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 | |
Marat Dukhan | 0184901 | 2020-04-27 19:28:32 -0700 | [diff] [blame] | 9 | #include <math.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 10 | #include <stdbool.h> |
| 11 | #include <stddef.h> |
| 12 | #include <stdint.h> |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 13 | #include <string.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 14 | |
Marat Dukhan | 57133c0 | 2020-04-13 00:54:59 -0700 | [diff] [blame] | 15 | #ifdef _WIN32 |
| 16 | #include <windows.h> |
| 17 | #else |
| 18 | #include <pthread.h> |
| 19 | #endif |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 20 | |
Marat Dukhan | d343c22 | 2019-10-07 09:22:14 -0700 | [diff] [blame] | 21 | #ifndef __EMSCRIPTEN__ |
| 22 | #include <cpuinfo.h> |
| 23 | #endif |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 24 | |
| 25 | #include <xnnpack.h> |
| 26 | #include <xnnpack/argmaxpool.h> |
| 27 | #include <xnnpack/avgpool.h> |
| 28 | #include <xnnpack/clamp.h> |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 29 | #include <xnnpack/common.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 30 | #include <xnnpack/conv.h> |
| 31 | #include <xnnpack/dwconv.h> |
| 32 | #include <xnnpack/gavgpool.h> |
| 33 | #include <xnnpack/gemm.h> |
Marat Dukhan | 4662b19 | 2020-05-21 15:52:03 -0700 | [diff] [blame] | 34 | #include <xnnpack/fill.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 35 | #include <xnnpack/hswish.h> |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 36 | #include <xnnpack/ibilinear.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 37 | #include <xnnpack/igemm.h> |
| 38 | #include <xnnpack/log.h> |
| 39 | #include <xnnpack/lut.h> |
| 40 | #include <xnnpack/maxpool.h> |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 41 | #include <xnnpack/memory.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 42 | #include <xnnpack/pad.h> |
| 43 | #include <xnnpack/params.h> |
| 44 | #include <xnnpack/pavgpool.h> |
| 45 | #include <xnnpack/prelu.h> |
Marat Dukhan | 1edc454 | 2020-01-27 12:40:13 -0800 | [diff] [blame] | 46 | #include <xnnpack/raddstoreexpminusmax.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 47 | #include <xnnpack/rmax.h> |
| 48 | #include <xnnpack/spmm.h> |
| 49 | #include <xnnpack/unpool.h> |
| 50 | #include <xnnpack/vadd.h> |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 51 | #include <xnnpack/vbinary.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 52 | #include <xnnpack/vmulcaddc.h> |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 53 | #include <xnnpack/vunary.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 54 | #include <xnnpack/zip.h> |
| 55 | |
| 56 | #ifndef XNN_ENABLE_ASSEMBLY |
| 57 | #define XNN_ENABLE_ASSEMBLY 1 |
| 58 | #endif |
| 59 | |
Marat Dukhan | 57133c0 | 2020-04-13 00:54:59 -0700 | [diff] [blame] | 60 | #ifdef _WIN32 |
| 61 | static INIT_ONCE init_guard = INIT_ONCE_STATIC_INIT; |
| 62 | #else |
| 63 | static pthread_once_t init_guard = PTHREAD_ONCE_INIT; |
| 64 | #endif |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 65 | |
| 66 | struct xnn_parameters xnn_params = { |
Marat Dukhan | 854fb6b | 2020-06-19 12:33:44 -0700 | [diff] [blame] | 67 | .init_flags = 0 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 68 | }; |
| 69 | |
Marat Dukhan | 0184901 | 2020-04-27 19:28:32 -0700 | [diff] [blame] | 70 | static void init(void) { |
Marat Dukhan | a199d49 | 2020-07-24 15:01:25 -0700 | [diff] [blame] | 71 | #if XNN_ARCH_WASM || XNN_ARCH_WASMSIMD |
Marat Dukhan | 0184901 | 2020-04-27 19:28:32 -0700 | [diff] [blame] | 72 | // Unlike most other architectures, on x86/x86-64 when floating-point instructions |
| 73 | // have no NaN arguments, but produce NaN output, the output NaN has sign bit set. |
| 74 | // We use it to distinguish x86/x86-64 from other architectures, by doing subtraction |
