@@ -579,6 +579,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
579579 ${KLEIDIAI_SRC} /kai/ukernels/
580580 ${KLEIDIAI_SRC} /kai/ukernels/matmul/
581581 ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/
582+ ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/
582583 ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/
583584 ${KLEIDIAI_SRC} /kai/ukernels/matmul/pack/)
584585
@@ -597,23 +598,34 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
597598 ${KLEIDIAI_SRC} /kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p4x8sb_f32_neon.c
598599 ${KLEIDIAI_SRC} /kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.c
599600 ${KLEIDIAI_SRC} /kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32_neon.c
600- ${KLEIDIAI_SRC} /kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.c)
601+ ${KLEIDIAI_SRC} /kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.c
602+ ${KLEIDIAI_SRC} /kai/ukernels/matmul/pack/kai_lhs_quant_pack_qai8dxp_f32.c
603+ ${KLEIDIAI_SRC} /kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi8cxp_qsi8cx_neon.c)
601604
602605 if (NOT DOTPROD_ENABLED MATCHES -1)
603606 list (APPEND GGML_KLEIDIAI_SOURCES
604607 ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c
605608 ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.c
606- ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c)
609+ ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c
610+ ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.c
611+ ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.c
612+ ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.c)
607613 endif ()
608614
609615 if (NOT I8MM_ENABLED MATCHES -1)
610- list (APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.c)
616+ list (APPEND GGML_KLEIDIAI_SOURCES
617+ ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.c
618+ ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.c)
611619 endif ()
612620
613621 if (NOT SME_ENABLED MATCHES -1)
614622 list (APPEND GGML_KLEIDIAI_SOURCES
615623 ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.c
616624 ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c
625+ ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa.c
626+ ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa_asm.S
627+ ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot.c
628+ ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot_asm.S
617629 ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.c
618630 ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa_asm.S
619631 ${KLEIDIAI_SRC} /kai/ukernels/matmul/pack/kai_lhs_pack_bf16p2vlx2_f32_sme.c
0 commit comments