Two related Phase 5 additions, both threaded the same way (a new
runtime field defaulting to the pre-existing behavior, settable
per-layer via the descriptor table or per-run via SET_BASE):
Configurable activation functions:
- neuron_parallel.v gains a 2-bit `activation` port (ACT_NONE =
linear + two-sided INT8 saturate, ACT_RELU = the original
hardwired behavior, kept as the default so every pre-existing
caller/testbench is unaffected), threaded through neuron_memory.v.
- spi_engine.v: SET_BASE sel=6 (single-layer path); the descriptor
table gains a 7th byte (multi-layer path).
- Verified in neuron_parallel_tb.v (negative pass-through + negative
saturation to -128) and end-to-end in
spi_neuron_top_runnetwork_tb.v (a real negative accumulator that
ACT_RELU would clamp to 0 comes through unclamped under ACT_NONE,
over real SPI/RAM).
Runtime network width (one bitstream, any topology up to its
build-time max, entirely host-configured over SPI):
- neuron_parallel.v gains n_inputs_real, bounding its MAC group loop
(n_inputs_real/PARALLEL groups instead of the fixed build-time
count). neuron_memory.v gains n_inputs_real/n_neurons_real,
bounding its X/W RAM-read loop and its neuron loop. All default to
the build-time max, so unconnected callers are unaffected.
n_inputs_real must stay a multiple of PARALLEL (same constraint
N_INPUTS itself is held to at elaboration time, now the caller's
runtime responsibility).
- spi_engine.v: SET_BASE sel=7/8 (single-layer path); the descriptor
table grows to 11 bytes/layer (+n_inputs_real +n_neurons_real,
multi-layer path) -- layer_sequencer.v also now copies only
n_neurons_real bytes into the ping-pong buffer, not the full
build width.
- This is real early termination, not bookkeeping: no RAM
zero-padding needed for the unused tail, and it measurably
completes faster. neuron_parallel_tb.v TEST 7: 3 cycles vs 6 for a
reduced-vs-full run, with garbage loaded into the skipped lanes to
prove they're never read. neuron_memory_tb.v TEST 5: through the
real PSRAM stack, 209 cycles vs 788. layer_sequencer_tb.v proves a
reduced n_neurons_real shortens the ping-pong copy-out itself
(bytes beyond the real count stay untouched, not just differing).
docs/FPGA-NeuralNetwork-Engine.md: §8.1 opcode/SET_BASE table, new
"Runtime network width" subsection, Phase 5 checklist, Current
Status table, and the "Core architectural principle" statement
updated to reflect that topology (not just trained parameters) is
now host-configured at runtime up to a build-time ceiling.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WQV3vS9TXaGDJ5cRfnfidt
Yosys + nextpnr-ecp5 synthesis of the extended spi_neuron_top.v
(layer_sequencer + mux + arbiter Port C wired in), speed grade -8,
N_INPUTS=32/N_NEURONS=1: PARALLEL=8 -> 40.57 MHz FAIL, PARALLEL=2 ->
42.54 MHz FAIL. Both worse than Phase 4 alone (~52.58/~55.85 MHz for
the identical configs) -- the Phase 5 wiring cost real timing
headroom. Critical path also shifted from Phase 4's saturation-
comparator finding to neuron_memory's x_mem/w_mem LUT-RAM read-mux
tree feeding the accumulator. Full analysis and candidate next steps
(block RAM for x_mem/w_mem, pipelining, a seed sweep to separate
placement noise from a structural bottleneck) recorded under Phase 7
in docs/FPGA-NeuralNetwork-Engine.md.
tools/fpga_benchmark.py: parametric Yosys + nextpnr-ecp5 benchmark
harness for the LFE5U-45F-8BG381 (speed grade -8, 80 MHz target),
sweeping PARALLEL over the neuron layer and parsing Fmax/LUT4/DFF/DSP
utilization out of the nextpnr report into JSON/CSV.
synth/ecp5/p2, p4, p8: real synthesis+PnR results backing the
same-price-tier FPGA comparison (P2: 87.88 MHz PASS, P4: 75.01 MHz
FAIL, P8: 147.62 MHz PASS -- non-monotonic, dominated by placement
noise since the whole design uses <2% of the device's LUT4 fabric at
every setting, and P8 notably maps to 0 DSP blocks vs 8/16 for P2/P4).
synth/ecp5/top.v: benchmark harness top-level, reworked to generate
deterministic non-constant X/weights/bias via `keep`-attributed
generate blocks so Yosys can't constant-fold the datapath away.
