SpatialMaxPooling failure case

[soumith]

SpatialMaxPooling failure case

[soumith]

this is a case that comes in Overfeat (fast) type network.

[soumith]

looks like this line is the culprit:
https://github.com/torch/cunn/blob/master/SpatialMaxPooling.cu#L242

I ran on a 3.5 capable card (titan black), but it still refuses to run. The x-dim of the blocks in this case becomes: 131072.
From this page:
http://en.wikipedia.org/wiki/CUDA#Version_features_and_specifications
xdim for 3.0+ can be upto 2^31 -1.
Anyways, cudnn doesn't have this bug, i am unlikely to even look into fixing this.

[jonathantompson]

Nice catch.

[soumith]

The reason here is that cunn is complied with 2.0 compability:

https://github.com/torch/cunn/blob/master/CMakeLists.txt#L19

If I switch that to 35, no problems

[russellfei]

@soumith Same catch here.

a = torch.rand(64,1024,12,12):cuda()
b = nn.SpatialMaxPooling(2,2,2,2):cuda():forward(a)
-- or -- 
b = nn.SpatialAdaptiveMaxPooling(8,8):cuda():forward(a)

gives the invalid argument error.
Maybe I should switch that arch to 35.
Thanks for instructions above.

--- updates
It works!

[bamos]

Hi, I have the same issue with SpatialAveragePooling and changing the arch to 35 in CMakeLists fixes the problem.