XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1 | // Copyright (c) Facebook, Inc. and its affiliates. |
| 2 | // All rights reserved. |
| 3 | // |
| 4 | // Copyright 2019 Google LLC |
| 5 | // |
| 6 | // This source code is licensed under the BSD-style license found in the |
| 7 | // LICENSE file in the root directory of this source tree. |
| 8 | |
| 9 | #include <stdbool.h> |
| 10 | #include <stddef.h> |
| 11 | #include <stdint.h> |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 12 | #include <string.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 13 | |
| 14 | #include <pthread.h> |
| 15 | |
Marat Dukhan | d343c22 | 2019-10-07 09:22:14 -0700 | [diff] [blame] | 16 | #ifndef __EMSCRIPTEN__ |
| 17 | #include <cpuinfo.h> |
| 18 | #endif |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 19 | |
| 20 | #include <xnnpack.h> |
| 21 | #include <xnnpack/argmaxpool.h> |
| 22 | #include <xnnpack/avgpool.h> |
Marat Dukhan | 6972249 | 2019-11-11 19:55:50 -0800 | [diff] [blame] | 23 | #include <xnnpack/bilinear.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 24 | #include <xnnpack/clamp.h> |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 25 | #include <xnnpack/common.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 26 | #include <xnnpack/conv.h> |
| 27 | #include <xnnpack/dwconv.h> |
| 28 | #include <xnnpack/gavgpool.h> |
| 29 | #include <xnnpack/gemm.h> |
| 30 | #include <xnnpack/hswish.h> |
| 31 | #include <xnnpack/igemm.h> |
| 32 | #include <xnnpack/log.h> |
| 33 | #include <xnnpack/lut.h> |
| 34 | #include <xnnpack/maxpool.h> |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 35 | #include <xnnpack/memory.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 36 | #include <xnnpack/pad.h> |
| 37 | #include <xnnpack/params.h> |
| 38 | #include <xnnpack/pavgpool.h> |
| 39 | #include <xnnpack/prelu.h> |
| 40 | #include <xnnpack/rmax.h> |
| 41 | #include <xnnpack/spmm.h> |
| 42 | #include <xnnpack/unpool.h> |
| 43 | #include <xnnpack/vadd.h> |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 44 | #include <xnnpack/vbinary.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 45 | #include <xnnpack/vmulcaddc.h> |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 46 | #include <xnnpack/vunary.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 47 | #include <xnnpack/zip.h> |
| 48 | |
| 49 | #ifndef XNN_ENABLE_ASSEMBLY |
| 50 | #define XNN_ENABLE_ASSEMBLY 1 |
| 51 | #endif |
| 52 | |
| 53 | static pthread_once_t init_guard = PTHREAD_ONCE_INIT; |
| 54 | |
| 55 | struct xnn_parameters xnn_params = { |
| 56 | .initialized = false |
| 57 | }; |
| 58 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 59 | #if XNN_ARCH_PNACL || XNN_ARCH_ASMJS || XNN_ARCH_WASM || XNN_ARCH_WASMSIMD |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 60 | extern uint32_t xnn_stub_wasm_f32_sub(uint32_t a, uint32_t b); |
| 61 | #endif |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 62 | #if XNN_ARCH_PNACL || XNN_ARCH_WASM || XNN_ARCH_WASMSIMD |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 63 | extern uint32_t xnn_stub_wasm_f32_min(uint32_t a, uint32_t b); |
| 64 | #endif |
| 65 | |
| 66 | static void init(void) { |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 67 | #if XNN_ARCH_ARM |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 68 | if (!cpuinfo_has_arm_neon()) { |
| 69 | xnn_log_error("XNNPACK initialization failed: NEON is not supported"); |
| 70 | return; |
| 71 | } |
| 72 | |
| 73 | /**************************** Q8 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 74 | #ifndef XNN_NO_Q8_OPERATORS |
| 75 | xnn_params.q8.gemm = (struct gemm_parameters) { |
| 76 | .gemm = (xnn_gemm_ukernel_function) xnn_q8_gemm_ukernel_4x8__neon, |
| 77 | .igemm = (xnn_igemm_ukernel_function) xnn_q8_igemm_ukernel_4x8__neon, |
| 78 | .mr = 4, |
| 79 | .nr = 8, |
| 80 | }; |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 81 | |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 82 | #if XNN_ENABLE_ASSEMBLY |
| 83 | xnn_params.q8.dwconv[0] = (struct dwconv_parameters) { |
| 84 | .up = (xnn_dwconv_up_ukernel_function) xnn_q8_dwconv_ukernel_up8x9__aarch32_neon, |
| 85 | .cr = 8, |
| 86 | .mr = 9, |
| 87 | }; |
| 88 | #else |
| 89 | xnn_params.q8.dwconv[0] = (struct dwconv_parameters) { |
| 90 | .up = (xnn_dwconv_up_ukernel_function) xnn_q8_dwconv_ukernel_up8x9__neon, |
| 91 | .cr = 8, |
| 92 | .mr = 9, |
| 93 | }; |
| 94 | #endif |
| 95 | xnn_params.q8.avgpool = (struct avgpool_parameters) { |
| 96 | .up = (xnn_avgpool_up_ukernel_function) xnn_q8_avgpool_ukernel_up9__neon, |
| 97 | .mp = (xnn_avgpool_mp_ukernel_function) xnn_q8_avgpool_ukernel_mp9p8q__neon, |
| 98 | .mr = 9, |
| 99 | .qr = 8, |
| 100 | }; |
| 101 | xnn_params.q8.gavgpool = (struct gavgpool_parameters) { |
| 102 | .up = (xnn_gavgpool_up_ukernel_function) xnn_q8_gavgpool_ukernel_up7__neon, |
| 103 | .mp = (xnn_gavgpool_mp_ukernel_function) xnn_q8_gavgpool_ukernel_mp7p7q__neon, |
| 104 | .mr = 7, |
| 105 | }; |
| 106 | xnn_params.q8.vadd = (xnn_vadd_ukernel_function) xnn_q8_vadd_ukernel__neon; |
| 107 | #endif // XNN_NO_Q8_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 108 | |
| 109 | /**************************** U8 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 110 | #ifndef XNN_NO_U8_OPERATORS |
| 111 | xnn_params.u8.maxpool = (struct maxpool_parameters) { |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 112 | .ukernel = (xnn_maxpool_ukernel_function) xnn_u8_maxpool_ukernel_9p8x__neon_c16, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 113 | .mr = 9, |
| 114 | .qr = 8, |
| 115 | }; |
| 116 | xnn_params.u8.clamp = (xnn_univector_ukernel_function) xnn_u8_clamp_ukernel__neon; |
| 117 | xnn_params.u8.rmax = xnn_u8_rmax_ukernel__neon; |
| 118 | xnn_params.u8.lut32norm = xnn_u8_lut32norm_ukernel__scalar; |
| 119 | #endif // XNN_NO_U8_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 120 | |
| 121 | /**************************** X8 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 122 | #ifndef XNN_NO_X8_OPERATORS |
| 123 | xnn_params.x8.lut = xnn_x8_lut_ukernel__scalar; |
| 124 | xnn_params.x8.zip = (struct zip_parameters) { |
| 125 | .x2 = (xnn_zipc_ukernel_function) xnn_x8_zip_x2_ukernel__neon, |
| 126 | .x3 = (xnn_zipc_ukernel_function) xnn_x8_zip_x3_ukernel__neon, |
| 127 | .x4 = (xnn_zipc_ukernel_function) xnn_x8_zip_x4_ukernel__neon, |
| 128 | .xm = (xnn_zipv_ukernel_function) xnn_x8_zip_xm_ukernel__neon, |
| 129 | }; |
| 130 | #endif // XNN_NO_X8_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 131 | |
| 132 | /**************************** F32 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 133 | #ifndef XNN_NO_F32_OPERATORS |
Frank Barchard | 3267092 | 2019-11-30 21:58:51 -0800 | [diff] [blame] | 134 | #if XNN_ENABLE_ASSEMBLY |
| 135 | xnn_params.f32.gemm = (struct gemm_parameters) { |
| 136 | .gemm = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_4x8__aarch32_neon_ld64, |
| 137 | .igemm = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_4x8__neon_lane_ld128, |
| 138 | .gemm1 = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_1x8__neon_lane_ld64, |
| 139 | .igemm1 = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_1x8__neon_lane_ld64, |
| 140 | .mr = 4, |
| 141 | .nr = 8, |
| 142 | }; |
| 143 | #else // XNN_ENABLE_ASSEMBLY |
| 144 | xnn_params.f32.gemm = (struct gemm_parameters) { |
| 145 | .gemm = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_4x8__neon_lane_ld128, |
| 146 | .igemm = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_4x8__neon_lane_ld128, |
| 147 | .gemm1 = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_1x8__neon_lane_ld64, |
| 148 | .igemm1 = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_1x8__neon_lane_ld64, |
| 149 | .mr = 4, |
| 150 | .nr = 8, |
| 151 | }; |
| 152 | #endif // XNN_ENABLE_ASSEMBLY |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 153 | xnn_params.f32.gemm2 = (struct gemm_parameters) { |
| 154 | .gemm = NULL, |
Frank Barchard | 91317c5 | 2019-11-22 10:54:35 -0800 | [diff] [blame] | 155 | .igemm = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_4x2__neon_lane_ld64, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 156 | .mr = 4, |
| 157 | .nr = 2, |
| 158 | }; |
| 159 | xnn_params.f32.dwconv[0] = (struct dwconv_parameters) { |
| 160 | .up = (xnn_dwconv_up_ukernel_function) xnn_f32_dwconv_ukernel_up4x4__psimd, |
| 161 | .cr = 4, |
| 162 | .mr = 4, |
| 163 | }; |
| 164 | xnn_params.f32.dwconv[1] = (struct dwconv_parameters) { |
| 165 | .up = (xnn_dwconv_up_ukernel_function) xnn_f32_dwconv_ukernel_up4x9__neon, |
| 166 | .cr = 4, |
| 167 | .mr = 9, |
| 168 | }; |
| 169 | xnn_params.f32.dwconv[2] = (struct dwconv_parameters) { |
| 170 | .up = (xnn_dwconv_up_ukernel_function) xnn_f32_dwconv_ukernel_up4x25__psimd, |
| 171 | .cr = 4, |
| 172 | .mr = 25, |
| 173 | }; |
| 174 | xnn_params.f32.avgpool = (struct avgpool_parameters) { |
| 175 | .up = (xnn_avgpool_up_ukernel_function) xnn_f32_avgpool_ukernel_up9__neon, |
| 176 | .mp = (xnn_avgpool_mp_ukernel_function) xnn_f32_avgpool_ukernel_mp9p8q__neon, |
| 177 | .mr = 9, |
| 178 | .qr = 8, |
| 179 | }; |
| 180 | xnn_params.f32.pavgpool = (struct pavgpool_parameters) { |
| 181 | .up = (xnn_pavgpool_up_ukernel_function) xnn_f32_pavgpool_ukernel_up9__neon, |
| 182 | .mp = (xnn_pavgpool_mp_ukernel_function) xnn_f32_pavgpool_ukernel_mp9p8q__neon, |
| 183 | .mr = 9, |
| 184 | .qr = 8, |
| 185 | }; |
| 186 | xnn_params.f32.gavgpool = (struct gavgpool_parameters) { |
| 187 | .up = (xnn_gavgpool_up_ukernel_function) xnn_f32_gavgpool_ukernel_up7__neon, |
| 188 | .mp = (xnn_gavgpool_mp_ukernel_function) xnn_f32_gavgpool_ukernel_mp7p7q__neon, |
| 189 | .mr = 7, |
| 190 | }; |
| 191 | xnn_params.f32.maxpool = (struct maxpool_parameters) { |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 192 | .ukernel = (xnn_maxpool_ukernel_function) xnn_f32_maxpool_ukernel_9p8x__psimd_c4, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 193 | .mr = 9, |
| 194 | .qr = 8, |
| 195 | }; |
| 196 | xnn_params.f32.argmaxpool[0] = (struct argmaxpool_parameters) { |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 197 | .up = (xnn_argmaxpool_up_ukernel_function) xnn_f32_argmaxpool_ukernel_4x__psimd_c4, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 198 | .mr = 4, |
| 199 | }; |
| 200 | xnn_params.f32.argmaxpool[1] = (struct argmaxpool_parameters) { |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 201 | .up = (xnn_argmaxpool_up_ukernel_function) xnn_f32_argmaxpool_ukernel_9x__psimd_c4, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 202 | .mr = 9, |
