ARM builds and SIMD-assembler prospects

[ET_]

Any hints about the addition of the prp option for small Mersenne exponents?

It would be optimal for our PIS 😀

[ewmayer]

@Luigi:

Work continues on PRP support ... but slowly. I am moving at end of August, so Mlucas work necessarily has been taking a back seat to all the work involved with that.

[ET_]

Originally Posted by ewmayer

@Luigi:

Work continues on PRP support ... but slowly. I am moving at end of August, so Mlucas work necessarily has been taking a back seat to all the work involved with that.

I didn't remember that, nor willing to look snappy, it wasn't my intention.

Hope everything goes smooth.

[ewmayer]

Originally Posted by ET_

I didn't remember that, nor willing to look snappy, it wasn't my intention.

You didn't remember that because I never previously mentioned it around here. :)

Thanks for the smoothness Glückwunsch!

[M344587487]

Originally Posted by ewmayer

Glad you got at least a nice chunk of the "missing FLOPS" back. I'm looking forward to buying an Odroid N1 once they go on sale, hopefully within a month - one of the beta-testers ran Mlucas benchmarks, and using all 6 cores (one 2-threaded job running on the 'big' 2-core a72 cpu, another 4-theaded one on the 'little' 4-core a53 cpu) he gets 2.2-2.3 the total throughput of an Odroid C2, which means ~3.5x the total throughput of a Pi3. N1 pricing is estimated at ~$110, i.e. about the same $/FLOP as the C2. We can ony hope this sparks a full-blown 'multi-socket war' amongst the various ARM-micro-PC manufacturers. :) Even for the N1 one still needs ~10 of them to match the LL-test throughput of a cutting-edge Intel quad, but things are getting close to the "interestingness" level as far as wide-scale adoption goes.

If ~10 RK3399 equals a quad core intel is true for all RK3399 boards then the "interestingness" level has improved due to prices (but it's still in x86's favour).

• NanoPi Neo4 is ~£37: http://linuxgizmos.com/nanopi-neo4-s...ize-and-price/
• RockPi 4 is meant to be slightly cheaper but I can't find it for the UK: http://linuxgizmos.com/rk3399-based-...or-even-lower/
• A bare as barebones can be intel quad core 8100 is ~£270 from ebuyer (mobo, CPU, 2x4GB 2400 RAM, explode-grade PSU), really it's £300-£320 once you go for better RAM and PSU
• An Odroid H2 (all-in-one intel quad core) is supposedly £150 ( https://www.odroid.co.uk/ODroid-H2 ), £60 for 2x4GB 2400 sodimm plus £15 for the PSU ( https://www.odroid.co.uk/index.php?r...product_id=863 ) = ~£225, eta god knows when

10x Neo4 + 10x SD cards = ~£400, PSUs probably ~£30, no need for a switch so total ~£430. Not too shabby, is there a stronger SoC than 2xA72+4xA53 that makes more bang for buck sense?

[nomead]

I doubt that 10x factor somewhat. On the Odroid forum someone ran a couple test runs on Mlucas and a test version of the cancelled Odroid N1. 2560K fft and the timing was 87.7 msec/iter running on the two "big" A72 cores. As a comparison my Ryzen 3 2200G is doing 5.07 ms/iter at 2560K when running mprime, 4 cores 1 worker (throughput is slightly higher for 4 workers) and aren't Intel chips supposed to be much faster than that?

Would someone please make more recent benchmarks available - the benchmark tables on both mersenne.org and mersenne.ca are a bit old now.

But anyway, I ordered a Rock960 board. It hasn't arrived yet, but it's also got that RK3399 chip on it. It's more expensive than those other boards, but it's just one board for testing and general fooling around, thermals and power consumption etc.

There are boards with e.g. Kirin 970 (4x A73, 4x A53) but they are really quite expensive now. Maybe after a while they'll get cheaper too. Or maybe something else will.

[ewmayer]

Re. the Odriod-N1 tests, I had a beta tester of one of those do an Mlucas build and timings on that - long story short, we got best total throughput running one 2-thread job on the dual A72 core and one 4-thread job on the quad A53, and said total throughput was ~2.2x what I get running on my quad-A53-core Odroid C2.

No idea whether the N2 will be appreciably better FLOPS-wise than the N1, and no word (AFAIK) on when the much-delayed N2 will finally be available, hopefully early in 2019.

Looking down the road to 2020, the best hope for an x86-competitive ARM implementation may be in Apple's PC roadmap. More frustrating waiting!

[M344587487]

Originally Posted by nomead

I doubt that 10x factor somewhat. On the Odroid forum someone ran a couple test runs on Mlucas and a test version of the cancelled Odroid N1. 2560K fft and the timing was 87.7 msec/iter running on the two "big" A72 cores. As a comparison my Ryzen 3 2200G is doing 5.07 ms/iter at 2560K when running mprime, 4 cores 1 worker (throughput is slightly higher for 4 workers) and aren't Intel chips supposed to be much faster than that?

Would someone please make more recent benchmarks available - the benchmark tables on both mersenne.org and mersenne.ca are a bit old now.
...

Taking your 87.7 ms/it for 2xA72 and my 160 msec/it for 2560K 4xA53 from the previous page we get ~17.65 it/sec, effectively ~56.65ms/it for an RK3399. Ewmayer's post on the Odroid forum suggests there may be a 10% slowdown running both clusters simultaneously. So ~11.2 RK3399 may be equivalent to a 2200G (~12.4 if 10% slowdown is present). From the mersenne.ca bench of an i3 8100 of 3.67ms/it for 2560K we can estimate that an 8100 roughly translates to ~15.4 RK3399 (~17.2 if 10% slowdown is present). That's a lot of compounded estimates so a big pinch of salt is required.

