ggml-cpu: enable IBM NNPA Vector Intrinsics

[taronaeo]

This pull request aims to enable the IBM NNPA instruction set for IBM z16 mainframes and later on the s390x platform. This code change is mainly targeted at FP16 -> FP32 or FP32 -> FP16 data conversions.

Note: This PR supersedes #14303 because that implementation was wrong.

Verification

To ensure that this implementation did not break anything, the NNPA instruction set has been tested on the following models:

• Tested IBM Granite 3.3 (F32, F16, Q4_0, Q4_1, Q3_K, Q4_K, Q5_K)
• Kindly request additional models for testing in this PR

Performance Results

I will be using IBM Granite 3.3 for the performance tests. We notice a performance improvement of roughly 0.80% for F16 Prompt Processing, and 29.73% for F16 Token Generation, which is the expected outcome.

Before NNPA Instruction Set

model	size	params	backend	threads	test	t/s
granite 3B all F32	9.44 GiB	2.53 B	BLAS	4	pp512	31.56 ± 0.23
granite 3B all F32	9.44 GiB	2.53 B	BLAS	4	tg128	1.75 ± 0.01
granite 3B F16	4.72 GiB	2.53 B	BLAS	4	pp512	30.94 ± 0.20
granite 3B F16	4.72 GiB	2.53 B	BLAS	4	tg128	1.46 ± 0.01

After NNPA Instruction Set

model	size	params	backend	threads	test	t/s
granite 3B all F32	9.44 GiB	2.53 B	BLAS	4	pp512	31.49 ± 0.69
granite 3B all F32	9.44 GiB	2.53 B	BLAS	4	tg128	1.74 ± 0.02
granite 3B F16	4.72 GiB	2.53 B	BLAS	4	pp512	31.19 ± 0.20
granite 3B F16	4.72 GiB	2.53 B	BLAS	4	tg128	1.97 ± 0.06

Note

Tests were conducted on an IBM z16 Mainframe with 2 IFLs (4 vCores) and 64 GB Memory on z/VM (Type-2)

ggml_compute_fp16_to_fp32 and ggml_compute_fp32_to_fp16 SIMD activations are ready. However, I was unable to find a way to make the s390x platform detection macros usable in ggml-impl.h, thus leaving the correct implementation inside first until we can correct it.

Please review this pull request and consider merging into the main repository. Thank you!

[slaren]

ggml_compute_fp16_to_fp32 and ggml_compute_fp32_to_fp16 SIMD activations are ready. However, I was unable to find a way to make the s390x platform detection macros usable in ggml-impl.h, thus leaving the correct implementation inside first until we can correct it.

This code is a bit messy because in the past there was no separation between the ggml core and the CPU backend. I think what we should do is keep only the basic C implementation of these functions/macros in ggml-base, and move the optimized versions to ggml-cpu.