Simon Layton

Ok, to clarify - the call to `cudnnAddTensor` is coming from the bias addition in a convolution operator? In which case you actually want (all in NCHW): (8, 512, 28,...

@pietern Still waiting on more information from @kangdongh otherwise I can't comment

I just downloaded CUDA 9.1.85 and built, with the only issues being some symbols in NCCL (I'm building against a locally-built version with CUDA 9) -- I don't see the...

Do you have an earlier compiler available to test with? I just noticed you're on 6.4, I'm on 5.4