Originally Posted by ewmayer

...
Looking down the road to 2020, the best hope for an x86-competitive ARM implementation may be in Apple's PC roadmap. More frustrating waiting!

Oh dear.

[nomead]

Just for completeness' sake, here are the full iteration timings for the Raspberry Pi 3B+ I'm running. Taking care of thermals gave a significant boost over no heatsink, but I got a further 3% by disconnecting the display and stopping X. Just having it run in the background doing "nothing" still apparently consumes a minor portion of CPU time.

Gentoo 64-bit "sakaki", probably version 1.21 of the package since I originally downloaded and installed it in August for other purposes. uname -a says

Linux pi64 4.14.44-v8-4fca48b7612d-bis+ #2 SMP PREEMPT Fri Jun 1 15:55:22 BST 2018 aarch64 GNU/Linux

Stock 1.4 GHz, and four Cortex-A53 cores on the BCM2837B0. No overclocking attempts have been made (or will be made). Running Mlucas_c2simd 17.1 (downloaded precompiled binary)

17.1
      1024  msec/iter =   47.12  ROE[avg,max] = [0.255116697, 0.375000000]  radices = 256  8 16 16  0  0  0  0  0  0
      1152  msec/iter =   54.31  ROE[avg,max] = [0.223389521, 0.281250000]  radices = 144 16 16 16  0  0  0  0  0  0
      1280  msec/iter =   58.89  ROE[avg,max] = [0.264171231, 0.375000000]  radices = 160 16 16 16  0  0  0  0  0  0
      1408  msec/iter =   68.59  ROE[avg,max] = [0.228616585, 0.312500000]  radices = 176 16 16 16  0  0  0  0  0  0
      1536  msec/iter =   75.89  ROE[avg,max] = [0.252626651, 0.343750000]  radices = 192 16 16 16  0  0  0  0  0  0
      1664  msec/iter =   83.16  ROE[avg,max] = [0.272233409, 0.406250000]  radices = 208 16 16 16  0  0  0  0  0  0
      1792  msec/iter =   90.47  ROE[avg,max] = [0.222731285, 0.312500000]  radices = 224 16 16 16  0  0  0  0  0  0
      1920  msec/iter =   98.88  ROE[avg,max] = [0.255165462, 0.375000000]  radices = 240 16 16 16  0  0  0  0  0  0
      2048  msec/iter =  105.43  ROE[avg,max] = [0.238688298, 0.312500000]  radices = 256 16 16 16  0  0  0  0  0  0
      2304  msec/iter =  123.05  ROE[avg,max] = [0.249503539, 0.312500000]  radices = 288 16 16 16  0  0  0  0  0  0
      2560  msec/iter =  143.92  ROE[avg,max] = [0.233106476, 0.312500000]  radices = 160 32 16 16  0  0  0  0  0  0
      2816  msec/iter =  165.29  ROE[avg,max] = [0.260105912, 0.375000000]  radices = 176 32 16 16  0  0  0  0  0  0
      3072  msec/iter =  181.54  ROE[avg,max] = [0.261096695, 0.375000000]  radices = 192 32 16 16  0  0  0  0  0  0
      3328  msec/iter =  198.91  ROE[avg,max] = [0.282578930, 0.375000000]  radices = 208 32 16 16  0  0  0  0  0  0
      3584  msec/iter =  215.88  ROE[avg,max] = [0.251145062, 0.375000000]  radices = 224 32 16 16  0  0  0  0  0  0
      3840  msec/iter =  235.33  ROE[avg,max] = [0.246073929, 0.343750000]  radices = 240 32 16 16  0  0  0  0  0  0
      4096  msec/iter =  253.36  ROE[avg,max] = [0.226999763, 0.281250000]  radices = 256 32 16 16  0  0  0  0  0  0
      4608  msec/iter =  294.28  ROE[avg,max] = [0.249245933, 0.375000000]  radices = 288 32 16 16  0  0  0  0  0  0
      5120  msec/iter =  330.03  ROE[avg,max] = [0.236507015, 0.312500000]  radices = 160 32 32 16  0  0  0  0  0  0
      5632  msec/iter =  376.08  ROE[avg,max] = [0.259536082, 0.343750000]  radices = 176 32 32 16  0  0  0  0  0  0
      6144  msec/iter =  417.59  ROE[avg,max] = [0.245978727, 0.343750000]  radices = 192 32 32 16  0  0  0  0  0  0
      6656  msec/iter =  456.34  ROE[avg,max] = [0.266108247, 0.375000000]  radices = 208 32 32 16  0  0  0  0  0  0
      7168  msec/iter =  498.72  ROE[avg,max] = [0.225733680, 0.312500000]  radices = 224 32 32 16  0  0  0  0  0  0
      7680  msec/iter =  545.44  ROE[avg,max] = [0.236645483, 0.312500000]  radices = 240 32 32 16  0  0  0  0  0  0

[thorken]

https://www.mersenneforum.org/showthread.php?t=23994

[ewmayer]

Originally Posted by thorken

https://www.mersenneforum.org/showthread.php?t=23994

If you're not willing to post source, you're going to continue being greeted with skepticism. All the major current clients used by GIMPSers allow users to inspect the source and to build directly from it, should they so desire.

On the other hand, were to do a build of Mlucas and post comparative timings vs your app on your Android hardware, that could be useful.