MatX - GPU-Accelerated Numerical Computing in Modern C++

MatX is a modern C++ library for numerical computing on NVIDIA GPUs and CPUs. Near-native performance can be achieved while using a simple syntax common in higher-level languages such as Python or MATLAB.

The above image shows the Python (Numpy) version of an FFT resampler next to the MatX version. The total runtimes of the NumPy version, CuPy version, and MatX version are shown below:

• Python/Numpy: 5360ms (Xeon(R) CPU E5-2698 v4 @ 2.20GHz)
• CuPy: 10.6ms (A100)
• MatX: 2.54ms (A100)

While the code complexity and length are roughly the same, the MatX version shows a 2100x over the Numpy version, and over 4x faster than the CuPy version on the same GPU.

Key features include:

• ⚡ MatX is fast. By using existing, optimized libraries as a backend, and efficient kernel generation when needed, no hand-optimizations are necessary

• 👐 MatX is easy to learn. Users familiar with high-level languages will pick up the syntax quickly

• 📑 MatX easily integrates with existing libraries and code

• 🎇 Visualize data from the GPU right on a web browser

• ↕️ IO capabilities for reading/writing files

Table of Contents

• Requirements
• Installation
  • Building MatX
  • Integrating MatX With Your Own Projects
• Documentation
  • Supported Data Types
• Unit Tests
• Quick Start Guide
• Release History
• Filing Issues
• Contributing Guide

Requirements

MatX support is currently limited to Linux only due to the time to test Windows. If you'd like to voice your support for native Windows support using Visual Studio, please comment on the issue here: NVIDIA#153.

Note: CUDA 12.0.0 through 12.2.0 have an issue that causes building MatX unit tests to show a compiler error or cause a segfault in the compiler. Please use CUDA 11.4-11.8 or CUDA 12.2.1+ with MatX.

MatX is using features in C++17 and the latest CUDA compilers and libraries. For this reason, when running with GPU support, CUDA 11.4 and g++9 or newer is required. You can download the CUDA Toolkit here.

MatX has been tested on and supports Pascal, Turing, Volta, Ampere, and Hopper GPU architectures. Jetson products are supported with Jetpack 5.0 or above.

Note for CPU/Host support: CPU/Host execution is considered beta. Only operator execution is supported right now, but no functions that require libraries (FFT/GEMM, etc). If you find a bug in an operator on CPU, please report it in the issues above.

Installation

MatX is a header-only library that does not require compiling for using in your applications. However, building unit tests, benchmarks, or examples must be compiled. CPM is used as a package manager for CMake to download and configure any dependencies. If MatX is to be used in an air-gapped environment, CPM can be configured to search locally for files. Depending on what options are enabled, compiling could take very long without parallelism enabled. Using the -j flag on make is suggested with the highest number your system will accommodate.

Building MatX

To build all components, issue the standard cmake build commands in a cloned repo:

mkdir build && cd build
cmake -DMATX_BUILD_TESTS=ON -DMATX_BUILD_BENCHMARKS=ON -DMATX_BUILD_EXAMPLES=ON -DMATX_BUILD_DOCS=OFF ..
make -j

By default CMake will target the GPU architecture(s) of the system you're compiling on. If you wish to target other architectures, pass the CMAKE_CUDA_ARCHITECTURES flag with a list of architectures to build for:

cmake .. -DCMAKE_CUDA_ARCHITECTURES="80;90"

By default nothing is compiled. If you wish to compile certain options, use the CMake flags below with ON or OFF values:

MATX_BUILD_TESTS
MATX_BUILD_BENCHMARKS
MATX_BUILD_EXAMPLES
MATX_BUILD_DOCS

For example, to enable unit test building:

mkdir build && cd build
cmake -DMATX_BUILD_TESTS=ON ..
make -j

Note that if documentation is selected all other build options are off. This eases the dependencies needed to build documentation so large libraries such as CUDA don't need to be installed.

Integrating MatX With Your Own Projects

MatX uses CMake as a first-class build generator, and therefore provides the proper config files to include into your own project. There are typically two ways to do this:

1. Adding MatX as a subdirectory
2. Installing MatX to the system

MatX as a Subdirectory

Adding the subdirectory is useful if you include the MatX source into the directory structure of your project. Using this method, you can simply add the MatX directory:

add_subdirectory(path/to/matx)

MatX Installed to the System

The other option is to install MatX and use the configuration file provided after building. This is typically done in a way similar to what is shown below:

cd /path/to/matx
mkdir build && cd build
cmake ..
make && make install

If you have the correct permissions, the headers and cmake packages will be installed on your system in the expected paths for your operating system. With the package installed you can use find_package as follows:

find_package(matx CONFIG REQUIRED)

An example of using this method can be found in the examples/cmake_sample_project directory

MatX CMake Targets

Once either of the two methods above are done, you can use the transitive target matx::matx in your library inside of target_link_libraries. MatX may add other optional targets in the future inside the matx:: namespace as well.

