Repository navigation
ng1: expand resident DID state on the GPU - #731
Conversation
|
very interestiong. this is not parallelizable at all. but probably still a win to decode on gpu to reduce traffic. |
|
Thanks for taking a look! You’re right that decoding the runs within each scanline is serial in the current implementation. There is a bit of parallelism across scanlines, though: the shader assigns one GPU invocation to each row, so different rows can run at the same time. Each invocation still checks earlier line-table entries to respect the early terminator, so there’s some duplicated work there. The hope is that avoiding the 2 MiB decoded-map upload when the map changes outweighs that extra work. I haven’t measured the tradeoff yet, so I’ll benchmark unchanged frames, cursor-only updates, and DID changes before claiming it’s faster. If it matters, I’ll also look at avoiding unnecessary re-expansion on cursor-only changes. |
Summary
Scope and limitations
This extends the test-only resident foundation from #569; it does not enable a production Newport GPU mode. Ordinary resident frames do not upload or read back raw framebuffer planes or silently fall back to CPU raster. DID metadata readback is used explicitly by the parity tests.
Initial resident scanout rejects retained CPU DID history that cannot be reconstructed from current VC2 state. Moving that history between devices or back to a CPU renderer remains future lifecycle work. The slice adapts the earlier candidate in Tzeentchnet#4 and does not depend on #698.
Validation
On Windows with an NVIDIA RTX 5070, using an isolated worktree with a refreshed ignored lockfile:
The full resident gate passed 63 tests, including the required Vulkan and D3D12 hardware runs. The existing main-worktree lockfile was left unchanged.