Term — GPU rendering benchmark

Results

Same device (a Huawei Pura 70), same corpus, A/B against the CPU baseline using the built-in Log frame stats developer switch.

logFrameStats · Huawei Pura 70 · 3 MB ANSI-SGR + CJK corpus on a loop · 2026-07
Workload	CPU path	GPU path	Speed-up
Full-screen scroll (firehose)	14.6 ms · 68 fps	5.3 ms · 188 fps	2.7×
Slow scroll (~5 rows/frame)	12.8 ms · 76 fps	6.0 ms · 167 fps	2.1×

GPU frame time is almost independent of how much scrolled — every frame is a full redraw, and that is cheap for a GPU. The CPU path’s sensitivity to the number of scrolled rows simply disappears.

GPU frame breakdown

About 1000 instanced quads per frame. Glyphs are rasterised into a texture atlas once, and every later frame only re-emits mesh instances.

Re-measured 2026-07-17 · after the build micro-optimisations (ASCII slot table, merged bg+glyph walk, allocation-free atlas signature)
Stage	Average	What it does
build	2.15 ms	CPU side: walk the grid → instance quads + rasterise new glyphs into the atlas + upload
draw	0.22 ms	GL instanced draw
swap	2.95 ms	eglSwapBuffers (present)
Frame total	5.32 ms	≈ 188 fps

Two 14 MB memory shuffles are gone: no more scroll-shifting the shadow bitmap (~6 ms) and no more copy into the dmabuf (~5.6 ms) — that ~11 ms is the entire saving. The GPU rasterises in place, straight onto the EGL surface.

Captures

The corpus is coloured ANSI-SGR mixed with Chinese, Korean and ASCII, written to the session’s TTY by the benchmark script — no need to touch the phone.

Full-screen firehose scroll: a full screen of coloured ANSI plus Chinese/Korean/ASCII scrolling continuously — **P1 · Full-screen scroll** — continuous scrolling, every frame a total redraw.

Idle: partial presentation frames driven only by the blinking cursor — **P2 · Idle** — only the cursor blinks; partial presentation.

Single-row updates: typing at 8 Hz with the rows above preserved pixel for pixel — **P3 · Single-row update** — typing at 8 Hz, rows above preserved pixel for pixel.

Benchmark script

The script automates all three phases (firehose scroll / idle blink / single-row update) by writing to the TTY of a session on the remote host — zero interaction on the phone.

gpubench.shDownload ↓

# corpus: coloured ANSI-SGR + Chinese/Korean/ASCII, 400 rows
corpus() {
  i=0
  while [ $i -lt 400 ]; do
    printf "\033[3%dm%04d 端末描画性能測定 中文渲染基准 가나다라마 \033[1mBOLD\033[0m ascii-abcdefghijklmnop %05d\n" \
      $((i % 8)) $i $((i * 37))
    i=$((i + 1))
  done
}
C="$(corpus)"

# P1 firehose scroll, 30s (full frames, total damage)
end=$((SECONDS + 30))
while [ $SECONDS -lt $end ]; do printf "%s\n" "$C" > "$T"; done

# P2 idle, 60s (blink only → partial presentation frames)
sleep 60

# P3 single-row update, 20s (~8 Hz, typing-shaped partial frames)
end=$((SECONDS + 20))
while [ $SECONDS -lt $end ]; do
  printf "\rtyping simulation %06d" $n > "$T"; n=$((n + 1)); sleep 0.12
done

How to reproduce

Open a session to the target host; copy gpubench.sh there and find the session’s TTY with who -u.
Run ./gpubench.sh /dev/ttysNNN (the session’s tty).
Collect: hdc -t <target> shell hilog | grep "connrender: gpu" — one stats window every 60 frames.

Note: the P2/P3 stats windows run at low DVFS clocks (only 2–9 draws per second keeps the clock at its lowest step), so they are not directly comparable with the firehose numbers — only compare like with like. P1’s first window includes a one-off burst of CJK glyph rasterisation, which is expected.