| 75 | // of two infinities (must produce NaN per IEEE 754 standard). |
| 76 | static const volatile float inf = INFINITY; |
| 77 | const bool is_wasm_x86 = signbit(inf - inf); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 78 | #endif |
Marat Dukhan | 854fb6b | 2020-06-19 12:33:44 -0700 | [diff] [blame] | 79 | uint32_t init_flags = XNN_INIT_FLAG_XNNPACK; |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 80 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 81 | #if XNN_ARCH_ARM |
Frank Barchard | bcdb1c1 | 2020-05-11 14:13:20 -0700 | [diff] [blame] | 82 | #if XNN_PLATFORM_MOBILE |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 83 | if (!cpuinfo_has_arm_neon()) { |
| 84 | xnn_log_error("XNNPACK initialization failed: NEON is not supported"); |
| 85 | return; |
| 86 | } |
| 87 | #else |
| 88 | if (!cpuinfo_has_arm_vfpv2() && !cpuinfo_has_arm_vfpv3()) { |
| 89 | xnn_log_error("XNNPACK initialization failed: VFP is not supported"); |
| 90 | return; |
| 91 | } |
| 92 | #endif |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 93 | |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 94 | if (cpuinfo_has_arm_neon()) { |
Marat Dukhan | f28cddf | 2020-08-10 14:05:02 -0700 | [diff] [blame] | 95 | /**************************** QS8 micro-kernels ****************************/ |
| 96 | #ifndef XNN_NO_QS8_OPERATORS |
| 97 | init_flags |= XNN_INIT_FLAG_QS8; |
| 98 | |
Frank Barchard | 146e999 | 2020-10-13 13:38:54 -0700 | [diff] [blame] | 99 | xnn_params.qs8.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane); |
| 100 | xnn_params.qs8.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_2x8__neon_mlal_lane); |
Marat Dukhan | f28cddf | 2020-08-10 14:05:02 -0700 | [diff] [blame] | 101 | xnn_params.qs8.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_1x8__neon_mlal_lane); |
| 102 | xnn_params.qs8.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_1x8__neon_mlal_lane); |
Frank Barchard | 146e999 | 2020-10-13 13:38:54 -0700 | [diff] [blame] | 103 | xnn_params.qs8.gemm.mr = 2; |
Marat Dukhan | f28cddf | 2020-08-10 14:05:02 -0700 | [diff] [blame] | 104 | xnn_params.qs8.gemm.nr = 8; |
| 105 | |
| 106 | xnn_params.qs8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qs8_dwconv_minmax_ukernel_up8x9__neon_mul16; |
| 107 | xnn_params.qs8.dwconv[0].channel_tile = 8; |
| 108 | xnn_params.qs8.dwconv[0].primary_tile = 9; |
| 109 | |
| 110 | xnn_params.qs8.gavgpool = (struct gavgpool_parameters) { |
| 111 | .up = (xnn_gavgpool_unipass_ukernel_function) xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c8_acc2, |
| 112 | .mp = (xnn_gavgpool_multipass_ukernel_function) xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c8_acc2, |
| 113 | .mr = 7, |
| 114 | }; |
Marat Dukhan | ff20948 | 2020-09-03 14:26:53 -0700 | [diff] [blame] | 115 | |
| 116 | xnn_params.qs8.vadd = (struct vbinary_parameters) { |
| 117 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vadd_minmax_ukernel__neon_ld64_x8, |
| 118 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x8, |
| 119 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x8, |
| 120 | .element_tile = 8, |
| 121 | }; |
Marat Dukhan | f28cddf | 2020-08-10 14:05:02 -0700 | [diff] [blame] | 122 | #endif // XNN_NO_QS8_OPERATORS |
| 123 | |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 124 | /*************************** QU8 micro-kernels ***************************/ |
| 125 | #ifndef XNN_NO_QU8_OPERATORS |
| 126 | init_flags |= XNN_INIT_FLAG_QU8; |
Marat Dukhan | 854fb6b | 2020-06-19 12:33:44 -0700 | [diff] [blame] | 127 | |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 128 | xnn_params.qu8.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_ukernel_4x8__neon); |
| 129 | xnn_params.qu8.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_ukernel_4x8__neon); |
| 130 | xnn_params.qu8.gemm.mr = 4; |
