How can I get the tile size of gemm for cudnn kernel name?

I profile a pytorch model, I get the backward kernel “void wgrad_alg1_nd_float_engine<float, float, 3, 0, 5, 7, 4, 3, 5, false, true>(int, int, int, float const*, int, float*, float const*, kernel_gradNd_params, unsigned long long, int, float, float, int, int, int*, int*)”

  1. It is a wgrad of a conv, it should be a gemm like forward, but I can’t get the tile size like :128x128 or 64x64 from the kernel name. Is it (float, float)?
  2. how can I get the help of wgrad_alg1_nd_float_engine?

The way pytorch is tiling would probably best be explained by the pytorch team. You may have luck on the Frameworks forum:

Or cuDNN

I’m not sure I understand this “how can I get the help of wgrad_alg1_nd_float_engine?” Can you elaborate on that question?

1 Like