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

1 files changed, 48 insertions, 0 deletions

diff --git a/src/utilities/utilities.cpp b/src/utilities/utilities.cpp
index 3d091b64..0f2661ad 100644
--- a/src/utilities/utilities.cpp
+++ b/src/utilities/utilities.cpp
@@ -353,6 +353,54 @@ void PopulateVector(std::vector<half> &vector, std::mt19937 &mt, std::uniform_re
 
 // =================================================================================================
 
+template <typename T, typename U>
+void DeviceToHost(const Arguments<U> &args, Buffers<T> &buffers, BuffersHost<T> &buffers_host,
+                  Queue &queue, const std::vector<std::string> &names) {
+  for (auto &name: names) {
+    if (name == kBufVecX) {buffers_host.x_vec = std::vector<T>(args.x_size, static_cast<T>(0)); buffers.x_vec.Read(queue, args.x_size, buffers_host.x_vec); }
+    else if (name == kBufVecY) { buffers_host.y_vec = std::vector<T>(args.y_size, static_cast<T>(0)); buffers.y_vec.Read(queue, args.y_size, buffers_host.y_vec); }
+    else if (name == kBufMatA) { buffers_host.a_mat = std::vector<T>(args.a_size, static_cast<T>(0)); buffers.a_mat.Read(queue, args.a_size, buffers_host.a_mat); }
+    else if (name == kBufMatB) { buffers_host.b_mat = std::vector<T>(args.b_size, static_cast<T>(0)); buffers.b_mat.Read(queue, args.b_size, buffers_host.b_mat); }
+    else if (name == kBufMatC) { buffers_host.c_mat = std::vector<T>(args.c_size, static_cast<T>(0)); buffers.c_mat.Read(queue, args.c_size, buffers_host.c_mat); }
+    else if (name == kBufMatAP) { buffers_host.ap_mat = std::vector<T>(args.ap_size, static_cast<T>(0)); buffers.ap_mat.Read(queue, args.ap_size, buffers_host.ap_mat); }
+    else if (name == kBufScalar) { buffers_host.scalar = std::vector<T>(args.scalar_size, static_cast<T>(0)); buffers.scalar.Read(queue, args.scalar_size, buffers_host.scalar); }
+    else { throw std::runtime_error("Invalid buffer name"); }
+  }
+}
+
+template <typename T, typename U>
+void HostToDevice(const Arguments<U> &args, Buffers<T> &buffers, BuffersHost<T> &buffers_host,
+                  Queue &queue, const std::vector<std::string> &names) {
+  for (auto &name: names) {
+    if (name == kBufVecX) { buffers.x_vec.Write(queue, args.x_size, buffers_host.x_vec); }
+    else if (name == kBufVecY) { buffers.y_vec.Write(queue, args.y_size, buffers_host.y_vec); }
+    else if (name == kBufMatA) { buffers.a_mat.Write(queue, args.a_size, buffers_host.a_mat); }
+    else if (name == kBufMatB) { buffers.b_mat.Write(queue, args.b_size, buffers_host.b_mat); }
+    else if (name == kBufMatC) { buffers.c_mat.Write(queue, args.c_size, buffers_host.c_mat); }
+    else if (name == kBufMatAP) { buffers.ap_mat.Write(queue, args.ap_size, buffers_host.ap_mat); }
+    else if (name == kBufScalar) { buffers.scalar.Write(queue, args.scalar_size, buffers_host.scalar); }
+    else { throw std::runtime_error("Invalid buffer name"); }
+  }
+}
+
+// Compiles the above functions
+template void DeviceToHost(const Arguments<half>&, Buffers<half>&, BuffersHost<half>&, Queue&, const std::vector<std::string>&);
+template void DeviceToHost(const Arguments<float>&, Buffers<float>&, BuffersHost<float>&, Queue&, const std::vector<std::string>&);
+template void DeviceToHost(const Arguments<double>&, Buffers<double>&, BuffersHost<double>&, Queue&, const std::vector<std::string>&);
+template void DeviceToHost(const Arguments<float>&, Buffers<float2>&, BuffersHost<float2>&, Queue&, const std::vector<std::string>&);
+template void DeviceToHost(const Arguments<double>&, Buffers<double2>&, BuffersHost<double2>&, Queue&, const std::vector<std::string>&);
+template void DeviceToHost(const Arguments<float2>&, Buffers<float2>&, BuffersHost<float2>&, Queue&, const std::vector<std::string>&);
+template void DeviceToHost(const Arguments<double2>&, Buffers<double2>&, BuffersHost<double2>&, Queue&, const std::vector<std::string>&);
+template void HostToDevice(const Arguments<half>&, Buffers<half>&, BuffersHost<half>&, Queue&, const std::vector<std::string>&);
+template void HostToDevice(const Arguments<float>&, Buffers<float>&, BuffersHost<float>&, Queue&, const std::vector<std::string>&);
+template void HostToDevice(const Arguments<double>&, Buffers<double>&, BuffersHost<double>&, Queue&, const std::vector<std::string>&);
+template void HostToDevice(const Arguments<float>&, Buffers<float2>&, BuffersHost<float2>&, Queue&, const std::vector<std::string>&);
+template void HostToDevice(const Arguments<double>&, Buffers<double2>&, BuffersHost<double2>&, Queue&, const std::vector<std::string>&);
+template void HostToDevice(const Arguments<float2>&, Buffers<float2>&, BuffersHost<float2>&, Queue&, const std::vector<std::string>&);
+template void HostToDevice(const Arguments<double2>&, Buffers<double2>&, BuffersHost<double2>&, Queue&, const std::vector<std::string>&);
+
+// =================================================================================================
+
 // Conversion between half and single-precision
 std::vector<float> HalfToFloatBuffer(const std::vector<half>& source) {
   auto result = std::vector<float>(source.size());