This is a port of LIBLINEAR.multicore to CUDA/GPU. Only linux is supported for the time being. Also, only dual l1/l2 classification is ported to the GPU.
On the GPU we use dense, not sparse data. This can be a very bad bottleneck, if the data is rather sparse. It is not uncommon for the GPU solver to be slower than the CPU version in this case! Please make sure to use this solver only for dense data.
Make sure you have all prerequisites, notably CUDA and openMP. LIBLINEAR.gpu uses CMake instead of simple Make. To compile, switch to the code directory and just issue:
$ cmake .
$ make all
This should produce a train binary under the current directory.
Simply execute
$ ./train heart_scale -s 1 -c 1 ./model
-
The GPU code might be still unoptimized, however it seems that the code is as fast as my 2x octa xeon e5-2670. As the GPU utiliziation (reported by nvidia-smi) seems to be very low (roughly 20%) it might be the serial code that keeps the GPU from flying.
-
Because of slight differences in floating point precision, the results of the GPU and CPU versions are not necessarily exactly the same. Also, the randomization is different, so one cannot expect the very same results.
-
Hyperthreading seems to hurt the multicore version of Liblinear. This should not hurt the GPU port though.
-
For now prediction is not touched at all, though it probably can be speed up too.
-
As small candidate sizes does not help gpu parallelization, we fixed this to be
- there seems to be little gain to start with a small size and increase it in our experiments.
- Only linux
- Only dual l1/l2 classification
This software is built upon the great liblinear-multicore library, which can be found here:
and uses CUDA in form of cuSPARSE (https://developer.nvidia.com/cusparse) and cuBLAS (https://developer.nvidia.com/cublas).
All code is licensed under the LIBLINEAR copyright:
Copyright (c) 2007-2015 The LIBLINEAR Project. All rights reserved.
Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met:
-
Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer.
-
Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution.
-
Neither name of copyright holders nor the names of its contributors may be used to endorse or promote products derived from this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.