WIP: feat(vk): a packed record layout for the mesh fallback's export and reader (#1477 BC-1a) #1561

Draft
omgitsgela wants to merge 36 commits from omgitsgela/Longshot:perf/1477-bc1-packed-export into main
Contributor

Type: feat · Risk: low · Tested on: Steam Deck LCD (VanGogh, RADV), desktop session, headless test profile, in game · Needs: the release Windows job to compile the touched files (see Must be tested)

TL;DR

BC-0 (#1531) measured that ~1.5 ms of the compute mesh fallback's 6.8 ms is our record-export traffic. BC-1a makes that traffic wide: the fallback's epilogue now writes a record's vertex lines four at a time (one 16-byte store instead of four strided scalar words) and the generated vertex stage reads them back the same way. On the Deck LCD, A/B/A/B on one build: +0.3 fps steady against a 0.2 spread, and each log's whole-frame GPU time line says −0.8 to −1.1 ms a frame — matching BC-0's bound, so the wide access is real on VANGOGH. New value RX_VK_MESH_FB_LAYOUT=packed4, off by default; both packed4 arms' pictures are visually clean.

Breakdown

  • Who it helps: Steam Deck and every other GPU that takes the compute fallback: the record handoff between our fallback's mesh stage and its generated vertex stage gets wide memory access instead of one word at a time.
  • What was wrong: the fallback's record export costs ~1.5 ms a frame (BC-0's measured split), and in the shipped soa layout a thread stores a record's lines as separate scalar words, strided by the record count — never wider than 4 bytes, on a hot vegetation workload whose records are 21 to 26 vertex lines.
  • What changes: a third layout, packed4, for programs whose records are only their vertex lines: lines grouped in fours, 16-byte stores and loads, no validity word. Everything else — header, bump allocator, indirect draw, counts — is untouched. A program whose records carry per-primitive lines cannot take it and is laid out soa, its own module exactly (one shared decision helper, LaidOutLayout, asked by both the epilogue and the vertex body of the same program, so a record's write and read cannot disagree whatever the switch said).
  • What does not change: unset or soa or aos: the emitter takes the same paths as before (the fork adds Packed4 only); native mesh is untouched; saves, ini, ABI unchanged. The switch's name is under the program cache stamp's RX_VK_MESH sweep, so a packed4 program is never read back by a run that did not ask for it, and mesh + vertex stages of a program are translated in one call and saved together.
  • How it was checked: 62 layout-model cases 0 failed with 48 fixture modules spirv-val clean (the packed4 fixture set is new; the model draws packed4's records exactly as aos/soa's, and the packed4-cannot-apply fixtures are byte-identical to plain soa); four Deck arms A/B/A/B on one build with the layout pinned per arm from the log's vk: mesh fallback records: line; both packed4 end-view.png frames inspected for addressing garbage (none); a delegated read-only review of the whole diff returned SHIP with zero findings, re-running the harness independently and checking 27,948 address expressions (cites in #1477 comment 48148).
  • What is still open: nothing unfinished in this PR. The keyed fb mesh prog <key> span A/B (does the fallback's own span shrink) needs #1460's span machinery, which is not in main — recorded in the docs row; it can run cheaply once #1460 lands.
  • Risk: low: opt-in, default path unchanged by construction; worst realistic case is a packed4 bug that renders displaced geometry — which the two packed4 arms' pictures and the model's draw-equality checks are the evidence against.

Why

#1477 BC-1: restructure the fallback's output after BC-0 bounded our export traffic at ~1.5 ms of the fallback's 6.8 ms (comment 47400). BC-2a came back disproven (the hot programs are rank-identity programs, comment 47796), and indexed/dedup turned out measured-and-dead from PR #906 (comment 48043 and 47993): so the remaining honest prize in this pool is the access width of the export and reader. Packed4 is the smallest shippable increment of that: no index buffer, no new passes, opt-in.

What changed

The layout enum gains Packed4 = 2, selected by mesh_rule::fb_layout on the word packed4. The emitter decides the layout once per program through LaidOutLayout(runtime_info, mesh) — Packed4 only where the program has no per-primitive lines, soa otherwise — and both the epilogue and the generated vertex body ask the same helper. Packed4 addressing: group g of record r at H + g*4*fit + r*4 (H the 8 header words, fit from the area's words over the vertex-line count — a few more records than soa holds, since its records carry no validity word), remainder lines in soa-shaped columns at H + l*fit + r; the writer uses the allocator's base + k as the record number, the reader's gl_Vertex indexes the same space because the draw's count is the records taken and every taken record wrote its lines.

  • gpu/vk/eden/shader_recompiler/backend/spirv/emit_spirv_mesh.cpp — LaidOutLayout, WritePackedRecords, vec4 BufferVecAt/LoadVec/StoreVec, the epilogue fork, the vertex body's packed reader.
  • gpu/vk/eden/shader_recompiler/backend/spirv/spirv_emit_context.{h,cpp} — MeshFallbackVec4: the PhysicalStorageBuffer block of four words the vec4 loads/stores address.
  • gpu/vk/eden/shader_recompiler/backend/spirv/emit_spirv.h — the enum value and MeshFallbackFitPacked4 (the host asks capacity through it).
  • gpu/vk/mesh_mode.h — packed4 as the word fb_layout accepts (the first cut documented it without parsing it: an arm would have run soa silently; caught before any arm ran, fixed and covered by the model test).
  • gpu/vk/vk_shader.cpp — the vk: mesh fallback records: log says which layout a start chose, and names the soa-fallback rule.
  • tests/mesh_fallback_layout_model.cpp / _ir.cpp / test_mesh_fallback_layout.py — the model draws packed4's records exactly as aos/soa's where capacity allows; packed4-cannot-apply fixtures byte-identical to soa; the switch-word parsing pinned (including that unknown words still fall to the default); 76 fixture modules across the harness, 48 validated under spv1.6 + vulkan1.3.
  • docs/development.md — the layout row gains packed4 with its rule and its measured provenance.

Measurements

Verdict: packed4 is real but small, exactly as BC-0's split bounds it: +0.3 fps steady against a 0.2 spread, +0.5/+1.3 over whole-drive windows, −0.8 to −1.1 ms a frame of whole-frame GPU time. The fallback pool's cheap half is now collected; 30-while-moving is not in this pool.

Hardware: Steam Deck LCD, AMD Custom GPU 0405 (RADV VANGOGH), Mesa 26.2.0-devel (git-035ae2f854), auto DPM (sclk median 1300 MHz), 16 GB shared.
Scene: Hyrule Field save (copied, slot 05), 1280x800 output, resolution=4 (640x400 internal), compute fallback, RX_VK_PASS_TIMING=2, no PROF_FINE, no fps cap; scripted route (warm, walks, pans, full-stick turns, circle) by mesh2245-arm3.sh.
Method: A/B/A/B on ONE overlay (rt-rc6-bc1, this branch at 195cd15d, so=df64452c, caches unlinked, h1 pays the cold translation, h2-h4 warm): h1/h3 RX_VK_MESH_FB_LAYOUT=soa, h2/h4 =packed4. Every arm's layout pinned from its own log line, not assumed; RX_VK_GM_LOOP at its default (off on RADV) in all four — dukenheimer's variant table (#1477) had not posted at arm time.
Baseline: soa arms vs packed4 arms on the same binary — main is not the comparison because both arms must share one build.

Before: this PR @ 195cd15d, RX_VK_MESH_FB_LAYOUT=soa (h1/h3) After, RX_VK_MESH_FB_LAYOUT=packed4: this PR @ 195cd15d (h2/h4)
steady play median 29.1 / 29.3 fps 29.4 / 29.6 fps
whole-drive frametime median (last 400 CSV rows, incl. turns) 23.7 / 24.3 fps 24.2 / 25.6 fps
play-window GPU time median 25.4 / 25.1 ms 24.3 / 24.3 ms
  • Steady: +0.3 fps both pairs, within-pair run-to-run spread 0.2.
  • Whole-frame GPU time (the log's own unprofiled line, 12-13 play-window reports per arm): 25.4/25.1 vs 24.3/24.3 ms — matches BC-0's ~1.5 ms export bound, which is the mechanism evidence.
  • Turning-tail windows too noisy to claim (13.4/12.9 vs 13.2/16.0); the CSV whole-drive medians (+0.5/+1.3) include turns and point the same way as GPU time but are noisier.
  • How to reproduce: Deck LCD, the copied save, ~/longshot-investigation/mesh2245-arm3.sh LABEL ~/longshot-nvme/rt-rc6-bc1 compute 4 RX_VK_MESH_FB_LAYOUT=packed4 (and =soa against it). Rows: docs/tested/2026-11-17-packed4-abba-omgitsgela-run19.md; raw run folders mesh2245-bc1a-h1..h4 on the Deck.

Flags and settings

This PR adds one value to an existing switch: packed4 for RX_VK_MESH_FB_LAYOUT, shipped unset (soa stays the default).

