Marat Dukhan | 9b474cf | 2021-05-25 16:37:48 -0700 | [diff] [blame] | 1 | # Copyright 2021 Google LLC |
| 2 | # |
| 3 | # This source code is licensed under the BSD-style license found in the |
| 4 | # LICENSE file in the root directory of this source tree. |
| 5 | |
Frank Barchard | 287952a | 2021-11-03 15:26:45 -0700 | [diff] [blame] | 6 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x8c4__neon_mlal_padal_dup |
| 7 | init: xnn_init_qs8_conv_minmax_fp32_neon_params |
| 8 | k-block: 16 |
| 9 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_padal_dup |
| 10 | init: xnn_init_qs8_conv_minmax_fp32_neon_params |
| 11 | k-block: 16 |
| 12 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x8c4__neonv8_mlal_padal_dup |
| 13 | init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| 14 | k-block: 16 |
| 15 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_padal_dup |
| 16 | init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| 17 | k-block: 16 |
| 18 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_padal_dup |
| 19 | init: xnn_init_qs8_conv_minmax_fp32_neon_params |
| 20 | k-block: 16 |
| 21 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2__neon_mlal_padal_dup |
| 22 | init: xnn_init_qs8_conv_minmax_fp32_neon_params |
| 23 | k-block: 16 |
| 24 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_padal_dup |
| 25 | init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| 26 | k-block: 16 |
| 27 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_padal_dup |
| 28 | init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| 29 | k-block: 16 |
Frank Barchard | 960ae34 | 2021-07-01 11:31:11 -0700 | [diff] [blame] | 30 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__aarch64_neon_mlal_padal |
| 31 | init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| 32 | k-block: 16 |
| 33 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm |
| 34 | init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| 35 | k-block: 16 |
| 36 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53 |
| 37 | init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| 38 | k-block: 16 |
| 39 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53 |
| 40 | init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| 41 | k-block: 16 |
Frank Barchard | 1663c0c | 2021-07-01 11:20:06 -0700 | [diff] [blame] | 42 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal |
| 43 | init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| 44 | k-block: 16 |
Frank Barchard | 960ae34 | 2021-07-01 11:31:11 -0700 | [diff] [blame] | 45 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__aarch64_neon_mlal_padal |
| 46 | init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| 47 | k-block: 16 |
| 48 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm |
| 49 | init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| 50 | k-block: 16 |
| 51 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53 |
| 52 | init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| 53 | k-block: 16 |
| 54 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53 |
| 55 | init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| 56 | k-block: 16 |
Frank Barchard | 98af05c | 2021-06-30 12:15:04 -0700 | [diff] [blame] | 57 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53 |
| 58 | init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| 59 | k-block: 8 |
| 60 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53 |
| 61 | init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| 62 | k-block: 8 |
Frank Barchard | 1a0b276 | 2021-06-29 18:37:59 -0700 | [diff] [blame] | 63 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_4x16c4__aarch64_neondot_cortex_a55 |
| 64 | init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| 65 | k-block: 16 |
| 66 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_4x16c4__aarch64_neondot_ld64 |
| 67 | init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| 68 | k-block: 8 |
| 69 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_4x16c4__aarch64_neondot_ld128 |
| 70 | init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| 71 | k-block: 16 |
Marat Dukhan | cf05585 | 2021-06-26 09:05:09 -0700 | [diff] [blame] | 72 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x16__neon_mlal_lane |
| 73 | init: xnn_init_qs8_conv_minmax_fp32_neon_params |
| 74 | k-block: 8 |
| 75 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_4x16__neon_mlal_lane |
| 76 | init: xnn_init_qs8_conv_minmax_fp32_neon_params |
| 77 | k-block: 8 |
| 78 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x16__neonv8_mlal_lane |
| 79 | init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| 80 | k-block: 8 |
| 81 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_4x16__neonv8_mlal_lane |
| 82 | init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| 83 | k-block: 8 |
| 84 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal_padal |
| 85 | init: xnn_init_qs8_conv_minmax_fp32_neon_params |
| 86 | k-block: 16 |
| 87 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal_padal |
| 88 | init: xnn_init_qs8_conv_minmax_fp32_neon_params |
| 89 | k-block: 16 |
| 90 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal_padal |
| 91 | init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| 92 | k-block: 16 |
| 93 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal_padal |
| 94 | init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| 95 | k-block: 16 |
Marat Dukhan | 18630de | 2021-06-02 22:20:01 -0700 | [diff] [blame] | 96 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x8c4__neondot |
| 97 | init: xnn_init_qs8_conv_minmax_fp32_neonv8_params |
| 98 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 99 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x4c2__sse2_ld64 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 100 | init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| 101 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 102 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x4c2__sse2_ld64 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 103 | init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| 104 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 105 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_3x4c2__sse2_ld64 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 106 | init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| 107 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 108 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_4x4c2__sse2_ld64 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 109 | init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| 110 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 111 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x4c2__sse41_ld64 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 112 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 113 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 114 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x4c2__sse41_ld64 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 115 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 116 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 117 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_3x4c2__sse41_ld64 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 118 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 