[iwoplaza]

Part of a stack that makes the metadata of 'use gpu' functions smaller and cheaper to load (#3246).

typegpu can now read metadata registered as [json, externals, name?] (v3), which the plugin starts emitting in #3246. v1 and v2 are still read, so code transformed by older plugin versions keeps working.

• json holds [params, statements]. It's parsed on first use (getFunctionMetadata), not when the bundle loads.
• Identifier params are plain strings ("a" instead of {"type":"i","name":"a"}). Destructured params are [name, alias][].
• getName reads the name from either shape without parsing the AST.

Cost: the reader adds 549–578 B minified / 175–222 B gzip to an app. That's measured on four docs examples built with Vite 8, against main (setup in #3248). #3246 saves more than that in every one of them.

[github-actions]

pkg.pr.new

packages
Ready to be installed by your favorite package manager ⬇️

https://pkg.pr.new/software-mansion/TypeGPU/tinyest@7be6355bed339b2b0b9027bcaf34879c45349038

https://pkg.pr.new/software-mansion/TypeGPU/typegpu@7be6355bed339b2b0b9027bcaf34879c45349038

https://pkg.pr.new/software-mansion/TypeGPU/unplugin-typegpu@7be6355bed339b2b0b9027bcaf34879c45349038

benchmark
view benchmark

commit
view commit

[github-actions]

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

🟢 Decreased	➖ Unchanged	🔴 Increased (max 79.48%)	❔ Unknown
0	36	316	0

import * as ... in PR vs import * as ... in target (did bundle size increase?):

Click to reveal the results table (149 entries).