Flag What it is Before this PR After this PR Effect Helps / doesn't help / untested Recommendation
RX_VK_MESH_FB_LAYOUT=packed4 the fallback's record layout, wide access existed with aos/soa only also accepts packed4 records packed in runs of four lines, no validity word; per-primitive-line programs stay soa helps: +0.3 fps steady, −1 ms GPU time on the Deck LCD (VanGogh) other GPUs unmeasured: leave unset until someone measures; campaign may flip the default after review

Testing

Tested: Steam Deck LCD desktop mode, in game, this branch at 195cd15d (four full scripted arms, A/B/A/B; no crash, no amdgpu kernel lines, layout pin correct in all four logs, both packed4 end views visually clean — checked for sheared grass, exploded quads, holes, wrong-colored patches). The branch head 0c32b62d (docs only past the arms' tree) plus merged main 1333302d: full python3 -I tests/run_all.py 610 passed / 36 failed / 128 skipped — and clean main at d1a307cb run in a separate worktree fails 39 of the same tests; the branch's 36 failures are a strict subset of main's (main additionally fails bug_report, crash_groups, hardware_report), so none is caused by this PR. tests/test_mesh_fallback_layout.py 62 cases 0 failed, 48 modules spirv-val clean (VM: g++, NDK spirv-val). tests/test_tested_configurations.py OK (it caught my first attempt at the docs rows and the rows were rewritten in the fragment format). tests/test_mesh_fallback_failed_once.py, stamp tests pass. A delegated read-only review of the diff (independent worktree, cited findings): SHIP, no CRITICAL/MAJOR/MINOR findings across writer/reader address agreement, validity-word removal (the allocator's gate is k < taken with the host barrier ordering stores before the indirect draw; no other production reader of the validity word), host↔shader capacity (both ask MeshFallbackFitPacked4; no separate host cap), vec4 alignment legality, cache stamping and paired-stage translation, soa byte-identity of packed4-cannot-apply programs, and count-vs-written records; it re-ran the harness clean and checked 27,948 address expressions (#1477 comment 48148).

Not tested: the keyed fb-span A/B on this lineage (needs #1460's span labels, not in main — the GPU-time line carries the mechanism instead); other GPUs (soa remains the measured-and-shipped default; packed4 stays opt-in); other titles; AMD's other architectures; the Game-Mode path.

Must be tested: Windows, llvm-mingw cross-compile (I have no Windows toolchain on this host): the build-windows.yml job on this PR (by hand: read the check on the PR page), expecting green for the touched emit_spirv_mesh.cpp, spirv_emit_context.{h,cpp}, vk_shader.cpp (all compiled on the Deck's clang and the VM's gcc here). Nothing else beyond the switch's own per-GPU measurement, which the flag table marks untested.

Risks

What could break: a packed4 addressing bug would show as displaced geometry in packed4 runs only — guarded by the model's draw-equality (packed4 draws the same primitives as aos/soa through capacity), 48 validated fixture modules, byte-identity where packed4 cannot apply, and both Deck arms' clean pictures. A capacity bug would overflow the area — guarded by the reviewer-checked base <= fit && taken <= fit-base allocator gate and the model's per-layout caps. Nothing for a normal run: the word is unset by default and the fork adds only the Packed4 path.

Turning it off: unset the variable (soa remains the default). The layout is part of the program cache stamp's RX_VK_MESH sweep, so switching back re-keys caches exactly like any RX_VK_MESH* change — one cold translation.

Compatibility: no save, ini, ABI or tool change; native-mesh GPUs never see a fallback program. The vec4 accesses declare Aligned 4 as every fallback access does — legal, with the intra-buffer offsets multiples of 16 bytes; whether a driver issues 16-byte transactions is driver-side, and the Deck's −1 ms GPU-time delta is the evidence for VANGOGH.

Issues

See #1477 (BC-1a of the campaign plan; the arms' table and the review outcome are comment 48148). Builds on BC-0's measured split (#1531) but does not depend on it: this branch is main + this change.

Checklist

Everyone

  • One topic.
  • The title is a Conventional Commit (feat(vk): ...) of the status line's type.
  • Why gives the reason plainly and links the issues that fit (the tracker searched).
  • Tested on my own PC where it runs (in the game, for a change the game runs): the Deck arms; what can't be run by hand is in Not tested with why.
  • Measurements has a verdict.
  • Nothing gets worse that I know of: the default path is unchanged; the switch's cost (a cold retranslation when set) is in Risks.
  • Testing names what was tested, what wasn't, and what must still be tested, where and how.
  • python -I -B tests/run_all.py passes (any failure that also fails on main is named). On the merged head (main d1a307cb merged as 1333302d): 610 passed / 36 failed / 128 skipped, and clean main at d1a307cb in its own worktree fails 39 of the same tests — the branch's 36 are a strict subset of main's (main also fails bug_report, crash_groups, hardware_report); none is caused by this PR.
  • Rebased on current main (merged 1333302d, main at d1a307cb).
  • No game files, dumps, build output or generated files.
  • third_party/ untouched (a new folder with a PROVENANCE.md instead).
  • The license is OK: GPL-3.0-or-later, nothing incompatible and nothing decompiled.

