cublasStatus_t status = cublasCreate(handle);
+  if (status != CUBLAS_STATUS_SUCCESS){
+    fprintf(stderr, "Error while initializing cuBLAS\n");
+    exit(1);
+  }
+
+  return handle;
+}
+#endif
+#+end_src
+
+#+begin_src c :tangle (eval h_private_func) :comments org
+#ifdef HAVE_CUBLAS_OFFLOAD
+void destroy_cublas_handle_interfaced(cublasHandle_t* handle);
+#endif
+#+end_src
+
+#+begin_src c :comments org :tangle (eval c) :exports none
+#ifdef HAVE_CUBLAS_OFFLOAD
+void destroy_cublas_handle_interfaced(cublasHandle_t* handle) {
+  if(handle != NULL) {
+    free(handle);
+  }
+}
+#endif
+#+end_src
+
+** DGEMM
+
+#+begin_src c :tangle (eval h_private_func) :comments org
+#ifdef HAVE_CUBLAS_OFFLOAD
+cublasStatus_t cublasDgemm_f(
+  cublasHandle_t* handle,
+  cublasOperation_t* transa, cublasOperation_t* transb,
+  int* m, int* n, int* k,
+  const double* alpha,
+  const double*A, int* lda,
+  const double* B, int* ldb,
+  const double* beta,
+  double*C, int* ldc
+);
+#endif
+#+end_src
+
+#+begin_src c :comments org :tangle (eval c) :exports none
+
+#ifdef HAVE_CUBLAS_OFFLOAD
+cublasStatus_t cublasDgemm_f(
+  cublasHandle_t* handle,
+  cublasOperation_t* transa, cublasOperation_t* transb,
+  int* m, int* n, int* k,
+  const double* alpha,
+  const double*A, int* lda,
+  const double* B, int* ldb,
+  const double* beta,
+  double*C, int* ldc
+) {
+  return cublasDgemm_f(
+    handle,
+    transa, transb,
+    m, n, k,
+    alpha, A, lda, B,ldb,
+    beta, C, ldc
+  );
+}
+#endif
+#+end_src
+
+
+
 * End of files                                                     :noexport:
 
 

org/qmckl_jastrow.org

@@ -5592,7 +5592,7 @@ integer function qmckl_compute_tmp_c_cublas_offload_f(context, cord_num, elec_nu
   LDB = size(een_rescaled_n,1)
   LDC = size(tmp_c,1)
 
-  ! Alloc and copy memory on device
+  ! TODO Replace with calls to cuBLAS
 
   do nw=1, walk_num
   do i=0, cord_num-1