The function "dispatch_workgroups_indirect" does not work

[panxinmiao]

Description

I tried to update the wgpu-native to the latest wgpu-core, but the test case of dispatch_workgroups_indirect in the downstream wgpu-py can't pass.

Then I tried to modify this example: "https://github.com/gfx-rs/wgpu/blob/master/wgpu/examples/hello-compute/main.rs" , create an indirect buffer and replace dispatch_workgroups with dispatch_workgroups_indirect for testing, the result is not correct also.

Repro steps

modify this example: "https://github.com/gfx-rs/wgpu/blob/master/wgpu/examples/hello-compute/main.rs", use dispatch_workgroups_indirect instead.

Expected vs observed behavior

The output value is always equal to the input. No computes seem to have been made.

Platform

wgpu master branch. Windows10, vulkan backend

[cwfitzgerald]

I can't seem to get this to reproduce on metal, what code did you use, I used

    let indirect_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
        label: None,
        contents: bytemuck::cast_slice(&[numbers.len() as u32, 1, 1]),
        usage: wgpu::BufferUsages::INDIRECT,
    });

    // A command encoder executes one or many pipelines.
    // It is to WebGPU what a command buffer is to Vulkan.
    let mut encoder =
        device.create_command_encoder(&wgpu::CommandEncoderDescriptor { label: None });
    {
        let mut cpass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor { label: None });
        cpass.set_pipeline(&compute_pipeline);
        cpass.set_bind_group(0, &bind_group, &[]);
        cpass.insert_debug_marker("compute collatz iterations");
        cpass.dispatch_workgroups_indirect(&indirect_buffer, 0); // Number of cells to run, the (x,y,z) size of item being processed
    }

[panxinmiao]

My code is similar to yours,
The results are as follows:

[cwfitzgerald]

I've got this reproducing on dx12 with a validation error about bad barriers - I know the problem here.