When it applies

  • AI assistance recorded: this branch's commits were written and measured with Hermes (Angela's agent), recorded here and in commit messages' bodies where the tool acted.
  • Numbers from the same PC, scene and settings, at least 2 interleaved runs each (A B B A), the median reported.
  • Switches documented (docs/development.md) and shipped at their recommended value (default soa; packed4 opt-in pending per-GPU measurement).
  • Host code builds on Linux gcc and the Deck's clang; Windows llvm-mingw is the Must-be-tested box above (no path handling changes).
  • A file in docs/tested/ for each new measured run, fragment format (the configurations checker passes).
  • Checks under tests/ that read changed source updated, not weakened: the layout model grew packed4's cases and the switch-word pins; nothing weakened (the harness's existing aos-as-main equality still passes).
  • Docs updated.
**Type:** feat · **Risk:** low · **Tested on:** Steam Deck LCD (VanGogh, RADV), desktop session, headless test profile, in game · **Needs:** the release Windows job to compile the touched files (see Must be tested) ## TL;DR BC-0 (#1531) measured that ~1.5 ms of the compute mesh fallback's 6.8 ms is our record-export traffic. BC-1a makes that traffic wide: the fallback's epilogue now writes a record's vertex lines four at a time (one 16-byte store instead of four strided scalar words) and the generated vertex stage reads them back the same way. On the Deck LCD, A/B/A/B on one build: +0.3 fps steady against a 0.2 spread, and each log's whole-frame `GPU time` line says −0.8 to −1.1 ms a frame — matching BC-0's bound, so the wide access is real on VANGOGH. New value `RX_VK_MESH_FB_LAYOUT=packed4`, off by default; both packed4 arms' pictures are visually clean. ## Breakdown - **Who it helps:** Steam Deck and every other GPU that takes the compute fallback: the record handoff between our fallback's mesh stage and its generated vertex stage gets wide memory access instead of one word at a time. - **What was wrong:** the fallback's record export costs ~1.5 ms a frame (BC-0's measured split), and in the shipped soa layout a thread stores a record's lines as separate scalar words, strided by the record count — never wider than 4 bytes, on a hot vegetation workload whose records are 21 to 26 vertex lines. - **What changes:** a third layout, `packed4`, for programs whose records are only their vertex lines: lines grouped in fours, 16-byte stores and loads, no validity word. Everything else — header, bump allocator, indirect draw, counts — is untouched. A program whose records carry per-primitive lines cannot take it and is laid out soa, its own module exactly (one shared decision helper, `LaidOutLayout`, asked by both the epilogue and the vertex body of the same program, so a record's write and read cannot disagree whatever the switch said). - **What does not change:** unset or `soa` or `aos`: the emitter takes the same paths as before (the fork adds Packed4 only); native mesh is untouched; saves, ini, ABI unchanged. The switch's name is under the program cache stamp's `RX_VK_MESH` sweep, so a packed4 program is never read back by a run that did not ask for it, and mesh + vertex stages of a program are translated in one call and saved together. - **How it was checked:** 62 layout-model cases 0 failed with 48 fixture modules spirv-val clean (the packed4 fixture set is new; the model draws packed4's records exactly as aos/soa's, and the packed4-cannot-apply fixtures are byte-identical to plain soa); four Deck arms A/B/A/B on one build with the layout pinned per arm from the log's `vk: mesh fallback records:` line; both packed4 end-view.png frames inspected for addressing garbage (none); a delegated read-only review of the whole diff returned SHIP with zero findings, re-running the harness independently and checking 27,948 address expressions (cites in #1477 comment 48148). - **What is still open:** nothing unfinished in this PR. The keyed `fb mesh prog <key>` span A/B (does the fallback's own span shrink) needs #1460's span machinery, which is not in main — recorded in the docs row; it can run cheaply once #1460 lands. - **Risk:** low: opt-in, default path unchanged by construction; worst realistic case is a packed4 bug that renders displaced geometry — which the two packed4 arms' pictures and the model's draw-equality checks are the evidence against. ## Why #1477 BC-1: restructure the fallback's output after BC-0 bounded our export traffic at ~1.5 ms of the fallback's 6.8 ms (comment 47400). BC-2a came back disproven (the hot programs are rank-identity programs, comment 47796), and indexed/dedup turned out measured-and-dead from PR #906 (comment 48043 and 47993): so the remaining honest prize in this pool is the access width of the export and reader. Packed4 is the smallest shippable increment of that: no index buffer, no new passes, opt-in. ## What changed The layout enum gains `Packed4 = 2`, selected by `mesh_rule::fb_layout` on the word `packed4`. The emitter decides the layout once per program through `LaidOutLayout(runtime_info, mesh)` — Packed4 only where the program has no per-primitive lines, soa otherwise — and both the epilogue and the generated vertex body ask the same helper. Packed4 addressing: group `g` of record `r` at `H + g*4*fit + r*4` (H the 8 header words, fit from the area's words over the vertex-line count — a few more records than soa holds, since its records carry no validity word), remainder lines in soa-shaped columns at `H + l*fit + r`; the writer uses the allocator's `base + k` as the record number, the reader's `gl_Vertex` indexes the same space because the draw's count is the records taken and every taken record wrote its lines. - `gpu/vk/eden/shader_recompiler/backend/spirv/emit_spirv_mesh.cpp` — `LaidOutLayout`, `WritePackedRecords`, vec4 `BufferVecAt`/`LoadVec`/`StoreVec`, the epilogue fork, the vertex body's packed reader. - `gpu/vk/eden/shader_recompiler/backend/spirv/spirv_emit_context.{h,cpp}` — `MeshFallbackVec4`: the PhysicalStorageBuffer block of four words the vec4 loads/stores address. - `gpu/vk/eden/shader_recompiler/backend/spirv/emit_spirv.h` — the enum value and `MeshFallbackFitPacked4` (the host asks capacity through it). - `gpu/vk/mesh_mode.h` — `packed4` as the word `fb_layout` accepts (the first cut documented it without parsing it: an arm would have run soa silently; caught before any arm ran, fixed and covered by the model test). - `gpu/vk/vk_shader.cpp` — the `vk: mesh fallback records:` log says which layout a start chose, and names the soa-fallback rule. - `tests/mesh_fallback_layout_model.cpp` / `_ir.cpp` / `test_mesh_fallback_layout.py` — the model draws packed4's records exactly as aos/soa's where capacity allows; packed4-cannot-apply fixtures byte-identical to soa; the switch-word parsing pinned (including that unknown words still fall to the default); 76 fixture modules across the harness, 48 validated under spv1.6 + vulkan1.3. - `docs/development.md` — the layout row gains `packed4` with its rule and its measured provenance. ## Measurements **Verdict: packed4 is real but small, exactly as BC-0's split bounds it: +0.3 fps steady against a 0.2 spread, +0.5/+1.3 over whole-drive windows, −0.8 to −1.1 ms a frame of whole-frame GPU time. The fallback pool's cheap half is now collected; 30-while-moving is not in this pool.** Hardware: Steam Deck LCD, AMD Custom GPU 0405 (RADV VANGOGH), Mesa 26.2.0-devel (git-035ae2f854), auto DPM (sclk median 1300 MHz), 16 GB shared. Scene: Hyrule Field save (copied, slot 05), 1280x800 output, resolution=4 (640x400 internal), compute fallback, `RX_VK_PASS_TIMING=2`, no PROF_FINE, no fps cap; scripted route (warm, walks, pans, full-stick turns, circle) by `mesh2245-arm3.sh`. Method: A/B/A/B on ONE overlay (`rt-rc6-bc1`, this branch at `195cd15d`, `so=df64452c`, caches unlinked, h1 pays the cold translation, h2-h4 warm): h1/h3 `RX_VK_MESH_FB_LAYOUT=soa`, h2/h4 `=packed4`. Every arm's layout pinned from its own log line, not assumed; `RX_VK_GM_LOOP` at its default (off on RADV) in all four — dukenheimer's variant table (#1477) had not posted at arm time. Baseline: soa arms vs packed4 arms on the same binary — main is not the comparison because both arms must share one build. | | Before: this PR @ 195cd15d, RX_VK_MESH_FB_LAYOUT=soa (h1/h3) | After, RX_VK_MESH_FB_LAYOUT=packed4: this PR @ 195cd15d (h2/h4) | |---|---|---| | steady play median | 29.1 / 29.3 fps | 29.4 / 29.6 fps | | whole-drive frametime median (last 400 CSV rows, incl. turns) | 23.7 / 24.3 fps | 24.2 / 25.6 fps | | play-window `GPU time` median | 25.4 / 25.1 ms | 24.3 / 24.3 ms | - Steady: +0.3 fps both pairs, within-pair run-to-run spread 0.2. - Whole-frame `GPU time` (the log's own unprofiled line, 12-13 play-window reports per arm): 25.4/25.1 vs 24.3/24.3 ms — matches BC-0's ~1.5 ms export bound, which is the mechanism evidence. - Turning-tail windows too noisy to claim (13.4/12.9 vs 13.2/16.0); the CSV whole-drive medians (+0.5/+1.3) include turns and point the same way as GPU time but are noisier. - How to reproduce: Deck LCD, the copied save, `~/longshot-investigation/mesh2245-arm3.sh LABEL ~/longshot-nvme/rt-rc6-bc1 compute 4 RX_VK_MESH_FB_LAYOUT=packed4` (and `=soa` against it). Rows: `docs/tested/2026-11-17-packed4-abba-omgitsgela-run19.md`; raw run folders `mesh2245-bc1a-h1..h4` on the Deck. ### Flags and settings This PR adds one value to an existing switch: `packed4` for `RX_VK_MESH_FB_LAYOUT`, shipped unset (soa stays the default). | Flag | What it is | Before this PR | After this PR | Effect | Helps / doesn't help / untested | Recommendation | |---|---|---|---|---|---|---| | `RX_VK_MESH_FB_LAYOUT=packed4` | the fallback's record layout, wide access | existed with `aos`/`soa` only | also accepts `packed4` | records packed in runs of four lines, no validity word; per-primitive-line programs stay soa | helps: +0.3 fps steady, −1 ms GPU time on the Deck LCD (VanGogh) | other GPUs unmeasured: leave unset until someone measures; campaign may flip the default after review | ## Testing **Tested:** Steam Deck LCD desktop mode, in game, this branch at `195cd15d` (four full scripted arms, A/B/A/B; no crash, no amdgpu kernel lines, layout pin correct in all four logs, both packed4 end views visually clean — checked for sheared grass, exploded quads, holes, wrong-colored patches). The branch head `0c32b62d` (docs only past the arms' tree) plus merged main `1333302d`: full `python3 -I tests/run_all.py` 610 passed / 36 failed / 128 skipped — and clean main at `d1a307cb` run in a separate worktree fails 39 of the same tests; **the branch's 36 failures are a strict subset of main's** (main additionally fails bug_report, crash_groups, hardware_report), so none is caused by this PR. `tests/test_mesh_fallback_layout.py` 62 cases 0 failed, 48 modules spirv-val clean (VM: g++, NDK spirv-val). `tests/test_tested_configurations.py` OK (it caught my first attempt at the docs rows and the rows were rewritten in the fragment format). `tests/test_mesh_fallback_failed_once.py`, stamp tests pass. A delegated read-only review of the diff (independent worktree, cited findings): SHIP, no CRITICAL/MAJOR/MINOR findings across writer/reader address agreement, validity-word removal (the allocator's gate is `k < taken` with the host barrier ordering stores before the indirect draw; no other production reader of the validity word), host↔shader capacity (both ask `MeshFallbackFitPacked4`; no separate host cap), vec4 alignment legality, cache stamping and paired-stage translation, soa byte-identity of packed4-cannot-apply programs, and count-vs-written records; it re-ran the harness clean and checked 27,948 address expressions (#1477 comment 48148). **Not tested:** the keyed `fb`-span A/B on this lineage (needs #1460's span labels, not in main — the GPU-time line carries the mechanism instead); other GPUs (soa remains the measured-and-shipped default; packed4 stays opt-in); other titles; AMD's other architectures; the Game-Mode path. **Must be tested:** Windows, llvm-mingw cross-compile (I have no Windows toolchain on this host): the `build-windows.yml` job on this PR (by hand: read the check on the PR page), expecting green for the touched `emit_spirv_mesh.cpp`, `spirv_emit_context.{h,cpp}`, `vk_shader.cpp` (all compiled on the Deck's clang and the VM's gcc here). Nothing else beyond the switch's own per-GPU measurement, which the flag table marks untested. ## Risks **What could break:** a packed4 addressing bug would show as displaced geometry in packed4 runs only — guarded by the model's draw-equality (packed4 draws the same primitives as aos/soa through capacity), 48 validated fixture modules, byte-identity where packed4 cannot apply, and both Deck arms' clean pictures. A capacity bug would overflow the area — guarded by the reviewer-checked `base <= fit && taken <= fit-base` allocator gate and the model's per-layout caps. Nothing for a normal run: the word is unset by default and the fork adds only the Packed4 path. **Turning it off:** unset the variable (soa remains the default). The layout is part of the program cache stamp's `RX_VK_MESH` sweep, so switching back re-keys caches exactly like any `RX_VK_MESH*` change — one cold translation. **Compatibility:** no save, ini, ABI or tool change; native-mesh GPUs never see a fallback program. The vec4 accesses declare `Aligned 4` as every fallback access does — legal, with the intra-buffer offsets multiples of 16 bytes; whether a driver issues 16-byte transactions is driver-side, and the Deck's −1 ms GPU-time delta is the evidence for VANGOGH. ## Issues See #1477 (BC-1a of the campaign plan; the arms' table and the review outcome are comment 48148). Builds on BC-0's measured split (#1531) but does not depend on it: this branch is main + this change. ## Checklist ### Everyone - [x] One topic. - [x] The title is a Conventional Commit (`feat(vk): ...`) of the status line's type. - [x] Why gives the reason plainly and links the issues that fit (the tracker searched). - [x] Tested on my own PC where it runs (in the game, for a change the game runs): the Deck arms; what can't be run by hand is in Not tested with why. - [x] Measurements has a verdict. - [x] Nothing gets worse that I know of: the default path is unchanged; the switch's cost (a cold retranslation when set) is in Risks. - [x] Testing names what was tested, what wasn't, and what must still be tested, where and how. - [x] `python -I -B tests/run_all.py` passes (any failure that also fails on `main` is named). On the merged head (main `d1a307cb` merged as `1333302d`): 610 passed / 36 failed / 128 skipped, and clean main at `d1a307cb` in its own worktree fails 39 of the same tests — the branch's 36 are a strict subset of main's (main also fails bug_report, crash_groups, hardware_report); none is caused by this PR. - [x] Rebased on current `main` (merged `1333302d`, main at `d1a307cb`). - [x] No game files, dumps, build output or generated files. - [x] `third_party/` untouched (a new folder with a `PROVENANCE.md` instead). - [x] The license is OK: GPL-3.0-or-later, nothing incompatible and nothing decompiled. ### When it applies - [x] AI assistance recorded: this branch's commits were written and measured with Hermes (Angela's agent), recorded here and in commit messages' bodies where the tool acted. - [x] Numbers from the same PC, scene and settings, at least 2 interleaved runs each (A B B A), the median reported. - [x] Switches documented (docs/development.md) and shipped at their recommended value (default soa; packed4 opt-in pending per-GPU measurement). - [x] Host code builds on Linux gcc and the Deck's clang; Windows llvm-mingw is the Must-be-tested box above (no path handling changes). - [x] A file in `docs/tested/` for each new measured run, fragment format (the configurations checker passes). - [x] Checks under `tests/` that read changed source updated, not weakened: the layout model grew packed4's cases and the switch-word pins; nothing weakened (the harness's existing aos-as-main equality still passes). - [x] Docs updated.
RX_VK_MESH_FB_LAYOUT=packed4: the compute fallback's mesh epilogue writes a
record's vertex lines four at a time (16-byte stores in the groups' region,
the remainder in soa's columns), and the generated vertex stage reads them
back four at a time. The validity word is gone: the draw's count is the
records the workgroups took, and every one of them wrote its lines. A
program whose records hold per-primitive lines is laid out soa, its own
module exactly, so a program's write and read cannot disagree whatever
the switch said. The header, allocator, indirect draw and counts are
unchanged; the fit count follows the layout, which the host asks through
the same LaidOutLayout answer the emitter gives.

