Q2k interleaving implementation - x86/x64 SIMD

[Srihari-mcw]

• The PR contains block interleaving approach for Q2_K quantization for x64/x86 AVX2/AVX512 SIMD Architecture
• AVX512 and AVX2 Versions are implemented for the GEMM function, whereas GEMV is implemented with AVX2 intrinsics
• The existing quantize_q8_K_4x8 function quantizes the float values to block_q8_Kx4 format
• repack_q2_K_to_q2_K_8_bl function rearranges the weight in Q2_K format to Q2_Kx8 format(block_q2_Kx8)

Block Interleaving Formats

Block_Q2_Kx8 :

• Used to contain data of 8 Q2_K blocks in interleaved fashion
• uint8 scales[128] - Scales and Mins from source Q2_K blocks are taken. Every 16 byte here is packed such that it contains scales and mins for corresponding sub blocks from Q2_K structure - There are 16 sub blocks in original Q2_K structure
• The d and dmin values from source Q2_K blocks are stored together in an array
• Quant values from the source Q2_K blocks are sequentially extracted and interleaved into groups of eight bytes

Performance Impact :

Gains of ~5.5 % seen with the AVX2 version and gains of ~25.5% seen with the AVX512 Version over the base commit with GCC Linux

GCC Linux :

Q2_K Model :

model	size	params	backend	threads	test	t/s	speedup	Commit id
phi3 3B Q2_K - Medium	1.32 GiB	3.82 B	CPU	6	pp 512	84.64 ± 0.20		38de3fb - Base Commit
phi3 3B Q2_K - Medium	1.32 GiB	3.82 B	CPU	6	pp 512	89.26 ± 0.21	5.45%	ef03580 - AVX2 Commit
phi3 3B Q2_K - Medium	1.32 GiB	3.82 B	CPU	6	pp 512	106.27 ± 0.32	25.54%	ef03580- AVX512 Commit
phi3 3B Q2_K - Medium	1.32 GiB	3.82 B	CPU	6	tg 128	37.81 ± 0.02		38de3fb - Base Commit
phi3 3B Q2_K - Medium	1.32 GiB	3.82 B	CPU	6	tg 128	37.80 ± 0.02	-0.03%	ef03580 - AVX2 Commit
phi3 3B Q2_K - Medium	1.32 GiB	3.82 B	CPU	6	tg 128	37.64 ± 0.01	-0.46%	ef03580 - AVX512 Commit

GCC Version = 12.3

Clang Linux:

More gains of ~26.3% seen with the AVX2 version and gains of ~53.9% seen with the AVX512 Version over the base commit with Clang Linux

Q2_K Model :

model	size	params	backend	threads	test	t/s	speedup	Commit id
phi3 3B Q2_K - Medium	1.32 GiB	3.82 B	CPU	6	pp 512	92.33 ± 0.20		38de3fb - Base Commit
phi3 3B Q2_K - Medium	1.32 GiB	3.82 B	CPU	6	pp 512	116.68 ± 0.40	26.37%	ef03580 - AVX2 Commit
phi3 3B Q2_K - Medium	1.32 GiB	3.82 B	CPU	6	pp 512	142.13 ± 0.63	53.93%	ef03580- AVX512 Commit
phi3 3B Q2_K - Medium	1.32 GiB	3.82 B	CPU	6	tg 128	38.26 ± 0.00		38de3fb - Base Commit
phi3 3B Q2_K - Medium	1.32 GiB	3.82 B	CPU	6	tg 128	38.11 ± 0.01	-0.38%	ef03580 - AVX2 Commit
phi3 3B Q2_K - Medium	1.32 GiB	3.82 B	CPU	6	tg 128	37.98 ± 0.01	-0.71%	ef03580 - AVX512 Commit

Clang Version = 20.1.0

The model tested was - https://huggingface.co/bartowski/Phi-3-mini-4k-instruct-GGUF

The PR was tested in AMD Ryzen 5 9600X which supports the following flags by default :

CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | OPENMP = 1 | AARCH64_REPACK = 1 |

Further the perplexity was tested and found to be similar with the Q2_K Model

The perplexity results are tabulated as follows :

model	perplexity (Final estimate PPL)	Commit id
phi3 3B Q2_K - Medium	9.5511 +/- 0.064212	38de3fb - Base Commit
phi3 3B Q2_K - Medium	9.5488 +/- 0.06419	ef03580 - Updated Commit

[slaren]

I tested this on a 13900k with gcc 13 and clang 19, but the improvement is not very significant. Repacking has a significant cost, since it increases load time and prevents usage of mmap, and as it is, I find this very hard to justify for AVX2. It may make sense for AVX512, but I cannot test that.

GCC-13:

Model	Threads	Test	t/s master	t/s q2k_interleaving_implementation	Speedup
llama 7B Q2_K_M	8	pp64	47.56	48.55	1.02
llama 7B Q2_K_M	8	tg32	19.92	19.38	0.97
llama 7B Q2_K_M	16	pp64	63.04	60.08	0.95
llama 7B Q2_K_M	16	tg32	21.22	20.44	0.96
llama 7B Q2_K_M	24	pp64	68.39	68.07	1.00
llama 7B Q2_K_M	24	tg32	19.72	19.76	1.00
llama 7B Q2_K_M	32	pp64	71.18	71.62	1.01
llama 7B Q2_K_M	32	tg32	17.87	17.51	0.98

Clang-19:

Model	Threads	Test	t/s master	t/s q2k_interleaving_implementation	Speedup
llama 7B Q2_K_M	8	pp64	48.28	52.27	1.08
llama 7B Q2_K_M	8	tg32	20.78	19.08	0.92
llama 7B Q2_K_M	16	pp64	65.23	61.42	0.94
llama 7B Q2_K_M	16	tg32	20.94	19.79	0.95
llama 7B Q2_K_M	24	pp64	69.69	71.17	1.02
llama 7B Q2_K_M	24	tg32	19.90	19.59	0.98
llama 7B Q2_K_M	32	pp64	71.04	75.26	1.06
llama 7B Q2_K_M	32	tg32	16.91	17.30	1.02

[Srihari-mcw]

Hi @slaren ,
Thanks for the reply. Based on your feedback and further internal testing, we have currently updated the patch to enable the repacking for machines that have AVX512 support alone, so that the patch can be considered for optimization in AVX512 based machines. We will continue investigating the AVX2 performance further

Thanks

[Srihari-mcw]

Hi @slaren , @ggerganov ,
With regards to the AVX512 changes, are there are any other steps to be done to get this PR merged? Kindly share your thoughts. Thanks

[ggerganov]

Deduplicate the generic GEMV and GEMM implementations following #14897.

After that, feel free to merge.

[Srihari-mcw]

Hi @slaren , @ggerganov ,

The code has been updated with de-duplication of generic code. Please let us know if the code is good for merging. Thanks