Documentation

Documentation for MatX can be built locally as shown above with the DBUILD_DOCS=ON cmake flag. Building documentation requires the following to be installed: doxygen, breathe, sphinx, sphinx-rtd-theme, libjs-mathjax, texlive-font-utils, flex, bison

• Current documentation can be found here
• A quick start guide can be found here
• Current library limitations are listed here
• A conversion from MATLAB and Python syntax is found here
• A self-guided Jupyer notebook training can be found here

MatX uses semantic versioning and reserve the right to introduce breaking API changes on major releases.

Supported Data Types

MatX supports all types that use standard C++ operators for math (+, -, etc). Unit tests are run against all common types shown below.

• Integer: int8_t, uint8_t, int16_t, uint16_t, int32_t, uint32_t, int64_t, uint64_t
• Floating Point: matxFp16 (fp16), matxBf16 (bfloat16), float, double
• Complex: matxfp16Complex, matxBf16Complex, cuda::std::complex<float>, cuda::std::complex<double>

Since CUDA half precision types (__half and __nv_bfloat16) do not support all C++ operators on the host side, MatX provides the matxFp16 and matxBf16 types for scalars, and matxFp16Complex and matxBf16Complex for complex types. These wrappers are needed so that tensor views can be evaluated on both the host and device, regardless of CUDA or hardware support. When possible, the half types will use hardware- accelerated intrinsics automatically. Existing code using __half and __nv_bfloat16 may be converted to the matx equivalent types directly and leverage all operators.

Unit Tests

MatX contains a suite of unit tests to test functionality of the primitive functions, plus end-to-end tests of example code. MatX uses pybind11 to generate some of the unit test inputs and outputs. This avoids the need to store large test vector files in git, and instead can be generated as-needed.

To run the unit tests, from the cmake build directory run:

test/matx_test

This will execute all unit tests defined. If you wish to execute a subset of tests, or run with different options, you may run test/matx_test directly with parameters defined by Google Test. To run matx_test directly, you must be inside the build/test directory for the correct paths to be set. For example, to run only tests with the name FFT:

cd build/test
./matx_test --gtest_filter="*FFT*"

Quick Start Guide

We provide a variety of training materials and examples to quickly learn the MatX API.

• A quick start guide can be found in the docs directory or from the main documentation site. The MatX quick start guide is modeled after NumPy's and demonstrates how to manipulate and create tensors.
• A set of MatX notebooks can be found in the docs directory. These four notebooks walk through the major MatX features and allow the developer to practice writing MatX code with guided examples and questions.
• Finally, for new MatX developers, browsing the example applications can provide familarity with the API and best practices.

Release Major Features

v0.5.0:

• Polyphase resampler
• Documentation overhaul with examples for each function

v0.4.0:

• SVD power iteration function
• sign, and index operators
• Batched QR for small matrices
• dlpack export support
• 16-bit float reductions
• Output iterator support in CUB

v0.3.0:

• Many new operators, including flatten, remap, lcollapse. rcollapse, fmod, clone, slice
• Extended N-D tensor support to more functions
• Allow operators on reduction inputs
• g++11 support
• NVTX support
• Many, many bug fixes

Discussions

We have an open discussions board here. We encourage any questions about the library to be posted here for other users to learn from and read through.

Filing Issues

We welcome and encourage the creation of issues against MatX. When creating a new issue, please use the following syntax in the title of your submission to help us prioritize responses and planned work.

• Bug Report: Append [BUG] to the beginning of the issue title, e.g. [BUG] MatX fails to build on P100 GPU
• Documentation Request: Append [DOC] to the beginning of the issue title
• Feature Request: Append [FEA] to the beginning of the issue title
• Submit a Question: Append [QST] to the beginning of the issue title

As with all issues, please be as verbose as possible and, if relevant, include a test script that demonstrates the bug or expected behavior. It's also helpful if you provide environment details about your system (bare-metal, cloud GPU, etc).

Contributing Guide

Please review the CONTRIBUTING.md file for information on how to contribute code and issues to MatX. We require all pull requests to have a linear history and rebase to main before merge.