wgpu issue: https://github.com/gfx-rs/wgpu/issues/2431

See also:

• https://github.com/gpuweb/gpuweb/issues/31
• https://github.com/gpuweb/gpuweb/issues/431
• https://github.com/gpuweb/gpuweb/issues/2189
• https://github.com/gpuweb/gpuweb/issues/257

Quoting Kvark:

Here is a rough plan:

    wgpu-hal expects platform-dependent draw/dispatch buffer.
    the only thing exposed to the outside is the size of that buffer (for each kind of draw/draw indexed/dispatch)
    wgpu-hal command encoder gets a new fn prepare_indirect(src_buffer, src_offset, dst_buffer, dst_offset, kind, count). Under the hood it will be a compute operation that does clamping as well as copies of first instance/vertex.
    wgpu-core spawns a separate encoder to prepare indirect calls whenever it sees one, and orders it accordingly. It also manages the destination buffer as a temporary thing.

To expand, if I understand correctly, What I'm planning to do is to:

• have wgpu-core store some proxy indirect buffers (somewhere, maybe on the device). These will be used as input for the draw indirect commands instead of what is in the command.
• in command_encoder_run_render_pass_impl, when running into indirect commands, we allocate ranges into the proxy buffers and use them in the render pass instead of the original buffer/range, and record that we need to sanitize the content of the original buffer at its requested offset into the new range of the proxy buffer.
• After we are done with the render pass, we see if we have recorded any ranges to sanitize, if so we create a compute pass that will be submitted just before the render pass. It will dispatch a simple compute shader that reads from the indirect buffer/ranges and write the sanitized output into the proxy buffer.

Validating into a separate buffer instead of in-place is so that the validation cannot be observed by other means than the indrect draw itself.

Look like Dawn does more or less what Kvark described: https://source.chromium.org/chromium/chromium/src/+/main:third_party/dawn/src/dawn/native/IndirectDrawValidationEncoder.cpp;l=238;drc=aaea55c708c63d53a89fb525484aa94747599714;bpv=1;bpt=1

Note that special care goes into batching validation of the everything in the same buffer into one dispatch of the validation shader. There's a scratch buffer helper that is allocated/resized on demand and the validation shader is lazily created.