Example #1
0
 // 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) {
   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;
 }
Example #2
0
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);
  FloatToHalfBuffer(buffers_host.b_mat, b_buffer2);
  return status;
}
Example #3
0
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.kernel_mode = args.kernel_mode;
  args2.channels = args.channels; args2.height = args.height; args2.width = args.width;
  args2.kernel_h = args.kernel_h; args2.kernel_w = args.kernel_w;
  args2.pad_h = args.pad_h; args2.pad_w = args.pad_w;
  args2.stride_h = args.stride_h; args2.stride_w = args.stride_w;
  args2.dilation_h = args.dilation_h; args2.dilation_w = args.dilation_w;
  args2.a_offset = args.a_offset; args2.b_offset = args.b_offset;
  auto status = RunReference(args2, buffers2);
  FloatToHalfBuffer(buffers_host.a_mat, buffers2.a_mat);
  return status;
}
Example #4
0
 static StatusCode RunReference2(const Arguments<T> &args, BuffersHost<T> &buffers_host, Queue&) {
   return RunReference(args, buffers_host);
 }