cpu: move draw resource resolution off the committer, from 25 to 10-15 µs per draw #4

Problem

With the recommended launch profile (APS5_NO_DRAW_RELOCATION=1, APS5_PARALLEL_DRAW_PREPARE=2) the frame (about 74 ms on the test machine) is set by the single committer thread: about 2,050 draws per frame at about 25 µs each, roughly 51 ms of work. 99 % of draws are template hits, yet the committer still resolves their whole bound state from guest memory, under the GPU lock, every time.

Where the 25 µs go on the template-hit path (APS5_PROFILE_DRAW), in µs per draw:

Details are in docs/research/cpu-chain.md.

What was tried

• A validity token per template (tracker generation, collection epoch, pending serial, storage-cache serial). It was correct, but worth only about −0.8 ms per frame. The remaining cost of the proof is concentrated in about 28 % of the proofs, at about 12.5 µs each, so it was archived.
• Lock-free per-texture validity. Inside the fast proof, CollectWrites is about 37 %. The ceiling of this piece is 3-4 % of the frame, so it was not prototyped.

The conclusion: small memos on the committer are not enough. The work has to leave the committer.

Direction being explored

The direction follows the command-stream model of DXVK, with ideas from vkd3d-proton and shadPS4:

• resolve a draw's resources on the draw preparation threads, outside the GPU lock, speculatively;
• validate the speculation at commit with cheap stamps;
• keep only recording under the lock.

A probe that counts what such a speculation would have kept, without changing behaviour, says it holds for 99.46 % of the draws in the reference scene (0.54 % fall back). The estimated committer cost after the step is about 13.5 µs per draw.

There are three prototypes, each behind its own switch, with a verification mode and tests with mutations. The first is in progress:

1. bindings;
2. vertex inputs, validation and pipeline;
3. the template proof.

Help wanted

• Review of the approach: what can go stale between preparation and commit (CPU writes, GPU writes, labels, pending imports), and which stamps catch it.
• Measurements on other CPUs: the split of the 25 µs may differ on AMD CPUs or CPUs with more cache. Use the Performance report form.
• Descriptors: pushing per-draw descriptors (maxPushDescriptors 32) or a descriptor heap instead of a new set per draw.

Related upstream work

• perf(agc): GPU descriptor heap (APS5_GPU_HEAP) boykopovar/AnyPS5#6947 (GPU descriptor heap) and feat(agc): look up table images at run time through a device image table on the GPU heap (APS5_GPU_PATH) boykopovar/AnyPS5#7642 (device image table on the GPU heap) go in the descriptor direction.
• perf(agc): read a sampled texture's DCC keys once per lookup boykopovar/AnyPS5#8117 reads a sampled texture's DCC keys once per lookup, which touches the fast proof.