is it needed for intel core i7 ?

[Sandy4321]

is it needed for intel core i7 ?

[fo40225]

No, those whl is build for AMD.

Anaconda build numpy link against mkl, it provided maximum performance on Intel CPU already.

[RoyiAvital]

Could you explain what this library does?
Do you patch MKL to chose a good code path based on CPU features instead of the manufacturer?

[fo40225]

@RoyiAvital
Those whl build with OpenMP instead of Intel TBB and link to Agner's MKL patch to prevent MKL check the CPUID is GenuineIntel or not.

[RoyiAvital]

Can you build MKL DLL like this?

I mean rebuilding something like mkl_rt.dll.

[fo40225]

@RoyiAvital
No, the way to patch mkl on windows is static link to the mkl and replace the detect function.

Maybe you can replace the function in the mkl_intel_thread.dll, mkl_avx2.dll, mkl_core.dll or mkl_sequential.dll.

[RoyiAvital]

I see.

Problem is many languages use the mkl_rt.dll.
I wonder if there is a way to patch it or something like that.

[fo40225]

@RoyiAvital
If you are using C++, checkout Agner's asmlib.

[RoyiAvital]

I know it. You seem to use its patch. I wonder how could one patch pre built DLL and not in compilation.

[fo40225]

@RoyiAvital

You can write a wrapper dll that export all function in MKL but static link the mkl.lib and patch.

It should work.