Also adds .gitignore for Python's __pycache__/*.pyc.
Wires the already-present layer_sequencer.v into the SPI stack:
- spi_engine.v: RUN_NETWORK opcode (0x23) + SET_BASE selectors for
table_base/buf_a_base/buf_b_base; STATUS.busy/done extended to
track the sequencer (seq_busy/seq_done) alongside neuron_memory
directly, so done latches on the last layer only.
- spi_neuron_top.v: instantiates layer_sequencer, muxes
neuron_memory's control inputs between it (while seq_busy) and
spi_engine's direct-drive path (legacy single-layer mode), wires
the sequencer's own RAM master to mem_arbiter's Port C.
Found and fixed a real race while writing the end-to-end test: STATUS's
sticky/clear-on-read done bit read its value live/combinationally
during transmission and cleared unconditionally on any STATUS read.
A done_event landing mid-transmission of a STATUS response byte could
be silently dropped -- the host would receive a stale byte while the
sticky bit was cleared regardless, hanging any host polling STATUS in
a loop. Present since Phase 4, not RUN_NETWORK-specific; only
surfaced under this test's continuous polling. Fixed by latching a
status_snapshot at opcode-accept time and gating the clear on what
was actually transmitted.
Tests: spi_engine_tb.v gains RUN_NETWORK/SET_BASE opcode tests (K/L);
new layer_sequencer_tb.v unit-tests the sequencer FSM directly
(descriptor table, ping-pong buffer addressing, byte-exact copy-out);
new spi_neuron_top_runnetwork_tb.v drives a real 2-layer network over
simulated SPI end to end (real neuron_memory + PSRAM, hand-computed
expected output) and confirms the legacy single-layer path still
works afterward. All existing testbenches still pass.
Implements the rest of the SPI interface (docs §8.1) on top of
spi_slave.v from the previous commit:
- rtl/spi_engine.v: opcode FSM + register bank, all 8 opcodes (NOP,
WRITE_RAM, READ_RAM, RESET, SET_BASE, START, STATUS, READ_OUTPUT,
READ_CONFIG). tx_byte is driven combinationally from live state
(not reactively on tx_byte_req), applying the prefetch-vs-consume
contract documented on spi_slave.v. STATUS.done is a sticky,
clear-on-read latch. RAM master port uses the same byte-level
convention as neuron_memory.v's external mem_* port.
- rtl/mem_arbiter.v: fixed-priority (neuron_memory > spi_engine)
grant-and-forward arbiter sharing one byte-level memory port
between spi_engine's WRITE_RAM/READ_RAM and neuron_memory's own
X/W/bias reads during a run.
- rtl/spi_neuron_top.v: full integration -- spi_slave -> spi_engine
-> mem_arbiter -> a single shared int8_memory_access ->
memory_interface -> psram_controller -> PSRAM pins. neuron_memory's
rst is global rst OR'd with the RESET opcode's soft-reset pulse.
The host has no direct electrical path to the RAM, only through
this chain.
Testing:
- sim/spi_engine_tb.v: 10 tests (one per opcode + WRITE_RAM/READ_RAM,
START idle-vs-busy, STATUS sticky/clear-on-read, extra-MOSI-bytes-
ignored, back-to-back transactions) against a synthetic 2-cycle-
latency RAM model, isolating the opcode FSM from PSRAM timing.
Found and fixed two testbench-only bugs (RTL needed no change):
the same delta-zero clock-edge race as spi_slave_tb.v (blocking
`nm_done=1` landing on the same sim time as a posedge -- fixed via
negedge-based pulsing) and a missing RAM sentinel initialization.
- sim/spi_neuron_top_tb.v: end-to-end test against the **real**
psram_model.v (not a mock) -- RESET/READ_CONFIG/WRITE_RAM/
READ_RAM/SET_BASE/START/STATUS/READ_OUTPUT all driven purely over
simulated SPI. 3/3 scenarios (sum, saturation, ReLU) pass on the
first attempt; confirms the arbiter and shared byte<->word bridge
are correct against real PSRAM timing, not just a synthetic mock.
