portFFT

Introduction

portFFT is a library implementing Fast Fourier Transforms using SYCL and C++. portFFT is in early stages of development and will support more options and optimizations in the future.

Pre-requisites

DPC++ oneAPI release 2024.0
- Nightly releases from intel/llvm should work but are not tested
- Other SYCL implementations are not tested
Level Zero drivers
- OpenCL drivers are not supported
CMake 3.20+
For tests and verifying benchmarks:
- Python
- Numpy

Getting Started

Building with CMake

Clone portFFT and run the following commands from the cloned repository.

Build using DPC++ 2024.0 as:

source /path/to/intel/oneapi/compiler/2024.0/env/vars.sh
cmake -Bbuild -DCMAKE_CXX_COMPILER=icpx -DPORTFFT_BUILD_TESTS=ON -DPORTFFT_BUILD_BENCHMARKS=ON
cmake --build build

Build using DPC++ nightlies as (SPIR64 target only):

cmake -Bbuild -DCMAKE_CXX_COMPILER=/path/to/dpcpp/bin/clang++ -DCMAKE_C_COMPILER=/path/to/dpcpp/bin/clang -DPORTFFT_BUILD_TESTS=ON -DPORTFFT_BUILD_BENCHMARKS=ON
cmake --build build

To compile AOT for a specific device, specify the target device with:

-DPORTFFT_DEVICE_TRIPLE=<T>[T1,..,Tn]

The list of available targets can be found on DPC++ compiler documentation page. Some AOT targets do not support double precision. To disable the building of tests and benchmarks using double precision, set -DPORTFFT_ENABLE_DOUBLE_BUILDS=OFF.

portFFT currently requires to set the subgroup size at compile time. Multiple sizes can be set and the first one that is supported by the device will be used. Depending on the device used you may need to set the subgroup size with -DPORTFFT_SUBGROUP_SIZES=<comma separated list of sizes>. By default only size 32 is used. If you run into the exception with the message None of the compiled subgroup sizes are supported by the device! then DPORTFFT_SUBGROUP_SIZES must be set to a different value(s) supported by the device.

Tests

Tests are build if the CMake setting PORTFFT_BUILD_TESTS is set to ON. Additionally, this enables clang-tidy checks if PORTFFT_CLANG_TIDY is at its default value of ON. Automatic fixing of some clang-tidy warnings can be enabled by setting PORTFFT_CLANG_TIDY_AUTOFIX to ON.

Run the tests from the build folder with:

ctest

portFFT benchmarks

Run pre-defined benchmarks from the build folder with:

./test/bench/bench_float

Run manual benchmarks from the build folder with for instance:

./test/bench/bench_manual_float d=cpx,n=5

Use the --help flag to print help message on the configuration syntax.

Supported configurations

portFFT is still in early development. The supported configurations are:

complex-to-complex transforms
interleaved complex and split complex (restricted to one dimension) storage
single and double precisions
forward and backward directions
in-place and out-of-place transforms
USM and buffer containers
batched transforms
1D transforms
multi-dimensional transforms with the following restrictions:
- default values for strides and distances
- size in each dimension must be supported by 1D transforms
Arbitrary forward and backward scales
Arbitrary forward and backward offsets
Arbitrary strides and distance where the problem size + auxilary data fits in the registers of a single subgroup.

Any 1D arbitrarily large input size that fits in global memory is supported, with a restriction that large input sizes should not have large prime factors. The largest prime factor depend on the device and the values set by PORTFFT_REGISTERS_PER_WI and PORTFFT_SUBGROUP_SIZES. For instance with PORTFFT_REGISTERS_PER_WI set to 128 (resp. 256) each work-item can hold a maximum of 27 (resp. 56) complex values, thus with PORTFFT_SUBGROUP_SIZES set to 32 the largest prime factor cannot exceed 27*32=864 (resp. 56*32=1792). portFFT may allocate up to 2 * PORTFFT_MAX_CONCURRENT_KERNELS * input_size scratch memory, depending on the configuration passed.

Any batch size is supported as long as the input and output data fits in global memory.

By default the library assumes subgroup size of 32 is used. If that is not supported by the device it is running on, the subgroup size can be set using PORTFFT_SUBGROUP_SIZES.

Configurations that attempt to read from the same memory address from two separate batches of a transform are not supported.

Known issues

portFFT relies on SYCL specialization constants which have some limitations currently:
- Some optimizations seem to be missing at JIT compile time. We are working to improve this.
- Specialization constants are currently emulated on Nvidia and AMD backends which prevents some optimizations.

Overall the work of optimizing portFFT is still in progress.

Troubleshooting

The library should compile without error on our supported platforms. If you run into trouble, or think you have found a bug, we have a support forum available through the developer website, or create an issue on GitHub.

Maintainers

This library is maintained by Codeplay Software Ltd. If you have any problems, please contact sycl@codeplay.com.

Contributions

This library is licensed under the Apache 2.0 license. Patches are very welcome! If you have an idea for a new feature or a fix, please get in contact.

codeplaysoftware/portFFT