Soa's threads store a word per line; a packed4 record of four lines is
one vec4 store, so the Field's hot programs (21 to 26 vertex lines) go
from 21 strided scalar stores a record to five vec4 stores plus a tail.
The Deck's A B B A arms follow on #1477 as run19's pair; the layout
model draws packed4's records exactly as aos and soa's, and 76 fixture
modules are spirv-val clean. Unmeasured against soa until those arms
run: the default stays soa.
The first BC-1a commit did not build: spirv_emit_context.cpp used
mesh_fallback_vec4_ptr before the header declared it; the epilogue's and
reader's Const(4) were ambiguous with the Const(u32,u32,..) composites
(need 4U); OpCompositeExtract wants a literal index, not the Id Const(c)
gave it; the record line's attribute u32 needs IR::Attribute{} at the
StoreFallbackOutput call; and the loop body left a bare block where the
If(...) closer belonged. The mesh translation unit now passes a syntax
check on the harness's own include set.
mesh_mode.h's doc block promised packed4 since the first BC-1a commit, but
fb_layout() only accepted aos and soa: an arm with RX_VK_MESH_FB_LAYOUT=
packed4 would have logged 'the default' and run soa in both arms, quietly
measuring nothing. A throwaway check of the pure parser (a binary of
mesh_mode.h's fb_layout over aos/soa/packed4/PACKED4/unknown words) says
the words take 0/1/2 and unknown words still fall to the default. It
matters for the arms: RX_VK_MESH_FB_LAYOUT is host-side, so a packed4 arm
against the so-far-built overlay needed no relink, but it needed this
commit to not be a no-op.
Two measurement rows for the #1477 campaign's BC-5 report (the ticket's
comment 48043 cites both): the run20/run21 A/B on the Deck LCD says the
fine profiler costs ~0.3 fps steady, so r18's span shares are real; and
the r18 row records the steady-play breakdown itself — passes 9.2 ms,
game compute 6.9, this fallback 6.8, barriers 1.4 — with the run18
caveat that its first steady block ran the command-buffer timestamp
budget out and the second, clean block is the one quoted. The named
program and dispatch keys are where the follow-ups start: dukenheimer's
skip pilot takes the five pass keys, the cold SPV probes are mine.
test_tested_configurations.py caught me writing prose where docs/tested
wants the standard header, separator and exactly one row: both rows are
now the eight-cell form, standing alone with their details repeated in
their own cells (no 'as above'), raw-run paths included. The checker
passes again.
docs(tested): BC-1a's A B B A on the Deck LCD — packed4 wins both pairs
Some checks failed
PR description / Description follows the template (pull_request_target) Successful in 4s
Website / Website on Cloudflare (pull_request_target) Has been skipped
CI / Conventional Commit title (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been cancelled
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been cancelled
CI / Workflow lint (actionlint) (pull_request) Has been cancelled
Guide screenshots / Guide images and local UI source review (pull_request) Has been cancelled
Third-party parts / As their notes say (pull_request) Has been cancelled
Third-party parts / The same as upstream (pull_request) Has been cancelled
Website / Website on Cloudflare (pull_request) Has been cancelled
0c32b62d0c
+0.3 fps steady against a 0.2 spread, +0.5/+1.3 over the whole drive
window, and each log's own GPU time line says -0.8 to -1.1 ms a frame
over play — the size BC-0 bounded the record export at, so the wide
access is real on VANGOGH. Both packed4 end views are clean Field
pictures; the turning tail stays too noisy to claim. The keyed fb spans
would have needed #1460's branch (not merged), so the GPU time line
carries the mechanism instead — recorded honestly in the row.
Author
Contributor

Tested in game: Steam Deck LCD desktop mode, four scripted A/B/A/B arms at 195cd15d (rt-rc6-bc1, so=df64452c): layout pin correct in all four logs, both packed4 end views visually clean, no crash or kernel lines

Tested in game: Steam Deck LCD desktop mode, four scripted A/B/A/B arms at 195cd15d (rt-rc6-bc1, so=df64452c): layout pin correct in all four logs, both packed4 end views visually clean, no crash or kernel lines
docs(tested): the three BC rows name the driver the logs actually printed
Some checks failed
PR description / Description follows the template (pull_request_target) Successful in 4s
CI / Conventional Commit title (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been cancelled
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been cancelled
CI / Workflow lint (actionlint) (pull_request) Has been cancelled
Guide screenshots / Guide images and local UI source review (pull_request) Has been cancelled
Third-party parts / As their notes say (pull_request) Has been cancelled
Third-party parts / The same as upstream (pull_request) Has been cancelled
Website / Website on Cloudflare (pull_request) Has been cancelled
8c31a9d9e0
The rows said Mesa 26.1.2; every arm log says `vk: driver: radv Mesa
26.2.0-devel (git-035ae2f854)` (run19 h1-h4, run20, run21). Checked each
log's own line, replaced in all three rows.
docs(tested): run19's arms ran pass timing off, not 2
Some checks failed
PR description / Description follows the template (pull_request_target) Successful in 4s
CI / Conventional Commit title (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been cancelled
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been cancelled
CI / Workflow lint (actionlint) (pull_request) Has been cancelled
Guide screenshots / Guide images and local UI source review (pull_request) Has been cancelled
Third-party parts / As their notes say (pull_request) Has been cancelled
Third-party parts / The same as upstream (pull_request) Has been cancelled
Website / Website on Cloudflare (pull_request) Has been cancelled
8640782968
busy-wrap3.sh forces PT=0, and h1-h4 launch manifests say
RX_VK_PASS_TIMING=0. The row said 2. The fps and GPU-time readings don't
depend on it; the row's Settings cell now says what the arms ran.
Contributor

Read ahead of a Steam Deck OLED run of this (head 864078296; built as main @ 9af4808f plus exactly this PR's own commits, which apply there unchanged with lines identical to the PR's diff; only longshot_vk.so differs from that main). The layout model gives 62 cases, 0 failed and 48 modules spirv-val clean here too, and the 7 failures of the changed set all fail on that main. The OLED arms are queued (the Field at 50% and at 100% standing, Kokiri at 50%, unset against packed4, the log's GPU time line per window); numbers will follow here. From the read:

Found sound

  • Address agreement: writer and reader use the same expressions and the same MeshFallbackFitPacked4(V); groups and tail columns cannot overlap (r < fit), and everything ends inside the host's area, whose size does not depend on the layout. Both stages ask LaidOutLayout, so a program's write and read cannot disagree.
  • Dropping the validity word: thread 0 reserves base, and only when the block fits does it add to the draw count and publish base; the workgroup then writes all it took with no early exit. Since the reservations are monotonic, every workgroup before the first that does not fit does fit and every one after has base > fit, so the draw's count is exactly the contiguous run of written records. Past capacity, zero-output meshlets and out-of-range slots add, write and draw nothing; a reservation with no write loop cannot happen.
  • Programs with per-primitive lines go through the unchanged soa emitter; no other code reads the layout value.

