Feature request - support for passing compiler options

[RajivKurian]

Thanks for the great project. It would be great if we could pass compiler options for our benchmark like gcc.godbolt.org. Right now I want to run my benchmark with -mavx2 and I can't find a way to do that.

[FredTingaud]

Hi,
I thought about it, but the problem with passing compiler options like that is that it would require some serious sanitizing, to avoid injections.
I am not very clear on how I could do it. I will have a look at godbolt's solution and see if I can use it.

[RajivKurian]

Yeah, looking at godbolt's solution would be my suggestion too. An alternative might be to have a few selectable options like -march=native etc, like you do for optimization level. Thanks!

…

On Mon, Aug 28, 2017 at 11:30 AM, Fred Tingaud ***@***.***> wrote: Hi, I thought about it, but the problem with passing compiler options like that is that it would require some serious sanitizing, to avoid injections. I am not very clear on how I could do it. I will have a look at godbolt's solution and see if I can use it. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#3 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AA0hpLgFb4pTSffAHADLyNPdHr9ezi-_ks5scwc0gaJpZM4PD3xZ> .

[didito]

Yes, being able to select architecture would help in benchmarking SIMD intrinsic code!

[kgariful]

Hi, -march option would be extremely useful. I make a benchmarking with SIMD instructions and autovectorization

[alexander-titov]

Hi,
It is a highly desirable feature!

[FredTingaud]

Hi,
Quick Bench runs on AWS, that doesn't guarantee any architecture for the kind of machines the project can afford.
Thus it would not possible to consistently run tests compiled with a given -march flags.

[didito]

Well, like somebody in the quick-bench-back-end issue already pointed out, this is not completely true.

You do have certain guarantees:
https://aws.amazon.com/ec2/instance-types/

Instance Type	CPU Architecture
T2	Intel Xeon (but does not say which one)
M4	2.3 GHz Intel Xeon E5-2686 v4 (Broadwell) or 2.4 GHz Intel Xeon E5-2676 v3 (Haswell)
M5	2.5 GHz Intel Xeon Platinum 8175 with AXV-512

So the minimum guaranteed is Haswell which means from the hardware side you have support for MMX, AES-NI, CLMUL, FMA3, SSE, SSE2, SSE3, SSSE3, SSE4, SSE4.1, SSE4.2, AVX up to AVX2. Broadwell did not change that, only with Skylake would we get AVX-512.

They do have a footnote saying

"AVX, AVX2 are only available on instances launched with HVM AMIs."

so not sure what is up with that. You could include a runtime check for that. But either way up until SSE4.2 you are fine. This would really make quickbench so much more useful to me!

[Yankes]

There is workaround, if you use __attribute__ ((target("fma"), optimize("-ffast-math"))) you can force GCC to change target processor per function and use SSE e.g.: http://quick-bench.com/SBnGN_2uHuwiFi5nIzekFVKK3ic
One important caveat is that test function need same attribute otherwise GCC will not inline it.

[didito]

Awesome, thank you @Yankes - that seems to work.

here are links to the documentation of the target attribute:
https://gcc.gnu.org/onlinedocs/gcc-8.2.0/gcc/x86-Function-Attributes.html#x86-Function-Attributes
https://clang.llvm.org/docs/AttributeReference.html#target-gnu-target

and here are the most relevant values:

• sse3
• ssse3
• sse4
• sse4.1
• sse4.2
• sse4a
• fma4
• avx

[FredTingaud]

Thanks for your comments!
Right now, I am super busy and can't really experiment with this, but I'll try to check that next time I work on an update.

[MattPD]

Support for -Ofast (and perhaps -Og) among the optim options would be valuable as well.

Edit: Added in 5bd1fe1 -- thank you!

[yvef]

As far as I see the topic is pretty old, but still relevant.
Is there any chance of changes?
Thanks.

[stephenberry]

You could always check a preset list of compiler flags.
It would be very helpful to be able to benchmark intrinsics (e.g. -mbmi2), otherwise we can't really test peak performance.

[Guiorgy]

@FredTingaud As a stop-gap, you could allow passing compiler options when run locally for those that need. Would that be difficult?