[lunks]

What and why

Ocarina of Time (Switch 2, 0400486024DEC000, profile in #1299): in Kokiri Forest the ground is missing wherever only the terrain covers it, the sky shows through, and from above the stone path floats. @aCidcOde traced it in #1326: the ten 260x260 R32F height tiles read as all zeros on the host while their storage holds data (DM_NVN_TEXINFO shows 141,674 to 204,571 non-zero bytes), so each terrain patch is a flat plane far below the player.

Cause: tex_cpu_content refuses an uncompressed texture when its pool has gpu_written set. The terrain's 202 MB pool (flags 0x24) also holds textures the GPU writes, so the whole pool counts as written. It is the Mario Kart World terrain case described above tex_cpu_content, in a pool the GPU writes too.

Change: mtl_tex_gpu_written records the storage range of each render target, clear, copy destination and image store (sorted, merged ranges). An uncompressed texture is refused, or dropped by cpu_unc_drop, only where its own storage overlaps a range. mtl_pool_gpu_written stays, as "the whole pool".

This is aCidcOde's patch from #1326 (credit: @aCidcOde), checked against this tree, with one change: ranges go with their pool (gpuw_drop). In the patch they never shrank, so a pool made later at those addresses would have counted as written, where pool_t.gpu_written starts at 0.

Checked

• Tests (Apple M5 Pro, macOS 27, niced; make -C dm/gfx/tests GAME=010015100B514000 ...): check_t_cpu_tex 65 checks, 0 failed (new: a texture beside a GPU write is taken and one over it is not; a write over its storage ends that; ranges merge and split; a pool that goes takes them along; no range means the whole pool); check_t_cpu_storage 32 checks, 0 failed (unchanged but for the function's name); test_docs.py OK; docs_symptoms.py --check current; lint_c.py tidy on nvn_mtl.m, hle_nvn.c, t_cpu_tex.m: 0 pairs above the baseline.
• Games run: none for this branch. The ground fix itself is aCidcOde's observation on their tree ("with it the ground draws at every camera pitch on the same route"), not repeated here. This branch was prepared from our own stack's commit fafad0e6 (stack-next, Ocarina of Time v131072 on an M5 Pro), whose commit message records no in-game measurement of its own.
• Not run: the rest of make -C dm/gfx/tests run, the host tests, regress.py. The change touches every game that records GPU writes, but only through when a CPU-written uncompressed texture is taken from its storage; Mario Kart World (0400C3F00006E000), whose note describes the earlier pool rule, was not run.
• Devices and OS: Apple M5 Pro, macOS 27.

Docs

• The finding is in docs/reference/nvn-objects.md ("Uncompressed storage the CPU wrote, where the GPU has not written", renamed from "...in a pool the GPU has not written": the two links to it in the Mario Kart World note and the symptom index follow) and python3 dm/tools/docs_symptoms.py was run
• No chain tool changed
• No game files, keys or data extracted from a game in the change

[bettterlabs]

Played on a 16 GB Mac mini M4, macOS 27.2, docked, v131072. Host only: david-on-spring's #1386 branch (562c1f4 on f3ac5ea, chain 84aede056dea) + #1396 + #1430 + #1367, run without and with this PR (16e86fd cherry-picked; the code applied cleanly, and main in t_cpu_tex.m needed both sides).

The same save was put back before each run, at a spot in Kokiri Forest where the ground was missing. I didn't touch the controller after the load, and DM_NVN_SHOT took a screenshot every 300 frames.

• Without it: the sky shows through under Link, and the grass edges float.
• With it: the dirt path, the grass and Link's shadow are there.

At the same frames (1500, 2400, 3300), the two runs differ by a mean of 9.7–10.2 levels, with 11.1–11.6% of pixels off by more than 32. Within one run, shots 300 frames apart differ by 4.3–5.2 (2.0–2.5%).

Frame time p50 was 20.4–20.7 ms in both runs. GPU p50 was 37.2–37.5 ms without and 36.1–36.2 ms with. Both logged only the title movie's AV1 RED and exited cleanly.

Tests on this stack: check_t_cpu_tex 71 checks, 2 failed. The same 2 fail without this PR on this base (62 checks, 2 failed: "finalized command snapshot uploads CPU-written pixels" and "replayed retired snapshot keeps its CPU content"), so its 9 new checks pass. check_t_cpu_storage 32 checks, 0 failed.

[seedds]

Superseded by e2b76d4 (from #1604), which keeps GPU-written storage as address ranges, as this PR does. It also protects each range's pages, so storage the CPU rewrites after a GPU write is taken from storage again. On Ocarina of Time's walk, 10 of 20 such reuses were terrain maps, which ranges kept until their pool goes would have left as zeros.

Thanks to you and to aCidcOde: the reference entry credits #1326 for finding the pool rule's part. Closing this one, so that one range table lands.

[lunks]

Closing: #1604 landed on main as e2b76d4 with the same idea (GPU-written storage as ranges instead of per pool) and fmilioni's page protection for a CPU rewrite, which #1390 lacked. Thank you fmilioni for the measurements and for being open to combining them.

[PaulEins]

Still needed on current main (0c7e88b9): without it Kokiri Forest's path renders white/missing again, and with it merged it's correct. It merges onto main with two small conflicts: the rec_copy line in nvnCommandBufferCopyBufferToTexture (keep main's line and add the mtl_tex_gpu_written call), and the test list in t_cpu_tex.m (add gpu_written_ranges()). Played 5 hours on top of it without a crash (M4 Pro, 48 GB, v131072, Mac app).

Drafted with Claude; reviewed by me.