Lines Matching refs:lhs_offset
8604 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_0_0_0_aligned() argument
8624 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_0_0_aligned()
8633 zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_0_0_aligned()
8659 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_0_0_1_aligned() argument
8679 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_0_1_aligned()
8688 zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_0_1_aligned()
8714 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_0_0_2_aligned() argument
8734 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_0_2_aligned()
8743 zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_0_2_aligned()
8769 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_0_0_3_aligned() argument
8789 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_0_3_aligned()
8798 zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_0_3_aligned()
8824 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_0_0_4_aligned() argument
8844 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_0_4_aligned()
8853 zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_0_4_aligned()
8879 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_0_0_5_aligned() argument
8899 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_0_5_aligned()
8908 zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_0_5_aligned()
8934 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_0_0_6_aligned() argument
8954 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_0_6_aligned()
8963 zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_0_6_aligned()
8989 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_0_0_7_aligned() argument
9009 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_0_7_aligned()
9018 zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_0_7_aligned()
9044 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_0_1_0_aligned() argument
9064 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_1_0_aligned()
9073 zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_1_0_aligned()
9077 zip_1x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_1_0_aligned()
9102 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_0_1_1_aligned() argument
9122 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_1_1_aligned()
9131 zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_1_1_aligned()
9135 zip_1x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_1_1_aligned()
9160 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_0_1_2_aligned() argument
9180 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_1_2_aligned()
9189 zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_1_2_aligned()
9193 zip_1x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_1_2_aligned()
9218 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_0_1_3_aligned() argument
9238 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_1_3_aligned()
9247 zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_1_3_aligned()
9251 zip_1x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_1_3_aligned()
9276 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_0_1_4_aligned() argument
9296 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_1_4_aligned()
9305 zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_1_4_aligned()
9309 zip_1x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_1_4_aligned()
9334 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_0_1_5_aligned() argument
9354 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_1_5_aligned()
9363 zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_1_5_aligned()
9367 zip_1x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_1_5_aligned()
9392 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_0_1_6_aligned() argument
9412 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_1_6_aligned()
9421 zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_1_6_aligned()
9425 zip_1x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_1_6_aligned()
9450 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_0_1_7_aligned() argument
9470 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_1_7_aligned()
9479 zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_1_7_aligned()
9483 zip_1x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_1_7_aligned()
9508 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_0_2_0_aligned() argument
9528 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_2_0_aligned()
9537 zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_2_0_aligned()
9541 zip_2x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_2_0_aligned()
9566 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_0_2_1_aligned() argument
9586 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_2_1_aligned()
9595 zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_2_1_aligned()
9599 zip_2x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_2_1_aligned()
9624 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_0_2_2_aligned() argument
9644 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_2_2_aligned()
9653 zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_2_2_aligned()
9657 zip_2x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_2_2_aligned()
9682 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_0_2_3_aligned() argument
9702 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_2_3_aligned()
9711 zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_2_3_aligned()
9715 zip_2x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_2_3_aligned()
9740 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_0_2_4_aligned() argument
9760 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_2_4_aligned()
9769 zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_2_4_aligned()
9773 zip_2x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_2_4_aligned()
9798 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_0_2_5_aligned() argument
9818 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_2_5_aligned()
9827 zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_2_5_aligned()
9831 zip_2x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_2_5_aligned()
9856 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_0_2_6_aligned() argument
9876 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_2_6_aligned()
9885 zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_2_6_aligned()
9889 zip_2x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_2_6_aligned()
9914 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_0_2_7_aligned() argument
9934 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_2_7_aligned()
9943 zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_2_7_aligned()
9947 zip_2x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_2_7_aligned()
9972 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_1_0_0_aligned() argument
9994 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_0_0_aligned()
10003 zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_0_0_aligned()
10042 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_1_0_1_aligned() argument
10064 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_0_1_aligned()
10073 zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_0_1_aligned()
10112 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_1_0_2_aligned() argument
10134 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_0_2_aligned()
10143 zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_0_2_aligned()
10182 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_1_0_3_aligned() argument
10204 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_0_3_aligned()
10213 zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_0_3_aligned()
10252 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_1_0_4_aligned() argument
10274 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_0_4_aligned()
10283 zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_0_4_aligned()
10322 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_1_0_5_aligned() argument
10344 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_0_5_aligned()
10353 zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_0_5_aligned()
10392 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_1_0_6_aligned() argument
10414 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_0_6_aligned()
10423 zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_0_6_aligned()
10462 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_1_0_7_aligned() argument
10484 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_0_7_aligned()
10493 zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_0_7_aligned()
10532 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_1_1_0_aligned() argument
10554 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_1_0_aligned()
10563 zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_1_0_aligned()
10567 zip_1x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_1_0_aligned()
10607 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_1_1_1_aligned() argument
10629 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_1_1_aligned()
10638 zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_1_1_aligned()
10642 zip_1x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_1_1_aligned()
10682 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_1_1_2_aligned() argument
10704 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_1_2_aligned()
10713 zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_1_2_aligned()
10717 zip_1x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_1_2_aligned()
10757 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_1_1_3_aligned() argument
10779 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_1_3_aligned()
10788 zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_1_3_aligned()
10792 zip_1x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_1_3_aligned()
10832 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_1_1_4_aligned() argument
10854 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_1_4_aligned()
10863 zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_1_4_aligned()
10867 zip_1x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_1_4_aligned()
10907 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_1_1_5_aligned() argument
10929 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_1_5_aligned()
10938 zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_1_5_aligned()
10942 zip_1x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_1_5_aligned()
10982 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_1_1_6_aligned() argument
11004 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_1_6_aligned()
11013 zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_1_6_aligned()
11017 zip_1x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_1_6_aligned()
11057 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_1_1_7_aligned() argument
11079 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_1_7_aligned()
11088 zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_1_7_aligned()
11092 zip_1x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_1_7_aligned()
11132 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_1_2_0_aligned() argument
11154 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_2_0_aligned()
11163 zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_2_0_aligned()
11167 zip_2x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_2_0_aligned()
11207 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_1_2_1_aligned() argument
11229 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_2_1_aligned()
11238 zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_2_1_aligned()
11242 zip_2x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_2_1_aligned()
11282 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_1_2_2_aligned() argument
11304 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_2_2_aligned()
11313 zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_2_2_aligned()
11317 zip_2x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_2_2_aligned()
11357 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_1_2_3_aligned() argument
11379 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_2_3_aligned()
11388 zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_2_3_aligned()
11392 zip_2x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_2_3_aligned()
11432 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_1_2_4_aligned() argument
11454 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_2_4_aligned()
11463 zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_2_4_aligned()
11467 zip_2x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_2_4_aligned()
11507 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_1_2_5_aligned() argument
11529 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_2_5_aligned()
11538 zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_2_5_aligned()
11542 zip_2x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_2_5_aligned()
11582 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_1_2_6_aligned() argument
11604 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_2_6_aligned()
11613 zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_2_6_aligned()
11617 zip_2x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_2_6_aligned()
11657 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_1_2_7_aligned() argument
11679 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_2_7_aligned()
11688 zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_2_7_aligned()
11692 zip_2x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_2_7_aligned()
11732 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_2_0_0_aligned() argument
11754 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_0_0_aligned()
11763 zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_0_0_aligned()
11802 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_2_0_1_aligned() argument
11824 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_0_1_aligned()
11833 zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_0_1_aligned()
11872 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_2_0_2_aligned() argument
11894 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_0_2_aligned()
11903 zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_0_2_aligned()
11942 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_2_0_3_aligned() argument
11964 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_0_3_aligned()
11973 zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_0_3_aligned()
12012 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_2_0_4_aligned() argument
12034 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_0_4_aligned()
12043 zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_0_4_aligned()
12082 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_2_0_5_aligned() argument
12104 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_0_5_aligned()
12113 zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_0_5_aligned()
12152 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_2_0_6_aligned() argument
12174 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_0_6_aligned()
12183 zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_0_6_aligned()
12222 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_2_0_7_aligned() argument
12244 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_0_7_aligned()
12253 zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_0_7_aligned()
12292 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_2_1_0_aligned() argument
12314 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_1_0_aligned()
12323 zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_1_0_aligned()
12327 zip_1x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_1_0_aligned()
12367 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_2_1_1_aligned() argument
12389 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_1_1_aligned()
12398 zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_1_1_aligned()
12402 zip_1x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_1_1_aligned()
12442 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_2_1_2_aligned() argument
12464 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_1_2_aligned()
12473 zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_1_2_aligned()
12477 zip_1x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_1_2_aligned()
12517 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_2_1_3_aligned() argument
12539 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_1_3_aligned()
12548 zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_1_3_aligned()
12552 zip_1x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_1_3_aligned()
12592 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_2_1_4_aligned() argument
12614 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_1_4_aligned()
12623 zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_1_4_aligned()
12627 zip_1x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_1_4_aligned()
12667 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_2_1_5_aligned() argument
12689 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_1_5_aligned()
12698 zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_1_5_aligned()
12702 zip_1x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_1_5_aligned()
12742 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_2_1_6_aligned() argument
12764 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_1_6_aligned()
12773 zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_1_6_aligned()
12777 zip_1x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_1_6_aligned()
12817 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_2_1_7_aligned() argument
12839 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_1_7_aligned()
12848 zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_1_7_aligned()
12852 zip_1x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_1_7_aligned()
12892 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_2_2_0_aligned() argument
12914 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_2_0_aligned()
12923 zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_2_0_aligned()
12927 zip_2x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_2_0_aligned()
12967 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_2_2_1_aligned() argument
12989 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_2_1_aligned()
12998 zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_2_1_aligned()