119 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 120 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_4x4c2__sse41_ld64 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 121 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 122 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 123 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x4c2__avx_ld64 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 124 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 125 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 126 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x4c2__avx_ld64 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 127 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 128 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 129 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_3x4c2__avx_ld64 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 130 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 131 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 132 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_4x4c2__avx_ld64 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 133 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 134 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 135 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x4c2__xop_ld64 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 136 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 137 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 138 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x4c2__xop_ld64 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 139 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 140 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 141 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_3x4c2__xop_ld64 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 142 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 143 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 144 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_4x4c2__xop_ld64 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 145 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 146 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 147 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x4c2__sse2_ld128 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 148 | init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| 149 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 150 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x4c2__sse2_ld128 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 151 | init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| 152 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 153 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_3x4c2__sse2_ld128 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 154 | init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| 155 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 156 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_4x4c2__sse2_ld128 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 157 | init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| 158 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 159 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x4c2__sse41_ld128 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 160 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 161 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 162 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x4c2__sse41_ld128 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 163 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 164 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 165 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_3x4c2__sse41_ld128 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 166 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 167 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 168 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_4x4c2__sse41_ld128 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 169 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 170 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 171 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x4c2__avx_ld128 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 172 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 173 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 174 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x4c2__avx_ld128 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 175 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 176 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 177 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_3x4c2__avx_ld128 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 178 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 179 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 180 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_4x4c2__avx_ld128 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 181 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 182 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 183 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x4c2__xop_ld128 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 184 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 185 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 186 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x4c2__xop_ld128 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 187 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 188 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 189 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_3x4c2__xop_ld128 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 190 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 191 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 192 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_4x4c2__xop_ld128 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 193 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 194 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 195 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x4c8__sse2_ld64 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 196 | init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| 197 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 198 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x4c8__sse2_ld64 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 199 | init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| 200 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 201 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_3x4c8__sse2_ld64 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 202 | init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| 203 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 204 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x4c8__ssse3_ld64 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 205 | init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| 206 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 207 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x4c8__ssse3_ld64 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 208 | init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| 