29.2 benchmark help #2 (Ryzen only)

[Prime95]

I've decided to implement different default FFT selections for Ryzen. To do this, I need just one user's benchmark of the all-complex FFTs. Again, I only need one Ryzen benchmark.

Linux executable:
https://www.dropbox.com/s/9n3cfkumuqykbbp/mprime?dl=0

Put this in prime.txt:
BenchComplex=2
AllBench=1 (actually Linux menus will now ask you about this option)

Do a throughput benchmark with settings 16 to 32768 sizes, "every FFT size", "all CPU cores", "one worker" and "maximum workers", 8 seconds should be sufficient.

When done delete mprime.

Thanks!

[db597]

I'd be happy to help with Ryzen benchmarks, but my machine is running Windows 10... any plans to release a Windows build?

[airsquirrels]

Running now...

[Prime95]

Here's the Windows link: https://www.dropbox.com/s/8tek7axg6v...ime95.exe?dl=0

It is untested, but ought to work. Be sure to add AllBench=1 as the benchmark dialog box has not been upgraded yet to include that option.

[db597]

I tried to run the windows executable, but got an error message (System Error):

"The code execution cannot proceed because libhwloc-5.dll was not found. Reinstalling the program may fix this problem."

[Prime95]

Originally Posted by db597

I tried to run the windows executable, but got an error message (System Error):

"The code execution cannot proceed because libhwloc-5.dll was not found. Reinstalling the program may fix this problem."

Thanks, but Airsquirrels is running now. One benchmark is sufficient.

[airsquirrels]

Originally Posted by Prime95

Thanks, but Airsquirrels is running now. One benchmark is sufficient.

I'm not sure if I miss-tagged one of the options to make this even more intense than intended,, but it is still running. Up around 5000K at the moment.

[tului]

Originally Posted by Prime95

Thanks, but Airsquirrels is running now. One benchmark is sufficient.

I've got it downloaded and ready if you need more than one hmu.

[Dubslow]

Originally Posted by Prime95

Thanks, but Airsquirrels is running now. One benchmark is sufficient.

It seems to me that in an attempt to empirically classify and measure a broad category of hardware, one should have a sample size greater than n=1.

For instance memory hardware, memory architecture, different chip levels (speed, core count, cache size etc), different motherboards. Unless you are testing only sizes that fit in cache? Which on the face of it I can't really justify in any way based on my limited and outdated knowledge.

[Prime95]

Originally Posted by Dubslow

It seems to me that in an attempt to empirically classify and measure a broad category of hardware, one should have a sample size greater than n=1.

This data will be to select the default FFT implementation for all-complex FFTs. If all goes well, in 29.2 you'll be able to run a benchmark to override the default selection with the best FFT implementation for your particular CPU / memory configuration.

[airsquirrels]

Here is the results.txt (bottom of this file). I left the other tests I had run on different versions in there for posterity.