[xTektor]

Changes

Twelve commits. Commits 1–4 are unchanged and 4 is now measured. Commits 5–12 are new and unmeasured. No FPS claims for them.

1. Native compute write marks (a47c72f3b0)

A Switch 2 compute storage pointer has no size, so a written binding was marked GPU-modified for up to 8 MiB. Ocarina skinning writes tens of kilobytes; the rest of that range is the CPU's own data. Fences copied it back, and the CPU's later writes waited.

The recompiler now proves how far each store reaches. A dispatch marks only those bytes. An unproven store still marks the whole binding. Reads are unchanged.

Legion result (MSVC 5fb53b44, Vulkan, docked, 1x, GPU accuracy Low, no DLSS, no Dynamic FPS): buffer downloads are 0 KiB in every gameplay window. Outdoor Kokiri was still about 10–16 fps. The download path is no longer the limiter.

2. Occlusion-report DMA reads (88f1b3d9ae)

The front was draining the whole back thread on every macro DMA read, about 27,000–30,000 reads a second. Those are occlusion reports. A read of a landed report, on a page shaders and copies have not written, now comes from guest memory. Anything else still drains.

Legion result (MSVC 88f1b3d9, same settings, two outdoor sessions including a warm cache): macro DMA read wait fell from about 270–373 ms/s to about 20–100 ms/s. Warm village under trees is typically 20–28 fps, where the previous build was 10–14. Walking is often still 12–14. Camera turns are typically 20–30. Overall outdoor is about 12–25, overlay average 15. Stalls to 0 fps remain. Buffer downloads stayed 0. The canopy haze is still there. Not playable 30.

Split drain stayed high (session 2 often 250–580 ms/s) because acquire waits did not move (160–490 ms/s, mean about 280).

3. CPU semaphore acquires (3863d03ef9)

zortos293#337 is not in this branch. CounterReport still treats a fence-flagged report as synchronized at GPU accuracy Low with Default fences, and the fence manager can still skip an empty fence. The acquire cost in the 88f1b3d9 profile is not the host-GPU wait zortos293#874 removes from zortos293#337. It is DrainBack on every acquire check, about 1,600 blocking drains a second, including polls of a semaphore the CPU releases.

zortos293#874 therefore does not apply here. zortos293#805 and zortos293#717 were not ported.

A read now waits only until a queued GPU release of that word has been executed. When none is queued, the host write has landed and the buffer cache is clean, the word is taken from guest memory and the back is not drained. A value that already matches is not trusted while a release of that word is still queued, which is the stale >= case from zortos293#231. A dirty buffer cache or an unfinished host query still drains and flushes.

Query ordering is unchanged, so this does not reintroduce the white flashes zortos293#337 fixed on trees that have that commit. A bug here would be a channel running past an acquire whose GPU release had not run.

Legion result (user play around Kokiri, same settings, window title 3863d03ef9): acquire site wait mean 93.6 ms/s versus 278.6 on 88f1b3d9. About 2,214 acquires/s, of which about 6.5% waited. Split drain mean 214.9 ms/s versus 412. Outdoor overlay on the verified frames is typically 19–24 fps (one 10 fps village sample, one 0 fps stall). Not stable 30. Canopy haze remains. Buffer downloads stayed 0.

What did not fall: writer wait mean 370 ms/s, list writers 276 ms/s, list GPU wait 199 ms/s, fence release 115 ms/s.

4. Command-list copy no longer blocks the front (ec8e93f750)

Those writer and list-GPU waits are the front standing in WaitSubmittedTick until the GPU finishes the copy of a command list the channel is about to read, after WaitForWriters. Upstream zortos293#752 measured the same pattern in Kokiri Forest: about 2,400 such waits a second, about a quarter of the time, with the back idle and more work already queued. This is that change, not the rest of zortos293#805.

A copy that is not finished suspends that channel, the same way a semaphore acquire does. The scheduler runs the other channels and the queued work. The channel parses the entry again once the copy's tick is done. The list is still read only after the GPU wrote it. Resuming skips the writer wait, because that wait already ran and waiting again would also wait for other channels' work queued during the suspension. If the step returns before that read, the skip is cleared.

No fps number for this commit when it was pushed.

Legion result for ec8e93f750 (outdoor Kokiri, same settings): 12–26 fps. Profile per second: front writer wait ~378 ms/s, list writers ~284, split drain ~232, pending sleep ~193, resume pending ~103, acquire ~93, macro DMA read ~102, buffer downloads 0, host_cpu_busy ~336. The old list GPU wait counter is gone, since the front no longer stands in the copy wait. Writer wait and list writers did not fall, which led to commit 5.

5. Command entries between methods skip the compute-writer wait (62a5a42b3)

Finishes the unpushed slice of the Cursor agent. The front ran WaitForWriters at every command-list entry. Now it does that only when the entry starts a new method (method_count == 0 skips it) or when a DMA copy writer is still pending (CommandEntryWaitsForWriters). The pending copy writer is tracked by NoteCopyWriter. A command-list read waits only for writers up to the last sync (WaitForSyncedWriters, a new list copy writers drain site). The GPU thread also publishes before the front sleeps on an empty queue. Credit: jpkto, and zortos293#805 (d4e8d85d4d4f). The -mtls-dialect=gnu2 change was left out. Unmeasured.

