Compare commits

...

1 Commits

Author SHA1 Message Date
coah
1a02886692 decomp(jpeg): refine grouped edge chroma declarations 2026-10-03 05:06:46 -05:00
4 changed files with 286 additions and 2 deletions

View File

@@ -284,7 +284,9 @@ static void TMCJPEGDEC_converterYUV411toRGB565edge(TMCCJPEGDecWork* work, s32 x,
s32 yEnd = y + height;
s32 lumaSkip = 32 - width;
s32 chromaSkip = lumaSkip >> 2;
s32 redOffset, greenOffset, blueOffset;
s32 redOffset;
s32 blueOffset;
s32 greenOffset;
ASSERTLINE((x & 3) == 0, __LINE__);
for (; y < yEnd; y++) {
s32 tileRow = (y >> 2) * tileWidth;

View File

@@ -641,7 +641,9 @@ static void TMCJPEGDEC_converterYUV420toRGBA8edge(TMCCJPEGDecWork* work, s32 x,
s32 xEnd = x + width;
s32 lumaSkip = 16 - width;
s32 chromaSkip = lumaSkip >> 1;
s32 redOffset, greenOffset, blueOffset;
s32 redOffset;
s32 blueOffset;
s32 greenOffset;
ASSERTLINE((x & 1) == 0, __LINE__);
for (; y < yEnd; y++) {
u16* output = (u16*)(texture + ((y & 3) << 3));

