Home | History | Annotate | Download | only in meta

Lines Matching refs:lhs_offset

8604                            std::int32_t lhs_offset, std::int32_t rhs_offset,
8624   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
8633     zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
8659                            std::int32_t lhs_offset, std::int32_t rhs_offset,
8679   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
8688     zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
8714                            std::int32_t lhs_offset, std::int32_t rhs_offset,
8734   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
8743     zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
8769                            std::int32_t lhs_offset, std::int32_t rhs_offset,
8789   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
8798     zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
8824                            std::int32_t lhs_offset, std::int32_t rhs_offset,
8844   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
8853     zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
8879                            std::int32_t lhs_offset, std::int32_t rhs_offset,
8899   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
8908     zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
8934                            std::int32_t lhs_offset, std::int32_t rhs_offset,
8954   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
8963     zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
8989                            std::int32_t lhs_offset, std::int32_t rhs_offset,
9009   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
9018     zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
9044                            std::int32_t lhs_offset, std::int32_t rhs_offset,
9064   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
9073     zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
9077   zip_1x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
9102                            std::int32_t lhs_offset, std::int32_t rhs_offset,
9122   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
9131     zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
9135   zip_1x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
9160                            std::int32_t lhs_offset, std::int32_t rhs_offset,
9180   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
9189     zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
9193   zip_1x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
9218                            std::int32_t lhs_offset, std::int32_t rhs_offset,
9238   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
9247     zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
9251   zip_1x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
9276                            std::int32_t lhs_offset, std::int32_t rhs_offset,
9296   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
9305     zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
9309   zip_1x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
9334                            std::int32_t lhs_offset, std::int32_t rhs_offset,
9354   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
9363     zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
9367   zip_1x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
9392                            std::int32_t lhs_offset, std::int32_t rhs_offset,
9412   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
9421     zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
9425   zip_1x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
9450                            std::int32_t lhs_offset, std::int32_t rhs_offset,
9470   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
9479     zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
9483   zip_1x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
9508                            std::int32_t lhs_offset, std::int32_t rhs_offset,
9528   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
9537     zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
9541   zip_2x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
9566                            std::int32_t lhs_offset, std::int32_t rhs_offset,
9586   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
9595     zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
9599   zip_2x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
9624                            std::int32_t lhs_offset, std::int32_t rhs_offset,
9644   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
9653     zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
9657   zip_2x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
9682                            std::int32_t lhs_offset, std::int32_t rhs_offset,
9702   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
9711     zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
9715   zip_2x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
9740                            std::int32_t lhs_offset, std::int32_t rhs_offset,
9760   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
9769     zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
9773   zip_2x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
9798                            std::int32_t lhs_offset, std::int32_t rhs_offset,
9818   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
9827     zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
9831   zip_2x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
9856                            std::int32_t lhs_offset, std::int32_t rhs_offset,
9876   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
9885     zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
9889   zip_2x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
9914                            std::int32_t lhs_offset, std::int32_t rhs_offset,
9934   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
9943     zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
9947   zip_2x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
9972                            std::int32_t lhs_offset, std::int32_t rhs_offset,
9994   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
10003     zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
10042                            std::int32_t lhs_offset, std::int32_t rhs_offset,
10064   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
10073     zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
10112                            std::int32_t lhs_offset, std::int32_t rhs_offset,
10134   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
10143     zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
10182                            std::int32_t lhs_offset, std::int32_t rhs_offset,
10204   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
10213     zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
10252                            std::int32_t lhs_offset, std::int32_t rhs_offset,
10274   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
10283     zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
10322                            std::int32_t lhs_offset, std::int32_t rhs_offset,
10344   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
10353     zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
10392                            std::int32_t lhs_offset, std::int32_t rhs_offset,
10414   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
10423     zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
10462                            std::int32_t lhs_offset, std::int32_t rhs_offset,
10484   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
10493     zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
10532                            std::int32_t lhs_offset, std::int32_t rhs_offset,
10554   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
10563     zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
10567   zip_1x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
10607                            std::int32_t lhs_offset, std::int32_t rhs_offset,
10629   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
10638     zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
10642   zip_1x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
10682                            std::int32_t lhs_offset, std::int32_t rhs_offset,
10704   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
10713     zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
10717   zip_1x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
10757                            std::int32_t lhs_offset, std::int32_t rhs_offset,
10779   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
10788     zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
10792   zip_1x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
10832                            std::int32_t lhs_offset, std::int32_t rhs_offset,
10854   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
10863     zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
10867   zip_1x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
10907                            std::int32_t lhs_offset, std::int32_t rhs_offset,
10929   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
10938     zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
10942   zip_1x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
10982                            std::int32_t lhs_offset, std::int32_t rhs_offset,
11004   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
11013     zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
11017   zip_1x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
11057                            std::int32_t lhs_offset, std::int32_t rhs_offset,
11079   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
11088     zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
11092   zip_1x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
11132                            std::int32_t lhs_offset, std::int32_t rhs_offset,
11154   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
11163     zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
11167   zip_2x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
11207                            std::int32_t lhs_offset, std::int32_t rhs_offset,
11229   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
11238     zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
11242   zip_2x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
11282                            std::int32_t lhs_offset, std::int32_t rhs_offset,
11304   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
11313     zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
11317   zip_2x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
11357                            std::int32_t lhs_offset, std::int32_t rhs_offset,
11379   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
11388     zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
11392   zip_2x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
11432                            std::int32_t lhs_offset, std::int32_t rhs_offset,
11454   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
11463     zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
11467   zip_2x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
11507                            std::int32_t lhs_offset, std::int32_t rhs_offset,
11529   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
11538     zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
11542   zip_2x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
11582                            std::int32_t lhs_offset, std::int32_t rhs_offset,
11604   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
11613     zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
11617   zip_2x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
11657                            std::int32_t lhs_offset, std::int32_t rhs_offset,
11679   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
11688     zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
11692   zip_2x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
11732                            std::int32_t lhs_offset, std::int32_t rhs_offset,
11754   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
11763     zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
11802                            std::int32_t lhs_offset, std::int32_t rhs_offset,
11824   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
11833     zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
11872                            std::int32_t lhs_offset, std::int32_t rhs_offset,
11894   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
11903     zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
11942                            std::int32_t lhs_offset, std::int32_t rhs_offset,
11964   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
11973     zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
12012                            std::int32_t lhs_offset, std::int32_t rhs_offset,
12034   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
12043     zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
12082                            std::int32_t lhs_offset, std::int32_t rhs_offset,
12104   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
12113     zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
12152                            std::int32_t lhs_offset, std::int32_t rhs_offset,
12174   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
12183     zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
12222                            std::int32_t lhs_offset, std::int32_t rhs_offset,
12244   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
12253     zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
12292                            std::int32_t lhs_offset, std::int32_t rhs_offset,
12314   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
12323     zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
12327   zip_1x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
12367                            std::int32_t lhs_offset, std::int32_t rhs_offset,
12389   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
12398     zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
12402   zip_1x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
12442                            std::int32_t lhs_offset, std::int32_t rhs_offset,
12464   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
12473     zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
12477   zip_1x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
12517                            std::int32_t lhs_offset, std::int32_t rhs_offset,
12539   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
12548     zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
12552   zip_1x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
12592                            std::int32_t lhs_offset, std::int32_t rhs_offset,
12614   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
12623     zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
12627   zip_1x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
12667                            std::int32_t lhs_offset, std::int32_t rhs_offset,
12689   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
12698     zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
12702   zip_1x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
12742                            std::int32_t lhs_offset, std::int32_t rhs_offset,
12764   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
12773     zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
12777   zip_1x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
12817                            std::int32_t lhs_offset, std::int32_t rhs_offset,
12839   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
12848     zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
12852   zip_1x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
12892                            std::int32_t lhs_offset, std::int32_t rhs_offset,
12914   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
12923     zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
12927   zip_2x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
12967                            std::int32_t lhs_offset, std::int32_t rhs_offset,
12989   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
12998     zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
13002   zip_2x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
13042                            std::int32_t lhs_offset, std::int32_t rhs_offset,
13064   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
13073     zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
13077   zip_2x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
13117                            std::int32_t lhs_offset, std::int32_t rhs_offset,
13139   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
13148     zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
13152   zip_2x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
13192                            std::int32_t lhs_offset, std::int32_t rhs_offset,
13214   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
13223     zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
13227   zip_2x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
13267                            std::int32_t lhs_offset, std::int32_t rhs_offset,
13289   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
13298     zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
13302   zip_2x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
