one ctx async 3.349727
two ctx async 1.713157
two ctx, 2 fct async 1.717468
two ctx, 2 fct with transfer 3.635517
two ctx, 2 fct async, 2 threads 1.731599
two ctx, 2 fct with transfer, 2 threads 2.071597

Fix call to gpu_from_host

Fix not using context_name in local_gpu_elemwise

Factoring inference of convolution output shape (continuation)

Fix #3190

Better error message.

Conv3d

Factoring inference of convolution output shape

Fixed error in doc

Fixes cl.exe not being seen on Windows

Port the speedup from #3163 to the new backend.

Enable cudnn for AbstractConv2d

comp, out and err evaluate to:

False
b"nvcc fatal   : Cannot find compiler 'cl.exe' in PATH\r\n"
b''

So I found it necessary to add the --compiler-bindir flag to the source, in order to actually use cudnn.

CorrMM optimizations

Allow optimizer_excluding=inplace to keep the move to GPU.

Clean up (outs|inps)_on_gpu to (outs|inps)_is_tensor