#1380

Coordinating who works on what (6)

OpenKoertLichtendonkopened 3h ago

This is the shared list of who works on what. @ItsNotRexed asked for it to be kept this way: claim work here before starting a PR; they check it when merging. Where two PRs overlap, the one that lands first is merged and the other is asked to rebase or close.

The tables are updated by a script from open PRs, fork branches and the commands below, and checked by @KoertLichtendonk. This issue continues #1218 (and before it #156, #374, #634, #1006), which reached 300 comments; claims made there still count.

Read CONTRIBUTING.md first (and AGENTS.md if you use a coding agent). No game data anywhere: archives, the executable, carved files, shaders in any form (DXBC, FSSM, disassembly, SPIR-V), textures, models, audio, videos, or logs and reports that quote them. No keys or decryption, and nothing online. Share outcomes (counts, pass/fail, error classes), not inputs.

Updated 2026-11-22 12:08 UTC by the coordinator (@KoertLichtendonk); corrections are welcome as comments.

Up for grabs

Unclaimed work, most useful first. Pick one, post claiming: <ID> here, and start. ⭐ = good first task. Size: S = an evening, M = a few days, L = a larger project or several PRs.

IDWhatWhy it mattersNeedsSize
TEST-57GPU results on an Intel Arc (A or B series) (no claim needed, several people can do this) (see #166)the one GPU family with only 128 bytes of push constants, so D3D-14's fallback has never run on real hardwarean Intel Arc GPU; the guide in docs/testing.md; no game filesS
SHADER-67Per-stage code-base slots in the translator and harness (open question 14; follow-up to the E_NOTIMPL in #291)the user-data layout it waited for is settled now that #291 is on main (#1083)translator and runtime knowledge; no game filesM
TEST-217GPU fixture for image atomics in a pixel shader: image_atomic_or/umin/umax on 2D and image_atomic_umax on 3D, with a fragment-stage storage-image binding (found by @aslan2142) (see finding)1,018 of the game's 1,038 shaders with image atomics are pixel shaders (3D umax is the most common form and never in compute); TEST-184's fixtures only run them in computeany Vulkan GPU; no game filesM
FUZZ-40tbuffer_load_format reading a channel the format lacks (e.g. xyzw on 16_16_SNORM): mem_diff's reference repeats a channel, the translator gives 0; measure DST_SEL on RDNA2 to decide which is right (found by @etcetera-yt) (see finding)docs/shader_validation.md lists tbuffer DST_SEL as unmeasured; mem_diff --programs 300 --seed 7 hits 3 such programsan AMD RDNA2 GPU (or the hardware oracle, FUZZ-23); no game filesS
RT-50runtime/src/vulkan_device_filter.h has no unit test. (earlier: @bobcatchris; no code found)the device filter decides which GPU the runtime uses; a regression there picks the wrong one silentlyC++; no GPU or game filesS
TEST-205ps_vcmpx_forms (#1374) fails on Intel UHD 770, 4096/4096 pixels; probably the 16-lane fragment subgroup case #1318 handled for ps_bitmask_perm (found by @kapitche) (see finding)a false failure on Intel GPUs hides real ones; the fix belongs in #1374's test before it mergesan Intel GPU to confirm; no game filesS

Plus 3 unclaimed D3D- items in #313, most of them small and without game files.

Blocked, not up for grabs yet: RT-5 Boot the game executable (waits for a developer with the game executable); RT-22 The game executable's preferred base and whether it has a .reloc directory (gta6rt --dry-run); without relocations the host must leave its base free, which matters under Wine (waits for waits for the game executable (@ayozetr's finding)); TEST-88 Replay valu_fuzz.py --family wave --seed 1 on a 32-wide GPU with #293 applied, and compare the mismatching programs with @VictorJulianiR's 13 predicted indices (waits for @esteban-dcp will run it once #293 is rebased (@simonbronk)).

Ready for review or a GPU test

Open PRs with no conflicts, oldest first, grouped by what a check needs (Win = Windows + MSVC, GPU = a Vulkan GPU, game copy = your own copy of the game). A second check from someone other than the author makes merging faster. Pick one that fits your hardware, then comment on the PR: Tested <commit> on <GPU, driver, OS>: pass (or what failed), with what you ran, or review the code. "checked" lists who already tested or reviewed it; a PR nobody has checked needs you most. CI "not run" usually means a first-time contributor's run waits for maintainer approval.

  • code read (CPU) (30): #379 (BUILD-36; CI not run; checked: @johnatan28), #408 (BUILD-67; CI not run; checked: @kapitche), #762 (PROC-26), #801 (PROC-42; checked: @oVolkra), #803 (BUILD-202; checked: @Dom-in), #971 (BUILD-192; CI running; checked: @kapitche, @UlteriorMotive00), #1412 (PROC-79; CI not run), #1474 (BUILD-426; CI running; checked: @LocranKing), #1477 (BUILD-418; CI not run), #1481 (TEST-208; CI running; checked: @dcastell1), #1485 (BUILD-430; CI running), #1486 (BUILD-418; CI not run), #1487 (SHADER-247; CI running), #1488 (BUILD-429; CI running), #1498 (BUILD-432; CI running), #1500 (D3D-178; CI not run), #1501 (BUILD-427; CI not run; checked: @copilot-pull-request-reviewer), #1503 (BUILD-434; CI running), #1507 (BUILD-436; CI running; checked: @x2Tweek), #1508 (BUILD-435; CI running), #1515 (BUILD-438; CI running), #1520 (BUILD-439; CI running), #1526 (BUILD-441; CI running), #1527 (FUZZ-41; CI not run), #1530 (BUILD-442; CI running), #1532 (BUILD-443; CI running), #1533 (BUILD-444; CI running), #1535 (ASSET-114; CI running), #1538 (BUILD-445; CI running), #1544 (TEST-216; CI running)
  • GPU (20): #319 (TEST-39; CI running; checked: @EchoTD, @tylerstraub), #1447 (FUZZ-35; CI running), #1479 (RT-89; CI running; checked: @bvtterbroot, @dcastell1), #1482 (SHADER-224; CI not run; checked: @tylerstraub, @Vonsek, @kapitche, @copilot-pull-request-reviewer), #1484 (SHADER-249; CI running; checked: @dcastell1), #1490 (SHADER-248; CI running; checked: @Rainmaker33, @kapitche, @dcastell1), #1492 (SHADER-241; CI running; checked: @Rainmaker33, @Dom-in, @kapitche), #1496 (SHADER-250; CI running; checked: @bvtterbroot), #1499 (SHADER-240; CI running; checked: @Rainmaker33, @bvtterbroot), #1505 (TEST-206; CI running; checked: @tylerstraub), #1509 (SHADER-239; CI running; checked: @tylerstraub), #1510 (SHADER-252; CI running; checked: @dcastell1), #1514 (SHADER-254; CI running; checked: @dcastell1), #1518 (TEST-214; CI not run; checked: @tylerstraub, @copilot-pull-request-reviewer), #1521 (SHADER-131; CI running; checked: @Itschotsch), #1525 (SHADER-255; CI not run), #1531 (SHADER-256; CI running; checked: @dcastell1), #1536 (SHADER-258; CI running), #1539 (SHADER-259; CI running), #1543 (SHADER-226; CI running)
  • Win + GPU (9): #899 (D3D-114; CI not run; checked: @UlteriorMotive00), #1324 (D3D-149; CI running; checked: @Rainmaker33, @kapitche, @kotorvr, @Parakeet-Wheat, @swam122yam-glitch), #1478 (SHADER-242; CI running; checked: @whattheflop), #1493 (D3D-172; CI running; checked: @Rainmaker33, @kapitche), #1511 (D3D-171; CI running; checked: @Parakeet-Wheat), #1513 (D3D-174; CI not run), #1519 (D3D-162; CI not run), #1524 (D3D-169; CI running), #1537 (TEST-209; CI running)
  • read (7): #1059 (PROC-49), #1475 (TEST-71; CI not run), #1497 (PROC-80; CI running), #1502 (D3D-56; CI running), #1516 (SHADER-234; CI running), #1534 (ASSET-115; CI running), #1542 (ASSET-107; CI running)
  • Win (3): #802 (RT-51; checked: @ItsJaved, @undeadhuman37, @quackottu), #1517 (RT-95; CI running), #1522 (RT-96; CI running)
  • AMD + 32-wide GPU (2): #1528 (BUILD-440; CI running), #1541 (SHADER-10; CI running)

Needs a rebase on main

Open PRs that conflict with main after the last merge round, by author. Rebasing brings them back into the review queue above.

