neural_processor_packed.v: split the original single "Stage 1" (packed
DSP48E1 multiply + INT8 unpack + register) into two real stages --
Stage 1a registers the raw DSP48E1 product with zero logic in between,
Stage 1b does the carry-heavy unpack (the real critical path EXP-0094
traced) from that already-registered value. Adds exactly one real
clock cycle of latency; throughput unaffected (real valid/ready
handshaking throughout, no fixed-latency assumption downstream).
Real verification: isolated bit-exact vs 2x real neural_processor.v
(18/18 PASS, testbench fixed to latch each core's result independently
since result_valid is a one-shot pulse and the DUT is now one cycle
deeper -- not an RTL bug). Full-system functional xsim on real DDR3:
32/32 PASS. Real, full P&R: WNS=+0.269ns, WHS=+0.026ns, 0 failing
setup or hold endpoints -- N=16 TIMING CLOSES.
Also root-caused (not an RTL bug, folded into CLAUDE.md): a real
Vivado incremental-synthesis quirk silently carried forward a
N_GROUPS=2 parameter binding from an earlier sweep run despite no
-generic override and an intervening reset_run -- fixed by always
passing -generic explicitly and confirming the real elaborated value
via a post-synth DSP48E1 count.
Isolated on this branch -- does not touch the physical board already
in fabrication on v3-artix7 (N=8, unmodified).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUG92aM9m68TRc4rG55BcC
New hardware/v3/ (Artix-7 port, branch v3-artix7): the compute engine
that makes the 100x-vs-ESP32 target theoretically reachable on
XC7A100T's 240 DSP48E1 budget.
mac2_dsp_packed.v: packs 2 INT8 MACs sharing one resident weight into
a single DSP48-shaped 25x18 multiply, exploiting this project's own
weight-stationary reuse pattern (layer_weight_buffer.v, EXP-0057/0058)
where one weight is genuinely multiplied against many different
activations. Verified exhaustively: 16,777,216/16,777,216
(weight,x0,x1) combinations, 0 errors.
Two real bugs found and fixed during that verification (both purely
arithmetic/RTL, not toolchain-related):
1. An off-by-one in a declared wire width caused Verilog's part-select
unsigned-by-default rule to corrupt sign extension on the upper
product field -- ~50% of vectors failed.
2. After fixing (1), still ~50% failed: concatenating two independently
sign-extended fields ({sext(x1,9), sext(x0,16)}) is NOT equivalent
to the real arithmetic sum x1*2^16+x0 whenever the lower field is
negative (its own two's-complement encoding "bleeds" an extra 2^16
into the concatenated value). Fixed by building the packed operand
with an explicit arithmetic shift-and-add instead of concatenation.
neural_processor_packed.v: full port of hardware/v2/rtl/
neural_processor.v's pipeline (same stage count/structure), doubled on
the accumulator/bias/activation/saturation side to process two
weight-reuse positions per weight-tile stream. Verified against TWO
real hardware/v2/rtl/neural_processor.v instances (job A / job B, same
shared weight, independent activations) -- 18/18 PASS, 0 errors,
covering the functional sweep, INT8 extremes, and back-to-back jobs.
A third real bug found in the process (in the new testbench, not the
RTL): clearing operand_valid/tile_last in the same simulation delta as
the handshake edge that should register tile_last=1 races against the
DUTs' own FSM evaluation of that same edge -- the same pulse-clearing
race class found three times already today in hardware/v2/sim (EXP-0058
and its follow-up commits). Fixed the same way: hold the pulse past the
edge with a real time delay (#1) before clearing.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MUG92aM9m68TRc4rG55BcC
micheleandClaude Sonnet 5