| 131 | xnn_params.qu8.gemm.nr = 8; |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 132 | |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 133 | xnn_params.qu8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qu8_dwconv_minmax_ukernel_up8x9__neon; |
| 134 | xnn_params.qu8.dwconv[0].channel_tile = 8; |
| 135 | xnn_params.qu8.dwconv[0].primary_tile = 9; |
| 136 | xnn_params.qu8.avgpool = (struct avgpool_parameters) { |
| 137 | .up = (xnn_avgpool_unipass_ukernel_function) xnn_qu8_avgpool_minmax_ukernel_9x__neon_c8, |
| 138 | .mp = (xnn_avgpool_multipass_ukernel_function) xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8, |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 139 | .mr = 9, |
| 140 | .qr = 8, |
| 141 | }; |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 142 | xnn_params.qu8.gavgpool = (struct gavgpool_parameters) { |
| 143 | .up = (xnn_gavgpool_unipass_ukernel_function) xnn_qu8_gavgpool_minmax_ukernel_7x__neon_c8, |
| 144 | .mp = (xnn_gavgpool_multipass_ukernel_function) xnn_qu8_gavgpool_minmax_ukernel_7p7x__neon_c8, |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 145 | .mr = 7, |
| 146 | }; |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 147 | xnn_params.qu8.vadd = (xnn_vadd_ukernel_function) xnn_qu8_vadd_minmax_ukernel__neon; |
| 148 | #endif // XNN_NO_QU8_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 149 | |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 150 | /**************************** U8 micro-kernels ****************************/ |
| 151 | #ifndef XNN_NO_U8_OPERATORS |
Marat Dukhan | 854fb6b | 2020-06-19 12:33:44 -0700 | [diff] [blame] | 152 | init_flags |= XNN_INIT_FLAG_U8; |
| 153 | |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 154 | xnn_params.u8.maxpool = (struct maxpool_parameters) { |
| 155 | .ukernel = (xnn_maxpool_ukernel_function) xnn_u8_maxpool_minmax_ukernel_9p8x__neon_c16, |
| 156 | .mr = 9, |
| 157 | .qr = 8, |
| 158 | }; |
| 159 | xnn_params.u8.clamp = (xnn_univector_ukernel_function) xnn_u8_clamp_ukernel__neon_x64; |
| 160 | xnn_params.u8.rmax = xnn_u8_rmax_ukernel__neon; |
| 161 | xnn_params.u8.lut32norm = xnn_u8_lut32norm_ukernel__scalar; |
| 162 | #endif // XNN_NO_U8_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 163 | |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 164 | /**************************** X8 micro-kernels ****************************/ |
| 165 | #ifndef XNN_NO_X8_OPERATORS |
Marat Dukhan | 854fb6b | 2020-06-19 12:33:44 -0700 | [diff] [blame] | 166 | init_flags |= XNN_INIT_FLAG_X8; |
| 167 | |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 168 | xnn_params.x8.lut = xnn_x8_lut_ukernel__scalar; |
| 169 | xnn_params.x8.zip = (struct zip_parameters) { |
| 170 | .x2 = (xnn_zipc_ukernel_function) xnn_x8_zip_x2_ukernel__neon, |
| 171 | .x3 = (xnn_zipc_ukernel_function) xnn_x8_zip_x3_ukernel__neon, |
| 172 | .x4 = (xnn_zipc_ukernel_function) xnn_x8_zip_x4_ukernel__neon, |
| 173 | .xm = (xnn_zipv_ukernel_function) xnn_x8_zip_xm_ukernel__neon, |
| 174 | }; |
| 175 | #endif // XNN_NO_X8_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 176 | |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 177 | /**************************** F32 micro-kernels ****************************/ |
| 178 | #ifndef XNN_NO_F32_OPERATORS |
Marat Dukhan | 854fb6b | 2020-06-19 12:33:44 -0700 | [diff] [blame] | 179 | init_flags |= XNN_INIT_FLAG_F32; |
| 180 | |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 181 | #if XNN_ENABLE_ASSEMBLY |
| 182 | switch (cpuinfo_get_uarch(0)->uarch) { |
| 183 | case cpuinfo_uarch_cortex_a5: |
| 184 | case cpuinfo_uarch_cortex_a7: |
Frank Barchard | 490febe | 2020-07-16 18:42:17 -0700 | [diff] [blame] | 185 | xnn_params.f32.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_cortex_a7); |
