# Copyright 2019 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.

# AArch32 assembly
- name: xnn_f32_gemm_minmax_ukernel_4x4__aarch32_vfp_ld64
  init: xnn_init_f32_minmax_scalar_params
  k-block: 2
  assembly: true
- name: xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_cortex_a7
  init: xnn_init_f32_minmax_scalar_params
  k-block: 2
  assembly: true
- name: xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_cortex_a53
  init: xnn_init_f32_minmax_scalar_params
  k-block: 4
  assembly: true
  pipelined: true
- name: xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_cortex_a55
  init: xnn_init_f32_minmax_scalar_params
  k-block: 4
  assembly: true
  pipelined: true
- name: xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_cortex_a75
  init: xnn_init_f32_minmax_scalar_params
  k-block: 4
  assembly: true
  pipelined: true
- name: xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_ld64
  init: xnn_init_f32_minmax_scalar_params
  k-block: 2
  assembly: true
- name: xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_prfm_cortex_a53
  init: xnn_init_f32_minmax_scalar_params
  k-block: 4
  assembly: true
  pipelined: true
- name: xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_prfm_cortex_a75
  init: xnn_init_f32_minmax_scalar_params
  k-block: 4
  assembly: true
  pipelined: true
# AArch64 assembly
- name: xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a53
  init: xnn_init_f32_minmax_scalar_params
  k-block: 8
  assembly: true
  pipelined: true
- name: xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a75
  init: xnn_init_f32_minmax_scalar_params
  k-block: 8
  assembly: true
  pipelined: true
- name: xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_ld64
  init: xnn_init_f32_minmax_scalar_params
  k-block: 2
  assembly: true
- name: xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_prfm_cortex_a53
  init: xnn_init_f32_minmax_scalar_params
  k-block: 8
  assembly: true
  pipelined: true
- name: xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_prfm_cortex_a75
  init: xnn_init_f32_minmax_scalar_params
  k-block: 8
  assembly: true
  pipelined: true
- name: xnn_f32_gemm_minmax_ukernel_1x12__aarch64_neonfma_cortex_a53
  init: xnn_init_f32_minmax_scalar_params
  k-block: 4
  assembly: true
  pipelined: true
- name: xnn_f32_gemm_minmax_ukernel_4x2__aarch64_neonfma_cortex_a75
  init: xnn_init_f32_minmax_scalar_params
  k-block: 8
  assembly: true
- name: xnn_f32_gemm_minmax_ukernel_4x2__aarch64_neonfma_ld64
  init: xnn_init_f32_minmax_scalar_params
  k-block: 2
  assembly: true
- name: xnn_f32_gemm_minmax_ukernel_4x2__aarch64_neonfma_prfm_cortex_a75
  init: xnn_init_f32_minmax_scalar_params
  k-block: 8
  assembly: true
- name: xnn_f32_gemm_minmax_ukernel_4x8__aarch64_neonfma_cortex_a53
  init: xnn_init_f32_minmax_scalar_params
  k-block: 4
  assembly: true
  pipelined: true
- name: xnn_f32_gemm_minmax_ukernel_4x8__aarch64_neonfma_cortex_a55
  init: xnn_init_f32_minmax_scalar_params
  k-block: 4
  assembly: true
  pipelined: true
- name: xnn_f32_gemm_minmax_ukernel_4x8__aarch64_neonfma_cortex_a75
  init: xnn_init_f32_minmax_scalar_params
  k-block: 8
  assembly: true
  pipelined: true
- name: xnn_f32_gemm_minmax_ukernel_4x8__aarch64_neonfma_ld64
  init: xnn_init_f32_minmax_scalar_params
  k-block: 2
  assembly: true
- name: xnn_f32_gemm_minmax_ukernel_4x8__aarch64_neonfma_ld128
  init: xnn_init_f32_minmax_scalar_params
  k-block: 4
  assembly: true
- name: xnn_f32_gemm_minmax_ukernel_4x8__aarch64_neonfma_prfm_cortex_a53
  init: xnn_init_f32_minmax_scalar_params
  k-block: 4
  assembly: true
  pipelined: true
- name: xnn_f32_gemm_minmax_ukernel_4x8__aarch64_neonfma_prfm_cortex_a75
  init: xnn_init_f32_minmax_scalar_params
  k-block: 8
  assembly: true
  pipelined: true
- name: xnn_f32_gemm_minmax_ukernel_4x12__aarch64_neonfma_cortex_a53
  init: xnn_init_f32_minmax_scalar_params
  k-block: 4
  assembly: true
  pipelined: true
- name: xnn_f32_gemm_minmax_ukernel_5x8__aarch64_neonfma_cortex_a75
  init: xnn_init_f32_minmax_scalar_params
  k-block: 8
  assembly: true
  pipelined: true
- name: xnn_f32_gemm_minmax_ukernel_5x8__aarch64_neonfma_prfm_cortex_a75
  init: xnn_init_f32_minmax_scalar_params
  k-block: 8
  assembly: true
  pipelined: true
- name: xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a53
  init: xnn_init_f32_minmax_scalar_params
  k-block: 4
  assembly: true
  pipelined: true