13342                            std::int32_t lhs_offset, std::int32_t rhs_offset,
13364   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
13373     zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
13377   zip_2x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
13417                            std::int32_t lhs_offset, std::int32_t rhs_offset,
13439   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
13448     zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
13452   zip_2x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
13491                    std::int32_t k, std::int32_t lhs_offset,
13510   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
13519     zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
13544                    std::int32_t k, std::int32_t lhs_offset,
13563   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
13572     zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
13597                    std::int32_t k, std::int32_t lhs_offset,
13616   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
13625     zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
13650                    std::int32_t k, std::int32_t lhs_offset,
13669   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
13678     zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
13703                    std::int32_t k, std::int32_t lhs_offset,
13722   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
13731     zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
13756                    std::int32_t k, std::int32_t lhs_offset,
13775 lhs_offset * rhs_offset * k + result_offset;
13784     zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
13809                    std::int32_t k, std::int32_t lhs_offset,
13828   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
13837     zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
13862                    std::int32_t k, std::int32_t lhs_offset,
13881   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
13890     zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
13915                    std::int32_t k, std::int32_t lhs_offset,
13934   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
13943     zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
13947   zip_1x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
13971                    std::int32_t k, std::int32_t lhs_offset,
13990   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
13999     zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
14003   zip_1x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
14027                    std::int32_t k, std::int32_t lhs_offset,
14046   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
14055     zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
14059   zip_1x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
14083                    std::int32_t k, std::int32_t lhs_offset,
14102   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
14111     zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
14115   zip_1x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
14139                    std::int32_t k, std::int32_t lhs_offset,
14158   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
14167     zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
14171   zip_1x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
14195                    std::int32_t k, std::int32_t lhs_offset,
14214   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
14223     zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
14227   zip_1x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
14251                    std::int32_t k, std::int32_t lhs_offset,
14270   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
14279     zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
14283   zip_1x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
14307                    std::int32_t k, std::int32_t lhs_offset,
14326   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
14335     zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
14339   zip_1x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
14363                    std::int32_t k, std::int32_t lhs_offset,
14382   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
14391     zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
14395   zip_2x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
14419                    std::int32_t k, std::int32_t lhs_offset,
14438   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
14447     zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
14451   zip_2x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
14475                    std::int32_t k, std::int32_t lhs_offset,
14494   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
14503     zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
14507   zip_2x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
14531                    std::int32_t k, std::int32_t lhs_offset,
14550   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
14559     zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
14563   zip_2x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
14587                    std::int32_t k, std::int32_t lhs_offset,
14606   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
14615     zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
14619   zip_2x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
14643                    std::int32_t k, std::int32_t lhs_offset,
14662   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
14671     zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
14675   zip_2x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
14699                    std::int32_t k, std::int32_t lhs_offset,
14718   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
14727     zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
14731   zip_2x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
14755                    std::int32_t k, std::int32_t lhs_offset,
14774   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
14783     zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
14787   zip_2x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
14811                    std::int32_t k, std::int32_t lhs_offset,
14832   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
14841     zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
14879                    std::int32_t k, std::int32_t lhs_offset,
14900   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
14909     zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
14947                    std::int32_t k, std::int32_t lhs_offset,
14968   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
14977     zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
15015                    std::int32_t k, std::int32_t lhs_offset,
15036   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
15045     zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
15083                    std::int32_t k, std::int32_t lhs_offset,
15104   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
15113     zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
15151                    std::int32_t k, std::int32_t lhs_offset,
15172   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
15181     zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
15219                    std::int32_t k, std::int32_t lhs_offset,
15240   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
15249     zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
15287                    std::int32_t k, std::int32_t lhs_offset,
15308   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
15317     zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
15355                    std::int32_t k, std::int32_t lhs_offset,
15376   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
15385     zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
15389   zip_1x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
15428                    std::int32_t k, std::int32_t lhs_offset,
15449   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
15458     zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
15462   zip_1x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
15501                    std::int32_t k, std::int32_t lhs_offset,
15522   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
15531     zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
15535   zip_1x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
15574                    std::int32_t k, std::int32_t lhs_offset,
15595   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
15604     zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
15608   zip_1x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
15647                    std::int32_t k, std::int32_t lhs_offset,
15668   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
15677     zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
15681   zip_1x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
15720                    std::int32_t k, std::int32_t lhs_offset,
15741   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
15750     zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
15754   zip_1x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
15793                    std::int32_t k, std::int32_t lhs_offset,
15814   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
15823     zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
15827   zip_1x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
15866                    std::int32_t k, std::int32_t lhs_offset,
15887   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
15896     zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
15900   zip_1x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
15939                    std::int32_t k, std::int32_t lhs_offset,
15960   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
15969     zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
15973   zip_2x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
16012                    std::int32_t k, std::int32_t lhs_offset,
16033   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
16042     zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
16046   zip_2x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
16085                    std::int32_t k, std::int32_t lhs_offset,
16106   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
16115     zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
16119   zip_2x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
16158                    std::int32_t k, std::int32_t lhs_offset,
16179   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
16188     zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
16192   zip_2x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
16231                    std::int32_t k, std::int32_t lhs_offset,
16252   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
16261     zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
16265   zip_2x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
16304                    std::int32_t k, std::int32_t lhs_offset,
16325   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
16334     zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
16338   zip_2x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
16377                    std::int32_t k, std::int32_t lhs_offset,
16398   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
16407     zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
16411   zip_2x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
16450                    std::int32_t k, std::int32_t lhs_offset,
16471   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
16480     zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
16484   zip_2x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
16523                    std::int32_t k, std::int32_t lhs_offset,
16544 lhs_offset * rhs_offset * k + result_offset;
16553     zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
16591                    std::int32_t k, std::int32_t lhs_offset,
16612   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
16621     zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
16659                    std::int32_t k, std::int32_t lhs_offset,
16680   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
16689     zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
16727                    std::int32_t k, std::int32_t lhs_offset,
16748   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
16757     zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
16795                    std::int32_t k, std::int32_t lhs_offset,
16816   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
16825     zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
16863                    std::int32_t k, std::int32_t lhs_offset,
16884   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
16893     zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
16931                    std::int32_t k, std::int32_t lhs_offset,
16952   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
16961     zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
16999                    std::int32_t k, std::int32_t lhs_offset,
17020   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
17029     zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
17067                    std::int32_t k, std::int32_t lhs_offset,
17088   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
17097     zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
17101   zip_1x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
17140                    std::int32_t k, std::int32_t lhs_offset,
17161   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
17170     zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
17174   zip_1x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
17213                    std::int32_t k, std::int32_t lhs_offset,
17234   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
17243     zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
17247   zip_1x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset
17286                    std::int32_t k, std::int32_t lhs_offset,
17307   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
17316     zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
17320   zip_1x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
17359                    std::int32_t k, std::int32_t lhs_offset,
17380   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
17389     zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
17393   zip_1x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
17432                    std::int32_t k, std::int32_t lhs_offset,
17453   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
17462     zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
17466   zip_1x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
17505                    std::int32_t k, std::int32_t lhs_offset,
17526   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
17535     zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
17539   zip_1x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
17578                    std::int32_t k, std::int32_t lhs_offset,
17599   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
17608     zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
17612   zip_1x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
17651                    std::int32_t k, std::int32_t lhs_offset,
17672   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
17681     zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
17685   zip_2x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
17724                    std::int32_t k, std::int32_t lhs_offset,
17745   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
17754     zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
17758   zip_2x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
17797                    std::int32_t k, std::int32_t lhs_offset,
17818   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
17827     zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
17831   zip_2x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
17870                    std::int32_t k, std::int32_t lhs_offset,
17891   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
17900     zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
17904   zip_2x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
17943                    std::int32_t k, std::int32_t lhs_offset,
17964   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
17973     zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
17977   zip_2x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
18016                    std::int32_t k, std::int32_t lhs_offset,
18037   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
18046     zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
18050   zip_2x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
18089                    std::int32_t k, std::int32_t lhs_offset,
18110   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
18119     zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
18123   zip_2x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
18162                    std::int32_t k, std::int32_t lhs_offset,
18183   const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset;
18192     zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
18196   zip_2x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
18236                             std::int32_t lhs_offset, std::int32_t rhs_offset,
18253   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
