VolumetricConvolution tests fail

[torch]

VolumetricConvolution tests fail

[soumith]

does it fail EVERY time, or occasionally? I am not able to reproduce this on my machine (ugh!). Did you upgrade nn and cunn packages?

[szagoruyko]

All tests fail in both Ubuntu and OS X, sure the packages are updated, nn.test() passes without errors. On OS X the error is different though:

th> nn.testcuda{'VolumetricConvolution_forward_single'}
Running 1 tests
*  ==> Done Completed 0 asserts in 1 tests with 1 errors
--------------------------------------------------------------------------------
VolumetricConvolution_forward_single
 Function call failed
/usr/local/share/lua/5.1/cunn/test.lua:2897: bad argument #1 (field finput does not exist)
stack traceback:
    [C]: in function 'forward'
    /usr/local/share/lua/5.1/cunn/test.lua:2897: in function </usr/local/share/lua/5.1/cunn/test.lua:2863>
    [C]: in function 'xpcall'
    /usr/local/share/lua/5.1/torch/Tester.lua:112: in function 'pcall'
    /usr/local/share/lua/5.1/torch/Tester.lua:169: in function 'run'
    /usr/local/share/lua/5.1/cunn/test.lua:3088: in function 'f'
    [string "local f = function() return nn.testcuda{'Volu..."]:1: in main chunk
    [C]: in function 'xpcall'
    /usr/local/share/lua/5.1/trepl/init.lua:553: in function </usr/local/share/lua/5.1/trepl/init.lua:489>

[soumith]

on OSX you clearly have an outdated nn package. I added the finput field in this recent commit:
torch/nn@2c6d601

On ubuntu, i reproduced your problem, i think it is a blas issue, looking into it.

[soumith]

ok fixed the issue via commit torch/nn@b071b6a

Turns out, inside an OpenMP for loop, doing something as simple as:

long j;
#pragma omp parallel for private(j)
for (j=0; j < 10; j++) {
    THTensor *inb = THTensor_(new)();
    THTensor_(free)(inb);
}

segfaults. not sure why, I tried to figure it out and debug with no luck. So disabled openmp for the volumetric batch looping (CPU)