13002 zip_2x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_2_1_aligned()
13042 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_2_2_2_aligned() argument
13064 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_2_2_aligned()
13073 zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_2_2_aligned()
13077 zip_2x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_2_2_aligned()
13117 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_2_2_3_aligned() argument
13139 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_2_3_aligned()
13148 zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_2_3_aligned()
13152 zip_2x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_2_3_aligned()
13192 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_2_2_4_aligned() argument
13214 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_2_4_aligned()
13223 zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_2_4_aligned()
13227 zip_2x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_2_4_aligned()
13267 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_2_2_5_aligned() argument
13289 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_2_5_aligned()
13298 zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_2_5_aligned()
13302 zip_2x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_2_5_aligned()
13342 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_2_2_6_aligned() argument
13364 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_2_6_aligned()
13373 zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_2_6_aligned()
13377 zip_2x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_2_6_aligned()
13417 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8_2_2_7_aligned() argument
13439 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_2_7_aligned()
13448 zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_2_7_aligned()
13452 zip_2x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_2_7_aligned()
13491 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_0_0_0() argument
13510 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_0_0()
13519 zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_0_0()
13544 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_0_0_1() argument
13563 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_0_1()
13572 zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_0_1()
13597 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_0_0_2() argument
13616 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_0_2()
13625 zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_0_2()
13650 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_0_0_3() argument
13669 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_0_3()
13678 zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_0_3()
13703 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_0_0_4() argument
13722 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_0_4()
13731 zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_0_4()
13756 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_0_0_5() argument
13775 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_0_5()
13784 zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_0_5()
13809 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_0_0_6() argument
13828 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_0_6()
13837 zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_0_6()
13862 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_0_0_7() argument
13881 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_0_7()
13890 zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_0_7()
13915 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_0_1_0() argument
13934 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_1_0()
13943 zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_1_0()
13947 zip_1x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_1_0()
13971 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_0_1_1() argument
13990 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_1_1()
13999 zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_1_1()
14003 zip_1x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_1_1()
14027 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_0_1_2() argument
14046 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_1_2()
14055 zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_1_2()
14059 zip_1x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_1_2()
14083 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_0_1_3() argument
14102 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_1_3()
14111 zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_1_3()
14115 zip_1x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_1_3()
14139 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_0_1_4() argument
14158 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_1_4()
14167 zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_1_4()
14171 zip_1x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_1_4()
14195 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_0_1_5() argument
14214 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_1_5()
14223 zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_1_5()
14227 zip_1x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_1_5()
14251 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_0_1_6() argument
14270 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_1_6()
14279 zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_1_6()
14283 zip_1x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_1_6()
14307 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_0_1_7() argument
14326 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_1_7()
14335 zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_1_7()
14339 zip_1x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_1_7()
14363 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_0_2_0() argument
14382 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_2_0()
14391 zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_2_0()
14395 zip_2x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_2_0()
14419 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_0_2_1() argument
14438 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_2_1()
14447 zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_2_1()
14451 zip_2x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_2_1()
14475 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_0_2_2() argument
14494 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_2_2()
14503 zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_2_2()
14507 zip_2x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_2_2()
14531 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_0_2_3() argument
14550 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_2_3()
14559 zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_2_3()
14563 zip_2x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_2_3()
14587 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_0_2_4() argument
14606 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_2_4()
14615 zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_2_4()
14619 zip_2x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_2_4()
14643 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_0_2_5() argument
14662 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_2_5()
14671 zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_2_5()
14675 zip_2x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_2_5()
14699 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_0_2_6() argument
14718 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_2_6()
14727 zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_2_6()
14731 zip_2x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_2_6()
14755 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_0_2_7() argument
14774 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_0_2_7()
14783 zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_2_7()
14787 zip_2x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_0_2_7()
14811 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_1_0_0() argument
14832 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_0_0()
14841 zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_0_0()
14879 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_1_0_1() argument
14900 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_0_1()
14909 zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_0_1()
14947 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_1_0_2() argument
14968 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_0_2()
14977 zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_0_2()
15015 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_1_0_3() argument
15036 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_0_3()
15045 zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_0_3()
15083 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_1_0_4() argument
15104 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_0_4()
15113 zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_0_4()
15151 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_1_0_5() argument
15172 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_0_5()
15181 zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_0_5()
15219 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_1_0_6() argument
15240 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_0_6()
15249 zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_0_6()
15287 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_1_0_7() argument
15308 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_0_7()
15317 zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_0_7()
15355 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_1_1_0() argument
15376 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_1_0()
15385 zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_1_0()
15389 zip_1x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_1_0()
15428 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_1_1_1() argument
15449 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_1_1()
15458 zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_1_1()
15462 zip_1x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_1_1()
15501 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_1_1_2() argument
15522 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_1_2()
15531 zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_1_2()
15535 zip_1x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_1_2()
15574 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_1_1_3() argument
15595 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_1_3()
15604 zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_1_3()
15608 zip_1x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_1_3()
15647 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_1_1_4() argument
15668 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_1_4()
15677 zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_1_4()
15681 zip_1x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_1_4()
15720 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_1_1_5() argument
15741 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_1_5()
15750 zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_1_5()
15754 zip_1x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_1_5()
15793 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_1_1_6() argument
15814 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_1_6()
15823 zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_1_6()
15827 zip_1x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_1_6()
15866 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_1_1_7() argument
15887 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_1_7()
15896 zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_1_7()
15900 zip_1x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_1_7()
15939 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_1_2_0() argument
15960 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_2_0()
15969 zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_2_0()
15973 zip_2x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_2_0()
16012 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_1_2_1() argument
16033 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_2_1()
16042 zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_2_1()
16046 zip_2x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_2_1()
16085 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_1_2_2() argument
16106 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_2_2()
16115 zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_2_2()
16119 zip_2x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_2_2()
16158 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_1_2_3() argument
16179 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_2_3()
16188 zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_2_3()
16192 zip_2x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_2_3()
16231 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_1_2_4() argument
16252 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_2_4()
16261 zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_2_4()
16265 zip_2x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_2_4()
16304 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_1_2_5() argument
16325 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_2_5()
16334 zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_2_5()
16338 zip_2x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_2_5()
16377 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_1_2_6() argument
16398 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_2_6()
16407 zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_2_6()
16411 zip_2x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_2_6()
16450 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_1_2_7() argument
16471 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_1_2_7()
16480 zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_2_7()
16484 zip_2x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_1_2_7()
16523 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_2_0_0() argument
16544 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_0_0()
16553 zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_0_0()
16591 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_2_0_1() argument
16612 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_0_1()
16621 zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_0_1()
16659 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_2_0_2() argument
16680 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_0_2()
16689 zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_0_2()
16727 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_2_0_3() argument
16748 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_0_3()
16757 zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_0_3()
16795 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_2_0_4() argument
16816 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_0_4()
16825 zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_0_4()
16863 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_2_0_5() argument
16884 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_0_5()
16893 zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_0_5()
16931 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_2_0_6() argument
16952 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_0_6()
16961 zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_0_6()
16999 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_2_0_7() argument
17020 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_0_7()
17029 zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_0_7()
17067 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_2_1_0() argument
17088 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_1_0()
17097 zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_1_0()
17101 zip_1x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_1_0()
17140 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_2_1_1() argument
17161 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_1_1()
17170 zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_1_1()
17174 zip_1x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_1_1()
17213 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_2_1_2() argument
17234 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_1_2()
17243 zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_1_2()
17247 zip_1x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_1_2()
17286 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_2_1_3() argument
17307 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_1_3()
17316 zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_1_3()
17320 zip_1x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_1_3()
17359 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_2_1_4() argument
17380 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_1_4()
17389 zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_1_4()
17393 zip_1x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_1_4()
17432 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_2_1_5() argument
17453 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_1_5()
17462 zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_1_5()
17466 zip_1x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_1_5()
17505 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_2_1_6() argument
17526 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_1_6()
17535 zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_1_6()
17539 zip_1x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_1_6()
17578 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_2_1_7() argument
17599 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_1_7()
17608 zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_1_7()
17612 zip_1x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_1_7()
17651 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_2_2_0() argument
17672 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_2_0()
17681 zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_2_0()
17685 zip_2x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_2_0()
17724 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_2_2_1() argument
