diff options
| -rw-r--r-- | README.md | 21 | ||||
| -rw-r--r-- | ROADMAP.md | 3 | ||||
| -rw-r--r-- | doc/details_gemm.md | 27 |
3 files changed, 42 insertions, 9 deletions
@@ -2,11 +2,11 @@ CLBlast: The tuned OpenCL BLAS library ================ -| | Build status | Tests on Intel CPU | Tests on NVIDIA GPU | Tests on AMD GPU | +| | Build status | Tests on Intel CPU | Tests on NVIDIA GPU | Tests on Intel GPU | |-----|-----|-----|-----|-----| -| Windows | [](https://ci.appveyor.com/project/CNugteren/clblast) | N/A | N/A | N/A | -| Linux | [](https://travis-ci.org/CNugteren/CLBlast/branches) | [](http://67.207.87.39:8010/#/builders/97) | [](http://67.207.87.39:8010/#/builders/98) | [](http://67.207.87.39:8010/#/builders/96) | -| OS X | [](https://travis-ci.org/CNugteren/CLBlast/branches) | [](http://67.207.87.39:8010/#/builders/101) | N/A | N/A | +| Windows | [](https://ci.appveyor.com/project/CNugteren/clblast) | [](http://ci.arrayfire.org:8010/#/builders/clblast-windows-intel-i7-4790k) | [](http://ci.arrayfire.org:8010/#/builders/clblast-windows-nvidia-k5000) | [](http://ci.arrayfire.org:8010/#/builders/clblast-windows-intel-HD4600) | +| Linux | [](https://travis-ci.org/CNugteren/CLBlast/branches) | [](http://ci.arrayfire.org:8010/#/builders/clblast-linux-intel-e5-2620-v4) | [](http://ci.arrayfire.org:8010/#/builders/clblast-linux-nvidia-k80) | N/A | +| OS X | [](https://travis-ci.org/CNugteren/CLBlast/branches) | [](http://ci.arrayfire.org:8010/#/builders/clblast-osx-intel-i5-4278U) | N/A | N/A | CLBlast is a modern, lightweight, performant and tunable OpenCL BLAS library written in C++11. It is designed to leverage the full performance potential of a wide variety of OpenCL devices from different vendors, including desktop and laptop GPUs, embedded GPUs, and other accelerators. CLBlast implements BLAS routines: basic linear algebra subprograms operating on vectors and matrices. See [the CLBlast website](https://cnugteren.github.io/clblast) for performance reports on various devices as well as the latest CLBlast news. @@ -78,6 +78,7 @@ More detailed documentation is available in separate files: * [Tuning for better performance](doc/tuning.md) * [Testing the library for correctness](doc/testing.md) * [Bindings / wrappers for other languages](doc/bindings.md) +* [More details on the GEMM kernel](doc/details_gemm.md) * [Glossary with some terms explained](doc/glossary.md) @@ -118,6 +119,7 @@ The main contributing authors (code, pull requests, testing) are: * [Dimitri Van Assche](https://github.com/dvasschemacq) * [Shehzan Mohammed](https://shehzan10.github.io) * [Marco Cianfriglia](https://github.com/mcian) +* [Kodonnell](https://github.com/kodonnell) * Everyone else listed as a [GitHub contributor](https://github.com/CNugteren/CLBlast/graphs/contributors) Tuning and testing on a variety of OpenCL devices was made possible by: @@ -127,11 +129,13 @@ Tuning and testing on a variety of OpenCL devices was made possible by: * [dividiti](http://www.dividiti.com) * [SURFsara HPC center](http://www.surfsara.com) * [ArrayFire](http://arrayfire.org) +* [TomTom](http://www.tomtom.com) * Everyone reporting [tuning results](https://github.com/CNugteren/CLBlast/issues/1) Hardware/software for this project was contributed by: -* [ArrayFire](http://arrayfire.org) for settings up and supporting Jenkins CI correctness tests on 7 platforms +* [HPC research group at the University of Bristol](http://uob-hpc.github.io/zoo/) for access to their GPU zoo +* [ArrayFire](http://arrayfire.org) for settings up and supporting Buildbot correctness tests on multiple platforms * [JetBrains](https://www.jetbrains.com/clion/) for supply a free CLion IDE license for CLBlast developers * [Travis CI](https://travis-ci.org/CNugteren/CLBlast/branches) and [AppVeyor](https://ci.appveyor.com/project/CNugteren/clblast) for free automated build tests for open-source projects @@ -141,12 +145,13 @@ More information Further information on CLBlast is available through the following links: -* A 20-minute presentation of CLBlast was given at the GPU Technology Conference in May 2017. A recording is available on the [GTC on-demand website](http://on-demand.gputechconf.com/gtc/2017/video/s7280-nugteren-clblast.mp4) (poor audio quality however) and a full slide-set is also available [as PDF](http://on-demand.gputechconf.com/gtc/2017/presentation/s7280-cedric-nugteren-clblast.pdf). -* More in-depth information and experimental results are also available in a scientific paper titled [CLBlast: A Tuned OpenCL BLAS Library](https://arxiv.org/abs/1705.05249) (May 2017). For CLTune, the inspiration for the included auto-tuner, see also the [CLTune: A Generic Auto-Tuner for OpenCL Kernels](https://arxiv.org/abs/1703.06503) paper. +* A 20-minute presentation of CLBlast was given at the GPU Technology Conference in May 2017. A recording is available on the [GTC on-demand website](http://on-demand.gputechconf.com/gtc/2017/video/s7280-nugteren-clblast.mp4) (poor