Question: is it possible to simulate SSE2 only arch on AVX2 cpu ? about highway HOT 9 CLOSED

If a machine lacks AVX2, HWY_DYNAMIC_DISPATCH will not call into avx2 code. But maybe there is some behavior change if the vector length differs?

If you want to force either SSE4 (if available) or else scalar, you can #define HWY_DISABLED_TARGETS (HWY_AVX2|HWY_AVX3). You could also call DisableTargets() at runtime.

from highway.

@jan-wassenberg , thanks. I think there is still a bug in NearestInt:
In debug build on my local machine, all unit tests pass, but in release mode, I get failures. And these are the unit tests that use NearestInt. Any ideas on what would cause optimized builds to fail?

If I disable AVX2, then all tests pass.

Do you have tests comparing NearestInt with AVX2 to C++ lrintf ?

from highway.

auto y = va_r*r + va_g*g;

gives a different result from

auto y = va_r*r;
y +=  + va_g*g;

The second calculation matches the C++ calculation.

from highway.

So, my conclusion is that vanilla C++ code compiled inside HWY method gives different results from vanilla C++ code
compiled outside of HWY method..

from highway.

Depending on optimization flags, the compiler may perform unsafe, non-IEEE compliant transformations.
Do you have -ffast-math enabled? I'd recommend against that.

that vanilla C++ code compiled inside HWY method gives different results from vanilla C++ code
compiled outside of HWY method..

This could be due to FMA - compilers may replace a*b+c with FMA (depending on fp-contract/similar flags). That might even be a good thing, though it will change results by 1E-5 or so. Presumably the results would match if you compiled everything with -ffp-contract=fast ?

from highway.

Thanks. I don't have -ffast-math enabled.
So, are you saying that the compiler settings for HWY method will differ from settings for non-HWY ?

from highway.

Yes, that is necessarily the case - on GCC and Clang, you cannot emit SIMD code without -mavx2 etc. (or the target-specific attributes used here). Given we also specify -mfma, I believe the compiler is within its rights to contract some muls.

This seems to be one platform difference we cannot abstract away - SSE4 and AVX2 are going to return different results because the former does not support FMA. Does that make sense?

from highway.

Thanks. I may just disable SSE4 in that case.

But, let's just take AVX2 for now. I am hoping that I can get the same outputs for
both AVX2 and scalar, so that the output is uniform across different CPUs.

Also, even in AVX2 case, I still need to run scalar on last part of a buffer sometimes depending
on the number of threads in the system, as I break up the buffer into chunks, one chunk per thread,
and there can be some leftover bytes.

from highway.

@jan-wassenberg thanks, I think I understand the issue now. I will have to think about the best strategy for my project.

from highway.