hiptensor is AMD's C++ library for accelerating tensor primitives based on the composable kernel library, through general purpose kernel languages, like HIP C++.

• AMD CDNA class GPU featuring matrix core support: gfx908, gfx90a as 'gfx9'

Note: Double precision FP64 datatype support requires gfx90a

• ROCm stack minimum version 5.7
• ROCm-cmake minimum version 0.8.0 for ROCm 5.7
• C++ 17
• CMake >=3.6
• Composable Kernel

Optional:

• doxygen (for building documentation)

Run the steps below to build documentation locally.

cd docs

pip3 install -r .sphinx/requirements.txt

python3 -m sphinx -T -E -b html -d _build/doctrees -D language=en . _build/html

Operations - Contraction Tensor Data Types - FP32 , FP64

1. Create and track a hipTensor fork.
2. Clone your fork:

git clone -b develop https://github.com/<your_fork>/hipTensor.git .
.githooks/install
git checkout -b <new_branch>
...
git add <new_work>
git commit -m "What was changed"
git push origin <new_branch>
...

3. Create a pull request to ROCmSoftwarePlatform/hipTensor develop branch.
4. Await CI and approval feedback.
5. Once approved, merge!

Note: Please don't forget to install the githooks as there are triggers for clang formatting in commits.

By default, the project is configured as Release mode. Here are some of the examples for the configuration:

After configuration, build with cmake --build <build_dir> -- -j<nproc>

• Target a specific GPU (e.g. gfx908:xnack-)
• Use lots of threads (e.g. -j64)

Tests API implementation of logger verbosity and functionality. o <build_dir>/bin/logger_test

Tests the API implementation of bilinear contraction algorithm with validation. o <build_dir>/bin/bilinear_contraction_f32_test o <build_dir>/bin/bilinear_contraction_f64_test

Tests the API implementation of scale contraction algorithm with validation. o <build_dir>/bin/scale_contraction_f32_test o <build_dir>/bin/scale_contraction_f64_test

These are stand-alone use-cases of the hipTensor contraction operations.

Demonstrates the API implementation of bilinear contraction operation without validation. o <build_dir>/bin/simple_contraction_bilinear_f32

Demonstrates the API implementation of scale contraction operation without validation. o <build_dir>/bin/simple_contraction_scale_f32

Client application links to hipTensor library, and therefore hipTensor library needs to be installed before building client applications.

mkdir -p samples/build
cd samples/build

cmake                                                                                                  \
-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc                                                              \
-D CMAKE_PREFIX_PATH="/opt/rocm;${PATH_TO_HIPTENSOR_INSTALL_DIRECTORY};${PATH_TO_CK_INSTALL_DIRECTORY} \
..

 make