From 1a82861a902e17f15486664b340c50530cce6542 Mon Sep 17 00:00:00 2001
From: cnugteren <web@cedricnugteren.nl>
Date: Sat, 2 Apr 2016 11:58:00 -0700
Subject: Added support for testing (performance and correctness) against a CPU
 BLAS library

---
 test/routines/level3/xsyrk.h | 55 ++++++++++++++++++++++++++++++++------------
 1 file changed, 40 insertions(+), 15 deletions(-)

(limited to 'test/routines/level3/xsyrk.h')

diff --git a/test/routines/level3/xsyrk.h b/test/routines/level3/xsyrk.h
index 8dacb5b3..7675e2aa 100644
--- a/test/routines/level3/xsyrk.h
+++ b/test/routines/level3/xsyrk.h
@@ -19,7 +19,12 @@
 #include <vector>
 #include <string>
 
-#include "wrapper_clblas.h"
+#ifdef CLBLAST_REF_CLBLAS
+  #include "wrapper_clblas.h"
+#endif
+#ifdef CLBLAST_REF_CBLAS
+  #include "wrapper_cblas.h"
+#endif
 
 namespace clblast {
 // =================================================================================================
@@ -69,7 +74,7 @@ class TestXsyrk {
   static Transposes GetBTransposes(const Transposes &) { return {}; } // N/A for this routine
 
   // Describes how to run the CLBlast routine
-  static StatusCode RunRoutine(const Arguments<T> &args, const Buffers<T> &buffers, Queue &queue) {
+  static StatusCode RunRoutine(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) {
     auto queue_plain = queue();
     auto event = cl_event{};
     auto status = Syrk(args.layout, args.triangle, args.a_transpose,
@@ -82,19 +87,39 @@ class TestXsyrk {
   }
 
   // Describes how to run the clBLAS routine (for correctness/performance comparison)
-  static StatusCode RunReference(const Arguments<T> &args, const Buffers<T> &buffers, Queue &queue) {
-    auto queue_plain = queue();
-    auto event = cl_event{};
-    auto status = clblasXsyrk(static_cast<clblasOrder>(args.layout),
-                              static_cast<clblasUplo>(args.triangle),
-                              static_cast<clblasTranspose>(args.a_transpose),
-                              args.n, args.k, args.alpha,
-                              buffers.a_mat(), args.a_offset, args.a_ld, args.beta,
-                              buffers.c_mat(), args.c_offset, args.c_ld,
-                              1, &queue_plain, 0, nullptr, &event);
-    clWaitForEvents(1, &event);
-    return static_cast<StatusCode>(status);
-  }
+  #ifdef CLBLAST_REF_CLBLAS
+    static StatusCode RunReference1(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) {
+      auto queue_plain = queue();
+      auto event = cl_event{};
+      auto status = clblasXsyrk(static_cast<clblasOrder>(args.layout),
+                                static_cast<clblasUplo>(args.triangle),
+                                static_cast<clblasTranspose>(args.a_transpose),
+                                args.n, args.k, args.alpha,
+                                buffers.a_mat(), args.a_offset, args.a_ld, args.beta,
+                                buffers.c_mat(), args.c_offset, args.c_ld,
+                                1, &queue_plain, 0, nullptr, &event);
+      clWaitForEvents(1, &event);
+      return static_cast<StatusCode>(status);
+    }
+  #endif
+
+  // Describes how to run the CPU BLAS routine (for correctness/performance comparison)
+  #ifdef CLBLAST_REF_CBLAS
+    static StatusCode RunReference2(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) {
+      std::vector<T> a_mat_cpu(args.a_size, static_cast<T>(0));
+      std::vector<T> c_mat_cpu(args.c_size, static_cast<T>(0));
+      buffers.a_mat.Read(queue, args.a_size, a_mat_cpu);
+      buffers.c_mat.Read(queue, args.c_size, c_mat_cpu);
+      cblasXsyrk(convertToCBLAS(args.layout),
+                 convertToCBLAS(args.triangle),
+                 convertToCBLAS(args.a_transpose),
+                 args.n, args.k, args.alpha,
+                 a_mat_cpu, args.a_offset, args.a_ld, args.beta,
+                 c_mat_cpu, args.c_offset, args.c_ld);
+      buffers.c_mat.Write(queue, args.c_size, c_mat_cpu);
+      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) {
-- 
cgit v1.2.3