Claimed for 3+ hours without a PR

Work here goes from claim to PR in minutes (median 6 minutes, 95% within 1.2 hours). At 6 hours without a PR the coordinator asks the claimer once; at 12 hours without a PR, draft PR or status: line the claim lapses and the item goes back to Up for grabs, with a link to the earlier claimer's code if their fork has a branch for it. Holding more than 3 claims without a PR: while you're not active (no comment here and no PR opened or pushed for 3 hours), the claims past your 3 newest get one reminder at 3 hours and lapse at 6.

TEST-55 @slopgobblercobbler (12h), SHADER-65 @Vonsek (11h), SHADER-83 @luisbaker (11h), D3D-74 @luisbaker (10h), SHADER-104 @BetterJdue (10h), D3D-99 @lahlousaad4 (9h), BUILD-139 @stayjuced (9h), D3D-102 @73N1X (9h), ASSET-60 @Vonsek (9h), BUILD-146 @DaniSBR1 (8h), RT-28 @JavierBlanco23 (8h), BUILD-153 @TheHoracio (8h), SHADER-124 @LouisRiverstone (8h), SHADER-125 @kapitche (8h), PROC-28 @fearkov (8h), TEST-114 @kapitche (8h), RT-29 @leozinhoszg (8h), SHADER-132 @shelbykb2 (8h), SHADER-133 @shelbykb2 (8h), SHADER-134 @shelbykb2 (8h), SHADER-135 @shelbykb2 (8h), SHADER-136 @shelbykb2 (8h), SHADER-137 @shelbykb2 (8h), BUILD-161 @shelbykb2 (8h), BUILD-162 @shelbykb2 (8h), BUILD-168 @adonizm (8h), TEST-120 @bobcatchris (7h), ASSET-62 @cxchhh (7h), BUILD-173 @cxchhh (7h), BUILD-174 @cxchhh (7h), BUILD-175 @traeb0n (7h), PROC-34 @leozinhoszg (7h), BUILD-176 @dea7hsmiless (7h), BUILD-177 @bobcatchris (7h), BUILD-179 @traeb0n (7h), RT-43 @bobcatchris (7h), RT-44 @bobcatchris (7h), SHADER-142 @bobcatchris (7h), BUILD-190 @gabrielsteglich (6h), BUILD-199 @JavierBlanco23 (6h), BUILD-201 @JavierBlanco23 (6h), BUILD-234 @JavierBlanco23 (6h), RT-61 @sombraSoft (5h), SHADER-157 @bobcatchris (5h), TEST-144 @ray3242 (5h), ASSET-69 @LocranKing (5h), RT-63 @JavierBlanco23 (5h), RT-65 @JavierBlanco23 (5h), BUILD-250 @toeeshchaudhary (5h), BUILD-252 @toeeshchaudhary (5h), SHADER-165 @InduwaraSMPN (5h), FUZZ-20 @quackottu (5h), BUILD-262 @rafaelpssz (5h), PROC-54 @LucasTayar-MIMO (5h), BUILD-275 @rafaelpssz (4h), BUILD-280 @lucastayaralvarenga-ui (4h), D3D-132 @quackottu (4h), BUILD-281 @lucastayaralvarenga-ui (4h), BUILD-292 @lucastayaralvarenga-ui (4h), SHADER-105 @edupode (4h), BUILD-310 @GennyDotJson (4h), ASSET-86 @Andretuta (4h), ASSET-70 @mesterember104-cpu (3h), SHADER-167 @Vonsek (3h), SHADER-188 @Mia-ld (3h), SHADER-186 @quackottu (3h)

How to help

You haveStart here
Any GPU, Windows or LinuxBuild runtime/ and run runtime/test/shaders/build.py, check.py and scripts/run_gpu_tests.py (CONTRIBUTING.md). Post the result in #166, not as a new issue. If your GPU is already well covered there, testing an open PR from "Ready for review or a GPU test" helps more. No game files needed.
C++ on WindowsA RT- (PE loader, shims) item, or a D3D- item: #313 lists the critical path and every D3D12 method that is still open.
Shader / GPU / compiler knowledgeA SHADER- item, or claim a FUZZ- instruction family and fuzz it against the CPU checks in tests/semantic; post each finding here as a new item before fixing it. A new translator rule goes into the contract table in docs/design.md, with how it was confirmed.
PythonBUILD- and TEST- items. Tests must run the repository's code (the translator or the tool), not a model written inside the test.
Your own copy of the gameCorpus counts, the package-gated tests (LEONIDA_GAME_DIR) and ASSET- items. Run them on your machine and post only outcomes.

Commands

Write these at the start of a line in a comment here. The tables are updated from them. The tables include every comment up to this one. Before you claim, check the comments after it: someone may have claimed the item since.

  • claiming: <ID or new item> before you start. Check the tables and the open PRs first. A claim counts once posted, unless an earlier claim or open PR covers the item. Agents: read this issue's current body (not a cached copy) and run gh pr list -R ItsNotRexed/Leonida --state all --search "<ID>" before claiming; a merged PR means the item is done.
  • status: <ID> <one line> for progress, and to keep a claim alive (a claim with no draft PR and no status: for 12 hours lapses: the item goes back to Up for grabs with a link to your code). A draft PR keeps it too.
  • blocked: <ID> by #N when you wait on another PR, hardware or the game executable.
  • releasing: <ID> if you stop; done: <ID> #PR when the PR is up.
  • finding: <one line> for work you found but won't do yourself: a bug, a missing method, a stale doc, a follow-up of your own PR. Below that line: where (file and line), why it matters, and whether it needs game files or particular hardware. The coordinator checks it against existing items and PRs, then adds it to Up for grabs with your name on it.
  • In the PR: Depends on: #N for every PR yours builds on or conflicts with, and Fixes #N for the issue it fixes.

