Added reference implementations for performance-testing against cuBLAS

1 files changed, 11 insertions, 0 deletions

diff --git a/test/routines/level1/xdot.hpp b/test/routines/level1/xdot.hpp
index cb3d7979..c95b16ef 100644
--- a/test/routines/level1/xdot.hpp
+++ b/test/routines/level1/xdot.hpp
@@ -110,6 +110,17 @@ class TestXdot {
     }
   #endif
 
+  // Describes how to run the cuBLAS routine (for correctness/performance comparison)
+  #ifdef CLBLAST_REF_CUBLAS
+    static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) {
+      auto status = cublasXdot(args.n,
+                               buffers.scalar, args.dot_offset,
+                               buffers.x_vec, args.x_offset, args.x_inc,
+                               buffers.y_vec, args.y_offset, args.y_inc);
+      if (status == CUBLAS_STATUS_SUCCESS) { return StatusCode::kSuccess; } else { return StatusCode::kUnknownError; }
+    }
+  #endif
+
   // Describes how to download the results of the computation (more importantly: which buffer)
   static std::vector<T> DownloadResult(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) {
     std::vector<T> result(args.scalar_size, static_cast<T>(0));