search

Smithsonian / sdbe

Code for the SWARM VLBI digital back-end
5 stars 2 forks source link

Preprocessor #5

Open krosenfeld opened 9 years ago

krosenfeld commented 9 years ago

Transform SWARM rate frequency domain data to R2DBE rate time domain data.

krosenfeld commented 9 years ago

Can we re-order the B-frame data before dequantizing? That would reduce the number of bits being shuffled around.

anyoung commented 9 years ago

I believe we can; the reordering boundaries are conveniently located on (four?)-byte boundaries.

On Wed, Apr 29, 2015 at 3:59 PM, Katherine Rosenfeld < notifications@github.com> wrote:

Can we re-order the B-frame data before dequantizing? That would reduce the number of bits being shuffled around.

— Reply to this email directly or view it on GitHub https://github.com/sma-wideband/sdbe/issues/5#issuecomment-97564454.

krosenfeld commented 9 years ago

@ruriktherus pointed out that that cufft callbacks would be useful here: http://devblogs.nvidia.com/parallelforall/cuda-pro-tip-use-cufft-callbacks-custom-data-processing/

This does not appear to be (easily) supported by scikits.cuda or pycuda, but the examples from the CUDA toolkit and parallelforall repo do run on hamster.

lindyblackburn commented 9 years ago

The callbacks work element-by-element? We may be working on 32bits (16 samples) at a time and unpacking into 16 floats.

krosenfeld commented 9 years ago

I believe that callbacks are called on a per-element basis.