| 186 | xnn_params.f32.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x8__aarch32_neon_cortex_a7); |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 187 | xnn_params.f32.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__neon_lane_ld64); |
| 188 | xnn_params.f32.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__neon_lane_ld64); |
| 189 | xnn_params.f32.gemm.mr = 4; |
| 190 | xnn_params.f32.gemm.nr = 8; |
Marat Dukhan | 05702cf | 2020-03-26 15:41:33 -0700 | [diff] [blame] | 191 | break; |
Marat Dukhan | 05702cf | 2020-03-26 15:41:33 -0700 | [diff] [blame] | 192 | |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 193 | case cpuinfo_uarch_cortex_a53: |
| 194 | case cpuinfo_uarch_cortex_a55r0: |
| 195 | xnn_params.f32.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_cortex_a53); |
| 196 | xnn_params.f32.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x8__aarch32_neon_cortex_a53); |
| 197 | xnn_params.f32.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__neon_lane_ld64); |
| 198 | xnn_params.f32.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__neon_lane_ld64); |
| 199 | xnn_params.f32.gemm.mr = 4; |
| 200 | xnn_params.f32.gemm.nr = 8; |
| 201 | break; |
| 202 | |
| 203 | case cpuinfo_uarch_cortex_a55: |
| 204 | xnn_params.f32.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_cortex_a55); |
| 205 | xnn_params.f32.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x8__aarch32_neon_cortex_a55); |
| 206 | xnn_params.f32.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__neon_lane_ld64); |
| 207 | xnn_params.f32.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__neon_lane_ld64); |
| 208 | xnn_params.f32.gemm.mr = 4; |
| 209 | xnn_params.f32.gemm.nr = 8; |
| 210 | break; |
| 211 | |
| 212 | case cpuinfo_uarch_cortex_a57: |
| 213 | case cpuinfo_uarch_cortex_a72: |
| 214 | case cpuinfo_uarch_cortex_a73: |
| 215 | xnn_params.f32.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_pld_cortex_a75); |
| 216 | xnn_params.f32.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x8__aarch32_neon_pld_cortex_a75); |
| 217 | xnn_params.f32.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__neon_lane_ld64); |
| 218 | xnn_params.f32.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__neon_lane_ld64); |
| 219 | xnn_params.f32.gemm.mr = 4; |
| 220 | xnn_params.f32.gemm.nr = 8; |
| 221 | break; |
| 222 | |
| 223 | case cpuinfo_uarch_krait: |
| 224 | default: |
| 225 | xnn_params.f32.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_cortex_a75); |
| 226 | xnn_params.f32.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x8__aarch32_neon_cortex_a75); |
| 227 | xnn_params.f32.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__neon_lane_ld64); |
| 228 | xnn_params.f32.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__neon_lane_ld64); |
| 229 | xnn_params.f32.gemm.mr = 4; |
| 230 | xnn_params.f32.gemm.nr = 8; |
| 231 | break; |
| 232 | } |
| 233 | #if XNN_MAX_UARCH_TYPES > 1 |
| 234 | { |
| 235 | /* Choose micro-kernels for little cores according to micro-kernel specification for the big core */ |
| 236 | const uint32_t mr = xnn_params.f32.gemm.mr; |
| 237 | const uint32_t nr = xnn_params.f32.gemm.nr; |
| 238 | const uint32_t log2_sr = xnn_params.f32.gemm.log2_sr; |
| 239 | for (size_t i = 1; i < XNN_MAX_UARCH_TYPES; i++) { |
| 240 | const struct cpuinfo_uarch_info* uarch_info = cpuinfo_get_uarch(i); |
| 241 | if (uarch_info == NULL) { |
| 242 | /* No more microarchitectures in the system */ |
Marat Dukhan | 05702cf | 2020-03-26 15:41:33 -0700 | [diff] [blame] | 243 | break; |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 244 | } |
| 245 | |
| 246 | switch (uarch_info->uarch) { |
| 247 | case cpuinfo_uarch_cortex_a53: |
| 248 | case cpuinfo_uarch_cortex_a55r0: |
| 249 | if (mr == 4 && nr == 8 && log2_sr == 0) { |
| 250 | xnn_params.f32.gemm.minmax.gemm.function[i] = (xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_cortex_a53; |
