Matrix-Vector Multiplication (MVM) is a fundamental memory-bound Level-2 BLAS operation. The kernel drives the performance of various scientific applications, including 1) seismic imaging to reveal the subsurface layers for better monitoring the permafrost degradation or mitigating exploration and drilling risks for oil and gas industries, and 2) ground-based computational astronomy for supporting real-time simulations necessary to outsmart the atmospheric turbulence and help identifying exoplanets. We further leverage the inherent data sparsity structure of the resulting covariance matrices using Tile Low-Rank (TLR) matrix approximations. Our TLR-MVM outperforms its dense counterpart on many vendor architectures with high productivity in mind and maintains the numerical robustness of the applications.
We strongly recommend using spack to install TLR-MVM dependencies for x86_64 systems.
This includes Intel CPU, AMD EPYC CPU, NVIDIA GPU, AMD GPU in the following
installation sections.
We use [email protected], MKL or BLIS or OpenBLAS or cuBLAS or rocBLAS,
[email protected] to
build the library.
MKLoroneAPI MKLcan be used to installINTEL CPUBLISis used to install 'AMD EPYC CPU'cuBLASis used to install 'NVIDIA GPU'rocBLASus used to install 'AMD GPU'
Check install folder sapck.yaml and use the file to install dependencies.
A single-threaded BLAS (matrix vector multiplication) implementation is required.
One can use MKL, OpenBLAS or BLAS that comes with compilers, set MKL_ROOT to
let library find it.
MPI is optional but strongly recommended, set MPI_ROOT to let library find it.
NCCL is required to build the library with NVIDIA GPU, set NCCL_ROOT to let library
find it.
mkdir build && cd build
CC=gcc CXX=g++ cmake -DCMAKE_INSTALL_PREFIX=./install \
-DCMAKE_BUILD_TYPE=Release -DUSE_MPI=ON -DUSE_MKL=ON -DBUILD_TEST=ON ..
CC=gcc CXX=g++ cmake -DCMAKE_INSTALL_PREFIX=./install \
-DCMAKE_BUILD_TYPE=Release -DUSE_MPI=ON -DUSE_MKL=ON -DBUILD_TEST=ON ..
CC=gcc CXX=g++ cmake -DCMAKE_INSTALL_PREFIX=./install \
-DCMAKE_BUILD_TYPE=Release -DUSE_MPI=ON -DUSE_BLIS=ON -DBUILD_TEST=ON ..
CC=mpincc CXX=mpinc++ cmake -DCMAKE_INSTALL_PREFIX=./install \
-DCMAKE_BUILD_TYPE=Release -DUSE_COMPILER_BLAS=ON -DBUILD_TEST=ON \
-DUSE_MPI=ON ..
CC=mpincc CXX=mpinc++ cmake -DCMAKE_INSTALL_PREFIX=./install \
-DCMAKE_BUILD_TYPE=Release -DUSE_COMPILER_BLAS=ON -DBUILD_TEST=ON \
-DUSE_MPI=ON ..
CC=gcc CXX=g++ cmake -DCMAKE_INSTALL_PREFIX=./install \
-DCMAKE_BUILD_TYPE=Release \
-DCMAKE_CUDA_COMPILER:PATH=$CUDAToolkit_ROOT/bin/nvcc \
-DUSE_MKL=ON -DBUILD_CUDA=ON -DBUILD_TEST=ON ..
CC=gcc CXX=g++ cmake -DCMAKE_INSTALL_PREFIX=./install \
-DCMAKE_BUILD_TYPE=Release \
-DCMAKE_CUDA_COMPILER:PATH=$CUDAToolkit_ROOT/bin/nvcc \
-DUSE_MKL=ON -DBUILD_CUDA=ON -DBUILD_TEST=ON ..
Compile and install
make install -j
You also need to download the dataset to run the experiments. dataset download url:
- seismic redatuming dataset https://zenodo.org/record/6582600
- MAVIS AO system matrcies dataset https://zenodo.org/record/7305622
after download, put it in a seperate folder and set WORK_ROOT to that folder.
in install/test folder, you can try to launch bash file. These are the test files.
benchmark folder offers dense matrix vector multiplication benchmark tools.
We offer Makefile to compile the benchmarks since TLR-MVM requires single-threaded
BLAS while benchmark use threaded version BLAS.
Please check the corresponding environment variables in the Makefile and compile.
Currently, we suggest one to use our python library with NVIDIA GPU. To install it,
BUILD_CUDA=ON python setup.py build
This will create a build directory and build library inside it. After installation, add the library path (build/libxxx) to your
Y. Hong, M. Ravasi, H. Ltaief, D. Keyes, Can tile low-rank compression live up to expectation? An application to 3D multi-dimensional deconvolution, 2023, SEG IMAGE 2023 International Meeting for Applied Geoscience & Energy (extended abstract).
Y. Hong, H. Ltaief, M. Ravasi, D. Keyes, HPC Seismic Redatuming by Inversion with Algebraic Compression and Multiple Precisions, 2023, KAUST Repo Preprint.
H. Ltaief, Y. Hong, L. Wilson, M. Jacquelin, M. Ravasi and D. Keyes, Scaling the “Memory Wall” for Multi-Dimensional Seismic Processing with Algebraic Compression on Cerebras CS-2 Systems, 2023, Proceedings of the International Conference for High Performance Computing, Networking, Storage, and Analysis (SC’23), IEEE Computer Society (ACM Gordon Bell Finalist).
H. Ltaief, Y. Hong, A. Dabah, R. Alomairy, S. Abdulah, C. Goreczny, P. Gepner, M. Ravasi, D. Gratadour and D. Keyes, Steering Customized AI Architectures for HPC Scientific Applications, 2023, Springer International Supercomputing Conference (ISC’23) (A. Bhatele et al., eds.), Lecture Notes in Computer Science, Vol. 13948, pp. 125–143, doi 10.1007/978-3-031-32041-5 7.
M. Ravasi, Y. Hong, H. Ltaief and D. Keyes, Tile-Low Rank Compressed Multi-Dimensional Convolution and Its Application to Seismic Redatuming Problems, 2022, 83rd EAGE Annual Conference, doi 10.3997/2214-4609.202210253 (extended abstract).
M. Ravasi, Y. Hong, H. Ltaief, D. Keyes and D. Vargas, Large-Scale Marchenko Imaging with Distance-Aware Matrix Reordering, Tile Low-Rank Compression, and Mixed-Precision Computations, 2022, SEG IMAGE 2022 International Meeting for Applied Geoscience & Energy (extended abstract).
H. Ltaief, J. Cranney, D. Gratadour, Y. Hong, L. Gatineau, and D. Keyes, Meeting the Real-Time Challenges of Ground-Based Telescopes Using Low-Rank Matrix Computations, 2021, ACM/IEEE International Conference for High Performance Computing, Networking, Storage, and Analysis (SC’21), doi 10.1145/3458817.3476225.
Y. Hong, H. Ltaief, M. Ravasi, L. Gatineau and D. E. Keyes, Accelerating Seismic Redatuming Using Tile Low-Rank Approximations on NEC SX-Aurora TSUBASA, 2021, Supercomputing Frontiers and Innovations 8:6–26, doi 0.14529/jsfi210201.
H. Zhang, J. Cranney, N. Doucet, Y. Hong, D. Gratadour, H. Ltaief, D. Keyes and F. Rigaut, Predictive Learn and Apply: MAVIS application – Learn, 2020, in Proceedings of SPIE 11448, Adaptive Optics Systems VII, 114482L (extended abstract).
If you have any troubles, please create an issue or send email to [email protected] / [email protected].