XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1 | # Copyright 2019 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. |
Frank Barchard | 21be34f | 2019-10-09 19:32:19 -0700 | [diff] [blame] | 5 | - name: xnn_f32_gemminc_ukernel_1x8__aarch64_neonfma_cortex_a53 |
| 6 | k-block: 8 |
| 7 | pipelined: true |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 8 | assembly: true |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 9 | - name: xnn_f32_gemminc_ukernel_1x8__aarch64_neonfma_cortex_a57 |
| 10 | k-block: 8 |
| 11 | pipelined: true |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 12 | assembly: true |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 13 | - name: xnn_f32_gemminc_ukernel_1x8__aarch64_neonfma_cortex_a75 |
| 14 | k-block: 8 |
| 15 | pipelined: true |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 16 | assembly: true |
Frank Barchard | 46fb807 | 2019-10-25 12:54:22 -0700 | [diff] [blame] | 17 | - name: xnn_f32_gemminc_ukernel_4x8__aarch64_neonfma_cortex_a53 |
Frank Barchard | 0ecc2ab | 2019-11-14 10:57:48 -0800 | [diff] [blame] | 18 | k-block: 4 |
| 19 | pipelined: true |
Frank Barchard | 46fb807 | 2019-10-25 12:54:22 -0700 | [diff] [blame] | 20 | assembly: true |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 21 | - name: xnn_f32_gemminc_ukernel_4x8__aarch64_neonfma_cortex_a57 |
| 22 | k-block: 8 |
| 23 | pipelined: true |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 24 | assembly: true |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 25 | - name: xnn_f32_gemminc_ukernel_4x8__aarch64_neonfma_cortex_a75 |
| 26 | k-block: 8 |
| 27 | pipelined: true |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 28 | assembly: true |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 29 | - name: xnn_f32_gemminc_ukernel_5x8__aarch64_neonfma_cortex_a75 |
| 30 | k-block: 8 |
| 31 | pipelined: true |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 32 | assembly: true |
Frank Barchard | a7fb855 | 2019-10-23 17:14:17 -0700 | [diff] [blame] | 33 | - name: xnn_f32_gemminc_ukernel_6x8__aarch64_neonfma_cortex_a53 |
Frank Barchard | 00bf68e | 2019-10-27 03:00:09 -0700 | [diff] [blame] | 34 | k-block: 4 |
Frank Barchard | e64f91a | 2019-11-11 13:18:00 -0800 | [diff] [blame] | 35 | pipelined: true |
Frank Barchard | a7fb855 | 2019-10-23 17:14:17 -0700 | [diff] [blame] | 36 | assembly: true |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 37 | - name: xnn_f32_gemminc_ukernel_6x8__aarch64_neonfma_cortex_a57 |
Frank Barchard | bd41971 | 2019-10-31 14:15:36 -0700 | [diff] [blame] | 38 | k-block: 4 |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 39 | assembly: true |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 40 | - name: xnn_f32_gemminc_ukernel_6x8__aarch64_neonfma_cortex_a73 |
| 41 | k-block: 8 |
| 42 | pipelined: true |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 43 | assembly: true |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 44 | - name: xnn_f32_gemminc_ukernel_6x8__aarch64_neonfma_cortex_a75 |
| 45 | k-block: 8 |
| 46 | pipelined: true |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 47 | assembly: true |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 48 | - name: xnn_f32_gemminc_ukernel_1x12__aarch64_neonfma_cortex_a53 |
| 49 | k-block: 4 |
| 50 | pipelined: true |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 51 | assembly: true |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 52 | - name: xnn_f32_gemminc_ukernel_4x12__aarch64_neonfma_cortex_a53 |
| 53 | k-block: 4 |
| 54 | pipelined: true |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 55 | assembly: true |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 56 | - name: xnn_f32_gemminc_ukernel_4x8__aarch64_neonfma_ld64 |
