search

torch / cutorch

A CUDA backend for Torch7
Other
336 stars 208 forks source link

Tegra K1 #65

Open culurciello opened 10 years ago

culurciello commented 10 years ago

Hi guys, we are trying to run some Torch code on the Tegra K1 board: http://www.nvidia.com/object/tegra-k1-processor.html We are running into problems with cunn allocating too many resources, possibly because it is targeted to larger GPUs. The Tegra only has 192 CUDA cores.

error in SpatialMaxSampling.updateOutput: too many resourced requested for launch

I was wondering if you can give me some guidance on how to lower the resources.

Note: SpatialConvolutionMM works, possibly because BLAS is adjusted to the GPU specifications, but SpatialMaxPooling (not SpatialMaxPoolingCUDA) gives the error.

Possibly there needs to be a check for GPU resources in that function.

soumith commented 10 years ago

the launch sizes of the kernels are probably not suited. did you try using CuDNN?