17745 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_2_1()
17754 zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_2_1()
17758 zip_2x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_2_1()
17797 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_2_2_2() argument
17818 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_2_2()
17827 zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_2_2()
17831 zip_2x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_2_2()
17870 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_2_2_3() argument
17891 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_2_3()
17900 zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_2_3()
17904 zip_2x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_2_3()
17943 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_2_2_4() argument
17964 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_2_4()
17973 zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_2_4()
17977 zip_2x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_2_4()
18016 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_2_2_5() argument
18037 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_2_5()
18046 zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_2_5()
18050 zip_2x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_2_5()
18089 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_2_2_6() argument
18110 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_2_6()
18119 zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_2_6()
18123 zip_2x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_2_6()
18162 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_2_2_7() argument
18183 const std::int32_t const_offset = lhs_offset * rhs_offset * k + result_offset; in gemm_q8_2_2_7()
18192 zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_2_7()
18196 zip_2x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_q8_2_2_7()
18236 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_0_0_0_aligned() argument
18253 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_0_0_aligned()
18259 zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_0_0_aligned()
18283 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_0_0_1_aligned() argument
18300 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_0_1_aligned()
18306 zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_0_1_aligned()
18330 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_0_0_2_aligned() argument
18347 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_0_2_aligned()
18353 zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_0_2_aligned()
18377 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_0_0_3_aligned() argument
18394 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_0_3_aligned()
18400 zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_0_3_aligned()
18424 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_0_0_4_aligned() argument
18441 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_0_4_aligned()
18447 zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_0_4_aligned()
18471 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_0_0_5_aligned() argument
18488 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_0_5_aligned()
18494 zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_0_5_aligned()
18518 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_0_0_6_aligned() argument
18535 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_0_6_aligned()
18541 zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_0_6_aligned()
18565 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_0_0_7_aligned() argument
18582 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_0_7_aligned()
18588 zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_0_7_aligned()
18612 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_0_1_0_aligned() argument
18629 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_1_0_aligned()
18635 zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_1_0_aligned()
18639 zip_1x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_1_0_aligned()
18663 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_0_1_1_aligned() argument
18680 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_1_1_aligned()
18686 zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_1_1_aligned()
18690 zip_1x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_1_1_aligned()
18714 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_0_1_2_aligned() argument
18731 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_1_2_aligned()
18737 zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_1_2_aligned()
18741 zip_1x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_1_2_aligned()
18765 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_0_1_3_aligned() argument
18782 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_1_3_aligned()
18788 zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_1_3_aligned()
18792 zip_1x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_1_3_aligned()
18816 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_0_1_4_aligned() argument
18833 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_1_4_aligned()
18839 zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_1_4_aligned()
18843 zip_1x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_1_4_aligned()
18867 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_0_1_5_aligned() argument
18884 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_1_5_aligned()
18890 zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_1_5_aligned()
18894 zip_1x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_1_5_aligned()
18918 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_0_1_6_aligned() argument
18935 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_1_6_aligned()
18941 zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_1_6_aligned()
18945 zip_1x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_1_6_aligned()
18969 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_0_1_7_aligned() argument
18986 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_1_7_aligned()
18992 zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_1_7_aligned()
18996 zip_1x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_1_7_aligned()
19020 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_0_2_0_aligned() argument
19037 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_2_0_aligned()
19043 zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_2_0_aligned()
19047 zip_2x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_2_0_aligned()
19071 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_0_2_1_aligned() argument
19088 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_2_1_aligned()
19094 zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_2_1_aligned()
19098 zip_2x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_2_1_aligned()
19122 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_0_2_2_aligned() argument
19139 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_2_2_aligned()
19145 zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_2_2_aligned()
19149 zip_2x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_2_2_aligned()
19173 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_0_2_3_aligned() argument
19190 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_2_3_aligned()
19196 zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_2_3_aligned()
19200 zip_2x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_2_3_aligned()
19224 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_0_2_4_aligned() argument
19241 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_2_4_aligned()
19247 zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_2_4_aligned()
19251 zip_2x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_2_4_aligned()
19275 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_0_2_5_aligned() argument
19292 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_2_5_aligned()
19298 zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_2_5_aligned()
19302 zip_2x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_2_5_aligned()
19326 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_0_2_6_aligned() argument
19343 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_2_6_aligned()
19349 zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_2_6_aligned()
19353 zip_2x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_2_6_aligned()
19377 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_0_2_7_aligned() argument
19394 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_2_7_aligned()
19400 zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_2_7_aligned()
19404 zip_2x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_2_7_aligned()
19428 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_1_0_0_aligned() argument
19447 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_0_0_aligned()
19453 zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_0_0_aligned()
19488 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_1_0_1_aligned() argument
19507 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_0_1_aligned()
19513 zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_0_1_aligned()
19548 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_1_0_2_aligned() argument
19567 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_0_2_aligned()
19573 zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_0_2_aligned()
19608 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_1_0_3_aligned() argument
19627 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_0_3_aligned()
19633 zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_0_3_aligned()
19668 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_1_0_4_aligned() argument
19687 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_0_4_aligned()
19693 zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_0_4_aligned()
19728 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_1_0_5_aligned() argument
19747 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_0_5_aligned()
19753 zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_0_5_aligned()
19788 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_1_0_6_aligned() argument
19807 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_0_6_aligned()
19813 zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_0_6_aligned()
19848 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_1_0_7_aligned() argument
19867 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_0_7_aligned()
19873 zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_0_7_aligned()
19908 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_1_1_0_aligned() argument
19927 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_1_0_aligned()
19933 zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_1_0_aligned()
19937 zip_1x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_1_0_aligned()
19975 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_1_1_1_aligned() argument
19994 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_1_1_aligned()
20000 zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_1_1_aligned()
20004 zip_1x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_1_1_aligned()
20042 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_1_1_2_aligned() argument
20061 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_1_2_aligned()
20067 zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_1_2_aligned()
20071 zip_1x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_1_2_aligned()
20109 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_1_1_3_aligned() argument
20128 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_1_3_aligned()
20134 zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_1_3_aligned()
20138 zip_1x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_1_3_aligned()
20176 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_1_1_4_aligned() argument
20195 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_1_4_aligned()
20201 zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_1_4_aligned()
20205 zip_1x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_1_4_aligned()
20243 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_1_1_5_aligned() argument
20262 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_1_5_aligned()
20268 zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_1_5_aligned()
20272 zip_1x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_1_5_aligned()
20310 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_1_1_6_aligned() argument
20329 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_1_6_aligned()
20335 zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_1_6_aligned()
20339 zip_1x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_1_6_aligned()
20377 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_1_1_7_aligned() argument
20396 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_1_7_aligned()
20402 zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_1_7_aligned()
20406 zip_1x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_1_7_aligned()
20444 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_1_2_0_aligned() argument
20463 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_2_0_aligned()
20469 zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_2_0_aligned()
20473 zip_2x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_2_0_aligned()
20511 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_1_2_1_aligned() argument
20530 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_2_1_aligned()
20536 zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_2_1_aligned()
20540 zip_2x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_2_1_aligned()
20578 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_1_2_2_aligned() argument
20597 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_2_2_aligned()
20603 zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_2_2_aligned()
20607 zip_2x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_2_2_aligned()
20645 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_1_2_3_aligned() argument
20664 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_2_3_aligned()
20670 zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_2_3_aligned()
20674 zip_2x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_2_3_aligned()
20712 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_1_2_4_aligned() argument
20731 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_2_4_aligned()
20737 zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_2_4_aligned()
20741 zip_2x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_2_4_aligned()
20779 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_1_2_5_aligned() argument
20798 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_2_5_aligned()
20804 zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_2_5_aligned()
20808 zip_2x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_2_5_aligned()
20846 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_1_2_6_aligned() argument
20865 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_2_6_aligned()
20871 zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_2_6_aligned()
20875 zip_2x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_2_6_aligned()
20913 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_1_2_7_aligned() argument
20932 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_2_7_aligned()
20938 zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_2_7_aligned()
20942 zip_2x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_2_7_aligned()
20980 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_2_0_0_aligned() argument
20999 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_0_0_aligned()
21005 zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_0_0_aligned()
21040 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_2_0_1_aligned() argument
21059 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_0_1_aligned()
21065 zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_0_1_aligned()
21100 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_2_0_2_aligned() argument
21119 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_0_2_aligned()
21125 zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_0_2_aligned()
21160 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_2_0_3_aligned() argument
21179 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_0_3_aligned()
21185 zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_0_3_aligned()
21220 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_2_0_4_aligned() argument
21239 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_0_4_aligned()
21245 zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_0_4_aligned()
21280 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_2_0_5_aligned() argument
21299 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_0_5_aligned()
21305 zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_0_5_aligned()
21340 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_2_0_6_aligned() argument
21359 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_0_6_aligned()
21365 zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_0_6_aligned()
21400 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_2_0_7_aligned() argument
21419 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_0_7_aligned()
21425 zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_0_7_aligned()
21460 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_2_1_0_aligned() argument
21479 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_1_0_aligned()
21485 zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_1_0_aligned()
21489 zip_1x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_1_0_aligned()
21527 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_2_1_1_aligned() argument
21546 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_1_1_aligned()
21552 zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_1_1_aligned()
21556 zip_1x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_1_1_aligned()
21594 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_2_1_2_aligned() argument
21613 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_1_2_aligned()
21619 zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_1_2_aligned()
21623 zip_1x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_1_2_aligned()
21661 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_2_1_3_aligned() argument
21680 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_1_3_aligned()
