[taronaeo]

Overview

This PR adds s390x vector intrinsic into various vector functions. Majority of the implementations were referenced from the ARM implementation. We see an at-best performance improvement of approximately 1.13x for pp512 and 1.10x for tg128.

Additional information

Perplexity

Details

Granite-3.0-1B-A400M

====== Perplexity statistics ======
Mean PPL(Q)                   :  10.604782 ±   0.189686
Mean PPL(base)                :  10.578525 ±   0.188078
Cor(ln(PPL(Q)), ln(PPL(base))):  99.78%
Mean ln(PPL(Q)/PPL(base))     :   0.002479 ±   0.001189
Mean PPL(Q)/PPL(base)         :   1.002482 ±   0.001191
Mean PPL(Q)-PPL(base)         :   0.026258 ±   0.012639
====== KL divergence statistics ======
Mean    KLD:   0.008938 ±   0.000227
Maximum KLD:   1.428281
99.9%   KLD:   0.532981
99.0%   KLD:   0.098483
95.0%   KLD:   0.021852
90.0%   KLD:   0.014622
Median  KLD:   0.003667
10.0%   KLD:   0.000036
 5.0%   KLD:   0.000006
 1.0%   KLD:  -0.000000
 0.1%   KLD:  -0.000003
Minimum KLD:  -0.000030
====== Token probability statistics ======
Mean    Δp:  0.018 ± 0.014 %
Maximum Δp: 29.722%
99.9%   Δp: 15.448%
99.0%   Δp:  6.623%
95.0%   Δp:  3.027%
90.0%   Δp:  1.672%
75.0%   Δp:  0.280%
Median  Δp:  0.000%
25.0%   Δp: -0.278%
10.0%   Δp: -1.598%
 5.0%   Δp: -2.923%
 1.0%   Δp: -6.366%
 0.1%   Δp: -14.088%
Minimum Δp: -63.157%
RMS Δp    :  2.163 ± 0.052 %
Same top p: 95.706 ± 0.127 %


Llama-3.2-1B
====== Perplexity statistics ======
Mean PPL(Q)                   :  18.020235 ±   0.743449
Mean PPL(base)                :  18.009705 ±   0.741459
Cor(ln(PPL(Q)), ln(PPL(base))):  99.97%
Mean ln(PPL(Q)/PPL(base))     :   0.000585 ±   0.000992
Mean PPL(Q)/PPL(base)         :   1.000585 ±   0.000993
Mean PPL(Q)-PPL(base)         :   0.010530 ±   0.017917
====== KL divergence statistics ======
Mean    KLD:   0.000871 ±   0.000016
Maximum KLD:   0.024187
99.9%   KLD:   0.012690
99.0%   KLD:   0.005106
95.0%   KLD:   0.002253
90.0%   KLD:   0.001615
Median  KLD:   0.000666
10.0%   KLD:   0.000085
 5.0%   KLD:   0.000022
 1.0%   KLD:  -0.000004
 0.1%   KLD:  -0.000036
Minimum KLD:  -0.000086
====== Token probability statistics ======
Mean    Δp:  0.009 ± 0.009 %
Maximum Δp:  6.510%
99.9%   Δp:  4.751%
99.0%   Δp:  1.990%
95.0%   Δp:  1.035%
90.0%   Δp:  0.598%
75.0%   Δp:  0.132%
Median  Δp:  0.000%
25.0%   Δp: -0.127%
10.0%   Δp: -0.576%
 5.0%   Δp: -1.010%
 1.0%   Δp: -1.964%
 0.1%   Δp: -4.064%
Minimum Δp: -4.786%
RMS Δp    :  0.664 ± 0.018 %
Same top p: 97.980 ± 0.197 %

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES, Claude Code + Opus 5.5