test.sh errors for ToT cunn

[borisfom]

Those are with Pascal:

VolumetricDilatedConvolution
error on weight (backward) with torch.CudaHalfTensor
LT failed: 2.3657379150391 >= 2.26
        ...fomitchev/git/distro/install/share/lua/5.1/cunn/test.lua:5100: in function 'v'
        ...fomitchev/git/distro/install/share/lua/5.1/cunn/test.lua:5690: in function <...fomitchev/git/distro/install/share/lua/5.1/cunn/test.lua:5688>
--------------------------------------------------------------------------------
VolumetricDilatedConvolution
error on bias (backward) with torch.CudaHalfTensor
LT failed: 2.3330039978027 >= 2.26
        ...fomitchev/git/distro/install/share/lua/5.1/cunn/test.lua:5103: in function 'v'
        ...fomitchev/git/distro/install/share/lua/5.1/cunn/test.lua:5690: in function <...fomitchev/git/distro/install/share/lua/5.1/cunn/test.lua:5688>
--------------------------------------------------------------------------------
SpatialDilatedConvolution_forward_batch
error on state (forward) with torch.CudaHalfTensor
LT failed: 0.020028591156006 >= 0.0101
        ...fomitchev/git/distro/install/share/lua/5.1/cunn/test.lua:1741: in function 'v'
        ...fomitchev/git/distro/install/share/lua/5.1/cunn/test.lua:5690: in function <...fomitchev/git/distro/install/share/lua/5.1/cunn

[gchanan]

what does ToT refer to?

[borisfom]

@gchanan : Top of the Tree, a.k.a HEAD :).
I am also getting 'floating point exception' on nn tests with P100.

[borisfom]

Update: after pulling master to all submodules today, this error is gone. I still do get intermittent failures in nn tests: segfaults.