rayTriIntersectIndexedInt (0x7C2C40): SSE Moller-Trumbore with deferred
divide, matching original's epsilon thresholds. Parity-tested against
original for edge hits, backfaces, parallel rays, and per-triangle t/uv.
JMP-patched in weirdperformance.
Added transform44 profiling hooks for ray_tri_indexed_int (0x7C2C40)
and ProcessStaticObjectsCulling (0x683BF0).
Bench: added rayTriIndexedInt bench with exhaustive parity tests.
PerformSpatialCulling (0x6B8C60): Zig rewrite with SSE outcode
computation. 1.4x speedup, JMP-patched in weirdperformance.
performCollisionDetection (0x6B88E0): fully inlined SSE Moller-Trumbore
ray-triangle intersection, eliminating 4 SetVector3 calls and the
ray_tri function pointer call per triangle. ~22 cyc/tri in bench.
Both graduated from transform44 A/B testing to production JMP patches.
entity_sse.zig: reimplementations of UpdateEntityAndChunksPositions and
updateEntitiesInBounds (A/B tested, 1.2x bench, not shipped - memory
bound with negligible real-world gain).
clickthrough: fixed CheckObjectTypePermissions hook from fastcall to
thiscall (ECX preservation), fixed ClntObjMgrObjectPtr from fastcall(4)
to fastcall(5) with correct arg count.
bench: added performCollisionDetection bench with synthetic mesh data
and patched FindOrCreateHashEntry stub.
- Ghidra C decompilation of RenderParticleSprites (422 lines) and 5
helper functions (calculateColorValues, matVec3Transform, etc.)
- particle_sse.zig with calcColorValues_SSE (10.7x bench but cache-miss
bound in-game — needs inlining into full function replacement)
- Bench harness for calcColorValues with correctness check
- build.zig: particle_sse as separate ReleaseFast compilation unit
- colorDetour reverted to pass-through (SSE has no in-game effect due
to L1 cache misses on scattered ColorCtx structs)
- bone_sse: deduplicate findInterpIdx calls in bone loop — rotation's
search result reused for scale/translation when tracks share temporal
structure (canReuseInterp guard). Est. ~23% bone loop cycle reduction.
- silicon: SSE replacement for processLinkedListCollision (0x6ABC40,
1.57% CPU). V4 AABB overlap test replaces 6 x87 FCOMP/FNSTSW.
Benched at 3.2x speedup (378→115 cyc/call, 8 nodes).
- transform44: remove A/B toggle, always use SSE path (teardown guard
kept). A/B infrastructure remains for other hooks.
- bench: add processLinkedListCollision benchmark with fake linked list
test fixture and stubbed addGeometryToBuffer.
si_ftol: 9-byte naked asm using FISTTP (SSE3 truncate-from-x87) replaces
the 39-byte FSTCW/FLDCW/FISTP rounding mode dance. 4 vs 7 cycles (1.7x).
13.2M calls/7.5s in-game -- ~13ms savings per period.
Benchmark: patch-in-place at mapped 0x40A2B0, test parity across 19 values,
best-of-5 timing with varying inputs. Framework for all silicon functions.
Also disabled h67 (ConvertPixelsToScreenAlt) probe -- game passes ECX=0
as valid input, thiscall probe crashes on null this.
mulMat3x4 / mulMat3x4InPlace: translation row had A*B operands swapped.
Original computes A_translation * B_rotation + B_translation, but our
code was doing A_rotation * B_translation + A_translation. Verified by
tracing x87 disassembly: first element loads A[9]*B[col] pattern.
Fixed in both silicon_sse.zig and silicon.zig.
packParticleColor: x87 rounds 127.5 to 128 (round-to-nearest), but SSE
@intFromFloat truncates to 127. Added @round() before @intFromFloat.
42/42 benchmarks now pass correctness. 0 MISMATCHes.
Inlined benchmarks use x87 inline asm vs direct Zig @Vector code with
no CALL/RET on either side. Shows the true instruction-level comparison
that in-place patching would achieve:
dotProduct: x87=3 SSE=1 -> 3.0x (was 0.3x when called)
evalPoly: x87=4 SSE=1 -> 4.0x (was 0.7x when called)
vec3MulScal: x87=3 SSE=2 -> 1.5x (was 0.8x when called)
Every "loser" from the called benchmarks flips to a winner when inlined.
The entire performance gap was function call overhead (~5 cycles), not
instruction quality. Confirms in-place patching as the right strategy.
Also added RADV_TEX_ANISO env var exploration item to release notes and
updated math polyfill section with full benchmark breakdown.
In-place functions (vec3MulAssign, scaleByVec, etc.) were showing fake
speedups (121x, 10x) because repeated application on the same data
caused values to overflow to Inf/NaN. x87 FPU traps on denormals while
SSE handles them in hardware, making the comparison meaningless.
Now each iteration resets from a template copy. Real results show these
functions are ~1.0x (neutral), not 100x wins. The actual winners are
rotMat3x3/4x4 at 2.1x, and several functions where our scalar-through-
pointers approach is slower than the original x87 pipeline (dotProduct
0.3x, evaluatePolynomial 0.4x) — candidates for real SIMD optimization.
Extracts original x87 FPU bytes from WoW.exe via Ghidra, mmaps them
executable, and benchmarks against our SSE replacements. Covers all 17
UnitXP polyfill functions with correctness validation and cycle counts.
Maps a page at 0x7ff000 for the float 1.0 constant referenced by
rotMat3x3/rotMat4x4/planeNormal via absolute address 0x7ff9d8.
Build: zig build bench / zig build run-bench
MarcelineVQ