| 203 | }; |
| 204 | xnn_params.f32.argmaxpool[2] = (struct argmaxpool_parameters) { |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 205 | .mp = (xnn_argmaxpool_mp_ukernel_function) xnn_f32_argmaxpool_ukernel_9p8x__psimd_c4, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 206 | .mr = 9, |
| 207 | .qr = 8, |
| 208 | }; |
Marat Dukhan | 6972249 | 2019-11-11 19:55:50 -0800 | [diff] [blame] | 209 | xnn_params.f32.bilinear = (struct bilinear_parameters) { |
| 210 | .ukernel = (xnn_bilinear_ukernel_function) xnn_f32_bilinear_ukernel__neon_c8, |
| 211 | .pixel_tile = 1, |
| 212 | .channel_tile = 8, |
| 213 | }; |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 214 | xnn_params.f32.clamp = (xnn_univector_ukernel_function) xnn_f32_clamp_ukernel__neon; |
| 215 | xnn_params.f32.hswish = (xnn_univector_ukernel_function) xnn_f32_hswish_ukernel__neon; |
| 216 | xnn_params.f32.prelu = (struct prelu_parameters) { |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 217 | .ukernel = (xnn_prelu_ukernel_function) xnn_f32_prelu_ukernel__neon_2x8, |
| 218 | .row_tile = 2, |
| 219 | .channel_tile = 8, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 220 | }; |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 221 | xnn_params.f32.vadd = (struct vbinary_parameters) { |
| 222 | .op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vadd_ukernel__neon_x8, |
| 223 | .opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_ukernel__neon_x8, |
| 224 | .ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_ukernel__neon_x8, |
| 225 | .element_tile = 8, |
| 226 | }; |
Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 227 | xnn_params.f32.vmax = (struct vbinary_parameters) { |
| 228 | .op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmax_ukernel__neon_x8, |
| 229 | .opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__neon_x8, |
| 230 | .ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__neon_x8, |
| 231 | .element_tile = 8, |
| 232 | }; |
| 233 | xnn_params.f32.vmin = (struct vbinary_parameters) { |
| 234 | .op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmin_ukernel__neon_x8, |
| 235 | .opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__neon_x8, |
| 236 | .ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__neon_x8, |
| 237 | .element_tile = 8, |
| 238 | }; |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 239 | xnn_params.f32.vmul = (struct vbinary_parameters) { |
| 240 | .op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmul_ukernel__neon_x8, |
| 241 | .opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_ukernel__neon_x8, |
| 242 | .ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_ukernel__neon_x8, |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 243 | .element_tile = 8, |
| 244 | }; |
Marat Dukhan | 05f3f6d | 2019-12-03 15:13:53 -0800 | [diff] [blame] | 245 | xnn_params.f32.vsub = (struct vbinary_parameters) { |
| 246 | .op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsub_ukernel__neon_x8, |
| 247 | .opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsubc_ukernel__neon_x8, |
| 248 | .ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrsubc_ukernel__neon_x8, |
| 249 | .element_tile = 8, |
| 250 | }; |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 251 | xnn_params.f32.vmulcaddc = (struct vmulcaddc_parameters) { |
Marat Dukhan | 49e6ee9 | 2019-11-06 15:55:29 -0800 | [diff] [blame] | 252 | .ukernel = (xnn_vmulcaddc_ukernel_function) xnn_f32_vmulcaddc_ukernel_c4__neon_2x, |
| 253 | .channel_tile = 4, |
| 254 | .row_tile = 2, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 255 | }; |
| 256 | #endif // XNN_NO_F32_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 257 | |
| 258 | /**************************** X32 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 259 | #ifndef XNN_NO_X32_OPERATORS |
| 260 | xnn_params.x32.pad = (struct pad_parameters) { |
| 261 | .ukernel = xnn_x32_pad_x2__neon, |
| 262 | .mr = 2, |
| 263 | }; |
| 264 | xnn_params.x32.unpool = (xnn_unpool_ukernel_function) xnn_x32_unpool_ukernel__psimd; |
| 265 | xnn_params.x32.zip = (struct zip_parameters) { |
| 266 | .x2 = (xnn_zipc_ukernel_function) xnn_x32_zip_x2_ukernel__neon, |
| 267 | .x3 = (xnn_zipc_ukernel_function) xnn_x32_zip_x3_ukernel__neon, |
| 268 | .x4 = (xnn_zipc_ukernel_function) xnn_x32_zip_x4_ukernel__neon, |
| 269 | .xm = (xnn_zipv_ukernel_function) xnn_x32_zip_xm_ukernel__neon, |
| 270 | }; |
| 271 | #endif // XNN_NO_X32_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 272 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 273 | #elif XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 274 | |
| 275 | /**************************** Q8 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 276 | #ifndef XNN_NO_Q8_OPERATORS |
| 277 | xnn_params.q8.gemm = (struct gemm_parameters) { |
| 278 | .gemm = (xnn_gemm_ukernel_function) xnn_q8_gemm_ukernel_8x8__neon, |
| 279 | .igemm = (xnn_igemm_ukernel_function) xnn_q8_igemm_ukernel_8x8__neon, |
| 280 | .mr = 8, |
| 281 | .nr = 8, |
| 282 | }; |
| 283 | xnn_params.q8.dwconv[0] = (struct dwconv_parameters) { |
| 284 | .up = (xnn_dwconv_up_ukernel_function) xnn_q8_dwconv_ukernel_up8x9__neon, |
| 285 | .cr = 8, |
| 286 | .mr = 9, |
| 287 | }; |
| 288 | xnn_params.q8.avgpool = (struct avgpool_parameters) { |
| 289 | .up = (xnn_avgpool_up_ukernel_function) xnn_q8_avgpool_ukernel_up9__neon, |
| 290 | .mp = (xnn_avgpool_mp_ukernel_function) xnn_q8_avgpool_ukernel_mp9p8q__neon, |
| 291 | .mr = 9, |
| 292 | .qr = 8, |
| 293 | }; |
| 294 | xnn_params.q8.gavgpool = (struct gavgpool_parameters) { |
| 295 | .up = (xnn_gavgpool_up_ukernel_function) xnn_q8_gavgpool_ukernel_up7__neon, |
| 296 | .mp = (xnn_gavgpool_mp_ukernel_function) xnn_q8_gavgpool_ukernel_mp7p7q__neon, |
| 297 | .mr = 7, |
| 298 | }; |
| 299 | xnn_params.q8.vadd = (xnn_vadd_ukernel_function) xnn_q8_vadd_ukernel__neon; |
| 300 | #endif // XNN_NO_Q8_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 301 | |
| 302 | /**************************** U8 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 303 | #ifndef XNN_NO_U8_OPERATORS |
| 304 | xnn_params.u8.maxpool = (struct maxpool_parameters) { |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 305 | .ukernel = (xnn_maxpool_ukernel_function) xnn_u8_maxpool_ukernel_9p8x__neon_c16, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 306 | .mr = 9, |
| 307 | .qr = 8, |
| 308 | }; |
| 309 | xnn_params.u8.clamp = (xnn_univector_ukernel_function) xnn_u8_clamp_ukernel__neon; |
| 310 | xnn_params.u8.lut32norm = xnn_u8_lut32norm_ukernel__scalar; |
| 311 | xnn_params.u8.rmax = xnn_u8_rmax_ukernel__neon; |
| 312 | #endif // XNN_NO_U8_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 313 | |
| 314 | /**************************** X8 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 315 | #ifndef XNN_NO_X8_OPERATORS |
| 316 | xnn_params.x8.lut = xnn_x8_lut_ukernel__scalar; |
| 317 | xnn_params.x8.zip = (struct zip_parameters) { |
| 318 | .x2 = (xnn_zipc_ukernel_function) xnn_x8_zip_x2_ukernel__neon, |
| 319 | .x3 = (xnn_zipc_ukernel_function) xnn_x8_zip_x3_ukernel__neon, |
| 320 | .x4 = (xnn_zipc_ukernel_function) xnn_x8_zip_x4_ukernel__neon, |
| 321 | .xm = (xnn_zipv_ukernel_function) xnn_x8_zip_xm_ukernel__neon, |
| 322 | }; |
| 323 | #endif // XNN_NO_X8_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 324 | |
| 325 | /**************************** F32 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 326 | #ifndef XNN_NO_F32_OPERATORS |
| 327 | #if XNN_ENABLE_ASSEMBLY |
| 328 | switch (cpuinfo_get_core(0)->uarch) { |
| 329 | case cpuinfo_uarch_kryo: |
| 330 | xnn_params.f32.gemm = (struct gemm_parameters) { |
| 331 | .gemm = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_4x8__aarch64_neonfma_cortex_a57, |
| 332 | .igemm = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_4x8__aarch64_neonfma_cortex_a75, |
| 333 | .gemm1 = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_1x8__aarch64_neonfma_cortex_a75, |
| 334 | .igemm1 = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_1x8__aarch64_neonfma_cortex_a75, |
| 335 | .mr = 4, |
| 336 | .nr = 8, |
| 337 | }; |
| 338 | break; |
| 339 | case cpuinfo_uarch_cortex_a57: |
| 340 | xnn_params.f32.gemm = (struct gemm_parameters) { |
| 341 | .gemm = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_6x8__aarch64_neonfma_cortex_a57, |
| 342 | .igemm = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_6x8__aarch64_neonfma_cortex_a57, |
| 343 | .gemm1 = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_1x8__aarch64_neonfma_cortex_a57, |
| 344 | .igemm1 = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_1x8__aarch64_neonfma_cortex_a57, |
| 345 | .mr = 6, |
| 346 | .nr = 8, |
| 347 | }; |
| 348 | break; |
| 349 | case cpuinfo_uarch_cortex_a72: |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 350 | xnn_params.f32.gemm = (struct gemm_parameters) { |
| 351 | .gemm = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_4x8__aarch64_neonfma_cortex_a75, |
| 352 | .igemm = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_4x8__aarch64_neonfma_cortex_a75, |
| 353 | .gemm1 = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_1x8__aarch64_neonfma_cortex_a75, |
| 354 | .igemm1 = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_1x8__aarch64_neonfma_cortex_a75, |
| 355 | .mr = 4, |
| 356 | .nr = 8, |
| 357 | }; |
| 358 | break; |
| 359 | case cpuinfo_uarch_cortex_a75: |
Frank Barchard | 263bb09 | 2019-10-28 15:28:46 -0700 | [diff] [blame] | 360 | case cpuinfo_uarch_cortex_a76: |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 361 | case cpuinfo_uarch_meerkat_m3: |
| 362 | case (cpuinfo_uarch_meerkat_m3 + 1): |
| 363 | xnn_params.f32.gemm = (struct gemm_parameters) { |
| 364 | .gemm = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_6x8__aarch64_neonfma_cortex_a75, |
| 365 | .igemm = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_6x8__aarch64_neonfma_cortex_a75, |
| 366 | .gemm1 = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_1x8__aarch64_neonfma_cortex_a75, |
| 367 | .igemm1 = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_1x8__aarch64_neonfma_cortex_a75, |
| 368 | .mr = 6, |
| 369 | .nr = 8, |
| 370 | }; |
| 371 | break; |
Frank Barchard | df06d80 | 2019-11-20 15:53:46 -0800 | [diff] [blame] | 372 | |
| 373 | case cpuinfo_uarch_mongoose_m1: |
| 374 | case cpuinfo_uarch_mongoose_m2: |
| 375 | xnn_params.f32.gemm = (struct gemm_parameters) { |