- name: xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a55
  init: xnn_init_f32_minmax_scalar_params
  k-block: 4
  assembly: true
  pipelined: true
- name: xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a73
  init: xnn_init_f32_minmax_scalar_params
  k-block: 8
  assembly: true
  pipelined: true
- name: xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a75
  init: xnn_init_f32_minmax_scalar_params
  k-block: 8
  assembly: true
  pipelined: true
- name: xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_ld64
  init: xnn_init_f32_minmax_scalar_params
  k-block: 2
  assembly: true
- name: xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_ld128
  init: xnn_init_f32_minmax_scalar_params
  k-block: 4
  assembly: true
- name: xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_prfm_cortex_a53
  init: xnn_init_f32_minmax_scalar_params
  k-block: 4
  assembly: true
  pipelined: true
- name: xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_prfm_cortex_a75
  init: xnn_init_f32_minmax_scalar_params
  k-block: 8
  assembly: true
  pipelined: true
# ARM NEON
- name: xnn_f32_gemm_minmax_ukernel_1x8__neon_dup_ld64
  init: xnn_init_f32_minmax_scalar_params
  k-block: 2
- name: xnn_f32_gemm_minmax_ukernel_1x8__neon_lane_ld64
  init: xnn_init_f32_minmax_scalar_params
  k-block: 2
- name: xnn_f32_gemm_minmax_ukernel_1x8__neonfma_dup_ld64
  init: xnn_init_f32_minmax_scalar_params
  k-block: 2
- name: xnn_f32_gemm_minmax_ukernel_1x8__neonfma_lane_ld64
  init: xnn_init_f32_minmax_scalar_params
  k-block: 2
  arch:
    - aarch64
- name: xnn_f32_gemm_minmax_ukernel_1x8s4__neon
  init: xnn_init_f32_minmax_scalar_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_1x8s4__neonfma
  init: xnn_init_f32_minmax_scalar_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_4x2__neon_lane_ld64
  init: xnn_init_f32_minmax_scalar_params
  k-block: 2
- name: xnn_f32_gemm_minmax_ukernel_4x2__neonfma_lane_ld64
  init: xnn_init_f32_minmax_scalar_params
  k-block: 2
  arch:
    - aarch64
- name: xnn_f32_gemm_minmax_ukernel_4x8__neon_dup_ld64
  init: xnn_init_f32_minmax_scalar_params
  k-block: 2
- name: xnn_f32_gemm_minmax_ukernel_4x8__neon_dup_ld128
  init: xnn_init_f32_minmax_scalar_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_4x8__neon_lane_ld64
  init: xnn_init_f32_minmax_scalar_params
  k-block: 2
- name: xnn_f32_gemm_minmax_ukernel_4x8__neon_lane_ld128
  init: xnn_init_f32_minmax_scalar_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_4x8__neonfma_dup_ld64
  init: xnn_init_f32_minmax_scalar_params
  k-block: 2
- name: xnn_f32_gemm_minmax_ukernel_4x8__neonfma_dup_ld128
  init: xnn_init_f32_minmax_scalar_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_4x8__neonfma_lane_ld64
  init: xnn_init_f32_minmax_scalar_params
  k-block: 2
  arch:
    - aarch64
- name: xnn_f32_gemm_minmax_ukernel_4x8__neonfma_lane_ld128
  init: xnn_init_f32_minmax_scalar_params
  k-block: 4
  arch:
    - aarch64
- name: xnn_f32_gemm_minmax_ukernel_4x8s4__neon
  init: xnn_init_f32_minmax_scalar_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_4x8s4__neonfma
  init: xnn_init_f32_minmax_scalar_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_5x8__neon_lane_ld64
  init: xnn_init_f32_minmax_scalar_params
  k-block: 2
- name: xnn_f32_gemm_minmax_ukernel_5x8__neonfma_lane_ld64
  init: xnn_init_f32_minmax_scalar_params
  k-block: 2
  arch:
    - aarch64
- name: xnn_f32_gemm_minmax_ukernel_6x2__neon_lane_ld64
  init: xnn_init_f32_minmax_scalar_params
  k-block: 2
- name: xnn_f32_gemm_minmax_ukernel_6x2__neonfma_lane_ld64
  init: xnn_init_f32_minmax_scalar_params
  k-block: 2
  arch:
    - aarch64
- name: xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld64
  init: xnn_init_f32_minmax_scalar_params
  k-block: 2
- name: xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld128
  init: xnn_init_f32_minmax_scalar_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64
  init: xnn_init_f32_minmax_scalar_params
  k-block: 2
- name: xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld128
  init: xnn_init_f32_minmax_scalar_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld64
  init: xnn_init_f32_minmax_scalar_params
  k-block: 2
- name: xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld128
  init: xnn_init_f32_minmax_scalar_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64
  init: xnn_init_f32_minmax_scalar_params
  k-block: 2
