Don't use cudnnConvolutionBiasActivationForward. #54
Conversation
It supports only the precomputed implicit GEMM implementation of convolution.
|
|
||
| // We pad the bias with leading dims of 1, since CUDNN's tensors appear to | ||
| // need a dimension of at least 3. | ||
| int bias_dims[3] = {1, 1, c.num_channels_out}; | ||
| int bias_stride[3] = {c.num_channels_out, c.num_channels_out, 1}; | ||
| int bias_dims[4] = {1, c.num_channels_out, 1, 1}; |
There was a problem hiding this comment.
Can you explain - preferably in a comment - why this specific pattern of dims is required? It's mysterious to me.
There was a problem hiding this comment.
Done. I can make the bias optional at some point later.
The ConvolutionComputation class does not know whether or not the bias is optional at initialization time. It will simply avoid using the bias if it is a nullptr.
|
@danpovey I made the bias optional in my latest commit. The ConvolutionComputation object doesn't know whether the bias is optional in my design. It will simply avoid running any bias-related routines if the bias argument points to null. If you want the ConvolutionComputation object to take a boolean specifying whether or not the bias is used, and check that that matches with whether the bias is null or not, feel free to implement it that way. |
|
Thanks-- merging. Try to get in the habit of changing the documentation when you make code changes! It's easier psychologically to do it right away than later. |
It supports only the precomputed implicit GEMM implementation of
convolution.
Also fix the bias layout.