21686 zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_1_3_aligned()
21690 zip_1x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_1_3_aligned()
21728 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_2_1_4_aligned() argument
21747 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_1_4_aligned()
21753 zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_1_4_aligned()
21757 zip_1x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_1_4_aligned()
21795 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_2_1_5_aligned() argument
21814 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_1_5_aligned()
21820 zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_1_5_aligned()
21824 zip_1x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_1_5_aligned()
21862 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_2_1_6_aligned() argument
21881 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_1_6_aligned()
21887 zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_1_6_aligned()
21891 zip_1x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_1_6_aligned()
21929 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_2_1_7_aligned() argument
21948 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_1_7_aligned()
21954 zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_1_7_aligned()
21958 zip_1x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_1_7_aligned()
21996 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_2_2_0_aligned() argument
22015 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_2_0_aligned()
22021 zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_2_0_aligned()
22025 zip_2x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_2_0_aligned()
22063 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_2_2_1_aligned() argument
22082 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_2_1_aligned()
22088 zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_2_1_aligned()
22092 zip_2x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_2_1_aligned()
22130 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_2_2_2_aligned() argument
22149 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_2_2_aligned()
22155 zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_2_2_aligned()
22159 zip_2x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_2_2_aligned()
22197 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_2_2_3_aligned() argument
22216 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_2_3_aligned()
22222 zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_2_3_aligned()
22226 zip_2x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_2_3_aligned()
22264 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_2_2_4_aligned() argument
22283 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_2_4_aligned()
22289 zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_2_4_aligned()
22293 zip_2x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_2_4_aligned()
22331 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_2_2_5_aligned() argument
22350 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_2_5_aligned()
22356 zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_2_5_aligned()
22360 zip_2x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_2_5_aligned()
22398 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_2_2_6_aligned() argument
22417 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_2_6_aligned()
22423 zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_2_6_aligned()
22427 zip_2x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_2_6_aligned()
22465 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32_2_2_7_aligned() argument
22484 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_2_7_aligned()
22490 zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_2_7_aligned()
22494 zip_2x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_2_7_aligned()
22531 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_0_0_0() argument
22549 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_0_0()
22555 zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_0_0()
22578 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_0_0_1() argument
22596 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_0_1()
22602 zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_0_1()
22625 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_0_0_2() argument
22643 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_0_2()
22649 zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_0_2()
22672 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_0_0_3() argument
22690 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_0_3()
22696 zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_0_3()
22719 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_0_0_4() argument
22737 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_0_4()
22743 zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_0_4()
22766 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_0_0_5() argument
22784 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_0_5()
22790 zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_0_5()
22813 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_0_0_6() argument
22831 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_0_6()
22837 zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_0_6()
22860 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_0_0_7() argument
22878 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_0_7()
22884 zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_0_7()
22907 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_0_1_0() argument
22925 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_1_0()
22931 zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_1_0()
22935 zip_1x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_1_0()
22958 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_0_1_1() argument
22976 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_1_1()
22982 zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_1_1()
22986 zip_1x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_1_1()
23009 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_0_1_2() argument
23027 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_1_2()
23033 zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_1_2()
23037 zip_1x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_1_2()
23060 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_0_1_3() argument
23078 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_1_3()
23084 zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_1_3()
23088 zip_1x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_1_3()
23111 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_0_1_4() argument
23129 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_1_4()
23135 zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_1_4()
23139 zip_1x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_1_4()
23162 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_0_1_5() argument
23180 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_1_5()
23186 zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_1_5()
23190 zip_1x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_1_5()
23213 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_0_1_6() argument
23231 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_1_6()
23237 zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_1_6()
23241 zip_1x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_1_6()
23264 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_0_1_7() argument
23282 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_1_7()
23288 zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_1_7()
23292 zip_1x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_1_7()
23315 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_0_2_0() argument
23333 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_2_0()
23339 zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_2_0()
23343 zip_2x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_2_0()
23366 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_0_2_1() argument
23384 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_2_1()
23390 zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_2_1()
23394 zip_2x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_2_1()
23417 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_0_2_2() argument
23435 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_2_2()
23441 zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_2_2()
23445 zip_2x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_2_2()
23468 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_0_2_3() argument
23486 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_2_3()
23492 zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_2_3()
23496 zip_2x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_2_3()
23519 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_0_2_4() argument
23537 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_2_4()
23543 zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_2_4()
23547 zip_2x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_2_4()
23570 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_0_2_5() argument
23588 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_2_5()
23594 zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_2_5()
23598 zip_2x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_2_5()
23621 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_0_2_6() argument
23639 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_2_6()
23645 zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_2_6()
23649 zip_2x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_2_6()
23672 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_0_2_7() argument
23690 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_0_2_7()
23696 zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_2_7()
23700 zip_2x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_0_2_7()
23723 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_1_0_0() argument
23743 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_0_0()
23749 zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_0_0()
23783 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_1_0_1() argument
23803 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_0_1()
23809 zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_0_1()
23843 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_1_0_2() argument
23863 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_0_2()
23869 zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_0_2()
23903 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_1_0_3() argument
23923 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_0_3()
23929 zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_0_3()
23963 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_1_0_4() argument
23983 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_0_4()
23989 zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_0_4()
24023 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_1_0_5() argument
24043 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_0_5()
24049 zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_0_5()
24083 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_1_0_6() argument
24103 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_0_6()
24109 zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_0_6()
24143 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_1_0_7() argument
24163 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_0_7()
24169 zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_0_7()
24203 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_1_1_0() argument
24223 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_1_0()
24229 zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_1_0()
24233 zip_1x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_1_0()
24270 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_1_1_1() argument
24290 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_1_1()
24296 zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_1_1()
24300 zip_1x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_1_1()
24337 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_1_1_2() argument
24357 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_1_2()
24363 zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_1_2()
24367 zip_1x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_1_2()
24404 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_1_1_3() argument
24424 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_1_3()
24430 zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_1_3()
24434 zip_1x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_1_3()
24471 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_1_1_4() argument
24491 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_1_4()
24497 zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_1_4()
24501 zip_1x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_1_4()
24538 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_1_1_5() argument
24558 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_1_5()
24564 zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_1_5()
24568 zip_1x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_1_5()
24605 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_1_1_6() argument
24625 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_1_6()
24631 zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_1_6()
24635 zip_1x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_1_6()
24672 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_1_1_7() argument
24692 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_1_7()
24698 zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_1_7()
24702 zip_1x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_1_7()
24739 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_1_2_0() argument
24759 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_2_0()
24765 zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_2_0()
24769 zip_2x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_2_0()
24806 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_1_2_1() argument
24826 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_2_1()
24832 zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_2_1()
24836 zip_2x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_2_1()
24873 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_1_2_2() argument
24893 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_2_2()
24899 zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_2_2()
24903 zip_2x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_2_2()
24940 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_1_2_3() argument
24960 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_2_3()
24966 zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_2_3()
24970 zip_2x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_2_3()
25007 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_1_2_4() argument
25027 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_2_4()
25033 zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_2_4()
25037 zip_2x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_2_4()
25074 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_1_2_5() argument
25094 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_2_5()
25100 zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_2_5()
25104 zip_2x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_2_5()
25141 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_1_2_6() argument
25161 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_2_6()
25167 zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_2_6()
25171 zip_2x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_2_6()
25208 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_1_2_7() argument
25228 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_1_2_7()
25234 zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_2_7()
25238 zip_2x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_1_2_7()
25275 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_2_0_0() argument
25295 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_0_0()
25301 zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_0_0()
25335 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_2_0_1() argument
25355 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_0_1()
25361 zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_0_1()
25395 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_2_0_2() argument
25415 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_0_2()
25421 zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_0_2()
25455 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_2_0_3() argument
25475 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_0_3()
25481 zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_0_3()
25515 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_2_0_4() argument
25535 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_0_4()
25541 zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_0_4()
25575 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_2_0_5() argument
25595 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_0_5()
25601 zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_0_5()
25635 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_2_0_6() argument
25655 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_0_6()
25661 zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_0_6()
25695 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_2_0_7() argument
25715 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_0_7()
25721 zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_0_7()
25755 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_2_1_0() argument
25775 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_1_0()
25781 zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_1_0()
25785 zip_1x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_1_0()
25822 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_2_1_1() argument
25842 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_1_1()
25848 zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_1_1()
25852 zip_1x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_1_1()
25889 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_2_1_2() argument
