A rudimentary wrapper around the fast Maxwell kernels for GEMM and convolution operations provided by nervanagpu
This is a work in progress, help is welcome! (see below)
Note that the Theano team at the LISA lab are separately working on integrating these kernels into the library themselves. However, they are focusing their efforts on the half-precision (fp16) kernels for now, and they are implementing half precision support (and support for these kernels) using the new Theano backend. I implemented these wrappers because I wanted something that was readily usable using the current Theano backend, so I've only wrapped the single precision (fp32) kernels.
check out this repository and add the nervana_theano module to your Python path.
You will need Theano (0.7 or the latest version from git) and nervanagpu.
You will need a NVIDIA Maxwell GPU for this code to run. These include all 900 series cards, and a select number of 700 series cards (as well as 800M series cards). However, the most popular 700 series cards like the 780Ti and the Titan are Kepler-based cards, so this code will not run on them.
If you wish to use the provided layer classes for Lasagne, you will need to install that as well.
You can use the gemm kernel in Theano as follows:
import theano.tensor as T
from nervana_theano.gemm import nervana_dot
x, y = T.matrices('x', 'y')
prod_cublas = T.dot(x, y)
prod_nervana = nervana_dot(x, y)
# these should give the same result
The Nervana convolution kernels support 1D, 2D and 3D convolutions. They use c01b or batch-size-last axis ordering, like the cuda-convnet kernels. This is different form Theano's default bc01 or batch-size-first ordering.
The nervana_conv wrapper function adds the necessary dimshuffles by default, so you can use it as a drop-in replacement for theano.tensor.nnet.conv.conv2d. However, this may degrade performance, so this behaviour can be turned off by passing the keyword argument dimshuffle=False. In that case, you will need to provide inputs with axes in c01b order.
You can use the convolution kernels in Theano as follows:
import theano.tensor as T
from nervana_theano.conv import nervana_conv
x = T.tensor4('x')
w = T.tensor4('w')
border_mode = 'valid' # or 'full', or...
# for nervana_conv this can also be a tuple of integers
conv_theano = T.nnet.conv.conv2d(x, w, border_mode=border_mode, subsample=strides)
conv_nervana = nervana_conv(x, w, padding=border_mode, strides=strides)
# or with manual shuffling:
x_shuffled = x.dimshuffle(1, 2, 3, 0) # batch size last
w_shuffled = w.dimshuffle(1, 2, 3, 0) # batch size last
conv_nervana = nervana_conv(x_shuffled, w_shuffled, padding=border_mode, strides=strides, dimshuffle=False)
Note that nervana_conv will perform a 1D, 2D or 3D convolution depending on whether the inputs are 3D, 4D or 5D tensors.
nervana_theano.layers.NervanaConvLayer is a Lasagne layer class that functions as a drop-in replacement for lasagne.layers.Conv1DLayer and lasagne.layers.Conv2DLayer (in addition it also supports 3D convolutions). Like lasagne.layers.cuda_convnet.Conv2DCCLayer it has a dimshuffle keyword argument to perform the necessary dimshuffles automatically. This is set to True by default and can be disabled as a performance optimization.
This is a work in progress, help / remarks / suggestions / pull requests are very welcome. I don't know if I will have much time to work on this in the near future. Here are some things left to do:
- implement the gradient for the GEMM wrapper
- add wrappers for the pooling kernels
- optimize the wrappers to avoid unnecessary computation and other slowdowns
- figure out what needs to change to optimize performance for the GTX 980 (instead of the GTX Titan X only), since more people have this card
- add support for the built-in ReLU operation of the kernels
- add Theano optimizations that replace theano.tensor.nnet.conv2d / theano.tensor.dot with the Nervana equivalents automatically
This code comes with no warranty or support, I just put it online because I figured other people might find it interesting / useful. Use it at your own risk. If something doesn't work you're welcome to submit an issue on GitHub, but I can't promise I'll be able to look at it or fix it.
Thanks to Scott Gray, Jan Schlüter and the Theano team for their help, and to Nervana Systems for open sourcing nervanagpu.