Open work

Shader translator (RDNA2 → SPIR-V) and its GPU checks

IDItemOwnerPRStatus
SHADER-65Corpus scan: append/consume UAV counters and typed buffer formats the translator doesn't model@Vonsek—results in #374: no UAV counters in any shader; the 5 immediate formats are all in the table; formats read from V# word 3 still open
SHADER-83v_ashrrev_i64 (@luisbaker's claim; the reported failure is most likely TEST-67's bit 22, see #374)@luisbaker—@luisbaker confirmed the test failure is TEST-67; what's left is a type cleanup
SHADER-104lds_fuzz.py --emulate wave64: vk_compute_test crashes (0xC0000005) on 59 of 100 seed-1 programs on an RX 6700 (AMD Windows); RTX 3060 and 5090 on Linux run them all@BetterJdue—@BetterJdue: AMD's shader compiler crashes in vkCreateComputePipelines (amdvlk64.dll); spirv-val accepts the SPIR-V; reducing it next; @whattheflop's Ryzen 9800X3D iGPU (RDNA2, driver 25.10.30) crashes the same way in amdvlk64.dll (33 of 100)
SHADER-105lds_fuzz on Intel (Arc B580 and Xe-LP UHD 770 alike): mismatches beyond ds_bpermute: DS atomics (max_i32/max_u32), ds_cmpst, small reads, and registers that keep their initial value@edupode—claimed in the tracker; @AlxcBTW's merged #1345 (LDS hardware oracle) says it answers this item
SHADER-124spirv_mini_interp.py runs 32-bit atomics on physical-storage-buffer pointers only, so the LDS atomics the translator already emits (ds_add/sub/min/max/and/or/xor, `ds@LouisRiverstone—claimed in the tracker
SHADER-125new SHADER item, the root cause of @BACONMANXDGOD's finding above (Intel ds_permute_b32 / cs_w32_permute case C, and lds_fuzz w32 programs 24 and 26).@kapitche—claimed in the tracker
SHADER-132RDNA2 measurement for #753 (SHADER-35): f16/f64 results under MODE.FP_ROUND toward zero vs nearest, raw kernels on an RX 6800 (gfx1030), wave32 and wave64.@shelbykb2—claimed in the tracker (mentions #753)
SHADER-133RDNA2 measurement of ds_permute_b32 (collisions, unaddressed lanes, EXEC, wave64 halves) for #628/#789, raw kernels on an RX 6800 (gfx1030).@shelbykb2—claimed in the tracker (mentions #628, #789)
SHADER-134RDNA2 measurement for #616 (SHADER-88): v_sin/v_cos f32/f16 at whole negative turns, near and past ±256 turns, specials, RX 6800 (gfx1030).@shelbykb2—claimed in the tracker (mentions #616)
SHADER-135RDNA2 measurement: v_fma_mix* inline constants with op_sel_hi, v_exp/v_log_f16, f16 NaN/denormal operands, and #718's VOP3 clamp/omod words, RX 6800 (gfx1030).@shelbykb2—claimed in the tracker (mentions #718)
SHADER-136RDNA2 measurement of docs/design.md's open f32 cases (v_rcp/v_rsq/v_log/v_sqrt specials, VOP3 clamp/omod on v_cvt_f32_*/v_rcp_iflag), RX 6800 (gfx1030).@shelbykb2—claimed in the tracker
SHADER-137RDNA2 hardware words for the integer rows checked only on the CPU (s_absdiff/s_mul_hi, s_lshl*_add SCC, v_mad_u64/i64 sdst, add/sub_nc clamp, SAD, s_sext/@shelbykb2—claimed in the tracker
SHADER-142#594 — v_sin_f32/v_cos_f32 (and the f16 pair) scale the unreduced turn operand by 2*pi before range reduction.@bobcatchris—claimed in the tracker (mentions #594)
SHADER-157#1032 investigation — pin down whether the ps_gather_sampler_lod minLOD failure on the RX 7900 XTX is in the translator or the AMD driver.@bobcatchris—claimed in the tracker (mentions #1032)
SHADER-165tools/shader_meta.py _split_operands omits read-modify-write SGPR ops (s_cmov_b32/b64, s_bitset0/1_b32/b64) from SGPR_RMW (dropping live user SGPRs from `@InduwaraSMPN—claimed in the tracker
SHADER-167AMD driver 23.20.44 (RDNA2 iGPU): typeless RTV clears read wrong, ms_clip crashes, ps_bary's raw barycentrics differ, and emulated ms_tri leaks 42 GiB while compiling; ours or the driver's?@Vonsek—claimed in the tracker
SHADER-186NaN under the VOP3 clamp bit: hardware gives +0 with DX10_CLAMP=1 (and keeps the NaN with DX10_CLAMP=0), while the translator emits FClamp, whose NaN result is undefined@quackottu—claimed in the tracker
SHADER-188v_cmp_lt_u32_e64 writing to vcc_lo / s[18:19] under a non-full exec mask produces 0x00000000 across all lanes instead of the active-lane comparison bitmask (`cr@Mia-ld—claimed in the tracker; may already be fixed by #1125 (SHADER-172, merged): retest on current main
SHADER-190Find the float mode the game's shaders start in (IEEE, DX10_CLAMP, FP denormals; the console's equivalent of PGM_RSRC1), e.g. in the FSSM header@quackottu—claimed in the tracker
SHADER-195LDS exchange inside one wave without s_barrier: RDNA runs a wave in lockstep, SPIR-V gives no ordering between invocations, so on NVIDIA a lane may read before another lane's write; insert subgroup-scope barriers@AlxcBTW—claimed in the tracker (second item of the same claim; shadPS4 #5285 fixed the same)
SHADER-200Gen.mask_src reads an SGPR holding plain data (not marked by lane_masks.py) as all lanes on; take the per-lane bit (v >> lane) & 1 (the SGPR half; @Vonsek has the immediate half, SHADER-198)@Pipodi—claimed in the tracker
SHADER-207#618, ds_swizzle_b32 from a source lane that isn't running (translator fix plus the ps_bitmask_perm check)@jakeefr—claimed in the tracker (mentions #618)
SHADER-20816-bit integer VOP3 ops with a literal operand zero the destination's high half@jakeefr—claimed in the tracker
SHADER-210corpus count for @fearkov's open question (https://github.com/ItsNotRexed/Leonida/issues/374#issuecomment-7073351907): does exp mrtN follow the SV_Target register, or@yAIndoo—claimed in the tracker
SHADER-213the fix: the translator maps the k-th colour export to the k-th declared SV_Target instead of location N, with a hand-written fixture and the contract-table row in `doc@yAIndoo—claimed in the tracker
SHADER-225v_ldexp_f16, v_frexp_mant_f16, v_frexp_exp_i16_f16 (rejected as Unsupported today; the f16 follow-up of SHADER-89)@dev-ewann—claimed in the tracker
SHADER-2292D image_sample with integer sampled views in pixel shaders under implicit minification, separate from the image_sample_lz fix in #1405.@Vonsek—claimed in the tracker (mentions #1405)
SHADER-231swizzled buffer V#s (WORD1 SWIZZLE_ENABLE bit 31, WORD3 ADD_TID_ENABLE bit 23) in MUBUF buffer_load*/buffer_store*/buffer_atomic*.@Vyral1978—claimed in the tracker
SHADER-232translator ignores the VOP3 clamp bit on v_mad_u32_u24/v_mad_i32_i24, v_mul_u32_u24_e64/v_mul_i32_i24_e64 and v_add/sub/subrev_co(_ci)_u32 (it wraps).@quackottu—claimed in the tracker
SHADER-236the ACCC part of the 29 DXBC shaders in amd_fidelityfx_upscaler_xbox.dll (payload tag FSPX, no DXIL; #864), decoded from the DLL and compared with the game's XBCH@omardev29—claimed in the tracker (mentions #864)
SHADER-246s_quadmask_b32 / s_quadmask_b64 / s_bitreplicate_b64_b32 on scalar data (rejected as unsupported today; a lane-mask operand stays Unsupported).@LouisRiverstone—claimed in the tracker
SHADER-251v_lshlrev_b16 with op_sel: the result half differs on a Radeon 680M (RDNA2) in valu_fuzz --family f16 seed 2 program 163, wave64 (#1506)@S1m0ne1997—claimed in the tracker
SHADER-257Typed-buffer FORMATS table (tools/rdna2spirv.py) has no SINT entries for GFX10 formats 63, 73 and 76 (the 32-bit SINT variants next to the UINT ones)@tylerstraub—claimed in the tracker; @rafaelpssz: 72/73 (32_32_32 UINT/SINT) are missing too, and SINT fits the raw-dword u32 path

Plus 44 with a PR (see the review and rebase lists above).