209 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 210 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_3x4c8__ssse3_ld64 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 211 | init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| 212 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 213 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x4c8__sse41_ld64 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 214 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 215 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 216 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x4c8__sse41_ld64 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 217 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 218 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 219 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_3x4c8__sse41_ld64 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 220 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 221 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 222 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x4c8__avx_ld64 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 223 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 224 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 225 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x4c8__avx_ld64 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 226 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 227 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 228 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_3x4c8__avx_ld64 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 229 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 230 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 231 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x4c8__xop_ld64 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 232 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 233 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 234 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x4c8__xop_ld64 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 235 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 236 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 237 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_3x4c8__xop_ld64 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 238 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 239 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 240 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x4c8__sse2_ld128 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 241 | init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| 242 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 243 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x4c8__sse2_ld128 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 244 | init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| 245 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 246 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_3x4c8__sse2_ld128 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 247 | init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| 248 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 249 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x4c8__ssse3_ld128 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 250 | init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| 251 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 252 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x4c8__ssse3_ld128 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 253 | init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| 254 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 255 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_3x4c8__ssse3_ld128 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 256 | init: xnn_init_qs8_conv_minmax_fp32_sse2_params |
| 257 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 258 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x4c8__sse41_ld128 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 259 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 260 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 261 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x4c8__sse41_ld128 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 262 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 263 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 264 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_3x4c8__sse41_ld128 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 265 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 266 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 267 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x4c8__avx_ld128 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 268 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 269 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 270 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x4c8__avx_ld128 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 271 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 272 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 273 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_3x4c8__avx_ld128 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 274 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 275 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 276 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x4c8__xop_ld128 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 277 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 278 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 279 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x4c8__xop_ld128 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 280 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 281 | k-block: 8 |
Marat Dukhan | 801d2c2 | 2021-06-02 21:25:05 -0700 | [diff] [blame] | 282 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_3x4c8__xop_ld128 |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 283 | init: xnn_init_qs8_conv_minmax_fp32_sse4_params |
| 284 | k-block: 8 |
Marat Dukhan | 9b474cf | 2021-05-25 16:37:48 -0700 | [diff] [blame] | 285 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__avx2 |
| 286 | init: xnn_init_qs8_conv_minmax_fp32_avx2_params |
| 287 | k-block: 8 |
| 288 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__avx2 |
| 289 | init: xnn_init_qs8_conv_minmax_fp32_avx2_params |
| 290 | k-block: 8 |
| 291 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_3x8c8__avx2 |
| 292 | init: xnn_init_qs8_conv_minmax_fp32_avx2_params |
| 293 | k-block: 8 |
Marat Dukhan | 71855ee | 2021-05-25 19:05:06 -0700 | [diff] [blame] | 294 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x16c8__avx512skx |
| 295 | init: xnn_init_qs8_conv_minmax_fp32_avx512_params |
| 296 | k-block: 8 |
| 297 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x16c8__avx512skx |
| 298 | init: xnn_init_qs8_conv_minmax_fp32_avx512_params |
| 299 | k-block: 8 |
| 300 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_3x16c8__avx512skx |
| 301 | init: xnn_init_qs8_conv_minmax_fp32_avx512_params |
| 302 | k-block: 8 |
| 303 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx |
