blob: 79c040ae2b67fb5efc3a4624c9f395028b8a9772 [file] [log] [blame]
# 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_2x8c8__aarch64_neon_mlal_padal
init: xnn_init_qs8_minmax_neon_params
k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm
init: xnn_init_qs8_minmax_neon_params
k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__aarch64_neon_mull_padal
init: xnn_init_qs8_minmax_neon_params
k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53
init: xnn_init_qs8_minmax_neon_params
k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53
init: xnn_init_qs8_minmax_neon_params
k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal
init: xnn_init_qs8_minmax_neon_params
k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__aarch64_neon_mlal_padal
init: xnn_init_qs8_minmax_neon_params
k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm
init: xnn_init_qs8_minmax_neon_params
k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53
init: xnn_init_qs8_minmax_neon_params
k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53
init: xnn_init_qs8_minmax_neon_params
k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53
init: xnn_init_qs8_minmax_neon_params
k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53
init: xnn_init_qs8_minmax_neon_params
k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x16c4__aarch64_neondot_ld32
init: xnn_init_qs8_minmax_neon_params
k-block: 4
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x16c4__aarch64_neondot_ld64
init: xnn_init_qs8_minmax_neon_params
k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x16c4__aarch64_neondot_ld32
init: xnn_init_qs8_minmax_neon_params
k-block: 4
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x16c4__aarch64_neondot_ld64
init: xnn_init_qs8_minmax_neon_params
k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x16c4__aarch64_neondot_ld128
init: xnn_init_qs8_minmax_neon_params
k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x16c4__aarch64_neondot_cortex_a55
init: xnn_init_qs8_minmax_neon_params
k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x16__neon_mlal_lane
init: xnn_init_qs8_minmax_neon_fp32_params
k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x16__neon_mlal_lane
init: xnn_init_qs8_minmax_neon_fp32_params
k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x16__neonv8_mlal_lane
init: xnn_init_qs8_minmax_neon_params
k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x16__neonv8_mlal_lane
init: xnn_init_qs8_minmax_neon_params
k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_padal_dup
init: xnn_init_qs8_minmax_neon_fp32_params
k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2__neon_mlal_padal_dup
init: xnn_init_qs8_minmax_neon_fp32_params
k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_padal_dup
init: xnn_init_qs8_minmax_neon_params
k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_padal_dup
init: xnn_init_qs8_minmax_neon_params
k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal_padal
init: xnn_init_qs8_minmax_neon_fp32_params
k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal_padal
init: xnn_init_qs8_minmax_neon_fp32_params
k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal_padal
init: xnn_init_qs8_minmax_neon_params
k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal_padal
init: xnn_init_qs8_minmax_neon_params
k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x8c4__neondot
init: xnn_init_qs8_minmax_neon_params
k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x8c4__neondot
init: xnn_init_qs8_minmax_neon_params
k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_6x8c4__neondot
init: xnn_init_qs8_minmax_neon_params
k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_8x8c4__neondot
init: xnn_init_qs8_minmax_neon_params
k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x16c4__neondot
init: xnn_init_qs8_minmax_neon_params
k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x16c4__neondot
init: xnn_init_qs8_minmax_neon_params
k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_6x16c4__neondot
init: xnn_init_qs8_minmax_neon_params
k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_8x16c4__neondot
init: xnn_init_qs8_minmax_neon_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_1x4c8__wasmsimd_ld64
init: xnn_init_qs8_minmax_wasmsimd_params
k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x4c8__wasmsimd_ld64
init: xnn_init_qs8_minmax_wasmsimd_params
k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__wasmsimd_ld64
init: xnn_init_qs8_minmax_wasmsimd_params
k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__wasmsimd_ld128
init: xnn_init_qs8_minmax_wasmsimd_params
k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x4c8__wasmsimd_ld128
init: xnn_init_qs8_minmax_wasmsimd_params
k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__wasmsimd_ld128
init: xnn_init_qs8_minmax_wasmsimd_params
k-block: 8
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrint
init: xnn_init_qs8_minmax_scalar_lrint_params
k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrint
init: xnn_init_qs8_minmax_scalar_lrint_params
k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrint
init: xnn_init_qs8_minmax_scalar_lrint_params
k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrint
init: xnn_init_qs8_minmax_scalar_lrint_params
k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrint
init: xnn_init_qs8_minmax_scalar_lrint_params
k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrint
init: xnn_init_qs8_minmax_scalar_lrint_params
k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrint
init: xnn_init_qs8_minmax_scalar_lrint_params
k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrint
init: xnn_init_qs8_minmax_scalar_lrint_params
k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_magic
init: xnn_init_qs8_minmax_scalar_magic_params
k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_magic
init: xnn_init_qs8_minmax_scalar_magic_params
k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_magic
init: xnn_init_qs8_minmax_scalar_magic_params
k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_magic
init: xnn_init_qs8_minmax_scalar_magic_params
k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_magic
init: xnn_init_qs8_minmax_scalar_magic_params
k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_magic
init: xnn_init_qs8_minmax_scalar_magic_params
k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_magic
init: xnn_init_qs8_minmax_scalar_magic_params
k-block: 1
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_magic
init: xnn_init_qs8_minmax_scalar_magic_params
k-block: 1