| 376 | .gemm = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_6x8s4__neonfma, |
| 377 | .igemm = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_6x8s4__neonfma, |
| 378 | .gemm1 = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_1x8s4__neonfma, |
| 379 | .igemm1 = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_1x8s4__neonfma, |
| 380 | .mr = 6, |
| 381 | .nr = 8, |
| 382 | .log2_sr = 2, |
| 383 | }; |
| 384 | break; |
| 385 | |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 386 | case cpuinfo_uarch_cortex_a53: |
| 387 | case cpuinfo_uarch_cortex_a55: |
| 388 | xnn_params.f32.gemm = (struct gemm_parameters) { |
Frank Barchard | bd1d5d9 | 2019-10-30 15:53:30 -0700 | [diff] [blame] | 389 | .gemm = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_6x8__aarch64_neonfma_cortex_a53, |
| 390 | .igemm = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_6x8__aarch64_neonfma_cortex_a53, |
| 391 | .gemm1 = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_1x8__aarch64_neonfma_cortex_a53, |
| 392 | .igemm1 = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_1x8__aarch64_neonfma_cortex_a53, |
| 393 | .mr = 6, |
| 394 | .nr = 8, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 395 | }; |
| 396 | break; |
| 397 | case cpuinfo_uarch_cortex_a73: |
| 398 | xnn_params.f32.gemm = (struct gemm_parameters) { |
| 399 | .gemm = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_6x8__aarch64_neonfma_cortex_a73, |
| 400 | .igemm = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_6x8__aarch64_neonfma_cortex_a73, |
| 401 | .gemm1 = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_1x8__aarch64_neonfma_cortex_a75, |
| 402 | .igemm1 = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_1x8__aarch64_neonfma_cortex_a75, |
| 403 | .mr = 6, |
| 404 | .nr = 8, |
| 405 | }; |
| 406 | break; |
| 407 | default: |
| 408 | xnn_params.f32.gemm = (struct gemm_parameters) { |
Frank Barchard | 91317c5 | 2019-11-22 10:54:35 -0800 | [diff] [blame] | 409 | .gemm = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_6x8__neonfma_lane_ld64, |
| 410 | .igemm = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_6x8__neonfma_lane_ld64, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 411 | .gemm1 = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_1x8__aarch64_neonfma_cortex_a75, |
| 412 | .igemm1 = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_1x8__aarch64_neonfma_cortex_a75, |
Frank Barchard | 2af471b | 2019-10-16 19:10:32 -0700 | [diff] [blame] | 413 | .mr = 6, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 414 | .nr = 8, |
| 415 | }; |
| 416 | break; |
| 417 | } |
| 418 | #else // XNN_ENABLE_ASSEMBLY |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 419 | xnn_params.f32.gemm = (struct gemm_parameters) { |
Frank Barchard | 91317c5 | 2019-11-22 10:54:35 -0800 | [diff] [blame] | 420 | .gemm = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_6x8__neonfma_lane_ld64, |
| 421 | .igemm = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_6x8__neonfma_lane_ld64, |
| 422 | .gemm1 = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_1x8__neonfma_lane_ld64, |
| 423 | .igemm1 = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_1x8__neonfma_lane_ld64, |
Frank Barchard | 2af471b | 2019-10-16 19:10:32 -0700 | [diff] [blame] | 424 | .mr = 6, |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 425 | .nr = 8, |
| 426 | }; |
Frank Barchard | 3267092 | 2019-11-30 21:58:51 -0800 | [diff] [blame] | 427 | #endif // XNN_ENABLE_ASSEMBLY |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 428 | |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 429 | xnn_params.f32.gemm2 = (struct gemm_parameters) { |
| 430 | .gemm = NULL, |
Frank Barchard | 91317c5 | 2019-11-22 10:54:35 -0800 | [diff] [blame] | 431 | .igemm = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_4x2__neonfma_lane_ld64, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 432 | .mr = 4, |
| 433 | .nr = 2, |
| 434 | }; |
| 435 | xnn_params.f32.dwconv[0] = (struct dwconv_parameters) { |
| 436 | .up = (xnn_dwconv_up_ukernel_function) xnn_f32_dwconv_ukernel_up4x4__psimd, |
| 437 | .cr = 4, |
| 438 | .mr = 4, |
| 439 | }; |
| 440 | switch (cpuinfo_get_core(0)->uarch) { |
| 441 | case cpuinfo_uarch_kryo: |
| 442 | xnn_params.f32.dwconv[1] = (struct dwconv_parameters) { |
| 443 | .up = (xnn_dwconv_up_ukernel_function) xnn_f32_dwconv_ukernel_up4x9__neonfma, |
| 444 | .cr = 4, |
| 445 | .mr = 9, |
| 446 | }; |
| 447 | break; |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 448 | #if XNN_ENABLE_ASSEMBLY |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 449 | case cpuinfo_uarch_cortex_a53: |
| 450 | case cpuinfo_uarch_cortex_a55: |
| 451 | xnn_params.f32.dwconv[1] = (struct dwconv_parameters) { |
| 452 | .up = (xnn_dwconv_up_ukernel_function) xnn_f32_dwconv_ukernel_up4x9__aarch64_neonfma_cortex_a55, |
| 453 | .cr = 4, |
| 454 | .mr = 9, |
| 455 | }; |
| 456 | break; |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 457 | #endif |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 458 | default: |
| 459 | xnn_params.f32.dwconv[1] = (struct dwconv_parameters) { |
| 460 | .up = (xnn_dwconv_up_ukernel_function) xnn_f32_dwconv_ukernel_up8x9__neonfma, |
| 461 | .cr = 8, |
| 462 | .mr = 9, |
| 463 | }; |
| 464 | break; |
| 465 | } |
| 466 | xnn_params.f32.dwconv[2] = (struct dwconv_parameters) { |
| 467 | .up = (xnn_dwconv_up_ukernel_function) xnn_f32_dwconv_ukernel_up4x25__psimd, |
| 468 | .cr = 4, |
| 469 | .mr = 25, |
| 470 | }; |
| 471 | xnn_params.f32.avgpool = (struct avgpool_parameters) { |
| 472 | .up = (xnn_avgpool_up_ukernel_function) xnn_f32_avgpool_ukernel_up9__neon, |
| 473 | .mp = (xnn_avgpool_mp_ukernel_function) xnn_f32_avgpool_ukernel_mp9p8q__neon, |
| 474 | .mr = 9, |
| 475 | .qr = 8, |
| 476 | }; |
| 477 | xnn_params.f32.pavgpool = (struct pavgpool_parameters) { |
| 478 | .up = (xnn_pavgpool_up_ukernel_function) xnn_f32_pavgpool_ukernel_up9__neon, |
| 479 | .mp = (xnn_pavgpool_mp_ukernel_function) xnn_f32_pavgpool_ukernel_mp9p8q__neon, |
| 480 | .mr = 9, |
| 481 | .qr = 8, |
| 482 | }; |
| 483 | xnn_params.f32.gavgpool = (struct gavgpool_parameters) { |
| 484 | .up = (xnn_gavgpool_up_ukernel_function) xnn_f32_gavgpool_ukernel_up7__neon, |
| 485 | .mp = (xnn_gavgpool_mp_ukernel_function) xnn_f32_gavgpool_ukernel_mp7p7q__neon, |
| 486 | .mr = 7, |
| 487 | }; |
| 488 | xnn_params.f32.maxpool = (struct maxpool_parameters) { |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 489 | .ukernel = (xnn_maxpool_ukernel_function) xnn_f32_maxpool_ukernel_9p8x__psimd_c4, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 490 | .mr = 9, |
| 491 | .qr = 8, |
| 492 | }; |
| 493 | xnn_params.f32.argmaxpool[0] = (struct argmaxpool_parameters) { |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 494 | .up = (xnn_argmaxpool_up_ukernel_function) xnn_f32_argmaxpool_ukernel_4x__psimd_c4, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 495 | .mr = 4, |
| 496 | }; |
| 497 | xnn_params.f32.argmaxpool[1] = (struct argmaxpool_parameters) { |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 498 | .up = (xnn_argmaxpool_up_ukernel_function) xnn_f32_argmaxpool_ukernel_9x__psimd_c4, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 499 | .mr = 9, |
| 500 | }; |
| 501 | xnn_params.f32.argmaxpool[2] = (struct argmaxpool_parameters) { |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 502 | .mp = (xnn_argmaxpool_mp_ukernel_function) xnn_f32_argmaxpool_ukernel_9p8x__psimd_c4, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 503 | .mr = 9, |
| 504 | .qr = 8, |
| 505 | }; |
Marat Dukhan | 6972249 | 2019-11-11 19:55:50 -0800 | [diff] [blame] | 506 | xnn_params.f32.bilinear = (struct bilinear_parameters) { |
| 507 | .ukernel = (xnn_bilinear_ukernel_function) xnn_f32_bilinear_ukernel__neonfma_c8, |
| 508 | .pixel_tile = 1, |
| 509 | .channel_tile = 8, |
| 510 | }; |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 511 | xnn_params.f32.clamp = (xnn_univector_ukernel_function) xnn_f32_clamp_ukernel__neon; |
| 512 | xnn_params.f32.hswish = (xnn_univector_ukernel_function) xnn_f32_hswish_ukernel__neonfma; |
Marat Dukhan | 14bec50 | 2019-11-18 11:35:31 -0800 | [diff] [blame] | 513 | xnn_params.f32.sigmoid = (xnn_univector_ukernel_function) xnn_f32_sigmoid_ukernel__neon_frac_p9_p10_nr1recps_x16; |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 514 | xnn_params.f32.prelu = (struct prelu_parameters) { |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 515 | .ukernel = (xnn_prelu_ukernel_function) xnn_f32_prelu_ukernel__neon_2x8, |
| 516 | .row_tile = 2, |
| 517 | .channel_tile = 8, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 518 | }; |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 519 | xnn_params.f32.vadd = (struct vbinary_parameters) { |
| 520 | .op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vadd_ukernel__neon_x8, |
| 521 | .opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_ukernel__neon_x8, |
| 522 | .ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_ukernel__neon_x8, |
| 523 | .element_tile = 8, |
| 524 | }; |
Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 525 | xnn_params.f32.vmax = (struct vbinary_parameters) { |
| 526 | .op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmax_ukernel__neon_x8, |
| 527 | .opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__neon_x8, |
| 528 | .ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__neon_x8, |
| 529 | .element_tile = 8, |
| 530 | }; |
| 531 | xnn_params.f32.vmin = (struct vbinary_parameters) { |
| 532 | .op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmin_ukernel__neon_x8, |
| 533 | .opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__neon_x8, |
| 534 | .ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__neon_x8, |
| 535 | .element_tile = 8, |
| 536 | }; |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 537 | xnn_params.f32.vmul = (struct vbinary_parameters) { |
| 538 | .op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmul_ukernel__neon_x8, |
| 539 | .opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_ukernel__neon_x8, |
| 540 | .ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_ukernel__neon_x8, |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 541 | .element_tile = 8, |
| 542 | }; |
Marat Dukhan | 05f3f6d | 2019-12-03 15:13:53 -0800 | [diff] [blame] | 543 | xnn_params.f32.vsub = (struct vbinary_parameters) { |
| 544 | .op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsub_ukernel__neon_x8, |
