Separated host-device and device-host memory copies from execution of the CBLAS reference code; for fair timing and code de-duplication

9 files changed, 50 insertions, 87 deletions

diff --git a/test/routines/level3/xgemm.hpp b/test/routines/level3/xgemm.hpp
index 1b12fb1c..a33cbfec 100644
--- a/test/routines/level3/xgemm.hpp
+++ b/test/routines/level3/xgemm.hpp
@@ -45,6 +45,8 @@ class TestXgemm {
             kArgAOffset, kArgBOffset, kArgCOffset,
             kArgAlpha, kArgBeta};
   }
+  static std::vector<std::string> BuffersIn() { return {kBufMatA, kBufMatB, kBufMatC}; }
+  static std::vector<std::string> BuffersOut() { return {kBufMatC}; }
 
   // Describes how to obtain the sizes of the buffers
   static size_t GetSizeA(const Arguments<T> &args) {
@@ -121,21 +123,14 @@ class TestXgemm {
 
   // 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> b_mat_cpu(args.b_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.b_mat.Read(queue, args.b_size, b_mat_cpu);
-      buffers.c_mat.Read(queue, args.c_size, c_mat_cpu);
+    static StatusCode RunReference2(const Arguments<T> &args, BuffersHost<T> &buffers_host, Queue &) {
       cblasXgemm(convertToCBLAS(args.layout),
                  convertToCBLAS(args.a_transpose),
                  convertToCBLAS(args.b_transpose),
                  args.m, args.n, args.k, args.alpha,
-                 a_mat_cpu, args.a_offset, args.a_ld,
-                 b_mat_cpu, args.b_offset, args.b_ld, args.beta,
-                 c_mat_cpu, args.c_offset, args.c_ld);
-      buffers.c_mat.Write(queue, args.c_size, c_mat_cpu);
+                 buffers_host.a_mat, args.a_offset, args.a_ld,
+                 buffers_host.b_mat, args.b_offset, args.b_ld, args.beta,
+                 buffers_host.c_mat, args.c_offset, args.c_ld);
       return StatusCode::kSuccess;
     }
   #endif
diff --git a/test/routines/level3/xhemm.hpp b/test/routines/level3/xhemm.hpp
index 76550b15..74029c7e 100644
--- a/test/routines/level3/xhemm.hpp
+++ b/test/routines/level3/xhemm.hpp
@@ -45,6 +45,8 @@ class TestXhemm {
             kArgAOffset, kArgBOffset, kArgCOffset,
             kArgAlpha, kArgBeta};
   }
+  static std::vector<std::string> BuffersIn() { return {kBufMatA, kBufMatB, kBufMatC}; }
+  static std::vector<std::string> BuffersOut() { return {kBufMatC}; }
 
   // Describes how to obtain the sizes of the buffers
   static size_t GetSizeA(const Arguments<T> &args) {
@@ -121,21 +123,14 @@ class TestXhemm {
 
   // 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> b_mat_cpu(args.b_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.b_mat.Read(queue, args.b_size, b_mat_cpu);
-      buffers.c_mat.Read(queue, args.c_size, c_mat_cpu);
+    static StatusCode RunReference2(const Arguments<T> &args, BuffersHost<T> &buffers_host, Queue &) {
       cblasXhemm(convertToCBLAS(args.layout),
                  convertToCBLAS(args.side),
                  convertToCBLAS(args.triangle),
                  args.m, args.n, args.alpha,
-                 a_mat_cpu, args.a_offset, args.a_ld,
-                 b_mat_cpu, args.b_offset, args.b_ld, args.beta,
-                 c_mat_cpu, args.c_offset, args.c_ld);
-      buffers.c_mat.Write(queue, args.c_size, c_mat_cpu);
+                 buffers_host.a_mat, args.a_offset, args.a_ld,
+                 buffers_host.b_mat, args.b_offset, args.b_ld, args.beta,
+                 buffers_host.c_mat, args.c_offset, args.c_ld);
       return StatusCode::kSuccess;
     }
   #endif
diff --git a/test/routines/level3/xher2k.hpp b/test/routines/level3/xher2k.hpp
index 5ca3aac6..ea13bbc1 100644
--- a/test/routines/level3/xher2k.hpp
+++ b/test/routines/level3/xher2k.hpp
@@ -45,6 +45,8 @@ class TestXher2k {
             kArgAOffset, kArgBOffset, kArgCOffset,
             kArgAlpha, kArgBeta};
   }
+  static std::vector<std::string> BuffersIn() { return {kBufMatA, kBufMatB, kBufMatC}; }
+  static std::vector<std::string> BuffersOut() { return {kBufMatC}; }
 