| 57 | k-block: 2 |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 58 | assembly: true |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 59 | - name: xnn_f32_gemminc_ukernel_4x8__aarch64_neonfma_ld128 |
| 60 | k-block: 4 |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 61 | assembly: true |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 62 | - name: xnn_f32_gemminc_ukernel_6x8__aarch64_neonfma_ld64 |
| 63 | k-block: 2 |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 64 | assembly: true |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 65 | - name: xnn_f32_gemminc_ukernel_6x8__aarch64_neonfma_ld128 |
| 66 | k-block: 4 |
Frank Barchard | 7e95597 | 2019-10-11 10:34:25 -0700 | [diff] [blame] | 67 | assembly: true |
Frank Barchard | 91317c5 | 2019-11-22 10:54:35 -0800 | [diff] [blame] | 68 | - name: xnn_f32_gemminc_ukernel_1x8__neon_lane_ld64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 69 | k-block: 2 |
Frank Barchard | 91317c5 | 2019-11-22 10:54:35 -0800 | [diff] [blame] | 70 | - name: xnn_f32_gemminc_ukernel_4x8__neon_lane_ld64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 71 | k-block: 2 |
Frank Barchard | 91317c5 | 2019-11-22 10:54:35 -0800 | [diff] [blame] | 72 | - name: xnn_f32_gemminc_ukernel_4x8__neon_lane_ld128 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 73 | k-block: 4 |
Frank Barchard | 91317c5 | 2019-11-22 10:54:35 -0800 | [diff] [blame] | 74 | - name: xnn_f32_gemminc_ukernel_5x8__neon_lane_ld64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 75 | k-block: 2 |
Frank Barchard | 91317c5 | 2019-11-22 10:54:35 -0800 | [diff] [blame] | 76 | - name: xnn_f32_gemminc_ukernel_6x8__neon_lane_ld64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 77 | k-block: 2 |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 78 | - name: xnn_f32_gemminc_ukernel_6x8__neon_lane_ld128 |
| 79 | k-block: 4 |
Frank Barchard | 91317c5 | 2019-11-22 10:54:35 -0800 | [diff] [blame] | 80 | - name: xnn_f32_gemminc_ukernel_1x8__neonfma_lane_ld64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 81 | k-block: 2 |
Frank Barchard | 91317c5 | 2019-11-22 10:54:35 -0800 | [diff] [blame] | 82 | arch: |
| 83 | - aarch64 |
| 84 | - name: xnn_f32_gemminc_ukernel_4x8__neonfma_lane_ld64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 85 | k-block: 2 |
Frank Barchard | 91317c5 | 2019-11-22 10:54:35 -0800 | [diff] [blame] | 86 | arch: |
| 87 | - aarch64 |
| 88 | - name: xnn_f32_gemminc_ukernel_4x8__neonfma_lane_ld128 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 89 | k-block: 4 |
Frank Barchard | 91317c5 | 2019-11-22 10:54:35 -0800 | [diff] [blame] | 90 | arch: |
| 91 | - aarch64 |
| 92 | - name: xnn_f32_gemminc_ukernel_5x8__neonfma_lane_ld64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 93 | k-block: 2 |
Frank Barchard | 91317c5 | 2019-11-22 10:54:35 -0800 | [diff] [blame] | 94 | arch: |
| 95 | - aarch64 |
| 96 | - name: xnn_f32_gemminc_ukernel_6x8__neonfma_lane_ld64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 97 | k-block: 2 |
Frank Barchard | 91317c5 | 2019-11-22 10:54:35 -0800 | [diff] [blame] | 98 | arch: |
| 99 | - aarch64 |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 100 | - name: xnn_f32_gemminc_ukernel_6x8__neonfma_lane_ld128 |
| 101 | k-block: 4 |
| 102 | arch: |
| 103 | - aarch64 |
Frank Barchard | 5243bb0 | 2019-11-22 16:37:50 -0800 | [diff] [blame] | 104 | - name: xnn_f32_gemminc_ukernel_1x8__neon_dup_ld64 |
| 105 | k-block: 2 |
| 106 | - name: xnn_f32_gemminc_ukernel_4x8__neon_dup_ld64 |
| 107 | k-block: 2 |
| 108 | - name: xnn_f32_gemminc_ukernel_4x8__neon_dup_ld128 |
| 109 | k-block: 4 |
| 110 | - name: xnn_f32_gemminc_ukernel_6x8__neon_dup_ld64 |
| 111 | k-block: 2 |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 112 | - name: xnn_f32_gemminc_ukernel_6x8__neon_dup_ld128 |
