Note: This video may require joining the NVIDIA Developer Program or login

GTC Silicon Valley-2019 ID:S9257:New FFT Library with Flexible C++ API

Lukasz Ligowski(NVIDIA)
We will introduce a new FFT library that extends and complements cuFFT. This new library allows to go beyond performance limits imposed by the current cuFFT API. During our talk, we will walk through new API concepts, showcase performance, and discuss planned future developments.Specifically, the new library aims to provide performance and flexibility improvements by allowing the merging of FFT calculations with other code at convenient levels of abstraction. Examples are: device functions that can be inlined in user CUDA kernels; functions that are able to generate custom FFT kernels at a higher level; and functions that can optimize the sequence of FFT transforms at the top level.Software handling FFT calculations would be able to leverage compile-time information from the user, thus enabling more efficient code generation and better integration with the user code.While the initial release of this new library is planned to directly provide only a subset of functionality of the cuFFT