   // Describes how to obtain the sizes of the buffers
   static size_t GetSizeA(const Arguments<U> &args) {
@@ -121,22 +123,15 @@ class TestXher2k {
 
   // Describes how to run the CPU BLAS routine (for correctness/performance comparison)
   #ifdef CLBLAST_REF_CBLAS
-    static StatusCode RunReference2(const Arguments<U> &args, Buffers<T> &buffers, Queue &queue) {
-      std::vector<T> a_mat_cpu(args.a_size, static_cast<T>(0));
-      std::vector<T> b_mat_cpu(args.b_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.b_mat.Read(queue, args.b_size, b_mat_cpu);
-      buffers.c_mat.Read(queue, args.c_size, c_mat_cpu);
+    static StatusCode RunReference2(const Arguments<U> &args, BuffersHost<T> &buffers_host, Queue&) {
       auto alpha2 = T{args.alpha, args.alpha};
       cblasXher2k(convertToCBLAS(args.layout),
                   convertToCBLAS(args.triangle),
                   convertToCBLAS(args.a_transpose),
                   args.n, args.k, alpha2,
-                  a_mat_cpu, args.a_offset, args.a_ld,
-                  b_mat_cpu, args.b_offset, args.b_ld, args.beta,
-                  c_mat_cpu, args.c_offset, args.c_ld);
-      buffers.c_mat.Write(queue, args.c_size, c_mat_cpu);
+                  buffers_host.a_mat, args.a_offset, args.a_ld,
+                  buffers_host.b_mat, args.b_offset, args.b_ld, args.beta,
+                  buffers_host.c_mat, args.c_offset, args.c_ld);
       return StatusCode::kSuccess;
     }
   #endif
diff --git a/test/routines/level3/xherk.hpp b/test/routines/level3/xherk.hpp
index e93d887a..b1ce83e0 100644
--- a/test/routines/level3/xherk.hpp
+++ b/test/routines/level3/xherk.hpp
@@ -45,6 +45,8 @@ class TestXherk {
             kArgAOffset, kArgCOffset,
             kArgAlpha, kArgBeta};
   }
+  static std::vector<std::string> BuffersIn() { return {kBufMatA, kBufMatC}; }
+  static std::vector<std::string> BuffersOut() { return {kBufMatC}; }
 
   // Describes how to obtain the sizes of the buffers
   static size_t GetSizeA(const Arguments<U> &args) {
@@ -110,18 +112,13 @@ class TestXherk {
 
   // Describes how to run the CPU BLAS routine (for correctness/performance comparison)
   #ifdef CLBLAST_REF_CBLAS
-    static StatusCode RunReference2(const Arguments<U> &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);
+    static StatusCode RunReference2(const Arguments<U> &args, BuffersHost<T> &buffers_host, Queue&) {
       cblasXherk(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);
+                 buffers_host.a_mat, args.a_offset, args.a_ld, args.beta,
+                 buffers_host.c_mat, args.c_offset, args.c_ld);
       return StatusCode::kSuccess;
     }
   #endif
diff --git a/test/routines/level3/xsymm.hpp b/test/routines/level3/xsymm.hpp
index 9d127e26..6ab644b8 100644
--- a/test/routines/level3/xsymm.hpp
+++ b/test/routines/level3/xsymm.hpp
@@ -45,6 +45,8 @@ class TestXsymm {
             kArgAOffset, kArgBOffset, kArgCOffset,
             kArgAlpha, kArgBeta};
   }
+  static std::vector<std::string> BuffersIn() { return {kBufMatA, kBufMatB, kBufMatC}; }
+  static std::vector<std::string> BuffersOut() { return {kBufMatC}; }
 