| 251 | xnn_params.f32.gemm.minmax.igemm.function[i] = (xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x8__aarch32_neon_cortex_a53; |
| 252 | xnn_params.f32.gemm.minmax.gemm1.function[i] = (xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__neon_lane_ld64; |
| 253 | xnn_params.f32.gemm.minmax.igemm1.function[i] = (xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__neon_lane_ld64; |
| 254 | } |
| 255 | break; |
| 256 | case cpuinfo_uarch_cortex_a55: |
| 257 | if (mr == 4 && nr == 8 && log2_sr == 0) { |
| 258 | xnn_params.f32.gemm.minmax.gemm.function[i] = (xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_cortex_a55; |
| 259 | xnn_params.f32.gemm.minmax.igemm.function[i] = (xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x8__aarch32_neon_cortex_a55; |
| 260 | xnn_params.f32.gemm.minmax.gemm1.function[i] = (xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__neon_lane_ld64; |
| 261 | xnn_params.f32.gemm.minmax.igemm1.function[i] = (xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__neon_lane_ld64; |
| 262 | } |
| 263 | break; |
| 264 | default: |
| 265 | break; |
| 266 | } |
Marat Dukhan | 05702cf | 2020-03-26 15:41:33 -0700 | [diff] [blame] | 267 | } |
| 268 | } |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 269 | #endif // XNN_MAX_UARCH_TYPES > 1 |
| 270 | #else // XNN_ENABLE_ASSEMBLY |
| 271 | xnn_params.f32.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x8__neon_lane_ld128); |
| 272 | xnn_params.f32.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x8__neon_lane_ld128); |
| 273 | xnn_params.f32.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__neon_lane_ld64); |
| 274 | xnn_params.f32.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__neon_lane_ld64); |
| 275 | xnn_params.f32.gemm.mr = 4; |
| 276 | xnn_params.f32.gemm.nr = 8; |
| 277 | #endif // XNN_ENABLE_ASSEMBLY |
| 278 | xnn_params.f32.gemm2.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x2__neon_lane_ld64); |
| 279 | xnn_params.f32.gemm2.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x2__neon_lane_ld64); |
| 280 | xnn_params.f32.gemm2.mr = 4; |
| 281 | xnn_params.f32.gemm2.nr = 2; |
| 282 | |
| 283 | xnn_params.f32.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up4x4__neon; |
| 284 | xnn_params.f32.dwconv[0].channel_tile = 4, |
| 285 | xnn_params.f32.dwconv[0].primary_tile = 4, |
| 286 | |
| 287 | xnn_params.f32.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up4x9__neon; |
| 288 | xnn_params.f32.dwconv[1].channel_tile = 4; |
| 289 | xnn_params.f32.dwconv[1].primary_tile = 9; |
| 290 | |
| 291 | xnn_params.f32.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2; |
| 292 | xnn_params.f32.dwconv[2].channel_tile = 4; |
| 293 | xnn_params.f32.dwconv[2].primary_tile = 25; |
| 294 | |
| 295 | xnn_params.f32.avgpool = (struct avgpool_parameters) { |
| 296 | .up = (xnn_avgpool_unipass_ukernel_function) xnn_f32_avgpool_minmax_ukernel_9x__neon_c4, |
| 297 | .mp = (xnn_avgpool_multipass_ukernel_function) xnn_f32_avgpool_minmax_ukernel_9p8x__neon_c4, |
| 298 | .mr = 9, |
| 299 | .qr = 8, |
| 300 | }; |
| 301 | xnn_params.f32.pavgpool = (struct pavgpool_parameters) { |
| 302 | .up = (xnn_pavgpool_unipass_ukernel_function) xnn_f32_pavgpool_minmax_ukernel_9x__neon_c4, |
| 303 | .mp = (xnn_pavgpool_multipass_ukernel_function) xnn_f32_pavgpool_minmax_ukernel_9p8x__neon_c4, |
| 304 | .mr = 9, |
| 305 | .qr = 8, |
| 306 | }; |
| 307 | xnn_params.f32.gavgpool = (struct gavgpool_parameters) { |
| 308 | .up = (xnn_gavgpool_unipass_ukernel_function) xnn_f32_gavgpool_minmax_ukernel_7x__neon_c4, |
| 309 | .mp = (xnn_gavgpool_multipass_ukernel_function) xnn_f32_gavgpool_minmax_ukernel_7p7x__neon_c4, |
| 310 | .mr = 7, |
| 311 | }; |
| 312 | xnn_params.f32.maxpool = (struct maxpool_parameters) { |
