[dean0396]

What and why

pkg_upload_ asks up_big_src for a texture's no-copy source buffer twice for nearly every staged texture that reads a large pool no table holds whole, or a pool no GPU address was given out in. The first call comes before the staging allocation. The second comes right after it and always runs, in case up_alloc flushed the batch that held the first buffer. up_alloc only flushes when the batch's staging would pass UP_BATCH_BYTES, so the second buffer is almost always a duplicate over the same pages in the same open batch. Each duplicate is one more residency add, and one more remove when the batch completes.

My fork's build has the same two calls but not main's big_full check, so it also made source buffers over large pools the raw table holds whole. On it, the residency site in up_big_src added 466 buffers (1,324 MB) for about 233 textures (662 MB) on the first turn into Kakariko in Ocarina of Time, and 482 buffers (1,395 MB) in Castle Town. Those counts can include textures in large pools held whole, which main leaves out, so main probably makes fewer there. How many duplicates main makes in a game is not measured.

Change:

• pkg_upload_ keeps g_up_seq after the first call. It calls up_big_src again only when that value changed, that is, when up_alloc flushed the batch and began a new one (up_begin is the only writer of g_up_seq).
• DM_NVN_UPSRC_DEDUPE=0 brings back the two calls, with a row in docs/env_switches.tsv.
• The upload staging blocks line also counts the source buffers made, those made by the second call, and the second calls left out.
• New GPU test t_upload_src with its Makefile rule (check_t_upload_src runs it with the switch on and off).
• One sentence added to the "Memory pools nothing gave an address to stay out of the raw pool table" entry in docs/reference/memory.md.

Main already has the big_full check in up_big_src (no source buffer over a large pool that raw global memory's table holds whole), so this only touches the second call.

Checked

• Tests (M3 Max, macOS 26.7.2, run on main 352e4a6; the branch is now rebased onto 8b55ebb, which changes only the shader build tools and docs, and t_upload_src gives the same 132 / 95 checks there; make -C dm/gfx/tests GAME=0400486024DEC000 check_<test>):
  • t_upload_src: 132 checks, 0 failed with the default; 95 checks, 0 failed with DM_NVN_UPSRC_DEDUPE=0.
    • A fresh batch makes 1 buffer and 1 residency add at up_big_src's site.
    • A batch that the texture's own staging must flush makes 2 buffers, one in each batch, and both batches complete.
    • With the switch off, a fresh batch makes 2.
    • A large pool held whole makes none, with the switch on or off.
    • Every texture reads back equal to the CPU reference.
    • Negative controls: I built the test against the change with the second call made unconditional again (main's behaviour), and 12 checks fail: two buffers in every fresh batch. With the second call never made, 16 fail: the flushed-batch and switch-off cases.
  • t_pool_chunks, t_upload_slots, t_upload_budget, t_direct_upload, t_cpu_tex, t_pool_retire, t_deswz_gpu, t_initial_uncached, t_gmem_words and t_bc_alias pass with the default and with DM_NVN_UPSRC_DEDUPE=0. The results match unchanged main.
  • The whole make -k run suite in dm/gfx/tests gives the same results on the branch as on unchanged main, apart from the new t_upload_src. One check, t_feedback's "with nothing running: a commit", fails on both trees: in main's suite run (the branch's passed), and in 5 of 6 reruns on the branch and 6 of 6 on main.
  • The Mac host (dmhost, GAME=0400486024DEC000) builds with the same warnings as main.
  • dm/tools/tests/test_docs.py passes. test_env_switches.py fails with identical output on unchanged main (four switches quoted in the docs but read nowhere, and DM_NVN_VCONV_KEEP has no row). The new switch has its row.
• Games run: none. I didn't run Ocarina of Time on this branch or on main. The numbers above are from my fork's build, before the change; I have no after measurement from a game.
• Not run: scripts/regress.py check (Ocarina of Time is the only game set up on this Mac, and I didn't run it), and lint_c.py (no LLVM 23 here).

Overlapping PRs (checked against main 352e4a6): 750 PRs were open, and 367 of them touched a file this PR changes (nvn_mtl.m 219, docs/env_switches.tsv 230, dm/gfx/tests/Makefile 99, docs/reference/memory.md 41, t_upload_src.m none). I dry-ran the gh pr diff of all 367 with git apply --check on main and on this branch, whole and file by file. I also merged each PR's version of each shared file (its base blob plus its diff) onto main's and this branch's version with git merge-file, and compared which functions each PR changes. Whole diffs: 83 apply on both, 282 on neither, 2 on main only. Since then, #2339 and #2340 change nvn_mtl.m (#2340 also docs/env_switches.tsv); both apply on main and on this branch.

