Submission draft (supersedes worker drafts) — counts from a second verified copy, native Linux
Proposed venue: their bug-report issue form (PR #12 adds it). Counts and environment only; no game data, no keys, no acquisition links, per CONTRIBUTING.
Proposed title
Full-corpus counts from a second verified copy on native Linux: 25,032 of 26,503 shaders translate clean; the README line reads as all 26,503
Environment (say what you tested, per CONTRIBUTING)
- Native Arch-family Linux (Arch-derived distro), not Windows, not WSL. No GPU stage run; CPU-only.
- Python 3.14.9 (runs with -I, stdlib only), distro LLVM 23.1.5 (llvm-mc assembles gfx1013; the WSL route in the docs was not needed), native spirv-tools (spirv-as / spirv-val), native .NET, submodule fetched via git on Linux.
What was run
- The copy verified against game/xbox-1.0.2.1.json: 108/108 files SHA-256 match on a full read; the executable slot is the manifest's only unfilled entry. (The repo's own check stage confirms the same: "all files present" plus the exe note.)
- Current main (793b06f, post PR #5 and PR #6): shaders stage, carve -> unpack -> disassemble -> translate -> merge, run as two shards as the tool intends.
- A second full run on the commit immediately before PR #5 produced identical translator numbers.
Counts (tool-reported, this copy)
| Stage | Result |
|---|---|
| Carve (shaders.rpf) | 26,503 index rows; 25,734 distinct files; 1,377 names occur more than once; 17,734 files carry the collision suffix; 767 rows share a file (exact-duplicate names overwrite) |
| Unpack | 25,734 outputs; 2 exceptions counted by the tool |
| Disassemble (gfx1013) | 26,503 shaders -> clean 25,030; invalid words 1,473 |
| Models (archives with entries) | 155,598 exported, 0 fail, 0 skip; no layout rejections on this copy (README: "not all layouts are decoded yet") |
| Textures (in-archive entries) | 6,738 exported, 0 fail, 0 skip |
All counts are tool-reported from this machine; intermediate files are prunable and every stage reruns from the copy, so any row here can be re-derived on request. | Translate (2 shards, current main) | shard 0: 13,251 -> 12,527 translated, 12,527 pass spirv-val, 724 "reachable instruction with normalized encoding"; shard 1: 13,252 -> 12,505 / 12,505, 747 same category; merged index 25,032 entries | | Second run (pre-PR-5 commit, 2 shards total) | 26,503 -> 25,032 translated, 25,032 pass spirv-val, 1,471 same category |
The one-line gap
The README claims all 26,503 carve/unpack/disassemble and all 26,503 translate and pass spirv-val with the per-stage split 2,236 VS / 12,101 PS / 2,689 CS / 9,282 mesh / 195 geometry. On this copy the translator leaves 1,471 shaders in the "reachable instruction with normalized encoding" bucket (the tool's own LLVM-skip re-anchor path) and disassembly flags 1,473 as invalid words. Since this copy is hash-identical to the manifest, the difference is presumably between this run and the maintainer's, and the obvious suspects are the name layer (1,377 duplicated names, 767 shared files, truncation at the fssm layer) and the case-insensitive filesystem the documented layout runs on. Offered as observation, not diagnosis.
Also true from this machine (optional second comment, or fold into #7 if reopened)
Native distro LLVM 23.1.5 on Arch-family Linux assembles gfx1013 fine and the native carve/unpack/ disassembly pipeline completes on it end to end; the docs' WSL note is Windows-side only.
What this is NOT
- Not the disasm-path layout bug: current main's PR #5 already fixed that; an earlier draft describing it is superseded.
- Not an MIT-license point: that's their decision (issue #4).
- Not any claim about GPU results: the GPU stage was never run here (no Windows host, and the harness tools load vulkan-1.dll; that harness port is the known gap).
Activity
3Update: full rerun on current main (c432cdc), per the coordinator's request on #634
Same copy, full SHA-256 re-verified before the run via the tool's own check stage: 108/108 match, the exe slot is still the manifest's only unfilled entry.
Environment, per CONTRIBUTING: native Arch-family Linux (CachyOS), not Windows, not WSL; distro LLVM 23.1.5 (llvm-mc, gfx1013); Python 3.14.9 with -I; native spirv-tools; NVIDIA RTX 4080, driver 620.73.04. The tool probed this GPU, saw the smallest maximum subgroup size is 32, and ran translate with wave64 emulation on. No GPU stage was run; everything below is the CPU pipeline.
| Stage | Result |
|---|---|
| Check (full) | 108/108 SHA-256 match |
| Carve (shaders.rpf) | 26,503 shaders carved, 1 candidate rejected |
| Unpack | 26,503 FSSM blobs, problems: none |
| Disassemble (gfx1013) | 26,503 shaders -> clean 25,030; invalid words 1,473, identical to the rows above |
| Translate (--wave64-emulate, 10 shards, 2,650-2,651 per shard) | every shard 100%: 26,503 translated, 26,503 pass spirv-val; merged index 26,503 entries |
The README's line reads as true on this machine on current main: the 1,471 gap reported above is exactly the LLVM 23 source-field fix (#223) the coordinator pointed at. Link to the ask: https://github.com/ItsNotRexed/Leonida/issues/634#issuecomment-7072547702
AI-assisted (Hermes); every number above is direct tool output from this run.
A Windows data point on a later main (43d0015, the same commit for every step): my own 1.0.2.1 copy, LLVM 23.1.3 llvm-mc (gfx1013), SPIRV-Tools v2026.4, Windows 11, --wave64-emulate (RTX 3060, 32-wide), 8 shards + --merge.
- Disassembly: clean 25,030, invalid words 1,473 (same as above).
- The 1,473 shaders with invalid words: all 1,473 translate; none is rejected for a normalized reachable instruction.
- Full corpus: 26,503 translated, 26,501 pass
spirv-val. The two that fail are #881's (The id bound is larger than the max id bound 4194303), not LLVM 23: #900 addresses them.
So with LLVM 23 the README line can read like LLVM 21's once #900 lands; until then both toolchains give 26,501 on current main.
Thanks. These rejections are expected for now. LLVM 23's disassembler normalizes more encodings than LLVM 21 does, for example by dropping default modifiers or choosing a shorter form. The translator re-assembles every instruction and rejects any reachable one whose bytes don't round-trip, so it doesn't silently lose a bit.
To widen
same_encodingsafely I need examples. Could you post, for a few of the 1,471 shaders, the first normalized instruction, i.e. the original dword(s) next to what LLVM 23 prints? The full list of rejected shader names would also help. Until then the README says the full-corpus result applies to LLVM 21.