| 304 | init: xnn_init_qs8_conv_minmax_fp32_avx512_params |
| 305 | k-block: 8 |
Marat Dukhan | 8dc106e | 2021-08-31 15:23:02 -0700 | [diff] [blame] | 306 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x4c2__wasmsimd_dot16x2_ld64 |
| 307 | init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| 308 | k-block: 8 |
| 309 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x4c2__wasmsimd_dot16x2_ld64 |
| 310 | init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| 311 | k-block: 8 |
| 312 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_3x4c2__wasmsimd_dot16x2_ld64 |
| 313 | init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| 314 | k-block: 8 |
| 315 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_4x4c2__wasmsimd_dot16x2_ld64 |
| 316 | init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| 317 | k-block: 8 |
| 318 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x4c2__wasmsimd_dot16x2_ld128 |
| 319 | init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| 320 | k-block: 8 |
| 321 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x4c2__wasmsimd_dot16x2_ld128 |
| 322 | init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| 323 | k-block: 8 |
| 324 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_3x4c2__wasmsimd_dot16x2_ld128 |
| 325 | init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| 326 | k-block: 8 |
| 327 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_4x4c2__wasmsimd_dot16x2_ld128 |
| 328 | init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| 329 | k-block: 8 |
| 330 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x4c8__wasmsimd_dot16x2_ld64 |
| 331 | init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| 332 | k-block: 8 |
| 333 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x4c8__wasmsimd_dot16x2_ld64 |
| 334 | init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| 335 | k-block: 8 |
| 336 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_3x4c8__wasmsimd_dot16x2_ld64 |
| 337 | init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| 338 | k-block: 8 |
| 339 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_4x4c8__wasmsimd_dot16x2_ld64 |
| 340 | init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| 341 | k-block: 8 |
| 342 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x4c8__wasmsimd_dot16x2_ld128 |
| 343 | init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| 344 | k-block: 8 |
| 345 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x4c8__wasmsimd_dot16x2_ld128 |
| 346 | init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| 347 | k-block: 8 |
| 348 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_3x4c8__wasmsimd_dot16x2_ld128 |
| 349 | init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| 350 | k-block: 8 |
| 351 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_4x4c8__wasmsimd_dot16x2_ld128 |
| 352 | init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| 353 | k-block: 8 |
Marat Dukhan | dfc2db0 | 2021-08-08 21:19:07 -0700 | [diff] [blame] | 354 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x4c8__wasmsimd_mul16_ld64 |
Marat Dukhan | 4741e41 | 2021-06-30 13:38:06 -0700 | [diff] [blame] | 355 | init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| 356 | k-block: 8 |
Marat Dukhan | dfc2db0 | 2021-08-08 21:19:07 -0700 | [diff] [blame] | 357 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x4c8__wasmsimd_mul16_ld64 |
Marat Dukhan | 4741e41 | 2021-06-30 13:38:06 -0700 | [diff] [blame] | 358 | init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| 359 | k-block: 8 |
Marat Dukhan | dfc2db0 | 2021-08-08 21:19:07 -0700 | [diff] [blame] | 360 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_3x4c8__wasmsimd_mul16_ld64 |
Marat Dukhan | 4741e41 | 2021-06-30 13:38:06 -0700 | [diff] [blame] | 361 | init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| 362 | k-block: 8 |
Marat Dukhan | dfc2db0 | 2021-08-08 21:19:07 -0700 | [diff] [blame] | 363 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x4c8__wasmsimd_mul16_ld128 |
Marat Dukhan | 4741e41 | 2021-06-30 13:38:06 -0700 | [diff] [blame] | 364 | init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| 365 | k-block: 8 |
Marat Dukhan | dfc2db0 | 2021-08-08 21:19:07 -0700 | [diff] [blame] | 366 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x4c8__wasmsimd_mul16_ld128 |
Marat Dukhan | 4741e41 | 2021-06-30 13:38:06 -0700 | [diff] [blame] | 367 | init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| 368 | k-block: 8 |
Marat Dukhan | dfc2db0 | 2021-08-08 21:19:07 -0700 | [diff] [blame] | 369 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_3x4c8__wasmsimd_mul16_ld128 |
Marat Dukhan | 4741e41 | 2021-06-30 13:38:06 -0700 | [diff] [blame] | 370 | init: xnn_init_qs8_conv_minmax_fp32_wasmsimd_params |
| 371 | k-block: 8 |
Marat Dukhan | 779b253 | 2021-06-29 14:14:13 -0700 | [diff] [blame] | 372 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_lrint |
| 373 | init: xnn_init_qs8_conv_minmax_fp32_scalar_lrint_params |
| 374 | k-block: 1 |
| 375 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_lrint |
| 376 | init: xnn_init_qs8_conv_minmax_fp32_scalar_lrint_params |
| 377 | k-block: 1 |
| 378 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_lrint |
| 379 | init: xnn_init_qs8_conv_minmax_fp32_scalar_lrint_params |
| 380 | k-block: 1 |
| 381 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_lrint |
| 382 | init: xnn_init_qs8_conv_minmax_fp32_scalar_lrint_params |
| 383 | k-block: 1 |
| 384 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrint |
| 385 | init: xnn_init_qs8_conv_minmax_fp32_scalar_lrint_params |
| 386 | k-block: 1 |
| 387 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_lrint |
| 388 | init: xnn_init_qs8_conv_minmax_fp32_scalar_lrint_params |
| 389 | k-block: 1 |
| 390 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrint |
| 391 | init: xnn_init_qs8_conv_minmax_fp32_scalar_lrint_params |
| 392 | k-block: 1 |
| 393 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_lrint |
| 394 | init: xnn_init_qs8_conv_minmax_fp32_scalar_lrint_params |
| 395 | k-block: 1 |
| 396 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_magic |
| 397 | init: xnn_init_qs8_conv_minmax_fp32_scalar_magic_params |
| 398 | k-block: 1 |
| 399 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_magic |
| 400 | init: xnn_init_qs8_conv_minmax_fp32_scalar_magic_params |
| 401 | k-block: 1 |
| 402 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_3x2__scalar_magic |
| 403 | init: xnn_init_qs8_conv_minmax_fp32_scalar_magic_params |
| 404 | k-block: 1 |
| 405 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_4x2__scalar_magic |
| 406 | init: xnn_init_qs8_conv_minmax_fp32_scalar_magic_params |
| 407 | k-block: 1 |
| 408 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_magic |
| 409 | init: xnn_init_qs8_conv_minmax_fp32_scalar_magic_params |
| 410 | k-block: 1 |
| 411 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_2x4__scalar_magic |
| 412 | init: xnn_init_qs8_conv_minmax_fp32_scalar_magic_params |
| 413 | k-block: 1 |
| 414 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_magic |
| 415 | init: xnn_init_qs8_conv_minmax_fp32_scalar_magic_params |
| 416 | k-block: 1 |
| 417 | - name: xnn_qs8_igemm_minmax_fp32_ukernel_4x4__scalar_magic |
| 418 | init: xnn_init_qs8_conv_minmax_fp32_scalar_magic_params |
| 419 | k-block: 1 |