| 113 | k-block: 4 |
Frank Barchard | 5243bb0 | 2019-11-22 16:37:50 -0800 | [diff] [blame] | 114 | - name: xnn_f32_gemminc_ukernel_1x8__neonfma_dup_ld64 |
| 115 | k-block: 2 |
| 116 | - name: xnn_f32_gemminc_ukernel_4x8__neonfma_dup_ld64 |
| 117 | k-block: 2 |
| 118 | - name: xnn_f32_gemminc_ukernel_4x8__neonfma_dup_ld128 |
| 119 | k-block: 4 |
| 120 | - name: xnn_f32_gemminc_ukernel_6x8__neonfma_dup_ld64 |
| 121 | k-block: 2 |
Frank Barchard | 69172d9 | 2019-11-26 16:22:39 -0800 | [diff] [blame] | 122 | - name: xnn_f32_gemminc_ukernel_6x8__neonfma_dup_ld128 |
| 123 | k-block: 4 |
Frank Barchard | 5243bb0 | 2019-11-22 16:37:50 -0800 | [diff] [blame] | 124 | - name: xnn_f32_gemminc_ukernel_1x8s4__neon |
| 125 | k-block: 4 |
| 126 | - name: xnn_f32_gemminc_ukernel_4x8s4__neon |
| 127 | k-block: 4 |
| 128 | - name: xnn_f32_gemminc_ukernel_6x8s4__neon |
| 129 | k-block: 4 |
| 130 | - name: xnn_f32_gemminc_ukernel_8x8s4__neon |
| 131 | k-block: 4 |
Frank Barchard | df06d80 | 2019-11-20 15:53:46 -0800 | [diff] [blame] | 132 | - name: xnn_f32_gemminc_ukernel_1x8s4__neonfma |
| 133 | k-block: 4 |
| 134 | - name: xnn_f32_gemminc_ukernel_4x8s4__neonfma |
| 135 | k-block: 4 |
| 136 | - name: xnn_f32_gemminc_ukernel_6x8s4__neonfma |
| 137 | k-block: 4 |
| 138 | - name: xnn_f32_gemminc_ukernel_8x8s4__neonfma |
| 139 | k-block: 4 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 140 | - name: xnn_f32_gemminc_ukernel_1x8__sse_load1 |
| 141 | k-block: 1 |
| 142 | - name: xnn_f32_gemminc_ukernel_4x8__sse_load1 |
| 143 | k-block: 1 |
| 144 | - name: xnn_f32_gemminc_ukernel_1x8__sse_dup |
| 145 | k-block: 4 |
| 146 | - name: xnn_f32_gemminc_ukernel_4x8__sse_dup |
| 147 | k-block: 4 |
| 148 | - name: xnn_f32_gemminc_ukernel_1x8s4__sse |
| 149 | k-block: 4 |
| 150 | - name: xnn_f32_gemminc_ukernel_4x8s4__sse |
| 151 | k-block: 4 |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 152 | - name: xnn_f32_gemminc_ukernel_1x8__avx_broadcast |
| 153 | k-block: 1 |
| 154 | - name: xnn_f32_gemminc_ukernel_4x8__avx_broadcast |
| 155 | k-block: 1 |
| 156 | - name: xnn_f32_gemminc_ukernel_5x8__avx_broadcast |
| 157 | k-block: 1 |
| 158 | - name: xnn_f32_gemminc_ukernel_6x8__avx_broadcast |
| 159 | k-block: 1 |
| 160 | - name: xnn_f32_gemminc_ukernel_7x8__avx_broadcast |
| 161 | k-block: 1 |
| 162 | - name: xnn_f32_gemminc_ukernel_1x8__fma3_broadcast |
| 163 | k-block: 1 |
| 164 | - name: xnn_f32_gemminc_ukernel_4x8__fma3_broadcast |
| 165 | k-block: 1 |
| 166 | - name: xnn_f32_gemminc_ukernel_5x8__fma3_broadcast |
| 167 | k-block: 1 |
| 168 | - name: xnn_f32_gemminc_ukernel_6x8__fma3_broadcast |
| 169 | k-block: 1 |
| 170 | - name: xnn_f32_gemminc_ukernel_7x8__fma3_broadcast |
| 171 | k-block: 1 |
| 172 | - name: xnn_f32_gemminc_ukernel_8x8__fma3_broadcast |
| 173 | k-block: 1 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 174 | - name: xnn_f32_gemminc_ukernel_1x8__psimd_loadsplat |
| 175 | k-block: 1 |
| 176 | - name: xnn_f32_gemminc_ukernel_4x8__psimd_loadsplat |
| 177 | k-block: 1 |
| 178 | - name: xnn_f32_gemminc_ukernel_6x8__psimd_loadsplat |
| 179 | k-block: 1 |
| 180 | - name: xnn_f32_gemminc_ukernel_1x8__psimd_splat |
| 181 | k-block: 4 |
| 182 | - name: xnn_f32_gemminc_ukernel_4x8__psimd_splat |
| 183 | k-block: 4 |
| 184 | - name: xnn_f32_gemminc_ukernel_6x8__psimd_splat |
| 185 | k-block: 4 |
| 186 | - name: xnn_f32_gemminc_ukernel_1x8s4__psimd |
| 187 | k-block: 4 |
| 188 | - name: xnn_f32_gemminc_ukernel_4x8s4__psimd |
| 189 | k-block: 4 |
| 190 | - name: xnn_f32_gemminc_ukernel_6x8s4__psimd |
| 191 | k-block: 4 |
| 192 | - name: xnn_f32_gemminc_ukernel_1x4__scalar |
| 193 | k-block: 1 |
| 194 | - name: xnn_f32_gemminc_ukernel_2x4__scalar |
| 195 | k-block: 1 |
| 196 | - name: xnn_f32_gemminc_ukernel_4x4__scalar |
| 197 | k-block: 1 |