Platform detection incorrect for AVX2 support

[techvintage]

The current code does not detect platform correctly for AVX2 support and hence uses slower functions on non-Intel platforms.This fix causes AVX2 related functions to be invoked correctly and giving the required 2.5X+ speed-up on AMD and possibly other platforms which support AVX2.

[tianjunwork]

Hi @techvintage Thanks for this PR.
Instead of hard code, could you make the change in can_use_intel_avx512()? You can change the name of this function too.

[techvintage]

Hi @tianjunwork ,this is not hardcoded but changing asm_type to highest level (=7) to begin with.If AVX2 functionality is not supported,it starts downgrading and uses corresponding slower functions instead.This is opposite approach to setting asm_type=0 (lowest value) and increasing it incrementally and fixes the bug.

[tianjunwork]

Hi @techvintage, I don't get it, if the logic after uint32_t asm_type = 7; works as expected on your platform, asm_type will be reassigned to a new value. How does your patch work?

[techvintage]

So can_use_intel_avx512() and can_use_Intel_4thGen_features function do not work correctly on non-Intel Platforms and proceeds with asm_type=7 and picks up the faster functions on AMD platforms correctly.We saw a 2.5X+ gain with this on Matisse with this.
Also, I see a FFmpeg patch failing on travis CI,My travis CI is synced to github but looks like cloning ffmpeg git repo timed out.Do we developers have a way of triggering the build again?

[tianjunwork]

I don't have a non-intel platform so don't know how these two functions behave. You mean can_use_intel_avx512() doesn't work but still return 1? Otherwise it runs to "asm_type = 1;".
I re-run the travis task. It is passed now.

[techvintage]

Yes,that's correct.So if we start with asm_type=7(highest),If can_use_intel_avx512() returns true for Intel platforms,it will be still at 7 else will start lowering asm_type based on support available.

[tianjunwork]

If can_use_intel_avx512() still returns 1 for non-intel platform, why do you need to set asm_type=7 at the beginning?
Intel platform works just fine. Your goal is to handle non-intel platform, right?
If so, instead of letting non-working can_use_intel_avx512() and can_use_intel_core4th_gen_features() mess up the logic. Can we make these two function more generic?

[Austin-Hu]

Hi @techvintage ,

Please consider to implemente the checking for AMD CPUs, like what CanUseIntelAVX512() and Check4thGenIntelCoreFeatures() do for Intel CPUs. Your change occationally can work on your platform which supports AVX2, but cannot for other (old) platforms which can only support SSEx or no intrinsics.
In theory, GetCpuAsmType() should check all the x86 CPUs with SSE and/or AVX intrinsics supported, including Intel, AMD (and VIA).

Thanks!