  arch:
    - aarch64
- name: xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld128
  init: xnn_init_f32_minmax_scalar_params
  k-block: 4
  arch:
    - aarch64
- name: xnn_f32_gemm_minmax_ukernel_6x8s4__neon
  init: xnn_init_f32_minmax_scalar_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_6x8s4__neonfma
  init: xnn_init_f32_minmax_scalar_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_8x8s4__neon
  init: xnn_init_f32_minmax_scalar_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_8x8s4__neonfma
  init: xnn_init_f32_minmax_scalar_params
  k-block: 4
# x86 SSE
- name: xnn_f32_gemm_minmax_ukernel_1x8__sse_dup
  init: xnn_init_f32_minmax_sse_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_1x8__sse_load1
  init: xnn_init_f32_minmax_sse_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_1x8__sse2_dup
  init: xnn_init_f32_minmax_sse_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_1x8s4__sse
  init: xnn_init_f32_minmax_sse_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_3x8__sse_dup
  init: xnn_init_f32_minmax_sse_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_3x8__sse_load1
  init: xnn_init_f32_minmax_sse_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_3x8__sse2_dup
  init: xnn_init_f32_minmax_sse_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_3x8s4__sse
  init: xnn_init_f32_minmax_sse_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_4x2c4__sse
  init: xnn_init_f32_minmax_sse_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_4x8__sse_dup
  init: xnn_init_f32_minmax_sse_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_4x8__sse_load1
  init: xnn_init_f32_minmax_sse_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_4x8__sse2_dup
  init: xnn_init_f32_minmax_sse_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_4x8s4__sse
  init: xnn_init_f32_minmax_sse_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_5x8__sse_dup
  init: xnn_init_f32_minmax_sse_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_5x8__sse_load1
  init: xnn_init_f32_minmax_sse_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_5x8__sse2_dup
  init: xnn_init_f32_minmax_sse_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_5x8s4__sse
  init: xnn_init_f32_minmax_sse_params
  k-block: 4
# x86 AVX
- name: xnn_f32_gemm_minmax_ukernel_1x8__avx_broadcast
  init: xnn_init_f32_minmax_avx_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_1x16__avx_broadcast
  init: xnn_init_f32_minmax_avx_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_3x16__avx_broadcast
  init: xnn_init_f32_minmax_avx_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_4x8__avx_broadcast
  init: xnn_init_f32_minmax_avx_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_4x16__avx_broadcast
  init: xnn_init_f32_minmax_avx_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_5x8__avx_broadcast
  init: xnn_init_f32_minmax_avx_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_5x16__avx_broadcast
  init: xnn_init_f32_minmax_avx_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_6x8__avx_broadcast
  init: xnn_init_f32_minmax_avx_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_7x8__avx_broadcast
  init: xnn_init_f32_minmax_avx_params
  k-block: 1
# x86 FMA3
- name: xnn_f32_gemm_minmax_ukernel_1x8__fma3_broadcast
  init: xnn_init_f32_minmax_avx_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_1x16__fma3_broadcast
  init: xnn_init_f32_minmax_avx_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_1x16s4__fma3_broadcast
  init: xnn_init_f32_minmax_avx_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_3x16__fma3_broadcast
  init: xnn_init_f32_minmax_avx_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_3x16s4__fma3_broadcast
  init: xnn_init_f32_minmax_avx_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_4x8__fma3_broadcast
  init: xnn_init_f32_minmax_avx_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_4x16__fma3_broadcast
  init: xnn_init_f32_minmax_avx_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_4x16s4__fma3_broadcast
  init: xnn_init_f32_minmax_avx_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_5x8__fma3_broadcast
  init: xnn_init_f32_minmax_avx_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_5x16__fma3_broadcast
  init: xnn_init_f32_minmax_avx_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_5x16s4__fma3_broadcast
