Fix for issue #13170

[shalinib-ibm]

Build fails with compilation error on power pc.
This patch fixes the same.

Tested with unit tests run via
cmake --build <build_dir> && cd <build_dir> && make test

Make sure to read the contributing guidelines before submitting a PR

[shalinib-ibm]

@slaren After the addition of faster kernels for depth wise 2D convolution in ggml-cpu with this commit, c6e8cc2, we see that build fails with compilation error on power pc . This patch fixes the same. can you please review this PR ?

[CISC]

@shalinib-ibm Any chance of adding PPC to the CI build to more easily catch this in the future?

[shalinib-ibm]

@CISC agree that it would be nice to have some sort of CI for power arch. Will have to work on it.

[slaren]

Can you explain why this change is necessary? There are other cases where GGML_F32_VEC is used as a single scalar instead of an array, so it is not clear to me why this is necessary here.

[shalinib-ibm]

@slaren, in this piece of code, sum is expected to be a vector.
llama.cpp/ggml/src/ggml-cpu/ops.cpp:6120
6120 | GGML_F32_VEC sum = GGML_F32_VEC_ZERO;
But power9 defines
GGML_F32_VEC as a vector float and GGM_F32_VEC_ZERO as a scalar
defined in lines 371, 372, 343,344 gml/src/ggml-cpu/simd-mappings.h#L344) .

So we are making. a vector of vector with size 1. This way it does not break x86 or arm.
Similar style code can be seen here:

llama.cpp/ggml/src/ggml-cpu/vec.cpp

Line 28 in 5a63980

GGML_F32_VEC sum[GGML_F32_ARR] = { GGML_F32_VEC_ZERO };

Hence this change is necessary to fix the bug on power architectures.

[slaren]

GGML_F32_VEC sum[GGML_F32_ARR] = { GGML_F32_VEC_ZERO };

With AVX2 this expands to __mm256 sum[4] = { _mm256_setzero_ps() };, which does not look right to me, it is only initializing the first element with _mm256_setzero_ps().

In any case, it seems to me that the solution would to define GGML_F32_VEC_ZERO to {0.0f} for this architecture, rather than just 0, since it is meant to initialize entire vectors.

[shalinib-ibm]

Hi slaren,
{ __m256_setzero_ps() } : This is the initializer for the sum[4] array. In C/C++, when an array is partially initialised, the remaining elements are automatically set to zero. With AVX2, sum[0] is explicitly initialised to a 256 bit register with all 8 floats set to 0.0f. The remaining elements sum[1], sum[2], sum[3] are implicitly initialized to zero.

As per your suggestion, I have defined GGML_F32_VEC_ZERO to {0.0f} for power. Can you please review ?