Worth a line in the description

  • The PR edits stamped translator files, so its build stamp differs from main's even with the switch unset: every player's caches retranslate once after it lands, whichever layout they run; and packed4 is a second stamp, so a tester needs two preparations on a Deck (an explicit =soa would be a third stamp text for the same code: arms that pin soa by leaving the switch unset run what a player gets).

Small things

  • With the validity word gone the reader loses soa's last line of defence (an unwritten record's position zeroed): a future bug that overshot the count would draw stale vertices, not degenerate ones. Not a defect today; a fixture that checks count equals records written at the capacity boundary would pin it.
  • The 16-byte accesses are declared Aligned 4: legal, and the buffer's base is in practice far more aligned, though nothing asserts it; Aligned 16 would say what is relied on.
  • A duplicated, truncated doc-comment block above LaidOutLayout.
Read ahead of a Steam Deck OLED run of this (head `864078296`; built as main @ 9af4808f plus exactly this PR's own commits, which apply there unchanged with lines identical to the PR's diff; only `longshot_vk.so` differs from that main). The layout model gives 62 cases, 0 failed and 48 modules `spirv-val` clean here too, and the 7 failures of the changed set all fail on that main. The OLED arms are queued (the Field at 50% and at 100% standing, Kokiri at 50%, unset against `packed4`, the log's `GPU time` line per window); numbers will follow here. From the read: **Found sound** - Address agreement: writer and reader use the same expressions and the same `MeshFallbackFitPacked4(V)`; groups and tail columns cannot overlap (`r < fit`), and everything ends inside the host's area, whose size does not depend on the layout. Both stages ask `LaidOutLayout`, so a program's write and read cannot disagree. - Dropping the validity word: thread 0 reserves `base`, and only when the block fits does it add to the draw count and publish `base`; the workgroup then writes all it took with no early exit. Since the reservations are monotonic, every workgroup before the first that does not fit does fit and every one after has `base > fit`, so the draw's count is exactly the contiguous run of written records. Past capacity, zero-output meshlets and out-of-range slots add, write and draw nothing; a reservation with no write loop cannot happen. - Programs with per-primitive lines go through the unchanged soa emitter; no other code reads the layout value. **Worth a line in the description** - The PR edits stamped translator files, so its build stamp differs from main's **even with the switch unset**: every player's caches retranslate once after it lands, whichever layout they run; and `packed4` is a second stamp, so a tester needs two preparations on a Deck (an explicit `=soa` would be a third stamp text for the same code: arms that pin soa by leaving the switch unset run what a player gets). **Small things** - With the validity word gone the reader loses soa's last line of defence (an unwritten record's position zeroed): a future bug that overshot the count would draw stale vertices, not degenerate ones. Not a defect today; a fixture that checks count equals records written at the capacity boundary would pin it. - The 16-byte accesses are declared `Aligned 4`: legal, and the buffer's base is in practice far more aligned, though nothing asserts it; `Aligned 16` would say what is relied on. - A duplicated, truncated doc-comment block above `LaidOutLayout`.
docs(tested): BC-3's out-of-band Mesa 26.3.0 pair on the Deck
Some checks failed
PR description / Description follows the template (pull_request_target) Successful in 4s
CI / Conventional Commit title (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been cancelled
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been cancelled
CI / Workflow lint (actionlint) (pull_request) Has been cancelled
Guide screenshots / Guide images and local UI source review (pull_request) Has been cancelled
Third-party parts / As their notes say (pull_request) Has been cancelled
Third-party parts / The same as upstream (pull_request) Has been cancelled
Website / Website on Cloudflare (pull_request) Has been cancelled
dcb7b870e4
Driver swap moves steady fps by +0.1 (inside a 0.3 pair spread) with GPU
time -0.3 ms and a mixed replay breakdown; first launch under the new
driver compiles 36,540 pipelines in about 27 minutes.
docs(tested): the turning-tail numbers are fps, not ms
Some checks failed
PR description / Description follows the template (pull_request_target) Successful in 4s
CI / Conventional Commit title (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been cancelled
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been cancelled
CI / Workflow lint (actionlint) (pull_request) Has been cancelled
Guide screenshots / Guide images and local UI source review (pull_request) Has been cancelled
Third-party parts / As their notes say (pull_request) Has been cancelled
Third-party parts / The same as upstream (pull_request) Has been cancelled
Website / Website on Cloudflare (pull_request) Has been cancelled
84d48de17a
docs(tested): the windowed against headless pair
Some checks failed
PR description / Description follows the template (pull_request_target) Successful in 6s
CI / Conventional Commit title (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been cancelled
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been cancelled
CI / Workflow lint (actionlint) (pull_request) Has been cancelled
Guide screenshots / Guide images and local UI source review (pull_request) Has been cancelled
Third-party parts / As their notes say (pull_request) Has been cancelled
Third-party parts / The same as upstream (pull_request) Has been cancelled
Website / Website on Cloudflare (pull_request) Has been cancelled
9af4d2321e
Same binary, same scene: windowed 29.6/29.6 against headless 29.4/29.2,
GPU time -0.2 ms. The compositor and real present cost sit below the
run-to-run spread here, so the campaign's headless tables represent
desktop play at these settings.
docs(tested): Kokiri's second scene and a pacing check
Some checks failed
PR description / Description follows the template (pull_request_target) Successful in 4s
CI / Conventional Commit title (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been cancelled
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been cancelled
CI / Workflow lint (actionlint) (pull_request) Has been cancelled
Guide screenshots / Guide images and local UI source review (pull_request) Has been cancelled
Third-party parts / As their notes say (pull_request) Has been cancelled
Third-party parts / The same as upstream (pull_request) Has been cancelled
Website / Website on Cloudflare (pull_request) Has been cancelled
51901c1c07
Kokiri needs ~10 ms less GPU a frame than the Field yet lands on the same
~29.5 fps, and RX_VK_PACE=0 changes nothing: the plateau is the game's own
cadence, so steady fps is saturated and GPU time is the signal.
docs(tested): one row per file for the Kokiri and pacing rows
Some checks failed
PR description / Description follows the template (pull_request_target) Successful in 5s
CI / Conventional Commit title (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been cancelled
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been cancelled
CI / Workflow lint (actionlint) (pull_request) Has been cancelled
Guide screenshots / Guide images and local UI source review (pull_request) Has been cancelled
Third-party parts / As their notes say (pull_request) Has been cancelled
Third-party parts / The same as upstream (pull_request) Has been cancelled
Website / Website on Cloudflare (pull_request) Has been cancelled
5125cbcb10
The reader wants a single row per fragment; the pacing check gets its own
file (run25) instead of riding in run24.
docs(tested): pin the pacing arm's switches from its own log
Some checks failed
PR description / Description follows the template (pull_request_target) Successful in 4s
CI / Conventional Commit title (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been cancelled
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been cancelled
CI / Workflow lint (actionlint) (pull_request) Has been cancelled
Guide screenshots / Guide images and local UI source review (pull_request) Has been cancelled
Third-party parts / As their notes say (pull_request) Has been cancelled
Third-party parts / The same as upstream (pull_request) Has been cancelled
Website / Website on Cloudflare (pull_request) Has been cancelled
3029721ea3
The switch list carries RX_VK_PACE=0 (player) and HOS_FPS_MAX=1000 (player),
so the row no longer leans on the uncapped test line alone.
docs(tested): the read-back wait, not a cadence, caps these frames
Some checks failed
PR description / Description follows the template (pull_request_target) Successful in 5s
CI / Conventional Commit title (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been cancelled
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been cancelled
CI / Workflow lint (actionlint) (pull_request) Has been cancelled
Guide screenshots / Guide images and local UI source review (pull_request) Has been cancelled
Third-party parts / As their notes say (pull_request) Has been cancelled
Third-party parts / The same as upstream (pull_request) Has been cancelled
Website / Website on Cloudflare (pull_request) Has been cancelled
60f6018e08
zeldaootr's check is real on the LCD too: RX_VK_READBACK_LATE=newest lifts
Kokiri standing from 29.1-29.5 to 32.3-32.9 across three arms against one
clean control. My run25 cadence claim is wrong and retracted; run24 gets
read with run26. The Field arm does not move, matching his 50% result.
docs(tested): packed4 against soa with the read-back wait probed off
Some checks failed
PR description / Description follows the template (pull_request_target) Successful in 4s
CI / Conventional Commit title (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been cancelled
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been cancelled
CI / Workflow lint (actionlint) (pull_request) Has been cancelled
Guide screenshots / Guide images and local UI source review (pull_request) Has been cancelled
Third-party parts / As their notes say (pull_request) Has been cancelled
Third-party parts / The same as upstream (pull_request) Has been cancelled
Website / Website on Cloudflare (pull_request) Has been cancelled
17dc9c46b0
Same null layout result under the probe: standing 32.9/32.9 soa against
33.1/33.5 packed4, GPU 16.0/15.0 against 14.5/14.4 ms. The probe lifts all
four arms to 32.6-33.9 where unprobed Kokiri sat at 29.1-29.6.
Contributor

Tester's first numbers on a Steam Deck OLED, the setting where its GPU binds: packed4 takes 0.4 ms of GPU time a frame off the Field standing at 100% (every packed4 run below every default run), and nothing measurable while turning. The Field at 50%, Kokiri, and both again with RX_VK_READBACK_LATE=newest (the wait out of the way, as in #1477) are running now.

