FBGEMM (Facebook GEneral Matrix Multiplication) is a low-precision, high-performance matrix-matrix multiplications and convolution library for server-side inference.
The library provides efficient low-precision general matrix multiplication for small batch sizes and support for accuracy-loss minimizing techniques such as row-wise quantization and outlier-aware quantization. FBGEMM also exploits fusion opportunities in order to overcome the unique challenges of matrix multiplication at lower precision with bandwidth-bound operations.
FBGEMM is used as a backend of Caffe2 and PyTorch quantized operators for x86 machines:
- Caffe2: https://github.com/pytorch/pytorch/tree/master/caffe2/quantization/server
- PyTorch: https://github.com/pytorch/pytorch/tree/master/aten/src/ATen/native/quantized/cpu
- New Features and Recent Improvements (January, 2020)
The tests (in test folder) and benchmarks (in bench folder) are some great examples of using FBGEMM. For instance, SpMDMTest test in test/PackedRequantizeAcc16Test.cc shows how to combine row offset calculations with packing of A (PackAWithRowOffset), how to pack B matrix (PackBMatrix) and construct output pipeline (sparse_matrix*dense_matrix --> requantization --> nop) fused with inner GEMM macro kernel.
FBGEMM uses the standard CMAKE-based build flow.
FBGEMM requires gcc 5+ and a CPU with support for avx2 instruction set or higher. It's been tested on Mac OS X and Linux.
With inner kernels, FBGEMM takes a “one size doesn't fit all” approach, so the implementation dynamically generates efficient matrix-shape specific vectorized code using a third-party library called asmjit. asmjit is required to build FBGEMM.
FBGEMM detects CPU instruction set support at runtime using cpuinfo library and dispatches optimized kernels for the detected instruction set. Therefore, cpuinfo is required to detect CPU type.
googletest is required to build and run FBGEMM's tests. googletest is not required if you don't want to run FBGEMM tests. By default, building of tests is on. Turn it off by setting FBGEMM_BUILD_TESTS to off.
You can download asmjit, cpuinfo, googletest and set ASMJIT_SRC_DIR, CPUINFO_SRC_DIR, GOOGLETEST_SOURCE_DIR respectively for cmake to find these libraries. If any of these variables is not set, cmake will build the git submodules found in the third_party directory.
FBGEMM, in general, does not have any dependency on Intel MKL. However, for performance comparison, some benchmarks use MKL functions. If MKL is found or MKL path is provided with INTEL_MKL_DIR benchmarks are built with MKL and performance numbers are reported for MKL functions as well. However, if MKL is not found, the benchmarks are not built.
General build instructions are as follows:
git clone --recursive https://github.com/pytorch/FBGEMM.git
cd FBGEMM
# if you are updating an existing checkout
git submodule sync
git submodule update --init --recursive
mkdir build && cd build
cmake ..
make
To run the tests after building FBGEMM (if tests are built), use the following command:
make test
make install
For a high-level overview, design philosophy and brief descriptions of various parts of FBGEMM please see our blog.
We have extensively used comments in our source files. The best and up-do-date documentation is available in the source files.
You can also turn on the option to generate the documentation (using Doxygen and Sphinx by setting FBGEMM_BUILD_DOCS to ON, and then follow the above cmake build process.
For those looking for the appropriate article to cite regarding FBGEMM, we recommend citing our paper:
@article{fbgemm,
title={FBGEMM: Enabling High-Performance Low-Precision Deep Learning Inference},
author={Khudia, Daya and Huang, Jianyu and Basu, Protonu and Deng, Summer and Liu, Haixin and Park, Jongsoo and Smelyanskiy, Mikhail},
journal={arXiv preprint arXiv:2101.05615},
year={2021}
}
See the CONTRIBUTING file for how to help out.
FBGEMM is BSD licensed, as found in the LICENSE file.