audio quality however) and a full slide-set is also available [as PDF](http://on-demand.gputechconf.com/gtc/2017/presentation/s7280-cedric-nugteren-clblast.pdf). An updated version was also presented at IWOCL in May 2018. The slide set can be found [here as PDF](https://cnugteren.github.io/downloads/CLBlastIWOCL18.pdf). +* More in-depth information and experimental results are also available in a scientific paper titled [CLBlast: A Tuned OpenCL BLAS Library](https://arxiv.org/abs/1705.05249) (v1 May 2017, updated to v2 in April 2018). For CLTune, the inspiration for the included auto-tuner, see also the [CLTune: A Generic Auto-Tuner for OpenCL Kernels](https://arxiv.org/abs/1703.06503) paper. How to cite this work: - C. Nugteren. CLBlast: A Tuned OpenCL BLAS Library. ArXiv pre-print 1705.05249, 2017. + Cedric Nugteren. CLBlast: A Tuned OpenCL BLAS Library. In IWOCL'18: International Workshop + on OpenCL. ACM, New York, NY, USA, 10 pages. 2018. https://doi.org/10.1145/3204919.3204924 Support us @@ -18,7 +18,8 @@ This file gives an overview of the main features planned for addition to CLBlast | [#223](https://github.com/CNugteren/CLBlast/issues/223) | Feb '18 | CNugteren | ✔ | Python OpenCL interface | | [#237](https://github.com/CNugteren/CLBlast/issues/237) | Mar '18 | CNugteren | ✔ | Making tuning possible from the CLBlast API | | [#228](https://github.com/CNugteren/CLBlast/issues/228) | Mar-Apr '18 | CNugteren | ✔ | Improving performance for Qualcomm Adreno GPUs | -| [#270](https://github.com/CNugteren/CLBlast/issues/270) | May '18 | CNugteren | | Implement col2im | | [#267](https://github.com/CNugteren/CLBlast/issues/267) | May '18 | CNugteren | | Merge im2col and GEMM into a direct kernel | +| [#270](https://github.com/CNugteren/CLBlast/issues/270) | July '18 | CNugteren | | Implement col2im | +| - | July '18 | CNugteren | | Add a SYCL interface to the library | | [#136](https://github.com/CNugteren/CLBlast/issues/136) | ?? | CNugteren | | Implement xAXPBY and xSET | | [#169](https://github.com/CNugteren/CLBlast/issues/169) | ?? | dividiti | | Problem-specific tuning parameter selection | diff --git a/doc/details_gemm.md b/doc/details_gemm.md new file mode 100644 index 00000000..d4666abb --- /dev/null +++ b/doc/details_gemm.md @@ -0,0 +1,27 @@ +CLBlast: Details on the GEMM routine and kernel +================ + +This document gives a bit more detail on how the GEMM routine is organised and implemented. For other information about CLBlast, see the [main README](../README.md). + + +GEMM: Two approaches +------------- + +CLBlast implements two approaches to GEMM: direct and indirect: + +* Direct GEMM: Computing GEMM using a single generic kernel which handles all cases (e.g. all kinds of matrix sizes). +* Indirect GEMM: Computing GEMM using multiple kernels: the main GEMM kernel and a few pre-processing and post-processing kernels. The main kernel makes several assumptions (e.g. sizes need to be multiples of 32), which the other kernels make sure are satisfied. The main kernel is often faster than the generic kernel of the direct approach, but the cost of pre-processing and post-processing kernels can sometimes be high for small sizes or particular devices. + + +GEMM: In-direct approach +------------- + +Similar to the work by Matsumoto et al. ("Performance Tuning of Matrix Multiplication in OpenCL on Different GPUs and CPUs"), the main GEMM kernel makes many assumptions on the input arguments, which are handled by pre-processing and post-processing kernels. These assumptions are e.g. matrix sizes are a multiple of the work-group sizes, offsets are zero, and matrix B is transposed. This is a good solution for larger problem sizes since O(n^2) data movement is typically cheaper than O(n^3) computation, but the hidden constant starts to play a role for smaller n. Therefore, there is also a single-kernel direct version available for those cases, but it shares most of the design and parameters as discussed below. + +The main kernel has 14 different parameters, of which some are illustrated in figure 1 in the [CLBlast paper](https://arxiv.org/pdf/1705.05249). The parameters define among others the work-group sizes in 2 dimensions (MWG, NWG), the 2D register tiling configuration (MWI, NWI), the vector widths of both input matrices (VWM, VWN), loop unroll factors (KWI), and whether or not and how to use the local memory. + + +GEMM: Direct approach +------------- + +This is a single-kernel approach that shared many of the parameters for the in-direct kernel. One of the differences is that within the kernel there are checks for incomplete tiles in the m/n/k dimensions, influenced by the tuning parameters and the matrix sizes. These incomplete tiles will run a different part of the code, as they for example cannot benefit from vectorisation. Another difference is that there are dedicated kernels for each a/b transpose requirement: NN, NT, TN, TT for non-transposed and transposed.
\ No newline at end of file |