avx on Core i5-2400

[kaufmannr]

Good evening,

I needed to deactivate avx compile flag for my Core i5-2400 to get it compile on Linux. Somebody else did implement the missing fp16 functions as code, but I was unable to find the right place for it in the macro hell.
I think this would be even better than just deactivating avx but I cannot find this code/comment anymore.

Would be cool, if your code would run on my CPU.

Thanks + best regards, Rainer

[anzz1]

The spec sheet for i5-2400 says that AVX should be supported on that processor?

It's probably AVX2 what you want to disable, but keep AVX.
Try:

mkdir build
cd build
cmake -DLLAMA_AVX2=OFF ..
cmake --build . --config Release

Okay thats for cmake, it seems you are using make.

However , your processor supports AVX but does not support F16C.
You are doing it the wrong way around, checking for F16C first and then for AVX?
Is there a problem of checking them separately like it currently does?
You should be able to enable AVX and only disable F16C , to enable AVX features.

It doesn't work with AVX enabled and F16C disabled?

[slaren]

If there are any cases where F16C is used without checking for the flag, that should be fixed instead.

[slaren]

I think ggml.c:1131 may be the culprit:

#define GGML_F32Cx8_STORE(x, y) _mm_storeu_si128((__m128i *)(x), _mm256_cvtps_ph(y, 0))

If I am not mistaken _mm256_cvtps_ph is only available with F16C, but this code only checks for AVX.

[anzz1]

I think ggml.c:1131 may be the culprit:

#define GGML_F32Cx8_STORE(x, y) _mm_storeu_si128((__m128i *)(x), _mm256_cvtps_ph(y, 0))

If I am not mistaken _mm256_cvtps_ph is only available with F16C, but this code only checks for AVX.

You are correct.
And load too

#define GGML_F32Cx8_LOAD(x)     _mm256_cvtph_ps(_mm_loadu_si128((__m128i *)(x)))

_mm256_cvtph_ps() = VCVTPH2PS
_mm256_cvtps_ph() = VCVTPS2PH

both are F16C/CVT16 instructions.
and should be guarded with __F16C__ check

using the SSE versions instead would fix that

#define GGML_F32Cx4_LOAD(x)     __sse_f16x4_load(x)
#define GGML_F32Cx4_STORE(x, y) __sse_f16x4_store(x, y)

[slaren]

Superseded by #563