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

4 files changed, 57 insertions, 72 deletions

diff --git a/test/routines/levelx/xaxpybatched.hpp b/test/routines/levelx/xaxpybatched.hpp
index ee15ff92..05141bbb 100644
--- a/test/routines/levelx/xaxpybatched.hpp
+++ b/test/routines/levelx/xaxpybatched.hpp
@@ -45,6 +45,8 @@ class TestXaxpyBatched {
             kArgXInc, kArgYInc,
             kArgBatchCount, kArgAlpha};
   }
+  static std::vector<std::string> BuffersIn() { return {kBufVecX, kBufVecY}; }
+  static std::vector<std::string> BuffersOut() { return {kBufVecY}; }
 
   // Helper for the sizes per batch
   static size_t PerBatchSizeX(const Arguments<T> &args) { return args.n * args.x_inc; }
@@ -123,17 +125,12 @@ class TestXaxpyBatched {
 
   // 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> x_vec_cpu(args.x_size, static_cast<T>(0));
-      std::vector<T> y_vec_cpu(args.y_size, static_cast<T>(0));
-      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 &) {
       for (auto batch = size_t{0}; batch < args.batch_count; ++batch) {
         cblasXaxpy(args.n, args.alphas[batch],
-                   x_vec_cpu, args.x_offsets[batch], args.x_inc,
-                   y_vec_cpu, args.y_offsets[batch], args.y_inc);
+                   buffers_host.x_vec, args.x_offsets[batch], args.x_inc,
+                   buffers_host.y_vec, args.y_offsets[batch], args.y_inc);
       }
-      buffers.y_vec.Write(queue, args.y_size, y_vec_cpu);
       return StatusCode::kSuccess;
     }
   #endif
diff --git a/test/routines/levelx/xgemmbatched.hpp b/test/routines/levelx/xgemmbatched.hpp
index 80a30e4d..ab5f20c5 100644
--- a/test/routines/levelx/xgemmbatched.hpp
+++ b/test/routines/levelx/xgemmbatched.hpp
@@ -45,6 +45,8 @@ class TestXgemmBatched {
             kArgAOffset, kArgBOffset, kArgCOffset,
             kArgBatchCount, kArgAlpha, kArgBeta};
   }
+  static std::vector<std::string> BuffersIn() { return {kBufMatA, kBufMatB, kBufMatC}; }
+  static std::vector<std::string> BuffersOut() { return {kBufMatC}; }
 
   // Helper for the sizes per batch
   static size_t PerBatchSizeA(const Arguments<T> &args) {
@@ -152,23 +154,16 @@ class TestXgemmBatched {
 
   // 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 &) {
       for (auto batch = size_t{0}; batch < args.batch_count; ++batch) {
         cblasXgemm(convertToCBLAS(args.layout),
                    convertToCBLAS(args.a_transpose),
                    convertToCBLAS(args.b_transpose),
                    args.m, args.n, args.k, args.alphas[batch],
-                   a_mat_cpu, args.a_offsets[batch], args.a_ld,
-                   b_mat_cpu, args.b_offsets[batch], args.b_ld, args.betas[batch],
-                   c_mat_cpu, args.c_offsets[batch], args.c_ld);
+                   buffers_host.a_mat, args.a_offsets[batch], args.a_ld,
+                   buffers_host.b_mat, args.b_offsets[batch], args.b_ld, args.betas[batch],
+                   buffers_host.c_mat, args.c_offsets[batch], args.c_ld);
       }
-      buffers.c_mat.Write(queue, args.c_size, c_mat_cpu);
       return StatusCode::kSuccess;
     }
   #endif
diff --git a/test/routines/levelx/xinvert.hpp b/test/routines/levelx/xinvert.hpp
index b470dbf3..ffb484b0 100644
--- a/test/routines/levelx/xinvert.hpp
+++ b/test/routines/levelx/xinvert.hpp
@@ -25,17 +25,10 @@ namespace clblast {
 // =================================================================================================
 
 template <typename T>
-StatusCode RunReference(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) {
+StatusCode RunReference(const Arguments<T> &args, BuffersHost<T> &buffers_host) {
   const bool is_upper = ((args.triangle == Triangle::kUpper && args.layout != Layout::kRowMajor) ||
                          (args.triangle == Triangle::kLower && args.layout == Layout::kRowMajor));
 