25909 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_1_2()
25915 zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_1_2()
25919 zip_1x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_1_2()
25956 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_2_1_3() argument
25976 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_1_3()
25982 zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_1_3()
25986 zip_1x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_1_3()
26023 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_2_1_4() argument
26043 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_1_4()
26049 zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_1_4()
26053 zip_1x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_1_4()
26090 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_2_1_5() argument
26110 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_1_5()
26116 zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_1_5()
26120 zip_1x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_1_5()
26157 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_2_1_6() argument
26177 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_1_6()
26183 zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_1_6()
26187 zip_1x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_1_6()
26224 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_2_1_7() argument
26244 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_1_7()
26250 zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_1_7()
26254 zip_1x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_1_7()
26291 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_2_2_0() argument
26311 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_2_0()
26317 zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_2_0()
26321 zip_2x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_2_0()
26358 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_2_2_1() argument
26378 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_2_1()
26384 zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_2_1()
26388 zip_2x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_2_1()
26425 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_2_2_2() argument
26445 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_2_2()
26451 zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_2_2()
26455 zip_2x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_2_2()
26492 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_2_2_3() argument
26512 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_2_3()
26518 zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_2_3()
26522 zip_2x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_2_3()
26559 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_2_2_4() argument
26579 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_2_4()
26585 zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_2_4()
26589 zip_2x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_2_4()
26626 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_2_2_5() argument
26646 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_2_5()
26652 zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_2_5()
26656 zip_2x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_2_5()
26693 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_2_2_6() argument
26713 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_2_6()
26719 zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_2_6()
26723 zip_2x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_2_6()
26760 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_2_2_7() argument
26780 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_i32_2_2_7()
26786 zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_2_7()
26790 zip_2x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_i32_2_2_7()
26828 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_0_0_0_aligned() argument
26846 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_0_0_aligned()
26852 zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_0_0_aligned()
26876 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_0_0_1_aligned() argument
26894 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_0_1_aligned()
26900 zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_0_1_aligned()
26924 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_0_0_2_aligned() argument
26942 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_0_2_aligned()
26948 zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_0_2_aligned()
26972 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_0_0_3_aligned() argument
26990 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_0_3_aligned()
26996 zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_0_3_aligned()
27020 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_0_0_4_aligned() argument
27038 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_0_4_aligned()
27044 zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_0_4_aligned()
27068 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_0_0_5_aligned() argument
27086 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_0_5_aligned()
27092 zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_0_5_aligned()
27116 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_0_0_6_aligned() argument
27134 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_0_6_aligned()
27140 zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_0_6_aligned()
27164 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_0_0_7_aligned() argument
27182 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_0_7_aligned()
27188 zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_0_7_aligned()
27212 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_0_1_0_aligned() argument
27230 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_1_0_aligned()
27236 zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_1_0_aligned()
27240 zip_1x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_1_0_aligned()
27264 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_0_1_1_aligned() argument
27282 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_1_1_aligned()
27288 zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_1_1_aligned()
27292 zip_1x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_1_1_aligned()
27316 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_0_1_2_aligned() argument
27334 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_1_2_aligned()
27340 zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_1_2_aligned()
27344 zip_1x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_1_2_aligned()
27368 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_0_1_3_aligned() argument
27386 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_1_3_aligned()
27392 zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_1_3_aligned()
27396 zip_1x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_1_3_aligned()
27420 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_0_1_4_aligned() argument
27438 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_1_4_aligned()
27444 zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_1_4_aligned()
27448 zip_1x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_1_4_aligned()
27472 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_0_1_5_aligned() argument
27490 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_1_5_aligned()
27496 zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_1_5_aligned()
27500 zip_1x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_1_5_aligned()
27524 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_0_1_6_aligned() argument
27542 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_1_6_aligned()
27548 zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_1_6_aligned()
27552 zip_1x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_1_6_aligned()
27576 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_0_1_7_aligned() argument
27594 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_1_7_aligned()
27600 zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_1_7_aligned()
27604 zip_1x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_1_7_aligned()
27628 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_0_2_0_aligned() argument
27646 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_2_0_aligned()
27652 zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_2_0_aligned()
27656 zip_2x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_2_0_aligned()
27680 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_0_2_1_aligned() argument
27698 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_2_1_aligned()
27704 zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_2_1_aligned()
27708 zip_2x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_2_1_aligned()
27732 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_0_2_2_aligned() argument
27750 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_2_2_aligned()
27756 zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_2_2_aligned()
27760 zip_2x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_2_2_aligned()
27784 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_0_2_3_aligned() argument
27802 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_2_3_aligned()
27808 zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_2_3_aligned()
27812 zip_2x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_2_3_aligned()
27836 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_0_2_4_aligned() argument
27854 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_2_4_aligned()
27860 zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_2_4_aligned()
27864 zip_2x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_2_4_aligned()
27888 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_0_2_5_aligned() argument
27906 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_2_5_aligned()
27912 zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_2_5_aligned()
27916 zip_2x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_2_5_aligned()
27940 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_0_2_6_aligned() argument
27958 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_2_6_aligned()
27964 zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_2_6_aligned()
27968 zip_2x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_2_6_aligned()
27992 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_0_2_7_aligned() argument
28010 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_2_7_aligned()
28016 zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_2_7_aligned()
28020 zip_2x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_2_7_aligned()
28044 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_1_0_0_aligned() argument
28064 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_0_0_aligned()
28070 zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_0_0_aligned()
28105 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_1_0_1_aligned() argument
28125 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_0_1_aligned()
28131 zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_0_1_aligned()
28166 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_1_0_2_aligned() argument
28186 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_0_2_aligned()
28192 zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_0_2_aligned()
28227 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_1_0_3_aligned() argument
28247 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_0_3_aligned()
28253 zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_0_3_aligned()
28288 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_1_0_4_aligned() argument
28308 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_0_4_aligned()
28314 zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_0_4_aligned()
28349 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_1_0_5_aligned() argument
28369 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_0_5_aligned()
28375 zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_0_5_aligned()
28410 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_1_0_6_aligned() argument
28430 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_0_6_aligned()
28436 zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_0_6_aligned()
28471 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_1_0_7_aligned() argument
28491 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_0_7_aligned()
28497 zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_0_7_aligned()
28532 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_1_1_0_aligned() argument
28552 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_1_0_aligned()
28558 zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_1_0_aligned()
28562 zip_1x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_1_0_aligned()
28600 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_1_1_1_aligned() argument
28620 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_1_1_aligned()
28626 zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_1_1_aligned()
28630 zip_1x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_1_1_aligned()
28668 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_1_1_2_aligned() argument
28688 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_1_2_aligned()
28694 zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_1_2_aligned()
28698 zip_1x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_1_2_aligned()
28736 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_1_1_3_aligned() argument
28756 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_1_3_aligned()
28762 zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_1_3_aligned()
28766 zip_1x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_1_3_aligned()
28804 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_1_1_4_aligned() argument
28824 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_1_4_aligned()
28830 zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_1_4_aligned()
28834 zip_1x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_1_4_aligned()
28872 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_1_1_5_aligned() argument
28892 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_1_5_aligned()
28898 zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_1_5_aligned()
28902 zip_1x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_1_5_aligned()
28940 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_1_1_6_aligned() argument
28960 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_1_6_aligned()
28966 zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_1_6_aligned()
28970 zip_1x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_1_6_aligned()
29008 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_1_1_7_aligned() argument
29028 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_1_7_aligned()
29034 zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_1_7_aligned()
29038 zip_1x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_1_7_aligned()
29076 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_1_2_0_aligned() argument
29096 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_2_0_aligned()
29102 zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_2_0_aligned()
29106 zip_2x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_2_0_aligned()
29144 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_1_2_1_aligned() argument
29164 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_2_1_aligned()
29170 zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_2_1_aligned()
29174 zip_2x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_2_1_aligned()
29212 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_1_2_2_aligned() argument
29232 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_2_2_aligned()
29238 zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_2_2_aligned()
29242 zip_2x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_2_2_aligned()
29280 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_1_2_3_aligned() argument
29300 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_2_3_aligned()
29306 zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_2_3_aligned()
29310 zip_2x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_2_3_aligned()
29348 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_1_2_4_aligned() argument
29368 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_2_4_aligned()
29374 zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_2_4_aligned()
29378 zip_2x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_2_4_aligned()
29416 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_1_2_5_aligned() argument
29436 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_2_5_aligned()
29442 zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_2_5_aligned()
29446 zip_2x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_2_5_aligned()
29484 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_1_2_6_aligned() argument
29504 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_2_6_aligned()
29510 zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_2_6_aligned()
29514 zip_2x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_2_6_aligned()
29552 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_1_2_7_aligned() argument
29572 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_2_7_aligned()
29578 zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_2_7_aligned()
29582 zip_2x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_2_7_aligned()
29620 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_2_0_0_aligned() argument
29640 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_0_0_aligned()
29646 zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_0_0_aligned()
29681 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_2_0_1_aligned() argument
29701 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_0_1_aligned()
29707 zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_0_1_aligned()
29742 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_2_0_2_aligned() argument
29762 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_0_2_aligned()
29768 zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_0_2_aligned()
29803 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_2_0_3_aligned() argument
29823 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_0_3_aligned()
29829 zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_0_3_aligned()
29864 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_2_0_4_aligned() argument
