SIGILL on many Intel CPUs

intel / DML

Intel® Data Mover Library (Intel® DML)

https://intel.github.io/DML/

MIT License

81 stars 17 forks source link

SIGILL on many Intel CPUs #3

Closed kilobyte closed 2 years ago

kilobyte commented 2 years ago

Hi! I'm afraid that the detection of supported CPU extensions is flaky. On a bunch of machines I've tested, most examples die at an AVX instruction not supported by the CPU. This includes even Pentium 4 which you explicitly list as supported!

Machines I've tested:

uarch	model
Gemini Lake	J4115	✗
Skylake	i7-6700K	✓
Pinnacle Ridge	TR 2990WX	✓
Pentium 4	f15 m4 s3	✗
Braswell	N3160	✗
Pentium 4	f15 m4 s1	DNC (32-bit)
Phenom 2	Thuban	✗

You check cpuid, but take only cache info. The presence of AVX/AVX2/AVX512/etc is readily available there.

Arvinos commented 2 years ago

Hello @kilobyte, Currently library doesn't support dynamic dispatcher. The problem will fixed as dispatcher will be added.

EgorKupaev commented 2 years ago

Hello, @kilobyte , currently the issue lies in the options used for compilation.

target_compile_options(dml_core_px
    PRIVATE $<$<C_COMPILER_ID:GNU>:-mavx2>
    PRIVATE $<$<C_COMPILER_ID:MSVC>:/arch:AVX2>)

And as prerequisites state:

Minimal CPU vector instruction set supported for not optimized code is Intel® Streaming SIMD Extensions 2 (Pentium® 4 processor or higher)

I think AVX2 instructions enabled wrongly.

As a workaround now you can remove these options from: sources/cores/CMakeLists.txt

EgorKupaev commented 2 years ago

As part of v0.1.4-beta was removed -avx2 flag for non-optimized version of the library, so basically when the library is built without -DDML_AVX512 it uses default compiler options.

Additionally, the library's system requirements were clarified in the documentation.