Translator correctness, one instruction family per fuzzer (claim a family before fuzzing it)

IDItemOwnerPRStatus
FUZZ-20cpu_ref.py _int_src: VOP3 abs/neg on integer sources must clear or flip bit 31, not take an integer abs or negate (v_cndmask_b32_e64)@quackottu—@quackottu has a failing test ready and offered it; asked @BBareth first, since cpu_ref.py is on their list
FUZZ-38crosslane_fuzz.py still accepts a program count below 1 (the last fuzzer that does; the sibling of FUZZ-34)@bobcatchris—claimed in the tracker
FUZZ-39RDNA1/RADV GPU verification sweep (open PRs #1405, #1393, #1365 where no RDNA1 check exists) + 8× V540 fallback/fuzz if time; no code changes@bobcatchris—claimed in the tracker (mentions #1405, #1393, #1365)

Plus 5 with a PR (see the review and rebase lists above).

Runtime: PE loader, shims, the package's own DLLs as tests

IDItemOwnerPRStatus
RT-28render_target_clear_color (runtime/src/vulkan_replay.h): convert the integer formats D3D-60 added (R8G8/R16G16/R16G16B16A16/R32G32 UINT/SINT, R10G10B10A2_UINT with clamping)@JavierBlanco23—claimed in the tracker
RT-29in-process startup trace for gta6rt (--trace FILE)@leozinhoszg—claimed in the tracker
RT-43gta6rt accepts --dry-run together with --run (or --init), so the "no image code runs" promise is broken: with both flags, run is true and the loader calls the e@bobcatchris—claimed in the tracker
RT-44new TEST item (from @StebanSchrader's finding) — runtime/test/vk_texture_test.cpp returns `two_mip ?@bobcatchris—claimed in the tracker
RT-61runtime CMake cleanup (no overlap with #800-#803, #888, #933, or #953):@sombraSoft—claimed in the tracker (mentions #800, #803, #888)
RT-63.github/workflows/runtime.yml ?@JavierBlanco23—claimed in the tracker
RT-65.github/workflows/runtime.yml ?@JavierBlanco23—claimed in the tracker
RT-93follow-up to RT-87 (#1384): LoadLibraryExA/W returns a shim pseudo-handle even for data-file loads.@ianneobot-cyber—claimed in the tracker (mentions #1384)
RT-97Console path mapping in gta6rt: CreateFile*/FindFirstFile*/GetFileAttributes* shims that map GAME: paths to --game-dir and send writes to an overlay; today file calls go straight to the host@HQasper—claimed in the tracker

Plus 13 with a PR (see the review and rebase lists above).

Xbox D3D12 → Vulkan (plan and milestones)

56 items in progress, tracked in #313: the critical path, who builds what, and which D3D12 methods are still open.

Assets and formats: models, textures, audio, movies, archives

IDItemOwnerPRStatus
ASSET-19The 5 format-11 (R16G16B16A16_UNORM) TTPF textures: the element-size table lacks an 8-byte entry@ThatBranModding—waits for breadbbbb/rpf9-explorer#4 to be merged upstream, then a submodule bump
ASSET-20Animation: join rigs to clips and decode one known clip (24-byte header, packed bitstream)@S1m0ne1997—claimed in the tracker
ASSET-60implicit ordering of legacy texture copy writes in Vulkan replay, following the synchronization-validation finding on #634 and the discussion on #741.@Vonsek—claimed in the tracker (mentions #634, #741)
ASSET-62rpf9-explorer, texture dictionaries match the drawable-dictionary header and become model candidates that never recover: 2,608 over the ten archives, including the 1,009@cxchhh—claimed in the tracker
ASSET-6911 drawable dictionaries in core_restofworld are rejected because their drawable has radius 0 (rpf9-explorer)@LocranKing—claimed in the tracker
ASSET-70Find the enclosing resource-root path for the eight ASSET-57 primary-LOD geometry lists (draw ownership, docs/restofworld.md in #924)@mesterember104-cpu—claimed in the tracker
ASSET-86decode the MoVE (version 19) container in anim.rpf (324 streams, 1.49 MB inflated; docs/animation.md maps only the 46-byte header prefix +0x00/+0x04/+0x14/`+0@Andretuta—claimed in the tracker
ASSET-102map lights and the vehicle lane graph in levels.rpf. Neither is documented: docs/world.md covers entity placement but has no light emitters, no vehicle path nodes@cxchhh—claimed in the tracker
ASSET-109audio.rpf: 1,184 streams carry companion tags 0x5a/0xd9 that no doc describes; find what they hold@abilliontokens—claimed in the tracker
ASSET-113rpf9-explorer AwcRecovery.DecodedSampleRate divides 32 kHz Opus by 32000, but the format block's +0x0C count is already 48 kHz frames; audio_manifest.json durations for those streams are 1.5x too long@springlesSpri123—claimed in the tracker

Plus 9 with a PR (see the review and rebase lists above).

Build pipeline (make_game.py, tools/)

