Optimize x86 fp16s innerproduct gemm to eliminate loop-carried stalls

[Edwardssss]

PR Description:

Overview
This PR significantly improves the performance of innerproduct_gemm_fp16s_sse on x86 architectures by mitigating severe loop-carried dependency stalls and replacing inefficient instructions in the microkernels.

Problem
The existing innerproduct_gemm_fp16s implementation had two major bottlenecks:

1. Loop-carried Stalls: FMA loop sequentially reused the same destination registers (e.g. _sum0 to _sum3) in accumulation iterations. Given the 4-5 cycle latency of typical FMA instructions, this caused severe processor pipeline stalls, wasting execution ports constraints.
2. Instruction Inefficiency: Used _mm256_extractf128_si256 for FP16->FP32 conversions which is slightly more expensive compared to memory operand fusion.

Solution

• Loop Unrolling & Accumulator Expansion: Fully unrolled the loops in elempack == 1 & num_output_elempack == 8 / 16 / 4 blocks. We now use up to 8 independent accumulation registers to completely break the false data dependency, hiding FMA latency.
• Memory Operand Fusion: Dropped _mm256_extractf128_si256 in favor of direct 128-bit loads (_mm_lddqu_si128) fused into _mm256_cvtph_ps, reducing register traffic.

Benchmark Results (AMD Family 17h, 12 Cores, FP16 mode)
Tested via benchncnn 4 6 2 -1:

Model	Original (min)	PR (min)	Improvement
vgg16	503.22	417.70	~17.0%
resnet50	326.63	284.35	~13.0%
resnet18	114.28	98.78	~13.5%

AMDuProf Analysis:
Profiling the single-process CPU runtime confirms that the innerproduct_gemm_fp16s_sse hotspot significantly dropped CPU_TIME:

• Before optimization: 67.87s
• After optimization: 55.14s (Total execution lowered by >12s).

Commit Split:

1. Optimize innerproduct x86 fp16s gemm using fused loads and fully unrolled FMA (Addressed the 1x8 case alongside the load logic rewrite).
2. Alleviate loop-carried stalls in innerproduct fp16s microkernels by unrolling (Extrapolated the fixes across 1x16 and 1x4 blocks respectively).

[tencent-adm]

All committers have signed the CLA.

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 93.99%. Comparing base (086dda4) to head (904955b).
⚠️ Report is 1 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #6682      +/-   ##
==========================================
+ Coverage   93.65%   93.99%   +0.34%     
==========================================
  Files         930      930              
  Lines      296508   298257    +1749     
==========================================
+ Hits       277688   280341    +2653     
+ Misses      18820    17916     -904     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.