Matrix multiplication (GEMM) is a core operation to numerous scientific applications. Traditional implementations of Strassen-like fast matrix multiplication (FMM) algorithms often do not perform well except for very large matrix sizes, due to the increased cost of memory movement, which is particularly noticeable for non-square matrices. Such implementations also require considerable workspace and modifications to the standard BLAS interface. We propose a code generator framework to automatically implement a large family of FMM algorithms suitable for multiplications of arbitrary matrix sizes and shapes. By representing FMM with a triple of matrices [U,V,W] that capture the linear combinations of submatrices that are formed, we can use the Kronecker product to define a multi-level representation of Strassen-like algorithms. Incorporating the matrix additions that must be performed for Strassen-like algorithms into the inherent packing and micro-kernel operations inside GEMM avoids extra workspace and reduces the cost of memory movement. Adopting the same loop structures as high-performance GEMM implementations allows parallelization of all FMM algorithms with simple but efficient data parallelism without the overhead of task parallelism. We present a simple performance model for general FMM algorithms and compare actual performance of 20+ FMM algorithms to modeled predictions. Our implementations demonstrate a performance benefit over conventional GEMM on single core and multi-core systems. This study shows that Strassen-like fast matrix multiplication can be incorporated into libraries for practical use.
- Strassen's Algorithm Reloaded
- Generating Families of Practical Fast Matrix Multiplication Algorithms
- A Framework for Practical Parallel Fast Matrix Multiplication
- Strassen's Algorithm for Tensor Contraction
- Go to
metadirectory:
$ cd fmm-gen/meta
- Set up environment variables:
Replace
$core_numwith the number of cores the user wants to run.
$ export OMP_NUM_THREADS=$core_num
$ export KMP_AFFINITY=compact
Note: if hyper-threading is enabled, the following alternative must be used:
$ export KMP_AFFINITY=compact,1
- Code generators:
If you want to generate the different implementations for a specific algorithm:
$ python control.py ${N} $m1n1p1 $L1 $m2n2p2 $L2 ...... $m{N}n{N}p{N} $L{N} ${pack_type} ${gen_path}
e.g.
$ python control.py 2 222 1 323 1 abc ${HOME}/fmm-gen
$ python control.py 1 222 2 abc ../
This script will generate the code and compile it.
To further execute the code, go to the generated code directory (e.g. ${HOME}/fmm-gen/222-1_333-1_abc, or ../222-2_abc).
When $core_num is equal to 1,
run
./test/test_xxx-x_st.x $m $n $k
When $core_num is greater than 1,
run
./test/test_xxx-x_mt.x $m $n $k
If you have access of a job submission system on a cluster, change the path_prefix variable in config.py, then:
$ python run_sbatch_script.py
This script will generate the code for all implementations, compile them, and submit the jobs to SLURM submission queue for execution.
- Hybrid partitions:
$ python control.py 1 222 1 abc
$ python control.py 1 222 2 abc
$ python control.py 1 232 1 abc
$ python control.py 1 232 2 abc
$ python control.py 1 333 1 abc
$ python control.py 1 333 2 abc
$ python control.py 2 222 1 232 1 abc
$ python control.py 2 222 1 333 1 abc
- Model:
$ python model_gen.py
This script will generate csv files for plotting the modeled performance curves.
- Evaluation and expected result: The output will include the following components:
- Input problem size.
- Running time (in seconds).
- Effective GFLOPS (circle{1} in Figure 5).
The user can compare the relative Effective GFLOPS for different implementations. The trend should match the performance curves shown in this paper. Since the machines may be different from ours, the absolute GFLOPS could be different.
For those of you looking for the appropriate article to cite regarding fmm-gen, we recommend citing our IPDPS17 paper:
@inproceedings{FMM:IPDPS17,
author = {Jianyu Huang and Leslie Rice and Devin A. Matthews and Robert A. {v}an~{d}e~{G}eijn},
title = {Generating Families of Practical Fast Matrix Multiplication Algorithms},
booktitle = {31st IEEE International Parallel and Distributed Processing Symposium (IPDPS 2017)},
year = 2017,
}
This material was partially sponsored by grants from the National Science Foundation (Awards ACI-1550493), by Intel Corporation through an Intel Parallel Computing Center grant, and by a gift from Qualcomm Foundation.
Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation (NSF).
Bugs can be reported to jianyu@cs.utexas.edu. Any feedback, feature requests, and comments are welcome.