What's New in WebGPU (Chrome 128)

François Beaufort
François Beaufort
GitHub

Experimenting with subgroups

The subgroups feature enables SIMD-level parallelism, allowing threads within a group to communicate and perform collective math operations (for example, calculating the sum of 16 numbers). This provides a highly efficient form of cross-thread data sharing.

A minimal implementation of the subgroups proposal is available for local testing behind the "Unsafe WebGPU Support" flag at chrome://flags/#enable-unsafe-webgpu.

You can also try subgroups on your site with real users by signing up for the origin trial. Read Get started with origin trials for instructions on how to prepare your site to use origin trials. The origin trial will run from Chrome 128 to 131 (ending February 19, 2025). See Intent to Experiment.

When the "subgroups" feature is available in a GPUAdapter, request a GPUDevice with this feature to get subgroups support in WGSL and check its minSubgroupSize and maxSubgroupSize limits.

You also need to explicitly enable this extension in your WGSL code with enable subgroups;. When enabled, you get access to the following additions:

  • subgroup_invocation_id: A built-in value for the index of the thread within the subgroup.
  • subgroup_size: A built-in value for subgroup size access.
  • subgroupBallot(value): Returns a set of bit fields where the bit corresponding to subgroup_invocation_id is 1 if value is true for that active invocation and 0 otherwise.
  • subgroupBroadcast(value, id): Broadcasts the value from the invocation with subgroup_invocation_id matching id to all invocations within the subgroup. Note: id must be a compile-time constant.

More built-in functions such as subgroupAdd, subgroupAll, subgroupElect, subgroupShuffle will be added in the future. See issue 354738715.

To allow f16 in subgroups operations, request a GPUDevice with the "subgroups", "subgroups-f16", and "shader-f16" features, then enable it in your WGSL code with enable f16, subgroups, subgroups_f16;.

The following code snippet provides a base to tinker with and discover the potential of subgroups.

const adapter = await navigator.gpu.requestAdapter();
if (!adapter.features.has("subgroups")) {
  throw new Error("Subgroups support is not available");
}
// Explicitly request subgroups support.
const device = await adapter.requestDevice({
  requiredFeatures: ["subgroups"],
});

const shaderModule = device.createShaderModule({ code: `
  enable subgroups;

  var<workgroup> wgmem : u32;

  @group(0) @binding(0)
  var<storage, read> inputs : array<u32>;

  @group(0) @binding(1)
  var<storage, read_write> output : array<u32>;

  @compute @workgroup_size(64)
  fn main(@builtin(subgroup_size) subgroupSize : u32,
          @builtin(subgroup_invocation_id) id : u32,
          @builtin(local_invocation_index) lid : u32) {
    // One thread per workgroup writes the value to workgroup memory.
    if (lid == 0) {
      wgmem = inputs[lid];
    }
    workgroupBarrier();
    var v = 0u;

    // One thread per subgroup reads the value from workgroup memory
    // and shares that value with every other thread in the subgroup
    // to reduce local memory bandwidth.
    if (id == 0) {
      v = wgmem;
    }
    v = subgroupBroadcast(v, 0);
    output[lid] = v;
  }`,
});

// Send the appropriate commands to the GPU...

Deprecate setting depth bias for lines and points

A WebGPU spec change makes it a validation error to set depthBias, depthBiasSlopeScale, and depthBiasClamp to a non-zero value when the topology for a render pipeline is a line or point type. To give developers enough time to update their code, a warning in the DevTools Console is shown about this upcoming validation while also forcing the values to 0 in these circumstances. See issue 352567424.

Hide uncaptured error DevTools warning if preventDefault

In the DevTools Console, warnings for uncapturederror events are no longer displayed if an event listener for uncapturederror has been registered and the Event preventDefault() method has been called within the event listener callback. This behaviour matches event handling in JavaScript. See the following example and issue 40263619.

const adapter = await navigator.gpu.requestAdapter();
const device = await adapter.requestDevice();

device.addEventListener("uncapturederror", (event) => {
  // Prevents browser warning to show up in the DevTools Console.
  event.preventDefault();

  // TODO: Handle event.error
});

WGSL interpolate sampling first and either

WGSL interpolate attribute lets you manage user-defined IO data interpolation. Now, new interpolate sampling parameters first (default) and either give you additional control: first uses the value from the primitive's first vertex, while either allows either the first or last vertex. See issue 340278447.

Dawn updates

The implementation of Dawn's WGPUFuture to handle asynchronous operations is now complete. Key concepts include wgpuInstanceProcessEvents for opportunistic event processing and WGPUCallbackMode for defining callback locations. WGPUFuture signifies one-time events with an infinite lifetime, and wgpuInstanceWaitAny awaits completion of any future or a timeout. See issue 42240932.

The CompositeAlphaMode::Auto value is now not reported by Surface::GetCapabilities(). It's still valid, and equivalent to Surface::GetCapabilities().alphaMode[0]. See issue 292.

The OpenGL backend now supports Surface with a y-flip blit for each Present() call. See issue 344814083.

The Adapter::GetProperties() method is deprecated in favor of using Adapter::GetInfo().

Jaswant, an external contributor, has rewritten all the CMake files, making them easier to update and allowing for pre-builds. Check out the quickstart guide for using Dawn in CMake projects.

This covers only some of the key highlights. Check out the exhaustive list of commits.

What's New in WebGPU

A list of everything that has been covered in the What's New in WebGPU series.

Chrome 131

Chrome 130

Chrome 129

Chrome 128

Chrome 127

Chrome 126

Chrome 125

Chrome 124

Chrome 123

Chrome 122

Chrome 121

Chrome 120

Chrome 119

Chrome 118

Chrome 117

Chrome 116

Chrome 115

Chrome 114

Chrome 113