| 313 | .ukernel = (xnn_maxpool_ukernel_function) xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4, |
| 314 | .mr = 9, |
| 315 | .qr = 8, |
| 316 | }; |
| 317 | xnn_params.f32.argmaxpool[0] = (struct argmaxpool_parameters) { |
Marat Dukhan | ef25c6d | 2020-07-24 00:59:40 -0700 | [diff] [blame] | 318 | .up = (xnn_argmaxpool_unipass_ukernel_function) xnn_f32_argmaxpool_ukernel_4x__neon_c4, |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 319 | .mr = 4, |
| 320 | }; |
| 321 | xnn_params.f32.argmaxpool[1] = (struct argmaxpool_parameters) { |
Marat Dukhan | ef25c6d | 2020-07-24 00:59:40 -0700 | [diff] [blame] | 322 | .up = (xnn_argmaxpool_unipass_ukernel_function) xnn_f32_argmaxpool_ukernel_9x__neon_c4, |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 323 | .mr = 9, |
| 324 | }; |
| 325 | xnn_params.f32.argmaxpool[2] = (struct argmaxpool_parameters) { |
Marat Dukhan | ef25c6d | 2020-07-24 00:59:40 -0700 | [diff] [blame] | 326 | .mp = (xnn_argmaxpool_multipass_ukernel_function) xnn_f32_argmaxpool_ukernel_9p8x__neon_c4, |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 327 | .mr = 9, |
| 328 | .qr = 8, |
| 329 | }; |
| 330 | xnn_params.f32.ibilinear = (struct ibilinear_parameters) { |
| 331 | .ukernel = (xnn_ibilinear_ukernel_function) xnn_f32_ibilinear_ukernel__neon_c8, |
| 332 | .pixel_tile = 1, |
| 333 | .channel_tile = 8, |
| 334 | }; |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 335 | xnn_params.f32.abs = (xnn_univector_ukernel_function) xnn_f32_vabs_ukernel__neon_x8; |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 336 | xnn_params.f32.clamp = (xnn_univector_ukernel_function) xnn_f32_clamp_ukernel__neon_x8; |
Marat Dukhan | 55dde5b | 2020-07-10 22:48:54 -0700 | [diff] [blame] | 337 | xnn_params.f32.hswish = (xnn_univector_ukernel_function) xnn_f32_hswish_ukernel__neon_x16; |
Marat Dukhan | 2881333 | 2020-06-10 18:05:38 -0700 | [diff] [blame] | 338 | xnn_params.f32.lrelu = (xnn_univector_ukernel_function) xnn_f32_vlrelu_ukernel__neon_x8; |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 339 | xnn_params.f32.neg = (xnn_univector_ukernel_function) xnn_f32_vneg_ukernel__neon_x8; |
Marat Dukhan | 64e5251 | 2020-06-09 13:41:16 -0700 | [diff] [blame] | 340 | if (cpuinfo_has_arm_neon_v8()) { |
| 341 | xnn_params.f32.rndne = (xnn_univector_ukernel_function) xnn_f32_vrndne_ukernel__neonv8_x8; |
| 342 | xnn_params.f32.rndz = (xnn_univector_ukernel_function) xnn_f32_vrndz_ukernel__neonv8_x8; |
| 343 | xnn_params.f32.rndu = (xnn_univector_ukernel_function) xnn_f32_vrndu_ukernel__neonv8_x8; |
| 344 | xnn_params.f32.rndd = (xnn_univector_ukernel_function) xnn_f32_vrndd_ukernel__neonv8_x8; |
| 345 | } else { |
| 346 | xnn_params.f32.rndne = (xnn_univector_ukernel_function) xnn_f32_vrndne_ukernel__neon_x8; |
| 347 | xnn_params.f32.rndz = (xnn_univector_ukernel_function) xnn_f32_vrndz_ukernel__neon_x8; |
| 348 | xnn_params.f32.rndu = (xnn_univector_ukernel_function) xnn_f32_vrndu_ukernel__neon_x8; |
| 349 | xnn_params.f32.rndd = (xnn_univector_ukernel_function) xnn_f32_vrndd_ukernel__neon_x8; |
| 350 | } |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 351 | xnn_params.f32.sigmoid = (xnn_univector_ukernel_function) xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x8; |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 352 | xnn_params.f32.sqr = (xnn_univector_ukernel_function) xnn_f32_vsqr_ukernel__neon_x8; |
Marat Dukhan | 6804bbd | 2020-06-30 19:26:11 -0700 | [diff] [blame] | 353 | xnn_params.f32.sqrt = (xnn_univector_ukernel_function) xnn_f32_vsqrt_ukernel__scalar_sqrt_x1; |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 354 | xnn_params.f32.prelu = (struct prelu_parameters) { |
| 355 | .ukernel = (xnn_prelu_ukernel_function) xnn_f32_prelu_ukernel__neon_2x8, |
| 356 | .row_tile = 2, |
| 357 | .channel_tile = 8, |
| 358 | }; |
| 359 | xnn_params.f32.raddstoreexpminusmax = xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8; |
