View Single Post
Old 2021-09-25, 18:13   #4
kriesel's Avatar
Mar 2017
US midwest

17D316 Posts

Right. In V7.x you'd have to give bounds with a PRP worktodo line, then kill the run after P-1 stage 2 finishes. (Possibly you could use -iters if that worked. I vaguely recall trying that -iters method without success. The iters option value would be about 1.45-1.5 times the B1 bound.) And it would be slower than V6.11-380 or -364 etc. That's why I referred to V6.11-380 in post 2, and gave links to speed comparisons. Many of us are still running V6.11-380 or lower for some purposes. When I was running multiple RadeonVIIs as a P-1 factory, that was all with modified bounds prepended, on V6.11-380; sometimes ~40 wavefront P-1 results/day/GPU with GPU72 row bounds indicated at
If you reread the worktodo entry formats post carefully in the Gpuowl section, you'll see that PFactor= is not one of the accepted forms for V7.x. That's intentional, not an omission. V7.x does not support standalone PFactor, nor LL; it's combined P-1/PRP only. In some cases that's faster than V6.11-380 standalone P-1 followed by PRP, in most cases not. Also note that v7.2-69 was not the fastest 7.2 version in my experience; generally that was 7.2-53.

All that is part of why the Gpuowl intro post includes "In either case, note that the computation types, hardware supported, fft size limits, file formats, etc have varied greatly and rapidly over the course of the hundreds of versions. Choose version according to what you want to run and what each offers."

Preda never let the past get in the way of trying to make improvements. That seems to have worked well overall. And presents an occasional pitfall for the unwary.
kriesel is offline   Reply With Quote