Windows: whole-surface compares of imported render targets dominate GTA V's draw path (measurement) #4432

docs/dev/TechnicalDebt.md says that on Windows, host-pointer imports of the guest mapping are compared instead of write-watched, and that "measuring the comparison cost" is not done yet. Here is a measurement from one title, where this cost is most of the frame.

Setup. GTA V (PPSA04264), Windows 11, RTX 4090 (NVIDIA), WinLibs GCC 15.2.0, Release, write watch on. Tree: main 325e0a8 plus the PRs the title needs to reach Story, plus #2802. The scene is the first minutes of the Story prologue.

Where the time goes. I added local timers inside StorageTexture::Refresh (summed over all threads, per 10 s):

• The draws bind resident render targets (3840x2160 and 1920x1080). Their memory is imported, so GuestMemory::ChangedBlocks reports it untracked and Refresh ends in compareUntracked over the whole surface.
• Without perf(agc): memoize untracked texture comparisons by the stores that reach the surface #2802: 11k to 15k refreshes, 8.3 to 8.6 s of whole-surface compare, 0.2 to 0.6 FPS.
• With perf(agc): memoize untracked texture comparisons by the stores that reach the surface #2802: 26k to 36k refreshes, about 7 s of compare, 1.1 to 1.35 FPS. The compare is still the largest cost by far.

Why the memo still misses with #2802. Out of 420k compareUntracked calls:

So almost every remaining compare is the one that guards against a CPU store into unwatched memory at a new ordering point. Splitting the compare into 64 KiB units would not change that, because at a new epoch every unit is a candidate. In this scene the compared surfaces were found unchanged about 98% of the time.

I don't have a proposal for how to keep these imports watched on Windows; the TechnicalDebt entry explains why the driver can dirty imported pages after the import window. I'm posting this so the cost is on record for whoever picks it up. Happy to run a branch on this title and report numbers.

AI-assisted (Claude Code).