Files
Dragon-s-Lair-X68k/tools/analysis/08_mode_map.py
T
prosolis 497f88b945 Rate control: rebuilt per-frame, wired in, and gated at zero drift
FINDINGS 26 stopped the session-5 rate controller before it shipped: it built a
lam-ladder of independent whole-sequence encodes and picked frames off it, so
SKIP blocks referenced reconstructions the decoder never saw -- 111 of 120
frames drifted. The fix is the structural one 26.1 said it had to be.

vq_hybrid is now frame-drivable -- frame_ctx / decide / paint -- and encode() is
a thin loop over it. Rate control drives the same three calls, bisects lam per
frame under the leaky bucket, and feeds back the frame it actually emitted. The
desync has no way to occur, and 09_ratectl_drift.py goes 111/120 -> 0/120. That
test is now part of check.sh, which is ~2 min rather than ~40 s.

Both overshoots on the worst sustained window are closed for under 1 dB, totals
including audio: sasi 137.4 -> 109.5 KB/s (-0.60 dB), scsi 381.6 -> 280.0 KB/s
(-0.91 dB). Zero frames hit the lam=800 cliff, so nothing was destroyed to get
there. Rate control also makes the display path cheaper -- scsi's median drops
53.6% -> 47.1% -- because raising lam moves blocks to SKIP and V1.

Two knobs measured rather than guessed. --rc-floor is worth 0.00 dB on that
window and defaults to the profile lam, so rate control cannot regress content
that already fits. --prefill defaults to 0 and is documented as a trap: it buys
a permission to overshoot of exactly bucket/nframes, and on a 14-frame clip it
disables the controller outright.

FINDINGS 26.5 was wrong in both halves and 27.6 records it. _paint was not the
bottleneck (14% of a frame, though vectorising it was still right at 17.1x) and
the ladder was never "minutes" -- those were k-means in build(). What makes
per-frame rate control affordable is that VQ.assign depends on neither lam nor
prev, so it is cached one frame deep: a 12-step search over 120 frames costs
0.31 s against 49.1 s.

Also caught: fixed-lam sasi was already 5% over target on 00020, the clip
everyone called easy. Nothing noticed because the profile table quotes PSNR and
not bitrate.

check.sh: ALL GREEN.

Claude-Session: https://claude.ai/code/session_01194oWYW8DQXK1SZ2DnChW6
2026-08-23 14:36:45 -07:00

