Lines Matching refs:Unaligned
34 * - otherwise perform unaligned loads only (-> NoneAligned case)
40 * - perform half aligned and half unaligned loads (-> EvenAligned case)
41 * - otherwise perform unaligned loads only (-> NoneAligned case)
53 * One might also wonder why in the EvenAligned case we perform unaligned loads instead of using the aligned-loads plus re-alignment
54 * strategy as in the FirstAligned case. The reason is that we observed that unaligned loads on a 8 byte boundary are not too slow
55 * compared to unaligned loads on a 4 byte boundary.
152 // Currently, it seems to be better to perform unaligned loads anyway
204 // process initial unaligned coeffs
223 _EIGEN_ACCUMULATE_PACKETS(Aligned,Unaligned,Aligned);
262 _EIGEN_ACCUMULATE_PACKETS(Aligned,Unaligned,Unaligned);
267 _EIGEN_ACCUMULATE_PACKETS(Unaligned,Unaligned,Unaligned);
296 // process first unaligned result's coeffs
305 pstore(&res[i], pcj.pmadd(lhs0.template load<LhsPacket, Unaligned>(i), ptmp0, pload<ResPacket>(&res[i])));
480 // process initial unaligned coeffs
499 _EIGEN_ACCUMULATE_PACKETS(Aligned,Unaligned,Aligned);
507 * the overhead of unaligned loads. Moreover unaligned loads are handled
509 * overlaping the desired unaligned packet. This is *much* more efficient
510 * than basic unaligned loads.
540 _EIGEN_ACCUMULATE_PACKETS(Aligned,Unaligned,Unaligned);
545 _EIGEN_ACCUMULATE_PACKETS(Unaligned,Unaligned,Unaligned);
579 // process first unaligned result's coeffs
592 ptmp0 = pcj.pmadd(lhs0.template load<LhsPacket, Unaligned>(j), rhs.getVectorMapper(j, 0).template load<RhsPacket, Aligned>(0), ptmp0);