Test	tsdown
d_unstruct.ts	3.03 kB ($${\color{red}+79.5\%}$$)
tgpu_slot.ts	3.08 kB ($${\color{red}+77.3\%}$$)
tgpu_unroll.ts	3.76 kB ($${\color{red}+55.5\%}$$)
d_struct.ts	5.18 kB ($${\color{red}+35.0\%}$$)
tgpu_computeFn.ts	45.58 kB ($${\color{red}+0.9\%}$$)
tgpu_fn.ts	45.78 kB ($${\color{red}+0.9\%}$$)
tgpu_vertexFn.ts	46.13 kB ($${\color{red}+0.9\%}$$)
tgpu_fragmentFn.ts	46.32 kB ($${\color{red}+0.9\%}$$)
d_bool.ts	20.26 kB ($${\color{red}+0.9\%}$$)
d_f16.ts	20.26 kB ($${\color{red}+0.9\%}$$)
d_f32.ts	20.26 kB ($${\color{red}+0.9\%}$$)
d_i32.ts	20.26 kB ($${\color{red}+0.9\%}$$)
d_u32.ts	20.26 kB ($${\color{red}+0.9\%}$$)
d_ref.ts	20.27 kB ($${\color{red}+0.9\%}$$)
d_u16.ts	20.28 kB ($${\color{red}+0.9\%}$$)
d_abstractInt.ts	20.33 kB ($${\color{red}+0.9\%}$$)
d_abstractFloat.ts	20.34 kB ($${\color{red}+0.9\%}$$)
d_textureDepth2d.ts	20.70 kB ($${\color{red}+0.8\%}$$)
d_textureDepthCube.ts	20.71 kB ($${\color{red}+0.8\%}$$)
d_texture1d.ts	20.71 kB ($${\color{red}+0.8\%}$$)
d_texture2d.ts	20.71 kB ($${\color{red}+0.8\%}$$)
d_texture3d.ts	20.71 kB ($${\color{red}+0.8\%}$$)
d_textureCube.ts	20.72 kB ($${\color{red}+0.8\%}$$)
d_textureDepth2dArray.ts	20.72 kB ($${\color{red}+0.8\%}$$)
d_textureDepthCubeArray.ts	20.72 kB ($${\color{red}+0.8\%}$$)
d_textureDepthMultisampled2d.ts	20.73 kB ($${\color{red}+0.8\%}$$)
d_texture2dArray.ts	20.73 kB ($${\color{red}+0.8\%}$$)
d_textureCubeArray.ts	20.74 kB ($${\color{red}+0.8\%}$$)
d_textureMultisampled2d.ts	20.74 kB ($${\color{red}+0.8\%}$$)
common_fullScreenTriangle.ts	49.01 kB ($${\color{red}+0.8\%}$$)
std_discard.ts	21.51 kB ($${\color{red}+0.8\%}$$)
tgpu_comptime.ts	21.61 kB ($${\color{red}+0.8\%}$$)
std_arrayLength.ts	21.73 kB ($${\color{red}+0.8\%}$$)
std_copy.ts	21.75 kB ($${\color{red}+0.8\%}$$)
std_extensionEnabled.ts	21.83 kB ($${\color{red}+0.8\%}$$)
std_range.ts	22.10 kB ($${\color{red}+0.8\%}$$)
std_isBeingTranspiled.ts	22.14 kB ($${\color{red}+0.8\%}$$)
std_getShaderStage.ts	22.19 kB ($${\color{red}+0.8\%}$$)
std_getTargetShaderLanguage.ts	22.20 kB ($${\color{red}+0.8\%}$$)
std_typeOf.ts	22.25 kB ($${\color{red}+0.8\%}$$)
d_disarrayOf.ts	22.27 kB ($${\color{red}+0.8\%}$$)
STATIC_common.ts	56.56 kB ($${\color{red}+0.7\%}$$)
std_atomicLoad.ts	26.38 kB ($${\color{red}+0.7\%}$$)
std_atomicCompareExchangeWeak.ts	26.39 kB ($${\color{red}+0.7\%}$$)
std_atomicExchange.ts	26.39 kB ($${\color{red}+0.7\%}$$)
std_atomicStore.ts	26.39 kB ($${\color{red}+0.7\%}$$)
std_workgroupBarrier.ts	26.39 kB ($${\color{red}+0.7\%}$$)
std_workgroupUniformLoad.ts	26.39 kB ($${\color{red}+0.7\%}$$)
std_atomicAdd.ts	26.39 kB ($${\color{red}+0.7\%}$$)
std_atomicAnd.ts	26.39 kB ($${\color{red}+0.7\%}$$)
std_atomicMax.ts	26.39 kB ($${\color{red}+0.7\%}$$)
std_atomicMin.ts	26.39 kB ($${\color{red}+0.7\%}$$)
std_atomicOr.ts	26.39 kB ($${\color{red}+0.7\%}$$)
std_atomicSub.ts	26.39 kB ($${\color{red}+0.7\%}$$)
std_atomicXor.ts	26.39 kB ($${\color{red}+0.7\%}$$)
std_storageBarrier.ts	26.39 kB ($${\color{red}+0.7\%}$$)
std_textureBarrier.ts	26.39 kB ($${\color{red}+0.7\%}$$)
d_uint8.ts	28.21 kB ($${\color{red}+0.6\%}$$)
d_float16.ts	28.22 kB ($${\color{red}+0.6\%}$$)
d_float16x2.ts	28.22 kB ($${\color{red}+0.6\%}$$)
d_float16x4.ts	28.22 kB ($${\color{red}+0.6\%}$$)
d_float32.ts	28.22 kB ($${\color{red}+0.6\%}$$)
d_float32x2.ts	28.22 kB ($${\color{red}+0.6\%}$$)
d_float32x3.ts	28.22 kB ($${\color{red}+0.6\%}$$)
d_float32x4.ts	28.22 kB ($${\color{red}+0.6\%}$$)
d_sint16.ts	28.22 kB ($${\color{red}+0.6\%}$$)
d_sint16x2.ts	28.22 kB ($${\color{red}+0.6\%}$$)
d_sint16x4.ts	28.22 kB ($${\color{red}+0.6\%}$$)
d_sint32.ts	28.22 kB ($${\color{red}+0.6\%}$$)
d_sint32x2.ts	28.22 kB ($${\color{red}+0.6\%}$$)
d_sint32x3.ts	28.22 kB ($${\color{red}+0.6\%}$$)
d_sint32x4.ts	28.22 kB ($${\color{red}+0.6\%}$$)
d_sint8.ts	28.22 kB ($${\color{red}+0.6\%}$$)
d_sint8x2.ts	28.22 kB ($${\color{red}+0.6\%}$$)
d_sint8x4.ts	28.22 kB ($${\color{red}+0.6\%}$$)
