Fix usage of F16C intrinsics in AVX code

[slaren]

Sandy bridge supports AVX but not F16C. Should fix #562

@kaufmannr can you verify if this fixes your problem?

[kaufmannr]

I can confirm that

#if defined(F16C)
// the _mm256_cvt intrinsics require F16C
#define GGML_F32Cx8_LOAD(x) _mm256_cvtph_ps(_mm_loadu_si128((__m128i *)(x)))
#define GGML_F32Cx8_STORE(x, y) _mm_storeu_si128((__m128i *)(x), _mm256_cvtps_ph(y, 0))
#else
static inline __m256 __sse_f16x8_load(ggml_fp16_t *x) {
float tmp[8];

for (int i = 0; i < 8; i++)
    tmp[i] = GGML_FP16_TO_FP32(x[i]);

return _mm256_loadu_ps(tmp);

}

static inline void __sse_f16x8_store(ggml_fp16_t *x, __m256 y) {
float arr[8];

_mm256_storeu_ps(arr, y);

for (int i = 0; i < 8; i++)
    x[i] = GGML_FP16_TO_FP32(arr[i]);

}
#define GGML_F32Cx8_LOAD(x) __sse_f16x8_load(x)
#define GGML_F32Cx8_STORE(x, y) __sse_f16x8_store(x, y)
#endif

compiles on the i5-2400 but was not able to test the program itself due to other issues.

[anzz1]

This should be resolved before merging. I couldn't edit the thing above so the post above is what I am referring to.

[anzz1]

There is one more problem regarding to this though, is that Makefile is configured to check for the processor features but CMake always declares it no matter what:
https://github.com/ggerganov/llama.cpp/blob/7e5395575a3360598f2565c73c8a2ec0c0abbdb8/CMakeLists.txt#L200

I'm not necessarily saying that CMake should be changed to check for the features but it should print information of the enabled features or a disclaimer of some sort to reduce the amount of trouble people are having.

Probably adding a option to CMakeLists for F16C in non-Windows-MSVC builds like there already are for FMA/AVX/etc. would be the right choice.

Any way, that's a problem for another day. Good that this one is fixed now. We can simply instruct people to use make as a primary option for now.

edit: #576 should fix this.