  init: xnn_init_f32_minmax_avx_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_6x8__fma3_broadcast
  init: xnn_init_f32_minmax_avx_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_7x8__fma3_broadcast
  init: xnn_init_f32_minmax_avx_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_8x8__fma3_broadcast
  init: xnn_init_f32_minmax_avx_params
  k-block: 1
# x86 AVX512
- name: xnn_f32_gemm_minmax_ukernel_1x16__avx512f_broadcast
  init: xnn_init_f32_minmax_scalar_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_4x16__avx512f_broadcast
  init: xnn_init_f32_minmax_scalar_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_5x16__avx512f_broadcast
  init: xnn_init_f32_minmax_scalar_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_6x16__avx512f_broadcast
  init: xnn_init_f32_minmax_scalar_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_7x16__avx512f_broadcast
  init: xnn_init_f32_minmax_scalar_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_8x16__avx512f_broadcast
  init: xnn_init_f32_minmax_scalar_params
  k-block: 1
# WAsm SIMD
- name: xnn_f32_gemm_minmax_ukernel_1x8__wasmsimd_arm_loadsplat
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_1x8__wasmsimd_arm_splat
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_1x8__wasmsimd_x86_loadsplat
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_1x8__wasmsimd_x86_splat
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_1x8s4__wasmsimd_arm
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_1x8s4__wasmsimd_x86
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_3x8__wasmsimd_arm_loadsplat
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_3x8__wasmsimd_arm_splat
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_3x8__wasmsimd_x86_loadsplat
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_3x8__wasmsimd_x86_splat
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_3x8s4__wasmsimd_arm
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_3x8s4__wasmsimd_x86
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_4x2c4__wasmsimd_arm
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_4x2c4__wasmsimd_x86
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_4x8__wasmsimd_arm_loadsplat
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_4x8__wasmsimd_arm_splat
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_4x8__wasmsimd_x86_loadsplat
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_4x8__wasmsimd_x86_splat
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_4x8s4__wasmsimd_arm
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_4x8s4__wasmsimd_x86
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_5x8__wasmsimd_arm_loadsplat
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_5x8__wasmsimd_arm_splat
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_5x8__wasmsimd_x86_splat
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_5x8s4__wasmsimd_arm
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_5x8s4__wasmsimd_x86
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_arm_loadsplat
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_arm_splat
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_x86_splat
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_6x8s4__wasmsimd_arm
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_6x8s4__wasmsimd_x86
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
# WAsm Relaxed SIMD
- name: xnn_f32_gemm_minmax_ukernel_1x8__wasmrelaxedsimd_fma_loadsplat
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_1x8__wasmrelaxedsimd_fma_splat
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_1x8__wasmrelaxedsimd_loadsplat
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_1x8__wasmrelaxedsimd_splat
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_1x8s4__wasmrelaxedsimd
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_1x8s4__wasmrelaxedsimd_fma
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_3x8__wasmrelaxedsimd_fma_loadsplat
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_3x8__wasmrelaxedsimd_fma_splat
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_3x8__wasmrelaxedsimd_loadsplat
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_3x8__wasmrelaxedsimd_splat
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_3x8s4__wasmrelaxedsimd
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_3x8s4__wasmrelaxedsimd_fma
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_4x2c4__wasmrelaxedsimd
