29.2 benchmark help

[Prime95]

Background: Prime95 selects pre-determined FFT implementations based on benchmarks from various machines I've owned over the years. The problem is that the pre-determined selection may not be best for your machine -- it may have faster memory, more or fewer cores, etc.

Here's the plan. I've added code that writes benchmark timings to gwnum.txt as well as code to read that data in. Next up is writing code that selects the best FFT implementation based on the data. I hope this results in a tiny performance boost on machines that the default FFT implementation is non-optimal.

If I create an executable with every possible FFT implementation, I'll end up with a very large executable. So, I need help selecting which FFT implementations to include and which ones can safely be excluded.

To start the process, I'd like to gather some data from 64-bit linux machines with FMA3 support using this executable: https://www.dropbox.com/s/9n3cfkumuqykbbp/mprime?dl=0

Download the executable, add the line "AllBench=1" to prime.txt. 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.

This will be time consuming
Your machine must support FMA3
Unusual machines are what I'm looking for. Especially Ryzen, OC'd machines, fast mem, 2 cores, 6 cores, fancy Xeons

When done email results.txt to me. Delete the executable - it is not for production work.

Thanks!

[Jayder]

If I ran this on a Windows machine in an Ubuntu VM, would that affect anything, or would that be okay?

[Mark Rose]

Originally Posted by Jayder

If I ran this on a Windows machine in an Ubuntu VM, would that affect anything, or would that be okay?

It would certainly affect things.

You could boot the Ubuntu Desktop install USB to Desktop and run it from there?

[Mark Rose]

Would you like us to benchmark hyperthreads?

[mackerel]

I'd love to help but the linux requirement adds enough friction that I may be selective and test limited systems. If it ran on Windows I could easily throw it on everything.

I could do any of following if others aren't doing it, and are of specific interest.
Ryzen 1700 stock and OC
OC'd machines - Broadwell (with 128MB L4 cache) and Skylake quad cores
Fast ram machines: Skylake quad cores
Fancy Xeons: 14 core Haswell
2 cores: i3 Haswell, Skylake

[Mark Rose]

I'm running it on:

i3-2120 @ 3.3
i3-4710 @ 3.7 with single rank DDR3-1600
i3-4710 @ 3.7 with dual rank DDR3-1600
i7-4770k @ 3.9 (all core, so overclocked) with dual rank DDR3-1600
i5-6600 @ 3.3 (underclocked) with dual rank DDR4-2133

[mackerel]

Mark, does i3-2120 support required FMA3? Also the other i3 is 4170 not 4710?

[Prime95]

Originally Posted by Mark Rose

Would you like us to benchmark hyperthreads?

no

[Prime95]

I'll create a Windows build tonight

[VictordeHolland]

Originally Posted by Mark Rose

I'm running it on:

i3-2120 @ 3.3
i3-4710 @ 3.7 with single rank DDR3-1600
i3-4710 @ 3.7 with dual rank DDR3-1600

i3-2120 is SandyBridge based, so has AVX, but doesn't have AVX2/FMA
i3-4170 is Haswell, which does have AVX, AVX2, FMA (it mostly Pentium and Celeron parts that have AVX disabled for market segmentation)

[Mark Rose]

Originally Posted by mackerel

Mark, does i3-2120 support required FMA3? Also the other i3 is 4170 not 4710?

Hmm.. I did this in the middle of the night, and you're right, i3-2120 doesn't support FMA3. I'll put it back to normal work.

And the two other i3's are 4170.