Ejemplo n.º 1
0
Archivo: xspr.hpp Proyecto: gpu/CLBlast
 // 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 = Spr(args.layout, args.triangle,
                       args.n, args.alpha,
                       buffers.x_vec(), args.x_offset, args.x_inc,
                       buffers.ap_mat(), args.ap_offset,
                       &queue_plain, &event);
     if (status == StatusCode::kSuccess) { clWaitForEvents(1, &event); clReleaseEvent(event); }
   #elif CUDA_API
     auto status = Spr(args.layout, args.triangle,
                       args.n, args.alpha,
                       buffers.x_vec(), args.x_offset, args.x_inc,
                       buffers.ap_mat(), args.ap_offset,
                       queue.GetContext()(), queue.GetDevice()());
     cuStreamSynchronize(queue());
   #endif
   return status;
 }
Ejemplo n.º 2
0
 // Describes how to run the CLBlast routine
 static StatusCode RunRoutine(const Arguments<U> &args, Buffers<T> &buffers, Queue &queue) {
   auto queue_plain = queue();
   auto event = cl_event{};
   auto status = Hpr(args.layout, args.triangle,
                     args.n, args.alpha,
                     buffers.x_vec(), args.x_offset, args.x_inc,
                     buffers.ap_mat(), args.ap_offset,
                     &queue_plain, &event);
   if (status == StatusCode::kSuccess) { clWaitForEvents(1, &event); clReleaseEvent(event); }
   return status;
 }
Ejemplo n.º 3
0
 // 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 = Hpr2(args.layout, args.triangle,
                      args.n, args.alpha,
                      buffers.x_vec(), args.x_offset, args.x_inc,
                      buffers.y_vec(), args.y_offset, args.y_inc,
                      buffers.ap_mat(), args.ap_offset,
                      &queue_plain, &event);
   clWaitForEvents(1, &event);
   return status;
 }