| 545 | .opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsubc_ukernel__neon_x8, |
| 546 | .ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrsubc_ukernel__neon_x8, |
| 547 | .element_tile = 8, |
| 548 | }; |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 549 | xnn_params.f32.vmulcaddc = (struct vmulcaddc_parameters) { |
Marat Dukhan | 49e6ee9 | 2019-11-06 15:55:29 -0800 | [diff] [blame] | 550 | .ukernel = (xnn_vmulcaddc_ukernel_function) xnn_f32_vmulcaddc_ukernel_c4__neonfma_2x, |
| 551 | .channel_tile = 4, |
| 552 | .row_tile = 2, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 553 | }; |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 554 | #ifndef XNN_NO_NCHW_OPERATORS |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 555 | xnn_params.f32.spmm = (struct spmm_parameters) { |
Erich Elsen | 9cdade3 | 2019-10-16 05:26:59 -0700 | [diff] [blame] | 556 | .ukernel = (xnn_spmm_ukernel_function) xnn_f32_spmm_ukernel_16x1__neonfma_pipelined, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 557 | .mr = 16, |
| 558 | .nr = 1, |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 559 | }; |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 560 | xnn_params.f32.spmm2 = (struct spmm_parameters) { |
| 561 | .ukernel = (xnn_spmm_ukernel_function) xnn_f32_spmm_ukernel_16x2__neonfma, |
| 562 | .mr = 16, |
| 563 | .nr = 2, |
| 564 | }; |
| 565 | xnn_params.f32.spmm4 = (struct spmm_parameters) { |
| 566 | .ukernel = (xnn_spmm_ukernel_function) xnn_f32_spmm_ukernel_16x4__neonfma, |
| 567 | .mr = 16, |
| 568 | .nr = 4, |
| 569 | }; |
| 570 | xnn_params.f32.hwc2spchw_dconv3x3c3s2 = (struct hwc2spchw_dconv_parameters) { |
| 571 | .ukernel_with_symm_padding = |
| 572 | (xnn_conv_hwc2spchw_ukernel_function) xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__neonfma_2x2, |
| 573 | .output_channel_tile = 4, |
| 574 | .output_height_tile = 2, |
| 575 | .output_width_tile = 2, |
| 576 | }; |
| 577 | xnn_params.f32.spchw_dwconv3x3 = (struct spchw_dwconv_parameters) { |
| 578 | .ukernel = (xnn_dwconv_spchw_ukernel_function) xnn_f32_dwconv_spchw_ukernel_3x3p1__neonfma, |
| 579 | .input_width_tile = 4, |
| 580 | .output_width_tile = 4, |
| 581 | .output_height_tile = 3, |
| 582 | }; |
| 583 | xnn_params.f32.spchw_dwconv3x3s2 = (struct spchw_dwconv_parameters) { |
| 584 | .ukernel = (xnn_dwconv_spchw_ukernel_function) xnn_f32_dwconv_spchw_ukernel_3x3s2p1__neonfma, |
| 585 | .input_width_tile = 4, |
| 586 | .output_width_tile = 4, |
| 587 | .output_height_tile = 1, |
| 588 | }; |
Marat Dukhan | a99918a | 2019-11-15 14:40:12 -0800 | [diff] [blame] | 589 | xnn_params.f32.spchw_dwconv5x5 = (struct spchw_dwconv_parameters) { |
| 590 | .ukernel = (xnn_dwconv_spchw_ukernel_function) xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma, |
| 591 | .input_width_tile = 4, |
| 592 | .output_width_tile = 4, |
Erich Elsen | 4ad5115 | 2019-11-19 13:11:53 -0800 | [diff] [blame] | 593 | .output_height_tile = 3, |
Marat Dukhan | a99918a | 2019-11-15 14:40:12 -0800 | [diff] [blame] | 594 | }; |
| 595 | xnn_params.f32.spchw_dwconv5x5s2 = (struct spchw_dwconv_parameters) { |
| 596 | .ukernel = (xnn_dwconv_spchw_ukernel_function) xnn_f32_dwconv_spchw_ukernel_5x5s2p2__neonfma, |
| 597 | .input_width_tile = 4, |
| 598 | .output_width_tile = 4, |
| 599 | .output_height_tile = 1, |
| 600 | }; |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 601 | xnn_params.f32.spchw_gavgpool = (struct spchw_gavgpool_parameters) { |
| 602 | .ukernel = (xnn_gavgpool_spchw_ukernel_function) xnn_f32_gavgpool_spchw_ukernel__neon_x4, |
| 603 | .channel_tile = 4, |
| 604 | }; |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 605 | #endif // XNN_NO_NCHW_OPERATORS |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 606 | #endif // XNN_NO_F32_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 607 | |
| 608 | /**************************** X32 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 609 | #ifndef XNN_NO_X32_OPERATORS |
| 610 | xnn_params.x32.pad = (struct pad_parameters) { |
| 611 | .ukernel = xnn_x32_pad_x2__neon, |
| 612 | .mr = 2, |
| 613 | }; |
| 614 | xnn_params.x32.unpool = (xnn_unpool_ukernel_function) xnn_x32_unpool_ukernel__psimd; |
| 615 | xnn_params.x32.zip = (struct zip_parameters) { |
| 616 | .x2 = (xnn_zipc_ukernel_function) xnn_x32_zip_x2_ukernel__neon, |
| 617 | .x3 = (xnn_zipc_ukernel_function) xnn_x32_zip_x3_ukernel__neon, |
| 618 | .x4 = (xnn_zipc_ukernel_function) xnn_x32_zip_x4_ukernel__neon, |
| 619 | .xm = (xnn_zipv_ukernel_function) xnn_x32_zip_xm_ukernel__neon, |
| 620 | }; |
| 621 | #endif // XNN_NO_X32_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 622 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 623 | #elif XNN_ARCH_X86 || XNN_ARCH_X86_64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 624 | if (!cpuinfo_has_x86_sse2()) { |
| 625 | xnn_log_error("XNNPACK initialization failed: SSE2 is not supported"); |
| 626 | return; |
| 627 | } |
| 628 | |
| 629 | /**************************** Q8 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 630 | #ifndef XNN_NO_Q8_OPERATORS |
| 631 | xnn_params.q8.gemm = (struct gemm_parameters) { |
| 632 | .gemm = (xnn_gemm_ukernel_function) xnn_q8_gemm_ukernel_4x4c2__sse2, |
| 633 | .igemm = (xnn_igemm_ukernel_function) xnn_q8_igemm_ukernel_4x4c2__sse2, |
| 634 | .mr = 4, |
| 635 | .nr = 4, |
| 636 | .log2_kr = 1, |
| 637 | }; |
| 638 | xnn_params.q8.dwconv[0] = (struct dwconv_parameters) { |
| 639 | .up = (xnn_dwconv_up_ukernel_function) xnn_q8_dwconv_ukernel_up8x9__sse2, |
| 640 | .cr = 8, |
| 641 | .mr = 9, |
| 642 | }; |
| 643 | xnn_params.q8.avgpool = (struct avgpool_parameters) { |
| 644 | .up = (xnn_avgpool_up_ukernel_function) xnn_q8_avgpool_ukernel_up9__sse2, |
| 645 | .mp = (xnn_avgpool_mp_ukernel_function) xnn_q8_avgpool_ukernel_mp9p8q__sse2, |
| 646 | .mr = 9, |
| 647 | .qr = 8, |
| 648 | }; |
| 649 | xnn_params.q8.gavgpool = (struct gavgpool_parameters) { |
| 650 | .up = (xnn_gavgpool_up_ukernel_function) xnn_q8_gavgpool_ukernel_up7__sse2, |
| 651 | .mp = (xnn_gavgpool_mp_ukernel_function) xnn_q8_gavgpool_ukernel_mp7p7q__sse2, |
| 652 | .mr = 7, |
| 653 | }; |
| 654 | xnn_params.q8.vadd = (xnn_vadd_ukernel_function) xnn_q8_vadd_ukernel__sse2; |
| 655 | #endif // XNN_NO_Q8_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 656 | |
| 657 | /**************************** U8 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 658 | #ifndef XNN_NO_U8_OPERATORS |
| 659 | xnn_params.u8.maxpool = (struct maxpool_parameters) { |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 660 | .ukernel = (xnn_maxpool_ukernel_function) xnn_u8_maxpool_ukernel_9p8x__sse2_c16, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 661 | .mr = 9, |
| 662 | .qr = 8, |
| 663 | }; |
| 664 | xnn_params.u8.clamp = (xnn_univector_ukernel_function) xnn_u8_clamp_ukernel__sse2; |
| 665 | xnn_params.u8.lut32norm = xnn_u8_lut32norm_ukernel__scalar; |
| 666 | xnn_params.u8.rmax = xnn_u8_rmax_ukernel__sse2; |
| 667 | #endif // XNN_NO_U8_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 668 | |
| 669 | /**************************** X8 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 670 | #ifndef XNN_NO_X8_OPERATORS |
| 671 | xnn_params.x8.lut = xnn_x8_lut_ukernel__scalar; |
| 672 | xnn_params.x8.zip = (struct zip_parameters) { |
| 673 | .x2 = (xnn_zipc_ukernel_function) xnn_x8_zip_x2_ukernel__sse2, |
| 674 | .x3 = (xnn_zipc_ukernel_function) xnn_x8_zip_x3_ukernel__sse2, |
| 675 | .x4 = (xnn_zipc_ukernel_function) xnn_x8_zip_x4_ukernel__sse2, |
| 676 | .xm = (xnn_zipv_ukernel_function) xnn_x8_zip_xm_ukernel__sse2, |
| 677 | }; |
| 678 | #endif // XNN_NO_X8_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 679 | |
| 680 | /**************************** F32 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 681 | #ifndef XNN_NO_F32_OPERATORS |
Marat Dukhan | 0f349c4 | 2019-11-27 11:58:54 -0800 | [diff] [blame] | 682 | if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx512f()) { |
| 683 | xnn_params.f32.gemm = (struct gemm_parameters) { |
| 684 | .gemm = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_7x16__avx512f_broadcast, |
| 685 | .igemm = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_7x16__avx512f_broadcast, |
| 686 | .gemm1 = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_1x16__avx512f_broadcast, |
| 687 | .igemm1 = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_1x16__avx512f_broadcast, |
| 688 | .mr = 7, |
| 689 | .nr = 16, |
| 690 | }; |
| 691 | } else if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_fma3()) { |
Marat Dukhan | 1025ea3 | 2019-11-21 16:01:08 -0800 | [diff] [blame] | 692 | xnn_params.f32.gemm = (struct gemm_parameters) { |
| 693 | .gemm = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_7x8__fma3_broadcast, |
| 694 | .igemm = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_7x8__fma3_broadcast, |
| 695 | .gemm1 = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_1x8__fma3_broadcast, |
| 696 | .igemm1 = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_1x8__fma3_broadcast, |
| 697 | .mr = 7, |
| 698 | .nr = 8, |
| 699 | }; |
| 700 | } else if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx()) { |
| 701 | xnn_params.f32.gemm = (struct gemm_parameters) { |
| 702 | .gemm = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_7x8__avx_broadcast, |
| 703 | .igemm = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_7x8__avx_broadcast, |
| 704 | .gemm1 = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_1x8__avx_broadcast, |
| 705 | .igemm1 = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_1x8__avx_broadcast, |
| 706 | .mr = 7, |
| 707 | .nr = 8, |
| 708 | }; |
| 709 | } else { |
| 710 | xnn_params.f32.gemm = (struct gemm_parameters) { |
| 711 | .gemm = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_4x8__sse_load1, |
| 712 | .igemm = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_4x8__sse_load1, |
| 713 | .gemm1 = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_1x8__sse_load1, |
| 714 | .igemm1 = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_1x8__sse_load1, |
| 715 | .mr = 4, |
| 716 | .nr = 8, |
| 717 | }; |
| 718 | } |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 719 | xnn_params.f32.gemm2 = (struct gemm_parameters) { |
| 720 | .gemm = NULL, |
| 721 | .igemm = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_4x2c4__sse, |
| 722 | .mr = 4, |
| 723 | .nr = 2, |
| 724 | .log2_kr = 2, |
| 725 | }; |