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_4x2c4__wasmrelaxedsimd_fma
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_4x8__wasmrelaxedsimd_fma_loadsplat
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_4x8__wasmrelaxedsimd_fma_splat
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_4x8__wasmrelaxedsimd_loadsplat
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_4x8__wasmrelaxedsimd_splat
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_4x8s4__wasmrelaxedsimd
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_4x8s4__wasmrelaxedsimd_fma
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_5x8__wasmrelaxedsimd_fma_loadsplat
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_5x8__wasmrelaxedsimd_fma_splat
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_5x8__wasmrelaxedsimd_loadsplat
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_5x8__wasmrelaxedsimd_splat
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_5x8s4__wasmrelaxedsimd
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_5x8s4__wasmrelaxedsimd_fma
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_6x8__wasmrelaxedsimd_fma_loadsplat
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_6x8__wasmrelaxedsimd_fma_splat
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_6x8__wasmrelaxedsimd_loadsplat
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_6x8__wasmrelaxedsimd_splat
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_6x8s4__wasmrelaxedsimd
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
- name: xnn_f32_gemm_minmax_ukernel_6x8s4__wasmrelaxedsimd_fma
  init: xnn_init_f32_minmax_wasmsimd_params
  k-block: 4
# WAsm
- name: xnn_f32_gemm_minmax_ukernel_1x4__wasm
  init: xnn_init_f32_minmax_scalar_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_2x4__wasm
  init: xnn_init_f32_minmax_scalar_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_4x2__wasm
  init: xnn_init_f32_minmax_scalar_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_4x4__wasm
  init: xnn_init_f32_minmax_scalar_params
  k-block: 1
# Scalar
- name: xnn_f32_gemm_minmax_ukernel_1x4__scalar
  init: xnn_init_f32_minmax_scalar_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_2x4__scalar
  init: xnn_init_f32_minmax_scalar_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_4x2__scalar
  init: xnn_init_f32_minmax_scalar_params
  k-block: 1
- name: xnn_f32_gemm_minmax_ukernel_4x4__scalar
  init: xnn_init_f32_minmax_scalar_params
  k-block: 1
# AArch32 JIT assembly
- name: xnn_generate_f32_gemm_ukernel_4x8__aarch32_neon_cortex_a7
  init: xnn_init_f32_minmax_scalar_params
  k-block: 2
  assembly: true
- name: xnn_generate_f32_gemm_ukernel_4x8__aarch32_neon_cortex_a53
  init: xnn_init_f32_minmax_scalar_params
  k-block: 4
  pipelined: true
  assembly: true
- name: xnn_generate_f32_gemm_ukernel_4x8__aarch32_neon_cortex_a55
  init: xnn_init_f32_minmax_scalar_params
  k-block: 4
  pipelined: true
- name: xnn_generate_f32_gemm_ukernel_4x8__aarch32_neon_cortex_a75
  init: xnn_init_f32_minmax_scalar_params
  k-block: 4
  pipelined: true
  assembly: true
- name: xnn_generate_f32_gemm_ukernel_4x8__aarch32_neon_ld64
  init: xnn_init_f32_minmax_scalar_params
  k-block: 2
  assembly: true
- name: xnn_generate_f32_gemm_ukernel_4x8__aarch32_neon_prfm_cortex_a75
  init: xnn_init_f32_minmax_scalar_params
  k-block: 4
  pipelined: true
  assembly: true
# AArch64 JIT assembly
- name: xnn_generate_f32_gemm_ukernel_upto6x8__aarch64_neonfma_cortex_a75
  init: xnn_init_f32_minmax_scalar_params
  k-block: 8
  pipelined: true
- name: xnn_generate_f32_gemm_ukernel_upto6x8__aarch64_neonfma_prfm_cortex_a75
  init: xnn_init_f32_minmax_scalar_params
  k-block: 8
  pipelined: true
- name: xnn_generate_f32_gemm_ukernel_1x8__aarch64_neonfma_cortex_a75
  init: xnn_init_f32_minmax_scalar_params
  k-block: 8
  pipelined: true
- name: xnn_generate_f32_gemm_ukernel_1x8__aarch64_neonfma_prfm_cortex_a75
  init: xnn_init_f32_minmax_scalar_params
  k-block: 8
  pipelined: true
- name: xnn_generate_f32_gemm_ukernel_4x8__aarch64_neonfma_cortex_a75
  init: xnn_init_f32_minmax_scalar_params
  k-block: 8
  pipelined: true
- name: xnn_generate_f32_gemm_ukernel_4x8__aarch64_neonfma_prfm_cortex_a75
  init: xnn_init_f32_minmax_scalar_params
  k-block: 8
  pipelined: true
- name: xnn_generate_f32_gemm_ukernel_6x8__aarch64_neonfma_ld128
  init: xnn_init_f32_minmax_scalar_params
  k-block: 4