android13/external/XNNPACK/test/qc8-gemm-minmax-fp32.yaml

# Copyright 2021 Google LLC
#
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.

- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x8c4__aarch32_neondot_cortex_a55
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
  assembly: true
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x8c4__aarch32_neondot_ld64
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
  assembly: true
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x8__aarch32_neonv8_mlal_lane_prfm_cortex_a53
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
  assembly: true
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x8__aarch32_neonv8_mlal_lane_cortex_a53
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
  assembly: true
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53
  init: xnn_init_qs8_minmax_neon_params
  k-block: 8
  assembly: true
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53
  init: xnn_init_qs8_minmax_neon_params
  k-block: 8
  assembly: true
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7
  init: xnn_init_qs8_minmax_neon_params
  k-block: 8
  assembly: true
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7
  init: xnn_init_qs8_minmax_neon_params
  k-block: 8
  assembly: true
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64
  init: xnn_init_qs8_minmax_neon_params
  k-block: 8
  assembly: true
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x8__aarch32_neon_mlal_lane_ld64
  init: xnn_init_qs8_minmax_neon_params
  k-block: 8
  assembly: true
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x8__aarch32_neonv8_mlal_lane_prfm_ld64
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
  assembly: true
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x8__aarch32_neonv8_mlal_lane_ld64
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
  assembly: true
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x8__neon_mlal_lane
  init: xnn_init_qs8_minmax_neon_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8__neon_mlal_lane
  init: xnn_init_qs8_minmax_neon_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x8__neon_mlal_lane
  init: xnn_init_qs8_minmax_neon_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x8__neon_mlal_lane
  init: xnn_init_qs8_minmax_neon_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neon_mlal_lane
  init: xnn_init_qs8_minmax_neon_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x16__neon_mlal_lane
  init: xnn_init_qs8_minmax_neon_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x16__neon_mlal_lane
  init: xnn_init_qs8_minmax_neon_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x16__neon_mlal_lane
  init: xnn_init_qs8_minmax_neon_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x16__neon_mlal_lane
  init: xnn_init_qs8_minmax_neon_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neon_mlal_lane
  init: xnn_init_qs8_minmax_neon_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x8__neon_mlal_lane_prfm
  init: xnn_init_qs8_minmax_neon_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8__neon_mlal_lane_prfm
  init: xnn_init_qs8_minmax_neon_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x8__neon_mlal_lane_prfm
  init: xnn_init_qs8_minmax_neon_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x8__neon_mlal_lane_prfm
  init: xnn_init_qs8_minmax_neon_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neon_mlal_lane_prfm
  init: xnn_init_qs8_minmax_neon_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x16__neon_mlal_lane_prfm
  init: xnn_init_qs8_minmax_neon_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x16__neon_mlal_lane_prfm
  init: xnn_init_qs8_minmax_neon_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x16__neon_mlal_lane_prfm
  init: xnn_init_qs8_minmax_neon_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x16__neon_mlal_lane_prfm
  init: xnn_init_qs8_minmax_neon_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neon_mlal_lane_prfm
  init: xnn_init_qs8_minmax_neon_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x8__neonv8_mlal_lane
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8__neonv8_mlal_lane
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x8__neonv8_mlal_lane
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x8__neonv8_mlal_lane
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x16__neonv8_mlal_lane
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x16__neonv8_mlal_lane
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x16__neonv8_mlal_lane
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x16__neonv8_mlal_lane
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x8__neonv8_mlal_lane_prfm
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8__neonv8_mlal_lane_prfm
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x8__neonv8_mlal_lane_prfm
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x8__neonv8_mlal_lane_prfm
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane_prfm
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x16__neonv8_mlal_lane_prfm
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x16__neonv8_mlal_lane_prfm
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x16__neonv8_mlal_lane_prfm
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x16__neonv8_mlal_lane_prfm
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane_prfm
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld1r
  init: xnn_init_qs8_minmax_neon_params
  k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2__neon_mlal_ld1r
  init: xnn_init_qs8_minmax_neon_params
  k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_ld1r
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld2r
  init: xnn_init_qs8_minmax_neon_params
  k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2__neon_mlal_ld2r
  init: xnn_init_qs8_minmax_neon_params
  k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld2r
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_ld2r
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld4r
  init: xnn_init_qs8_minmax_neon_params
  k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2__neon_mlal_ld4r
  init: xnn_init_qs8_minmax_neon_params
