Extracting timing from cosim

The co-simulation produces a VCD; calibration turns it into datapoints. This is the FIR-specific instance of the general calibration playbook and the AXI-MM burst extraction tools — driven by fir_calibrate.py (measure_cosim).

What to measure, and where

The FIR command has two bus-visible operations, each a top-level burst: the X-read and the Y-write. From the VCD those are the read and write bursts on the m_axi port. The begin/end events bracket exactly these, so the sim and the cosim measure the same spans.

from waveflow.utils.vcd import VcdParser, AximmBeatType

def transfer_beats(burst):
    return sum(1 for b in burst["beat_type"] if b == AximmBeatType.TRANSFER)

vp = VcdParser(vcd); clk = vp.add_clock_signal()
sigs, _ = vp.add_aximm_signals(prefix=gmem_prefix, dir="both")
writes, reads, clk_ns = vp.extract_aximm_bursts(clk_name=clk, aximm_sigs=sigs)

read_words  = sum(transfer_beats(b) for b in reads)    # X (+ taps) — true occupancy
write_words = sum(transfer_beats(b) for b in writes)   # Y

The important discipline: occupancy is the transfer-beat count, not the wall-clock span. The idle beats inside a burst are the upstream stall (the channel waiting on compute), which belongs to the compute block, not the bus — see transfer-beat occupancy. That separation is what makes the FIR primitives physical: the read/write occupancy comes out exactly n_row·(n_col+T) and n_row·(n_col−T+1) words (the kernel re-reads the T taps per row), with slope 1.

Sweep into a corpus

One cosim run is one datapoint; a sweep over sizes gives the corpus. FIR sweeps n_row ∈ {1, 2, 4, 8} × n_col ∈ {64, 256, 1024} — 12 points — recording for each the X-read span, the Y-write span, the fill, the whole-kernel, and the transfer-beat counts:

db = CalibDataFrame(columns=["n_row", "n_col", "x_read_span_cyc", "y_write_span_cyc",
                             "whole_kernel_cyc", "read_words", "write_words"])
for n_row in [1, 2, 4, 8]:
    for n_col in [64, 256, 1024]:
        db.add_datapoint(measure_cosim(n_row, n_col))   # one RTL cosim -> one row
db.save("results/cosim_grid.json")                       # the committed corpus

The result is the committed grid results/cosim_grid.json — 12 rows of ground-truth measurements. Choosing the grid deliberately (≥ 3 values per feature, an interior held-out point) is what lets the fit prove it generalizes rather than memorizes.

The corpus is the input to the model. The next page fits the one calibrated term from it, composes the whole-kernel, and shows the result.

Extracting timing from cosim

What to measure, and where

Sweep into a corpus

Next