d_snorm16.ts	28.22 kB ($${\color{red}+0.6\%}$$)
d_snorm16x2.ts	28.22 kB ($${\color{red}+0.6\%}$$)
d_snorm16x4.ts	28.22 kB ($${\color{red}+0.6\%}$$)
d_snorm8.ts	28.22 kB ($${\color{red}+0.6\%}$$)
d_snorm8x2.ts	28.22 kB ($${\color{red}+0.6\%}$$)
d_snorm8x4.ts	28.22 kB ($${\color{red}+0.6\%}$$)
d_uint16.ts	28.22 kB ($${\color{red}+0.6\%}$$)
d_uint16x2.ts	28.22 kB ($${\color{red}+0.6\%}$$)
d_uint16x4.ts	28.22 kB ($${\color{red}+0.6\%}$$)
d_uint32.ts	28.22 kB ($${\color{red}+0.6\%}$$)
d_uint32x2.ts	28.22 kB ($${\color{red}+0.6\%}$$)
d_uint32x3.ts	28.22 kB ($${\color{red}+0.6\%}$$)
d_uint32x4.ts	28.22 kB ($${\color{red}+0.6\%}$$)
d_uint8x2.ts	28.22 kB ($${\color{red}+0.6\%}$$)
d_uint8x4.ts	28.22 kB ($${\color{red}+0.6\%}$$)
d_unorm10_10_10_2.ts	28.22 kB ($${\color{red}+0.6\%}$$)
d_unorm16.ts	28.22 kB ($${\color{red}+0.6\%}$$)
d_unorm16x2.ts	28.22 kB ($${\color{red}+0.6\%}$$)
d_unorm16x4.ts	28.22 kB ($${\color{red}+0.6\%}$$)
d_unorm8.ts	28.22 kB ($${\color{red}+0.6\%}$$)
d_unorm8x2.ts	28.22 kB ($${\color{red}+0.6\%}$$)
d_unorm8x4.ts	28.22 kB ($${\color{red}+0.6\%}$$)
d_unorm8x4_bgra.ts	28.22 kB ($${\color{red}+0.6\%}$$)
d_formatToWGSLType.ts	28.21 kB ($${\color{red}+0.6\%}$$)
d_packedFormats.ts	28.24 kB ($${\color{red}+0.6\%}$$)
d_isPackedData.ts	28.28 kB ($${\color{red}+0.6\%}$$)
d_alignmentOf.ts	29.17 kB ($${\color{red}+0.6\%}$$)
tgpu_accessor.ts	68.34 kB ($${\color{red}+0.6\%}$$)
tgpu_mutableAccessor.ts	68.35 kB ($${\color{red}+0.6\%}$$)
d_isBuiltin.ts	31.89 kB ($${\color{red}+0.5\%}$$)
d_sizeOf.ts	31.94 kB ($${\color{red}+0.5\%}$$)
d_isContiguous.ts	31.95 kB ($${\color{red}+0.5\%}$$)
d_getLongestContiguousPrefix.ts	31.96 kB ($${\color{red}+0.5\%}$$)
std_subgroupAdd.ts	32.32 kB ($${\color{red}+0.5\%}$$)
std_quadBroadcast.ts	32.33 kB ($${\color{red}+0.5\%}$$)
std_quadSwapDiagonal.ts	32.33 kB ($${\color{red}+0.5\%}$$)
std_quadSwapX.ts	32.33 kB ($${\color{red}+0.5\%}$$)
std_quadSwapY.ts	32.33 kB ($${\color{red}+0.5\%}$$)
std_subgroupAll.ts	32.33 kB ($${\color{red}+0.5\%}$$)
std_subgroupAnd.ts	32.33 kB ($${\color{red}+0.5\%}$$)
std_subgroupAny.ts	32.33 kB ($${\color{red}+0.5\%}$$)
std_subgroupBallot.ts	32.33 kB ($${\color{red}+0.5\%}$$)
std_subgroupBroadcast.ts	32.33 kB ($${\color{red}+0.5\%}$$)
std_subgroupBroadcastFirst.ts	32.33 kB ($${\color{red}+0.5\%}$$)
std_subgroupElect.ts	32.33 kB ($${\color{red}+0.5\%}$$)
std_subgroupExclusiveAdd.ts	32.33 kB ($${\color{red}+0.5\%}$$)
std_subgroupExclusiveMul.ts	32.33 kB ($${\color{red}+0.5\%}$$)
std_subgroupInclusiveAdd.ts	32.33 kB ($${\color{red}+0.5\%}$$)
std_subgroupInclusiveMul.ts	32.33 kB ($${\color{red}+0.5\%}$$)
std_subgroupMax.ts	32.33 kB ($${\color{red}+0.5\%}$$)
std_subgroupMin.ts	32.33 kB ($${\color{red}+0.5\%}$$)
std_subgroupMul.ts	32.33 kB ($${\color{red}+0.5\%}$$)
std_subgroupOr.ts	32.33 kB ($${\color{red}+0.5\%}$$)
std_subgroupShuffle.ts	32.33 kB ($${\color{red}+0.5\%}$$)
std_subgroupShuffleDown.ts	32.33 kB ($${\color{red}+0.5\%}$$)
std_subgroupShuffleUp.ts	32.33 kB ($${\color{red}+0.5\%}$$)
std_subgroupShuffleXor.ts	32.33 kB ($${\color{red}+0.5\%}$$)
std_subgroupXor.ts	32.33 kB ($${\color{red}+0.5\%}$$)
d_size.ts	33.63 kB ($${\color{red}+0.5\%}$$)
d_align.ts	33.63 kB ($${\color{red}+0.5\%}$$)
d_location.ts	33.63 kB ($${\color{red}+0.5\%}$$)
d_interpolate.ts	33.64 kB ($${\color{red}+0.5\%}$$)
d_arrayOf.ts	33.69 kB ($${\color{red}+0.5\%}$$)
tgpu_vertexLayout.ts	33.95 kB ($${\color{red}+0.5\%}$$)
d_invariant.ts	34.01 kB ($${\color{red}+0.5\%}$$)
std_identity2.ts	34.42 kB ($${\color{red}+0.5\%}$$)
std_identity3.ts	34.42 kB ($${\color{red}+0.5\%}$$)
std_identity4.ts	34.42 kB ($${\color{red}+0.5\%}$$)
std_rotationX4.ts	34.42 kB ($${\color{red}+0.5\%}$$)
std_rotationY4.ts	34.42 kB ($${\color{red}+0.5\%}$$)
std_rotationZ4.ts	34.42 kB ($${\color{red}+0.5\%}$$)
std_scaling4.ts	34.42 kB ($${\color{red}+0.5\%}$$)
std_translation4.ts	34.42 kB ($${\color{red}+0.5\%}$$)
common_writeSoA.ts	34.55 kB ($${\color{red}+0.5\%}$$)