  k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld4r
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_ld4r
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x8c4s2__neon_mlal
  init: xnn_init_qs8_minmax_neon_params
  k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4s2__neon_mlal
  init: xnn_init_qs8_minmax_neon_params
  k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x8c4s2__neonv8_mlal
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4s2__neonv8_mlal
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2s4__neon_mlal
  init: xnn_init_qs8_minmax_neon_params
  k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2s4__neon_mlal
  init: xnn_init_qs8_minmax_neon_params
  k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2s4__neonv8_mlal
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2s4__neonv8_mlal
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x8c4__neon_mlal_dup
  init: xnn_init_qs8_minmax_neon_params
  k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup
  init: xnn_init_qs8_minmax_neon_params
  k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x8c4__neonv8_mlal_dup
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x8c4__neon_mlal_ld1r
  init: xnn_init_qs8_minmax_neon_params
  k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld1r
  init: xnn_init_qs8_minmax_neon_params
  k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x8c4__neonv8_mlal_ld1r
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld1r
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x8c4__neon_mlal_ld2r
  init: xnn_init_qs8_minmax_neon_params
  k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r
  init: xnn_init_qs8_minmax_neon_params
  k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x8c4__neonv8_mlal_ld2r
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_dup
  init: xnn_init_qs8_minmax_neon_params
  k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2__neon_mlal_dup
  init: xnn_init_qs8_minmax_neon_params
  k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_dup
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_dup
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__aarch64_neon_mlal
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 16
  assembly: true
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__aarch64_neon_mlal_prfm
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 16
  assembly: true
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__aarch64_neon_mull
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
  assembly: true
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__aarch64_neon_mlal_cortex_a53
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 16
  assembly: true
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__aarch64_neon_mlal_prfm_cortex_a53
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 16
  assembly: true
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 16
  assembly: true
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__aarch64_neon_mlal
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 16
  assembly: true
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__aarch64_neon_mlal_prfm
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 16
  assembly: true
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__aarch64_neon_mlal_cortex_a53
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 16
  assembly: true
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__aarch64_neon_mlal_prfm_cortex_a53
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 16
  assembly: true
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
  assembly: true
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
  assembly: true
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x16__aarch64_neon_mlal_lane_prfm_ld64
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
  assembly: true
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x16__aarch64_neon_mlal_lane_ld64
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
  assembly: true
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x16c4__aarch64_neondot_ld32
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 4
  assembly: true
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x16c4__aarch64_neondot_ld64
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
  assembly: true
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x16c4__aarch64_neondot_ld32
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 4
  assembly: true
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x16c4__aarch64_neondot_ld64
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
  assembly: true
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x16c4__aarch64_neondot_ld128
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 16
  assembly: true
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x16c4__aarch64_neondot_cortex_a55
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 16
  assembly: true
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal
  init: xnn_init_qs8_minmax_neon_params
  k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal
  init: xnn_init_qs8_minmax_neon_params
  k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x8c4__neondot
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x8c4__neondot
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_6x8c4__neondot
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_8x8c4__neondot
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x16c4__neondot
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x16c4__neondot
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_6x16c4__neondot
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_8x16c4__neondot
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2__sse2_ld64
  init: xnn_init_qs8_minmax_sse2_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x4c2__sse2_ld64
  init: xnn_init_qs8_minmax_sse2_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2__sse2_ld64
  init: xnn_init_qs8_minmax_sse2_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x4c2__sse2_ld64
  init: xnn_init_qs8_minmax_sse2_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2__sse41_ld64