29884 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_0_4_aligned()
29890 zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_0_4_aligned()
29925 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_2_0_5_aligned() argument
29945 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_0_5_aligned()
29951 zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_0_5_aligned()
29986 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_2_0_6_aligned() argument
30006 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_0_6_aligned()
30012 zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_0_6_aligned()
30047 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_2_0_7_aligned() argument
30067 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_0_7_aligned()
30073 zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_0_7_aligned()
30108 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_2_1_0_aligned() argument
30128 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_1_0_aligned()
30134 zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_1_0_aligned()
30138 zip_1x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_1_0_aligned()
30176 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_2_1_1_aligned() argument
30196 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_1_1_aligned()
30202 zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_1_1_aligned()
30206 zip_1x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_1_1_aligned()
30244 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_2_1_2_aligned() argument
30264 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_1_2_aligned()
30270 zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_1_2_aligned()
30274 zip_1x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_1_2_aligned()
30312 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_2_1_3_aligned() argument
30332 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_1_3_aligned()
30338 zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_1_3_aligned()
30342 zip_1x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_1_3_aligned()
30380 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_2_1_4_aligned() argument
30400 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_1_4_aligned()
30406 zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_1_4_aligned()
30410 zip_1x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_1_4_aligned()
30448 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_2_1_5_aligned() argument
30468 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_1_5_aligned()
30474 zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_1_5_aligned()
30478 zip_1x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_1_5_aligned()
30516 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_2_1_6_aligned() argument
30536 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_1_6_aligned()
30542 zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_1_6_aligned()
30546 zip_1x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_1_6_aligned()
30584 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_2_1_7_aligned() argument
30604 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_1_7_aligned()
30610 zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_1_7_aligned()
30614 zip_1x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_1_7_aligned()
30652 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_2_2_0_aligned() argument
30672 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_2_0_aligned()
30678 zip_3x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_2_0_aligned()
30682 zip_2x8_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_2_0_aligned()
30720 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_2_2_1_aligned() argument
30740 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_2_1_aligned()
30746 zip_3x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_2_1_aligned()
30750 zip_2x8_1_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_2_1_aligned()
30788 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_2_2_2_aligned() argument
30808 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_2_2_aligned()
30814 zip_3x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_2_2_aligned()
30818 zip_2x8_2_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_2_2_aligned()
30856 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_2_2_3_aligned() argument
30876 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_2_3_aligned()
30882 zip_3x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_2_3_aligned()
30886 zip_2x8_3_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_2_3_aligned()
30924 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_2_2_4_aligned() argument
30944 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_2_4_aligned()
30950 zip_3x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_2_4_aligned()
30954 zip_2x8_4_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_2_4_aligned()
30992 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_2_2_5_aligned() argument
31012 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_2_5_aligned()
31018 zip_3x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_2_5_aligned()
31022 zip_2x8_5_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_2_5_aligned()
31060 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_2_2_6_aligned() argument
31080 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_2_6_aligned()
31086 zip_3x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_2_6_aligned()
31090 zip_2x8_6_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_2_6_aligned()
31128 std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f_2_2_7_aligned() argument
31148 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_2_7_aligned()
31154 zip_3x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_2_7_aligned()
31158 zip_2x8_7_aligned(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_2_7_aligned()
31195 std::int32_t k, std::int32_t lhs_offset, in gemm_f_0_0_0() argument
31213 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_0_0()
31219 zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_0_0()
31242 std::int32_t k, std::int32_t lhs_offset, in gemm_f_0_0_1() argument
31260 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_0_1()
31266 zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_0_1()
31289 std::int32_t k, std::int32_t lhs_offset, in gemm_f_0_0_2() argument
31307 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_0_2()
31313 zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_0_2()
31336 std::int32_t k, std::int32_t lhs_offset, in gemm_f_0_0_3() argument
31354 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_0_3()
31360 zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_0_3()
31383 std::int32_t k, std::int32_t lhs_offset, in gemm_f_0_0_4() argument
31401 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_0_4()
31407 zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_0_4()
31430 std::int32_t k, std::int32_t lhs_offset, in gemm_f_0_0_5() argument
31448 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_0_5()
31454 zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_0_5()
31477 std::int32_t k, std::int32_t lhs_offset, in gemm_f_0_0_6() argument
31495 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_0_6()
31501 zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_0_6()
31524 std::int32_t k, std::int32_t lhs_offset, in gemm_f_0_0_7() argument
31542 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_0_7()
31548 zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_0_7()
31571 std::int32_t k, std::int32_t lhs_offset, in gemm_f_0_1_0() argument
31589 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_1_0()
31595 zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_1_0()
31599 zip_1x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_1_0()
31622 std::int32_t k, std::int32_t lhs_offset, in gemm_f_0_1_1() argument
31640 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_1_1()
31646 zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_1_1()
31650 zip_1x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_1_1()
31673 std::int32_t k, std::int32_t lhs_offset, in gemm_f_0_1_2() argument
31691 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_1_2()
31697 zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_1_2()
31701 zip_1x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_1_2()
31724 std::int32_t k, std::int32_t lhs_offset, in gemm_f_0_1_3() argument
31742 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_1_3()
31748 zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_1_3()
31752 zip_1x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_1_3()
31775 std::int32_t k, std::int32_t lhs_offset, in gemm_f_0_1_4() argument
31793 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_1_4()
31799 zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_1_4()
31803 zip_1x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_1_4()
31826 std::int32_t k, std::int32_t lhs_offset, in gemm_f_0_1_5() argument
31844 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_1_5()
31850 zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_1_5()
31854 zip_1x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_1_5()
31877 std::int32_t k, std::int32_t lhs_offset, in gemm_f_0_1_6() argument
31895 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_1_6()
31901 zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_1_6()
31905 zip_1x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_1_6()
31928 std::int32_t k, std::int32_t lhs_offset, in gemm_f_0_1_7() argument
31946 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_1_7()
31952 zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_1_7()
31956 zip_1x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_1_7()
31979 std::int32_t k, std::int32_t lhs_offset, in gemm_f_0_2_0() argument
31997 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_2_0()
32003 zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_2_0()
32007 zip_2x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_2_0()
32030 std::int32_t k, std::int32_t lhs_offset, in gemm_f_0_2_1() argument
32048 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_2_1()
32054 zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_2_1()
32058 zip_2x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_2_1()
32081 std::int32_t k, std::int32_t lhs_offset, in gemm_f_0_2_2() argument
32099 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_2_2()
32105 zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_2_2()
32109 zip_2x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_2_2()
32132 std::int32_t k, std::int32_t lhs_offset, in gemm_f_0_2_3() argument
32150 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_2_3()
32156 zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_2_3()
32160 zip_2x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_2_3()
32183 std::int32_t k, std::int32_t lhs_offset, in gemm_f_0_2_4() argument
32201 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_2_4()
32207 zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_2_4()
32211 zip_2x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_2_4()
32234 std::int32_t k, std::int32_t lhs_offset, in gemm_f_0_2_5() argument
32252 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_2_5()
32258 zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_2_5()
32262 zip_2x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_2_5()
32285 std::int32_t k, std::int32_t lhs_offset, in gemm_f_0_2_6() argument
32303 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_2_6()
32309 zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_2_6()
32313 zip_2x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_2_6()
32336 std::int32_t k, std::int32_t lhs_offset, in gemm_f_0_2_7() argument
32354 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_0_2_7()
32360 zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_2_7()
32364 zip_2x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_0_2_7()
32387 std::int32_t k, std::int32_t lhs_offset, in gemm_f_1_0_0() argument
32407 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_0_0()
32413 zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_0_0()
32447 std::int32_t k, std::int32_t lhs_offset, in gemm_f_1_0_1() argument
32467 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_0_1()
32473 zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_0_1()
32507 std::int32_t k, std::int32_t lhs_offset, in gemm_f_1_0_2() argument
32527 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_0_2()
32533 zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_0_2()
32567 std::int32_t k, std::int32_t lhs_offset, in gemm_f_1_0_3() argument
32587 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_0_3()
32593 zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_0_3()
32627 std::int32_t k, std::int32_t lhs_offset, in gemm_f_1_0_4() argument
32647 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_0_4()
32653 zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_0_4()
32687 std::int32_t k, std::int32_t lhs_offset, in gemm_f_1_0_5() argument
32707 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_0_5()
32713 zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_0_5()
32747 std::int32_t k, std::int32_t lhs_offset, in gemm_f_1_0_6() argument
32767 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_0_6()
32773 zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_0_6()
32807 std::int32_t k, std::int32_t lhs_offset, in gemm_f_1_0_7() argument
32827 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_0_7()
32833 zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_0_7()
32867 std::int32_t k, std::int32_t lhs_offset, in gemm_f_1_1_0() argument
32887 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_1_0()
32893 zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_1_0()
32897 zip_1x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_1_0()
32934 std::int32_t k, std::int32_t lhs_offset, in gemm_f_1_1_1() argument
32954 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_1_1()
32960 zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_1_1()
32964 zip_1x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_1_1()
33001 std::int32_t k, std::int32_t lhs_offset, in gemm_f_1_1_2() argument
33021 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_1_2()
33027 zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_1_2()
33031 zip_1x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_1_2()
33068 std::int32_t k, std::int32_t lhs_offset, in gemm_f_1_1_3() argument
33088 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_1_3()
33094 zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_1_3()
33098 zip_1x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_1_3()
33135 std::int32_t k, std::int32_t lhs_offset, in gemm_f_1_1_4() argument
33155 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_1_4()
33161 zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_1_4()
33165 zip_1x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_1_4()
33202 std::int32_t k, std::int32_t lhs_offset, in gemm_f_1_1_5() argument
33222 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_1_5()
33228 zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_1_5()
33232 zip_1x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_1_5()
33269 std::int32_t k, std::int32_t lhs_offset, in gemm_f_1_1_6() argument
33289 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_1_6()
33295 zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_1_6()
33299 zip_1x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_1_6()
33336 std::int32_t k, std::int32_t lhs_offset, in gemm_f_1_1_7() argument
33356 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_1_7()
33362 zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_1_7()
33366 zip_1x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_1_7()
33403 std::int32_t k, std::int32_t lhs_offset, in gemm_f_1_2_0() argument
33423 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_2_0()
33429 zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_2_0()
33433 zip_2x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_2_0()
33470 std::int32_t k, std::int32_t lhs_offset, in gemm_f_1_2_1() argument
33490 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_2_1()
33496 zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_2_1()
33500 zip_2x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_2_1()
33537 std::int32_t k, std::int32_t lhs_offset, in gemm_f_1_2_2() argument
33557 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_2_2()
33563 zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_2_2()
33567 zip_2x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_2_2()
33604 std::int32_t k, std::int32_t lhs_offset, in gemm_f_1_2_3() argument
33624 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_2_3()
33630 zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_2_3()
33634 zip_2x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_2_3()
33671 std::int32_t k, std::int32_t lhs_offset, in gemm_f_1_2_4() argument
33691 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_2_4()
33697 zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_2_4()
33701 zip_2x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_2_4()
33738 std::int32_t k, std::int32_t lhs_offset, in gemm_f_1_2_5() argument
33758 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_2_5()
33764 zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_2_5()
33768 zip_2x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_2_5()
33805 std::int32_t k, std::int32_t lhs_offset, in gemm_f_1_2_6() argument
33825 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_2_6()
33831 zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_2_6()
33835 zip_2x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_2_6()
33872 std::int32_t k, std::int32_t lhs_offset, in gemm_f_1_2_7() argument
33892 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_1_2_7()
33898 zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_2_7()
33902 zip_2x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_1_2_7()
33939 std::int32_t k, std::int32_t lhs_offset, in gemm_f_2_0_0() argument
33959 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_0_0()
33965 zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_0_0()
33999 std::int32_t k, std::int32_t lhs_offset, in gemm_f_2_0_1() argument
34019 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_0_1()
34025 zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_0_1()