18259     zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
18283                             std::int32_t lhs_offset, std::int32_t rhs_offset,
18300   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
18306     zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
18330                             std::int32_t lhs_offset, std::int32_t rhs_offset,
18347   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
18353     zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
18377                             std::int32_t lhs_offset, std::int32_t rhs_offset,
18394   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
18400     zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
18424                             std::int32_t lhs_offset, std::int32_t rhs_offset,
18441   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
18447     zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
18471                             std::int32_t lhs_offset, std::int32_t rhs_offset,
18488   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
18494     zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
18518                             std::int32_t lhs_offset, std::int32_t rhs_offset,
18535   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
18541     zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
18565                             std::int32_t lhs_offset, std::int32_t rhs_offset,
18582   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
18588     zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
18612                             std::int32_t lhs_offset, std::int32_t rhs_offset,
18629   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
18635     zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
18639   zip_1x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
18663                             std::int32_t lhs_offset, std::int32_t rhs_offset,
18680   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
18686     zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
18690   zip_1x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
18714                             std::int32_t lhs_offset, std::int32_t rhs_offset,
18731   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
18737     zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
18741   zip_1x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
18765                             std::int32_t lhs_offset, std::int32_t rhs_offset,
18782   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
18788     zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
18792   zip_1x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
18816                             std::int32_t lhs_offset, std::int32_t rhs_offset,
18833   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
18839     zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
18843   zip_1x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
18867                             std::int32_t lhs_offset, std::int32_t rhs_offset,
18884   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
18890     zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
18894   zip_1x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
18918                             std::int32_t lhs_offset, std::int32_t rhs_offset,
18935   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
18941     zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
18945   zip_1x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
18969                             std::int32_t lhs_offset, std::int32_t rhs_offset,
18986   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
18992     zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
18996   zip_1x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
19020                             std::int32_t lhs_offset, std::int32_t rhs_offset,
19037   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
19043     zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
19047   zip_2x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
19071                             std::int32_t lhs_offset, std::int32_t rhs_offset,
19088   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
19094     zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
19098   zip_2x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
19122                             std::int32_t lhs_offset, std::int32_t rhs_offset,
19139   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
19145     zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
19149   zip_2x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
19173                             std::int32_t lhs_offset, std::int32_t rhs_offset,
19190   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
19196     zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
19200   zip_2x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
19224                             std::int32_t lhs_offset, std::int32_t rhs_offset,
19241   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
19247     zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
19251   zip_2x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
19275                             std::int32_t lhs_offset, std::int32_t rhs_offset,
19292   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
19298     zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
19302   zip_2x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
19326                             std::int32_t lhs_offset, std::int32_t rhs_offset,
19343   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
19349     zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
19353   zip_2x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
19377                             std::int32_t lhs_offset, std::int32_t rhs_offset,
19394   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
19400     zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
19404   zip_2x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
19428                             std::int32_t lhs_offset, std::int32_t rhs_offset,
19447   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
19453     zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
19488                             std::int32_t lhs_offset, std::int32_t rhs_offset,
19507   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
19513     zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
19548                             std::int32_t lhs_offset, std::int32_t rhs_offset,
19567   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
19573     zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
19608                             std::int32_t lhs_offset, std::int32_t rhs_offset,
19627   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
19633     zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
19668                             std::int32_t lhs_offset, std::int32_t rhs_offset,
19687   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
19693     zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
19728                             std::int32_t lhs_offset, std::int32_t rhs_offset,
19747   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
19753     zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
19788                             std::int32_t lhs_offset, std::int32_t rhs_offset,
19807   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
19813     zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
19848                             std::int32_t lhs_offset, std::int32_t rhs_offset,
19867   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
19873     zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
19908                             std::int32_t lhs_offset, std::int32_t rhs_offset,
19927   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
19933     zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
19937   zip_1x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
19975                             std::int32_t lhs_offset, std::int32_t rhs_offset,
19994   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
20000     zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
20004   zip_1x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
20042                             std::int32_t lhs_offset, std::int32_t rhs_offset,
20061   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
20067     zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
20071   zip_1x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
20109                             std::int32_t lhs_offset, std::int32_t rhs_offset,
20128   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
20134     zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
20138   zip_1x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
20176                             std::int32_t lhs_offset, std::int32_t rhs_offset,
20195   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
20201     zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
20205   zip_1x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
20243                             std::int32_t lhs_offset, std::int32_t rhs_offset,
20262   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
20268     zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
20272   zip_1x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
20310                             std::int32_t lhs_offset, std::int32_t rhs_offset,
20329   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
20335     zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
20339   zip_1x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
20377                             std::int32_t lhs_offset, std::int32_t rhs_offset,
20396   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
20402     zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
20406   zip_1x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
20444                             std::int32_t lhs_offset, std::int32_t rhs_offset,
20463 lhs_offset * rhs_offset * k;
20469     zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
20473   zip_2x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
20511                             std::int32_t lhs_offset, std::int32_t rhs_offset,
20530   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
20536     zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
20540   zip_2x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
20578                             std::int32_t lhs_offset, std::int32_t rhs_offset,
20597   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
20603     zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
20607   zip_2x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
20645                             std::int32_t lhs_offset, std::int32_t rhs_offset,
20664   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
20670     zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
20674   zip_2x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
20712                             std::int32_t lhs_offset, std::int32_t rhs_offset,
20731   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
20737     zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
20741   zip_2x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
20779                             std::int32_t lhs_offset, std::int32_t rhs_offset,
20798   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
20804     zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
20808   zip_2x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
20846                             std::int32_t lhs_offset, std::int32_t rhs_offset,
20865   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
20871     zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
20875   zip_2x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
20913                             std::int32_t lhs_offset, std::int32_t rhs_offset,
20932   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
20938     zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
20942   zip_2x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
20980                             std::int32_t lhs_offset, std::int32_t rhs_offset,
20999   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
21005     zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
21040                             std::int32_t lhs_offset, std::int32_t rhs_offset,
21059   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
21065     zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
21100                             std::int32_t lhs_offset, std::int32_t rhs_offset,
21119   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
21125     zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
21160                             std::int32_t lhs_offset, std::int32_t rhs_offset,
21179   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
21185 lhs_offset, 0);
21220                             std::int32_t lhs_offset, std::int32_t rhs_offset,
21239   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
21245     zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
21280                             std::int32_t lhs_offset, std::int32_t rhs_offset,
21299   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
21305     zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
21340                             std::int32_t lhs_offset, std::int32_t rhs_offset,
21359   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
21365     zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
21400                             std::int32_t lhs_offset, std::int32_t rhs_offset,
21419   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
21425     zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
21460                             std::int32_t lhs_offset, std::int32_t rhs_offset,
21479   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
21485     zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
21489   zip_1x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
21527                             std::int32_t lhs_offset, std::int32_t rhs_offset,
21546   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
21552     zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
21556   zip_1x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
21594                             std::int32_t lhs_offset, std::int32_t rhs_offset,
21613   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
21619     zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
21623   zip_1x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
21661                             std::int32_t lhs_offset, std::int32_t rhs_offset,
21680   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
21686     zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
21690   zip_1x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
21728                             std::int32_t lhs_offset, std::int32_t rhs_offset,
21747   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
21753     zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
21757   zip_1x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
21795                             std::int32_t lhs_offset, std::int32_t rhs_offset,
21814   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
21820     zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
21824   zip_1x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
21862                             std::int32_t lhs_offset, std::int32_t rhs_offset,
21881   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
21887     zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
21891   zip_1x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
21929                             std::int32_t lhs_offset, std::int32_t rhs_offset,
21948   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
21954     zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
21958   zip_1x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
21996                             std::int32_t lhs_offset, std::int32_t rhs_offset,
22015   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
22021     zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
22025   zip_2x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
22063                             std::int32_t lhs_offset, std::int32_t rhs_offset,
22082   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
22088     zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
22092   zip_2x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
22130                             std::int32_t lhs_offset, std::int32_t rhs_offset,
22149   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
22155     zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
22159   zip_2x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
22197                             std::int32_t lhs_offset, std::int32_t rhs_offset,
22216   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
22222     zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
22226   zip_2x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
22264                             std::int32_t lhs_offset, std::int32_t rhs_offset,
22283   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
22289     zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
22293   zip_2x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
22331                             std::int32_t lhs_offset, std::int32_t rhs_offset,
22350   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
22356     zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
22360   zip_2x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
22398                             std::int32_t lhs_offset, std::int32_t rhs_offset,
