@@ -590,6 +590,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
590590 ${KLEIDIAI_SRC} /kai/ukernels/
591591 ${KLEIDIAI_SRC} /kai/ukernels/matmul/
592592 ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/
593+ ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/
593594 ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/
594595 ${KLEIDIAI_SRC} /kai/ukernels/matmul/pack/)
595596
@@ -608,23 +609,34 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
608609 ${KLEIDIAI_SRC} /kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p4x8sb_f32_neon.c
609610 ${KLEIDIAI_SRC} /kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.c
610611 ${KLEIDIAI_SRC} /kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32_neon.c
611- ${KLEIDIAI_SRC} /kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.c)
612+ ${KLEIDIAI_SRC} /kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.c
613+ ${KLEIDIAI_SRC} /kai/ukernels/matmul/pack/kai_lhs_quant_pack_qai8dxp_f32.c
614+ ${KLEIDIAI_SRC} /kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi8cxp_qsi8cx_neon.c)
612615
613616 if (NOT DOTPROD_ENABLED MATCHES -1)
614617 list (APPEND GGML_KLEIDIAI_SOURCES
615618 ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c
616619 ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.c
617- ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c)
620+ ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c
621+ ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.c
622+ ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.c
623+ ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.c)
618624 endif ()
619625
620626 if (NOT I8MM_ENABLED MATCHES -1)
621- 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)
627+ list (APPEND GGML_KLEIDIAI_SOURCES
628+ ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.c
629+ ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.c)
622630 endif ()
623631
624632 if (NOT SME_ENABLED MATCHES -1)
625633 list (APPEND GGML_KLEIDIAI_SOURCES
626634 ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.c
627635 ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c
636+ ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa.c
637+ ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa_asm.S
638+ ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot.c
639+ ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot_asm.S
628640 ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.c
629641 ${KLEIDIAI_SRC} /kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa_asm.S
630642 ${KLEIDIAI_SRC} /kai/ukernels/matmul/pack/kai_lhs_pack_bf16p2vlx2_f32_sme.c
0 commit comments