From 6b625f8915ce0596d65187bd3a8eb47e91a0084e Mon Sep 17 00:00:00 2001
From: Cedric Nugteren <web@cedricnugteren.nl>
Date: Mon, 10 Apr 2017 22:54:14 +0200
Subject: Added reference implementations for performance-testing against
 cuBLAS

---
 test/routines/levelx/xgemmbatched.hpp | 17 +++++++++++++++++
 1 file changed, 17 insertions(+)

(limited to 'test/routines/levelx/xgemmbatched.hpp')

diff --git a/test/routines/levelx/xgemmbatched.hpp b/test/routines/levelx/xgemmbatched.hpp
index e13e9382..ae8630c0 100644
--- a/test/routines/levelx/xgemmbatched.hpp
+++ b/test/routines/levelx/xgemmbatched.hpp
@@ -160,6 +160,23 @@ class TestXgemmBatched {
     }
   #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 &) {
+      for (auto batch = size_t{0}; batch < args.batch_count; ++batch) {
+        auto status = cublasXgemm(args.layout,
+                                  convertToCUBLAS(args.a_transpose),
+                                  convertToCUBLAS(args.b_transpose),
+                                  args.m, args.n, args.k, args.alphas[batch],
+                                  buffers.a_mat, args.a_offsets[batch], args.a_ld,
+                                  buffers.b_mat, args.b_offsets[batch], args.b_ld, args.betas[batch],
+                                  buffers.c_mat, args.c_offsets[batch], args.c_ld);
+      if (status != CUBLAS_STATUS_SUCCESS) { return StatusCode::kUnknownError; }
+      }
+      return StatusCode::kSuccess;
+    }
+  #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.c_size, static_cast<T>(0));
-- 
cgit v1.2.3