burn
burn copied to clipboard
tracel-ai
Crash on Chrome WebGPU for kernels that bind with aliasing.
Describe the bug
If some operations use a tensor multiple times, Chrome can crash complaining it's a violation of the WebGPU specification to bind aliased buffers.
To Reproduce
let tensor = Tensor::<Backend, 2>::zeros([2, 1], &device);
let test: Tensor<Backend, 2> = tensor.clone().mul(tensor).sum_dim(1); // fails!
// for some reason works if tensor is 1D? Or if not summing / presumable some other operation?
// It might be slightly more involved than I thought!
// The following does work:
// let test: Tensor<Backend, 2> = tensor.clone().powf_scalar(2.0).sum_dim(1);
Run this code on WebGPU on Chrome (firefox nightly seems fine with it - though I've had a host of other issues on firefox so generally it seems less complete).
The output looks like:
Writable storage buffer binding aliasing found between [BindGroup (unlabeled)] set at bind group index 0, binding index 0, and [BindGroup (unlabeled)] set at bind group index 0, binding index 1, with overlapping ranges (offset: 0, size: 8) and (offset: 0, size: 8) in [Buffer (unlabeled)].
- While encoding [ComputePassEncoder (unlabeled)].DispatchWorkgroups(1, 1, 1).
127.0.0.1/:1 [Invalid CommandBuffer "Command Encoder" from CommandEncoder "Command Encoder"] is invalid.
- While calling [Queue].Submit([[Invalid CommandBuffer "Command Encoder" from CommandEncoder "Command Encoder"]])
Expected behavior Don't crash :) How to achieve that seems trickier. You'd need different WGSL versions of the kernel depending on whether inputs/outputs are the same or not. CubeCL might be able to handle that, but I'm not entirey sure what 's best here!
Not sure if there is an option to disable that check, if not we will have to compile more kernels automatically based on handle ids I guess.
