fabi/Xenia-Canary
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

[x64] Fix AVX-512 vctuxs NaN handling

Browse Source 
(using ordered comparison predicate)
This commit is contained in:
Herman S.
2026-02-12 21:39:53 +09:00
parent f95ebb9c55
commit 70c1092195
1 changed files with 5 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
src/xenia/cpu/backend/x64/x64_seq_vector.cc
View File

@@ -101,13 +101,12 @@ struct VECTOR_CONVERT_F2I
if (i.instr->flags & ARITHMETIC_UNSIGNED) { if (i.instr->flags & ARITHMETIC_UNSIGNED) {
if (e.IsFeatureEnabled(kX64EmitAVX512Ortho)) { if (e.IsFeatureEnabled(kX64EmitAVX512Ortho)) {
Opmask mask = e.k1; Opmask mask = e.k1;
// Mask positive values and unordered values // Mask non-negative, non-NaN values (ordered, >= 0)
// _CMP_NLT_UQ // _CMP_GE_OQ
e.vcmpps(mask, i.src1, e.GetXmmConstPtr(XMMZero), 0x15); e.vcmpps(mask, i.src1, e.GetXmmConstPtr(XMMZero), 0x1D);
// vcvttps2udq will saturate overflowing positive values and unordered // vcvttps2udq will saturate overflowing positive values to UINT_MAX.
// values to UINT_MAX. Mask registers will write zero everywhere // Zero-masking writes zero for negative values and NaN
// else (negative values)
e.vcvttps2udq(i.dest.reg() | mask | e.T_z, i.src1); e.vcvttps2udq(i.dest.reg() | mask | e.T_z, i.src1);
return; return;
} }