WebGPU Hello World

This sample is the first bring-up checkpoint for the WebGPU target. It keeps the scope intentionally small: compile one tensor program to a WebGPU VMFB, dump the compiler-generated WGSL, and ask Dawn to validate that WGSL as a compute pipeline.

The sample does not use samples/simple_embedding. That sample exercises a synchronous embedding flow with blocking host transfers, which is the wrong shape for the current inline WebGPU host because JavaScript promises cannot settle while the wasm thread is blocked in C.

Build the VMFB

Enable WebGPU SPIR-V compiler support in the Bazel configuration before building the sample:

IREE_TARGET_BACKEND_WEBGPU_SPIRV=ON build_tools/bin/iree-bazel-configure

build_tools/bin/iree-bazel-build \
  //samples/webgpu/hello_world:hello_world_bytecode_module_webgpu

Dump and Validate WGSL

The validator uses the webgpu Node package. Install it outside of the checkout so npm does not create package.json, package-lock.json, or node_modules/ in the repository root:

npm install --prefix /tmp/iree-webgpu-validator webgpu

mkdir -p /tmp/iree-webgpu-hello-world

bazel-bin/tools/iree-compile \
  samples/webgpu/hello_world/hello_world.mlir \
  --iree-hal-target-device=webgpu \
  --iree-hal-dump-executable-binaries-to=/tmp/iree-webgpu-hello-world \
  -o=/tmp/iree-webgpu-hello-world/hello_world.vmfb

IREE_WEBGPU_PACKAGE_ROOT=/tmp/iree-webgpu-validator \
  node samples/webgpu/hello_world/validate_wgsl.mjs \
  /tmp/iree-webgpu-hello-world/*.wgsl