impr: Utilize subgroups in MNIST Inference when possible

[reczkok]

Requires chrome 134

EDIT: It is now available by default in chrome and pretty fast. Those are the measured times but I would't take them at face value - some query sets for subgroup version resulted in 0ns inference time which probably means it's below the precision of timestamp queries.

	No Subgroups	Subgroups
Measured Time

EDIT2: Since we now enable timing utilities in TypeGPU timing could become the default in this example

[github-actions]

pkg.pr.new

packages

pnpm i https://pkg.pr.new/software-mansion/TypeGPU/typegpu@986
pnpm i https://pkg.pr.new/software-mansion/TypeGPU/typegpu@b7d4c5809428f9c01d781055e3110e00ca480a54

benchmark
view benchmark

commit
view commit

[Copilot]

Pull Request Overview

This pull request improves the MNIST inference example by leveraging GPU subgroup operations when available, which should enhance performance.

• The initialization code now detects subgroup support using navigator.gpu.
• A new subgroup compute shader is introduced alongside a fallback shader for devices that do not support subgroups.
• The compute pipeline dynamically selects the appropriate shader based on the detected capabilities.

Comments suppressed due to low confidence (2)

apps/typegpu-docs/src/content/examples/algorithms/mnist-inference/index.ts:23

• [nitpick] Consider renaming 'fallbackShader' to something like 'nonSubgroupShader' to more explicitly indicate its purpose.

const fallbackShader = `

apps/typegpu-docs/src/content/examples/algorithms/mnist-inference/index.ts:66

• The workgroup size of 64 is assumed to be divisible by the subgroup size (ssize). Consider adding a static assertion or additional documentation to ensure this invariant holds on all target devices.

var<workgroup> subgroupSums: array<f32, 64 / 4>;