Marat Dukhan | 479f87e | 2019-11-27 15:17:06 -0800 | [diff] [blame] | 726 | if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx512f()) { |
| 727 | xnn_params.f32.dwconv[0] = (struct dwconv_parameters) { |
| 728 | .up = (xnn_dwconv_up_ukernel_function) xnn_f32_dwconv_ukernel_up16x4__avx512f, |
| 729 | .cr = 16, |
| 730 | .mr = 4, |
| 731 | }; |
| 732 | xnn_params.f32.dwconv[1] = (struct dwconv_parameters) { |
| 733 | .up = (xnn_dwconv_up_ukernel_function) xnn_f32_dwconv_ukernel_up16x9__avx512f, |
| 734 | .cr = 16, |
| 735 | .mr = 9, |
| 736 | }; |
| 737 | xnn_params.f32.dwconv[2] = (struct dwconv_parameters) { |
| 738 | .up = (xnn_dwconv_up_ukernel_function) xnn_f32_dwconv_ukernel_up16x25__avx512f, |
| 739 | .cr = 16, |
| 740 | .mr = 25, |
| 741 | }; |
| 742 | } else if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_fma3()) { |
Marat Dukhan | 17ec5f3 | 2019-11-22 13:34:16 -0800 | [diff] [blame] | 743 | xnn_params.f32.dwconv[0] = (struct dwconv_parameters) { |
| 744 | .up = (xnn_dwconv_up_ukernel_function) xnn_f32_dwconv_ukernel_up16x4__fma3, |
| 745 | .cr = 16, |
| 746 | .mr = 4, |
| 747 | }; |
| 748 | xnn_params.f32.dwconv[1] = (struct dwconv_parameters) { |
| 749 | .up = (xnn_dwconv_up_ukernel_function) xnn_f32_dwconv_ukernel_up16x9__fma3, |
| 750 | .cr = 16, |
| 751 | .mr = 9, |
| 752 | }; |
| 753 | xnn_params.f32.dwconv[2] = (struct dwconv_parameters) { |
| 754 | .up = (xnn_dwconv_up_ukernel_function) xnn_f32_dwconv_ukernel_up8x25__fma3, |
| 755 | .cr = 8, |
| 756 | .mr = 25, |
| 757 | }; |
| 758 | } else if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx()) { |
| 759 | xnn_params.f32.dwconv[0] = (struct dwconv_parameters) { |
| 760 | .up = (xnn_dwconv_up_ukernel_function) xnn_f32_dwconv_ukernel_up16x4__avx, |
| 761 | .cr = 16, |
| 762 | .mr = 4, |
| 763 | }; |
| 764 | xnn_params.f32.dwconv[1] = (struct dwconv_parameters) { |
| 765 | .up = (xnn_dwconv_up_ukernel_function) xnn_f32_dwconv_ukernel_up16x9__avx, |
| 766 | .cr = 16, |
| 767 | .mr = 9, |
| 768 | }; |
| 769 | xnn_params.f32.dwconv[2] = (struct dwconv_parameters) { |
| 770 | .up = (xnn_dwconv_up_ukernel_function) xnn_f32_dwconv_ukernel_up8x25__avx, |
| 771 | .cr = 8, |
| 772 | .mr = 25, |
| 773 | }; |
| 774 | } else { |
| 775 | xnn_params.f32.dwconv[0] = (struct dwconv_parameters) { |
| 776 | .up = (xnn_dwconv_up_ukernel_function) xnn_f32_dwconv_ukernel_up8x4__sse, |
| 777 | .cr = 8, |
| 778 | .mr = 4, |
| 779 | }; |
| 780 | xnn_params.f32.dwconv[1] = (struct dwconv_parameters) { |
| 781 | .up = (xnn_dwconv_up_ukernel_function) xnn_f32_dwconv_ukernel_up8x9__sse, |
| 782 | .cr = 8, |
| 783 | .mr = 9, |
| 784 | }; |
| 785 | xnn_params.f32.dwconv[2] = (struct dwconv_parameters) { |
| 786 | .up = (xnn_dwconv_up_ukernel_function) xnn_f32_dwconv_ukernel_up8x25__sse, |
| 787 | .cr = 8, |
| 788 | .mr = 25, |
| 789 | }; |
| 790 | } |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 791 | xnn_params.f32.avgpool = (struct avgpool_parameters) { |
| 792 | .up = (xnn_avgpool_up_ukernel_function) xnn_f32_avgpool_ukernel_up9__sse, |
| 793 | .mp = (xnn_avgpool_mp_ukernel_function) xnn_f32_avgpool_ukernel_mp9p8q__sse, |
| 794 | .mr = 9, |
| 795 | .qr = 8, |
| 796 | }; |
| 797 | xnn_params.f32.pavgpool = (struct pavgpool_parameters) { |
| 798 | .up = (xnn_pavgpool_up_ukernel_function) xnn_f32_pavgpool_ukernel_up9__sse, |
| 799 | .mp = (xnn_pavgpool_mp_ukernel_function) xnn_f32_pavgpool_ukernel_mp9p8q__sse, |
| 800 | .mr = 9, |
| 801 | .qr = 8, |
| 802 | }; |
| 803 | xnn_params.f32.gavgpool = (struct gavgpool_parameters) { |
| 804 | .up = (xnn_gavgpool_up_ukernel_function) xnn_f32_gavgpool_ukernel_up7__sse, |
| 805 | .mp = (xnn_gavgpool_mp_ukernel_function) xnn_f32_gavgpool_ukernel_mp7p7q__sse, |
| 806 | .mr = 7, |
| 807 | }; |
| 808 | xnn_params.f32.maxpool = (struct maxpool_parameters) { |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 809 | .ukernel = (xnn_maxpool_ukernel_function) xnn_f32_maxpool_ukernel_9p8x__sse_c4, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 810 | .mr = 9, |
| 811 | .qr = 8, |
| 812 | }; |
| 813 | xnn_params.f32.argmaxpool[0] = (struct argmaxpool_parameters) { |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 814 | .up = (xnn_argmaxpool_up_ukernel_function) xnn_f32_argmaxpool_ukernel_4x__sse2_c4, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 815 | .mr = 4, |
| 816 | }; |
| 817 | xnn_params.f32.argmaxpool[1] = (struct argmaxpool_parameters) { |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 818 | .up = (xnn_argmaxpool_up_ukernel_function) xnn_f32_argmaxpool_ukernel_9x__sse2_c4, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 819 | .mr = 9, |
| 820 | }; |
| 821 | xnn_params.f32.argmaxpool[2] = (struct argmaxpool_parameters) { |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 822 | .mp = (xnn_argmaxpool_mp_ukernel_function) xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 823 | .mr = 9, |
| 824 | .qr = 8, |
| 825 | }; |
Marat Dukhan | 6972249 | 2019-11-11 19:55:50 -0800 | [diff] [blame] | 826 | xnn_params.f32.bilinear = (struct bilinear_parameters) { |
| 827 | .ukernel = (xnn_bilinear_ukernel_function) xnn_f32_bilinear_ukernel__sse_c8, |
| 828 | .pixel_tile = 1, |
| 829 | .channel_tile = 8, |
| 830 | }; |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 831 | if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx512f()) { |
| 832 | xnn_params.f32.clamp = (xnn_univector_ukernel_function) xnn_f32_clamp_ukernel__avx512f; |
| 833 | } else if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx()) { |
| 834 | xnn_params.f32.clamp = (xnn_univector_ukernel_function) xnn_f32_clamp_ukernel__avx; |
| 835 | } else { |
| 836 | xnn_params.f32.clamp = (xnn_univector_ukernel_function) xnn_f32_clamp_ukernel__sse; |
| 837 | } |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 838 | xnn_params.f32.hswish = (xnn_univector_ukernel_function) xnn_f32_hswish_ukernel__sse; |
Marat Dukhan | 7bee751 | 2019-11-18 15:15:48 -0800 | [diff] [blame] | 839 | xnn_params.f32.sigmoid = (xnn_univector_ukernel_function) xnn_f32_sigmoid_ukernel__sse2_p5_div_x16; |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 840 | xnn_params.f32.prelu = (struct prelu_parameters) { |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 841 | .ukernel = (xnn_prelu_ukernel_function) xnn_f32_prelu_ukernel__sse2_2x8, |
| 842 | .row_tile = 2, |
| 843 | .channel_tile = 8, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 844 | }; |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 845 | xnn_params.f32.vadd = (struct vbinary_parameters) { |
| 846 | .op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vadd_ukernel__sse_x8, |
| 847 | .opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_ukernel__sse_x8, |
| 848 | .ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_ukernel__sse_x8, |
| 849 | .element_tile = 8, |
| 850 | }; |
Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 851 | xnn_params.f32.vmax = (struct vbinary_parameters) { |
| 852 | .op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmax_ukernel__sse_x8, |
| 853 | .opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__sse_x8, |
| 854 | .ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__sse_x8, |
| 855 | .element_tile = 8, |
| 856 | }; |
| 857 | xnn_params.f32.vmin = (struct vbinary_parameters) { |
| 858 | .op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmin_ukernel__sse_x8, |
| 859 | .opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__sse_x8, |
| 860 | .ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__sse_x8, |
| 861 | .element_tile = 8, |
| 862 | }; |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 863 | xnn_params.f32.vmul = (struct vbinary_parameters) { |
| 864 | .op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmul_ukernel__sse_x8, |
| 865 | .opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_ukernel__sse_x8, |
| 866 | .ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_ukernel__sse_x8, |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 867 | .element_tile = 8, |
| 868 | }; |
Marat Dukhan | 05f3f6d | 2019-12-03 15:13:53 -0800 | [diff] [blame] | 869 | xnn_params.f32.vsub = (struct vbinary_parameters) { |
| 870 | .op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsub_ukernel__sse_x8, |
| 871 | .opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsubc_ukernel__sse_x8, |
| 872 | .ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrsubc_ukernel__sse_x8, |
| 873 | .element_tile = 8, |
| 874 | }; |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 875 | xnn_params.f32.vmulcaddc = (struct vmulcaddc_parameters) { |
Marat Dukhan | 49e6ee9 | 2019-11-06 15:55:29 -0800 | [diff] [blame] | 876 | .ukernel = (xnn_vmulcaddc_ukernel_function) xnn_f32_vmulcaddc_ukernel_c4__sse_2x, |
| 877 | .channel_tile = 4, |
| 878 | .row_tile = 2, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 879 | }; |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 880 | #ifndef XNN_NO_NCHW_OPERATORS |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 881 | xnn_params.f32.spmm = (struct spmm_parameters) { |
| 882 | .ukernel = (xnn_spmm_ukernel_function) xnn_f32_spmm_ukernel_4x1__sse, |
| 883 | .mr = 4, |
| 884 | .nr = 1, |
| 885 | }; |
| 886 | xnn_params.f32.spchw_dwconv3x3 = (struct spchw_dwconv_parameters) { |
| 887 | .ukernel = (xnn_dwconv_spchw_ukernel_function) xnn_f32_dwconv_spchw_ukernel_3x3p1__sse, |
| 888 | .input_width_tile = 4, |
| 889 | .output_width_tile = 4, |
| 890 | .output_height_tile = 1, |
| 891 | }; |
| 892 | xnn_params.f32.spchw_dwconv3x3s2 = (struct spchw_dwconv_parameters) { |
| 893 | .ukernel = (xnn_dwconv_spchw_ukernel_function) xnn_f32_dwconv_spchw_ukernel_3x3s2p1__sse, |
| 894 | .input_width_tile = 4, |
| 895 | .output_width_tile = 4, |
| 896 | .output_height_tile = 1, |
| 897 | }; |
| 898 | xnn_params.f32.spchw_gavgpool = (struct spchw_gavgpool_parameters) { |
| 899 | .ukernel = (xnn_gavgpool_spchw_ukernel_function) xnn_f32_gavgpool_spchw_ukernel__sse_x4, |
| 900 | .channel_tile = 4, |
| 901 | }; |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 902 | #endif // XNN_NO_NCHW_OPERATORS |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 903 | #endif // XNN_NO_F32_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 904 | |
| 905 | /**************************** X32 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 906 | #ifndef XNN_NO_X32_OPERATORS |
| 907 | xnn_params.x32.pad = (struct pad_parameters) { |
| 908 | .ukernel = xnn_x32_pad_x2__sse2, |
| 909 | .mr = 2, |
| 910 | }; |