This PR's code on plain main 9af4808f, one binary for both arms; A = RX_VK_MESH_FB_LAYOUT unset (soa), B = packed4, each on its own prepared cache, order A B A B B A. Hyrule Field, Lost Woods tunnel mouth at night, 1280x800 native (Render 100%), the game's own upscaler, vsync off, HOS_FPS_MAX=1000, 45 s standing then two full-stick pans. SteamOS 3.9.5, Mesa 26.2.0-devel, GPU at manual 1600 MHz before and after every run.

A1 A2 A3 B1 B2 B3 overlap
GPU time a frame, standing (ms) 24.15 24.2 24.3 23.8 23.8 23.8 none: -0.4 ms (-1.7%)
GPU time a frame, turning (ms) 21.4 21.9 21.5 21.4 21.1 21.8 yes
standing fps 26.45 26.43 26.42 26.62 26.70 26.66 none: +0.9%
turning fps 21.46 22.33 21.95 22.89 22.73 22.15 yes
turning p99 (ms) 104.2 95.3 82.5 81.3 105.7 95.7 yes
worst 1-s fps in the turn 16 17 17 17 15 16 yes
  • Arms pinned four ways: the layout line from each arm's cold preparation (line after line (soa) (the default) / packed four lines (#1477 BC-1a; …)), the running game's environment, one longshot_vk.so hash in all six runs, and each run keeping its own arm's cache (47,798 program keys kept, no preparation). Draws and dispatches a frame are equal in A and B (1,091-1,095 and 145 standing).
  • Size: about 40% of your LCD saving. This is 1600 MHz on a night scene, so the fallback's share of the frame differs.
  • The turn: GPU time is the same in both arms and the dip numbers overlap completely; B's turning median is 3.6% higher but inside A's spread. Consistent with the turn here being bound on the host as well.
  • Pictures: standing and after-turn shots of all six runs, vegetation by eye: tufts, ferns and the tunnel's foliage intact, no sheared blades, spikes, holes or wrong colours; what differs between runs (ground mist, Link's tunic in shadow) differs between A runs as much. Flicker during the turn can't be judged from stills.
  • One thing to know: the cold preparation with packed4 took 811 s against 639 s for the default and reached 159 MiB of available memory (605 MiB for the default), the lowest a preparation has reached on this Deck. It completed. One preparation each, so that is an observation, not a measured difference.
Tester's first numbers on a Steam Deck OLED, the setting where its GPU binds: **packed4 takes 0.4 ms of GPU time a frame off the Field standing at 100% (every packed4 run below every default run), and nothing measurable while turning.** The Field at 50%, Kokiri, and both again with `RX_VK_READBACK_LATE=newest` (the wait out of the way, as in https://git.magicbean.dev/MagicBeanDev/Longshot/issues/1477) are running now. This PR's code on plain main `9af4808f`, one binary for both arms; A = `RX_VK_MESH_FB_LAYOUT` unset (soa), B = `packed4`, each on its own prepared cache, order A B A B B A. Hyrule Field, Lost Woods tunnel mouth at night, 1280x800 native (Render 100%), the game's own upscaler, vsync off, `HOS_FPS_MAX=1000`, 45 s standing then two full-stick pans. SteamOS 3.9.5, Mesa 26.2.0-devel, GPU at manual 1600 MHz before and after every run. | | A1 | A2 | A3 | B1 | B2 | B3 | overlap | |---|---|---|---|---|---|---|---| | GPU time a frame, standing (ms) | 24.15 | 24.2 | 24.3 | 23.8 | 23.8 | 23.8 | **none: -0.4 ms (-1.7%)** | | GPU time a frame, turning (ms) | 21.4 | 21.9 | 21.5 | 21.4 | 21.1 | 21.8 | yes | | standing fps | 26.45 | 26.43 | 26.42 | 26.62 | 26.70 | 26.66 | none: +0.9% | | turning fps | 21.46 | 22.33 | 21.95 | 22.89 | 22.73 | 22.15 | yes | | turning p99 (ms) | 104.2 | 95.3 | 82.5 | 81.3 | 105.7 | 95.7 | yes | | worst 1-s fps in the turn | 16 | 17 | 17 | 17 | 15 | 16 | yes | - **Arms pinned four ways:** the layout line from each arm's cold preparation (`line after line (soa) (the default)` / `packed four lines (#1477 BC-1a; …)`), the running game's environment, one `longshot_vk.so` hash in all six runs, and each run keeping its own arm's cache (47,798 program keys kept, no preparation). Draws and dispatches a frame are equal in A and B (1,091-1,095 and 145 standing). - **Size:** about 40% of your LCD saving. This is 1600 MHz on a night scene, so the fallback's share of the frame differs. - **The turn:** GPU time is the same in both arms and the dip numbers overlap completely; B's turning median is 3.6% higher but inside A's spread. Consistent with the turn here being bound on the host as well. - **Pictures:** standing and after-turn shots of all six runs, vegetation by eye: tufts, ferns and the tunnel's foliage intact, no sheared blades, spikes, holes or wrong colours; what differs between runs (ground mist, Link's tunic in shadow) differs between A runs as much. Flicker during the turn can't be judged from stills. - **One thing to know:** the cold preparation with packed4 took 811 s against 639 s for the default and reached **159 MiB** of available memory (605 MiB for the default), the lowest a preparation has reached on this Deck. It completed. One preparation each, so that is an observation, not a measured difference.
docs(tested): packed4 at 100% render shows its ms as frames
Some checks failed
PR description / Description follows the template (pull_request_target) Successful in 4s
CI / Conventional Commit title (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been cancelled
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been cancelled
CI / Workflow lint (actionlint) (pull_request) Has been cancelled
Guide screenshots / Guide images and local UI source review (pull_request) Has been cancelled
Third-party parts / As their notes say (pull_request) Has been cancelled
Third-party parts / The same as upstream (pull_request) Has been cancelled
Website / Website on Cloudflare (pull_request) Has been cancelled
bb8ec12a9b
soa 25.1/25.3 against packed4 25.6/25.5 standing, GPU 31.9/31.9 against
31.0/31.0 ms: the tightest layout spread of the campaign, because at
100% the Field frame is GPU-limited rather than wait-limited.
Author
Contributor

Two more measured rows on this branch since the description was written, both strengthening the case:

  • docs/tested/2026-11-18-probed-layout-ab-omgitsgela-run27.md: with the host-side read-back wait probed out (RX_VK_READBACK_LATE=newest), Kokiri arms lift to ~33 fps and the layout stays fps-null while GPU time keeps its ~1 ms (16.0/15.0 against 14.5/14.4). The ms are robust to the regime.
  • docs/tested/2026-11-18-field100-packed4-abba-omgitsgela-run28.md: at Render 100% in Hyrule Field the ms convert to frames — standing soa 25.1/25.3 against packed4 25.6/25.5, window groups barely overlapping, GPU time 31.9/31.9 against 31.0/31.0 ms. At 100% the Field frame is GPU-bound (below the read-back plateau), which is where users play when they complain about detail loss at 50%.

So the fps claim in the TL;DR should read: invisible at 50% render (wait-limited), visible at 100% render (GPU-limited), −0.9 to −1.1 ms whole-frame GPU time in both. Full thread: #1477 comments 48564, 48600, 48643.

