Skylake X linux test please

[Prime95]

Can someone with a Skylake X CPU please run a quick test. You'll need linux and gcc installed.

See https://www.mersenneforum.org/showpo...0&postcount=29

[GP2]

For Ernst's main.c from the linked post, I compiled with gcc -O3 -march=skylake-avx512 main.c using gcc --version = gcc (GCC) 7.3.1 20180303 (Red Hat 7.3.1-5)

Running a.out gives no output.

This is on an AWS machine, Intel(R) Xeon(R) Platinum 8124M CPU @ 3.00GHz

I don't know if "Skylake X" is different from some other kind of Skylake???

[ATH]

You need to run:
time ./a.out

I did on a c5d.large in the other thread:

On a EC2 c5d.large (1 core / 2vCPU Xeon 8124M 3.00 Ghz, gcc 7.2):

real 0m0.675s
user 0m0.668s
sys 0m0.000s

Let us know if you need testing on more cores?

[GP2]

On a one-core (2 vCPU) c5d.large on AWS I got:

real    0m0.765s
user    0m0.765s
sys     0m0.000s

On a two-core (4 vCPU) c5d.xlarge on AWS I got:

real    0m0.707s
user    0m0.707s
sys     0m0.000s

and on a one-core n1-highcpu-2 on Google Compute Engine with gcc 8.2.0 I got:

real    0m0.961s
user    0m0.957s
sys     0m0.004s

On the latter, /proc/cpuinfo just gives generic "Intel(R) Xeon(R) CPU @ 2.00GHz"

There is a discrepancy with ATH's timing on the identical c5d.large instance. I did pause mprime while I ran the test. Perhaps different compiler versions, or different kernels with Spectre fixes that slow down execution? I am using the latest version of Amazon Linux 2.

Perhaps the likeliest explanation is simply the shared tenancy with other AWS users on the same physical hardware. I have some instances running mprime about 10% faster than others on the same instance type, it's the luck of the draw.

[ewmayer]

Originally Posted by GP2

For Ernst's main.c from the linked post, I compiled with gcc -O3 -march=skylake-avx512 main.c using gcc --version = gcc (GCC) 7.3.1 20180303 (Red Hat 7.3.1-5)

Running a.out gives no output.

Note, no output means 'error free' in this case - I probably should've added a 'success!' printf in that case, but just FYI.

Presumably George is hoping to find some hw/sw configuration which fails the test due to the same issue he encountered.

[fivemack]

model name : Intel(R) Core(TM) i9-7940X CPU @ 3.10GHz
pineapple@pineapple:/tmp$ time ./a.out

real 0m0.755s
user 0m0.755s
sys 0m0.000s

model name : Intel(R) Xeon(R) Silver 4114 CPU @ 2.20GHz
oak@oak:~$ time ./mayer-test

real 0m0.879s
user 0m0.876s
sys 0m0.004s

[Prime95]

Originally Posted by fivemack

model name : Intel(R) Core(TM) i9-7940X CPU @ 3.10GHz
pineapple@pineapple:/tmp$ time ./a.out

real 0m0.755s
user 0m0.755s
sys 0m0.000s

That's the one closest to my Skylake-X. Now we'll have to wait several days to see if my Skylake-X works.

[ewmayer]

Originally Posted by Prime95

That's the one closest to my Skylake-X. Now we'll have to wait several days to see if my Skylake-X works.

Well, assuming it's not a hardware issue, where would you look next? You posted the following over in the original Skylake-X swizzle thread:

Originally Posted by Prime95

The problematic "vpmovzxbq zmm31, xmm0" code is 62h, 62h, 7dh, 48h, 32h, 0f8h

The Intel manual says vpmovzxbq maps to the following opcode:

EVEX.256.66.0F38.WIG 32 /r

You have more experience with this low-level opcode-parsing than I - can you discern from the numbers you posted whether your assembler is producing the low-4-bytes-of-input-xmm-to-low-4-doubles-of-output-ymm half-width version of the instruction?

[Prime95]

Originally Posted by ewmayer

The Intel manual says vpmovzxbq maps to the following opcode:

EVEX.256.66.0F38.WIG 32 /r

You have more experience with this low-level opcode-parsing than I

Actually, we have the same level of experience -- none. I have not tried to figure out what "EVEX.256.66.0F38.WIG 32" means. I got the "62h, 62h, 7dh, 48h, 32h, 0f8h" bytes frome the defuse.ca web site, which uses the GNU tool chain.

[ewmayer]

Originally Posted by Prime95

Actually, we have the same level of experience -- none. I have not tried to figure out what "EVEX.256.66.0F38.WIG 32" means. I got the "62h, 62h, 7dh, 48h, 32h, 0f8h" bytes frome the defuse.ca web site, which uses the GNU tool chain.

What input format does that tool need? I have .s and .o files from compilation of the little test code I posted, perhaps seeing what the assembly for that code's vpmovzxbq instruction can be compared versus yours above. Here are gzipped .s and .o from my code, in case that might be of help:

[ldesnogu]

I tried with two different disassemblers, the one from GNU and the one that comes with LLVM. I don't know if it's of any help...

$ llvm-objdump-3.8 -d main.o  | grep vpmov
     337:    62 62 7d 48 32 f0     vpmovzxbq    %xmm0, %zmm30
     33d:    62 62 7d 48 32 f9     vpmovzxbq    %xmm1, %zmm31
$ objdump -d main.o  | grep vpmov
 337:    62 62 7d 48 32 f0        vpmovzxbq %xmm0,%zmm30
 33d:    62 62 7d 48 32 f9        vpmovzxbq %xmm1,%zmm31