| 911 | xnn_params.x32.unpool = (xnn_unpool_ukernel_function) xnn_x32_unpool_ukernel__psimd; |
| 912 | xnn_params.x32.zip = (struct zip_parameters) { |
| 913 | .x2 = (xnn_zipc_ukernel_function) xnn_x32_zip_x2_ukernel__sse2, |
| 914 | .x3 = (xnn_zipc_ukernel_function) xnn_x32_zip_x3_ukernel__sse2, |
| 915 | .x4 = (xnn_zipc_ukernel_function) xnn_x32_zip_x4_ukernel__sse2, |
| 916 | .xm = (xnn_zipv_ukernel_function) xnn_x32_zip_xm_ukernel__sse2, |
| 917 | }; |
| 918 | #endif // XNN_NO_X32_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 919 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 920 | #elif XNN_ARCH_PNACL || XNN_ARCH_WASMSIMD |
Marat Dukhan | 466b523 | 2019-10-09 11:22:20 -0700 | [diff] [blame] | 921 | // Unlike most other architectures, on x86/x86-64 when floating-point instructions |
| 922 | // have no NaN arguments, but produce NaN output, the output NaN has sign bit set. |
| 923 | // We use it to distinguish x86/x86-64 from other architectures, by doing subtraction |
| 924 | // of two infinities (must produce NaN per IEEE 754 standard). |
| 925 | static volatile uint32_t minus_inf = UINT32_C(0xFF800000); |
| 926 | const bool is_wasm_x86 = (int32_t) xnn_stub_wasm_f32_sub(minus_inf, minus_inf) < 0; |
| 927 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 928 | /**************************** Q8 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 929 | #ifndef XNN_NO_Q8_OPERATORS |
| 930 | xnn_params.q8.gemm = (struct gemm_parameters) { |
| 931 | .gemm = (xnn_gemm_ukernel_function) xnn_q8_gemm_ukernel_2x2__scalar, |
| 932 | .igemm = (xnn_igemm_ukernel_function) xnn_q8_igemm_ukernel_2x2__scalar, |
| 933 | .mr = 2, |
| 934 | .nr = 2, |
| 935 | }; |
| 936 | xnn_params.q8.dwconv[0] = (struct dwconv_parameters) { |
| 937 | .up = (xnn_dwconv_up_ukernel_function) xnn_q8_dwconv_ukernel_up1x9__scalar, |
| 938 | .cr = 1, |
| 939 | .mr = 9, |
| 940 | }; |
| 941 | xnn_params.q8.avgpool = (struct avgpool_parameters) { |
| 942 | .up = (xnn_avgpool_up_ukernel_function) xnn_q8_avgpool_ukernel_up9__scalar, |
| 943 | .mp = (xnn_avgpool_mp_ukernel_function) xnn_q8_avgpool_ukernel_mp9p8q__scalar, |
| 944 | .mr = 9, |
| 945 | .qr = 8, |
| 946 | }; |
| 947 | xnn_params.q8.gavgpool = (struct gavgpool_parameters) { |
| 948 | .up = (xnn_gavgpool_up_ukernel_function) xnn_q8_gavgpool_ukernel_up7__scalar, |
| 949 | .mp = (xnn_gavgpool_mp_ukernel_function) xnn_q8_gavgpool_ukernel_mp7p7q__scalar, |
| 950 | .mr = 7, |
| 951 | }; |
| 952 | xnn_params.q8.vadd = (xnn_vadd_ukernel_function) xnn_q8_vadd_ukernel__scalar; |
| 953 | #endif // XNN_NO_Q8_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 954 | |
| 955 | /**************************** U8 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 956 | #ifndef XNN_NO_U8_OPERATORS |
| 957 | xnn_params.u8.maxpool = (struct maxpool_parameters) { |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 958 | .ukernel = (xnn_maxpool_ukernel_function) xnn_u8_maxpool_ukernel_9p8x__scalar_c1, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 959 | .mr = 9, |
| 960 | .qr = 8, |
| 961 | }; |
| 962 | xnn_params.u8.clamp = (xnn_univector_ukernel_function) xnn_u8_clamp_ukernel__scalar; |
| 963 | xnn_params.u8.lut32norm = xnn_u8_lut32norm_ukernel__scalar; |
| 964 | xnn_params.u8.rmax = xnn_u8_rmax_ukernel__scalar; |
| 965 | #endif // XNN_NO_U8_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 966 | |
| 967 | /**************************** X8 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 968 | #ifndef XNN_NO_X8_OPERATORS |
| 969 | xnn_params.x8.lut = xnn_x8_lut_ukernel__scalar; |
| 970 | xnn_params.x8.zip = (struct zip_parameters) { |
| 971 | .x2 = (xnn_zipc_ukernel_function) xnn_x8_zip_x2_ukernel__scalar, |
| 972 | .x3 = (xnn_zipc_ukernel_function) xnn_x8_zip_x3_ukernel__scalar, |
| 973 | .x4 = (xnn_zipc_ukernel_function) xnn_x8_zip_x4_ukernel__scalar, |
| 974 | .xm = (xnn_zipv_ukernel_function) xnn_x8_zip_xm_ukernel__scalar, |
| 975 | }; |
| 976 | #endif // XNN_NO_X8_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 977 | |
| 978 | /**************************** F32 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 979 | #ifndef XNN_NO_F32_OPERATORS |
| 980 | if (is_wasm_x86) { |
| 981 | xnn_params.f32.gemm = (struct gemm_parameters) { |
Marat Dukhan | cb80197 | 2019-10-23 02:10:33 -0700 | [diff] [blame] | 982 | .gemm = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_4x8__psimd_splat, |
| 983 | .igemm = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_4x8__psimd_splat, |
| 984 | .gemm1 = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_1x8__psimd_splat, |
| 985 | .igemm1 = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_1x8__psimd_splat, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 986 | .mr = 4, |
| 987 | .nr = 8, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 988 | }; |
| 989 | } else { |
| 990 | xnn_params.f32.gemm = (struct gemm_parameters) { |
Marat Dukhan | cd945c6 | 2019-10-25 11:59:50 -0700 | [diff] [blame] | 991 | .gemm = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_6x8s4__psimd, |
| 992 | .igemm = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_6x8s4__psimd, |
| 993 | .gemm1 = (xnn_gemm_ukernel_function) xnn_f32_igemm_ukernel_1x8s4__psimd, |
| 994 | .igemm1 = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_1x8s4__psimd, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 995 | .mr = 6, |
| 996 | .nr = 8, |
Marat Dukhan | cd945c6 | 2019-10-25 11:59:50 -0700 | [diff] [blame] | 997 | .log2_sr = 2, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 998 | }; |
| 999 | } |
| 1000 | xnn_params.f32.gemm2 = (struct gemm_parameters) { |
| 1001 | .gemm = NULL, |
| 1002 | .igemm = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_4x2c4__psimd, |
Marat Dukhan | 466b523 | 2019-10-09 11:22:20 -0700 | [diff] [blame] | 1003 | .mr = 4, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1004 | .nr = 2, |
| 1005 | .log2_kr = 2, |
Marat Dukhan | 466b523 | 2019-10-09 11:22:20 -0700 | [diff] [blame] | 1006 | }; |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1007 | xnn_params.f32.dwconv[0] = (struct dwconv_parameters) { |
Marat Dukhan | 5098c3e | 2019-11-07 12:01:19 -0800 | [diff] [blame] | 1008 | .up = (xnn_dwconv_up_ukernel_function) xnn_f32_dwconv_ukernel_up4x4__psimd_acc2, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1009 | .cr = 4, |
| 1010 | .mr = 4, |
Marat Dukhan | 466b523 | 2019-10-09 11:22:20 -0700 | [diff] [blame] | 1011 | }; |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1012 | xnn_params.f32.dwconv[1] = (struct dwconv_parameters) { |
Marat Dukhan | 5098c3e | 2019-11-07 12:01:19 -0800 | [diff] [blame] | 1013 | .up = (xnn_dwconv_up_ukernel_function) xnn_f32_dwconv_ukernel_up4x9__psimd_acc2, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1014 | .cr = 4, |
| 1015 | .mr = 9, |
| 1016 | }; |
| 1017 | xnn_params.f32.dwconv[2] = (struct dwconv_parameters) { |
Marat Dukhan | 5098c3e | 2019-11-07 12:01:19 -0800 | [diff] [blame] | 1018 | .up = (xnn_dwconv_up_ukernel_function) xnn_f32_dwconv_ukernel_up4x25__psimd_acc2, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1019 | .cr = 4, |
| 1020 | .mr = 25, |
| 1021 | }; |
| 1022 | xnn_params.f32.avgpool = (struct avgpool_parameters) { |
| 1023 | .up = (xnn_avgpool_up_ukernel_function) xnn_f32_avgpool_ukernel_up9__psimd, |
| 1024 | .mp = (xnn_avgpool_mp_ukernel_function) xnn_f32_avgpool_ukernel_mp9p8q__psimd, |
| 1025 | .mr = 9, |
| 1026 | .qr = 8, |
| 1027 | }; |
| 1028 | xnn_params.f32.pavgpool = (struct pavgpool_parameters) { |
| 1029 | .up = (xnn_pavgpool_up_ukernel_function) xnn_f32_pavgpool_ukernel_up9__psimd, |
| 1030 | .mp = (xnn_pavgpool_mp_ukernel_function) xnn_f32_pavgpool_ukernel_mp9p8q__psimd, |
| 1031 | .mr = 9, |
| 1032 | .qr = 8, |
| 1033 | }; |
| 1034 | xnn_params.f32.gavgpool = (struct gavgpool_parameters) { |
| 1035 | .up = (xnn_gavgpool_up_ukernel_function) xnn_f32_gavgpool_ukernel_up7__psimd, |
| 1036 | .mp = (xnn_gavgpool_mp_ukernel_function) xnn_f32_gavgpool_ukernel_mp7p7q__psimd, |
| 1037 | .mr = 7, |
| 1038 | }; |
| 1039 | xnn_params.f32.maxpool = (struct maxpool_parameters) { |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 1040 | .ukernel = (xnn_maxpool_ukernel_function) xnn_f32_maxpool_ukernel_9p8x__psimd_c4, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1041 | .mr = 9, |
| 1042 | .qr = 8, |
| 1043 | }; |
| 1044 | xnn_params.f32.argmaxpool[0] = (struct argmaxpool_parameters) { |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 1045 | .up = (xnn_argmaxpool_up_ukernel_function) xnn_f32_argmaxpool_ukernel_4x__psimd_c4, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1046 | .mr = 4, |
| 1047 | }; |
| 1048 | xnn_params.f32.argmaxpool[1] = (struct argmaxpool_parameters) { |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 1049 | .up = (xnn_argmaxpool_up_ukernel_function) xnn_f32_argmaxpool_ukernel_9x__psimd_c4, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1050 | .mr = 9, |
| 1051 | }; |
| 1052 | xnn_params.f32.argmaxpool[2] = (struct argmaxpool_parameters) { |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 1053 | .mp = (xnn_argmaxpool_mp_ukernel_function) xnn_f32_argmaxpool_ukernel_9p8x__psimd_c4, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1054 | .mr = 9, |
| 1055 | .qr = 8, |
| 1056 | }; |
Marat Dukhan | 6972249 | 2019-11-11 19:55:50 -0800 | [diff] [blame] | 1057 | xnn_params.f32.bilinear = (struct bilinear_parameters) { |
| 1058 | .ukernel = (xnn_bilinear_ukernel_function) xnn_f32_bilinear_ukernel__psimd_c8, |
| 1059 | .pixel_tile = 1, |
| 1060 | .channel_tile = 8, |
| 1061 | }; |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1062 | xnn_params.f32.clamp = (xnn_univector_ukernel_function) xnn_f32_clamp_ukernel__psimd; |
| 1063 | xnn_params.f32.hswish = (xnn_univector_ukernel_function) xnn_f32_hswish_ukernel__psimd; |
| 1064 | xnn_params.f32.prelu = (struct prelu_parameters) { |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 1065 | .ukernel = (xnn_prelu_ukernel_function) xnn_f32_prelu_ukernel__psimd_2x8, |
| 1066 | .row_tile = 2, |
| 1067 | .channel_tile = 8, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1068 | }; |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 1069 | xnn_params.f32.vadd = (struct vbinary_parameters) { |
| 1070 | .op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vadd_ukernel__psimd_x8, |
| 1071 | .opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_ukernel__psimd_x8, |
| 1072 | .ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_ukernel__psimd_x8, |
| 1073 | .element_tile = 8, |
