// Describes how to run the CLBlast routine
static StatusCode RunRoutine(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) {
auto queue_plain = queue();
auto event = cl_event{};
auto status = Gerc(args.layout,
args.m, args.n, args.alpha,
buffers.x_vec(), args.x_offset, args.x_inc,
buffers.y_vec(), args.y_offset, args.y_inc,
buffers.a_mat(), args.a_offset, args.a_ld,
&queue_plain, &event);
clWaitForEvents(1, &event);
return status;
}
// Describes how to run the CLBlast routine
static StatusCode RunRoutine(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) {
#ifdef OPENCL_API
auto queue_plain = queue();
auto event = cl_event{};
auto status = Gerc(args.layout,
args.m, args.n, args.alpha,
buffers.x_vec(), args.x_offset, args.x_inc,
buffers.y_vec(), args.y_offset, args.y_inc,
buffers.a_mat(), args.a_offset, args.a_ld,
&queue_plain, &event);
if (status == StatusCode::kSuccess) { clWaitForEvents(1, &event); clReleaseEvent(event); }
#elif CUDA_API
auto status = Gerc(args.layout,
args.m, args.n, args.alpha,
buffers.x_vec(), args.x_offset, args.x_inc,
buffers.y_vec(), args.y_offset, args.y_inc,
buffers.a_mat(), args.a_offset, args.a_ld,
queue.GetContext()(), queue.GetDevice()());
cuStreamSynchronize(queue());
#endif
return status;
}
