devtools/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2026-04-10 11:34:33 +08:00
Code Issues Packages Projects Releases Wiki Activity

Fix RowMajor gemm_pack_lhs for backends without half/quarter packets

Browse Source 
libeigen/eigen!2344

Co-authored-by: Rasmus Munk Larsen <rmlarsen@gmail.com>
This commit is contained in:
Rasmus Munk Larsen
2026-03-23 23:33:42 -07:00
parent f574cb9b18
commit 002229ce47
1 changed files with 8 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
Eigen/src/Core/products/GeneralBlockPanelKernel.h
View File

@@ -2050,12 +2050,15 @@ EIGEN_DONT_INLINE void gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, Pa
// pack those half ones until they match the number expected on the
// last peeling loop at this point (for the rhs).
//
// When there are no half/quarter packet types, the interleaved
// groups cannot be consumed by the GEBP micro-kernel's half/quarter
// loops. Fall through to the scalar loop instead.
if (Pack2 < PacketSize && !gone_last && (HasHalf || HasQuarter)) {
// When there are no half/quarter packet types (HasHalf and HasQuarter
// are both false), last_lhs_progress can exceed Pack2, producing
// interleaved groups that the GEBP micro-kernel cannot consume. In
// that case we use exactly Pack2 rows per group so the kernel's main
// loop (which reads Pack2 = LhsProgress values via ploaddup) can
// handle them; remaining rows fall through to the scalar loop below.
if (Pack2 < PacketSize && !gone_last) {
gone_last = true;
psize = pack = left & ~1;
psize = pack = (HasHalf || HasQuarter) ? (left & ~1) : Pack2;
}
}
}