IDItemOwnerPRStatus
BUILD-139the reported scripts/build_llvm_mc.ps1 cleanup failure after a successful build.@stayjuced—claimed in the tracker
BUILD-146barycentric fallback for devices without VK_KHR_fragment_shader_barycentric (RDNA1 and older on RADV, Intel Gen9, llvmpipe).@DaniSBR1—claimed in the tracker
BUILD-153ID3D12GraphicsCommandList1: OMSetDepthBounds, SetSamplePositions, SetViewInstanceMask on Vulkan (depth bounds as dynamic state plus the PSO `DepthBoundsTestEnable@TheHoracio—claimed in the tracker
BUILD-161RDNA2 measurement for #806: buffer_load_format_d16 / d16 stores of UNORM/SNORM formats, raw kernels on an RX 6800 (gfx1030).@shelbykb2—the typed-store part was released earlier and measured by @Vonsek on gfx1036 (result); the rest stays with @shelbykb2
BUILD-162RDNA2 measurement of DPP16/DPP8 controls (bound_ctrl, fi, row/bank masks) for #238 (SHADER-23), raw kernels on an RX 6800 (gfx1030).@shelbykb2—claimed in the tracker (mentions #238)
BUILD-168new D3D item — block-compressed CopyTextureRegion between mips whose edge falls inside a block on one side only (a 4x4 box into a 3x3 mip, a 6x6 mip into a 12x12 one) i@adonizm—claimed in the tracker
BUILD-173rpf9-explorer, fragments whose header has null +16/+32/+40 or unequal +48..+54 fields never become scan candidates (LooksLikeFragment).@cxchhh—claimed in the tracker; @sombraSoft counted 71 unrecovered LooksLikeFragment candidates left in levels.rpf at e293917 (finding)
BUILD-174rpf9dump manifest Scan { Candidates, Recovered, NotRecovered, Entries }, so candidates rejected before decoding show up (the #30 review follow-up from Kam471, resubmitt@cxchhh—claimed in the tracker (mentions #30)
BUILD-175the first fix from my PreparedList finding above (#878): VulkanQueue::execute keeps its PreparedList storage between executions instead of building a fresh vector o@traeb0n—claimed in the tracker (mentions #878)
BUILD-176@TheHoracio's vertex-format finding (https://github.com/ItsNotRexed/Leonida/issues/634#issuecomment-7074071972) — add R16_SNORM, R16_SINT, R8G8_SNORM, R8G8_SINT,@dea7hsmiless—claimed in the tracker
BUILD-177#881 — the two largest translated shaders exceed the SPIR-V id bound (4,194,303) since 178a4b4 (#755).@bobcatchris—claimed in the tracker (mentions #881, #755)
BUILD-179the next replay CPU follow-ups from replay_cpu_bench (#878), stacked on #893: no heap allocation per barrier, clear or copy in prepare/emit.@traeb0n—claimed in the tracker (mentions #878, #893)
BUILD-190D3D-20 follow-up — sparse tile mappings for reserved buffers.@gabrielsteglich—claimed in the tracker
BUILD-199@bobcatchris's swizzle_search.py numeric-argument finding (https://github.com/ItsNotRexed/Leonida/issues/634#issuecomment-7074746033) ?@JavierBlanco23—claimed in the tracker
BUILD-201@rafaelpssz's tools/meshlet_page.py finding (https://github.com/ItsNotRexed/Leonida/issues/634#issuecomment-7074553429) ?@JavierBlanco23—claimed in the tracker
BUILD-234bound RPF9 --limit archive scans to [pos, end) in tools/rpf9_carve.py, tools/page_index.py, tools/restofworld_index.py and tools/rbf0_index.py (following #712@JavierBlanco23—mostly on main already: 63f23a0d (merged 05:26 UTC) bounds rpf9_carve, restofworld_index, rbf0_index, anim_index and bvh_index; only page_index.py is left
BUILD-250extend same_encoding() to forgive the plain-VOP3 src2 op_sel_hi[2] bit (dword1 bit 25) that LLVM 23.1.5 re-encodes, which is the cause of the 1,471 LLVM-23 rejections@toeeshchaudhary—claimed in the tracker (mentions #644)
BUILD-252extend same_encoding() to forgive the plain-VOP3 src2 op_sel_hi[2] bit (dword1 bit 25) that LLVM 23.1.5 re-encodes, which is the cause of the 1,471 LLVM-23 rejections@toeeshchaudhary—claimed in the tracker (mentions #644)
BUILD-262one more BUILD item, follow-up of #679 / ASSET-52 (which fixed the 64 KB extent), reproduced with a synthetic header on main 48ad23f:@rafaelpssz—claimed in the tracker (mentions #679)
BUILD-275one more BUILD item, found on my own 1.0.2.1 copy (counts only) and reproduced with a synthetic index on main 48ad23f:@rafaelpssz—claimed in the tracker
BUILD-280bvh_index.py accepts the same destination for --out (JSON) and --csv, then silently overwrites the JSON with CSV.@lucastayaralvarenga-ui—claimed in the tracker
BUILD-281shader_profile.py attempts to read directories ending in .s and aborts instead of profiling the regular .s files.@lucastayaralvarenga-ui—claimed in the tracker
BUILD-292shader_disasm.py counts directories ending in .fssm against --limit and may send them to a worker.@lucastayaralvarenga-ui—claimed in the tracker
BUILD-310Build.bat window title + banner@GennyDotJson—claimed in the tracker
BUILD-336tools/navmesh_decode.py writes its reports with a plain open(), so when --out or --csv names a file inside a folder that does not exist the scan finishes and then@bobcatchris—claimed in the tracker
BUILD-422run the valu_fuzz hardware oracle (#1365) on RDNA2 for both families, on an RDNA2 iGPU (gfx1036, Windows HIP backend) with AMD clang 23 (rocm-sdk-core 10.0.0).@Sanonymous01—claimed in the tracker (mentions #1365)
BUILD-437Apple M2 — diagnose the 14 exact-32-lane MoltenVK skips + exact-bit check of #1407's f16 interp fallback; no code changes@bobcatchris—claimed in the tracker (mentions #1407)

Plus 48 with a PR (see the review and rebase lists above).

Tests, test runners and CI

IDItemOwnerPRStatus
TEST-55test_make_game_submodule: addCleanup instead of enterContext@slopgobblercobblerslopgobblercobbler:fix/submodule-test-entercontextbranch, not claimed
TEST-114the runner half of my cs_barrier_divergent finding above.@kapitche—claimed in the tracker
TEST-120copy_layout back-to-back copy regression test (new item).@bobcatchris—claimed in the tracker (mentions #775)
TEST-144hashdict.py validates each archive index.json and reports a bad one as one ERROR line (#1043), with a synthetic test in tests/test_hashdict.py.@ray3242—claimed in the tracker (mentions #1043, #1040)
TEST-198a ctest that proves Vulkan synchronization validation is active when LEONIDA_VK_VALIDATION=1.@AimLogin—claimed in the tracker
TEST-199RDNA1/RADV 8× V540 validation farm — verify DRI_PRIME isolates one Navi 12 per PCI BDF, then shard a private-corpus pipeline check across all eight cards if a corpus exis@bobcatchris—claimed in the tracker
TEST-212vk_texture_test: the framebuffer readback misses a subpass-to-transfer dependency@Vonsek—claimed in the tracker
TEST-215GPU test cs_w32_salu64: one wave32 lane stores a word of each of s_brev_b64, s_bfe_u64, s_bfe_i64, s_bitset1_b64, s_flbit_i32_b64, s_flbit_i32_i64, `s_cmov_@LouisRiverstone—claimed in the tracker

Plus 35 with a PR (see the review and rebase lists above).

Process and contributor docs

IDItemOwnerPRStatus
PROC-28check.py never runs ps_cube_sample (vk_texture_test cube) or ps_texture_sample_d16, which only run_gpu_tests.py checks.@fearkov—claimed in the tracker
PROC-34new DOC/TOOL item — ImHex patterns for RPF9, RBF0 and the object graph (tools/imhex/*.hexpat)@leozinhoszg—claimed in the tracker
PROC-54docs-only #644 follow-up — update the README shader translation status to cite the reported LLVM 23 reruns and distinguish translation from SPIR-V validation.@LucasTayar-MIMO—claimed in the tracker (mentions #644)
PROC-75docs/design.md has 11 contract rows listed two or three times again; dedupe plus a test@jakeefr—claimed in the tracker; @kotorvr listed seven repeated row titles (finding); @MIGHTYLUV also claimed it as PROC-77

Plus 9 with a PR (see the review and rebase lists above).

Recently done: PROC-70 (#1245 merged), PROC-71 (#1251 merged), PROC-72 (#1286 merged), PROC-73 (#1292 merged), PROC-74 (#1305 merged), PROC-76 (#1389 merged), PROC-77 (#1362 merged), PROC-78 (#1358 merged)

Several PRs for one issue

Open PRs (or fork branches) that fix the same issue. Please compare notes in the issue before both continue; the maintainer picks.