22417   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
22423     zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
22427   zip_2x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
22465                             std::int32_t lhs_offset, std::int32_t rhs_offset,
22484   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
22490     zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
22494   zip_2x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
22531                     std::int32_t k, std::int32_t lhs_offset,
22549   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
22555     zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
22578                     std::int32_t k, std::int32_t lhs_offset,
22596   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
22602     zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
22625                     std::int32_t k, std::int32_t lhs_offset,
22643   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
22649     zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
22672                     std::int32_t k, std::int32_t lhs_offset,
22690   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
22696     zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
22719                     std::int32_t k, std::int32_t lhs_offset,
22737   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
22743     zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
22766                     std::int32_t k, std::int32_t lhs_offset,
22784   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
22790     zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
22813                     std::int32_t k, std::int32_t lhs_offset,
22831   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
22837     zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
22860                     std::int32_t k, std::int32_t lhs_offset,
22878   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
22884     zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
22907                     std::int32_t k, std::int32_t lhs_offset,
22925   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
22931     zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
22935   zip_1x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
22958                     std::int32_t k, std::int32_t lhs_offset,
22976   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
22982     zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
22986   zip_1x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
23009                     std::int32_t k, std::int32_t lhs_offset,
23027   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
23033     zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
23037   zip_1x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
23060                     std::int32_t k, std::int32_t lhs_offset,
23078   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
23084     zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
23088   zip_1x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
23111                     std::int32_t k, std::int32_t lhs_offset,
23129   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
23135     zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
23139   zip_1x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
23162                     std::int32_t k, std::int32_t lhs_offset,
23180   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
23186     zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
23190   zip_1x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
23213                     std::int32_t k, std::int32_t lhs_offset,
23231   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
23237     zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
23241   zip_1x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
23264                     std::int32_t k, std::int32_t lhs_offset,
23282   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
23288     zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
23292   zip_1x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
23315                     std::int32_t k, std::int32_t lhs_offset,
23333   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
23339     zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
23343   zip_2x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
23366                     std::int32_t k, std::int32_t lhs_offset,
23384   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
23390     zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
23394   zip_2x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
23417                     std::int32_t k, std::int32_t lhs_offset,
23435   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
23441     zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
23445   zip_2x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
23468                     std::int32_t k, std::int32_t lhs_offset,
23486   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
23492     zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
23496   zip_2x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
23519                     std::int32_t k, std::int32_t lhs_offset,
23537   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
23543     zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
23547   zip_2x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
23570                     std::int32_t k, std::int32_t lhs_offset,
23588   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
23594     zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
23598   zip_2x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
23621                     std::int32_t k, std::int32_t lhs_offset,
23639   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
23645     zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
23649   zip_2x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
23672                     std::int32_t k, std::int32_t lhs_offset,
23690   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
23696     zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
23700   zip_2x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
23723                     std::int32_t k, std::int32_t lhs_offset,
23743   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
23749     zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
23783                     std::int32_t k, std::int32_t lhs_offset,
23803   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
23809     zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
23843                     std::int32_t k, std::int32_t lhs_offset,
23863   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
23869     zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
23903                     std::int32_t k, std::int32_t lhs_offset,
23923   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
23929     zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
23963                     std::int32_t k, std::int32_t lhs_offset,
23983   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
23989     zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
24023                     std::int32_t k, std::int32_t lhs_offset,
24043   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
24049     zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
24083                     std::int32_t k, std::int32_t lhs_offset,
24103   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
24109     zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
24143 lhs_offset,
24163   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
24169     zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
24203                     std::int32_t k, std::int32_t lhs_offset,
24223   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
24229     zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
24233   zip_1x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
24270                     std::int32_t k, std::int32_t lhs_offset,
24290   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
24296     zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
24300   zip_1x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
24337                     std::int32_t k, std::int32_t lhs_offset,
24357   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
24363     zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
24367   zip_1x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
24404                     std::int32_t k, std::int32_t lhs_offset,
24424   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
24430     zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
24434   zip_1x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
24471                     std::int32_t k, std::int32_t lhs_offset,
24491   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
24497     zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
24501   zip_1x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
24538                     std::int32_t k, std::int32_t lhs_offset,
24558   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
24564     zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
24568   zip_1x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
24605                     std::int32_t k, std::int32_t lhs_offset,
24625   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
24631     zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
24635   zip_1x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
24672                     std::int32_t k, std::int32_t lhs_offset,
24692   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
24698     zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
24702   zip_1x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
24739                     std::int32_t k, std::int32_t lhs_offset,
24759   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
24765     zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
24769   zip_2x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
24806                     std::int32_t k, std::int32_t lhs_offset,
24826   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
24832     zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
24836   zip_2x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
24873                     std::int32_t k, std::int32_t lhs_offset,
24893   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
24899     zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
24903   zip_2x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
24940                     std::int32_t k, std::int32_t lhs_offset,
24960   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
24966     zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
24970   zip_2x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
25007                     std::int32_t k, std::int32_t lhs_offset,
25027   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
25033     zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
25037   zip_2x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
25074                     std::int32_t k, std::int32_t lhs_offset,
25094   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
25100     zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
25104   zip_2x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
25141                     std::int32_t k, std::int32_t lhs_offset,
25161   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
25167     zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
25171   zip_2x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
25208                     std::int32_t k, std::int32_t lhs_offset,
25228   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
25234     zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
25238   zip_2x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
25275                     std::int32_t k, std::int32_t lhs_offset,
25295   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
25301     zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
25335                     std::int32_t k, std::int32_t lhs_offset,
25355   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
25361     zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
25395                     std::int32_t k, std::int32_t lhs_offset,
25415   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
25421     zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
25455                     std::int32_t k, std::int32_t lhs_offset,
25475   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
25481     zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
25515                     std::int32_t k, std::int32_t lhs_offset,
25535   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
25541     zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
25575                     std::int32_t k, std::int32_t lhs_offset,
25595   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
25601     zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
25635                     std::int32_t k, std::int32_t lhs_offset,
25655   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
25661     zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
25695                     std::int32_t k, std::int32_t lhs_offset,
25715   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
25721     zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
25755                     std::int32_t k, std::int32_t lhs_offset,
25775   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
25781     zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
25785   zip_1x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
25822                     std::int32_t k, std::int32_t lhs_offset,
25842   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
25848     zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
25852   zip_1x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
25889                     std::int32_t k, std::int32_t lhs_offset,
25909   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
25915     zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
25919   zip_1x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
25956                     std::int32_t k, std::int32_t lhs_offset,
25976   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
25982     zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
25986   zip_1x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
26023                     std::int32_t k, std::int32_t lhs_offset,
26043   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
26049     zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
26053   zip_1x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
26090                     std::int32_t k, std::int32_t lhs_offset,
26110   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
26116     zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
26120   zip_1x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
26157                     std::int32_t k, std::int32_t lhs_offset,
26177   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
26183     zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
26187   zip_1x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
26224                     std::int32_t k, std::int32_t lhs_offset,
26244   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
26250     zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
26254   zip_1x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
26291                     std::int32_t k, std::int32_t lhs_offset,
26311   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
26317     zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
26321   zip_2x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
26358                     std::int32_t k, std::int32_t lhs_offset,
26378   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
26384     zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
26388   zip_2x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
26425                     std::int32_t k, std::int32_t lhs_offset,
26445   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
26451     zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
26455   zip_2x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
26492                     std::int32_t k, std::int32_t lhs_offset,
26512   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
26518     zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
26522   zip_2x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
26559                     std::int32_t k, std::int32_t lhs_offset,
26579   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
26585     zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
26589   zip_2x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
26626                     std::int32_t k, std::int32_t lhs_offset,
26646   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
26652     zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
26656   zip_2x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
26693                     std::int32_t k, std::int32_t lhs_offset,
26713   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
26719     zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
26723   zip_2x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
26760                     std::int32_t k, std::int32_t lhs_offset,
26780   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
26786     zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
26790   zip_2x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
26828                           std::int32_t lhs_offset, std::int32_t rhs_offset,
26846   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
26852     zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
