How cudnn v8 benchmark different convolution algorithms


I have some questions about cudnn v8 benchmarking for convolution algorithms.

  1. Does the “valid_plans” pass to time_sorted_plan represent different convolution algorithms?

  2. If that’s the case, is it because the desc parameter in this function is of type cudnnBackendDescriptorType_t with the value CUDNN_BACKEND_OPERATION_CONVOLUTION_FORWARD_DESCRIPTOR ? and then generate an opGraph with corresponding operation?

Thanks for your help!!