Create CPU only Version

[sguada]

No description provided.

[Yangqing]

IMHO I wouldn't put this at a high priority. Plus it would be pretty nontrivial to implement this, requiring a lot of #ifdef macros as well as substantially rewriting the syncedmem class. We would also need to consider how we would like caffe::Caffe::mode to be... All these might make the codebase rather messy so I am not sure if it is the right move - at the end of day, it probably does not hurt too much to just ship with cuda libraries - unlike MKL, they are freely redistributable.

[sguada]

I agree that is not high priority, but I think that separating the cpu and gpu parts of code would be helpful for maintenance and eventually to make a CPU only version without needing CUDA to compile or run it.

[Yangqing]

The current code actually allows one to run without a physical GPU (as long
as cuda runtime is distributed - again it is freely allowed), which is what
I am planning to deploy on the ICSI cluster.

For developers working on caffe I assume they will at least have cuda
compilers.

Yangqing

On Tue, Nov 26, 2013 at 12:21 PM, Sergio Guadarrama <
notifications@github.com> wrote:

I agree that is not high priority, but I think that separating the cpu and
gpu parts of code would be helpful for maintenance and eventually to make a
CPU only version without needing CUDA to compile or run it.

—
Reply to this email directly or view it on GitHubhttps://github.com/Yangqing/caffe/issues/3#issuecomment-29330167
.

[tdomhan]

on Mac OS X 10.9 the cuda libaries are linked to libstdc++ while everything else on the system is linked to libc++ by default, due to the switch from gcc to clang. This way compiling caffe on OS X 10.9 is a huge mess right now, because you need to make sure that all the libraries you link to are linked to libstdc++, which mostly means you need to manually compile all the libraries and set the correct flags. The worst part is that you won't notice directly during compilation, but only when you try to run the program afterwards.
If there was a CPU only version, you could at least compile only the CPU code without a hassle, until nvcc will work with libc++ instead of stdlibc++.

[junwang4]

Hi tdomhan,
Have you succeeded to install it on your Mac OS X 10.9? I got an error:
clang: error: unsupported option '-dumpspecs'
Then it stops to compile.

[shelhamer]

The OSX 10.9 situation hasn't yet changed since @tdomhan's nice summary of the issue. OS X 10.9 is not currently a feasible compilation target for GPU mode. The -dumpspecs error you mention is due to CUDA/clang incompatibility, but it is not currently possible to compile with gcc either.

For CUDA compatibility on OS X it seems to mostly be a matter of waiting for a new version of CUDA to ship that links to libc++.

[sguada]

The problem of compiling with gcc seems to be that when linking it detects
duplicate functions between the cpp and the cuda code. If you figure out
how to fix that, then we will be a bit closer to compile with gcc in OS X
10.9

duplicate symbol caffe::LRNLayer::LRNLayer(caffe::LayerParameter
const&)in:
src/caffe/layers/lrn_layer.o
src/caffe/layers/lrn_layer.cuo

Sergio

2014-01-31 Evan Shelhamer notifications@github.com:

The OSX 10.9 situation hasn't yet changed since @tdomhanhttps://github.com/tdomhan's
nice summary of the issue. OS X 10.9 is not currently a feasible
compilation target for GPU mode. The -dumpspecs error you mention is due
to CUDA/clang incompatibility, but it is not currently possible to compile
with gcc either.

For CUDA compatibility on OS X it seems to mostly be a matter of waiting
for a new version of CUDA to ship that links to libc++.

Reply to this email directly or view it on GitHubhttps://github.com//issues/3#issuecomment-33861457
.

[tdomhan]

I managed to build and run caffe under Mac OS X. I'll post the details later. It's a little hacky of course.

[tdomhan]

Already I finally had time to add instructions for OS X 10.9, see: f0f594c
Hope this helps.

[junwang4]

Thanks, Tobias! Following your instructions, I succeeded in installing Caffe on OS X 10.9.

I did a test on the MNIST demo, but found that the GPU setting (running time: 275s) doesn't have any advantage over the CPU setting (running time: 284s) on a latest iMac. I only changed the setting "solver_mode: 1/0" in data/lenet_solver.prototxt. Do you have any comparison between the GPU and the CPU setting?

[Yangqing]

The MNIST demo is not likely to show the advantage of GPU over CPU, since
the model is very small and the overhead of e.g. data transfer and CPU side
control codes is big enough that GPU and CPU takes approximately the same
time. For larger models like ImageNet the GPU advantage will become clearer.

Yangqing

On Thu, Feb 6, 2014 at 8:39 PM, junwang4 notifications@github.com wrote:

Thanks, Tobias! Following your instructions, I succeeded in installing
Caffe on OS X 10.9.

I did a test on the MNIST demo, but found that the GPU setting (running
time: 275s) doesn't have any advantage over the CPU setting (running time:
284s) on a latest iMac. I only changed the setting "solver_mode: 1/0" in
data/lenet_solver.prototxt. Do you have any comparison between the GPU and
the CPU setting?

Reply to this email directly or view it on GitHubhttps://github.com//issues/3#issuecomment-34403732
.

[junwang4]

Yangqing,
Thanks for your clarification! Is there any comparison data of the running time between the CPU vs. GPU on ImageNet or CIFAR?

[Yangqing]

I have not got detailed analysis yet. On imagenet with GPUs, full
forward+backward using Alex Krizhevsky's network takes about 7ms, and
forward only takes about 2.5ms on my desktop with a K20 (when computation
are carried out in a batch fashion). CPUs are about 10 times slower than
that, but given the multiple choices of specific CPU types it might be hard
to say exactly what the speed it is.

Yangqing

On Thu, Feb 6, 2014 at 8:59 PM, junwang4 notifications@github.com wrote:

Yangqing,
Thanks for your clarification! Is there any comparison data of the running
time between the CPU vs. GPU on ImageNet or CIFAR?

Reply to this email directly or view it on GitHubhttps://github.com//issues/3#issuecomment-34404390
.

[junwang4]

Thanks - just noticed this from your CAFFE presentation
https://docs.google.com/presentation/d/1lzyXMRQFlOYE2Jy0lCNaqltpcCIKuRzKJxQ7vCuPRc8/edit#slide=id.g3c48e0e3ae4e94440

[kloudkl]

A not very accurate benchmark of CPU and GPU mode is done in #85. Even the result is not fair for GPU, it still shows great speed advantage.

[shelhamer]

Done in #561!