6. Transcode cache and dead-data downloads (6dad7ae89)

From javocsoft#17 (edab3cc934). Transcoded textures are kept in host memory, and the GC skips downloads of GPU data nothing will read again. The log reports with dead data deletes and the transcode cache stats. Unmeasured.

7. Back macros, opt-in (82713612c)

From Frizti#24 by mkooner (2f82fd2d5755, a port of jpkto a442c38fe6). Macros that read memory run on the split's back thread, and AnalyzeEffects proves which registers they write. Off unless EDEN_NX2_BACK_MACROS=1. Unmeasured.

8. OoT zstd on the host (a6fe18b57)

From zortos293#869 by Xiphias (4e3142e4f912). Ocarina of Time's zstd decompression, streamed blocks included, runs on the host via an SVC hook. It is gated on the main build ID E0F3CCFA… (matches the user's log) and the expected original code. Otherwise it does nothing. EDEN_NX2_OOT_HOST_ZSTD=0 turns it off. Unmeasured.

9. UsageTracker full pages (322e4ddfd)

From Spikol02/Nyamu#15 by Spiks. One bit per page records pages already full, so they are skipped 64 at a time. EDEN_NX2_USAGE_TRACKER_SKIP_FULL=0 turns it off. Unmeasured.

10. Native graphics SSBO sizing (f76899f23)

From Spikol02/Nyamu#21, itself from Frizti/Fritchii 82a57c333175. The cost was found in zortos293#101. A native graphics storage pointer takes the next word as its size instead of binding up to 8 MiB. This was not already in the tree. EDEN_NX2_GRAPHICS_SSBO_BARE=1 restores the old binding. Unmeasured.

11. Cheaper kernel locks and memory traps (0d57a3921)

The kernel-lock commit of javocsoft#16 (e0ae6b08b99f), from zortos293#717 by Stephen Jayakar. It covers KSpinLock, the KAddressArbiter spin, PhysicalCore idle, the KThread context guard, and GPU memory trap caches. Occlusion Off and the rest of zortos293#16 are not taken. Unmeasured.

12. FFMA/HFMA2 FMZ with SAT (c684f4af8), visual

From zortos293#825 by pvns1108 (fixes zortos293#745). zortos293#825 was picked over zortos293#811. FMZ+SAT no longer drops the addend, and the F16 FMZ select is fixed. This is the candidate for the canopy haze. Not checked in game. EDEN_NX2_OOT_VISUAL_WORKAROUND is not in this tree. It is a temporary opt-in per-shader edit in selevet-eglantine and was not ported. Unmeasured.

New switches

• EDEN_NX2_BACK_MACROS=1: enables commit 7. Off by default.
• EDEN_NX2_OOT_HOST_ZSTD=0: disables commit 8.
• EDEN_NX2_USAGE_TRACKER_SKIP_FULL=0: disables commit 9.
• EDEN_NX2_GRAPHICS_SSBO_BARE=1: reverts commit 10.

What this does not claim

Commits 5–12 are unmeasured. For commit 5, watch writer wait (~378 ms/s) and list writers (~284 ms/s). A wrong resume would show up as a command list read before its copy finished: garbage methods, a channel stuck on a release, or popped geometry. Not a claim of sustained 30.

Testing

• storage_write_extent: 13 tests, 61 assertions, passed.
• async_report_settled: an issued slot blocks until it completes. Passed.
• semaphore_acquire: a CPU word is read from memory; a queued release is waited for even when the old word would match; a dirty cache or unfinished query drains; the newer of two releases is the one waited for. 4 tests, 8 assertions, passed.
• list_read_resume: a resumed read skips the writer wait once; clearing the flag before the read does not skip a later wait. 2 tests, 4 assertions, passed.
• dma_pusher.cpp, vk_rasterizer.cpp, and the Vulkan BufferCache instantiation build with -Werror.
• command_entry_wait (new), async_report_settled, semaphore_acquire, list_read_resume: passed.
• transcode_cache, range_cover (new): passed.
• turing_macro_effects (new): passed, 6,785 assertions.
• oot_zstd_hook.cpp, svc.cpp, nso.cpp: compile. No upstream test.
• usage_tracker: 7 cases, 279,715 assertions, passed with and without EDEN_NX2_USAGE_TRACKER_SKIP_FULL=0.
• kernel_locks (new): 2 cases, passed. Kernel TUs build.
• maxwell_fmz (new): 4 cases, passed. 3 of them fail against the old translation.
• Tests were linked standalone against Catch2 and run on Linux with GCC 14; the Windows package is built by nup-romfs-windows.
• Not run: the game.

Legion check

Same settings as the 3863d03ef9 user-play run. Outdoor Kokiri, trees and walking, not the title screen.

1. EDEN_NX2_GPU_PROFILE=1. Compare against ec8e93f750: writer wait ~378, list writers ~284, split ~232, acquire ~93, DMA ~102 ms/s, downloads 0.
2. Overlay fps on village, rocks, and the stream. Still no playable-30 claim until that measurement.
3. Watch for a list read that ran early: unbound subchannels, a channel that never resumes, or geometry popping in.
4. Visual: canopy haze A/B (commit 12), and missing geometry or vegetation (commit 10; try EDEN_NX2_GRAPHICS_SSBO_BARE=1).