Thanks @RafaelCintron for all the explanation on D3D12 optimizations.

In the call yesterday you said you wanted to think more about bundles. I'd like to clarify again what the different uses for bundles are:

• Parallel recording of the inside of a render pass. D3D12 can do this using command lists because (until recently) there was no constraints on commands for being inside or outside a render pass. In WebGPU we need to have different types because you can either record something "at the top level" or "inside a render pass". Recording the inside of render pass in parallel is important because for many application they will have "main" render pass that contains thousands of objects and will dwarf the rest of the encoding work.
• Reordering inside a render pass. I don't have an example at hand for render bundles, but Babylon.js found it very useful to be able to record two command buffers in a single traversal of their scenegraph. They encode both upload commands to update uniforms, and the actual draw commands, then at submission time they just need to order uploads before draws. With reordering they'd need to do two passes over their scenegraph. Bundle doing the same type of thing with the inside of a render pass.
• Reuse of parts of the commands inside a render pass. Many applications will have part of their scene stay similar across frames. Reusing bundles allows saving the CPU cost of walking through the scenegraph subtree, encoding the commands and validating them.

My understanding is you have concerns that WebGPU implementation using bundles will have to use the DESCRIPTORS_VOLATILE flag. That's likely necessary in all cases because we can't guarantee that the content of uniform buffers will not change so we don't want the D3D driver to inline part of it in the root signature.

However, the D3D team's experience so far suggests they are rarely a performance win, and in many cases a pessimisation. Their programming model is too restricted to map well to typical app usage patterns, and they inhibit optimizations that drivers would otherwise perform.

Dawn made an implementation choice to always replay commands from software command buffer, but some other implementation can choose to back WebGPU bundles with D3D12 bundles, at which point there is the same performance tradeoff that you mentioned in your comment.

WDYT? Does it make you more comfortable with this PR?

@beaufortfrancois executeBundles isn't like setPipeline or setBindGroup because it essentially inlines all the commands in the bundles in particular draw or dispatch.

Discussed at 8 July 2019 teleconference

Merging as discussed in yesterday's call.