Two more measured rows on this branch since the description was written, both strengthening the case: - `docs/tested/2026-11-18-probed-layout-ab-omgitsgela-run27.md`: with the host-side read-back wait probed out (`RX_VK_READBACK_LATE=newest`), Kokiri arms lift to ~33 fps and the layout stays fps-null while `GPU time` keeps its ~1 ms (16.0/15.0 against 14.5/14.4). The ms are robust to the regime. - `docs/tested/2026-11-18-field100-packed4-abba-omgitsgela-run28.md`: **at Render 100% in Hyrule Field the ms convert to frames** — standing soa 25.1/25.3 against packed4 25.6/25.5, window groups barely overlapping, `GPU time` 31.9/31.9 against 31.0/31.0 ms. At 100% the Field frame is GPU-bound (below the read-back plateau), which is where users play when they complain about detail loss at 50%. So the fps claim in the TL;DR should read: invisible at 50% render (wait-limited), visible at 100% render (GPU-limited), −0.9 to −1.1 ms whole-frame GPU time in both. Full thread: #1477 comments 48564, 48600, 48643.
docs(tested): RX_VK_TEX_AHEAD damps the turning retirement storm
Some checks failed
PR description / Description follows the template (pull_request_target) Successful in 4s
CI / Conventional Commit title (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been cancelled
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been cancelled
CI / Workflow lint (actionlint) (pull_request) Has been cancelled
Guide screenshots / Guide images and local UI source review (pull_request) Has been cancelled
Third-party parts / As their notes say (pull_request) Has been cancelled
Third-party parts / The same as upstream (pull_request) Has been cancelled
Website / Website on Cloudflare (pull_request) Has been cancelled
e67135b4d2
Six arms at Render 100%: standing identical, but the off arms burst to
11 textures made a frame with descriptors to 30 ms and dips to 6-15 fps;
the on arms cap bursts at 3.3-8.9 and never dip below 16.
docs(tested): the MESH_NOCULL workaround costs ~0.7 ms, under the BC-4 gate
Some checks failed
PR description / Description follows the template (pull_request_target) Successful in 4s
CI / Conventional Commit title (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been cancelled
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been cancelled
CI / Workflow lint (actionlint) (pull_request) Has been cancelled
Guide screenshots / Guide images and local UI source review (pull_request) Has been cancelled
Third-party parts / As their notes say (pull_request) Has been cancelled
Third-party parts / The same as upstream (pull_request) Has been cancelled
Website / Website on Cloudflare (pull_request) Has been cancelled
d3e3414fce
Kokiri at Render 100%, AABB: standing identical, GPU 24.5/24.8 ms off
against 25.8/25.3 ms on. The draw-everything workaround is real but
~0.7 ms, below the ~1 ms pre-registered threshold, so BC-4 and its lite
variant both stay parked.
TEX_AHEAD at 50% render: no cost, storm columns shrink there too.
Some checks failed
PR description / Description follows the template (pull_request_target) Successful in 4s
CI / Conventional Commit title (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been cancelled
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been cancelled
CI / Workflow lint (actionlint) (pull_request) Has been cancelled
Guide screenshots / Guide images and local UI source review (pull_request) Has been cancelled
Third-party parts / As their notes say (pull_request) Has been cancelled
Third-party parts / The same as upstream (pull_request) Has been cancelled
Website / Website on Cloudflare (pull_request) Has been cancelled
b7e6cd6304
mesh ceiling at 100% is +3 fps standing but the turn survives it; Auto dynres latches 75% and buys the turn nothing.
Some checks failed
PR description / Description follows the template (pull_request_target) Successful in 4s
CI / Conventional Commit title (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been cancelled
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been cancelled
CI / Workflow lint (actionlint) (pull_request) Has been cancelled
Guide screenshots / Guide images and local UI source review (pull_request) Has been cancelled
Third-party parts / As their notes say (pull_request) Has been cancelled
Third-party parts / The same as upstream (pull_request) Has been cancelled
Website / Website on Cloudflare (pull_request) Has been cancelled
adbc1c99a0
vk: demand a 16-aligned records base so packed4's vectors can promise 16 (#1477 CS-1)
Some checks failed
PR description / Description follows the template (pull_request_target) Successful in 4s
CI / Conventional Commit title (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been cancelled
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been cancelled
CI / Workflow lint (actionlint) (pull_request) Has been cancelled
Guide screenshots / Guide images and local UI source review (pull_request) Has been cancelled
Third-party parts / As their notes say (pull_request) Has been cancelled
Third-party parts / The same as upstream (pull_request) Has been cancelled
Website / Website on Cloudflare (pull_request) Has been cancelled
5e64d69191
the turn's memory-reclaim storm reproduces on the LCD: 20k pages/s, 28% stall windows
Some checks failed
PR description / Description follows the template (pull_request_target) Successful in 4s
CI / Conventional Commit title (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been cancelled
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been cancelled
CI / Workflow lint (actionlint) (pull_request) Has been cancelled
Guide screenshots / Guide images and local UI source review (pull_request) Has been cancelled
Third-party parts / As their notes say (pull_request) Has been cancelled
Third-party parts / The same as upstream (pull_request) Has been cancelled
Website / Website on Cloudflare (pull_request) Has been cancelled
099de67aa5
omgitsgela changed title from feat(vk): a packed record layout for the mesh fallback's export and reader (#1477 BC-1a) to WIP: feat(vk): a packed record layout for the mesh fallback's export and reader (#1477 BC-1a) 2026-11-18 13:55:53 +00:00
Author
Contributor

Heads-up: the branch now carries one more code commit past the reviewed packed4 tree 195cd15d: 5e64d691 (#1477 CS-1) enforces a 16-byte-aligned records buffer base (create_buffer gains an alignment request; fallback_buffers_of demands and checks it on new and reused buffers) and raises the packed4 vector loads/stores' SPIR-V alignment declaration from 4 to 16, following a delegated arithmetic audit (offsets land at base+16k; the base was unpromised). Docs rows run27-run34 also sit on the branch. I am putting the PR back to WIP: the CS-1 change is not yet Deck-measured, and the earlier review covered only packed4 at 195cd15d. Deck arms (cs1 overlay build running now) + a fresh review of the new diff will restore ready; measurement decides whether CS-1 ships with this PR or as its own follow-up.

Heads-up: the branch now carries one more code commit past the reviewed packed4 tree `195cd15d`: `5e64d691` (#1477 CS-1) enforces a 16-byte-aligned records buffer base (`create_buffer` gains an alignment request; `fallback_buffers_of` demands and checks it on new and reused buffers) and raises the packed4 vector loads/stores' SPIR-V alignment declaration from 4 to 16, following a delegated arithmetic audit (offsets land at base+16k; the base was unpromised). Docs rows run27-run34 also sit on the branch. I am putting the PR back to WIP: the CS-1 change is not yet Deck-measured, and the earlier review covered only packed4 at `195cd15d`. Deck arms (cs1 overlay build running now) + a fresh review of the new diff will restore ready; measurement decides whether CS-1 ships with this PR or as its own follow-up.
vk: ask VMA for the records' 16-byte alignment the way VMA takes it, on a pool (#1477 CS-1)
Some checks failed
PR description / Description follows the template (pull_request_target) Successful in 4s
CI / Conventional Commit title (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been cancelled
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been cancelled
CI / Workflow lint (actionlint) (pull_request) Has been cancelled
Guide screenshots / Guide images and local UI source review (pull_request) Has been cancelled
Third-party parts / As their notes say (pull_request) Has been cancelled
Third-party parts / The same as upstream (pull_request) Has been cancelled
Website / Website on Cloudflare (pull_request) Has been cancelled
3ecd17a0be
VmaAllocationCreateInfo has no alignment member (the pinned VMA 3.3.0,
checked against the upstream header): the first CS-1 commit compiled on
neither machine and the Deck's compiler said so. VmaPoolCreateInfo::
minAllocationAlignment is the documented request, and a pool's blocks
need the device-address flag passed down through pMemoryAllocateNext
(plain allocations get it from vma_init's flag). One pool for the one
alignment ever asked for; no pool, the plain path, vk_draw.cpp's address
check still rejects a crooked base. test_mesh_fallback.py now pins the
Aligned-16 vector declarations and the demand-and-check on the base.
omgitsgela force-pushed perf/1477-bc1-packed-export from 3ecd17a0be
Some checks failed
PR description / Description follows the template (pull_request_target) Successful in 4s
CI / Conventional Commit title (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been cancelled
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been cancelled
CI / Workflow lint (actionlint) (pull_request) Has been cancelled
Guide screenshots / Guide images and local UI source review (pull_request) Has been cancelled
Third-party parts / As their notes say (pull_request) Has been cancelled
Third-party parts / The same as upstream (pull_request) Has been cancelled
Website / Website on Cloudflare (pull_request) Has been cancelled
to 0c7b08e9e3
Some checks failed
PR description / Description follows the template (pull_request_target) Successful in 4s
CI / Conventional Commit title (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been cancelled
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been cancelled
CI / Workflow lint (actionlint) (pull_request) Has been cancelled
Guide screenshots / Guide images and local UI source review (pull_request) Has been cancelled
Third-party parts / As their notes say (pull_request) Has been cancelled
Third-party parts / The same as upstream (pull_request) Has been cancelled
Website / Website on Cloudflare (pull_request) Has been cancelled
2026-11-18 14:18:45 +00:00
Compare
tests: the CS-1 pins read their own file's function, not the whole draw (#1477 CS-1)
Some checks failed
PR description / Description follows the template (pull_request_target) Successful in 5s
CI / Conventional Commit title (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been cancelled
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been cancelled
CI / Workflow lint (actionlint) (pull_request) Has been cancelled
Guide screenshots / Guide images and local UI source review (pull_request) Has been cancelled
Third-party parts / As their notes say (pull_request) Has been cancelled
Third-party parts / The same as upstream (pull_request) Has been cancelled
Website / Website on Cloudflare (pull_request) Has been cancelled
63cb71dbcc
vk: the aligned pool names its memory tier on the pool (#1477 CS-1)
Some checks failed
PR description / Description follows the template (pull_request_target) Successful in 4s
CI / Conventional Commit title (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been cancelled
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been cancelled
CI / Workflow lint (actionlint) (pull_request) Has been cancelled
Guide screenshots / Guide images and local UI source review (pull_request) Has been cancelled
Third-party parts / As their notes say (pull_request) Has been cancelled
Third-party parts / The same as upstream (pull_request) Has been cancelled
Website / Website on Cloudflare (pull_request) Has been cancelled
dbc6c429ae
A block's priority is fixed at vmaCreatePool; vkSetAllocationPriority
reaches allocations, not pool blocks, and the repo puts everything at
the tiers' rest (0.5, memory_priority.h: the driver's own default) —
unlike the images' pools, which already name their tier, an unnamed
pool would have parked these arenas at VMA's 0.0.
vk+tests: an aligned request never travels with a dedicated allocation; the pool mechanism gets pinned (#1477 CS-1)
Some checks failed
PR description / Description follows the template (pull_request_target) Successful in 4s
CI / Conventional Commit title (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been cancelled
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been cancelled
CI / Workflow lint (actionlint) (pull_request) Has been cancelled
Guide screenshots / Guide images and local UI source review (pull_request) Has been cancelled
Third-party parts / As their notes say (pull_request) Has been cancelled
Third-party parts / The same as upstream (pull_request) Has been cancelled
Website / Website on Cloudflare (pull_request) Has been cancelled
e53b99c40b
DEDICATED_MEMORY_BIT and a pool together are VMA's vmaCreateBufferWithPool
assert (13156/13218): a call that asked for both would abort the process
when asserts are on and take the pool silently when off. A dedicated
request skips the pool and the address check alone fail-closes. The test
now also pins the delivery itself (pool + minAllocationAlignment + the
device-address pNext): the first CS-1 attempt carried a request VMA 3.3.0
cannot take, and a Deck compile is a slow way to find that out.
vk: vmaCreateBufferWithAlignment takes the records' alignment request; the pool detour goes (#1477 CS-1)
Some checks failed
PR description / Description follows the template (pull_request_target) Successful in 8s
CI / Conventional Commit title (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been cancelled
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been cancelled
CI / Workflow lint (actionlint) (pull_request) Has been cancelled
Guide screenshots / Guide images and local UI source review (pull_request) Has been cancelled
Third-party parts / As their notes say (pull_request) Has been cancelled
Third-party parts / The same as upstream (pull_request) Has been cancelled
Website / Website on Cloudflare (pull_request) Has been cancelled
41ff29c5b9
The second review of the pool mechanism found it defective: the allocator
already prepends its own device-address flags (vma_init's flag, 12979),
so the pool's pMemoryAllocateNext repeated a structure type in
vkAllocateMemory's chain, and the lazily-made pool raced the parallel
replay's first callers. VMA 3.3.0 has the per-buffer API the docs point
at instead: vmaCreateBufferWithAlignment raises the requirements'
alignment before allocating, no chain, no pool, no priority policy
change, works with dedicated and host-visible alike.