| 360 | xnn_params.f32.rmax = xnn_f32_rmax_ukernel__neon; |
| 361 | xnn_params.f32.vadd = (struct vbinary_parameters) { |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 362 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vadd_minmax_ukernel__neon_x8, |
| 363 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_minmax_ukernel__neon_x8, |
| 364 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_minmax_ukernel__neon_x8, |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 365 | .element_tile = 8, |
| 366 | }; |
| 367 | xnn_params.f32.vdiv = (struct vbinary_parameters) { |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 368 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdiv_minmax_ukernel__scalar_x2, |
| 369 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdivc_minmax_ukernel__scalar_x2, |
| 370 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrdivc_minmax_ukernel__scalar_x2, |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 371 | .element_tile = 2, |
| 372 | }; |
| 373 | xnn_params.f32.vmax = (struct vbinary_parameters) { |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 374 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmax_ukernel__neon_x8, |
| 375 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__neon_x8, |
| 376 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__neon_x8, |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 377 | .element_tile = 8, |
| 378 | }; |
| 379 | xnn_params.f32.vmin = (struct vbinary_parameters) { |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 380 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmin_ukernel__neon_x8, |
| 381 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__neon_x8, |
| 382 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__neon_x8, |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 383 | .element_tile = 8, |
| 384 | }; |
| 385 | xnn_params.f32.vmul = (struct vbinary_parameters) { |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 386 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmul_minmax_ukernel__neon_x8, |
| 387 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_minmax_ukernel__neon_x8, |
| 388 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_minmax_ukernel__neon_x8, |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 389 | .element_tile = 8, |
| 390 | }; |
| 391 | xnn_params.f32.vsub = (struct vbinary_parameters) { |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 392 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsub_minmax_ukernel__neon_x8, |
| 393 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsubc_minmax_ukernel__neon_x8, |
| 394 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrsubc_minmax_ukernel__neon_x8, |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 395 | .element_tile = 8, |
| 396 | }; |
Marat Dukhan | f739926 | 2020-06-05 10:58:44 -0700 | [diff] [blame] | 397 | xnn_params.f32.vsqrdiff = (struct vbinary_parameters) { |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 398 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiff_ukernel__neon_x8, |
| 399 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiffc_ukernel__neon_x8, |
| 400 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiffc_ukernel__neon_x8, |
Marat Dukhan | f739926 | 2020-06-05 10:58:44 -0700 | [diff] [blame] | 401 | .element_tile = 8, |
| 402 | }; |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 403 | xnn_params.f32.vmulcaddc = (struct vmulcaddc_parameters) { |
Marat Dukhan | 9531e9f | 2020-07-24 15:25:02 -0700 | [diff] [blame] | 404 | .ukernel = (xnn_vmulcaddc_ukernel_function) xnn_f32_vmulcaddc_minmax_ukernel_c4__neon_2x, |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 405 | .channel_tile = 4, |
| 406 | .row_tile = 2, |
| 407 | }; |
| 408 | #endif // XNN_NO_F32_OPERATORS |
| 409 | |
