cuda - CUFFT alternatives (for FFT of half-float input) -

cuda - CUFFT alternatives (for FFT of half-float input) -

i need fft info stored in float16 format (on device). that, have have float buffer, kernel copy-convert info float16 array buffer, , phone call cufft routines.

would nice have cuda fft library can straight take float16 input.

are there such libraries or open-source cuda fft libraries, modified accordingly?

cuda gpu fft