  • none

Merge chains

Open PRs that build on or conflict with each other, in the suggested order (from Depends on lines and agreements here):

  • #469 → #572
  • #521 → #589 → #621
  • #615 → #663
  • #831 → #848
  • #893 → #901
  • #907 → #1489
  • #1500 → #1525

Files changed by more than one open PR

The PR that lands first is merged; the others rebase. Left out: files most PRs in an area touch (README.md, docs/design.md, docs/d3d12_vulkan.md, runtime/CMakeLists.txt, the translator tools/rdna2spirv.py), a person's own PRs, and fork branches from before most merges.

  • runtime/shims/d3d12_vulkan_device.cpp: 43 open PRs and branches (a hot spot: rebase right after each merge)
  • scripts/run_gpu_tests.py: 40 open PRs and branches (a hot spot: rebase right after each merge)
  • runtime/test/shaders/check.py: 37 open PRs and branches (a hot spot: rebase right after each merge)
  • runtime/src/vulkan_replay.h: 32 open PRs and branches (a hot spot: rebase right after each merge)
  • runtime/src/vulkan_backend.h: 30 open PRs and branches (a hot spot: rebase right after each merge)
  • runtime/shims/d3d12_com.cpp: 28 open PRs and branches (a hot spot: rebase right after each merge)
  • runtime/test/vk_test_device.h: 28 open PRs and branches (a hot spot: rebase right after each merge)
  • runtime/shims/d3d12_com.h: 27 open PRs and branches (a hot spot: rebase right after each merge)
  • runtime/src/command_list.h: 27 open PRs and branches (a hot spot: rebase right after each merge)
  • runtime/test/command_list_test.cpp: 24 open PRs and branches (a hot spot: rebase right after each merge)
  • runtime/test/d3d12_state_map_test.cpp: 24 open PRs and branches (a hot spot: rebase right after each merge)
  • runtime/src/shader_library.h: 22 open PRs and branches (a hot spot: rebase right after each merge)
  • runtime/test/d3d12_compute_test.cpp: 22 open PRs and branches (a hot spot: rebase right after each merge)
  • runtime/test/d3d12_draw_test.cpp: 22 open PRs and branches (a hot spot: rebase right after each merge)
  • runtime/test/d3d12_test_shaders.h: 22 open PRs and branches (a hot spot: rebase right after each merge)
  • runtime/test/glsl/cs_user_data.comp: 22 open PRs and branches (a hot spot: rebase right after each merge)
  • runtime/test/glsl/ps_draw.frag: 22 open PRs and branches (a hot spot: rebase right after each merge)
  • runtime/test/glsl/vs_draw.vert: 22 open PRs and branches (a hot spot: rebase right after each merge)
  • runtime/test/shader_library_test.cpp: 22 open PRs and branches (a hot spot: rebase right after each merge)
  • runtime/src/vulkan_present.h: 20 open PRs and branches (a hot spot: rebase right after each merge)
  • runtime/test/d3d12_present_test.cpp: 18 open PRs and branches (a hot spot: rebase right after each merge)
  • tests/test_shader_checks.py: 17 open PRs and branches (a hot spot: rebase right after each merge)
  • tests/semantic/regen_fixtures.py: 16 open PRs and branches (a hot spot: rebase right after each merge)
  • runtime/src/vulkan_memory.h: 14 open PRs and branches (a hot spot: rebase right after each merge)
  • runtime/test/d3d12_vulkan_device_test.cpp: 14 open PRs and branches (a hot spot: rebase right after each merge)
  • runtime/test/glsl/vs_fullscreen.vert: 14 open PRs and branches (a hot spot: rebase right after each merge)
  • runtime/test/d3d12_com_test.cpp: 12 open PRs and branches (a hot spot: rebase right after each merge)
  • tests/semantic/test_value_semantics.py: 10 open PRs and branches (a hot spot: rebase right after each merge)
  • docs/testing.md: #736, #773, #837, #1475, #1497, #1512, #1518, jakeefr:fix/ds-swizzle-missing-lanes-v2
  • runtime/test/vk_pipeline_check.cpp: #367, #436, #664, #773, #1096, #1099, #1482
  • docs/shader_validation.md: #833, #1396, #1447, #1487, #1521
  • runtime/test/d3d12_descriptor_test.cpp: #759, #760, #765, #1325, #1478
  • tests/semantic/README.md: #896, #1299, #1398, #1495, #1525
  • tests/test_run_gpu_tests.py: #615, #663, #773, #1144, #1518
  • tools/mem_diff.py: #833, #1396, #1428, #1435, #1521
  • .github/workflows/runtime.yml: #770, #801, #802, #1096
  • make_game.py: #379, #408, #1101, #1541
  • runtime/test/vk_harness.cpp: #436, #906, #1482, #1541
  • tests/semantic/test_f32_valu.py: #1116, #1281, HQasper:pr-1281-merged, crastinator:f32-specials-main-merge
  • tests/test_translate.py: #664, #1116, #1482, crastinator:f32-specials-main-merge
  • docs/audio.md: #1421, #1449, #1483
  • runtime/src/gfx10_descriptors.h: #693, #760, #773
  • runtime/src/pe_loader.cpp: #335, #395, #1522
  • runtime/test/vk_store_test.cpp: #773, #1402, #1419
  • tests/test_f32_fuzz.py: #1281, HQasper:pr-1281-merged, crastinator:f32-specials-main-merge
  • tests/test_mem_diff.py: #833, #1396, #1521
  • tests/test_rcp_iflag_mods.py: #1281, HQasper:pr-1281-merged, crastinator:f32-specials-main-merge
  • tools/awc_opus.py: #1428, #1449, #1453
  • tools/bink_names.py: #1428, #1503, #1532
  • tools/cf_diff.py: #615, #663, #1428
  • tools/f32_fuzz.py: #1281, #1428, HQasper:pr-1281-merged
  • tools/lds_fuzz.py: #789, #1428, #1447
  • tools/manifest_verifier.py: #582, #1428, #1474
  • tools/root_sig_contract.py: #427, #1428, #1544
  • tools/saves.py: #1428, #1438, #1515
  • tools/ttpf_mips.py: #939, #1428, #1458
  • tools/world_index.py: #971, #1097, #1428
  • .github/workflows/jactionlint.yml: #770, #801
  • docs/models.md: #1534, #1542
  • runtime/shims/d3d12_xs_device.cpp: #558, #1325
  • runtime/test/d3d12_copy_compare.h: #860, #1325
  • runtime/test/shaders/cs_bvh32.asm: #1402, #1419
  • runtime/test/shaders/cs_cvt_pk_u8.asm: #409, #1129
  • runtime/test/shaders/cs_f32_specials.asm: #1116, crastinator:f32-specials-main-merge
  • runtime/test/shaders/cs_imgatomic.asm: #1402, #1419
  • runtime/test/shaders/ps_get_lod.asm: #1402, #1419
  • runtime/test/vk_bindless_test.cpp: #436, #773
  • runtime/test/vk_command_replay_test.cpp: #519, #1324
  • runtime/test/vk_compute_test.cpp: #1419, #1525
  • runtime/test/vk_texture_test.cpp: #436, #1402
  • runtime/test/xbox_present_test.cpp: #1523, #1524
  • tests/semantic/fixtures/disasm/ps_f32_omod_zero.s: #1281, HQasper:pr-1281-merged
  • tests/semantic/spirv_mini_interp.py: #1482, #1525
  • tests/test_awc_opus.py: #1449, #1453
  • tests/test_f64_conversion.py: #1281, HQasper:pr-1281-merged
  • tests/test_lds_fuzz.py: #789, #1447
  • tests/test_manifest_verifier.py: #582, #1474
  • tests/test_root_sig_contract.py: #427, #1544
  • tests/test_saves.py: #1438, #1515
  • tests/test_ttpf_mips.py: #939, #1458
  • tests/test_world_index.py: #971, #1097
  • tools/anim_clips.py: #1428, #1530
  • tools/awc_manifest.py: #1428, #1453
  • tools/bink_index.py: #1428, #1538
  • tools/crosslane_fuzz.py: #1428, #1447
  • tools/fssm_unpack.py: #1101, #1428
  • tools/gamedata_index.py: #1428, #1488
  • tools/gamedata_maps.py: #1428, #1533
  • tools/gen_d3d12_stubs.py: #335, #1428
  • tools/hashdict.py: #1428, #1498
  • tools/image_fuzz.py: #1428, #1527
  • tools/mesh_compute_vs.py: #1428, #1507
  • tools/pe_inspect.py: #1428, #1486
  • tools/ps_interp_modes.py: #1428, #1501
  • tools/psin_index.py: #1428, #1535
  • tools/rpf9_names.py: #1428, #1520
  • tools/rsc8_audio.py: #1428, #1483
  • tools/shader_diff.py: #453, #1428
  • tools/shader_disasm.py: #1101, #1428
  • tools/shader_profile.py: #1428, #1468
  • tools/tex_export.py: #1062, #1428
  • tools/valu_fuzz.py: #1428, #1447
  • tools/world_build.py: #1428, #1508
  • tools/world_census.py: #1428, #1477