View File

@@ -0,0 +1,149 @@
# RGB565411 edge grouped offset declarations
Base: `17130b4b` (merged RGB565411 normal improvement), 43U only.
Branch: `agent/bittle/jpeg-rgb411edge-triple`.
Target: `TMCJPEGDEC_converterYUV411toRGB565edge` in
`libs/RVLMiddleware/TMC_JPEG/src/texturecvtr/Texture_MCUtoRGB565.c`.
## Scope and historical coverage
The only source change splits the existing line 287 declaration in place and
orders its three uninitialized `s32` scalars as redOffset, blueOffset,
greenOffset. All assignments, initializers, expressions, types, scopes, helpers,
assertions and traversal remain unchanged. There are no new locals.
These are the three chroma contributions shared by each four-pixel group. The
existing x-alignment assertion guarantees entry at a chroma-group boundary.
The existing branch assigns all three offsets, in its original red/green/blue
assignment order, before the first pixel reads them. No address escapes.
Declaration order changes no arithmetic or observation order.
Current original function text SHA256 (definition through closing brace and
newline):
`ed020f92da0ea0584d9d7fc4ecbcad9f11f8f7621a8829cfcd9fabf135361207`.
This exact function hash also occurs at `180e620c`, `be7ded2f`, `6ec48a1b`
and `cdc22045`; this is not a newly changed expression context.
Historical round4 lines 1538-1544 establish the signed chroma-pointer context
and 95.37736% baseline. Round5 lines 251, 302-306 and 698 record endpoint,
output-pointer and row-template probes. The tex1 log at lines 137-143, 243-246,
346-352 and 520-525 records structural, input-type and lifetime trials. Its
statement that declaration permutations were inert provides no individual
order manifest. The grouped-declaration inventory also cited wider split
searches containing these offsets, but did not establish exhaustive coverage
of the current internal triple with every surrounding declaration fixed.
This trial closes that specific finite measurement gap. Possible historical
order overlap is acknowledged; no claim that these six orders were never
compiled before is made. No wider declaration search or continuation occurred.
## Bounded experiment
1. Preserve prior branch refs; create a fresh branch at merged `17130b4b`.
2. Copy the parent-verified idle main cache, configure the approved local
wrapper/tool overrides, and run ordinary default Ninja, explicit 43U `ok`,
and a fresh full official objdiff report.
3. Compile the same-order red/green/blue split first. The entire object and
full 1027-unit official report are exactly identical to baseline. This
prerequisite passes before any permutation is measured.
4. Enumerate exactly the six permutations of these three existing declarations.
Each source is rebuilt by ordinary Ninja with unchanged compiler arguments;
each call emits the MWCC object-build line and updates the object mtime.
Exact source stability across each compile is checked. Save each source,
object, SHA256, build output and fresh official full report.
5. Select the first highest official-fuzzy order passing the sibling/code/data
gates; then freshly rebuild the selected source with default Ninja and the
explicit 43U `ok` target. No additional variant is tested.
The independent same-order neutrality compile is separate from the six
permutation compiles. The final selected-source rebuild is validation, not an
additional searched order. No compiler call is suppressed or replaced.
| Order | Function fuzzy | Unit fuzzy | Sibling/exact/code/data gates |
| --- | ---: | ---: | --- |
| red, green, blue | 95.37736 | 96.84735 | pass |
| red, blue, green | 95.42453 | 96.850586 | pass; retained |
| green, red, blue | 95.37736 | 96.84735 | pass |
| green, blue, red | 95.42453 | 96.850586 | pass |
| blue, red, green | 95.42453 | 96.850586 | pass |
| blue, green, red | 95.42453 | 96.850586 | pass |
All six raw full reports were compared against baseline. All twelve sibling
function records and raw bodies remain unchanged in every permutation; all
seven existing exact functions remain raw-byte exact and at 100%. Exact code,
