WebGPU 1.0: Query Result Timing for Indirect Draw Pagination

Answer

According to the WebGPU 1.0 specification, a query result buffer can be read only after all GPU work that writes to that buffer has finished. This means the application must submit the command encoder that recorded the render pass and then wait for the queue to become idle before calling mapReadAsync on the query buffer.

Confirmed facts from the spec

• The spec defines that a GPUQuerySet result becomes available when the corresponding GPUCommandBuffer that contains the query has completed execution.
• Mapping a buffer (mapReadAsync) is allowed only when the buffer is not in use by the GPU; otherwise the call is rejected or yields undefined data.
• Therefore a separate queue submission (or an explicit idle wait) is required after the render pass ends.

Likely explanation of observed browser differences

Some browsers internally flush the command buffer when a render pass ends, making the query result appear immediately mappable. This behavior is not guaranteed by the spec and relying on it can lead to intermittent failures or stale data on other implementations.

Recommended pattern for pagination

1. Create a GPUQuerySet with one occlusion query and a matching GPUBuffer for the result.
2. Begin a render pass, set the occlusion query, draw the scene (or bounding geometry), end the pass.
3. Finish encoding, submit the command encoder via GPUQueue.submit.
4. Wait for the queue to idle: await queue.onSubmittedWorkDone() (or queue.waitForIdle()).
5. Call queryBuffer.mapReadAsync, await the promise, then copy the mapped value into a readable ArrayBuffer (or use getMappedRange).
6. Encode a second command encoder that contains an indirect draw pass. Set the indirect buffer’s offset/size based on the read count, then submit it.
7. Repeat as needed for further pages.

Verification checklist

• Submit the first pass, then attempt to map the query buffer before waiting for idle – the call should reject or return stale data.
• After waiting for idle, mapping succeeds and the returned count matches the number of primitives that passed the occlusion test.
• Using that count in an indirect draw renders only the visible primitives.