Benchmarks
Two kinds of numbers: the CPU cost of what Helios does per frame (JMH, against a GL backend that does nothing), and the GPU cost of real workloads in the game (the harness). Report what you measured, including regressions.
How to run
./gradlew :helios:jmh # all CPU benchmarks, GC profiler on
./gradlew :helios:jmh -Pjmh.include=textLine # one benchmark (on Windows, avoid | in the pattern)
./gradlew :harness:runHarness25 -Pbench -PjvmArgs="-Dlazarus.gl=core" # GPU timings in the game
JMH results land in helios/build/results/jmh/results.json. The harness prints HELIOS-BENCH lines: each workload
draws into an offscreen framebuffer the size of the window between two glFinish calls, 20 warm-up runs, median and
p90 of 30 runs; each run counts as one frame for the streamed vertex buffers. GPU numbers include the CPU work of
building and uploading the batch. They are noisy at the 1 ms level; the first
workload of a run tends to read high.
CPU (JMH)
Windows 10, Java 25.0.4, measured 2026-11-19. ns/op and B/op (allocated bytes per operation). Shape benchmarks
are per shape, amortized over a 1000-shape batch and its flush.
| Benchmark | What | Before | After |
|---|---|---|---|
rectSolid |
one rounded rect, one colour | 68.7 ns, 0.5 B | 56.3 ns, 0.2 B |
rectGradient |
one rounded rect, linear gradient | 119.5 ns, 128 B | 97.4 ns, 0.2 B |
line |
one line with round caps | 69.6 ns, 0.5 B | 61.6 ns, 0.2 B |
textLine |
62-character line, layout + kerning + glyphs | 4509 ns, 4808 B | 2971 ns, 208 B |
textLineShadow |
same, classic shadow (twice the glyphs) | 9134 ns, 9072 B | 6010 ns, 208 B |
textFormatted |
line with colour and format codes | 4426 ns, 4440 B | 2699 ns, 208 B |
textWidth |
measure a line | 1477 ns, 4264 B | 128 ns, 0 B |
stateGuard |
open a guard, touch six slots, restore | 41.7 ns, 544 B | 12.8 ns, 160 B |
blurBehindCpu |
blur-behind bookkeeping (GL calls are no-ops) | 857 ns, 5928 B | 446 ns, 2824 B |
preprocessShader |
expand shape.fsh (load time only) |
39.6 µs | 38.7 µs |
Added with the utilities (H3), same machine:
| Benchmark | What | Result |
|---|---|---|
colorLerpHsb |
HSB interpolation of two colours | 19.7 ns, 0 B |
projectBox |
project an entity-sized box to its 2D bounds | 68.2 ns, 0 B |
animationValue |
read an eased animation (includes the clock read) | 32.5 ns, 0 B |
smoothUpdate |
advance a smoothed value (includes the clock read) | 33.4 ns, 0 B |
“Before” is the first measurement of the H2 code. The changes: kerning lookups for ASCII pairs go through a dense
table instead of a boxed HashMap key; StateGuard keeps saved state in fields and bit masks instead of an
EnumSet, a List and eight arrays; linear gradients no longer build a corner array; vertex colours are written as
one int each. The 208 B left per text line is the state guard of the batch flush.
Core profile (H3.5)
Helios moved to the OpenGL 4.1 core profile (Lazarus core mode): indexed triangles from its own vertex arrays, uploads through a fenced ring of buffers, matrices as uniforms. Same machine, measured 2026-11-18.
CPU (JMH), against the no-op backend:
| Benchmark | H3 (legacy) | H3.5 (core) |
|---|---|---|
rectSolid |
56.3 ns, 0.2 B | 54.3 ns, 0.1 B |
rectGradient |
97.4 ns, 0.2 B | 92.3 ns, 0.1 B |
line |
61.6 ns, 0.2 B | 66.9 ns, 0.1 B |
textLine |
2971 ns, 208 B | 2966 ns, 184 B |
textLineShadow |
6010 ns, 208 B | 6528 ns, 184 B |
textFormatted |
2699 ns, 208 B | 2919 ns, 184 B |
stateGuard |
12.8 ns, 160 B | 11.4 ns, 136 B |
blurBehindCpu |
446 ns, 2824 B | 524 ns, 2584 B |
preprocessShader |
38.7 µs | 42.5 µs |
Within noise, except that a flush now also reads the game’s two matrices and sets three more uniforms; the guard is smaller (fewer slots).
GPU (harness), Lazarus feat/core-gl at fce8f84, -Dlazarus.gl=core -Dlazarus.gl.debug=true:
| Workload | Median | p90 | With OptiFine HD U M5 |
|---|---|---|---|
| 10000 rounded rects (24x12, translucent) | 1.96 ms | 2.46 ms | 5.12 ms (first workload, see above) |
| 10000 circles with a radial gradient | 2.00 ms | 2.45 ms | 1.49 ms |
| 2000 glyphs of 9 px text (32 lines) | 0.23 ms | 0.34 ms | 0.22 ms |
| One MSDF text batch (1 line, 1 draw call) | 0.029 ms | 0.031 ms | 0.024 ms |
| One blur pass (1 iteration, full screen) | 0.057 ms | 0.059 ms | 0.056 ms |
| Full-screen blur behind, 4 iterations | 0.20 ms | 0.31 ms | 0.17 ms |
| Full-screen bloom, 3 iterations | 0.15 ms | 0.17 ms | 0.14 ms |
Faster than the legacy path across the board on this machine.
macOS, Apple M2 (“4.1 Metal”), 1708x960 Retina framebuffer, Java 25, Lazarus feat/core-gl at 0e47336, one run per
column. map is -Dhelios.stream=map (unsynchronized glMapBufferRange into the fenced ring):
| Workload | glBufferSubData (default) |
map |
|---|---|---|
| 10000 rounded rects | 7.99 ms | 8.24 ms |
| 10000 circles with a radial gradient | 4.81 ms | 8.06 ms |
| 2000 glyphs of 9 px text (32 lines) | 0.53 ms | 0.60 ms |
| One MSDF text batch (1 line, 1 draw call) | 0.21 ms | 0.21 ms |
| One blur pass (1 iteration, full screen) | 0.54 ms | 0.77 ms |
| Full-screen blur behind, 4 iterations | 1.27 ms | 1.82 ms |
| Full-screen bloom, 3 iterations | 1.21 ms | 1.25 ms |
Mapping is nowhere faster on the M2, so glBufferSubData stays the default. The Mac framebuffer has 1.8 times the
pixels of the Windows one, which accounts for most of the gap in the full-screen passes.
GPU (harness, legacy context, H2)
Windows 10, NVIDIA GeForce RTX 4070 (driver 581.80), 1281x720, Java 25, measured 2026-11-19.
| Workload | Median | p90 | With OptiFine HD U M5 |
|---|---|---|---|
| 10000 rounded rects (24x12, translucent) | 3.50 ms | 7.54 ms | 3.47 ms |
| 10000 circles with a radial gradient | 2.45 ms | 5.72 ms | 2.17 ms |
| 2000 glyphs of 9 px text (32 lines) | 0.35 ms | 0.80 ms | 0.27 ms |
| Full-screen blur behind, 4 iterations | 0.42 ms | 0.84 ms | 0.33 ms |
| Full-screen bloom, 3 iterations | 0.36 ms | 0.86 ms | 0.29 ms |
Font generation
fontgen builds the bundled Inter atlas (201 glyphs, 4224 kerning pairs, 512x512) in about 0.9 s on the same
machine, as part of ./gradlew build.