cudnn.LogSoftMax seems to be broken about cudnn.torch

[szagoruyko]

This is what I get after running your snippet:

cudnn.SoftMax
 0.5000
 0.5000
[torch.CudaTensor of size 2]

nn.SoftMax
 0.5000
 0.5000
[torch.CudaTensor of size 2]

cudnn.LogSoftMax
-0.6931
-0.6931
[torch.CudaTensor of size 2]

nn.LogSoftMax
-0.6931
-0.6931
[torch.CudaTensor of size 2]

are you sure you're on the latest nn/cunn/cudnn? do nn.test(), nn.testcuda() and cudnn/test/test.lua work?

[Cadene]

nn.test() -- ==> Done Completed 2411 asserts in 143 tests with 0 errors
nn.testcuda() -- ==> Done Completed 560 asserts in 127 tests with 0 errors  

I just reinstalled cudnn R3 and now it works perfectly.
I still don't know why this function was broken.

Thanks, someone can close de issue.