data and completion measures do not decrease. No other unit changes. The
six-order space is exhausted in this exact surrounding source context; this
says nothing about broader neighborhoods or eventual exact matching.
## Fresh selected-source gates
- Function: **95.37736% -> 95.42453%**, 424/424 bytes, 106/106 instructions
- Normalized positional ctxdiff: **62 -> 61**
- Unit fuzzy: **96.84735% -> 96.850586%**
- Global fuzzy: **99.727745% -> 99.72777%**
- Exact functions: **7/13**, exact code **2648/6184**, unchanged
- Merged RGB565411 normal gain remains **90.89862%**, raw body unchanged
- All 1027 units and every reported function checked: no regressions
- Fresh selected whole object equals saved permutation-1 object exactly
- Fresh final full report equals saved permutation-1 full report exactly
- Seven exact functions independently checked against original bytes and
normalized disassembly: ctxdiff 0 for each
- Baseline, candidate and original have zero allocated non-code bytes; empty
pool identical. No symbol/data/relocation/configuration change
- Literal audit: 12 functions analyzed, zero candidates/errors; setter skipped
because its source/target sizes differ, as at baseline
- Ordinary default build/progress plus explicit `build/43U/ok` pass
- DOL SHA1: `26116613f624061ba99c8d1a299aaa6efa85670d`
- `git diff --check` passes
## Behavior and ordering
The existing `jpeg_differential_audit.py` passes **9,456 paired cases** across
28 JPEG functions, including preserved GPR14-GPR31, stack pointer and LR.
A focused independent scalar RGB565411-edge model checks **12,689 vectors**
against candidate, original and saved baseline objects, for **38,067 object
executions**:
- 1,257 vectors span every edge width and height at supported scales 1/2/4/8,
including zero extents, partial four-pixel chroma groups, aligned x origins,
tiled boundaries and differing row residues
- 256 vectors exercise both dimension-selection branches independently with
random and constant samples
- 6,400 vectors cover all 256 luminance values for all pairs of signed chroma
values in {-128, -1, 0, 1, 127}, including saturation boundaries and partial
groups across rows
- 4,776 overlap vectors cover all fitting even destination bases for the
specified full/partial groups, pitches and row starts; every store remains
within the actual 0x184-byte conversion-buffer allocation
All non-stack bytes, including guards, agree with the sequential scalar model.
All three objects reach **106/106 instructions**. The overlap oracle reads from
its evolving buffer, preserving read/write observation order; state and
descriptor objects remain separate. These overlap arrangements are additional
machine-level diagnostics, not a new public API alias contract. Input samples
and bounded dimensions keep intermediate arithmetic within signed integer
range. The converter contains only integer arithmetic; no floating-point
reassociation is involved. The unchanged assignment/evaluation order is also
verified by the exact source replacement check.
This is bounded execution in the repository's PPC integer interpreter, not a
formal proof or a Wii hardware run.
## Preserved evidence
`/tmp/jpeg-rgb411edge-triple/` contains baseline source/object/report, frozen
original object, `split-neutral.*`, `permutation-0.*` through
`permutation-5.*`, `manifest.json`, `selection.json`, `trial.py`,
`trial-output.txt`, build logs, final report, `verify.py`, `verification.txt`,
`candidate.ctx`, literal output, `audit.py`, `audit-results.json`, audit output
and the differential result log. All six source/object snapshots are retained.
Baseline whole object SHA256:
`fec999a214f95c1f348632de53cac7672c15d70cbaea9b3566d7a7383a4e1b70`.
Retained whole object SHA256:
`f82e514864af9c85e1857e2fd59bfd976deb0aba1f651662b322730ced490ad6`.
Retained source SHA256:
`8fb0c89b7b0bcb7b0467ca1be3dfcdb70dfa43b1ec085d926d028e156470749a`.
Retained function-text SHA256:
`fff97c26d8afa6d734a253ae3b9aeaa648b3a28b682c0ef98fb8aeb7b07bb5ca`.

View File

@@ -0,0 +1,131 @@
# RGBA420 edge inner offset triple: six official-fuzzy-scored orders
Initial base `180e620c`; branch `agent/bittle/rgba420edge-offset-orders`.
Target `TMCJPEGDEC_converterYUV420toRGBA8edge`, 43U only. The sole editable source
was `s32 redOffset, greenOffset, blueOffset;` at original line 644, inside its
existing scope. No initializer, expression, assignment, type, other declaration,
helper, or scope moved. The final source only splits this declaration in place
and orders the three variables redOffset, blueOffset, greenOffset.
## Why the finite check was not already established
The official 87-build search in `fz20.attempts.md:1030..1054` covered the eight
leading declarations, not this inner triple. Its body preceded both the later
OR-expression change and `e8241734`'s red-owned luminance reconstruction. Wider
historical split/declaration trials mentioning these names do not retain the
six exact current-body, fixed-surroundings orders with official scores. The
read-only grouped inventory likewise states that this coverage was unproven.
No prior exhaustive equivalent finite coverage was found; this does not claim
none of the individual orders had ever appeared in a broader old trial.
Current baseline function SHA256:
`0c7dcbff2b82ffca9a9b1c5cf20396ab8bc6a5191844527a87a9ac13aa2d0b47`.
It is byte-identical at `e8241734` and the initial base. Original whole-source
SHA256: `d4651ae1f7dd68082c84008ecce8b12c8f9d1053bc09ec1c0694acbb114b9cba`.
## Neutral split and exactly six ordinary permutations
A fresh baseline object/full report reproduced 87.458336%, 480 bytes. Splitting
in the same red/green/blue order first produced an entirely byte-identical
object and complete 1,027-unit report. Only after that neutral prerequisite,
exactly six distinct permutations were compiled with ordinary real Ninja.
Each exact source/order/hash and compiler output was saved; every call invoked
MWCC successfully and source bytes were stable across compilation.
All six successful snapshots were scored with official objdiff, without
`--deduplicate`, using a private copy of the current project configuration,
preserving its options/metadata and freezing the original object. Only unit
names and target/base paths changed for the six snapshot units. Raw target
fuzzy selected the winner; batch aggregate measures were not used.
- red, green, blue: 87.458336%
- red, blue, green: 87.5% (retained, first minimal tied winner)
- green, red, blue: 87.458336%
- green, blue, red: 87.5%
- blue, red, green: 87.5%
- blue, green, red: 87.5%
Every order passes sibling-byte/report, exact-sibling, allocated-data, and
unit exact/code/data/link gates. This exhausts all six internal orders with
surrounding declarations fixed. It does not cover moves across other locals,
other scopes, initialized declarations, or other source forms. No further
order trial, converter run, or restart occurred.
## Initial fresh gates
- Function fuzzy: 87.458336% -> 87.5%; unit: 90.58702% -> 90.59006%
- Global fuzzy on initial base: 99.72766% -> 99.72768%
- Instructions stay 120/120; structural/positional score stays (12,77)
- The gain is visible in official fuzzy despite the unchanged coarse score
- Exact stays 2/13; matched code stays 660/6596; link/data totals unchanged
- All 12 sibling bodies unchanged; both YUV400 functions raw-byte exact,
official 100%, ctxdiff zero (76/76 and 89/89 instructions)
- No 1,027-unit report regression; no allocated non-code sections
- Empty pools identical; literal audit has zero candidates/errors, twelve
functions analyzed and unchanged unequal-size setter skipped
- Fresh selected-order rebuild equals its saved snapshot object
- Full default build, explicit progress/report/43U-ok targets, and diff check pass
- DOL SHA1: `26116613f624061ba99c8d1a299aaa6efa85670d`
Global completion remains incomplete; this is partial fuzzy progress only.
Retained function SHA256:
`c06a346c894ec20ff2a228498a58327c09e19c94c3c235180feee962c1e12d43`.
Retained whole-source SHA256:
`b58c82974159e852c29f437365d55fec5330676a775fc1f5e81ce418ef224978`.
## Independent 4:2:0 edge model
The scalar model passes 12,039 vectors / 36,117 object executions against
selected candidate, original, and saved baseline; every one of 120 instructions
is covered in all three. It models Y[16][16], Cb[8][8], Cr[8][8] at offsets
4/260/324, horizontal pixel pairs, absolute output-row-parity vertical reuse,
edge dimension branches, and real sequential reads/stores when input aliases
output. All x origins satisfy the existing even-x assertion.
Coverage: 1,212 complete edge-size/scale/origin cases, 256 dimension-branch
cases, 64 additional row/column-residue cases, 6,400 saturation vectors, and
4,107 halfword-aligned overlaps wholly within the actual 0x184-byte convBuf.
State/descriptors remain separate. Additional odd-row tests bound height so
chroma reads remain in the logical planes. Every byte below the reserved stack
region, including guards, matches; preserved GPRs, SP, and LR pass. Additional
residue/alias cases are machine-level checks, not API-validity claims. These
bounded interpreter tests are not formal equivalence or Wii runtime tests.
The repository differential audit independently passes all 9,456 paired cases.
## Artifacts
`build/43U/rgba420edge-offset-orders/` holds original and split source/object/
full reports, frozen original/config, all six source/object snapshots and
compiler logs, `enumerate.py`, `orders.json`, official scoring project/report,
`results.json`, final fresh gate files, `verify.py`, `audit.py`, and audit output.
No remote writes, linking changes, or previously frozen branch edits.
## Clean rebase and final current-main comparison
Initial validated commit `6d9027a5` was cleanly rebased onto `17130b4b`, producing
source/evidence commit `62a0cc5d`. That base includes the merged RGB565411 gain.
No additional order was tried. The exact base's original grouped RGBA source
was built with the rebased surrounding tree to obtain a fresh baseline, then
the already-selected source was restored and freshly built. Both RGBA baseline
and selected objects remain byte-identical to their pre-rebase counterparts.
The merged RGB565 source is byte-identical to `17130b4b`; its entire baseline
and final unit report is unchanged (unit fuzzy 96.84735%). The fresh 1,027-unit
comparison changes only RGBA420edge's fuzzy score; all other units/functions
and exact/code/data/link measures are unchanged. RGBA420edge remains
87.458336% -> 87.5%; RGBA unit 90.58702% -> 90.59006%. Current-base global fuzzy
is 99.727745% -> 99.72777%. The source remains only the grouped triple split and
red/blue/green order, with no expression, initializer, type, or scope change.
Full default build, explicit progress/report/43U-ok targets, sibling exactness,
pool/literal checks, source/byte checks, and diff checks pass again. DOL SHA1:
`26116613f624061ba99c8d1a299aaa6efa85670d`. The independent model again passes
all 12,039 vectors / 36,117 executions, including 4,107 contained overlaps and
all 120 instructions in all three objects. The repository audit again passes
all 9,456 paired cases. Previous bounded-test limitations still apply.
Current-base artifacts are in `build/43U/rgba420edge-offset-orders-rebased/`,
including fresh baseline/candidate source and objects, complete baseline/final
reports, build/progress logs, `verification.log`, `final.ctx`, exact/pool/literal/
status/completion results, and both independent/repository audit outputs. The
original six-order artifacts remain untouched in the first evidence directory.