aCe - Data-Centric Parallel Programming

Decoupling domain science from performance optimization.

DaCe is a parallel programming framework that takes code in Python/NumPy and other programming languages, and maps it to high-performance CPU, GPU, and FPGA programs, which can be optimized to achieve state-of-the-art. Internally, DaCe uses the Stateful DataFlow multiGraph (SDFG) data-centric intermediate representation: A transformable, interactive representation of code based on data movement. Since the input code and the SDFG are separate, it is possible to optimize a program without changing its source, so that it stays readable. On the other hand, transformations are customizable and user-extensible, so they can be written once and reused in many applications. With data-centric parallel programming, we enable direct knowledge transfer of performance optimization, regardless of the application or the target processor.

DaCe generates high-performance programs for:

Multi-core CPUs (tested on Intel, IBM POWER9, and ARM with SVE)
NVIDIA GPUs and AMD GPUs (with HIP)
Xilinx and Intel FPGAs

DaCe can be written inline in Python and transformed in the command-line/Jupyter Notebooks or SDFGs can be interactively modified using the Data-centric Interactive Optimization Development Environment Visual Studio Code extension.

For more information, see our paper.

See an example SDFG in the standalone viewer (SDFV).

Quick Start

Install DaCe with pip: pip install dace

Having issues? See Troubleshooting

Using DaCe in Python is as simple as adding a @dace decorator:

import dace
import numpy as np

@dace
def myprogram(a):
    for i in range(a.shape[0]):
        a[i] += i
    return np.sum(a)

Calling myprogram with any NumPy array or __{cuda_}array_interface__-supporting tensor (e.g., PyTorch, Numba) will generate data-centric code, compile, and run it. From here on out, you can optimize (interactively or automatically), instrument, and distribute your code. The code creates a shared library (DLL/SO file) that can readily be used in any C ABI compatible language (C/C++, FORTRAN, etc.).

For more information on how to use DaCe, see the samples or tutorials below:

Dependencies

Runtime dependencies:

A C++14-capable compiler (e.g., gcc 5.3+)
Python 3.7 or newer (Python 3.6 is supported but not actively tested)
CMake 3.15 or newer

Running

Python scripts: Run DaCe programs (in implicit or explicit syntax) using Python directly.

SDFV (standalone SDFG viewer): To view SDFGs separately, run the sdfv installed script with the .sdfg file as an argument. Alternatively, you can use the link or open dace/viewer/webclient/sdfv.html directly and choose a file in the browser.

Visual Studio Code extension: Install from the VSCode marketplace or open an .sdfg file for interactive SDFG viewing and transformation.

The sdfgcc tool: Compile .sdfg files with sdfgcc program.sdfg. Interactive command-line optimization is possible with the --optimize flag.

Jupyter Notebooks: DaCe is Jupyter-compatible. If a result is an SDFG or a state, it will show up directly in the notebook. See the tutorials for examples.

Octave scripts (experimental): .m files can be run using the installed script dacelab, which will create the appropriate SDFG file.

Note for Windows/Visual C++ users: If compilation fails in the linkage phase, try setting the following environment variable to force Visual C++ to use Multi-Threaded linkage:

X:\path\to\dace> set _CL_=/MT

Publication

If you use DaCe, cite us:

@inproceedings{dace,
  author    = {Ben-Nun, Tal and de~Fine~Licht, Johannes and Ziogas, Alexandros Nikolaos and Schneider, Timo and Hoefler, Torsten},
  title     = {Stateful Dataflow Multigraphs: A Data-Centric Model for Performance Portability on Heterogeneous Architectures},
  year      = {2019},
  booktitle = {Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis},
  series = {SC '19}
}

Troubleshooting

If you are using DaCe from the git repository and getting missing dependencies or missing include files, make sure you cloned the repository recursively (with git clone --recursive) and that the submodules are up to date.
If you are running on Mac OS and getting compilation errors when calling DaCe programs, make sure you have OpenMP installed and configured with Apple Clang. Otherwise, you can use GCC to compile the code by following these steps:
- Run brew install gcc
- Set your ~/.dace.conf compiler configuration to use the installed GCC. For example, if you installed version 9 (brew install gcc@9), run which g++-9 and set the config entry called compiler.cpu.executable (empty string by default) to the resulting path
- Remove any .dacecache folders to clear the cache

Other issues? Look for similar issues or start a discussion on our GitHub Discussions!

Configuration

DaCe creates a file called .dace.conf in the user's home directory. It provides useful settings that can be modified either directly in the file (YAML) or overridden on a case-by-case basis using environment variables that begin with DACE_ and specify the setting (where categories are separated by underscores). The full configuration schema is located here.

The priority order for configuration files is as follows:

If a DACE_* environment variable is found, its value will always be used
If with dace.config.set_temporary(...) is used (see example here)
A .dace.conf located in the current working directory
The .dace.conf located in the user's home directory or the path pointed to by the DACE_CONFIG environment variable

If no configuration file can be created in any of the above paths, default settings will be used.

Useful environment variable configurations include:

DACE_CONFIG (default: ~/.dace.conf): Override DaCe configuration file choice.

General configuration:

DACE_debugprint (default: False): Print debugging information.
DACE_compiler_use_cache (default: False): Uses DaCe program cache instead of re-optimizing and compiling programs.
DACE_compiler_default_data_types (default: Python): Chooses default types for integer and floating-point values. If Python is chosen, int and float are both 64-bit wide. If C is chosen, int and float are 32-bit wide.

Profiling:

DACE_profiling (default: False): Enables profiling measurement of the DaCe program runtime in milliseconds. Produces a log file and prints out median runtime.
DACE_treps (default: 100): Number of repetitions to run a DaCe program when profiling is enabled.

GPU programming and debugging:

DACE_compiler_cuda_backend (default: cuda): Chooses the GPU backend to use (can be cuda for NVIDIA GPUs or hip for AMD GPUs).
DACE_compiler_cuda_syncdebug (default: False): If True, calls device-synchronization after every GPU kernel and checks for errors. Good for checking crashes or invalid memory accesses.

FPGA programming:

DACE_compiler_fpga_vendor: (default: xilinx): Can be xilinx for Xilinx FPGAs, or intel_fpga for Intel FPGAs.

SDFG interactive transformation:

DACE_optimizer_transform_on_call (default: False): Uses the transformation command line interface every time a @dace function is called.
DACE_optimizer_interface (default: dace.transformation.optimizer.SDFGOptimizer): Controls the SDFG optimization process if transform_on_call is enabled. By default, uses the transformation command line interface.
DACE_optimizer_automatic_simplification (default: True): If False, skips automatic simplification in the Python frontend (see transformations tutorial for more information).

Contributing

DaCe is an open-source project. We are happy to accept Pull Requests with your contributions! Please follow the contribution guidelines before submitting a pull request.

License

DaCe is published under the New BSD license, see LICENSE.

ajppp/dace