CODEWITHSUNDEEP
macosdockercaffe
WestCoastProjects
WestCoastProjects

Reputation: 63062

Error when Building GPU docker image for caffe: Unsupported gpu architecture 'compute_60'

An attempt to build the gpu based caffe docker image was unsuccessful. The command run:

docker build -t caffe:gpu standalone/gpu

It ran for about twenty minutes including compiling many numpy/scikit scripts to c-code via cython. However the build finally failed in an nvidia nvcc step:

[  1%] [  1%] Building NVCC (Device) object src/caffe/CMakeFiles/cuda_compile.dir/layers/./cuda_compile_generated_cudnn_pooling_layer.cu.o
Building NVCC (Device) object src/caffe/CMakeFiles/cuda_compile.dir/util/./cuda_compile_generated_im2col.cu.o
nvcc fatal   : Unsupported gpu architecture 'compute_60'
CMake Error at cuda_compile_generated_cudnn_pooling_layer.cu.o.cmake:206 (message):
  Error generating
  /opt/caffe/build/src/caffe/CMakeFiles/cuda_compile.dir/layers/./cuda_compile_generated_cudnn_pooling_layer.cu.o


make[2]: *** [src/caffe/CMakeFiles/cuda_compile.dir/layers/./cuda_compile_generated_cudnn_pooling_layer.cu.o] Error 1
make[2]: *** Waiting for unfinished jobs....
nvcc fatal   : Unsupported gpu architecture 'compute_60'
CMake Error at cuda_compile_generated_im2col.cu.o.cmake:206 (message):
  Error generating

/opt/caffe/build/src/caffe/CMakeFiles/cuda_compile.dir/util/./cuda_compile_generated_im2col.cu.o


make[2]: *** [src/caffe/CMakeFiles/cuda_compile.dir/util/./cuda_compile_generated_im2col.cu.o] Error 1
make[1]: *** [src/caffe/CMakeFiles/caffe.dir/all] Error 2
make: *** [all] Error 2

I am on El Capitan and running on a late 2013 MBPro that has the GT750m nvidia chipset.

Note: The cpu based caffe docker image from their github repo: https://github.com/BVLC/caffe docker was successfully built inside the caffe/docker directory on my mac:

docker build -t caffe:cpu standalone/cpu

.. and about ten minutes later ..

Successfully built bf71c19501e5

There is some mention of a different project arrayfire encountering similar error https://github.com/arrayfire/arrayfire/issues/1535 . However I lack the specific knowledge to troubleshoot the BVLC/caffe/docker commands to try to replicate what appears to be cuda7.5 vs cuda8.0 versioning issues.

Has anyone successfully built the gpu version of Caffe on docker container from mac?

Upvotes: 2

Views: 3096

Answers (1)

cerebrou
cerebrou

Reputation: 5540

In the Makefile.example, try commenting out the *_60 and *_61 lines (for compatibility with CUDA < 8.0). 

CUDA_ARCH := -gencode arch=compute_20,code=sm_20 \
        -gencode arch=compute_20,code=sm_21 \
        -gencode arch=compute_30,code=sm_30 \
        -gencode arch=compute_35,code=sm_35 \
        -gencode arch=compute_50,code=sm_50 \
        -gencode arch=compute_52,code=sm_52 \
        #-gencode arch=compute_60,code=sm_60 \
        #-gencode arch=compute_61,code=sm_61 \
        #-gencode arch=compute_61,code=compute_61

Upvotes: 2

Related Questions