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x4c2__sse41_ld64
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2__sse41_ld64
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x4c2__sse41_ld64
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2__avx_ld64
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x4c2__avx_ld64
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2__avx_ld64
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x4c2__avx_ld64
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2__xop_ld64
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x4c2__xop_ld64
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2__xop_ld64
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x4c2__xop_ld64
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2__sse2_ld128
  init: xnn_init_qs8_minmax_sse2_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x4c2__sse2_ld128
  init: xnn_init_qs8_minmax_sse2_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2__sse2_ld128
  init: xnn_init_qs8_minmax_sse2_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x4c2__sse2_ld128
  init: xnn_init_qs8_minmax_sse2_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2__sse41_ld128
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x4c2__sse41_ld128
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2__sse41_ld128
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x4c2__sse41_ld128
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2__avx_ld128
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x4c2__avx_ld128
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2__avx_ld128
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x4c2__avx_ld128
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2__xop_ld128
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x4c2__xop_ld128
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2__xop_ld128
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x4c2__xop_ld128
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__sse2_ld64
  init: xnn_init_qs8_minmax_sse2_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x4c8__sse2_ld64
  init: xnn_init_qs8_minmax_sse2_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__sse2_ld64
  init: xnn_init_qs8_minmax_sse2_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__sse41_ld64
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x4c8__sse41_ld64
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__sse41_ld64
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__avx_ld64
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x4c8__avx_ld64
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__avx_ld64
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__xop_ld64
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x4c8__xop_ld64
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__xop_ld64
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__sse2_ld128
  init: xnn_init_qs8_minmax_sse2_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x4c8__sse2_ld128
  init: xnn_init_qs8_minmax_sse2_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__sse2_ld128
  init: xnn_init_qs8_minmax_sse2_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__sse41_ld128
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x4c8__sse41_ld128
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__sse41_ld128
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__avx_ld128
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x4c8__avx_ld128
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__avx_ld128
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__xop_ld128
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x4c8__xop_ld128
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__xop_ld128
  init: xnn_init_qs8_minmax_sse4_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__avx2
  init: xnn_init_qs8_minmax_avx2_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__avx2
  init: xnn_init_qs8_minmax_avx2_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x8c8__avx2
  init: xnn_init_qs8_minmax_avx2_params
  k-block: 8
- name: xnn_qc8_gemm_xw_minmax_fp32_ukernel_1x8c8__avx2
  init: xnn_init_qs8_minmax_avx2_params
  k-block: 8
- name: xnn_qc8_gemm_xw_minmax_fp32_ukernel_2x8c8__avx2
  init: xnn_init_qs8_minmax_avx2_params
  k-block: 8
- name: xnn_qc8_gemm_xw_minmax_fp32_ukernel_3x8c8__avx2
  init: xnn_init_qs8_minmax_avx2_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x16c8__avx512skx
  init: xnn_init_qs8_minmax_avx512_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x16c8__avx512skx
  init: xnn_init_qs8_minmax_avx512_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x16c8__avx512skx
  init: xnn_init_qs8_minmax_avx512_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx
  init: xnn_init_qs8_minmax_avx512_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2__wasmsimd_dot16x2_ld64
  init: xnn_init_qs8_minmax_wasmsimd_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x4c2__wasmsimd_dot16x2_ld64
  init: xnn_init_qs8_minmax_wasmsimd_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2__wasmsimd_dot16x2_ld64
  init: xnn_init_qs8_minmax_wasmsimd_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x4c2__wasmsimd_dot16x2_ld64
  init: xnn_init_qs8_minmax_wasmsimd_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2__wasmsimd_dot16x2_ld128
  init: xnn_init_qs8_minmax_wasmsimd_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x4c2__wasmsimd_dot16x2_ld128
  init: xnn_init_qs8_minmax_wasmsimd_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2__wasmsimd_dot16x2_ld128
  init: xnn_init_qs8_minmax_wasmsimd_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x4c2__wasmsimd_dot16x2_ld128
  init: xnn_init_qs8_minmax_wasmsimd_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2s4__wasmsimd_dot16x2_ld64
  init: xnn_init_qs8_minmax_wasmsimd_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x4c2s4__wasmsimd_dot16x2_ld64
  init: xnn_init_qs8_minmax_wasmsimd_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2s4__wasmsimd_dot16x2_ld64
  init: xnn_init_qs8_minmax_wasmsimd_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x4c2s4__wasmsimd_dot16x2_ld64
