[reczkok]

(FYI this is not the feature discussed in #2816 proposal, this is the actual WebGPU feature)
d.bytes(n) resolves to the WGSL buffer<n> type, and std.bufferView, std.bufferArrayView and std.bufferLength read and write its bytes as any host-shareable type at runtime offsets

[github-actions]

pkg.pr.new

packages
Ready to be installed by your favorite package manager ⬇️

https://pkg.pr.new/software-mansion/TypeGPU/typegpu@45d9dd2e0ad25b3e4227b15809c11a13281a7ccf

https://pkg.pr.new/software-mansion/TypeGPU/@typegpu/gl@45d9dd2e0ad25b3e4227b15809c11a13281a7ccf

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 [0.76, 1.46, 3.14, 4.74, 5.58, 9.73, 19.60, 17.48]
  line [0.77, 1.44, 2.92, 5.28, 5.97, 8.80, 19.46, 19.88]
  line [0.77, 1.50, 3.17, 4.98, 6.06, 9.61, 18.64, 16.43]

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 [0.21, 0.38, 0.55, 0.65, 0.91, 1.01, 1.20, 1.37]
  line [0.24, 0.42, 0.59, 0.70, 0.88, 0.90, 1.07, 1.25]
  line [0.26, 0.42, 0.61, 0.69, 0.92, 0.98, 1.16, 1.26]

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 [0.79, 1.72, 2.75, 5.36, 10.26, 20.65, 42.67, 89.29]
  line [0.74, 1.67, 2.67, 5.22, 9.41, 19.85, 43.31, 87.44]
  line [0.61, 1.69, 3.09, 6.00, 9.73, 20.29, 42.88, 89.77]

Loading

[github-actions]

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

🟢 Decreased	➖ Unchanged	🔴 Increased (max 12.84%)	❔ Unknown
0	312	43	4

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

Test	tsdown
STATIC_std.ts	114.58 kB ($${\color{red}+12.8\%}$$)
d_struct.ts	4.01 kB ($${\color{red}+4.7\%}$$)
STATIC_allImports.ts	319.70 kB ($${\color{red}+1.4\%}$$)
STATIC_d.ts	72.00 kB ($${\color{red}+1.0\%}$$)
tgpu_resolveWithContext.ts	161.78 kB ($${\color{red}+0.9\%}$$)
tgpu_resolve.ts	161.84 kB ($${\color{red}+0.9\%}$$)
d_arrayOf.ts	33.72 kB ($${\color{red}+0.7\%}$$)
d_builtin.ts	36.19 kB ($${\color{red}+0.7\%}$$)
tgpu_initFromDevice.ts	256.84 kB ($${\color{red}+0.7\%}$$)
tgpu_init.ts	257.38 kB ($${\color{red}+0.7\%}$$)
STATIC_tgpu.ts	268.61 kB ($${\color{red}+0.7\%}$$)
common_fullScreenTriangle.ts	48.82 kB ($${\color{red}+0.5\%}$$)
d_bytes.ts	1011 B
std_bufferArrayView.ts	40.15 kB
std_bufferLength.ts	41.29 kB
std_bufferView.ts	40.15 kB

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

[reczkok]

@pullfrog do a preliminary review

[pullfrog]

ℹ️ No critical issues — one minor validation gap worth a look.

Reviewed changes

• d.bytes(n) schema (data/bytes.ts) — new WgslBuffer schema resolving to WGSL buffer<n> (or runtime-sized buffer for 0), with size/alignment/equality/layout/serialization wiring and struct/array/address-space guards.
• std.bufferView / std.bufferArrayView / std.bufferLength (std/bufferView.ts) — codegen-only builtins that reinterpret buffer bytes as host-shareable types, with alignment, fit, atomic, host-shareable and scope validation, plus std exports.
• buffer_view language-feature plumbing (resolutionCtx.ts, wgslExtensions.ts, resolveData.ts, types.ts) — requireLanguageFeature and a requires buffer_view; directive emitted ahead of the module.
• For-of over pointer iterables (tgsl/forOfUtils.ts) — getRangeSnippets now unwraps pointer data types (enables iterating a bufferView-returned array).
• Docs + example + tests — new buffer-views.mdx, the shape-stream example, and bytes/bufferView/serialization/generator tests.

I read the full diff, verified the relevant source paths (dualImpl, pinEref, ptr, array, resolveData, forOfUtils, dataIO, compiledIO, schemaMemoryLayout, schemasEqual), ran the new tests (bytes, std/bufferView, serializeDataSchema), the tgsl suite, and tsc --noEmit — all pass. I also checked the claims against the WGSL spec: the requires buffer_view; directive/feature name is correct, directive ordering relative to enable is unconstrained, and the alignment/MinTypeSize/invalid-pointer semantics line up with assertViewFits/minSizeOf.

The buffer<N> size rule is the one gap I found; detail inline.

  ^{｜ Fix all ➔ ｜ Fix 👍s ➔ ｜ View workflow run ｜ Using deepseek-v4.1-flash (free via Pullfrog for OSS) ｜ 𝕏}

[pullfrog]

Suggested change

	export function bytes(byteSize = 0): WgslBuffer {
	if (!Number.isInteger(byteSize) || byteSize < 0) {
	export function bytes(byteSize = 0): WgslBuffer {
	if (!Number.isInteger(byteSize) || byteSize < 0) {
	throw new Error(`Cannot create a buffer schema of ${byteSize} bytes.`);
	}
	if (byteSize !== 0 && byteSize % 2 !== 0) {
	throw new Error(
	`Cannot create a buffer schema of ${byteSize} bytes, WGSL requires a buffer size to be a multiple of 2.`,
	);
	}",

[pullfrog]

Disregard the auto-suggestion above — its 2-line range was wrong and would leave a duplicate throw behind. The intended change is simply to reject sizes WGSL can't represent: after the existing integer/negative check in bytes(), add if (byteSize !== 0 && byteSize % 2 !== 0) throw new Error(...), and ideally a resolution-time % 4 check when f16 is disabled (WGSL allows buffer<N> only for N a multiple of 4, or of 2 when f16 is enabled; d.bytes(6) currently emits buffer<6>, a shader-creation error).

  ^{｜ View workflow run ｜ via Pullfrog ｜ Using deepseek-v4.1-flash (free via Pullfrog for OSS) ｜ 𝕏}