Real-toolchain verification (Yosys + nextpnr-ecp5 + ecppack):
spi_slave.v and spi_engine.v synthesize clean and comfortably clear
80 MHz in isolation (403 MHz / 191 MHz, no DSP usage). The full
spi_neuron_top.v integration, however, does NOT meet 80 MHz
(~52-56 MHz depending on PARALLEL) -- the critical path is entirely
inside neuron_parallel.v's existing saturation comparator (no
contribution from the new SPI/arbiter logic), but its routed delay
is ~57% worse than in the isolated benchmark due to placement/
routing congestion once SPI + PSRAM logic shares the fabric with
it, not resource exhaustion (2% DSP usage). Documented as a Phase
4/7 finding in docs/FPGA-NeuralNetwork-Engine.md -- a floorplanning/
pipelining problem for Phase 7, not a functional-correctness issue
(verified independently in simulation against real PSRAM timing).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WQV3vS9TXaGDJ5cRfnfidt
First real RTL piece of the SPI interface (docs §8.1 protocol
draft): the physical layer only -- Mode 0 (CPOL=0, CPHA=0),
MSB-first, byte-level shift register with a 3-stage CDC synchronizer
for SCLK/MOSI/CS_N (the SPI master clock is asynchronous to the
FPGA system clock). Exposes rx_byte/rx_valid, tx_byte/tx_byte_req,
and cs_active/cs_start/cs_end to the (not yet written) protocol
engine.
Documented an important consumer contract on tx_byte_req: it is a
prefetch hint (fires once extra after the last byte of every
transaction, since the slave cannot know in advance whether the
master will keep clocking), not a "byte consumed" event -- a
consumer must advance any stateful pointer (e.g. a RAM read address)
on rx_valid instead, which fires exactly once per real byte
transferred.
sim/spi_slave_tb.v: bit-banged SPI master BFM (4 tests: single byte,
multi-byte in one CS period, back-to-back transactions, slower
SCLK). Two testbench-only bugs found and fixed during bring-up (RTL
itself needed no functional change beyond the tx_byte_req contract
comment): the BFM was advancing its tx queue on tx_byte_req instead
of rx_valid (see contract above), and inter-test reset pulses raced
against posedge clk (blocking `rst=1` landing on the same simulation
time as a clock edge) -- fixed by asserting/deasserting reset on
negedge clk instead.
Verified two ways: Icarus Verilog (4/4 tests pass) and the real
ECP5 toolchain used for prior benchmarks (Yosys 0.68 synth: 0
problems, 41 FF / 55 LUT4, no latches; nextpnr-ecp5 --45k --package
CABGA381 --speed 8 --freq 80: PASS, Fmax 403.23 MHz; ecppack:
bitstream generated with no errors).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WQV3vS9TXaGDJ5cRfnfidt
Phase 4 (SPI Interface) only had a high-level conceptual sequence
(RESET/CONFIGURE/LOAD.../START/WAIT/READ) with no concrete opcodes,
framing, or register map -- not enough to start RTL from. Added
docs/FPGA-NeuralNetwork-Engine.md §8.1 with a concrete v1 draft:
- SPI Mode 0, MSB-first, one opcode byte per CS-low transaction.
- Explicit length field on WRITE_RAM/READ_RAM (chosen over
CS-edge-delimited streaming: simpler controller, just a byte
counter).
- READ_CONFIG opcode exposing N_INPUTS/N_NEURONS/PARALLEL/
ADDR_WIDTH/DATA_WIDTH at runtime, so one host firmware build can
target different bitstreams.
- RESET kept as its own opcode (0x0F), distinct from NOP.
- STATUS.done documented as required to be a STICKY, clear-on-read
bit in the SPI register bank: neuron_memory.done is a one-cycle
pulse that a slow SPI poll would almost certainly miss otherwise.
Opcode values themselves are marked explicitly as draft/example,
not frozen -- only the framing rules and the two decisions above are
meant to stick going into Phase 4 RTL work.
No RTL or testbench changes in this commit; design-only.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WQV3vS9TXaGDJ5cRfnfidt
neuron_memory.v only handled a single neuron. Added an N_NEURONS
parameter (default 1, fully backward compatible) and a memory-bound
neuron loop: X is read once (shared layer input), and for each
neuron in turn W and bias are re-read from PSRAM and fed to a
single, reused neuron_parallel instance -- no change to the
validated compute datapath (neuron_parallel/mac8/mac_unit).
Addressing follows layer.v's neuron-major convention: neuron n's
weights live at w_base + n*N_INPUTS bytes, its bias at
bias_addr + n. Output changed from a single `y` port to a packed
`y_bus` (DATA_WIDTH*N_NEURONS bits, neuron-major), matching
layer.v's y_bus.
- rtl/neuron_memory.v: N_NEURONS parameter, neuron_index/
w_group_base/bias_group_addr tracking, y_reg[] array assembled
into y_bus, STATE_WAIT_N now loops back to STATE_READ_W for the
next neuron instead of finishing after one.