  init: xnn_init_qs8_minmax_wasmsimd_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2s4__wasmsimd_dot16x2_ld128
  init: xnn_init_qs8_minmax_wasmsimd_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x4c2s4__wasmsimd_dot16x2_ld128
  init: xnn_init_qs8_minmax_wasmsimd_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2s4__wasmsimd_dot16x2_ld128
  init: xnn_init_qs8_minmax_wasmsimd_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x4c2s4__wasmsimd_dot16x2_ld128
  init: xnn_init_qs8_minmax_wasmsimd_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__wasmsimd_dot16x2_ld64
  init: xnn_init_qs8_minmax_wasmsimd_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x4c8__wasmsimd_dot16x2_ld64
  init: xnn_init_qs8_minmax_wasmsimd_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__wasmsimd_dot16x2_ld64
  init: xnn_init_qs8_minmax_wasmsimd_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x4c8__wasmsimd_dot16x2_ld64
  init: xnn_init_qs8_minmax_wasmsimd_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__wasmsimd_dot16x2_ld128
  init: xnn_init_qs8_minmax_wasmsimd_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x4c8__wasmsimd_dot16x2_ld128
  init: xnn_init_qs8_minmax_wasmsimd_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__wasmsimd_dot16x2_ld128
  init: xnn_init_qs8_minmax_wasmsimd_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x4c8__wasmsimd_dot16x2_ld128
  init: xnn_init_qs8_minmax_wasmsimd_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__wasmsimd_mul16_ld64
  init: xnn_init_qs8_minmax_wasmsimd_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x4c8__wasmsimd_mul16_ld64
  init: xnn_init_qs8_minmax_wasmsimd_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__wasmsimd_mul16_ld64
  init: xnn_init_qs8_minmax_wasmsimd_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__wasmsimd_mul16_ld128
  init: xnn_init_qs8_minmax_wasmsimd_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x4c8__wasmsimd_mul16_ld128
  init: xnn_init_qs8_minmax_wasmsimd_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__wasmsimd_mul16_ld128
  init: xnn_init_qs8_minmax_wasmsimd_params
  k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x2__wasm_fmagic
  init: xnn_init_qs8_minmax_scalar_fmagic_params
  k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x2__wasm_fmagic
  init: xnn_init_qs8_minmax_scalar_fmagic_params
  k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x2__wasm_fmagic
  init: xnn_init_qs8_minmax_scalar_fmagic_params
  k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x2__wasm_fmagic
  init: xnn_init_qs8_minmax_scalar_fmagic_params
  k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x4__wasm_fmagic
  init: xnn_init_qs8_minmax_scalar_fmagic_params
  k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x4__wasm_fmagic
  init: xnn_init_qs8_minmax_scalar_fmagic_params
  k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x4__wasm_fmagic
  init: xnn_init_qs8_minmax_scalar_fmagic_params
  k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x4__wasm_fmagic
  init: xnn_init_qs8_minmax_scalar_fmagic_params
  k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_fmagic
  init: xnn_init_qs8_minmax_scalar_fmagic_params
  k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_fmagic
  init: xnn_init_qs8_minmax_scalar_fmagic_params
  k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_fmagic
  init: xnn_init_qs8_minmax_scalar_fmagic_params
  k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_fmagic
  init: xnn_init_qs8_minmax_scalar_fmagic_params
  k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_fmagic
  init: xnn_init_qs8_minmax_scalar_fmagic_params
  k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_fmagic
  init: xnn_init_qs8_minmax_scalar_fmagic_params
  k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_fmagic
  init: xnn_init_qs8_minmax_scalar_fmagic_params
  k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_fmagic
  init: xnn_init_qs8_minmax_scalar_fmagic_params
  k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic
  init: xnn_init_qs8_minmax_scalar_imagic_params
  k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic
  init: xnn_init_qs8_minmax_scalar_imagic_params
  k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic
  init: xnn_init_qs8_minmax_scalar_imagic_params
  k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic
  init: xnn_init_qs8_minmax_scalar_imagic_params
  k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic
  init: xnn_init_qs8_minmax_scalar_imagic_params
  k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic
  init: xnn_init_qs8_minmax_scalar_imagic_params
  k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic
  init: xnn_init_qs8_minmax_scalar_imagic_params
  k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic
  init: xnn_init_qs8_minmax_scalar_imagic_params
  k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf
  init: xnn_init_qs8_minmax_scalar_lrintf_params
  k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf
  init: xnn_init_qs8_minmax_scalar_lrintf_params
  k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf
  init: xnn_init_qs8_minmax_scalar_lrintf_params
  k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf
  init: xnn_init_qs8_minmax_scalar_lrintf_params
  k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf
  init: xnn_init_qs8_minmax_scalar_lrintf_params
  k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf
  init: xnn_init_qs8_minmax_scalar_lrintf_params
  k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf
  init: xnn_init_qs8_minmax_scalar_lrintf_params
  k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf
  init: xnn_init_qs8_minmax_scalar_lrintf_params
  k-block: 1
- name: xnn_generate_qc8_gemm_fp32_ukernel_4x8c4__aarch32_neondot_ld64
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
  assembly: true
- name: xnn_generate_qc8_gemm_fp32_ukernel_4x8__aarch32_neonv8_mlal_lane_prfm_ld64
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
  assembly: true
- name: xnn_generate_qc8_gemm_fp32_ukernel_4x8__aarch32_neonv8_mlal_lane_ld64
  init: xnn_init_qs8_minmax_neonv8_params
  k-block: 8
  assembly: true