Using caffe without cuDNN

[ghost]

Hi there. Is it possible to use Caffe as a backend without cuDNN? (since it requires a Nvidia compute capability of 3.0, my gtx580 can't use it).

I tried commenting backend.cudnn_ctx = CuDNN.create() in backend.jl, but it, uhh, didn't work.

Thanks in advance.

$ julia mnist.jl 
12-Feb 19:33:24:INFO:root:Configuring Mocha...
12-Feb 19:33:24:INFO:root: * CUDA       enabled (MOCHA_USE_CUDA environment variable detected)
12-Feb 19:33:24:INFO:root: * Native Ext disabled by default
12-Feb 19:33:24:INFO:root:Mocha configured, continue loading module...
12-Feb 19:33:27:INFO:root:Initializing CuDNN backend...
12-Feb 19:33:27:INFO:root:CUDA backend initialized without cuDNN.
12-Feb 19:33:28:INFO:root:Constructing net MNIST-train on GPUBackend...
12-Feb 19:33:28:INFO:root:Topological sorting 8 layers...
12-Feb 19:33:28:INFO:root:Setup layers...
12-Feb 19:33:29:INFO:root:Network constructed!
12-Feb 19:33:29:INFO:root:Constructing net MNIST-test on GPUBackend...
12-Feb 19:33:29:INFO:root:Topological sorting 8 layers...
12-Feb 19:33:29:INFO:root:Setup layers...
12-Feb 19:33:29:DEBUG:root:ConvolutionLayer(conv1): sharing filters and bias
12-Feb 19:33:29:DEBUG:root:ConvolutionLayer(conv2): sharing filters and bias
12-Feb 19:33:29:DEBUG:root:InnerProductLayer(ip1): sharing weights and bias
12-Feb 19:33:29:DEBUG:root:InnerProductLayer(ip2): sharing weights and bias
12-Feb 19:33:29:INFO:root:Network constructed!
12-Feb 19:33:30:DEBUG:root:Checking network topology for back-propagation
12-Feb 19:33:30:DEBUG:root:Init network MNIST-train
12-Feb 19:33:30:DEBUG:root:Init parameter filter for layer conv1
12-Feb 19:33:30:DEBUG:root:Init parameter bias for layer conv1
12-Feb 19:33:30:DEBUG:root:Init parameter filter for layer conv2
12-Feb 19:33:30:DEBUG:root:Init parameter bias for layer conv2
12-Feb 19:33:30:DEBUG:root:Init parameter weight for layer ip1
12-Feb 19:33:30:DEBUG:root:Init parameter bias for layer ip1
12-Feb 19:33:30:DEBUG:root:Init parameter weight for layer ip2
12-Feb 19:33:30:DEBUG:root:Init parameter bias for layer ip2

signal (11): Segmentation fault
unknown function (ip: -1140687215)
cudnnConvolutionForward at /usr/local/lib/libcudnn.so.6.5 (unknown line)
convolution_forward at /home/notnikolaos/.julia/v0.3/Mocha/src/cuda/cudnn.jl:41
jlcall_convolution_forward_20757 at  (unknown line)
jl_apply_generic at /usr/bin/../lib/x86_64-linux-gnu/julia/libjulia.so (unknown line)
forward at /home/notnikolaos/.julia/v0.3/Mocha/src/cuda/layers/convolution.jl:83
jlcall_forward_20752 at  (unknown line)
jl_apply_generic at /usr/bin/../lib/x86_64-linux-gnu/julia/libjulia.so (unknown line)
forward at /home/notnikolaos/.julia/v0.3/Mocha/src/net.jl:137
jlcall_forward_20729 at  (unknown line)
jl_apply_generic at /usr/bin/../lib/x86_64-linux-gnu/julia/libjulia.so (unknown line)
solve at /home/notnikolaos/.julia/v0.3/Mocha/src/solvers.jl:222
jl_apply_generic at /usr/bin/../lib/x86_64-linux-gnu/julia/libjulia.so (unknown line)
unknown function (ip: -917840712)
unknown function (ip: -917844544)
unknown function (ip: -917777094)
unknown function (ip: -917774387)
jl_load at /usr/bin/../lib/x86_64-linux-gnu/julia/libjulia.so (unknown line)
include at ./boot.jl:245
jl_apply_generic at /usr/bin/../lib/x86_64-linux-gnu/julia/libjulia.so (unknown line)
include_from_node1 at loading.jl:128
jl_apply_generic at /usr/bin/../lib/x86_64-linux-gnu/julia/libjulia.so (unknown line)
process_options at ./client.jl:285
_start at ./client.jl:354
jlcall__start_17289 at /usr/bin/../lib/x86_64-linux-gnu/julia/sys.so (unknown line)
jl_apply_generic at /usr/bin/../lib/x86_64-linux-gnu/julia/libjulia.so (unknown line)
unknown function (ip: 4200623)
julia_trampoline at /usr/bin/../lib/x86_64-linux-gnu/julia/libjulia.so (unknown line)
unknown function (ip: 4199613)
__libc_start_main at /lib/x86_64-linux-gnu/libc.so.6 (unknown line)
unknown function (ip: 4199667)
unknown function (ip: 0)
Segmentation fault (core dumped)

[pluskid]

@notnikolaos Did you mean using GPU (instead of Caffe) as backend without cuDNN? Currently it is impossible to do GPU computation without cuDNN (unless you do not use convolution and pooling).

But I have a tentative plan for implementing fallback operations for CUDA when cuDNN is not available. It will be slower than the cuDNN but still faster than CPU backends.

[ghost]

Oops, yes, sorry, I wrongly thought Mocha called Caffe instead of cuDNN. You actually state it very clearly in the Readme, so I'm closing this 😊.