Multiple (prp) test at once?

[R. Gerbicz]

When you are loading from a precomputed cos/sin table at fft or compute these in-fly then you could reuse these values. Basically hide/halve this computation cost [or just the cost of load] if you'd compute multiple fft test at once. Ofcourse use "close" N values, so close p values [for Mersenne numbers] to have the same FFT size for the problems.

Say you'd compute
cos(d)*a1[i]+sin(d)*b1[i]
cos(d)*a2[i]+sin(d)*b2[i]

or for more than 2 tests:
cos(d)*a3[i]+sin(d)*b3[i]
cos(d)*a4[i]+sin(d)*b4[i]
etc.

Got the idea?

[chalsall]

Originally Posted by R. Gerbicz

Got the idea?

Yes. Thank you. Clear.

[preda]

Originally Posted by R. Gerbicz

When you are loading from a precomputed cos/sin table at fft or compute these in-fly then you could reuse these values. Basically hide/halve this computation cost [or just the cost of load] if you'd compute multiple fft test at once. Ofcourse use "close" N values, so close p values [for Mersenne numbers] to have the same FFT size for the problems.

Say you'd compute
cos(d)*a1[i]+sin(d)*b1[i]
cos(d)*a2[i]+sin(d)*b2[i]

or for more than 2 tests:
cos(d)*a3[i]+sin(d)*b3[i]
cos(d)*a4[i]+sin(d)*b4[i]
etc.

Got the idea?

On the GPU, we are limited by the small number of "VGPRs" (registers) per workgroup that are available. Because we're operating at the upper limit of VGPRs, there's no much room to operate on two datasets in parallel in the same workgroup.