izzy/xenia-edge
0
0
Fork 0
mirror of https://github.com/izzy2lost/xenia-edge.git synced 2026-07-06 00:20:26 -07:00
Code Issues Packages Projects Releases Wiki Activity

[A64/Vector] Use numeric min/max for D3DCOLOR pack

Browse Source 
- Prefer FMAXNM/FMINNM to avoid NaN propagation differences

- Match x86 and Windows ARM64 saturation behavior
This commit is contained in:
Will Martin
2026-01-21 11:47:25 +09:00
parent 1de36083c8
commit 283f3d7db7
1 changed files with 6 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/xenia/cpu/backend/a64/a64_seq_vector.cc
+6 -2
View File
@@ -1371,11 +1371,15 @@ struct PACK : Sequence<PACK, I<OPCODE_PACK, V128Op, V128Op, V128Op>> {
// are valid - max before min to pack NaN as zero (5454082B is heavily
// affected by the order - packs 0xFFFFFFFF in matrix code to get a 0
// constant).
// Use FMAXNM/FMINNM (numeric max/min) instead of FMAX/FMIN to handle NaN
// correctly. FMAXNM returns the non-NaN value when one operand is NaN,
// matching x86 behavior and Windows ARM64 expectations.
e.LDR(Q0, VConstData, e.GetVConstOffset(V3333));
e.FMAX(i.dest.reg().S4(), i.dest.reg().S4(), Q0.S4());
e.FMAXNM(i.dest.reg().S4(), src.S4(), Q0.S4());
e.LDR(Q0, VConstData, e.GetVConstOffset(VPackD3DCOLORSat));
e.FMIN(i.dest.reg().S4(), src.S4(), Q0.S4());
e.FMINNM(i.dest.reg().S4(), i.dest.reg().S4(), Q0.S4());
// Extract bytes.
// RGBA (XYZW) -> ARGB (WXYZ)
// w = ((src1.uw & 0xFF) << 24) | ((src1.ux & 0xFF) << 16) |