-  // Data transfer from OpenCL to std::vector
-  std::vector<T> a_mat_cpu(args.a_size, T{0.0});
-  buffers.a_mat.Read(queue, args.a_size, a_mat_cpu);
-
-  // Creates the output buffer
-  std::vector<T> b_mat_cpu(args.b_size, T{0.0});
-
   // Helper variables
   const auto block_size = args.m;
   const auto num_blocks = CeilDiv(args.n, block_size);
@@ -60,11 +53,11 @@ StatusCode RunReference(const Arguments<T> &args, Buffers<T> &buffers, Queue &qu
       auto a_value = T{1.0};
       if (args.diagonal == Diagonal::kNonUnit) {
         if (i + block_id * block_size < args.n) {
-          if (a_mat_cpu[i * a_ld + i + a_offset] == T{0.0}) { return StatusCode::kUnknownError; }
-          a_value = T{1.0} / a_mat_cpu[i * a_ld + i + a_offset];
+          if (buffers_host.a_mat[i * a_ld + i + a_offset] == T{0.0}) { return StatusCode::kUnknownError; }
+          a_value = T{1.0} / buffers_host.a_mat[i * a_ld + i + a_offset];
         }
       }
-      b_mat_cpu[i * b_ld + i + b_offset] = a_value;
+      buffers_host.b_mat[i * b_ld + i + b_offset] = a_value;
     }
 
     // Inverts the upper triangle row by row
@@ -75,11 +68,11 @@ StatusCode RunReference(const Arguments<T> &args, Buffers<T> &buffers, Queue &qu
           for (auto k = i + 1; k <= j; ++k) {
             auto a_value = T{0.0};
             if ((i + block_id * block_size < args.n) && (k + block_id * block_size < args.n)) {
-              a_value = a_mat_cpu[k * a_ld + i + a_offset];
+              a_value = buffers_host.a_mat[k * a_ld + i + a_offset];
             }
-            sum += a_value * b_mat_cpu[j * b_ld + k + b_offset];
+            sum += a_value * buffers_host.b_mat[j * b_ld + k + b_offset];
           }
-          b_mat_cpu[j * b_ld + i + b_offset] = - sum * b_mat_cpu[i * b_ld + i + b_offset];
+          buffers_host.b_mat[j * b_ld + i + b_offset] = - sum * buffers_host.b_mat[i * b_ld + i + b_offset];
         }
       }
     }
@@ -92,35 +85,32 @@ StatusCode RunReference(const Arguments<T> &args, Buffers<T> &buffers, Queue &qu
           for (auto k = j; k < i; ++k) {
             auto a_value = T{0.0};
             if ((i + block_id * block_size < args.n) && (k + block_id * block_size < args.n)) {
-              a_value = a_mat_cpu[k * a_ld + i + a_offset];
+              a_value = buffers_host.a_mat[k * a_ld + i + a_offset];
             }
-            sum += a_value * b_mat_cpu[j * b_ld + k + b_offset];
+            sum += a_value * buffers_host.b_mat[j * b_ld + k + b_offset];
           }
-          b_mat_cpu[j * b_ld + i + b_offset] = - sum * b_mat_cpu[i * b_ld + i + b_offset];
+          buffers_host.b_mat[j * b_ld + i + b_offset] = - sum * buffers_host.b_mat[i * b_ld + i + b_offset];
         }
       }
     }
   }
-
-  // Data transfer back to OpenCL
-  buffers.b_mat.Write(queue, args.b_size, b_mat_cpu);
   return StatusCode::kSuccess;
 }
 