   // Describes how to obtain the sizes of the buffers
   static size_t GetSizeA(const Arguments<T> &args) {
@@ -121,21 +123,14 @@ class TestXsymm {
 
   // 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> b_mat_cpu(args.b_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.b_mat.Read(queue, args.b_size, b_mat_cpu);
-      buffers.c_mat.Read(queue, args.c_size, c_mat_cpu);
+    static StatusCode RunReference2(const Arguments<T> &args, BuffersHost<T> &buffers_host, Queue &) {
       cblasXsymm(convertToCBLAS(args.layout),
                  convertToCBLAS(args.side),
                  convertToCBLAS(args.triangle),
                  args.m, args.n, args.alpha,
-                 a_mat_cpu, args.a_offset, args.a_ld,
-                 b_mat_cpu, args.b_offset, args.b_ld, args.beta,
-                 c_mat_cpu, args.c_offset, args.c_ld);
-      buffers.c_mat.Write(queue, args.c_size, c_mat_cpu);
+                 buffers_host.a_mat, args.a_offset, args.a_ld,
+                 buffers_host.b_mat, args.b_offset, args.b_ld, args.beta,
+                 buffers_host.c_mat, args.c_offset, args.c_ld);
       return StatusCode::kSuccess;
     }
   #endif
diff --git a/test/routines/level3/xsyr2k.hpp b/test/routines/level3/xsyr2k.hpp
index d1bdac56..1400c4e2 100644
--- a/test/routines/level3/xsyr2k.hpp
+++ b/test/routines/level3/xsyr2k.hpp
@@ -45,6 +45,8 @@ class TestXsyr2k {
             kArgAOffset, kArgBOffset, kArgCOffset,
             kArgAlpha, kArgBeta};
   }
+  static std::vector<std::string> BuffersIn() { return {kBufMatA, kBufMatB, kBufMatC}; }
+  static std::vector<std::string> BuffersOut() { return {kBufMatC}; }
 
   // Describes how to obtain the sizes of the buffers
   static size_t GetSizeA(const Arguments<T> &args) {
@@ -119,21 +121,14 @@ class TestXsyr2k {
 
   // 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> b_mat_cpu(args.b_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.b_mat.Read(queue, args.b_size, b_mat_cpu);
-      buffers.c_mat.Read(queue, args.c_size, c_mat_cpu);
+    static StatusCode RunReference2(const Arguments<T> &args, BuffersHost<T> &buffers_host, Queue &) {
       cblasXsyr2k(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,
-                  b_mat_cpu, args.b_offset, args.b_ld, args.beta,
-                  c_mat_cpu, args.c_offset, args.c_ld);
-      buffers.c_mat.Write(queue, args.c_size, c_mat_cpu);
+                  buffers_host.a_mat, args.a_offset, args.a_ld,
+                  buffers_host.b_mat, args.b_offset, args.b_ld, args.beta,
+                  buffers_host.c_mat, args.c_offset, args.c_ld);
       return StatusCode::kSuccess;
     }
   #endif
diff --git a/test/routines/level3/xsyrk.hpp b/test/routines/level3/xsyrk.hpp
index 1330924e..2df8d6b0 100644
--- a/test/routines/level3/xsyrk.hpp
+++ b/test/routines/level3/xsyrk.hpp
@@ -45,6 +45,8 @@ class TestXsyrk {
             kArgAOffset, kArgCOffset,
             kArgAlpha, kArgBeta};
   }
+  static std::vector<std::string> BuffersIn() { return {kBufMatA, kBufMatC}; }
+  static std::vector<std::string> BuffersOut() { return {kBufMatC}; }
 
   // Describes how to obtain the sizes of the buffers
   static size_t GetSizeA(const Arguments<T> &args) {
@@ -110,18 +112,13 @@ class TestXsyrk {
 
