silk/arm/NSQ_del_dec_neon_intr.c — Multiple optimizations

[rl123567]

silk/arm/NSQ_del_dec_neon_intr.c — Multiple optimizations

• LPC prediction: Replace 16 individual vld1q_s32 loads with ldp (load-pair) inline assembly, loading 8 data vectors + 2 coefficient vectors per block. Use vqdmulhq_laneq_s32 to index directly into 128-bit coefficient vectors, eliminating vget_low/high_s32 split overhead.
• Allpass sections: Fully unroll the allpass loop via macro with compile-time branching on shapingLPCOrder, enabling precise register allocation and removing loop overhead.
• Common subexpression elimination: Precompute Tilt_Q14_Q16 and LF_shp_Q14_Q15 outside the per-sample loop; cache LF_AR_Q14 to avoid redundant vector loads.
• AR_shp_Q28 widening: Remove the scalar tail loop since MAX_SHAPE_LPC_ORDER (24) is exactly divisible by 8.
• State replacement copy: Unroll the numOthers copy loop 16× with __builtin_prefetch to hide memory latency during the candidate state swap.

[rl123567]

Hi, @jmvalin Could you please help review this PR when you have a moment? It includes targeted NSQ_del_dec_neon_intr.c Multiple optimizations，for details, see the commit message. Please let me know if you need any additional information or if there are any changes required. Thanks in advance for your time!

[lpi]

Tested head: 330eae9f51f144ec195000e5554d0ebbef8724f9.

I tested safe ideas independently on top of #503. Their approximate combined production-c10 encoder improvements versus stock were:

• paired/laneq: +4.0820% CPU / +4.0825% wall
• Tilt/LF hoists: +4.0543% / +4.0692%
• LF_AR cache: +4.1450% / +4.1417%
• Reduce state-copy pressure in the hottest SILK mode #503 alone: +4.0713% / +4.0826%

Positive means faster/lower per-call cost. These stock-relative numbers are compounded cross-run estimates, not direct stock arms. Only LF_AR caching added a small increment over #503: +0.0768% CPU / +0.0616% wall.

I did not time the submitted bundle wholesale because its inline assembly does not model pointed-to memory, its prefetch pointer formation extends beyond the derived state-row region, and its state-copy rewrite overlaps #503. The allpass-unroll and AR-tail ideas compiled to equivalent Release code. The tested isolated variants were output-exact and sanitizer-clean.