Reference material discussion thread

[axn]

Originally Posted by kriesel

Added to https://www.mersenneforum.org/showpo...9&postcount=29.

Thanks. It looks like it is doing the right thing and should be, in theory, little bit (10-15%) faster than v6 stage 2.

I believe you said there was a general FFT regression from v6 to v7. Could this explain why v7 stage 2 is slower than v6, despite having a better stage 2 algorithm?

[kriesel]

Originally Posted by axn

I believe you said there was a general FFT regression from v6 to v7.

It depends on the fft length apparently, per https://www.mersenneforum.org/showpo...35&postcount=2

Note due to the time required for even a single PRP benchmark per fft length across all supported lengths per selected Gpuowl version and single OS version, the tabulated values are single measurements each

Could this explain why v7 stage 2 is slower than v6, despite having a better stage 2 algorithm?

Partially. It does not explain the full magnitude of discrepancy in V7/v6 P-1 performance ratio, between ~112% expectation and ~75% observation.