Already on main, still open on GitHub

The maintainer merges batches locally, so a PR can stay open after its work landed. Commits pushed to the PR after that are not on main.

  • #833: on main; 2 newer commit(s) on the PR are not
  • #1396: on main; 1 newer commit(s) on the PR are not
  • #1421: on main; 1 newer commit(s) on the PR are not
  • #1428: on main; 1 newer commit(s) on the PR are not
  • #1449: on main; 2 newer commit(s) on the PR are not

Fork branches ahead of main without a PR

  • crastinator:f32-specials-main-merge (3 commits, 2026-11-22): Tests: follow the shared f32 transcendental helpers in the CPU checks
  • springlesSpri123:asset-113-opus-duration (1 commit, 2026-11-22): Rpf9DumpDiagnostics: AWC Opus durations use 48 kHz frames (ASSET-113)
  • HQasper:pr-1281-merged (11 commits, 2026-11-22): Merge remote-tracking branch 'origin/main' into pr-1281
  • aslan2142:pr291-on-main (14 commits, 2026-11-22): Merge main (bb2e184) into runtime/compute-path
  • jakeefr:fix/ds-swizzle-missing-lanes-v2 (1 commit, 2026-11-22): rdna2spirv: ds_swizzle_b32 reads 0 from a source lane that isn't running (#618)
  • jakeefr:fix/design-duplicate-rows-v2 (1 commit, 2026-11-22): design.md: drop the contract rows the merge rounds left twice, and test for it
  • jakeefr:fix/u16-vop3-literal-v2 (1 commit, 2026-11-22): rdna2spirv: a 16-bit integer VOP3 with a literal operand keeps the destination's high half

Plus 126 older branches, more than 50 commits behind main (not listed, 65k body limit).

Decisions for the maintainer

  • CI approval for fork PRs.
  • Tiled resources (D3D-20): which packed-mip split should the layer follow? AMD RDNA2 and RDNA4 agree with each other; NVIDIA (the layer today) and WARP (never packs) differ. The console is RDNA2. @AC-Storm-YT will make the change once decided (#374).
  • CI check is red for every PR and on main since b1b30d8d (@bobcatchris bisected it): the merge took #1396's translator half (a FORMAT=INVALID V# reads zeros) but not its mem_diff reference half, so tests/test_mem_diff.py fails. #1396 is still open with that half at 9582e459; it merges cleanly onto bb2e1842 and the test passes. Merging it turns check green for everyone.

Process

  1. Claim before coding with the commands above (agreed by @ItsNotRexed).
  2. Overlapping PRs: the one that lands first is merged; the other rebases or closes.
  3. GPU test results go in #166, one comment per GPU and driver.

Activity

30
StebanSchrader3h ago

done: hw_oracle setup on 2024 drivers and LLVM 23 (claim) #1373. Tested on a Radeon 780M (gfx1103): 21,454 programs match, 0 unexplained mismatches. CI needs a maintainer's approval (first PR). AI-assisted (Claude Code).

Zebraine3h ago

claiming: include the four generated synthetic SPIR-V files for vk_binary_occlusion and vk_overlapping_occlusion in Runtime CI's runtime-windows artifact. With the same Windows executables under Wine 11.22 + llvmpipe, the current payload passes 0/2 and the payload with these files passes 2/2. This claim is limited to the two occlusion tests. No game files. AI-assisted.

Vonsek3h ago

claiming: 2D image_sample_lz with integer views, continuing my claim from the previous board. A source-built RGBA32 nearest-filter texture on native gfx1036 returned all original UINT and SINT 32-bit channels and saturated the D16 forms. I am checking the full-width translated pixels with mip and channel controls.

etcetera-yt3h ago

status: RT-84 overlap, for the coordinator. #1350 (@EmreO33) and my #1360 make the same fix: game arguments stay wide through process::set_command_line. My claim came first (#1218 7077774302, 10:24) and #1350 was opened first (10:29, mine at 10:33). Whichever you prefer is fine with me, and I'll close #1360 if #1350 is picked.

The one thing #1360 adds is a ctest, run_tiny_game_non_acp_argument, which runs gta6rt --run tiny_game -- hello 42 "α中文" and fails with 0xC0000409 on main's code. If #1350 lands without an end-to-end check like that, it could go in there. AI-assisted (Claude Code).

XFlawless13h ago

claiming: new BUILD item — tools/vertex_types.py and tools/instance_lists.py: --json given one of the input .rpf archives replaces the archive with the report (exit 0); a directory, or a path in a folder that doesn't exist, ends in a traceback after the full scan.

Reproduced on main f85582e with a hand-built archive: the RPF9 file becomes { "streams": 1, ... and the run exits 0. Fix mirrors navmesh_decode's same_report_path + _make_parent (and bvh_index's same_output_path): refuse before the scan, create the parent, one cannot write line on failure. Synthetic tests only, no game files. AI-assisted (Claude Code).

Dom-in3h ago

status: RT item from #1218 (comment 7077955344, GDK modules reach xgameruntime.dll through LoadLibrary/GetProcAddress): PR #1384.

finding + claiming: d3d12_r9g9b9e5_uav (from #1287) fails on main (f85582e2) on an AMD Radeon RX 9070 (RDNA4, 26.11.2) at line 399: (fs.Support1 & D3D12_FORMAT_SUPPORT1_RENDER_TARGET) == 0. This GPU's VK_FORMAT_E5B9G9R9_UFLOAT_PACK32 advertises COLOR_ATTACHMENT and COLOR_ATTACHMENT_BLEND, and CheckFeatureSupport(FORMAT_SUPPORT) in runtime/shims/d3d12_vulkan_device.cpp passes those through as RENDER_TARGET/BLENDABLE. D3D12 never offers R9G9B9E5_SHAREDEXP as a render target, so the layer answers differently from the PC runtime on any GPU whose driver renders to it. Plan: mask RT, blend, MSAA and depth-stencil for that format, and verify on this GPU. No game files. AI-assisted (Claude Code).

