ggml-cpu: BMI2 is only available on amd64

[candrews]

MSVC doesn't define BMI2 so it must be defined by by CMake. It was being set whenever GGML_BMI2 is set, but that results in a failure (error LNK2019: unresolved external symbol __pdep_u64) when targeting x86 (aka 32 bit) Windows. The fix is to only set BMI2 when targeting amd64.

Make sure to read the contributing guidelines before submitting a PR

[angt]

Thanks!
I don’t know about MSVC, but all the features I see around BMI2 are also AMD64 only.
Can you share more about the target configuration and the build failure?

[candrews]

Here's a build log in GitHub Actions which I believe provides all that info, demonstrating the failure when targeting win32:
https://github.com/candrews/pywhispercpp/actions/runs/19691916887/job/56409457483#step:5:431

[angt]

Ok i guess it's because of this line:

llama.cpp/ggml/CMakeLists.txt

Line 151 in 4abef75

option(GGML_BMI2 "ggml: enable BMI2" ${INS_ENB})

As a quick hack, maybe you can set the env SOURCE_DATE_EPOCH and select the flags you want?

A normal build gives:

-- x86 detected
-- Adding CPU backend variant ggml-cpu: -msse4.2;-mf16c;-mfma;-mbmi2;-mavx;-mavx2 GGML_SSE42;GGML_F16C;GGML_FMA;GGML_BMI2;GGML_AVX;GGML_AVX2

But with SOURCE_DATE_EPOCH=1:

-- x86 detected
-- Adding CPU backend variant ggml-cpu:

Then add the flags you want (GGML_AVX , etc)

edit: not GGML_AVX2 :)

[candrews]

That's a workaround, but that won't fix the GGML_CPU_ALL_VARIANTS=1 GGML_BACKEND_DL=1 case, which I'm eager to use :-)

[angt]

after taking a look, i think its just because of the missing -DGGML_NATIVE=OFF flag:

https://github.com/angt/pywhispercpp/actions/runs/19744492579/job/56575850622

code: angt/pywhispercpp@d287338

[angt]

@candrews, can we close this PR?

[candrews]

@candrews, can we close this PR?

If there's an approach I should be taking that would work, I'm not understanding it, sorry.

I'm trying: CIBW_ENVIRONMENT: CMAKE_ARGS="-DGGML_NATIVE=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON" which I believe it what #17528 (comment) advises. However, that still fails:
On Linux, the error is:

  /project/whisper.cpp/ggml/src/ggml-cpu/arch/arm/cpu-feats.cpp:36:35: error: ‘HWCAP2_SVE2’ was not declared in this scope; did you mean ‘HWCAP_SVE’?

https://github.com/candrews/pywhispercpp/actions/runs/19846422930/job/56864733852#step:5:449

And on Windows it fails with:

error LNK2019: unresolved external symbol __pdep_u64 referenced in function _ggml_vec_dot_iq1_m_q8_K [D:\a\pywhispercpp\pywhispercpp\build\temp.win32-cpython-39\Release\_pywhispercpp\whisper.cpp\ggml\src\ggml-cpu-skylakex.vcxproj]

https://github.com/candrews/pywhispercpp/actions/runs/19846422930/job/56864733855#step:5:534

Which is the 32 bit Windows problem.

The combination of -DGGML_NATIVE=OFF and -DGGML_CPU_ALL_VARIANTS=ON doesn't make sense to me that seems contradictory. One flag means "make a generic binary, not optimized for the host system" and the other means "make many binaries optimized for separate, specific target systems"

I'm happy to try and test things - thank you very much for working with me on this effort :)

[angt]

Here is the diff between our branches:

candrews/pywhispercpp@cpu-all-variants...angt:pywhispercpp:cpu-all-variants

I just made a PR to fix the HWCAP2_SVE2 issue, and the macOS issue is already known. I already had this one on FreeBSD: #17474

[candrews]

Does #17474 need to make to https://github.com/ggml-org/whisper.cpp as well?

[angt]

Does #17474 need to make to https://github.com/ggml-org/whisper.cpp as well?

Yes, we need to wait the next ggml sync