winograd configuration for CUDNN_CONVOLUTION_FWD_ALGO_WINOGRAD

AI & Data Science Deep Learning (Training & Inference) cuDNN
1

Hi,

Can the winograd transform be parameterised in CUDNN_CONVOLUTION_FWD_ALGO_WINOGRAD. The winograd operator is often defined given the filter size (e.g. 3x3) and the output tile size (e.g. 4x4, 6x6, 8x8). Can the later be specified when calling the algorithm? Is there some documentation about this?

Thanks!