hipFFT

hipFFT is an FFT marshalling library that supports rocFFT and cuFFT backends.

hipFFT exports an interface that doesn't require the client to change, regardless of the chosen backend. It sits between your application and the backend FFT library, where it marshals inputs to the backend and marshals results back to your application.

Documentation

Documentation for hipFFT is available at https://rocm.docs.amd.com/projects/hipFFT/en/latest/.

To build our documentation locally, run the following code:

cd docs

pip3 install -r sphinx/requirements.txt

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

Build and install

You can download pre-built packages from the ROCm package servers.

If you're using Ubuntu, you can run: sudo apt update && sudo apt install hipfft.

Transitioning from rocFFT

If you're transitioning from the hipFFT version included in rocFFT to the standalone hipFFT version, modify your build as shown in the following code:

• Previous method:

  `hipcc hipfft_1d_z2z.cpp -L/opt/rocm/lib -lrocfft`

• New method:

  `hipcc -I/opt/rocm/hipfft/include hipfft_1d_z2z.cpp -L/opt/rocm/lib -lhipfft -lrocfft`

Building from source

To build hipFFT from source, follow these steps:

1. Install the library build dependencies:

   • On AMD platforms, you must install rocFFT.
   • On NVIDIA platforms, you must install cuFFT.

2. Install the client build dependencies:

   • The clients (samples, tests, etc) included with the hipFFT source depend on FFTW, GoogleTest, and boost program options.

3. Build hipFFT:

   To show all build options:

     mkdir build && cd build
     cmake -LH ..

Here are some CMake build examples:

• AMD GPU
  • Case: Build a project using HIP language APIs + hipFFT with standard host compiler
    • Code: cmake -DCMAKE_CXX_COMPILER=g++ -DCMAKE_BUILD_TYPE=Release -L ..
  • Case: Build a project using HIP language APIs + hipFFT + device kernels with HIP-Clang
    • Code: cmake -DCMAKE_CXX_COMPILER=hipcc -DCMAKE_BUILD_TYPE=Release -DBUILD_CLIENTS=ON -L ..
• NVIDIA GPU
  • Case: Build a project using HIP language APIs + hipFFT with standard host compiler
    • Code: cmake -DCMAKE_CXX_COMPILER=g++ -DCMAKE_BUILD_TYPE=Release -DBUILD_WITH_LIB=CUDA -L ..
  • Case: Build a project using HIP language APIs + hipFFT + device kernels with HIP-NVCC
    • Code: HIP_PLATFORM=nvidia cmake -DCMAKE_CXX_COMPILER=hipcc -DCMAKE_BUILD_TYPE=Release -DBUILD_CLIENTS=ON -L ..

The `-DBUILD_CLIENTS=ON` option is only allowed with the HIPCC compiler.

Porting from CUDA

If you have existing CUDA code and want to transition to HIP, follow these steps:

1. HIPIFY your code and fix all unsupported CUDA features and user-defined macros
2. Build with HIP-NVCC to run on an NVIDIA device
3. Build with HIP-Clang to run on an AMD device

More information about porting to HIP is available in the HIP porting guide.