| 1074 | }; |
Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 1075 | xnn_params.f32.vmax = (struct vbinary_parameters) { |
| 1076 | .op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmax_ukernel__psimd_x8, |
| 1077 | .opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__psimd_x8, |
| 1078 | .ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__psimd_x8, |
| 1079 | .element_tile = 8, |
| 1080 | }; |
| 1081 | xnn_params.f32.vmin = (struct vbinary_parameters) { |
| 1082 | .op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmin_ukernel__psimd_x8, |
| 1083 | .opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__psimd_x8, |
| 1084 | .ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__psimd_x8, |
| 1085 | .element_tile = 8, |
| 1086 | }; |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 1087 | xnn_params.f32.vmul = (struct vbinary_parameters) { |
| 1088 | .op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmul_ukernel__psimd_x8, |
| 1089 | .opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_ukernel__psimd_x8, |
| 1090 | .ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_ukernel__psimd_x8, |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 1091 | .element_tile = 8, |
| 1092 | }; |
Marat Dukhan | 05f3f6d | 2019-12-03 15:13:53 -0800 | [diff] [blame] | 1093 | xnn_params.f32.vsub = (struct vbinary_parameters) { |
| 1094 | .op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsub_ukernel__psimd_x8, |
| 1095 | .opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsubc_ukernel__psimd_x8, |
| 1096 | .ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrsubc_ukernel__psimd_x8, |
| 1097 | .element_tile = 8, |
| 1098 | }; |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1099 | xnn_params.f32.vmulcaddc = (struct vmulcaddc_parameters) { |
Marat Dukhan | 49e6ee9 | 2019-11-06 15:55:29 -0800 | [diff] [blame] | 1100 | .ukernel = (xnn_vmulcaddc_ukernel_function) xnn_f32_vmulcaddc_ukernel_c4__psimd_2x, |
| 1101 | .channel_tile = 4, |
| 1102 | .row_tile = 2, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1103 | }; |
| 1104 | #endif // XNN_NO_F32_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1105 | |
| 1106 | /**************************** X32 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1107 | #ifndef XNN_NO_X32_OPERATORS |
| 1108 | xnn_params.x32.pad = (struct pad_parameters) { |
| 1109 | .ukernel = xnn_x32_pad_x2__psimd, |
| 1110 | .mr = 2, |
| 1111 | }; |
| 1112 | xnn_params.x32.unpool = (xnn_unpool_ukernel_function) xnn_x32_unpool_ukernel__psimd; |
| 1113 | xnn_params.x32.zip = (struct zip_parameters) { |
| 1114 | .x2 = (xnn_zipc_ukernel_function) xnn_x32_zip_x2_ukernel__psimd, |
| 1115 | .x3 = (xnn_zipc_ukernel_function) xnn_x32_zip_x3_ukernel__psimd, |
| 1116 | .x4 = (xnn_zipc_ukernel_function) xnn_x32_zip_x4_ukernel__psimd, |
| 1117 | .xm = (xnn_zipv_ukernel_function) xnn_x32_zip_xm_ukernel__psimd, |
| 1118 | }; |
| 1119 | #endif // XNN_NO_X32_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1120 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 1121 | #elif XNN_ARCH_WASM || XNN_ARCH_ASMJS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1122 | // Unlike most other architectures, on x86/x86-64 when floating-point instructions |
| 1123 | // have no NaN arguments, but produce NaN output, the output NaN has sign bit set. |
| 1124 | // We use it to distinguish x86/x86-64 from other architectures, by doing subtraction |
| 1125 | // of two infinities (must produce NaN per IEEE 754 standard). |
| 1126 | static volatile uint32_t minus_inf = UINT32_C(0xFF800000); |
| 1127 | const bool is_wasm_x86 = (int32_t) xnn_stub_wasm_f32_sub(minus_inf, minus_inf) < 0; |
| 1128 | |
| 1129 | /**************************** Q8 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1130 | #ifndef XNN_NO_Q8_OPERATORS |
| 1131 | xnn_params.q8.gemm = (struct gemm_parameters) { |
| 1132 | .gemm = (xnn_gemm_ukernel_function) xnn_q8_gemm_ukernel_2x2__scalar, |
| 1133 | .igemm = (xnn_igemm_ukernel_function) xnn_q8_igemm_ukernel_2x2__scalar, |
| 1134 | .mr = 2, |
| 1135 | .nr = 2, |
| 1136 | }; |
| 1137 | xnn_params.q8.dwconv[0] = (struct dwconv_parameters) { |
| 1138 | .up = (xnn_dwconv_up_ukernel_function) xnn_q8_dwconv_ukernel_up1x9__scalar, |
| 1139 | .cr = 1, |
| 1140 | .mr = 9, |
| 1141 | }; |
| 1142 | xnn_params.q8.avgpool = (struct avgpool_parameters) { |
| 1143 | .up = (xnn_avgpool_up_ukernel_function) xnn_q8_avgpool_ukernel_up9__scalar, |
| 1144 | .mp = (xnn_avgpool_mp_ukernel_function) xnn_q8_avgpool_ukernel_mp9p8q__scalar, |
| 1145 | .mr = 9, |
| 1146 | .qr = 8, |
| 1147 | }; |
| 1148 | xnn_params.q8.gavgpool = (struct gavgpool_parameters) { |
| 1149 | .up = (xnn_gavgpool_up_ukernel_function) xnn_q8_gavgpool_ukernel_up7__scalar, |
| 1150 | .mp = (xnn_gavgpool_mp_ukernel_function) xnn_q8_gavgpool_ukernel_mp7p7q__scalar, |
| 1151 | .mr = 7, |
| 1152 | }; |
| 1153 | xnn_params.q8.vadd = (xnn_vadd_ukernel_function) xnn_q8_vadd_ukernel__scalar; |
| 1154 | #endif // XNN_NO_Q8_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1155 | |
| 1156 | /**************************** U8 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1157 | #ifndef XNN_NO_U8_OPERATORS |
| 1158 | xnn_params.u8.maxpool = (struct maxpool_parameters) { |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 1159 | .ukernel = (xnn_maxpool_ukernel_function) xnn_u8_maxpool_ukernel_9p8x__scalar_c1, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1160 | .mr = 9, |
| 1161 | .qr = 8, |
| 1162 | }; |
| 1163 | xnn_params.u8.clamp = (xnn_univector_ukernel_function) xnn_u8_clamp_ukernel__scalar; |
| 1164 | xnn_params.u8.lut32norm = xnn_u8_lut32norm_ukernel__scalar; |
| 1165 | xnn_params.u8.rmax = xnn_u8_rmax_ukernel__scalar; |
| 1166 | #endif // XNN_NO_U8_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1167 | |
| 1168 | /**************************** X8 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1169 | #ifndef XNN_NO_X8_OPERATORS |
| 1170 | xnn_params.x8.lut = xnn_x8_lut_ukernel__scalar; |
| 1171 | xnn_params.x8.zip = (struct zip_parameters) { |
| 1172 | .x2 = (xnn_zipc_ukernel_function) xnn_x8_zip_x2_ukernel__scalar, |
| 1173 | .x3 = (xnn_zipc_ukernel_function) xnn_x8_zip_x3_ukernel__scalar, |
| 1174 | .x4 = (xnn_zipc_ukernel_function) xnn_x8_zip_x4_ukernel__scalar, |
| 1175 | .xm = (xnn_zipv_ukernel_function) xnn_x8_zip_xm_ukernel__scalar, |
| 1176 | }; |
| 1177 | #endif // XNN_NO_X8_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1178 | |
| 1179 | /**************************** F32 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1180 | #ifndef XNN_NO_F32_OPERATORS |
| 1181 | if (is_wasm_x86) { |
| 1182 | xnn_params.f32.gemm = (struct gemm_parameters) { |
| 1183 | .gemm = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_2x4__scalar, |
| 1184 | .igemm = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_2x4__scalar, |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 1185 | .gemm1 = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_1x4__wasm, |
| 1186 | .igemm1 = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_1x4__wasm, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1187 | .mr = 2, |
| 1188 | .nr = 4, |
| 1189 | }; |
| 1190 | } else { |
| 1191 | xnn_params.f32.gemm = (struct gemm_parameters) { |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 1192 | .gemm = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_4x4__wasm, |
| 1193 | .igemm = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_4x4__wasm, |
| 1194 | .gemm1 = (xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_1x4__wasm, |
| 1195 | .igemm1 = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_1x4__wasm, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1196 | .mr = 4, |
| 1197 | .nr = 4, |
| 1198 | }; |
| 1199 | } |
| 1200 | xnn_params.f32.gemm2 = (struct gemm_parameters) { |
| 1201 | .gemm = NULL, |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 1202 | .igemm = (xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_4x2__wasm, |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1203 | .mr = 4, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1204 | .nr = 2, |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1205 | }; |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1206 | xnn_params.f32.dwconv[0] = (struct dwconv_parameters) { |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 1207 | .up = (xnn_dwconv_up_ukernel_function) xnn_f32_dwconv_ukernel_up1x4__wasm_acc2, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1208 | .cr = 1, |
| 1209 | .mr = 4, |
| 1210 | }; |
| 1211 | xnn_params.f32.dwconv[1] = (struct dwconv_parameters) { |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 1212 | .up = (xnn_dwconv_up_ukernel_function) xnn_f32_dwconv_ukernel_up1x9__wasm_acc2, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1213 | .cr = 1, |
| 1214 | .mr = 9, |
| 1215 | }; |
| 1216 | xnn_params.f32.dwconv[2] = (struct dwconv_parameters) { |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 1217 | .up = (xnn_dwconv_up_ukernel_function) xnn_f32_dwconv_ukernel_up1x25__wasm_acc2, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1218 | .cr = 1, |
| 1219 | .mr = 25, |
| 1220 | }; |
| 1221 | xnn_params.f32.avgpool = (struct avgpool_parameters) { |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 1222 | .up = (xnn_avgpool_up_ukernel_function) xnn_f32_avgpool_ukernel_up9__wasm, |
| 1223 | .mp = (xnn_avgpool_mp_ukernel_function) xnn_f32_avgpool_ukernel_mp9p8q__wasm, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1224 | .mr = 9, |
| 1225 | .qr = 8, |
| 1226 | }; |
| 1227 | xnn_params.f32.pavgpool = (struct pavgpool_parameters) { |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 1228 | .up = (xnn_pavgpool_up_ukernel_function) xnn_f32_pavgpool_ukernel_up9__wasm, |
| 1229 | .mp = (xnn_pavgpool_mp_ukernel_function) xnn_f32_pavgpool_ukernel_mp9p8q__wasm, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1230 | .mr = 9, |
| 1231 | .qr = 8, |
| 1232 | }; |
| 1233 | xnn_params.f32.gavgpool = (struct gavgpool_parameters) { |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 1234 | .up = (xnn_gavgpool_up_ukernel_function) xnn_f32_gavgpool_ukernel_up7__wasm, |
| 1235 | .mp = (xnn_gavgpool_mp_ukernel_function) xnn_f32_gavgpool_ukernel_mp7p7q__wasm, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1236 | .mr = 7, |
| 1237 | }; |
| 1238 | xnn_params.f32.maxpool = (struct maxpool_parameters) { |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 1239 | .ukernel = (xnn_maxpool_ukernel_function) xnn_f32_maxpool_ukernel_9p8x__wasm_c1, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1240 | .mr = 9, |
