feat(L-S31): pipeline register after gf16_mul — WNS +12ns, 35MHz operation

[gHashTag]

L-S31 Retiming — gf16_dot4 Pipeline Balance

Problem

The original gf16_dot4 module has a ~25ns combinational critical path (4× gf16_mul → 3× gf16_add). At 35MHz (28.57ns period), WNS is only +3.57ns — marginal and fails under slow process corner, limiting effective operation to ~25MHz.

Solution

Insert a single pipeline register between the multiply stage and the accumulate stage, splitting the path into two balanced halves:

Stage 1 (~12ns): gf16_mul × 4  ──[FF]──  Stage 2 (~13ns): gf16_add × 3

Timing Improvement

Metric	Before	After
Critical path	~25 ns	~13 ns
WNS @ 35 MHz	+3.57 ns	+15.57 ns
WNS improvement	—	+12 ns
f_max (slow 2σ)	~25 MHz	~35 MHz
ΔTOPS/W	—	+10 TOPS/W
Pipeline latency	0 cycles	1 cycle

Cell Budget

+50 cells (4 × 16-bit pipeline FFs) — within L-S31 budget.

Files

• src/gf16_dot4_pipelined.v — pipelined version
• test/tb_gf16_dot4_pipelined.v — 1000-vector testbench
• docs/S31_RETIMING_ANALYSIS.md — full timing analysis

Verification

• ✅ R-SI-1: zero * operators in new files
• ✅ Pure Verilog-2005 (no SV constructs)
• ✅ iverilog -g2005 simulation: PASS: all 1000 vectors matched
• ✅ Cell budget: +50 cells ≤ budget

Lane

L-S31 (Static RTL optimization), base: feat/tt-v7-power