 // Half-precision version calling the above reference implementation after conversions
 template <>
-StatusCode RunReference<half>(const Arguments<half> &args, Buffers<half> &buffers, Queue &queue) {
-  auto a_buffer2 = HalfToFloatBuffer(buffers.a_mat, queue());
-  auto b_buffer2 = HalfToFloatBuffer(buffers.b_mat, queue());
-  auto dummy = clblast::Buffer<float>(0);
-  auto buffers2 = Buffers<float>{dummy, dummy, a_buffer2, b_buffer2, dummy, dummy, dummy};
+StatusCode RunReference<half>(const Arguments<half> &args, BuffersHost<half> &buffers_host) {
+  auto a_buffer2 = HalfToFloatBuffer(buffers_host.a_mat);
+  auto b_buffer2 = HalfToFloatBuffer(buffers_host.b_mat);
+  auto dummy = std::vector<float>(0);
+  auto buffers2 = BuffersHost<float>{dummy, dummy, a_buffer2, b_buffer2, dummy, dummy, dummy};
   auto args2 = Arguments<float>();
   args2.a_size = args.a_size; args2.b_size = args.b_size;
   args2.a_ld = args.a_ld; args2.m = args.m; args2.n = args.n;
   args2.a_offset = args.a_offset;
   args2.layout = args.layout; args2.triangle = args.triangle; args2.diagonal = args.diagonal;
-  auto status = RunReference(args2, buffers2, queue);
-  FloatToHalfBuffer(buffers.b_mat, b_buffer2, queue());
+  auto status = RunReference(args2, buffers2);
+  FloatToHalfBuffer(buffers_host.b_mat, b_buffer2);
   return status;
 }
 
@@ -140,6 +130,8 @@ class TestXinvert {
             kArgLayout, kArgTriangle, kArgDiagonal,
             kArgALeadDim, kArgAOffset};
   }
+  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) {
@@ -190,11 +182,15 @@ class TestXinvert {
   // Describes how to run a naive version of the routine (for correctness/performance comparison).
   // Note that a proper clBLAS or CPU BLAS comparison is not available for non-BLAS routines.
   static StatusCode RunReference1(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) {
-    return RunReference(args, buffers, queue);
+    auto buffers_host = BuffersHost<T>();
+    DeviceToHost(args, buffers, buffers_host, queue, BuffersIn());
+    const auto status = RunReference(args, buffers_host);
+    HostToDevice(args, buffers, buffers_host, queue, BuffersOut());
+    return status;
   }
 
-  static StatusCode RunReference2(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) {
-    return RunReference(args, buffers, queue);
+  static StatusCode RunReference2(const Arguments<T> &args, BuffersHost<T> &buffers_host, Queue&) {
+    return RunReference(args, buffers_host);
   }
 
   // Describes how to download the results of the computation (more importantly: which buffer)
diff --git a/test/routines/levelx/xomatcopy.hpp b/test/routines/levelx/xomatcopy.hpp
index d1064d0c..d5973b4c 100644
--- a/test/routines/levelx/xomatcopy.hpp
+++ b/test/routines/levelx/xomatcopy.hpp
@@ -23,13 +23,7 @@ namespace clblast {
 // =================================================================================================
 
 template <typename T>
-StatusCode RunReference(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) {
-
-  // Data transfer from OpenCL to std::vector
-  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);
+StatusCode RunReference(const Arguments<T> &args, BuffersHost<T> &buffers_host) {
 
   // Checking for invalid arguments
   const auto a_rotated = (args.layout == Layout::kRowMajor);
@@ -40,8 +34,8 @@ StatusCode RunReference(const Arguments<T> &args, Buffers<T> &buffers, Queue &qu
   if ((args.m == 0) || (args.n == 0)) { return StatusCode::kInvalidDimension; }
   if ((args.a_ld < args.m && !a_rotated) || (args.a_ld < args.n && a_rotated)) { return StatusCode::kInvalidLeadDimA; }
   if ((args.b_ld < args.m && !b_rotated) || (args.b_ld < args.n && b_rotated)) { return StatusCode::kInvalidLeadDimB; }
-  if (buffers.a_mat.GetSize() < (a_base + args.a_offset) * sizeof(T)) { return StatusCode::kInsufficientMemoryA; }
-  if (buffers.b_mat.GetSize() < (b_base + args.b_offset) * sizeof(T)) { return StatusCode::kInsufficientMemoryB; }
+  if (buffers_host.a_mat.size() * sizeof(T) < (a_base + args.a_offset) * sizeof(T)) { return StatusCode::kInsufficientMemoryA; }
+  if (buffers_host.b_mat.size() * sizeof(T) < (b_base + args.b_offset) * sizeof(T)) { return StatusCode::kInsufficientMemoryB; }
 