34059 std::int32_t k, std::int32_t lhs_offset, in gemm_f_2_0_2() argument
34079 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_0_2()
34085 zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_0_2()
34119 std::int32_t k, std::int32_t lhs_offset, in gemm_f_2_0_3() argument
34139 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_0_3()
34145 zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_0_3()
34179 std::int32_t k, std::int32_t lhs_offset, in gemm_f_2_0_4() argument
34199 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_0_4()
34205 zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_0_4()
34239 std::int32_t k, std::int32_t lhs_offset, in gemm_f_2_0_5() argument
34259 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_0_5()
34265 zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_0_5()
34299 std::int32_t k, std::int32_t lhs_offset, in gemm_f_2_0_6() argument
34319 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_0_6()
34325 zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_0_6()
34359 std::int32_t k, std::int32_t lhs_offset, in gemm_f_2_0_7() argument
34379 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_0_7()
34385 zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_0_7()
34419 std::int32_t k, std::int32_t lhs_offset, in gemm_f_2_1_0() argument
34439 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_1_0()
34445 zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_1_0()
34449 zip_1x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_1_0()
34486 std::int32_t k, std::int32_t lhs_offset, in gemm_f_2_1_1() argument
34506 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_1_1()
34512 zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_1_1()
34516 zip_1x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_1_1()
34553 std::int32_t k, std::int32_t lhs_offset, in gemm_f_2_1_2() argument
34573 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_1_2()
34579 zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_1_2()
34583 zip_1x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_1_2()
34620 std::int32_t k, std::int32_t lhs_offset, in gemm_f_2_1_3() argument
34640 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_1_3()
34646 zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_1_3()
34650 zip_1x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_1_3()
34687 std::int32_t k, std::int32_t lhs_offset, in gemm_f_2_1_4() argument
34707 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_1_4()
34713 zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_1_4()
34717 zip_1x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_1_4()
34754 std::int32_t k, std::int32_t lhs_offset, in gemm_f_2_1_5() argument
34774 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_1_5()
34780 zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_1_5()
34784 zip_1x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_1_5()
34821 std::int32_t k, std::int32_t lhs_offset, in gemm_f_2_1_6() argument
34841 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_1_6()
34847 zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_1_6()
34851 zip_1x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_1_6()
34888 std::int32_t k, std::int32_t lhs_offset, in gemm_f_2_1_7() argument
34908 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_1_7()
34914 zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_1_7()
34918 zip_1x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_1_7()
34955 std::int32_t k, std::int32_t lhs_offset, in gemm_f_2_2_0() argument
34975 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_2_0()
34981 zip_3x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_2_0()
34985 zip_2x8(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_2_0()
35022 std::int32_t k, std::int32_t lhs_offset, in gemm_f_2_2_1() argument
35042 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_2_1()
35048 zip_3x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_2_1()
35052 zip_2x8_1(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_2_1()
35089 std::int32_t k, std::int32_t lhs_offset, in gemm_f_2_2_2() argument
35109 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_2_2()
35115 zip_3x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_2_2()
35119 zip_2x8_2(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_2_2()
35156 std::int32_t k, std::int32_t lhs_offset, in gemm_f_2_2_3() argument
35176 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_2_3()
35182 zip_3x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_2_3()
35186 zip_2x8_3(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_2_3()
35223 std::int32_t k, std::int32_t lhs_offset, in gemm_f_2_2_4() argument
35243 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_2_4()
35249 zip_3x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_2_4()
35253 zip_2x8_4(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_2_4()
35290 std::int32_t k, std::int32_t lhs_offset, in gemm_f_2_2_5() argument
35310 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_2_5()
35316 zip_3x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_2_5()
35320 zip_2x8_5(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_2_5()
35357 std::int32_t k, std::int32_t lhs_offset, in gemm_f_2_2_6() argument
35377 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_2_6()
35383 zip_3x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_2_6()
35387 zip_2x8_6(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_2_6()
35424 std::int32_t k, std::int32_t lhs_offset, in gemm_f_2_2_7() argument
35444 const std::int32_t const_offset = lhs_offset * rhs_offset * k; in gemm_f_2_2_7()
35450 zip_3x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_2_7()
35454 zip_2x8_7(rhs_chunk, k, k, zipped_rhs_chunk, lhs_offset, 0); in gemm_f_2_2_7()
35493 std::int32_t k, std::int32_t lhs_offset, in gemm_q8_strided() argument
35513 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35519 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35525 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35531 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35537 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35543 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35549 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35555 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35565 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35571 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35577 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35583 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35589 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35595 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35601 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35607 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35617 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35623 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35629 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35635 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35641 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35647 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35653 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35659 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35673 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35679 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35685 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35691 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35697 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35703 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35709 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35715 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35725 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35731 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35737 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35743 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35749 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35755 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35761 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35767 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35777 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35783 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35789 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35795 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35801 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35807 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35813 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35819 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35833 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35839 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35845 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35851 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35857 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35863 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35869 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35875 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35885 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35891 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35897 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35903 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35909 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35915 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35921 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35927 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35937 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35943 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35949 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35955 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35961 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35967 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35973 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35979 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8_strided()
35995 internal::gemm_q8_0_0_0(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36001 internal::gemm_q8_0_0_1(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36007 internal::gemm_q8_0_0_2(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36013 internal::gemm_q8_0_0_3(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36019 internal::gemm_q8_0_0_4(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36025 internal::gemm_q8_0_0_5(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36031 internal::gemm_q8_0_0_6(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36037 internal::gemm_q8_0_0_7(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36047 internal::gemm_q8_0_1_0(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36053 internal::gemm_q8_0_1_1(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36059 internal::gemm_q8_0_1_2(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36065 internal::gemm_q8_0_1_3(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36071 internal::gemm_q8_0_1_4(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36077 internal::gemm_q8_0_1_5(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36083 internal::gemm_q8_0_1_6(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36089 internal::gemm_q8_0_1_7(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36099 internal::gemm_q8_0_2_0(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36105 internal::gemm_q8_0_2_1(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36111 internal::gemm_q8_0_2_2(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36117 internal::gemm_q8_0_2_3(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36123 internal::gemm_q8_0_2_4(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36129 internal::gemm_q8_0_2_5(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36135 internal::gemm_q8_0_2_6(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36141 internal::gemm_q8_0_2_7(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36155 internal::gemm_q8_1_0_0(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36161 internal::gemm_q8_1_0_1(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36167 internal::gemm_q8_1_0_2(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36173 internal::gemm_q8_1_0_3(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36179 internal::gemm_q8_1_0_4(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36185 internal::gemm_q8_1_0_5(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36191 internal::gemm_q8_1_0_6(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36197 internal::gemm_q8_1_0_7(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36207 internal::gemm_q8_1_1_0(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36213 internal::gemm_q8_1_1_1(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36219 internal::gemm_q8_1_1_2(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36225 internal::gemm_q8_1_1_3(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36231 internal::gemm_q8_1_1_4(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36237 internal::gemm_q8_1_1_5(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36243 internal::gemm_q8_1_1_6(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36249 internal::gemm_q8_1_1_7(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36259 internal::gemm_q8_1_2_0(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36265 internal::gemm_q8_1_2_1(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36271 internal::gemm_q8_1_2_2(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36277 internal::gemm_q8_1_2_3(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36283 internal::gemm_q8_1_2_4(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36289 internal::gemm_q8_1_2_5(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36295 internal::gemm_q8_1_2_6(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36301 internal::gemm_q8_1_2_7(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36315 internal::gemm_q8_2_0_0(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36321 internal::gemm_q8_2_0_1(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36327 internal::gemm_q8_2_0_2(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36333 internal::gemm_q8_2_0_3(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36339 internal::gemm_q8_2_0_4(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36345 internal::gemm_q8_2_0_5(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36351 internal::gemm_q8_2_0_6(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36357 internal::gemm_q8_2_0_7(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36367 internal::gemm_q8_2_1_0(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36373 internal::gemm_q8_2_1_1(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36379 internal::gemm_q8_2_1_2(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36385 internal::gemm_q8_2_1_3(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36391 internal::gemm_q8_2_1_4(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36397 internal::gemm_q8_2_1_5(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36403 internal::gemm_q8_2_1_6(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36409 internal::gemm_q8_2_1_7(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36419 internal::gemm_q8_2_2_0(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36425 internal::gemm_q8_2_2_1(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36431 internal::gemm_q8_2_2_2(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36437 internal::gemm_q8_2_2_3(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36443 internal::gemm_q8_2_2_4(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36449 internal::gemm_q8_2_2_5(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36455 internal::gemm_q8_2_2_6(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36461 internal::gemm_q8_2_2_7(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_q8_strided()
36476 std::int32_t k, std::int32_t lhs_offset, in gemm_i32_strided() argument
36491 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36496 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36501 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36506 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36511 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36516 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36521 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36526 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36535 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36540 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36545 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36550 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36555 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36560 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36565 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36570 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36579 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36584 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36589 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36594 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36599 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36604 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36609 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36614 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36627 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36632 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36637 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36642 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36647 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36652 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36657 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36662 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36671 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36676 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36681 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36686 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36691 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36696 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36701 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36706 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36715 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36720 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36725 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36730 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36735 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36740 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36745 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36750 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36763 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36768 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36773 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36778 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36783 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36788 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36793 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36798 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36807 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36812 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36817 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36822 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36827 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36832 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36837 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36842 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36851 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36856 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36861 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36866 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36871 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36876 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36881 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36886 lhs_offset, rhs_offset, result, in gemm_i32_strided()
