DD-AVX Library: Library of High Precision Sparse Matrix Operations Accelerated by SIMD

About

DD-AVX_v3 is SIMD accelerated simple interface high precision BLAS Lv.1 and Sparse BLAS operations Library.

These operations can be performed by combining double and double-double precision.

Multi-threading (OpenMP) and SIMD AVX / AVX2 / AVX512 acceleration are available.

This library provides an easy way to implement a fast and accurate Krylov subspace method.

This library is extensions of Lis_DD_AVXKernals and DD-AVX_v2 (archived).

Interface

This library provides BLAS Lv.1 and Sparse BLAS functions.

For acceleration, this library provides AoS format double-double precision vector.

This library provides the following six types.

Scalar

• d_real (alias of double)
• dd_real (provided by the QD Library)

Vector

• d_real_vector
• dd_real_vector (AoS format)

Sparse matrix (CRS format)

• d_real_SpMat

It has BLAS Lv.1 and Sparse BLAS functions for these types.

All combinations of BLAS functions are implemented. It works for both D and DD types.

See the axpy sample code and CG method sample code for more information on how to use it.

Build and Install

This library requires the QD library for scalar operations as a submodule. The QD library is downloaded and built automatically by make.

You can specify the destination directory with DD_AVX_DIR and make . The QD libraries are installed in the same directory.

You can build and install the QD library with the following commands:

System Requirements

• g++ 7.1 or higher (C++17)
• gfortran (to build QD)
• GNU make

Build for each SIMD instruction

AVX

make avx

make install

AVX2

make avx2

make install

AVX512

make avx512

make install

AVX512 testing is inadequate.

Since github actions's runner does not support AVX512, github actions only verifies the build. We performed manual testing on an Intel® Xeon® Platinum 8280.

Document

https://hishinuma-t.dev/DD-AVX_v3/ (generated by Doxygen)

Current Status and Restrictions

This is a beta version, and there are some restrictions and changes planned.

The detailed todo is discussed in Issue

• SIMD and OpenMP cannot be disabled. (If you want to change the number of OpenMP threads, do so in the environment variable.)
• The class design will be modified to implement element/row/column operations in the sparse matrix class.
• (SIMD_REG class is difficult to share with Scalar, so I want to change it to REG class.)
• The conversion routine to BCRS format doesn't work because I'm currently trying to rework it to make it multi-threaded.

Test and sample

test

We have a complete set of tests for each feature in the test/ directory.

cd test/

make

make test_all

sample

We have implemented an example of the CG method, which is available in sample/.

Publications

• Hishinuma_t's Dr. thesis
• Implementation of AVX512 thanks to Mr. Itsuki Doi
• and more