Long vectors work well for me

[fivemack]

After spending most of yesterday shaving yaks (mainly building gcc-8.2.0, since gcc-5.4 doesn't generate vector instructions for the VBITS=256 code in msieve-lacuda), I have 256-bit-wide vectors working on my Skylake Xeon box, and they're substantially faster than 64-bit-wide vectors: about 40 hours rather than about 60 hours for a 13M matrix.

[pinhodecarlos]

Windows binaries in here:

http://www.mersenneforum.org/showpos...2&postcount=61

Think I’m using the 256 bits version as well.

[wombatman]

Would using a 128 or 256 bit compilation of msieve help on an ivybridge?

[ATH]

Originally Posted by pinhodecarlos

Windows binaries in here:

http://www.mersenneforum.org/showpos...2&postcount=61

Think I’m using the 256 bits version as well.

Are you sure those work properly? They were not compiled with GCC 8.2.0 but with GCC 7.3.0. Will these 256 bit vectors only work on Skylake processors?

[pinhodecarlos]

My last 10 msieve post processing were done with it, no issues. Laptop has an ivy bridge processor

[fivemack]

The makefile does -march=native so a build on a Skylake Xeon might well not work on IVB or even on a Skylake non-Xeon. I’ll try doing a build on my IVB machine tomorrow and see if I can produce some timings.

[pinhodecarlos]

Apologies, I was typing through my iphone. I've just confirmed checking my folder that the msieve version I've been using lately is msieve-svn1018-vbits128-sandybridge. At least that's the only one I have unzipped in a separate folder although I have on the root the three versions.

[ATH]

Originally Posted by pinhodecarlos

Apologies, I was typing through my iphone. I've just confirmed checking my folder that the msieve version I've been using lately is msieve-svn1018-vbits128-sandybridge. At least that's the only one I have unzipped in a separate folder although I have on the root the three versions.

What I meant was does the 128 bit vectors give any improvement over 64 bit when it was compiled on GCC 7.3.0 ?

[pinhodecarlos]

Originally Posted by ATH

What I meant was does the 128 bit vectors give any improvement over 64 bit when it was compiled on GCC 7.3.0 ?

This is what I recall:

http://www.mersenneforum.org/showpos...2&postcount=14

[fivemack]

I fixed up the code so that VBITS=512 compiled and successfully started the linear algebra, but (on a machine with AVX512) it is less than half the speed of VBITS=256; my suspicion is that all the time is now spent in the B*N * N*B matrix multiply, which probably can be improved but would be a bit of work.

Ah, doing an objdump on the executable indicates that it's using pairs of ymm registers rather than single zmm registers; will need to fiddle around more with compilation options. Simply using -march=skylake-avx512 doesn't help: still nothing in 'objdump -d msieve | grep zmm'

[fivemack]

Here are some timings on an Ivy Bridge machine (six-core machine, running -t6, 13.52M matrix from C234_138_83 because that's what I had available)

SIZE	PREFETCH	Mdim/day
64	N		3.35
128	N		4.07
256	N		3.86
64	Y		3.41
128	Y		4.15
256	Y		4.00

So on this machine, running msieve-lacuda SVN1022 compiled with gcc-8.2.0 -march=native, MANUAL_PREFETCH is a noticeable plus, 128-bit vectors are better than 64-bit, but 256-bit vectors are less good than 128-bit