26876                           std::int32_t lhs_offset, std::int32_t rhs_offset,
26894   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
26900     zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
26924                           std::int32_t lhs_offset, std::int32_t rhs_offset,
26942   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
26948     zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
26972                           std::int32_t lhs_offset, std::int32_t rhs_offset,
26990   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
26996     zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
27020                           std::int32_t lhs_offset, std::int32_t rhs_offset,
27038   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
27044     zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
27068                           std::int32_t lhs_offset, std::int32_t rhs_offset,
27086   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
27092     zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
27116                           std::int32_t lhs_offset, std::int32_t rhs_offset,
27134   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
27140     zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
27164                           std::int32_t lhs_offset, std::int32_t rhs_offset,
27182   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
27188     zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
27212                           std::int32_t lhs_offset, std::int32_t rhs_offset,
27230   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
27236     zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
27240   zip_1x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
27264                           std::int32_t lhs_offset, std::int32_t rhs_offset,
27282   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
27288     zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
27292   zip_1x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
27316                           std::int32_t lhs_offset, std::int32_t rhs_offset,
27334   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
27340     zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
27344   zip_1x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
27368                           std::int32_t lhs_offset, std::int32_t rhs_offset,
27386   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
27392     zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
27396   zip_1x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
27420                           std::int32_t lhs_offset, std::int32_t rhs_offset,
27438   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
27444     zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
27448   zip_1x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
27472                           std::int32_t lhs_offset, std::int32_t rhs_offset,
27490   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
27496     zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
27500   zip_1x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
27524                           std::int32_t lhs_offset, std::int32_t rhs_offset,
27542   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
27548     zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
27552   zip_1x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
27576                           std::int32_t lhs_offset, std::int32_t rhs_offset,
27594   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
27600     zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
27604   zip_1x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
27628                           std::int32_t lhs_offset, std::int32_t rhs_offset,
27646   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
27652     zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
27656   zip_2x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
27680                           std::int32_t lhs_offset, std::int32_t rhs_offset,
27698   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
27704     zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
27708   zip_2x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
27732                           std::int32_t lhs_offset, std::int32_t rhs_offset,
27750   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
27756     zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
27760   zip_2x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
27784                           std::int32_t lhs_offset, std::int32_t rhs_offset,
27802   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
27808     zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
27812   zip_2x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
27836                           std::int32_t lhs_offset, std::int32_t rhs_offset,
27854   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
27860     zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
27864   zip_2x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
27888                           std::int32_t lhs_offset, std::int32_t rhs_offset,
27906   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
27912     zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
27916   zip_2x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
27940                           std::int32_t lhs_offset, std::int32_t rhs_offset,
27958   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
27964     zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
27968   zip_2x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
27992                           std::int32_t lhs_offset, std::int32_t rhs_offset,
28010   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
28016     zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
28020   zip_2x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
28044                           std::int32_t lhs_offset, std::int32_t rhs_offset,
28064   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
28070     zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
28105                           std::int32_t lhs_offset, std::int32_t rhs_offset,
28125   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
28131     zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
28166                           std::int32_t lhs_offset, std::int32_t rhs_offset,
28186   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
28192     zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
28227                           std::int32_t lhs_offset, std::int32_t rhs_offset,
28247   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
28253     zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
28288                           std::int32_t lhs_offset, std::int32_t rhs_offset,
28308   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
28314     zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
28349                           std::int32_t lhs_offset, std::int32_t rhs_offset,
28369   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
28375     zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
28410                           std::int32_t lhs_offset, std::int32_t rhs_offset,
28430   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
28436     zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
28471                           std::int32_t lhs_offset, std::int32_t rhs_offset,
28491   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
28497     zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
28532                           std::int32_t lhs_offset, std::int32_t rhs_offset,
28552   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
28558     zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
28562   zip_1x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
28600                           std::int32_t lhs_offset, std::int32_t rhs_offset,
28620   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
28626     zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
28630   zip_1x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
28668                           std::int32_t lhs_offset, std::int32_t rhs_offset,
28688   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
28694     zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
28698   zip_1x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
28736                           std::int32_t lhs_offset, std::int32_t rhs_offset,
28756   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
28762     zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
28766   zip_1x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
28804                           std::int32_t lhs_offset, std::int32_t rhs_offset,
28824   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
28830     zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
28834   zip_1x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
28872                           std::int32_t lhs_offset, std::int32_t rhs_offset,
28892   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
28898     zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
28902   zip_1x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
28940                           std::int32_t lhs_offset, std::int32_t rhs_offset,
28960   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
28966     zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
28970   zip_1x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
29008                           std::int32_t lhs_offset, std::int32_t rhs_offset,
29028   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
29034     zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
29038   zip_1x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
29076                           std::int32_t lhs_offset, std::int32_t rhs_offset,
29096   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
29102     zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
29106   zip_2x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
29144                           std::int32_t lhs_offset, std::int32_t rhs_offset,
29164   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
29170     zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
29174   zip_2x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
29212                           std::int32_t lhs_offset, std::int32_t rhs_offset,
29232   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
29238     zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
29242   zip_2x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
29280                           std::int32_t lhs_offset, std::int32_t rhs_offset,
29300   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
29306     zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
29310   zip_2x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
29348                           std::int32_t lhs_offset, std::int32_t rhs_offset,
29368   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
29374     zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
29378   zip_2x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
29416                           std::int32_t lhs_offset, std::int32_t rhs_offset,
29436   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
29442     zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
29446   zip_2x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
29484                           std::int32_t lhs_offset, std::int32_t rhs_offset,
29504   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
29510     zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
29514   zip_2x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
29552                           std::int32_t lhs_offset, std::int32_t rhs_offset,
29572   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
29578     zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
29582   zip_2x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
29620                           std::int32_t lhs_offset, std::int32_t rhs_offset,
29640   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
29646     zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
29681                           std::int32_t lhs_offset, std::int32_t rhs_offset,
29701   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
29707     zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
29742                           std::int32_t lhs_offset, std::int32_t rhs_offset,
29762   const std::int32_t const_offset = lhs_offset
29768     zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
29803                           std::int32_t lhs_offset, std::int32_t rhs_offset,
29823   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
29829     zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
29864                           std::int32_t lhs_offset, std::int32_t rhs_offset,
29884   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
29890     zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
29925                           std::int32_t lhs_offset, std::int32_t rhs_offset,
29945   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
29951     zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
29986                           std::int32_t lhs_offset, std::int32_t rhs_offset,
30006   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
30012     zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
30047                           std::int32_t lhs_offset, std::int32_t rhs_offset,
30067   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
30073     zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
30108                           std::int32_t lhs_offset, std::int32_t rhs_offset,
30128   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
30134     zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
30138   zip_1x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
30176                           std::int32_t lhs_offset, std::int32_t rhs_offset,
30196   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
30202     zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
30206   zip_1x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
30244                           std::int32_t lhs_offset, std::int32_t rhs_offset,
30264   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
30270     zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
30274   zip_1x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
30312                           std::int32_t lhs_offset, std::int32_t rhs_offset,
30332   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
30338     zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
30342   zip_1x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
30380                           std::int32_t lhs_offset, std::int32_t rhs_offset,
30400   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
30406     zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
30410   zip_1x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
30448                           std::int32_t lhs_offset, std::int32_t rhs_offset,
30468   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
30474     zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
30478   zip_1x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
30516                           std::int32_t lhs_offset, std::int32_t rhs_offset,
30536   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
30542     zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
30546   zip_1x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
30584                           std::int32_t lhs_offset, std::int32_t rhs_offset,
30604   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
30610     zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
30614   zip_1x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
30652                           std::int32_t lhs_offset, std::int32_t rhs_offset,
30672   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
30678     zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
30682   zip_2x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
30720                           std::int32_t lhs_offset, std::int32_t rhs_offset,
30740   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
30746     zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
30750   zip_2x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
30788                           std::int32_t lhs_offset, std::int32_t rhs_offset,
30808   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
30814     zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
30818   zip_2x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
30856                           std::int32_t lhs_offset, std::int32_t rhs_offset,
30876   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
30882     zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
30886   zip_2x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
30924                           std::int32_t lhs_offset, std::int32_t rhs_offset,
30944   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
30950     zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
30954   zip_2x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
30992                           std::int32_t lhs_offset, std::int32_t rhs_offset,
31012   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
31018     zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
31022   zip_2x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
31060                           std::int32_t lhs_offset, std::int32_t rhs_offset,
31080   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
31086     zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
31090   zip_2x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