It aligns where the buffer sits inside its block; the block's base is
Vulkan's to choose, which is why vk_draw.cpp's address checks stay. The
test now pins each of the three address guards separately (removing any
one fails: mutation-checked) and the delivery call itself, flattened
against reflow.
tests: the cleanup fixture grows into the alignment cases (#1477 CS-1)
Some checks failed
PR description / Description follows the template (pull_request_target) Successful in 5s
CI / Conventional Commit title (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been cancelled
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been cancelled
CI / Workflow lint (actionlint) (pull_request) Has been cancelled
Guide screenshots / Guide images and local UI source review (pull_request) Has been cancelled
Third-party parts / As their notes say (pull_request) Has been cancelled
Third-party parts / The same as upstream (pull_request) Has been cancelled
Website / Website on Cloudflare (pull_request) Has been cancelled
c63639de9a
Its mock create_buffer took six arguments; the real call now passes eight, and the
fixture extracts the real function verbatim, so it had to grow anyway (first review
round, finding 6). While in there, it becomes the fault-injection home for CS-1: the
mock REQUIRES the records allocation (the one with the indirect-header usage bit) to
demand 16 — the pre-CS-1 source now fails the fixture instead of passing it; a
crooked (4 mod 16) base is injected and must reject clean, destroy the buffer and
recover on the next draw; and a payload growth must share the first base, still
16-aligned. 13 cases, g++ -Werror clean.
tests: the fixture's logf stub stays quiet against the pre-CS-1 source too (#1477 CS-1)
Some checks failed
PR description / Description follows the template (pull_request_target) Successful in 4s
CI / Conventional Commit title (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been cancelled
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been cancelled
CI / Workflow lint (actionlint) (pull_request) Has been cancelled
Guide screenshots / Guide images and local UI source review (pull_request) Has been cancelled
Third-party parts / As their notes say (pull_request) Has been cancelled
Third-party parts / The same as upstream (pull_request) Has been cancelled
Website / Website on Cloudflare (pull_request) Has been cancelled
6a7f22cf45
docs: the TEX_AHEAD rows learn their mechanism correction (#1477)
Some checks failed
PR description / Description follows the template (pull_request_target) Successful in 6s
CI / Conventional Commit title (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been cancelled
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been cancelled
CI / Workflow lint (actionlint) (pull_request) Has been cancelled
Guide screenshots / Guide images and local UI source review (pull_request) Has been cancelled
Third-party parts / As their notes say (pull_request) Has been cancelled
Third-party parts / The same as upstream (pull_request) Has been cancelled
Website / Website on Cloudflare (pull_request) Has been cancelled
7cb2af8633
zeldaootr read main: RX_VK_TEX_AHEAD is the loading-screen pass over the
textures drawn before (90 s rule), not retirement deferral — that is their
unmerged RX_VK_EVICT_IN_USE_S, whose numbers I quoted at the switch in the
run31 post. Both rows now say what the arms actually caught: this route's
first pan is ~44 s after the anchor, inside the 90 s rule, so the storm
shrink is a load-window benefit (first look after load/warp), and the
105 s-pan pair shows no fps change. run31's cost side stands as measured.
docs: CS-1 measures null on the GPU line; the late pan's storm is worse than the early one (#1477 run36, run37)
Some checks failed
PR description / Description follows the template (pull_request_target) Successful in 4s
CI / Conventional Commit title (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been cancelled
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been cancelled
CI / Workflow lint (actionlint) (pull_request) Has been cancelled
Guide screenshots / Guide images and local UI source review (pull_request) Has been cancelled
Third-party parts / As their notes say (pull_request) Has been cancelled
Third-party parts / The same as upstream (pull_request) Has been cancelled
Website / Website on Cloudflare (pull_request) Has been cancelled
c572596c04
docs: run35's counter sampler answers the memory thread — the load ramp compacts hardest, turns churn with headroom (#1477)
Some checks failed
PR description / Description follows the template (pull_request_target) Successful in 5s
CI / Conventional Commit title (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been cancelled
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been cancelled
CI / Workflow lint (actionlint) (pull_request) Has been cancelled
Guide screenshots / Guide images and local UI source review (pull_request) Has been cancelled
Third-party parts / As their notes say (pull_request) Has been cancelled
Third-party parts / The same as upstream (pull_request) Has been cancelled
Website / Website on Cloudflare (pull_request) Has been cancelled
f975a10f6e
docs: the matched #1553 pair is a VanGogh wash, the kernel knobs and late TEX_AHEAD are null, bucketing halves churn but not fps (#1477 run38-run41)
Some checks failed
PR description / Description follows the template (pull_request_target) Successful in 5s
CI / Conventional Commit title (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been cancelled
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been cancelled
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been cancelled
CI / Workflow lint (actionlint) (pull_request) Has been cancelled
Guide screenshots / Guide images and local UI source review (pull_request) Has been cancelled
Third-party parts / As their notes say (pull_request) Has been cancelled
Third-party parts / The same as upstream (pull_request) Has been cancelled
Website / Website on Cloudflare (pull_request) Has been cancelled
611f32fad2
docs: the pool env levers cut churn 2-3x and change nothing — on the LCD at 100% the frame does not pay the allocator (#1477 run42)
Some checks are pending
CI / Conventional Commit title (pull_request) Blocked by required conditions
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Blocked by required conditions
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Blocked by required conditions
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Blocked by required conditions
CI / Workflow lint (actionlint) (pull_request) Blocked by required conditions
Guide screenshots / Guide images and local UI source review (pull_request) Blocked by required conditions
Third-party parts / As their notes say (pull_request) Blocked by required conditions
Third-party parts / The same as upstream (pull_request) Blocked by required conditions
Website / Website on Cloudflare (pull_request) Blocked by required conditions
PR description / Description follows the template (pull_request_target) Successful in 5s
cdef6ae71d
Some checks are pending
CI / Conventional Commit title (pull_request) Blocked by required conditions
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Blocked by required conditions
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Blocked by required conditions
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Blocked by required conditions
CI / Workflow lint (actionlint) (pull_request) Blocked by required conditions
Guide screenshots / Guide images and local UI source review (pull_request) Blocked by required conditions
Third-party parts / As their notes say (pull_request) Blocked by required conditions
Third-party parts / The same as upstream (pull_request) Blocked by required conditions
Website / Website on Cloudflare (pull_request) Blocked by required conditions
PR description / Description follows the template (pull_request_target) Successful in 5s
This pull request is marked as a work in progress.
Some workflows are waiting to be reviewed.
This branch is out-of-date with the base branch
View command line instructions

Checkout

From your project repository, check out a new branch and test the changes.
git fetch -u 'perf/1477-bc1-packed-export:omgitsgela-perf/1477-bc1-packed-export'
git switch 'omgitsgela-perf/1477-bc1-packed-export'
Sign in to join this conversation.
No reviewers
No milestone
No project
No assignees
2 participants
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
MagicBeanDev/Longshot!1561
No description provided.