| 1241 | .qr = 8, |
| 1242 | }; |
| 1243 | xnn_params.f32.argmaxpool[0] = (struct argmaxpool_parameters) { |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 1244 | .up = (xnn_argmaxpool_up_ukernel_function) xnn_f32_argmaxpool_ukernel_4x__scalar_c1, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1245 | .mr = 4, |
| 1246 | }; |
| 1247 | xnn_params.f32.argmaxpool[1] = (struct argmaxpool_parameters) { |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 1248 | .up = (xnn_argmaxpool_up_ukernel_function) xnn_f32_argmaxpool_ukernel_9x__scalar_c1, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1249 | .mr = 9, |
| 1250 | }; |
| 1251 | xnn_params.f32.argmaxpool[2] = (struct argmaxpool_parameters) { |
Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 1252 | .mp = (xnn_argmaxpool_mp_ukernel_function) xnn_f32_argmaxpool_ukernel_9p8x__scalar_c1, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1253 | .mr = 9, |
| 1254 | .qr = 8, |
| 1255 | }; |
Marat Dukhan | 6972249 | 2019-11-11 19:55:50 -0800 | [diff] [blame] | 1256 | xnn_params.f32.bilinear = (struct bilinear_parameters) { |
| 1257 | .ukernel = (xnn_bilinear_ukernel_function) xnn_f32_bilinear_ukernel__scalar_c2, |
| 1258 | .pixel_tile = 1, |
| 1259 | .channel_tile = 2, |
| 1260 | }; |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 1261 | xnn_params.f32.clamp = (xnn_univector_ukernel_function) xnn_f32_clamp_ukernel__wasm; |
| 1262 | xnn_params.f32.hswish = (xnn_univector_ukernel_function) xnn_f32_hswish_ukernel__wasm; |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1263 | xnn_params.f32.prelu = (struct prelu_parameters) { |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 1264 | .ukernel = (xnn_prelu_ukernel_function) xnn_f32_prelu_ukernel__wasm_2x4, |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 1265 | .row_tile = 4, |
| 1266 | .channel_tile = 4, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1267 | }; |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 1268 | xnn_params.f32.vadd = (struct vbinary_parameters) { |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 1269 | .op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vadd_ukernel__wasm_x4, |
| 1270 | .opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_ukernel__wasm_x4, |
| 1271 | .ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_ukernel__wasm_x4, |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 1272 | .element_tile = 8, |
| 1273 | }; |
Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 1274 | xnn_params.f32.vmax = (struct vbinary_parameters) { |
| 1275 | .op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmax_ukernel__wasm_x4, |
| 1276 | .opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__wasm_x4, |
| 1277 | .ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__wasm_x4, |
| 1278 | .element_tile = 8, |
| 1279 | }; |
| 1280 | xnn_params.f32.vmin = (struct vbinary_parameters) { |
| 1281 | .op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmin_ukernel__wasm_x4, |
| 1282 | .opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__wasm_x4, |
| 1283 | .ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__wasm_x4, |
| 1284 | .element_tile = 8, |
| 1285 | }; |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 1286 | xnn_params.f32.vmul = (struct vbinary_parameters) { |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 1287 | .op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmul_ukernel__wasm_x4, |
| 1288 | .opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_ukernel__wasm_x4, |
| 1289 | .ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_ukernel__wasm_x4, |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 1290 | .element_tile = 8, |
| 1291 | }; |
Marat Dukhan | 05f3f6d | 2019-12-03 15:13:53 -0800 | [diff] [blame] | 1292 | xnn_params.f32.vsub = (struct vbinary_parameters) { |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 1293 | .op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsub_ukernel__wasm_x4, |
| 1294 | .opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsubc_ukernel__wasm_x4, |
| 1295 | .ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrsubc_ukernel__wasm_x4, |
Marat Dukhan | 05f3f6d | 2019-12-03 15:13:53 -0800 | [diff] [blame] | 1296 | .element_tile = 8, |
| 1297 | }; |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1298 | xnn_params.f32.vmulcaddc = (struct vmulcaddc_parameters) { |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 1299 | .ukernel = (xnn_vmulcaddc_ukernel_function) xnn_f32_vmulcaddc_ukernel_c1__wasm_2x, |
Marat Dukhan | 49e6ee9 | 2019-11-06 15:55:29 -0800 | [diff] [blame] | 1300 | .channel_tile = 1, |
| 1301 | .row_tile = 2, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1302 | }; |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 1303 | #ifndef XNN_NO_NCHW_OPERATORS |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1304 | xnn_params.f32.spmm = (struct spmm_parameters) { |
Marat Dukhan | bff791e | 2019-10-24 11:05:37 -0700 | [diff] [blame] | 1305 | .ukernel = (xnn_spmm_ukernel_function) xnn_f32_spmm_ukernel_8x1__scalar, |
| 1306 | .mr = 8, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1307 | .nr = 1, |
| 1308 | }; |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 1309 | xnn_params.f32.spmm2 = (struct spmm_parameters) { |
| 1310 | .ukernel = (xnn_spmm_ukernel_function) xnn_f32_spmm_ukernel_8x2__scalar, |
| 1311 | .mr = 8, |
| 1312 | .nr = 2, |
| 1313 | }; |
| 1314 | xnn_params.f32.spmm4 = (struct spmm_parameters) { |
| 1315 | .ukernel = (xnn_spmm_ukernel_function) xnn_f32_spmm_ukernel_8x4__scalar, |
| 1316 | .mr = 8, |
| 1317 | .nr = 4, |
| 1318 | }; |
Marat Dukhan | 14fe0b2 | 2019-10-23 21:20:07 -0700 | [diff] [blame] | 1319 | xnn_params.f32.hwc2spchw_dconv3x3c3s2 = (struct hwc2spchw_dconv_parameters) { |
| 1320 | .ukernel_with_symm_padding = |
| 1321 | (xnn_conv_hwc2spchw_ukernel_function) xnn_f32_conv_hwc2spchw_ukernel_3x3s2p1c3x4__scalar_1x1, |
| 1322 | .output_channel_tile = 4, |
| 1323 | .output_height_tile = 1, |
| 1324 | .output_width_tile = 1, |
| 1325 | }; |
| 1326 | xnn_params.f32.spchw_dwconv3x3 = (struct spchw_dwconv_parameters) { |
| 1327 | .ukernel = (xnn_dwconv_spchw_ukernel_function) xnn_f32_dwconv_spchw_ukernel_3x3p1__scalar, |
| 1328 | .input_width_tile = 1, |
| 1329 | .output_width_tile = 1, |
| 1330 | .output_height_tile = 1, |
| 1331 | }; |
| 1332 | xnn_params.f32.spchw_dwconv3x3s2 = (struct spchw_dwconv_parameters) { |
| 1333 | .ukernel = (xnn_dwconv_spchw_ukernel_function) xnn_f32_dwconv_spchw_ukernel_3x3s2p1__scalar, |
| 1334 | .input_width_tile = 1, |
| 1335 | .output_width_tile = 1, |
| 1336 | .output_height_tile = 1, |
| 1337 | }; |
Marat Dukhan | a99918a | 2019-11-15 14:40:12 -0800 | [diff] [blame] | 1338 | xnn_params.f32.spchw_dwconv5x5 = (struct spchw_dwconv_parameters) { |
| 1339 | .ukernel = (xnn_dwconv_spchw_ukernel_function) xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar, |
| 1340 | .input_width_tile = 1, |
| 1341 | .output_width_tile = 1, |
| 1342 | .output_height_tile = 1, |
| 1343 | }; |
| 1344 | xnn_params.f32.spchw_dwconv5x5s2 = (struct spchw_dwconv_parameters) { |
| 1345 | .ukernel = (xnn_dwconv_spchw_ukernel_function) xnn_f32_dwconv_spchw_ukernel_5x5s2p2__scalar, |
| 1346 | .input_width_tile = 1, |
| 1347 | .output_width_tile = 1, |
| 1348 | .output_height_tile = 1, |
| 1349 | }; |
Marat Dukhan | 14fe0b2 | 2019-10-23 21:20:07 -0700 | [diff] [blame] | 1350 | xnn_params.f32.spchw_gavgpool = (struct spchw_gavgpool_parameters) { |
| 1351 | .ukernel = (xnn_gavgpool_spchw_ukernel_function) xnn_f32_gavgpool_spchw_ukernel__scalar_x1, |
| 1352 | .channel_tile = 1, |
| 1353 | }; |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 1354 | #endif // XNN_NO_NCHW_OPERATORS |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1355 | #endif // XNN_NO_F32_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1356 | |
| 1357 | /**************************** X32 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1358 | #ifndef XNN_NO_X32_OPERATORS |
| 1359 | xnn_params.x32.pad = (struct pad_parameters) { |
| 1360 | .ukernel = xnn_x32_pad_x2__scalar, |
| 1361 | .mr = 2, |
| 1362 | }; |
| 1363 | xnn_params.x32.unpool = (xnn_unpool_ukernel_function) xnn_x32_unpool_ukernel__scalar; |
| 1364 | xnn_params.x32.zip = (struct zip_parameters) { |
| 1365 | .x2 = (xnn_zipc_ukernel_function) xnn_x32_zip_x2_ukernel__scalar, |
| 1366 | .x3 = (xnn_zipc_ukernel_function) xnn_x32_zip_x3_ukernel__scalar, |
| 1367 | .x4 = (xnn_zipc_ukernel_function) xnn_x32_zip_x4_ukernel__scalar, |
| 1368 | .xm = (xnn_zipv_ukernel_function) xnn_x32_zip_xm_ukernel__scalar, |
| 1369 | }; |
| 1370 | #endif // XNN_NO_X32_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1371 | |
| 1372 | #else |
| 1373 | #error "Unsupported architecture" |
| 1374 | #endif |
| 1375 | xnn_params.initialized = true; |
| 1376 | } |
| 1377 | |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1378 | enum xnn_status xnn_initialize(const struct xnn_allocator* allocator) { |
Marat Dukhan | d343c22 | 2019-10-07 09:22:14 -0700 | [diff] [blame] | 1379 | #ifndef __EMSCRIPTEN__ |
| 1380 | if (!cpuinfo_initialize()) { |
| 1381 | return xnn_status_out_of_memory; |
| 1382 | } |
| 1383 | #endif |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1384 | pthread_once(&init_guard, &init); |
| 1385 | if (xnn_params.initialized) { |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 1386 | if (allocator != NULL) { |
| 1387 | memcpy(&xnn_params.allocator, allocator, sizeof(struct xnn_allocator)); |
| 1388 | } else { |
| 1389 | xnn_params.allocator.allocate = &xnn_allocate; |
| 1390 | xnn_params.allocator.reallocate = &xnn_reallocate; |
| 1391 | xnn_params.allocator.deallocate = &xnn_deallocate; |
| 1392 | xnn_params.allocator.aligned_allocate = &xnn_aligned_allocate; |
| 1393 | xnn_params.allocator.aligned_deallocate = &xnn_aligned_deallocate; |
| 1394 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1395 | return xnn_status_success; |
| 1396 | } else { |
| 1397 | return xnn_status_unsupported_hardware; |
| 1398 | } |
| 1399 | } |
| 1400 | |
| 1401 | enum xnn_status xnn_deinitialize(void) { |
Marat Dukhan | d343c22 | 2019-10-07 09:22:14 -0700 | [diff] [blame] | 1402 | #ifndef __EMSCRIPTEN__ |
| 1403 | cpuinfo_deinitialize(); |
| 1404 | #endif |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1405 | return xnn_status_success; |
| 1406 | } |