31128                           std::int32_t lhs_offset, std::int32_t rhs_offset,
31148   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
31154     zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
31158   zip_2x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
31195                   std::int32_t k, std::int32_t lhs_offset,
31213   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
31219     zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
31242                   std::int32_t k, std::int32_t lhs_offset,
31260   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
31266     zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
31289                   std::int32_t k, std::int32_t lhs_offset,
31307   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
31313     zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
31336                   std::int32_t k, std::int32_t lhs_offset,
31354   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
31360     zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
31383                   std::int32_t k, std::int32_t lhs_offset,
31401   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
31407     zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
31430                   std::int32_t k, std::int32_t lhs_offset,
31448   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
31454     zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
31477                   std::int32_t k, std::int32_t lhs_offset,
31495   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
31501     zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
31524                   std::int32_t k, std::int32_t lhs_offset,
31542   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
31548     zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
31571                   std::int32_t k, std::int32_t lhs_offset,
31589   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
31595     zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
31599   zip_1x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
31622                   std::int32_t k, std::int32_t lhs_offset,
31640   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
31646     zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
31650   zip_1x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
31673                   std::int32_t k, std::int32_t lhs_offset,
31691   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
31697     zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
31701   zip_1x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
31724                   std::int32_t k, std::int32_t lhs_offset,
31742   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
31748     zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
31752   zip_1x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
31775                   std::int32_t k, std::int32_t lhs_offset,
31793   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
31799     zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
31803   zip_1x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
31826                   std::int32_t k, std::int32_t lhs_offset,
31844   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
31850     zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
31854   zip_1x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
31877                   std::int32_t k, std::int32_t lhs_offset,
31895   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
31901     zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
31905   zip_1x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
31928                   std::int32_t k, std::int32_t lhs_offset,
31946   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
31952     zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
31956   zip_1x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
31979                   std::int32_t k, std::int32_t lhs_offset,
31997   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
32003     zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
32007   zip_2x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
32030                   std::int32_t k, std::int32_t lhs_offset,
32048   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
32054     zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
32058   zip_2x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset
32081                   std::int32_t k, std::int32_t lhs_offset,
32099   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
32105     zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
32109   zip_2x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
32132                   std::int32_t k, std::int32_t lhs_offset,
32150   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
32156     zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
32160   zip_2x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
32183                   std::int32_t k, std::int32_t lhs_offset,
32201   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
32207     zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
32211   zip_2x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
32234                   std::int32_t k, std::int32_t lhs_offset,
32252   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
32258     zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
32262   zip_2x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
32285                   std::int32_t k, std::int32_t lhs_offset,
32303   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
32309     zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
32313   zip_2x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
32336                   std::int32_t k, std::int32_t lhs_offset,
32354   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
32360     zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
32364   zip_2x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
32387                   std::int32_t k, std::int32_t lhs_offset,
32407   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
32413     zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
32447                   std::int32_t k, std::int32_t lhs_offset,
32467   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
32473     zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
32507                   std::int32_t k, std::int32_t lhs_offset,
32527   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
32533     zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
32567                   std::int32_t k, std::int32_t lhs_offset,
32587   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
32593     zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
32627                   std::int32_t k, std::int32_t lhs_offset,
32647   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
32653     zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
32687                   std::int32_t k, std::int32_t lhs_offset,
32707   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
32713     zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
32747                   std::int32_t k, std::int32_t lhs_offset,
32767   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
32773     zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
32807                   std::int32_t k, std::int32_t lhs_offset,
32827   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
32833     zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
32867                   std::int32_t k, std::int32_t lhs_offset,
32887   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
32893     zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
32897   zip_1x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
32934                   std::int32_t k, std::int32_t lhs_offset,
32954   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
32960     zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
32964   zip_1x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
33001                   std::int32_t k, std::int32_t lhs_offset,
33021   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
33027     zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
33031   zip_1x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
33068                   std::int32_t k, std::int32_t lhs_offset,
33088   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
33094     zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
33098   zip_1x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
33135                   std::int32_t k, std::int32_t lhs_offset,
33155   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
33161     zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
33165   zip_1x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
33202                   std::int32_t k, std::int32_t lhs_offset,
33222   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
33228     zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
33232   zip_1x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
33269                   std::int32_t k, std::int32_t lhs_offset,
33289   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
33295     zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
33299   zip_1x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
33336                   std::int32_t k, std::int32_t lhs_offset,
33356   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
33362     zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
33366   zip_1x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
33403                   std::int32_t k, std::int32_t lhs_offset,
33423   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
33429     zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
33433   zip_2x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
33470                   std::int32_t k, std::int32_t lhs_offset,
33490   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
33496     zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
33500   zip_2x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
33537                   std::int32_t k, std::int32_t lhs_offset,
33557   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
33563     zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
33567   zip_2x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
33604                   std::int32_t k, std::int32_t lhs_offset,
33624   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
33630     zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
33634   zip_2x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
33671                   std::int32_t k, std::int32_t lhs_offset,
33691   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
33697     zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
33701   zip_2x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
33738                   std::int32_t k, std::int32_t lhs_offset,
33758   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
33764     zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
33768   zip_2x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
33805                   std::int32_t k, std::int32_t lhs_offset,
33825   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
33831     zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
33835   zip_2x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
33872                   std::int32_t k, std::int32_t lhs_offset,
33892   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
33898     zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
33902   zip_2x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
33939                   std::int32_t k, std::int32_t lhs_offset,
33959   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
33965     zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
33999                   std::int32_t k, std::int32_t lhs_offset,
34019   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
34025     zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
34059                   std::int32_t k, std::int32_t lhs_offset,
34079   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
34085     zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
34119                   std::int32_t k, std::int32_t lhs_offset,
34139   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
34145     zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
34179                   std::int32_t k, std::int32_t lhs_offset,
34199   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
34205     zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
34239                   std::int32_t k, std::int32_t lhs_offset,
34259   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
34265     zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
34299                   std::int32_t k, std::int32_t lhs_offset,
34319   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
34325     zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
34359                   std::int32_t k, std::int32_t lhs_offset,
34379   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
34385     zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
34419                   std::int32_t k, std::int32_t lhs_offset,
34439   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
34445     zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
34449   zip_1x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
34486                   std::int32_t k, std::int32_t lhs_offset,
34506   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
34512     zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
34516   zip_1x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
34553                   std::int32_t k, std::int32_t lhs_offset,
34573   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
34579     zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
34583   zip_1x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
34620                   std::int32_t k, std::int32_t lhs_offset,
34640   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
34646     zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
34650   zip_1x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
34687                   std::int32_t k, std::int32_t lhs_offset,
34707   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
34713     zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
34717   zip_1x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
34754                   std::int32_t k, std::int32_t lhs_offset,
34774   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
34780     zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
34784   zip_1x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
34821                   std::int32_t k, std::int32_t lhs_offset,
34841   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
34847     zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
34851   zip_1x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
34888                   std::int32_t k, std::int32_t lhs_offset,
34908   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
34914     zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
34918   zip_1x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
34955                   std::int32_t k, std::int32_t lhs_offset,
34975   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
34981     zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
34985   zip_2x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
35022                   std::int32_t k, std::int32_t lhs_offset,
35042   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
35048     zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
35052   zip_2x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
35089                   std::int32_t k, std::int32_t lhs_offset,
35109   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
35115     zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
35119   zip_2x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
35156                   std::int32_t k, std::int32_t lhs_offset,
35176   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
35182     zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
35186   zip_2x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
35223                   std::int32_t k, std::int32_t lhs_offset,
35243   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
35249     zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
35253   zip_2x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
35290                   std::int32_t k, std::int32_t lhs_offset,
35310   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
35316     zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
35320   zip_2x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
35357                   std::int32_t k, std::int32_t lhs_offset,
35377   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
35383     zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
35387   zip_2x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
35424                   std::int32_t k, std::int32_t lhs_offset,
35444   const std::int32_t const_offset = lhs_offset * rhs_offset * k;
35450     zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
35454   zip_2x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0);
35493                      std::int32_t k, std::int32_t lhs_offset,