   // Matrix copy, scaling, and/or transpose
   for (auto id1 = size_t{0}; id1 < args.m; ++id1) {
@@ -52,30 +46,27 @@ StatusCode RunReference(const Arguments<T> &args, Buffers<T> &buffers, Queue &qu
       const auto b_two = (b_rotated) ? id1 : id2;
       const auto a_index = a_two * args.a_ld + a_one + args.a_offset;
       const auto b_index = b_two * args.b_ld + b_one + args.b_offset;
-      b_mat_cpu[b_index] = args.alpha * a_mat_cpu[a_index];
+      buffers_host.b_mat[b_index] = args.alpha * buffers_host.a_mat[a_index];
     }
   }
-
-  // Data transfer back to OpenCL
-  buffers.b_mat.Write(queue, args.b_size, b_mat_cpu);
   return StatusCode::kSuccess;
 }
 
 // Half-precision version calling the above reference implementation after conversions
 template <>
-StatusCode RunReference<half>(const Arguments<half> &args, Buffers<half> &buffers, Queue &queue) {
-  auto a_buffer2 = HalfToFloatBuffer(buffers.a_mat, queue());
-  auto b_buffer2 = HalfToFloatBuffer(buffers.b_mat, queue());
-  auto dummy = clblast::Buffer<float>(0);
-  auto buffers2 = Buffers<float>{dummy, dummy, a_buffer2, b_buffer2, dummy, dummy, dummy};
+StatusCode RunReference<half>(const Arguments<half> &args, BuffersHost<half> &buffers_host) {
+  auto a_buffer2 = HalfToFloatBuffer(buffers_host.a_mat);
+  auto b_buffer2 = HalfToFloatBuffer(buffers_host.b_mat);
+  auto dummy = std::vector<float>(0);
+  auto buffers2 = BuffersHost<float>{dummy, dummy, a_buffer2, b_buffer2, dummy, dummy, dummy};
   auto args2 = Arguments<float>();
   args2.a_size = args.a_size; args2.b_size = args.b_size;
   args2.a_ld = args.a_ld; args2.b_ld = args.b_ld; args2.m = args.m; args2.n = args.n;
   args2.a_offset = args.a_offset; args2.b_offset = args.b_offset;
   args2.layout = args.layout; args2.a_transpose = args.a_transpose;
   args2.alpha = HalfToFloat(args.alpha);
-  auto status = RunReference(args2, buffers2, queue);
-  FloatToHalfBuffer(buffers.b_mat, b_buffer2, queue());
+  auto status = RunReference(args2, buffers2);
+  FloatToHalfBuffer(buffers_host.b_mat, b_buffer2);
   return status;
 }
 
@@ -97,6 +88,8 @@ class TestXomatcopy {
             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) {
@@ -148,11 +141,15 @@ class TestXomatcopy {
   // Describes how to run a naive version of the routine (for correctness/performance comparison).
   // Note that a proper clBLAS or CPU BLAS comparison is not available for non-BLAS routines.
   static StatusCode RunReference1(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) {
-    return RunReference(args, buffers, queue);
+    auto buffers_host = BuffersHost<T>();
+    DeviceToHost(args, buffers, buffers_host, queue, BuffersIn());
+    const auto status = RunReference(args, buffers_host);
+    HostToDevice(args, buffers, buffers_host, queue, BuffersOut());
+    return status;
   }
 
-  static StatusCode RunReference2(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) {
-    return RunReference(args, buffers, queue);
+  static StatusCode RunReference2(const Arguments<T> &args, BuffersHost<T> &buffers_host, Queue&) {
+    return RunReference(args, buffers_host);
   }
 
   // Describes how to download the results of the computation (more importantly: which buffer)