[iwoplaza]

Every 'use gpu' function ships twice: as its JS implementation, and as its AST in the metadata. The JS implementation is only needed when the function is called on the CPU, which most of them never are.

With unstable_stripJsImpl: true, the plugin replaces the implementation with a stub. The first time the stub is called, it recreates the function from the AST.

__tgpu_meta(function __tgpu_stub(a0, a1, ...r) {
  return (__tgpu_stub.j ??= __tgpu_revive(__tgpu_stub))(a0, a1, ...r);
}, '[["a","b"],[[10,[1,"a","+","b"]]]]', {}, "add")

• reviveFunction (in typegpu/~internal) prints the AST back to JS with tseynit, rewrites + - * / % into __tsover_* calls like the plugin does, and builds the function with new Function. Externals that are property paths (config.nested.offset) become nested objects.
• The plugin imports reviveFunction only into modules that contain stubs, so apps that don't enable the option pay nothing.
• The stub keeps the original .length. TypeGPU relies on it: createGuardedComputePipeline infers its dimensionality from callback.length. An earlier prototype with a (...args) stub silently changed jelly-slider's compute shader from @workgroup_size(16, 16, 1) to (1, 1, 1).
• Stubs have no 'use gpu' directive, so later plugin passes leave them alone.

Caveats, which is why it's opt-in and unstable_:

• new Function requires a CSP that allows 'unsafe-eval'.
• Externals are read on the first call, not on every call.
• The recreated function has no source mapping.

Measurements

Same four apps and setup as #3246 (lab in #3248). Compared against #3246 with the option off (bytes, minified / gzip / brotli):

App	Off	On	Change
monocular-light-injection	505,710 / 124,788 / 104,119	473,574 / 115,910 / 96,707	−32,136 / −8,878 (−7.1%) / −7,412
selfie-segmentation	359,922 / 96,240 / 80,688	353,684 / 93,896 / 78,743	−6,238 / −2,344 (−2.4%) / −1,945
os-awards	459,476 / 124,605 / 104,049	446,767 / 119,392 / 99,585	−12,709 / −5,213 (−4.2%) / −4,464
jelly-slider	429,710 / 114,127 / 95,538	416,463 / 108,729 / 90,883	−13,247 / −5,398 (−4.7%) / −4,655

Together with unstable_obfuscate + unstable_minify, gzip drops by another 2,356–8,900 B compared to those options alone. The compressed gains are much larger than #3246's, because the removed JS is information the AST already holds, rather than syntax gzip can compress.

With the option off, bundles are byte-identical to #3246's.

Checks:

• A test app calls 'use gpu' functions on the CPU (vector operators, dotted externals, destructured params, loops) and prints their WGSL. The output is identical with and without the option, also with obfuscation.
• In Chromium, the four apps' shader modules are byte-identical to main's.

[github-actions]

pkg.pr.new

packages
Ready to be installed by your favorite package manager ⬇️

https://pkg.pr.new/software-mansion/TypeGPU/tinyest@af321c64c894de80cad0342a6b507129ca519be3

https://pkg.pr.new/software-mansion/TypeGPU/typegpu@af321c64c894de80cad0342a6b507129ca519be3

https://pkg.pr.new/software-mansion/TypeGPU/unplugin-typegpu@af321c64c894de80cad0342a6b507129ca519be3

benchmark
view benchmark

commit
view commit

[github-actions]

Resolution Time Benchmark

---
config:
  themeVariables:
    xyChart:
      plotColorPalette: "#E63946, #3B82F6, #059669"
---
xychart
  title "Random Branching (🔴 PR | 🔵 main | 🟢 release)"
  x-axis "max depth" [1, 2, 3, 4, 5, 6, 7, 8]
  y-axis "time (ms)"
  line [1.02, 2.04, 4.26, 6.91, 7.73, 12.34, 26.03, 25.71]
  line [1.01, 2.11, 4.57, 6.79, 8.12, 13.43, 26.98, 27.17]
  line [0.93, 1.90, 4.18, 6.75, 7.70, 12.85, 26.06, 26.00]

Loading

---
config:
  themeVariables:
    xyChart:
      plotColorPalette: "#E63946, #3B82F6, #059669"
---
xychart
  title "Linear Recursion (🔴 PR | 🔵 main | 🟢 release)"
  x-axis "max depth" [1, 2, 3, 4, 5, 6, 7, 8]
  y-axis "time (ms)"
  line [0.35, 0.59, 0.74, 0.88, 1.25, 1.26, 1.60, 1.76]
  line [0.33, 0.54, 0.74, 0.95, 1.21, 1.26, 1.51, 1.67]
  line [0.27, 0.50, 0.72, 0.87, 1.21, 1.19, 1.41, 1.57]

Loading

---
config:
  themeVariables:
    xyChart:
      plotColorPalette: "#E63946, #3B82F6, #059669"
---
xychart
  title "Full Tree (🔴 PR | 🔵 main | 🟢 release)"
  x-axis "max depth" [1, 2, 3, 4, 5, 6, 7, 8]
  y-axis "time (ms)"
  line [0.82, 2.23, 3.53, 6.98, 13.68, 29.17, 57.82, 121.85]
  line [0.98, 2.30, 3.77, 7.13, 13.52, 28.55, 61.65, 122.73]
  line [0.82, 2.10, 3.43, 7.05, 12.85, 26.73, 56.49, 113.94]

Loading

[github-actions]

Bundle size comparison (import * as ... in PR vs import * as ... in target):

🟢 Decreased	➖ Unchanged	🔴 Increased	❔ Unknown
0	352	0	0

No notable changes.

If you wish to run a comparison for other, slower bundlers, run the 'Tree-shake test' from the GitHub Actions menu.