35513                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35519                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35525                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35531                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35537                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35543                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35549                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35555                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35565                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35571                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35577                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35583                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35589                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35595                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35601                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35607                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35617                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35623                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35629                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35635                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35641                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35647                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35653                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35659                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35673                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35679                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35685                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35691                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35697                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35703                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35709                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35715                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35725                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35731                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35737                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35743                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35749                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35755                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35761                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35767                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35777                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35783                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35789                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35795                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35801                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35807                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35813                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35819                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35833                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35839                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35845                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35851                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35857                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35863                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35869                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35875                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35885                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35891                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35897                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35903                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35909                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35915                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35921                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35927                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35937                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35943                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35949                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35955                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35961                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35967                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35973                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35979                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
35995                 internal::gemm_q8_0_0_0(scratch, lhs, rhs, m, n, k, lhs_offset,
36001                 internal::gemm_q8_0_0_1(scratch, lhs, rhs, m, n, k, lhs_offset,
36007                 internal::gemm_q8_0_0_2(scratch, lhs, rhs, m, n, k, lhs_offset,
36013                 internal::gemm_q8_0_0_3(scratch, lhs, rhs, m, n, k, lhs_offset,
36019                 internal::gemm_q8_0_0_4(scratch, lhs, rhs, m, n, k, lhs_offset,
36025                 internal::gemm_q8_0_0_5(scratch, lhs, rhs, m, n, k, lhs_offset,
36031                 internal::gemm_q8_0_0_6(scratch, lhs, rhs, m, n, k, lhs_offset,
36037                 internal::gemm_q8_0_0_7(scratch, lhs, rhs, m, n, k, lhs_offset,
36047                 internal::gemm_q8_0_1_0(scratch, lhs, rhs, m, n, k, lhs_offset,
36053                 internal::gemm_q8_0_1_1(scratch, lhs, rhs, m, n, k, lhs_offset,
36059                 internal::gemm_q8_0_1_2(scratch, lhs, rhs, m, n, k, lhs_offset,
36065                 internal::gemm_q8_0_1_3(scratch, lhs, rhs, m, n, k, lhs_offset,
36071                 internal::gemm_q8_0_1_4(scratch, lhs, rhs, m, n, k, lhs_offset,
36077                 internal::gemm_q8_0_1_5(scratch, lhs, rhs, m, n, k, lhs_offset,
36083                 internal::gemm_q8_0_1_6(scratch, lhs, rhs, m, n, k, lhs_offset,
36089                 internal::gemm_q8_0_1_7(scratch, lhs, rhs, m, n, k, lhs_offset,
36099                 internal::gemm_q8_0_2_0(scratch, lhs, rhs, m, n, k, lhs_offset,
36105                 internal::gemm_q8_0_2_1(scratch, lhs, rhs, m, n, k, lhs_offset,
36111                 internal::gemm_q8_0_2_2(scratch, lhs, rhs, m, n, k, lhs_offset,
36117                 internal::gemm_q8_0_2_3(scratch, lhs, rhs, m, n, k, lhs_offset,
36123                 internal::gemm_q8_0_2_4(scratch, lhs, rhs, m, n, k, lhs_offset,
36129                 internal::gemm_q8_0_2_5(scratch, lhs, rhs, m, n, k, lhs_offset,
36135                 internal::gemm_q8_0_2_6(scratch, lhs, rhs, m, n, k, lhs_offset,
36141                 internal::gemm_q8_0_2_7(scratch, lhs, rhs, m, n, k, lhs_offset,
36155                 internal::gemm_q8_1_0_0(scratch, lhs, rhs, m, n, k, lhs_offset,
36161                 internal::gemm_q8_1_0_1(scratch, lhs, rhs, m, n, k, lhs_offset,
36167                 internal::gemm_q8_1_0_2(scratch, lhs, rhs, m, n, k, lhs_offset,
36173                 internal::gemm_q8_1_0_3(scratch, lhs, rhs, m, n, k, lhs_offset,
36179                 internal::gemm_q8_1_0_4(scratch, lhs, rhs, m, n, k, lhs_offset,
36185                 internal::gemm_q8_1_0_5(scratch, lhs, rhs, m, n, k, lhs_offset,
36191                 internal::gemm_q8_1_0_6(scratch, lhs, rhs, m, n, k, lhs_offset,
36197                 internal::gemm_q8_1_0_7(scratch, lhs, rhs, m, n, k, lhs_offset,
36207                 internal::gemm_q8_1_1_0(scratch, lhs, rhs, m, n, k, lhs_offset,
36213                 internal::gemm_q8_1_1_1(scratch, lhs, rhs, m, n, k, lhs_offset,
36219                 internal::gemm_q8_1_1_2(scratch, lhs, rhs, m, n, k, lhs_offset,
36225                 internal::gemm_q8_1_1_3(scratch, lhs, rhs, m, n, k, lhs_offset,
36231                 internal::gemm_q8_1_1_4(scratch, lhs, rhs, m, n, k, lhs_offset,
36237                 internal::gemm_q8_1_1_5(scratch, lhs, rhs, m, n, k, lhs_offset,
36243                 internal::gemm_q8_1_1_6(scratch, lhs, rhs, m, n, k, lhs_offset,
36249                 internal::gemm_q8_1_1_7(scratch, lhs, rhs, m, n, k, lhs_offset,
36259                 internal::gemm_q8_1_2_0(scratch, lhs, rhs, m, n, k, lhs_offset,
36265                 internal::gemm_q8_1_2_1(scratch, lhs, rhs, m, n, k, lhs_offset,
36271                 internal::gemm_q8_1_2_2(scratch, lhs, rhs, m, n, k, lhs_offset,
36277                 internal::gemm_q8_1_2_3(scratch, lhs, rhs, m, n, k, lhs_offset,
36283                 internal::gemm_q8_1_2_4(scratch, lhs, rhs, m, n, k, lhs_offset,
36289                 internal::gemm_q8_1_2_5(scratch, lhs, rhs, m, n, k, lhs_offset,
36295                 internal::gemm_q8_1_2_6(scratch, lhs, rhs, m, n, k, lhs_offset,
36301                 internal::gemm_q8_1_2_7(scratch, lhs, rhs, m, n, k, lhs_offset,
36315                 internal::gemm_q8_2_0_0(scratch, lhs, rhs, m, n, k, lhs_offset,
36321                 internal::gemm_q8_2_0_1(scratch, lhs, rhs, m, n, k, lhs_offset,
36327                 internal::gemm_q8_2_0_2(scratch, lhs, rhs, m, n, k, lhs_offset,
36333                 internal::gemm_q8_2_0_3(scratch, lhs, rhs, m, n, k, lhs_offset,
36339                 internal::gemm_q8_2_0_4(scratch, lhs, rhs, m, n, k, lhs_offset,
36345                 internal::gemm_q8_2_0_5(scratch, lhs, rhs, m, n, k, lhs_offset,
36351                 internal::gemm_q8_2_0_6(scratch, lhs, rhs, m, n, k, lhs_offset,
36357                 internal::gemm_q8_2_0_7(scratch, lhs, rhs, m, n, k, lhs_offset,
36367                 internal::gemm_q8_2_1_0(scratch, lhs, rhs, m, n, k, lhs_offset,
36373                 internal::gemm_q8_2_1_1(scratch, lhs, rhs, m, n, k, lhs_offset,
36379                 internal::gemm_q8_2_1_2(scratch, lhs, rhs, m, n, k, lhs_offset,
36385                 internal::gemm_q8_2_1_3(scratch, lhs, rhs, m, n, k, lhs_offset,
36391                 internal::gemm_q8_2_1_4(scratch, lhs, rhs, m, n, k, lhs_offset,
36397                 internal::gemm_q8_2_1_5(scratch, lhs, rhs, m, n, k, lhs_offset,
36403                 internal::gemm_q8_2_1_6(scratch, lhs, rhs, m, n, k, lhs_offset,
36409                 internal::gemm_q8_2_1_7(scratch, lhs, rhs, m, n, k, lhs_offset,
36419                 internal::gemm_q8_2_2_0(scratch, lhs, rhs, m, n, k, lhs_offset,
36425                 internal::gemm_q8_2_2_1(scratch, lhs, rhs, m, n, k, lhs_offset,
36431                 internal::gemm_q8_2_2_2(scratch, lhs, rhs, m, n, k, lhs_offset,
36437                 internal::gemm_q8_2_2_3(scratch, lhs, rhs, m, n, k, lhs_offset,
36443                 internal::gemm_q8_2_2_4(scratch, lhs, rhs, m, n, k, lhs_offset,
36449                 internal::gemm_q8_2_2_5(scratch, lhs, rhs, m, n, k, lhs_offset,
36455                 internal::gemm_q8_2_2_6(scratch, lhs, rhs, m, n, k, lhs_offset,
36461                 internal::gemm_q8_2_2_7(scratch, lhs, rhs, m, n, k, lhs_offset,
36476                       std::int32_t k, std::int32_t lhs_offset,
36491                                                  lhs_offset, rhs_offset, result,
36496                                                  lhs_offset, rhs_offset, result,
36501                                                  lhs_offset, rhs_offset, result,
36506                                                  lhs_offset, rhs_offset, result,
36511                                                  lhs_offset, rhs_offset, result,
36516                                                  lhs_offset, rhs_offset, result,
36521                                                  lhs_offset, rhs_offset, result,
36526                                                  lhs_offset, rhs_offset, result,
36535                                                  lhs_offset, rhs_offset, result,
36540                                                  lhs_offset, rhs_offset, result,
36545                                                  lhs_offset, rhs_offset, result,
36550                                                  lhs_offset, rhs_offset, result,
36555                                                  lhs_offset, rhs_offset, result,
36560                                                  lhs_offset, rhs_offset, result,
36565                                                  lhs_offset, rhs_offset, result,
36570                                                  lhs_offset, rhs_offset, result,
36579                                                  lhs_offset, rhs_offset, result,
36584                                                  lhs_offset, rhs_offset, result,
36589                                                  lhs_offset, rhs_offset, result,
36594                                                  lhs_offset, rhs_offset, result,
36599                                                  lhs_offset, rhs_offset, result,
36604                                                  lhs_offset, rhs_offset, result,
36609                                                  lhs_offset, rhs_offset, result,
36614                                                  lhs_offset, rhs_offset, result,
36627                                                  lhs_offset, rhs_offset, result,
36632                                                  lhs_offset, rhs_offset, result,
36637                                                  lhs_offset, rhs_offset, result,
36642                                                  lhs_offset, rhs_offset, result,
36647                                                  lhs_offset, rhs_offset, result,
36652                                                  lhs_offset, rhs_offset, result,
36657                                                  lhs_offset, rhs_offset, result,
36662                                                  lhs_offset, rhs_offset, result,
36671                                                  lhs_offset, rhs_offset, result,
36676                                                  lhs_offset, rhs_offset, result,
36681                                                  lhs_offset, rhs_offset, result,
36686                                                  lhs_offset, rhs_offset, result,
36691                                                  lhs_offset, rhs_offset, result,
36696                                                  lhs_offset, rhs_offset, result,
36701                                                  lhs_offset, rhs_offset, result,
36706                                                  lhs_offset, rhs_offset, result,
36715                                                  lhs_offset, rhs_offset, result,
36720                                                  lhs_offset, rhs_offset, result,
36725                                                  lhs_offset, rhs_offset, result,
36730                                                  lhs_offset, rhs_offset, result,
36735                                                  lhs_offset, rhs_offset, result,
36740                                                  lhs_offset, rhs_offset, result,
36745                                                  lhs_offset, rhs_offset, result,
36750                                                  lhs_offset, rhs_offset, result,
36763                                                  lhs_offset, rhs_offset, result,
36768                                                  lhs_offset, rhs_offset, result,
36773                                                  lhs_offset, rhs_offset, result,
36778                                                  lhs_offset, rhs_offset, result,
36783                                                  lhs_offset, rhs_offset, result,
36788                                                  lhs_offset, rhs_offset, result,
36793                                                  lhs_offset, rhs_offset, result,
36798                                                  lhs_offset, rhs_offset, result,
36807                                                  lhs_offset, rhs_offset, result,
36812                                                  lhs_offset, rhs_offset, result,
36817                                                  lhs_offset, rhs_offset, result,
36822                                                  lhs_offset, rhs_offset, result,
36827                                                  lhs_offset, rhs_offset, result,
36832                                                  lhs_offset, rhs_offset, result,
36837                                                  lhs_offset, rhs_offset, result,
36842                                                  lhs_offset, rhs_offset, result,
36851                                                  lhs_offset, rhs_offset, result,
36856                                                  lhs_offset, rhs_offset, result,
36861                                                  lhs_offset, rhs_offset, result,
36866                                                  lhs_offset, rhs_offset, result,
36871                                                  lhs_offset, rhs_offset, result,
36876                                                  lhs_offset, rhs_offset, result,
36881                                                  lhs_offset, rhs_offset, result,
36886                                                  lhs_offset, rhs_offset, result,
36901                 internal::gemm_i32_0_0_0(scratch, lhs, rhs, m, n, k, lhs_offset,
36905                 internal::gemm_i32_0_0_1(scratch, lhs, rhs, m, n, k, lhs_offset,
36909 lhs_offset,
36913                 internal::gemm_i32_0_0_3(scratch, lhs, rhs, m, n, k, lhs_offset,
36917                 internal::gemm_i32_0_0_4(scratch, lhs, rhs, m, n, k, lhs_offset,