36901 internal::gemm_i32_0_0_0(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
36905 internal::gemm_i32_0_0_1(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
36909 internal::gemm_i32_0_0_2(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
36913 internal::gemm_i32_0_0_3(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
36917 internal::gemm_i32_0_0_4(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
36921 internal::gemm_i32_0_0_5(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
36925 internal::gemm_i32_0_0_6(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
36929 internal::gemm_i32_0_0_7(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
36937 internal::gemm_i32_0_1_0(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
36941 internal::gemm_i32_0_1_1(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
36945 internal::gemm_i32_0_1_2(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
36949 internal::gemm_i32_0_1_3(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
36953 internal::gemm_i32_0_1_4(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
36957 internal::gemm_i32_0_1_5(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
36961 internal::gemm_i32_0_1_6(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
36965 internal::gemm_i32_0_1_7(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
36973 internal::gemm_i32_0_2_0(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
36977 internal::gemm_i32_0_2_1(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
36981 internal::gemm_i32_0_2_2(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
36985 internal::gemm_i32_0_2_3(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
36989 internal::gemm_i32_0_2_4(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
36993 internal::gemm_i32_0_2_5(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
36997 internal::gemm_i32_0_2_6(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37001 internal::gemm_i32_0_2_7(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37013 internal::gemm_i32_1_0_0(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37017 internal::gemm_i32_1_0_1(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37021 internal::gemm_i32_1_0_2(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37025 internal::gemm_i32_1_0_3(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37029 internal::gemm_i32_1_0_4(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37033 internal::gemm_i32_1_0_5(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37037 internal::gemm_i32_1_0_6(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37041 internal::gemm_i32_1_0_7(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37049 internal::gemm_i32_1_1_0(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37053 internal::gemm_i32_1_1_1(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37057 internal::gemm_i32_1_1_2(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37061 internal::gemm_i32_1_1_3(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37065 internal::gemm_i32_1_1_4(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37069 internal::gemm_i32_1_1_5(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37073 internal::gemm_i32_1_1_6(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37077 internal::gemm_i32_1_1_7(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37085 internal::gemm_i32_1_2_0(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37089 internal::gemm_i32_1_2_1(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37093 internal::gemm_i32_1_2_2(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37097 internal::gemm_i32_1_2_3(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37101 internal::gemm_i32_1_2_4(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37105 internal::gemm_i32_1_2_5(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37109 internal::gemm_i32_1_2_6(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37113 internal::gemm_i32_1_2_7(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37125 internal::gemm_i32_2_0_0(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37129 internal::gemm_i32_2_0_1(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37133 internal::gemm_i32_2_0_2(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37137 internal::gemm_i32_2_0_3(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37141 internal::gemm_i32_2_0_4(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37145 internal::gemm_i32_2_0_5(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37149 internal::gemm_i32_2_0_6(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37153 internal::gemm_i32_2_0_7(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37161 internal::gemm_i32_2_1_0(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37165 internal::gemm_i32_2_1_1(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37169 internal::gemm_i32_2_1_2(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37173 internal::gemm_i32_2_1_3(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37177 internal::gemm_i32_2_1_4(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37181 internal::gemm_i32_2_1_5(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37185 internal::gemm_i32_2_1_6(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37189 internal::gemm_i32_2_1_7(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37197 internal::gemm_i32_2_2_0(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37201 internal::gemm_i32_2_2_1(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37205 internal::gemm_i32_2_2_2(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37209 internal::gemm_i32_2_2_3(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37213 internal::gemm_i32_2_2_4(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37217 internal::gemm_i32_2_2_5(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37221 internal::gemm_i32_2_2_6(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37225 internal::gemm_i32_2_2_7(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_i32_strided()
37238 std::int32_t k, std::int32_t lhs_offset, in gemm_f_strided() argument
37253 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37258 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37263 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37268 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37273 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37278 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37283 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37288 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37297 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37302 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37307 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37312 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37317 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37322 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37327 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37332 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37341 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37346 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37351 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37356 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37361 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37366 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37371 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37376 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37389 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37394 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37399 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37404 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37409 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37414 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37419 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37424 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37433 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37438 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37443 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37448 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37453 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37458 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37463 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37468 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37477 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37482 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37487 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37492 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37497 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37502 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37507 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37512 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37525 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37530 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37535 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37540 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37545 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37550 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37555 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37560 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37569 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37574 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37579 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37584 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37589 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37594 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37599 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37604 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37613 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37618 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37623 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37628 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37633 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37638 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37643 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37648 scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f_strided()
37663 internal::gemm_f_0_0_0(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37668 internal::gemm_f_0_0_1(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37673 internal::gemm_f_0_0_2(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37678 internal::gemm_f_0_0_3(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37683 internal::gemm_f_0_0_4(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37688 internal::gemm_f_0_0_5(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37693 internal::gemm_f_0_0_6(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37698 internal::gemm_f_0_0_7(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37707 internal::gemm_f_0_1_0(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37712 internal::gemm_f_0_1_1(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37717 internal::gemm_f_0_1_2(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37722 internal::gemm_f_0_1_3(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37727 internal::gemm_f_0_1_4(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37732 internal::gemm_f_0_1_5(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37737 internal::gemm_f_0_1_6(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37742 internal::gemm_f_0_1_7(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37751 internal::gemm_f_0_2_0(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37756 internal::gemm_f_0_2_1(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37761 internal::gemm_f_0_2_2(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37766 internal::gemm_f_0_2_3(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37771 internal::gemm_f_0_2_4(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37776 internal::gemm_f_0_2_5(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37781 internal::gemm_f_0_2_6(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37786 internal::gemm_f_0_2_7(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37799 internal::gemm_f_1_0_0(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37804 internal::gemm_f_1_0_1(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37809 internal::gemm_f_1_0_2(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37814 internal::gemm_f_1_0_3(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37819 internal::gemm_f_1_0_4(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37824 internal::gemm_f_1_0_5(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37829 internal::gemm_f_1_0_6(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37834 internal::gemm_f_1_0_7(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37843 internal::gemm_f_1_1_0(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37848 internal::gemm_f_1_1_1(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37853 internal::gemm_f_1_1_2(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37858 internal::gemm_f_1_1_3(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37863 internal::gemm_f_1_1_4(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37868 internal::gemm_f_1_1_5(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37873 internal::gemm_f_1_1_6(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37878 internal::gemm_f_1_1_7(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37887 internal::gemm_f_1_2_0(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37892 internal::gemm_f_1_2_1(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37897 internal::gemm_f_1_2_2(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37902 internal::gemm_f_1_2_3(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37907 internal::gemm_f_1_2_4(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37912 internal::gemm_f_1_2_5(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37917 internal::gemm_f_1_2_6(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37922 internal::gemm_f_1_2_7(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37935 internal::gemm_f_2_0_0(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37940 internal::gemm_f_2_0_1(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37945 internal::gemm_f_2_0_2(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37950 internal::gemm_f_2_0_3(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37955 internal::gemm_f_2_0_4(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37960 internal::gemm_f_2_0_5(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37965 internal::gemm_f_2_0_6(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37970 internal::gemm_f_2_0_7(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37979 internal::gemm_f_2_1_0(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37984 internal::gemm_f_2_1_1(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37989 internal::gemm_f_2_1_2(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37994 internal::gemm_f_2_1_3(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
37999 internal::gemm_f_2_1_4(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
38004 internal::gemm_f_2_1_5(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
38009 internal::gemm_f_2_1_6(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
38014 internal::gemm_f_2_1_7(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
38023 internal::gemm_f_2_2_0(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
38028 internal::gemm_f_2_2_1(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
38033 internal::gemm_f_2_2_2(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
38038 internal::gemm_f_2_2_3(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
38043 internal::gemm_f_2_2_4(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
38048 internal::gemm_f_2_2_5(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
38053 internal::gemm_f_2_2_6(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
38058 internal::gemm_f_2_2_7(scratch, lhs, rhs, m, n, k, lhs_offset, in gemm_f_strided()
38072 std::int32_t k, std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_q8() argument
38075 gemm_q8_strided(scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_q8()
38081 std::int32_t k, std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_i32() argument
38083 gemm_i32_strided(scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, result, in gemm_i32()
38089 std::int32_t k, std::int32_t lhs_offset, std::int32_t rhs_offset, in gemm_f() argument
38091 gemm_f_strided(scratch, lhs, rhs, m, n, k, lhs_offset, rhs_offset, in gemm_f()