Library implementation suggestions about rocfft HOT 3 CLOSED

create internal counterparts for rocfft_plan_t, rocfft_plan_create and rocfft_plan_destroy that has more flexibility with the use of internal flags, vector of batches and so forth.
It might be good to eliminate internal calls to those public APIs.

from rocfft.

In his paper "Supercomputer-style FFT library for Apple G4" (link below), Richard Crandall presents an API proposal for an FFT library, comprising:

1. Complex FFT;
2. Real-signal FFT;
3. Inverse FFTs of the above;
4. Complex cyclic convolution;
5. Real-signal cyclic convolution;
6. 2-dimensional complex FFT;
7. 2-dimensional real-signal FFT;
8. Inverse FFTs for these 2-dimensional cases;
9. 2-dimensional complex convolution;
10. 2-dimensional real-signal convolution.

Offering "convolution" as an API primitive offers important optimization opportunities, for example by keeping the intermediary fft-transformed data in a transposed representation (saving transpose operations), and by merging multiple steps in a single kernel when possible (saving global memory round-trips).

https://pdfs.semanticscholar.org/25e3/6f4bdb1298416813ca3beaaf5f4b49105501.pdf?_ga=2.110805581.976118015.1512954370-353035247.1512954370

from rocfft.

Closing due to no new activity.

from rocfft.