36921                 internal::gemm_i32_0_0_5(scratch, lhs, rhs, m, n, k, lhs_offset,
36925                 internal::gemm_i32_0_0_6(scratch, lhs, rhs, m, n, k, lhs_offset,
36929                 internal::gemm_i32_0_0_7(scratch, lhs, rhs, m, n, k, lhs_offset,
36937                 internal::gemm_i32_0_1_0(scratch, lhs, rhs, m, n, k, lhs_offset,
36941                 internal::gemm_i32_0_1_1(scratch, lhs, rhs, m, n, k, lhs_offset,
36945                 internal::gemm_i32_0_1_2(scratch, lhs, rhs, m, n, k, lhs_offset,
36949                 internal::gemm_i32_0_1_3(scratch, lhs, rhs, m, n, k, lhs_offset,
36953                 internal::gemm_i32_0_1_4(scratch, lhs, rhs, m, n, k, lhs_offset,
36957                 internal::gemm_i32_0_1_5(scratch, lhs, rhs, m, n, k, lhs_offset,
36961                 internal::gemm_i32_0_1_6(scratch, lhs, rhs, m, n, k, lhs_offset,
36965                 internal::gemm_i32_0_1_7(scratch, lhs, rhs, m, n, k, lhs_offset,
36973                 internal::gemm_i32_0_2_0(scratch, lhs, rhs, m, n, k, lhs_offset,
36977                 internal::gemm_i32_0_2_1(scratch, lhs, rhs, m, n, k, lhs_offset,
36981                 internal::gemm_i32_0_2_2(scratch, lhs, rhs, m, n, k, lhs_offset,
36985                 internal::gemm_i32_0_2_3(scratch, lhs, rhs, m, n, k, lhs_offset,
36989                 internal::gemm_i32_0_2_4(scratch, lhs, rhs, m, n, k, lhs_offset,
36993                 internal::gemm_i32_0_2_5(scratch, lhs, rhs, m, n, k, lhs_offset,
36997                 internal::gemm_i32_0_2_6(scratch, lhs, rhs, m, n, k, lhs_offset,
37001                 internal::gemm_i32_0_2_7(scratch, lhs, rhs, m, n, k, lhs_offset,
37013                 internal::gemm_i32_1_0_0(scratch, lhs, rhs, m, n, k, lhs_offset,
37017                 internal::gemm_i32_1_0_1(scratch, lhs, rhs, m, n, k, lhs_offset,
37021                 internal::gemm_i32_1_0_2(scratch, lhs, rhs, m, n, k, lhs_offset,
37025                 internal::gemm_i32_1_0_3(scratch, lhs, rhs, m, n, k, lhs_offset,
37029                 internal::gemm_i32_1_0_4(scratch, lhs, rhs, m, n, k, lhs_offset,
37033                 internal::gemm_i32_1_0_5(scratch, lhs, rhs, m, n, k, lhs_offset,
37037                 internal::gemm_i32_1_0_6(scratch, lhs, rhs, m, n, k, lhs_offset,
37041                 internal::gemm_i32_1_0_7(scratch, lhs, rhs, m, n, k, lhs_offset,
37049                 internal::gemm_i32_1_1_0(scratch, lhs, rhs, m, n, k, lhs_offset,
37053                 internal::gemm_i32_1_1_1(scratch, lhs, rhs, m, n, k, lhs_offset,
37057                 internal::gemm_i32_1_1_2(scratch, lhs, rhs, m, n, k, lhs_offset,
37061                 internal::gemm_i32_1_1_3(scratch, lhs, rhs, m, n, k, lhs_offset,
37065                 internal::gemm_i32_1_1_4(scratch, lhs, rhs, m, n, k, lhs_offset,
37069                 internal::gemm_i32_1_1_5(scratch, lhs, rhs, m, n, k, lhs_offset,
37073                 internal::gemm_i32_1_1_6(scratch, lhs, rhs, m, n, k, lhs_offset,
37077                 internal::gemm_i32_1_1_7(scratch, lhs, rhs, m, n, k, lhs_offset,
37085                 internal::gemm_i32_1_2_0(scratch, lhs, rhs, m, n, k, lhs_offset,
37089                 internal::gemm_i32_1_2_1(scratch, lhs, rhs, m, n, k, lhs_offset,
37093                 internal::gemm_i32_1_2_2(scratch, lhs, rhs, m, n, k, lhs_offset,
37097                 internal::gemm_i32_1_2_3(scratch, lhs, rhs, m, n, k, lhs_offset,
37101                 internal::gemm_i32_1_2_4(scratch, lhs, rhs, m, n, k, lhs_offset,
37105                 internal::gemm_i32_1_2_5(scratch, lhs, rhs, m, n, k, lhs_offset,
37109                 internal::gemm_i32_1_2_6(scratch, lhs, rhs, m, n, k, lhs_offset,
37113                 internal::gemm_i32_1_2_7(scratch, lhs, rhs, m, n, k, lhs_offset,
37125                 internal::gemm_i32_2_0_0(scratch, lhs, rhs, m, n, k, lhs_offset,
37129                 internal::gemm_i32_2_0_1(scratch, lhs, rhs, m, n, k, lhs_offset,
37133                 internal::gemm_i32_2_0_2(scratch, lhs, rhs, m, n, k, lhs_offset,
37137                 internal::gemm_i32_2_0_3(scratch, lhs, rhs, m, n, k, lhs_offset,
37141                 internal::gemm_i32_2_0_4(scratch, lhs, rhs, m, n, k, lhs_offset,
37145                 internal::gemm_i32_2_0_5(scratch, lhs, rhs, m, n, k, lhs_offset,
37149                 internal::gemm_i32_2_0_6(scratch, lhs, rhs, m, n, k, lhs_offset,
37153                 internal::gemm_i32_2_0_7(scratch, lhs, rhs, m, n, k, lhs_offset,
37161                 internal::gemm_i32_2_1_0(scratch, lhs, rhs, m, n, k, lhs_offset,
37165                 internal::gemm_i32_2_1_1(scratch, lhs, rhs, m, n, k, lhs_offset,
37169                 internal::gemm_i32_2_1_2(scratch, lhs, rhs, m, n, k, lhs_offset,
37173                 internal::gemm_i32_2_1_3(scratch, lhs, rhs, m, n, k, lhs_offset,
37177                 internal::gemm_i32_2_1_4(scratch, lhs, rhs, m, n, k, lhs_offset,
37181                 internal::gemm_i32_2_1_5(scratch, lhs, rhs, m, n, k, lhs_offset,
37185                 internal::gemm_i32_2_1_6(scratch, lhs, rhs, m, n, k, lhs_offset,
37189                 internal::gemm_i32_2_1_7(scratch, lhs, rhs, m, n, k, lhs_offset,
37197                 internal::gemm_i32_2_2_0(scratch, lhs, rhs, m, n, k, lhs_offset,
37201                 internal::gemm_i32_2_2_1(scratch, lhs, rhs, m, n, k, lhs_offset,
37205                 internal::gemm_i32_2_2_2(scratch, lhs, rhs, m, n, k, lhs_offset,
37209                 internal::gemm_i32_2_2_3(scratch, lhs, rhs, m, n, k, lhs_offset,
37213                 internal::gemm_i32_2_2_4(scratch, lhs, rhs, m, n, k, lhs_offset,
37217                 internal::gemm_i32_2_2_5(scratch, lhs, rhs, m, n, k, lhs_offset,
37221                 internal::gemm_i32_2_2_6(scratch, lhs, rhs, m, n, k, lhs_offset,
37225                 internal::gemm_i32_2_2_7(scratch, lhs, rhs, m, n, k, lhs_offset,
37238                     std::int32_t k, std::int32_t lhs_offset,
37253                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37258                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37263                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37268                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37273                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37278                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37283                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37288                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37297                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37302                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37307                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37312                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37317                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37322                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37327                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37332                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37341                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37346                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37351                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37356                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37361                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37366                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37371                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37376                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37389                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37394                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37399                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37404                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37409                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37414                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37419                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37424                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37433                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37438                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37443                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37448                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37453                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37458                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37463                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37468                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37477                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37482                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37487                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37492                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37497                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37502                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37507                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37512                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37525                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37530                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37535                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37540                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37545                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37550                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37555                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37560                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37569                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37574                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37579                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37584                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37589                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37594                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37599                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37604                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37613                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37618                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37623                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37628                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37633                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37638                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37643                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37648                     scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
37663 lhs_offset,
37668                 internal::gemm_f_0_0_1(scratch, lhs, rhs, m, n, k, lhs_offset,
37673                 internal::gemm_f_0_0_2(scratch, lhs, rhs, m, n, k, lhs_offset,
37678                 internal::gemm_f_0_0_3(scratch, lhs, rhs, m, n, k, lhs_offset,
37683                 internal::gemm_f_0_0_4(scratch, lhs, rhs, m, n, k, lhs_offset,
37688                 internal::gemm_f_0_0_5(scratch, lhs, rhs, m, n, k, lhs_offset,
37693                 internal::gemm_f_0_0_6(scratch, lhs, rhs, m, n, k, lhs_offset,
37698                 internal::gemm_f_0_0_7(scratch, lhs, rhs, m, n, k, lhs_offset,
37707                 internal::gemm_f_0_1_0(scratch, lhs, rhs, m, n, k, lhs_offset,
37712                 internal::gemm_f_0_1_1(scratch, lhs, rhs, m, n, k, lhs_offset,
37717                 internal::gemm_f_0_1_2(scratch, lhs, rhs, m, n, k, lhs_offset,
37722                 internal::gemm_f_0_1_3(scratch, lhs, rhs, m, n, k, lhs_offset,
37727                 internal::gemm_f_0_1_4(scratch, lhs, rhs, m, n, k, lhs_offset,
37732                 internal::gemm_f_0_1_5(scratch, lhs, rhs, m, n, k, lhs_offset,
37737                 internal::gemm_f_0_1_6(scratch, lhs, rhs, m, n, k, lhs_offset,
37742                 internal::gemm_f_0_1_7(scratch, lhs, rhs, m, n, k, lhs_offset,
37751                 internal::gemm_f_0_2_0(scratch, lhs, rhs, m, n, k, lhs_offset,
37756                 internal::gemm_f_0_2_1(scratch, lhs, rhs, m, n, k, lhs_offset,
37761                 internal::gemm_f_0_2_2(scratch, lhs, rhs, m, n, k, lhs_offset,
37766                 internal::gemm_f_0_2_3(scratch, lhs, rhs, m, n, k, lhs_offset,
37771                 internal::gemm_f_0_2_4(scratch, lhs, rhs, m, n, k, lhs_offset,
37776                 internal::gemm_f_0_2_5(scratch, lhs, rhs, m, n, k, lhs_offset,
37781                 internal::gemm_f_0_2_6(scratch, lhs, rhs, m, n, k, lhs_offset,
37786                 internal::gemm_f_0_2_7(scratch, lhs, rhs, m, n, k, lhs_offset,
37799                 internal::gemm_f_1_0_0(scratch, lhs, rhs, m, n, k, lhs_offset,
37804                 internal::gemm_f_1_0_1(scratch, lhs, rhs, m, n, k, lhs_offset,
37809                 internal::gemm_f_1_0_2(scratch, lhs, rhs, m, n, k, lhs_offset,
37814                 internal::gemm_f_1_0_3(scratch, lhs, rhs, m, n, k, lhs_offset,
37819                 internal::gemm_f_1_0_4(scratch, lhs, rhs, m, n, k, lhs_offset,
37824                 internal::gemm_f_1_0_5(scratch, lhs, rhs, m, n, k, lhs_offset,
37829                 internal::gemm_f_1_0_6(scratch, lhs, rhs, m, n, k, lhs_offset,
37834                 internal::gemm_f_1_0_7(scratch, lhs, rhs, m, n, k, lhs_offset,
37843                 internal::gemm_f_1_1_0(scratch, lhs, rhs, m, n, k, lhs_offset,
37848                 internal::gemm_f_1_1_1(scratch, lhs, rhs, m, n, k, lhs_offset,
37853                 internal::gemm_f_1_1_2(scratch, lhs, rhs, m, n, k, lhs_offset,
37858                 internal::gemm_f_1_1_3(scratch, lhs, rhs, m, n, k, lhs_offset,
37863                 internal::gemm_f_1_1_4(scratch, lhs, rhs, m, n, k, lhs_offset,
37868                 internal::gemm_f_1_1_5(scratch, lhs, rhs, m, n, k, lhs_offset,
37873                 internal::gemm_f_1_1_6(scratch, lhs, rhs, m, n, k, lhs_offset,
37878                 internal::gemm_f_1_1_7(scratch, lhs, rhs, m, n, k, lhs_offset,
37887                 internal::gemm_f_1_2_0(scratch, lhs, rhs, m, n, k, lhs_offset,
37892                 internal::gemm_f_1_2_1(scratch, lhs, rhs, m, n, k, lhs_offset,
37897                 internal::gemm_f_1_2_2(scratch, lhs, rhs, m, n, k, lhs_offset,
37902                 internal::gemm_f_1_2_3(scratch, lhs, rhs, m, n, k, lhs_offset,
37907                 internal::gemm_f_1_2_4(scratch, lhs, rhs, m, n, k, lhs_offset,
37912                 internal::gemm_f_1_2_5(scratch, lhs, rhs, m, n, k, lhs_offset,
37917                 internal::gemm_f_1_2_6(scratch, lhs, rhs, m, n, k, lhs_offset,
37922                 internal::gemm_f_1_2_7(scratch, lhs, rhs, m, n, k, lhs_offset,
37935                 internal::gemm_f_2_0_0(scratch, lhs, rhs, m, n, k, lhs_offset,
37940                 internal::gemm_f_2_0_1(scratch, lhs, rhs, m, n, k, lhs_offset,
37945                 internal::gemm_f_2_0_2(scratch, lhs, rhs, m, n, k, lhs_offset,
37950                 internal::gemm_f_2_0_3(scratch, lhs, rhs, m, n, k, lhs_offset,
37955                 internal::gemm_f_2_0_4(scratch, lhs, rhs, m, n, k, lhs_offset,
37960                 internal::gemm_f_2_0_5(scratch, lhs, rhs, m, n, k, lhs_offset,
37965                 internal::gemm_f_2_0_6(scratch, lhs, rhs, m, n, k, lhs_offset,
37970                 internal::gemm_f_2_0_7(scratch, lhs, rhs, m, n, k, lhs_offset,
37979                 internal::gemm_f_2_1_0(scratch, lhs, rhs, m, n, k, lhs_offset,
37984                 internal::gemm_f_2_1_1(scratch, lhs, rhs, m, n, k, lhs_offset,
37989                 internal::gemm_f_2_1_2(scratch, lhs, rhs, m, n, k, lhs_offset,
37994                 internal::gemm_f_2_1_3(scratch, lhs, rhs, m, n, k, lhs_offset,
37999                 internal::gemm_f_2_1_4(scratch, lhs, rhs, m, n, k, lhs_offset,
38004                 internal::gemm_f_2_1_5(scratch, lhs, rhs, m, n, k, lhs_offset,
38009                 internal::gemm_f_2_1_6(scratch, lhs, rhs, m, n, k, lhs_offset,
38014                 internal::gemm_f_2_1_7(scratch, lhs, rhs, m, n, k, lhs_offset,
38023                 internal::gemm_f_2_2_0(scratch, lhs, rhs, m, n, k, lhs_offset,
38028                 internal::gemm_f_2_2_1(scratch, lhs, rhs, m, n, k, lhs_offset,
38033                 internal::gemm_f_2_2_2(scratch, lhs, rhs, m, n, k, lhs_offset,
38038                 internal::gemm_f_2_2_3(scratch, lhs, rhs, m, n, k, lhs_offset,
38043                 internal::gemm_f_2_2_4(scratch, lhs, rhs, m, n, k, lhs_offset,
38048                 internal::gemm_f_2_2_5(scratch, lhs, rhs, m, n, k, lhs_offset,
38053                 internal::gemm_f_2_2_6(scratch, lhs, rhs, m, n, k, lhs_offset,
38058                 internal::gemm_f_2_2_7(scratch, lhs, rhs, m, n, k, lhs_offset,
38072              std::int32_t k, std::int32_t lhs_offset, std::int32_t rhs_offset,
38075   gemm_q8_strided(scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,
38081               std::int32_t k, std::int32_t lhs_offset, std::int32_t rhs_offset,
38083   gemm_i32_strided(scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, result,
38089             std::int32_t k, std::int32_t lhs_offset, std::int32_t rhs_offset,
38091   gemm_f_strided(scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset,