WeirdUtils

143 Commits 12 Branches 16 Tags

3 Commits

Include renames

Author	SHA1	Message	Date
MarcelineVQ	0671ffce02	math_sse: SIMD optimizations for planeNormal, transformAABox, and others Added V4 vector helpers (loadV3, storeV3, dot3) and rewrote several functions to use @Vector(4, f32) operations instead of scalar f64: - planeNormal: 0.7x -> 1.7x (79 -> 33 cyc) -- V4 cross + normalize - transformAABox: 0.7x -> 1.2x (118 -> 69 cyc) -- f32 + @min/@max - evaluatePolynomial: 0.4x -> 0.6x (24 -> 16 cyc) -- drop f64 promotion - vec3MulScalar: 0.7x -> 0.8x -- V4 splat multiply - dotProduct/squaredMagnitude: rewritten with dot3 helper crossProduct reverted from shuffle-SIMD back to scalar -- shuffles added latency that the x87 pipeline doesn't have (16 vs 13 cyc). Remaining losers (dotProduct 0.4x, evalPoly 0.6x) are at the function call overhead floor -- the original x87 is 5-11 cycles, which is close to bare call/ret cost.	2026-03-14 21:57:14 -07:00
MarcelineVQ	8307e525d9	math_sse: fix calling conventions from Ghidra disassembly verification Hooks 1,3 (vecMulMat4, quatMulMat4): thiscall->fastcall. Assembly confirms EDX is read as a parameter (FMUL [EDX+...]) before any write. Detour types TC2r->FC3r, AB wrappers abTC2r->abFC3r. Hook 15 (0x699330): removed entirely. Was misidentified as vectorNormalize but Ghidra shows it's a vec3 componentwise >= comparison returning u32. Silicon-only function (not UnitXP), stub already in silicon.zig. Our normalize implementation would have silently corrupted comparison results. Also fixed comment accuracy: hook 8 RET 0x4 (not plain RET), hook 14 __thiscall (not __fastcall). Bitmask selection replaces MATH_TEST_HOOK single-select with proper bit flags. 17 hooks total, mask 0x77FFE.	2026-03-14 18:44:22 -07:00
MarcelineVQ	2b9b5bc043	bone_sse: assembly-verified reimplementation of transformMatrix4x4 13 bugs fixed by comparing against full assembly dump (5317 instructions): - Emitter check: this+0x188 -> this+0x1D8 - Animation time: added FILDtime_scale pattern for both primary (+0xB0) and secondary (+0xDC) slots - Conditional multiply: bone_local = *(bone_rt+0xF0) was missing - Billboard post-processing: 4 switch cases (types 8/16/32/64) implemented - Color animation loop bound: model_hdr+0x64 -> +0x6C - Bone keyframe data stride: 0x24 -> 0x54 - Ribbon emitter output stride: 0x15C -> 0x170 - Particle data/output strides: 0x1FC/0x17C -> 0x1F8/0x16C - Child SceneObject offsets: attach_idx +0x184->+0x1D4, next +0x190->+0x1E4 - Root bone parent: identity -> this+0xFC New files: - BONE_SSE_PROGRESS.md: section-by-section verification status - t44_full_asm.txt: complete function assembly (ground truth) - t44_helpers_asm.txt: all 12 helper function assemblies - math_sse.zig: 18 x87->SSE polyfill stubs (VanillaFixes integration) Also: OnWorldUpdate hook for true per-frame counting, DUMP_FRAMES=450. SSE dispatch currently disabled while particle sections are being verified.	2026-03-14 17:34:37 -07:00

Author

SHA1

Message

Date

MarcelineVQ

0671ffce02

math_sse: SIMD optimizations for planeNormal, transformAABox, and others

Added V4 vector helpers (loadV3, storeV3, dot3) and rewrote several
functions to use @Vector(4, f32) operations instead of scalar f64:

- planeNormal: 0.7x -> 1.7x (79 -> 33 cyc) -- V4 cross + normalize
- transformAABox: 0.7x -> 1.2x (118 -> 69 cyc) -- f32 + @min/@max
- evaluatePolynomial: 0.4x -> 0.6x (24 -> 16 cyc) -- drop f64 promotion
- vec3MulScalar: 0.7x -> 0.8x -- V4 splat multiply
- dotProduct/squaredMagnitude: rewritten with dot3 helper

crossProduct reverted from shuffle-SIMD back to scalar -- shuffles added
latency that the x87 pipeline doesn't have (16 vs 13 cyc).

Remaining losers (dotProduct 0.4x, evalPoly 0.6x) are at the function
call overhead floor -- the original x87 is 5-11 cycles, which is close
to bare call/ret cost.

2026-03-14 21:57:14 -07:00

MarcelineVQ

8307e525d9

math_sse: fix calling conventions from Ghidra disassembly verification

Hooks 1,3 (vecMulMat4, quatMulMat4): thiscall->fastcall. Assembly
confirms EDX is read as a parameter (FMUL [EDX+...]) before any write.
Detour types TC2r->FC3r, AB wrappers abTC2r->abFC3r.

Hook 15 (0x699330): removed entirely. Was misidentified as vectorNormalize
but Ghidra shows it's a vec3 componentwise >= comparison returning u32.
Silicon-only function (not UnitXP), stub already in silicon.zig. Our
normalize implementation would have silently corrupted comparison results.

Also fixed comment accuracy: hook 8 RET 0x4 (not plain RET), hook 14
__thiscall (not __fastcall). Bitmask selection replaces MATH_TEST_HOOK
single-select with proper bit flags. 17 hooks total, mask 0x77FFE.

2026-03-14 18:44:22 -07:00

MarcelineVQ

2b9b5bc043

bone_sse: assembly-verified reimplementation of transformMatrix4x4

13 bugs fixed by comparing against full assembly dump (5317 instructions):
- Emitter check: this+0x188 -> this+0x1D8
- Animation time: added FILD*time_scale pattern for both primary (+0xB0)
  and secondary (+0xDC) slots
- Conditional multiply: bone_local *= *(bone_rt+0xF0) was missing
- Billboard post-processing: 4 switch cases (types 8/16/32/64) implemented
- Color animation loop bound: model_hdr+0x64 -> +0x6C
- Bone keyframe data stride: 0x24 -> 0x54
- Ribbon emitter output stride: 0x15C -> 0x170
- Particle data/output strides: 0x1FC/0x17C -> 0x1F8/0x16C
- Child SceneObject offsets: attach_idx +0x184->+0x1D4, next +0x190->+0x1E4
- Root bone parent: identity -> this+0xFC

New files:
- BONE_SSE_PROGRESS.md: section-by-section verification status
- t44_full_asm.txt: complete function assembly (ground truth)
- t44_helpers_asm.txt: all 12 helper function assemblies
- math_sse.zig: 18 x87->SSE polyfill stubs (VanillaFixes integration)

Also: OnWorldUpdate hook for true per-frame counting, DUMP_FRAMES=450.
SSE dispatch currently disabled while particle sections are being verified.

2026-03-14 17:34:37 -07:00