LibreTT - Tensor Transpose Library

LibreTT is a free Tensor Transpose Library for nVidia, AMD, and Intel GPU accelerators.

About

LibreTT is a portable Tensor Transpose library that incorporates and superseds the original cuTT library, https://github.com/ap-hynninen/cutt, which is no longer supported due to tragic death of its author, Antti-Pekka Hynninen in a car accident. Enhancements include adding thread safety by Valeyev Lab at Virginia Tech, https://github.com/ValeevGroup/cutt, incorporating AMD GPU support by Dmitry Lyakh, Oak Ridge National Laboratory, https://github.com/DmitryLyakh/hipTT followed by implementation of warpsize 64 by Alessandro Fanfarillo, Advanced Micro Devices, Inc., and adding SYCL support by Argonne National Laboratory.

Credits

cuTT     Design the algorithms: Dmitry Lyakh (ORNL), Antti-Pekka Hynninen (ORNL)
cuTT     Make the original implementation: Antti-Pekka Hynninen (ORNL)
cuTT     Implement thread safety: Eduard Valeyev (VT), Dmitry Lyakh (ORNL)
hipTT    Make the initial port of cuTT to HIP: Luke Roskop (HPE), Dmitry Lyakh (ORNL)
hipTT    Finalize the working version: Alessandro Fanfarillo (AMD)
LibreTT  SYCL port and CUDA,AMD,INTEL platform integration: Victor Anisimov (ANL)
LibreTT  CMake enhancements and complex-double datatype: Ajay Panyala (PNNL)
LibreTT  SYCL portability and remove dependency on DPCT: Abhishek Bagusetty (ANL)
LibreTT  Enhancements for HIP build: Eduard Valeyev (VT)
LibreTT  Fix the code for deprecations in C++17 and SYCL: Abhishek Bagusetty (ANL)

Directory Structure

CMakeLists.txt      CMake script
LICENSE.md          License to use this software
Makefile            Top-level Makefile
Makefile.cuda       Makefile for CUDA platform
Makefile.hip        Makefile for HIP  platform
Makefile.sycl       Makefile for SYCL platform
README.md           This file
src/                Source code directory

Manual Compilation

make cuda [tests=all]

make hip [tests=all]

make sycl [groupsize=16 | groupsize=32] [tests=all]

The default mode make cuda | hip | sycl will compile portable tests only, i.e. tests 1, 2, and 3.

The option tests=all instructs the build process to compile all tests in librett_test.cpp.

CUDA: Use make cuda tests=all to build all tests or make cuda to build only portable tests.

HIP: Use make hip to build only portable tests since non-portable tests 4 and 5 will fail for HIP platform.

SYCL: Use make sycl to build only portable tests. The groupsize=16 is default for SYCL platform.

The non-portable test 5 will fail with make sycl groupsize=16 tests=all.

Use make sycl groupsize=32 tests=all to check that all tests pass.

The outcome of manual build

• bin/ librett_test, librett_bench, example
• build/ temporary placeholder for object files
• include/ librett.h
• lib/ librett.a

CMake compilation:

Example of CUDA compilation: cmake -H. -Bbuild -DENABLE_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=70

Example of HIP compilation: cmake -H. -Bbuild -DENABLE_HIP=ON -DCMAKE_CXX_COMPILER=hipcc -DCMAKE_HIP_ARCHITECTURES=gfx908

Example of SYCL compilation: cmake -H. -Bbuild -DENABLE_SYCL=ON -DCMAKE_CXX_COMPILER=icpx

Testing options: -DENABLE_TESTS=ON (default)

Testing

Manual build: Execute bin/librett_test without arguments.
CMake build: Execute ctest

Description

To be added.