117 lines
5.5 KiB
Python

#!/usr/bin/env python3
"""Render what the codec is actually DOING, per block, per frame.
Three panels at 12 fps: the palettised source (the real quality ceiling, not
1080p -- FINDINGS 11), the decoded output, and a block-mode map.
The mode map is not decoration. FINDINGS 24.5 makes the per-frame non-SKIP
fraction the number that selects the decoder's inner loop, and a percentile
cannot show you that the non-SKIP blocks are CLUSTERED (a moving character on a
held background) rather than scattered. Clustering is what a run-length over
the mode headers would exploit.
SKIP left as the previous frame, costs the 68000 nothing
V1 one 4x4 codeword, 1 byte
V4 four 2x2 codewords, 4 bytes
RAW 16 literal palette indices -- the escape that makes lam=0 pixel-exact
Usage: python3 tools/analysis/08_mode_map.py <frames_dir> <out.webm>
[--profile sasi|scsi] [--scale N] [--lossless] [--fixed-lam]
--fixed-lam renders the pre-session-6 encoder (no rate control) instead.
Output format follows the extension. Prefer .webm: GIF re-quantises to 256
colours, which is a poor fit for output whose subject is colour fidelity.
"""
import sys, os
sys.path.insert(0, os.path.join(os.path.dirname(os.path.abspath(__file__)), "..", "encoder"))
import numpy as np
from PIL import Image
import vq as VQ, vq_hybrid as H, ratectl as RC
MODE_RGB = np.array([[ 20, 22, 30], # SKIP - near black, costs nothing
[ 60, 150, 230], # V1 - blue
[ 80, 200, 120], # V4 - green
[235, 90, 70]], # RAW - red, the expensive escape
dtype=np.uint8)
LABEL = ["SKIP", "V1", "V4", "RAW"]
SCALE = 1
LOSSLESS = False
def main():
global SCALE, LOSSLESS
LOSSLESS = "--lossless" in sys.argv
src, out = sys.argv[1], sys.argv[2]
if "--scale" in sys.argv:
SCALE = int(sys.argv[sys.argv.index("--scale")+1])
prof = RC.PROFILES[sys.argv[sys.argv.index("--profile")+1]
if "--profile" in sys.argv else "sasi"]
m = H.build(src, k1=prof["k1"], k4=prof["k4"])
# Rate-controlled by default, so the map shows the mode decisions that
# actually ship. --fixed-lam renders the pre-session-6 encoder instead;
# the difference is visible as V4/RAW collapsing to V1/SKIP on peak frames.
if "--fixed-lam" in sys.argv:
enc = H.encode(m, lam=prof["lam"])
else:
enc = RC.encode_rate_controlled(m, prof["kbps"],
lam_lo=prof["lam"])
pal, H_, W_ = m["pal"], m["H"], m["W"]
nbx, nby = W_ // 4, H_ // 4
frames, stats = [], []
for f, (rec, mode) in enumerate(zip(enc["recon"], enc["modes"])):
srcp = pal[m["idx"][f]]
decp = pal[rec]
mmap = MODE_RGB[mode.reshape(nby, nbx)].repeat(4, 0).repeat(4, 1)
# tint the mode map with the decoded luma so the action stays legible
luma = decp.mean(2, keepdims=True) / 255.0
mmap = (mmap * (0.45 + 0.55 * luma)).astype(np.uint8)
gap = np.full((H_, 3, 3), 60, np.uint8)
panel = np.hstack([srcp, gap, decp, gap, mmap])
im = Image.fromarray(panel)
if SCALE != 1:
im = im.resize((panel.shape[1]*SCALE, H_*SCALE), Image.NEAREST)
frames.append(im)
stats.append([(mode == i).mean() for i in range(4)])
if out.endswith(".webm") or out.endswith(".mp4"):
# Preferred. GIF would impose its own 256-colour palette on top of
# output whose entire subject is colour fidelity, and costs ~4x the
# bytes doing it. -lossless keeps the panels pixel-exact.
import subprocess
w, h = frames[0].size
cmd = ["ffmpeg", "-v", "error", "-y", "-f", "rawvideo", "-pix_fmt", "rgb24",
"-s", f"{w}x{h}", "-r", "12", "-i", "-"]
# yuv444p, not 420: the mode map is flat saturated colour on a 4-pixel
# grid, and chroma subsampling smears exactly those edges. Lossless is
# available but runs larger than the GIF on this content; crf 18 in 444
# is visually clean at a quarter the size.
if out.endswith(".webm"):
cmd += ["-c:v", "libvpx-vp9", "-pix_fmt", "yuv444p", "-row-mt", "1"]
cmd += ["-lossless", "1"] if LOSSLESS else ["-crf", "18", "-b:v", "0"]
else:
cmd += ["-c:v", "libx264", "-crf", "12", "-pix_fmt", "yuv444p"]
p = subprocess.Popen(cmd + [out], stdin=subprocess.PIPE)
for f in frames:
p.stdin.write(np.asarray(f.convert("RGB")).tobytes())
p.stdin.close(); p.wait()
else:
# One shared adaptive palette: per-frame palettes are what make a naive
# GIF of this enormous, and a stable palette also stops the mode-map
# colours shimmering between frames.
shared = frames[0].quantize(colors=192, method=Image.MEDIANCUT)
q = [f.quantize(palette=shared, dither=Image.NONE) for f in frames]
q[0].save(out, save_all=True, append_images=q[1:],
duration=1000//12, loop=0, optimize=True)
st = np.array(stats)
print(f"{len(frames)} frames -> {out} ({os.path.getsize(out)/1024:.0f} KB)")
print(" panels: palettised source | decoded | block mode map")
for i, n in enumerate(LABEL):
print(f" {n:4s} mean {100*st[:,i].mean():5.1f}% "
f"per-frame range {100*st[:,i].min():5.1f}% .. {100*st[:,i].max():5.1f}%")
ns = 100 * (1 - st[:, 0])
print(f" non-SKIP: median {np.median(ns):.1f}% p90 {np.percentile(ns,90):.1f}%")
if __name__ == "__main__":
main()