- sim/neuron_memory_tb.v: updated to the new y_bus port
(N_NEURONS=1 explicit); all 5 existing tests still pass unchanged,
confirming backward compatibility.
- sim/neuron_memory_multi_tb.v: new end-to-end test (full
memory_interface + psram_controller + psram_model stack) with
N_NEURONS=3, validating per-neuron addressing and a single done
pulse at the end of the sequence (scale, larger value, ReLU).
- Full regression re-run: all existing testbenches still pass.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WQV3vS9TXaGDJ5cRfnfidt
Both Phase 2 findings (docs/FPGA-NeuralNetwork-Engine.md) shared one
root cause: GROUPS = N_INPUTS / PARALLEL is integer division. When
N_INPUTS is not an exact multiple of PARALLEL, the remainder inputs
were silently dropped from the accumulation (wrong result, no
error); when PARALLEL > N_INPUTS, GROUPS = 0 and the controller's
terminal condition was never met, hanging the neuron forever.
Added a single elaboration-time guard to rtl/neuron_parallel.v: a
`generate` block instantiates a deliberately undefined module when
N_INPUTS % PARALLEL != 0, forcing a hard failure in both simulation
and synthesis instead of a silent wrong answer or a deadlock. Valid
configurations are unaffected (the branch is never elaborated). The
validated datapath (mac8/mac_unit/accumulation/ReLU/saturation) is
untouched -- this is authorized as a scoped exception to the
"core is fixed, do not touch" project policy, for this guard only.
- sim/neuron_parallel_guard_negative_nonmultiple_tb.v and
sim/neuron_parallel_guard_negative_degenerate_tb.v: negative tests
that must fail to elaborate; verified both fail with the expected
"Unknown module type" error.
- sim/parameter_sweep_tb.v: rewritten to valid-configs-only (the
three configs that used to demonstrate truncation/hang no longer
compile, by design); added PARALLEL=2 and PARALLEL=4 configs,
the two best-performing values from
docs/FPGA-Neural-Datapatch-Benchmark.md.
- Full regression re-run after the RTL change: all existing
testbenches still pass unchanged.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WQV3vS9TXaGDJ5cRfnfidt
Roadmap Phase 2 asks to validate N_INPUTS/N_NEURONS/PARALLEL
combinations, including non-exact-multiple configurations. Added
sim/parameter_sweep_tb.v with 5 configs (two exact-multiple sanity
checks, two non-exact-multiple, one degenerate PARALLEL>N_INPUTS),
using a cycle-count watchdog instead of a blocking wait so a hanging
config is reported rather than hanging the simulation.
Findings (RTL unchanged, core datapath left untouched):
- GROUPS = N_INPUTS / PARALLEL truncates: when N_INPUTS is not an
exact multiple of PARALLEL, the remainder inputs are silently
never summed (confirmed 30/8 -> 6 dropped, 20/16 -> 4 dropped).
- PARALLEL > N_INPUTS gives GROUPS=0, and the controller's
group_index == GROUPS-1 terminal condition is never met: the
neuron hangs forever (confirmed via watchdog timeout).
Documented both as findings under Phase 2 in
docs/FPGA-NeuralNetwork-Engine.md for follow-up in Phase 3/7.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WQV3vS9TXaGDJ5cRfnfidt
Both testbenches instantiated their DUTs with a FRAC_BITS parameter and
Q8.8 fixed-point 16-bit values, which no longer exist in rtl/neuron_parallel.v
(now plain INT8, DATA_WIDTH=8, hardcoded +127 saturation, ReLU-only clamp).
This made both tests fail elaboration ("parameter FRAC_BITS not found").
Rewrote both benches with integer INT8 stimuli and expectations matching
the current core (no RTL changes): neuron_parallel_tb covers a mixed
vector, ReLU, positive saturation, and mixed values with a boundary
negative bias; layer_tb covers 8 neurons exercising scale, ReLU,
saturation, bias-only, and a sparse weight pattern across groups.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WQV3vS9TXaGDJ5cRfnfidt
preload_vector and preload_weights wrote using base as a word address
instead of a byte address (base + (k>>1) instead of (base>>1) + (k>>1)),
misaligning X/W data in PSRAM relative to what int8_memory_access expects.
Also adds a PATTERN test (X=1..32) to exercise mixed even/odd byte reads
and catch this class of bug going forward.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WQV3vS9TXaGDJ5cRfnfidt
micheleandClaude Sonnet 5