From b84d2296b87ac212474af855d916b12adf96bdb7 Mon Sep 17 00:00:00 2001
From: Cedric Nugteren <web@cedricnugteren.nl>
Date: Sat, 1 Apr 2017 13:36:24 +0200
Subject: Separated host-device and device-host memory copies from execution of
 the CBLAS reference code; for fair timing and code de-duplication

---
 test/routines/level2/xspmv.hpp | 17 ++++++-----------
 1 file changed, 6 insertions(+), 11 deletions(-)

(limited to 'test/routines/level2/xspmv.hpp')

diff --git a/test/routines/level2/xspmv.hpp b/test/routines/level2/xspmv.hpp
index 352c8cfd..560f5baa 100644
--- a/test/routines/level2/xspmv.hpp
+++ b/test/routines/level2/xspmv.hpp
@@ -45,6 +45,8 @@ class TestXspmv {
             kArgAPOffset, kArgXOffset, kArgYOffset,
             kArgAlpha, kArgBeta};
   }
+  static std::vector<std::string> BuffersIn() { return {kBufMatAP, kBufVecX, kBufVecY}; }
+  static std::vector<std::string> BuffersOut() { return {kBufVecY}; }
 
   // Describes how to obtain the sizes of the buffers
   static size_t GetSizeX(const Arguments<T> &args) {
@@ -112,20 +114,13 @@ class TestXspmv {
 
   // 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> ap_mat_cpu(args.ap_size, static_cast<T>(0));
-      std::vector<T> x_vec_cpu(args.x_size, static_cast<T>(0));
-      std::vector<T> y_vec_cpu(args.y_size, static_cast<T>(0));
-      buffers.ap_mat.Read(queue, args.ap_size, ap_mat_cpu);
-      buffers.x_vec.Read(queue, args.x_size, x_vec_cpu);
-      buffers.y_vec.Read(queue, args.y_size, y_vec_cpu);
+    static StatusCode RunReference2(const Arguments<T> &args, BuffersHost<T> &buffers_host, Queue &) {
       cblasXspmv(convertToCBLAS(args.layout),
                  convertToCBLAS(args.triangle),
                  args.n, args.alpha,
-                 ap_mat_cpu, args.ap_offset,
-                 x_vec_cpu, args.x_offset, args.x_inc, args.beta,
-                 y_vec_cpu, args.y_offset, args.y_inc);
-      buffers.y_vec.Write(queue, args.y_size, y_vec_cpu);
+                 buffers_host.ap_mat, args.ap_offset,
+                 buffers_host.x_vec, args.x_offset, args.x_inc, args.beta,
+                 buffers_host.y_vec, args.y_offset, args.y_inc);
       return StatusCode::kSuccess;
     }
   #endif
-- 
cgit v1.2.3