XFlawless13h ago

done: new BUILD item (vertex_types/instance_lists --json over an input archive, claim), #1385. 9 new tests fail on main and pass with the fix; Windows 11, Python 3.12, synthetic archives only. AI-assisted (Claude Code).

whattheflop3h ago

done: TEST-187 is on main via #1296, commit 311c37d0. It closes the GPU runner metadata index after decoding. I verified the equivalent merged code and 43 focused Windows tests, then closed the superseded #1307. AI-assisted.

liang-kai-shui3h ago

done: new BUILD item (gamedata_maps.py: wrong-typed index files) #1369. done: new BUILD item (gamedata_graph.py: malformed carve index) #1370. done: new BUILD item (gamedata_index.py: report paths) #1372. done: new BUILD item (gamedata_xml.py: report paths) #1383. done: new PROC item (docs/i18n/README.md: duplicate stale zh-CN row) #1386. AI-assisted.

MIGHTYLUV3h ago

claiming: new BUILD item — tools/page_index.py: protect inputs and carve manifests against destination overwrite, and create missing parent directories for the output report.

Reproduced on main f85582e with synthetic fixtures (CPU only, no game files):

  • <carve> <carve>/index.json overwrites the carve manifest from rpf9_carve.py with exit 0.
  • Specifying an output in a new directory (<carve> new_dir/out.json) errors with new_dir is not a folder instead of creating the parent folder.
  • No same_output_path() check exists between source and out (catching hard links, case variations, or identical paths).

Plan:

  • Add same_output_path(first, second) catching identical paths and hard links.
  • Refuse an out path that overwrites the source archive, any scanned file, or <carve>/index.json.
  • Create missing parent directories via os.makedirs(folder, exist_ok=True).
  • Add unit tests covering each case with synthetic fixtures in tests/test_page_index.py. Synthetic fixtures only, no game files. AI-assisted.
Vonsek3h ago

done: SHADER-215 is in #1396, from main c2967a5e. Clearing only the V# format bits previously caused 2,544 incorrect pixels in a separate synthetic D3D12 guest on each GPU. Both bound and unbound cases now match their CPU images on RTX 5080 and Radeon.

The SHADER-212 GPU fixture follow-up is in #1394. The filtered cs_store run checks both BGRA words on the merged production implementation. AI-assisted.

bobcatchris3h ago

done: new BUILD item (tools/shader_diff.py runs zero programs and reports a clean pass, claim) — #1397. --programs, --limit and --max-ops now require at least 1: one argparse line naming the option and exit 2, instead of 0 programs ... 0 mismatch with exit 0 (and instead of ValueError: empty range in randrange(1, 1) for --max-ops 0). tests/test_shader_diff_exit.py gains 8 tests: 22 passed across both shader_diff suites here, and the negative control against main f85582e2 fails exactly the 6 new refusal tests (6 failed, 4 passed). Defaults, output format and the 2/1/0 exit codes unchanged. CPU only, Windows 11, CPython 3.12.13, no GPU and no game data. Checked before claiming: gh pr list --state all --search shader_diff and the open-PR list have nothing touching tools/shader_diff.py.

MIGHTYLUV3h ago

claiming: new BUILD item — tools/story_index.py: protect inputs and carve manifests against destination overwrite.

Reproduced on main f85582e with synthetic fixtures (CPU only, no game files):

  • <carve> <carve>/index.json overwrites the carve manifest from rpf9_carve.py with exit 0.
  • No same_output_path() check exists between source and out (catching hard links, case variations, or identical paths).

Plan:

  • Add same_output_path(first, second) catching identical paths and hard links.
  • Refuse an out path that overwrites the source archive, any scanned file, or <carve>/index.json.
  • Add unit tests covering each case with synthetic fixtures in tests/test_story_index.py. Synthetic fixtures only, no game files. AI-assisted.
ezwinxdlmao3h ago

done: SHADER-217 #1398 (Depends on: #1334)

tbuffer_store_format_* and its d16 forms translate with the instruction's format. Every expected word of the GPU test was measured natively on a Radeon 680M, and the test passes on an RTX 3050 and the 680M. A FORMAT-0 V# disables the store, as measured; @Vonsek and @opit80 cover the loads and raw accesses.

#1371 (@shadshuvo) makes the same translator change; the PR says what this one adds.

AI-assisted (Claude Code).

abilliontokens3h ago

claiming: new D3D item — RSSetDepthBias (ID3D12GraphicsCommandList9, the console RSSetDepthBiasX's public form per docs/d3d12_xs_surface.md §5.1) as dynamic depth-bias state on the Vulkan layer (vkCmdSetDepthBias), stacked on #291 (Depends on: #291). IASetIndexBufferStripCutValue (the other GCL9 method) stays with D3D-57's strip-cut work. No game files; tests on Windows + Vulkan GPU with a host-D3D12 comparison.

etcetera-yt3h ago

@Vonsek @KoertLichtendonk @Dom-in Yes, that split matches what I see. #1368 doesn't choose an RTV behaviour. It only skips the WARP reference when the local WARP itself lacks typed R9G9B9E5 UAVs, and it can sit beside either #1375 or #1388.

A data point for the RTV question, from native D3D12 on Windows 10 19045 (a throwaway probe, CheckFeatureSupport(FORMAT_SUPPORT) and a 1x ALLOW_RENDER_TARGET CreateCommittedResource for R9G9B9E5_SHAREDEXP):

adapterSupport1Support2RENDER_TARGETTYPED_UAV1x RT texture
NVIDIA GeForce RTX 5060 Ti (driver 621.14)0x008013f00x00000200nono0x80070057
WARP0x008013f00x00000200nono0x80070057

So on this Windows 10 runtime the hardware adapter reports neither, the same as WARP (and as your RTX 5080 rejecting RTV creation). The layer's R9G9B9E5 UAV mode here is the emulated mutable R32_UINT (Support1=0x020003f1). That fits your device-dependent proposal: RTVs only where the device (and runtime) offers them. AI-assisted (Claude Code).

MIGHTYLUV3h ago

done: new BUILD item (story_index.py: report paths, claim) #1401. Refuses report paths matching the source archive, hard-linked aliases, the carve manifest (<carve>/index.json), or input payloads. 23/23 tests pass on Windows 11 (3 new tests). Synthetic fixtures only, no game files. AI-assisted.

nikolas-sturm3h ago

claiming: name the Bink 2 streams from the mount lists. New stdlib tool tools/bink_names.py aligns each pack's CDataFileMgr__ContentChangeSetDef movie list with the bink_index.py streams (the archive stores them in sorted-list order; entries held by other packs are skipped), with a per-row confidence. Resolves the "no stream has a name" item in docs/movies.md. Synthetic tests only, no game data; PR follows shortly.

henriquevinagre3h ago

done: TEST-184 #1402

XFlawless13h ago

claiming: new BUILD item — tools/awc_manifest.py: refuse a report destination that replaces an input AWC bank.

Reproduced on main c2967a5e, using a hand-built bank only: python -I tools/awc_manifest.py synthetic.awc --out synthetic.awc exits 0, prints 1 banks, 3 streams, and replaces the bank with CSV. The same risk applies to a bank found recursively below a directory source.

I will check the destination against the actual input-bank list before reading banks or writing a report, using realpath/normcase and samefile like the existing report tools. Tests will cover direct paths, symlink/hard-link aliases, recursively scanned banks, malformed banks, and normal CSV/JSON reports. Checked the current coordination body and comments plus all-state awc_manifest PR search; no matching claim or PR. CPU only; synthetic fixtures, no game files. AI-assisted (Codex).