Apply on main but not on this branch (4):

• Soft textures: a large block-compressed texture kept without its finest level, its level-0 pages back #1599, whole diff: it adds !keep_cpu && to the line with the first up_big_src call, which this PR also changes (src_seq). git merge-file gives one conflict there.
• Partial texture re-uploads, one page check per submission #2263, whole diff: the hunk that adds pkg_written between up_big_src and pkg_upload_ has up_big_src's res_add line as context, and this PR changes that line. git merge-file merges it cleanly.
• the iPhone 17 holds Ocarina of Time's 30 fps cap through the first dungeon: smaller textures, staging released and MetalFX for its upscaler on small devices, and memory pressure frees what no frame needs #2270, docs/env_switches.tsv part (its whole diff already fails on main): it rewrites the DM_NVN_UPLOAD_REUSE row, right above the new DM_NVN_UPSRC_DEDUPE row. git merge-file gives one conflict. It also changes pkg_upload_'s level loops; its nvn_mtl.m merges with one conflict on main and on this branch alike.
• Upload staging shrinks back when it sits idle on iOS: the three slots no longer keep 192 MB of private GPU memory through steady play #1491, docs/env_switches.tsv part (its whole diff already fails on main): it adds DM_NVN_UPLOAD_IDLE_MS above DM_NVN_UPLOAD_REUSE, with DM_NVN_VBUF_CACHE as context. git merge-file merges it cleanly.

Change the same functions (up_diag, up_big_src, pkg_upload_, or the declarations and comments just before up_diag and up_big_src): 18 PRs, counting #1599, #2263 and #2270 above.

• up_big_src itself (6). None of them applies to main's nvn_mtl.m.
  • nvn: preserve texture upgrades, BC render targets, and retired descriptor bounds #1731 returns a buffer already in g_up_keep that covers the range, rounded to 8 MB windows. That would also stop a second buffer within one batch, by scanning the batch's buffers on every call.
  • Large pools held once in the residency set: segments take covering chunks' buffers, one buffer per pool past its size, a grown pool keeps it (stacked on #1407; PRAGMATA on an 8 GB Mac) #1614 makes the buffer through nocopy_buf, and returns nil over write-protected pages in sparse safe mode.
  • FIFA 23: add EA shader archive support and game profile #1696 replaces its res_add call with up_res_add.
  • WIP: DM_NVN_TEX_DEDUPE gives back the guest copy of uploaded textures (Ocarina of Time: about 1.9 GB less footprint) #1386, Raw global memory's pool table lists only the pools the game took a buffer address in: every pool was wired for it, which slowed Ocarina of Time's frames to 60-65 ms #1394 and Ocarina of Time: experimental Mac playability and performance checkpoint #1663 give it and both of its calls a texture_only argument in place of main's pool lookup.
  • Merged from their own base, this branch adds one conflict region to WIP: DM_NVN_TEX_DEDUPE gives back the guest copy of uploaded textures (Ocarina of Time: about 1.9 GB less footprint) #1386, Raw global memory's pool table lists only the pools the game took a buffer address in: every pool was wired for it, which slowed Ocarina of Time's frames to 60-65 ms #1394, Large pools held once in the residency set: segments take covering chunks' buffers, one buffer per pool past its size, a grown pool keeps it (stacked on #1407; PRAGMATA on an 8 GB Mac) #1614 and FIFA 23: add EA shader archive support and game profile #1696, and three to Ocarina of Time: experimental Mac playability and performance checkpoint #1663. nvn: preserve texture upgrades, BC render targets, and retired descriptor bounds #1731 has 28 conflict regions with main and 28 with this branch.
• Elsewhere in pkg_upload_ (9), with the same git merge-file result on main and on this branch. All but nvn: reactive memory watchdog, 16-way view cache, and texture upload batching #2008 already fail to apply on main.
  • Rhythm Heaven Groove: A16 legacy support, touch pad and ambient bars #29, Textures: upload CPU-written linear and unflushed sampled textures #167, Switch2 merge #998, WIP: translate one submission on several cores #1205, Allow up to three compressed texture mip levels to be omitted #1418, Ocarina of Time (Switch 2): GPU-written texture ranges fix the terrain holes; opt-in texture release saves ~2.3 GB; a page on playing it #1447.
  • nvn: reactive memory watchdog, 16-way view cache, and texture upload batching #2008 applies on both.
  • WIP: Mario Kart World 1.8.0 on iOS - rendering, memory and presentation #1383 also adds a function, up_source_use, between up_begin and up_diag, and calls it again after up_alloc for an upload source of its own.
  • Split nvn_mtl.m into topic parts with a script that re-cuts it on any main (mechanical, same machine code) #479 moves the whole file into parts with a script, which re-cuts it after any change.

No open PR changes the docs/reference/memory.md entry this PR edits, and none adds a t_upload_src.m. All 99 dm/gfx/tests/Makefile parts give the same result on this branch as on main.

Docs

• The finding is in its docs/reference/memory.md entry, and python3 dm/tools/docs_symptoms.py was run (the index is unchanged)
• No game files, keys or data extracted from a game in the change or the attachments

[stan69b]

Hey, I cleanup my PR to handle the conflicts with main : #2270

it should now merge if you need my PR in yours for testing.

also check this one, The iPhone PR is a follow up of that PR : #1971

[seedds]

Thanks: you found this cause first. The fix landed on main as 0c23e0b, from #2378, with the batch recorded right after the first up_big_src call, as you did. It has no switch: going back to the duplicate buffer has no use, so DM_NVN_UPSRC_DEDUPE wasn't needed. The extra counters stayed out too, and this PR no longer applied to main. On Mario Kart World, up_big_src buffers over a 200 s run went from 24,696 (5,429 MB) to 12,082 (2,424 MB).