LATER

Linear Algebra on TEnsoRcore; see http://www2.cs.uh.edu/~panruowu/later.html

Prerequisites

• CMake 3.12+
• CUDA 10.1+
• CUTLASS 2.1+

Build

On Linux,

$ git clone git@github.com:Orgline/LATER.git
$ cd LATER && mkdir build && cd build
$ export CUDACXX=/usr/local/cuda-10.1/bin/nvcc
$ export CUDA_PATH=/usr/local/cuda-10.1
$ export CUTLASS_DIR=<CUTLASS Diretory> 
$ # for exmaple ~/cutlass-2.1.0
$ cmake .. -DCMAKE_CUDA_FLAGS="-gencode=arch=compute_80,code=sm_80" -DCUDA_ARCH="Ampere"
$ # On Volta, 75->70, Turing->Volta
$ cmake --build .

Change the CUDACXX and CUDA_PATH environment variables to match your system's CUDA installation directory.

On Windows,

$ git clone git@github.com:Orgline/LATER.git
$ cd LATER && mkdir build && cd build
$ cmake .. -A x64
$ cmake --build .

Run tests

On Linux

$ cd test
$ ./test_qr 1 16384 16384 -check

On Windows:

$ cd test/debug
$ test_qr.exe 1 16384 16384 -check

Tested GPUs and Platforms

• V100 (on RHEL Linux 7, CUDA 10.1, GCC 8)
• Titan V (on Ubuntu 18.04 Linux, CUDA 10.1, GCC 7.5.0)
• GeForce RTX 2060 (on Windows 10, CUDA 10.2, Visual Studio 2017)
• GeForce RTX 2080 Super (Ubuntu 18.04 Linux, CUDA 10.2, GCC 7.5.0)