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

[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 [1.09, 2.21, 4.56, 7.42, 8.91, 13.62, 27.60, 27.69]
  line [1.14, 2.24, 4.96, 6.94, 9.29, 15.65, 27.57, 28.01]
  line [1.07, 2.25, 4.47, 7.52, 9.20, 12.87, 28.15, 27.23]

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 [0.31, 0.54, 0.75, 0.92, 1.24, 1.29, 1.53, 1.66]
  line [0.32, 0.54, 0.79, 0.98, 1.32, 1.38, 1.56, 1.76]
  line [0.34, 0.56, 0.77, 0.91, 1.23, 1.30, 1.57, 1.69]

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 [1.00, 2.96, 4.76, 8.55, 15.08, 28.88, 61.79, 125.31]
  line [0.93, 2.29, 4.38, 7.88, 14.49, 30.22, 63.34, 127.36]
  line [0.95, 2.35, 3.91, 7.91, 14.47, 28.57, 60.51, 126.17]

Loading

[vende11s]

@iwoplaza

Identifier params are plain strings ("a" instead of {"type":"i","name":"a"}). Destructured params are [name, alias][].

I'm introducing that in #2913

nodes/types look like that:

export type ObjectPattern = readonly [
  type: NodeTypeCatalog['objectPattern'],
  props: readonly [name: string, binding: BindingElement][],
];

export type BindingPattern = ObjectPattern;
export type BindingName = Identifier | BindingPattern;
export type BindingElement = BindingName;

export type FuncParameter = LegacyFuncParameter | BindingElement;

Destructured params are [name, binding] (and not [string, string]) because it leaves room for nested destructuring (not implemented yet)

It lets us add new kinds of destructuring, default parameters, and rest element without breaking changes in AST, finally it could look like that:

export type ObjectPatternEntry =
  | readonly [name: string, binding: BindingElement]
  | RestElement<Identifier>;

export type ObjectPattern = readonly [
  type: NodeTypeCatalog['objectPattern'],
  props: readonly ObjectPatternEntry[],
];

export type BindingPattern = ObjectPattern | ArrayPattern;
export type BindingName = Identifier | BindingPattern;
export type BindingElement = BindingName | AssignmentPattern;

export type FuncParameter = BindingElement | RestElement;

tldr

Should I include this change in #2913, or leave it for this PR?

[iwoplaza]

@vende11s Thanks! I think #2913 should focus on the destructuring params, and this PR will just extend it with the rest.