   // 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);
+    static StatusCode RunReference2(const Arguments<T> &args, BuffersHost<T> &buffers_host, Queue &) {
       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);
+                 buffers_host.a_mat, args.a_offset, args.a_ld, args.beta,
+                 buffers_host.c_mat, args.c_offset, args.c_ld);
       return StatusCode::kSuccess;
     }
   #endif
diff --git a/test/routines/level3/xtrmm.hpp b/test/routines/level3/xtrmm.hpp
index 7c5bd842..84adc6e0 100644
--- a/test/routines/level3/xtrmm.hpp
+++ b/test/routines/level3/xtrmm.hpp
@@ -45,6 +45,8 @@ class TestXtrmm {
             kArgAOffset, kArgBOffset,
             kArgAlpha};
   }
+  static std::vector<std::string> BuffersIn() { return {kBufMatA, kBufMatB}; }
+  static std::vector<std::string> BuffersOut() { return {kBufMatB}; }
 
   // Describes how to obtain the sizes of the buffers
   static size_t GetSizeA(const Arguments<T> &args) {
@@ -112,20 +114,15 @@ class TestXtrmm {
 
   // 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> b_mat_cpu(args.b_size, static_cast<T>(0));
-      buffers.a_mat.Read(queue, args.a_size, a_mat_cpu);
-      buffers.b_mat.Read(queue, args.b_size, b_mat_cpu);
+    static StatusCode RunReference2(const Arguments<T> &args, BuffersHost<T> &buffers_host, Queue &) {
       cblasXtrmm(convertToCBLAS(args.layout),
                  convertToCBLAS(args.side),
                  convertToCBLAS(args.triangle),
                  convertToCBLAS(args.a_transpose),
                  convertToCBLAS(args.diagonal),
                  args.m, args.n, args.alpha,
-                 a_mat_cpu, args.a_offset, args.a_ld,
-                 b_mat_cpu, args.b_offset, args.b_ld);
-      buffers.b_mat.Write(queue, args.b_size, b_mat_cpu);
+                 buffers_host.a_mat, args.a_offset, args.a_ld,
+                 buffers_host.b_mat, args.b_offset, args.b_ld);
       return StatusCode::kSuccess;
     }
   #endif
diff --git a/test/routines/level3/xtrsm.hpp b/test/routines/level3/xtrsm.hpp
index a70ef03f..de5b307d 100644
--- a/test/routines/level3/xtrsm.hpp
+++ b/test/routines/level3/xtrsm.hpp
@@ -47,6 +47,8 @@ class TestXtrsm {
             kArgAOffset, kArgBOffset,
             kArgAlpha};
   }
+  static std::vector<std::string> BuffersIn() { return {kBufMatA, kBufMatB}; }
+  static std::vector<std::string> BuffersOut() { return {kBufMatB}; }
 
   // Describes how to obtain the sizes of the buffers
   static size_t GetSizeA(const Arguments<T> &args) {
@@ -124,20 +126,15 @@ class TestXtrsm {
 
   // 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> b_mat_cpu(args.b_size, static_cast<T>(0));
-      buffers.a_mat.Read(queue, args.a_size, a_mat_cpu);
-      buffers.b_mat.Read(queue, args.b_size, b_mat_cpu);
+    static StatusCode RunReference2(const Arguments<T> &args, BuffersHost<T> &buffers_host, Queue &) {
       cblasXtrsm(convertToCBLAS(args.layout),
                  convertToCBLAS(args.side),
                  convertToCBLAS(args.triangle),
                  convertToCBLAS(args.a_transpose),
                  convertToCBLAS(args.diagonal),
                  args.m, args.n, args.alpha,
-                 a_mat_cpu, args.a_offset, args.a_ld,
-                 b_mat_cpu, args.b_offset, args.b_ld);
-      buffers.b_mat.Write(queue, args.b_size, b_mat_cpu);
+                 buffers_host.a_mat, args.a_offset, args.a_ld,
+                 buffers_host.b_mat, args.b_offset, args.b_ld);
       return StatusCode::kSuccess;
     }
   #endif