Run soumith's benchmarking code on a K520, so can use a K520 as a proxy

[hughperkins]

Run soumith's benchmarking code on a K520, so can use a K520 as a proxy for guesstimating results on a Titan/Titan-x in the future.

[hughperkins]

On a K520, the output in results.txt of running python/benchmarking/deepcl_benchmark.py is:

ConvolutionalLayer{ LayerDimensions{ inputPlanes=3 inputImageSize=128 numFilters=96 filterSize=11 outputImageSize=118 padZeros=0 biased=1 skip=0} LINEAR }, forward: 1979.45349216ms
ConvolutionalLayer{ LayerDimensions{ inputPlanes=3 inputImageSize=128 numFilters=96 filterSize=11 outputImageSize=118 padZeros=0 biased=1 skip=0} LINEAR }, backward: 6590.74971676ms
ConvolutionalLayer{ LayerDimensions{ inputPlanes=64 inputImageSize=64 numFilters=128 filterSize=9 outputImageSize=56 padZeros=0 biased=1 skip=0} LINEAR }, forward: 4468.90640259ms
ConvolutionalLayer{ LayerDimensions{ inputPlanes=64 inputImageSize=64 numFilters=128 filterSize=9 outputImageSize=56 padZeros=0 biased=1 skip=0} LINEAR }, backward: 19538.6034012ms
ConvolutionalLayer{ LayerDimensions{ inputPlanes=128 inputImageSize=32 numFilters=128 filterSize=9 outputImageSize=24 padZeros=0 biased=1 skip=0} LINEAR }, forward: 1447.01929092ms
ConvolutionalLayer{ LayerDimensions{ inputPlanes=128 inputImageSize=32 numFilters=128 filterSize=9 outputImageSize=24 padZeros=0 biased=1 skip=0} LINEAR }, backward: 2514.27898407ms
ConvolutionalLayer{ LayerDimensions{ inputPlanes=128 inputImageSize=16 numFilters=128 filterSize=7 outputImageSize=10 padZeros=0 biased=1 skip=0} LINEAR }, forward: 212.709188461ms
ConvolutionalLayer{ LayerDimensions{ inputPlanes=128 inputImageSize=16 numFilters=128 filterSize=7 outputImageSize=10 padZeros=0 biased=1 skip=0} LINEAR }, backward: 272.564506531ms
ConvolutionalLayer{ LayerDimensions{ inputPlanes=384 inputImageSize=13 numFilters=384 filterSize=3 outputImageSize=11 padZeros=0 biased=1 skip=0} LINEAR }, forward: 635.085391998ms
ConvolutionalLayer{ LayerDimensions{ inputPlanes=384 inputImageSize=13 numFilters=384 filterSize=3 outputImageSize=11 padZeros=0 biased=1 skip=0} LINEAR }, backward: 1474.16718006ms

Now I just need to interpret this, and relate it to Soumith's results on a Titan.

[hughperkins]

• created a Jenkins job to run this
• created a gh-pages location to display the results ie, http://hughperkins.github.io/DeepCL/benchmarking/
• next up: probably modify the jenkins job to run from a new branch 'jenkins-benchmarking', and to upload the results automatically, to ${version}-${git hash}

[hughperkins]

Got the jenkins to save results automatically, but probably need to order the results by date somehow.

[hughperkins]

Results in correct order now.