| 410 | /**************************** X32 micro-kernels ****************************/ |
| 411 | #ifndef XNN_NO_X32_OPERATORS |
Marat Dukhan | 854fb6b | 2020-06-19 12:33:44 -0700 | [diff] [blame] | 412 | init_flags |= XNN_INIT_FLAG_X32; |
| 413 | |
Marat Dukhan | 4662b19 | 2020-05-21 15:52:03 -0700 | [diff] [blame] | 414 | xnn_params.x32.fill = (struct fill_parameters) { |
| 415 | .ukernel = (xnn_fill_ukernel_function) xnn_x32_fill_ukernel__neon, |
| 416 | .row_tile = 1, |
| 417 | }; |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 418 | xnn_params.x32.pad = (struct pad_parameters) { |
Marat Dukhan | 63523d4 | 2020-05-22 17:07:33 -0700 | [diff] [blame] | 419 | .ukernel = (xnn_pad_ukernel_function) xnn_x32_pad_ukernel__neon, |
| 420 | .row_tile = 1, |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 421 | }; |
| 422 | xnn_params.x32.unpool = (xnn_unpool_ukernel_function) xnn_x32_unpool_ukernel__neon; |
| 423 | xnn_params.x32.zip = (struct zip_parameters) { |
| 424 | .x2 = (xnn_zipc_ukernel_function) xnn_x32_zip_x2_ukernel__neon, |
| 425 | .x3 = (xnn_zipc_ukernel_function) xnn_x32_zip_x3_ukernel__neon, |
| 426 | .x4 = (xnn_zipc_ukernel_function) xnn_x32_zip_x4_ukernel__neon, |
| 427 | .xm = (xnn_zipv_ukernel_function) xnn_x32_zip_xm_ukernel__neon, |
| 428 | }; |
| 429 | #endif // XNN_NO_X32_OPERATORS |
| 430 | } else if (!XNN_PLATFORM_MOBILE) { |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 431 | /*************************** QU8 micro-kernels ***************************/ |
| 432 | #ifndef XNN_NO_QU8_OPERATORS |
| 433 | init_flags |= XNN_INIT_FLAG_QU8; |
Marat Dukhan | 854fb6b | 2020-06-19 12:33:44 -0700 | [diff] [blame] | 434 | |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 435 | xnn_params.qu8.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_ukernel_2x2__scalar); |
| 436 | xnn_params.qu8.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_ukernel_2x2__scalar); |
| 437 | xnn_params.qu8.gemm.mr = 2; |
| 438 | xnn_params.qu8.gemm.nr = 2; |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 439 | |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 440 | xnn_params.qu8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qu8_dwconv_minmax_ukernel_up1x9__scalar; |
| 441 | xnn_params.qu8.dwconv[0].channel_tile = 1; |
| 442 | xnn_params.qu8.dwconv[0].primary_tile = 9; |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 443 | |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 444 | xnn_params.qu8.avgpool = (struct avgpool_parameters) { |
| 445 | .up = (xnn_avgpool_unipass_ukernel_function) xnn_qu8_avgpool_minmax_ukernel_9x__scalar_c1, |
| 446 | .mp = (xnn_avgpool_multipass_ukernel_function) xnn_qu8_avgpool_minmax_ukernel_9p8x__scalar_c1, |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 447 | .mr = 9, |
| 448 | .qr = 8, |
| 449 | }; |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 450 | xnn_params.qu8.gavgpool = (struct gavgpool_parameters) { |
| 451 | .up = (xnn_gavgpool_unipass_ukernel_function) xnn_qu8_gavgpool_minmax_ukernel_7x__scalar_c1, |
| 452 | .mp = (xnn_gavgpool_multipass_ukernel_function) xnn_qu8_gavgpool_minmax_ukernel_7p7x__scalar_c1, |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 453 | .mr = 7, |
| 454 | }; |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 455 | xnn_params.qu8.vadd = (xnn_vadd_ukernel_function) xnn_qu8_vadd_minmax_ukernel__scalar; |
| 456 | #endif // XNN_NO_QU8_OPERATORS |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 457 | |
| 458 | /**************************** U8 micro-kernels ****************************/ |
| 459 | #ifndef XNN_NO_U8_OPERATORS |
Marat Dukhan | 854fb6b | 2020-06-19 12:33:44 -0700 | [diff] [blame] | 460 | init_flags |= XNN_INIT_FLAG_U8; |
| 461 | |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 462 | xnn_params.u8.maxpool = (struct maxpool_parameters) { |
| 463 | .ukernel = (xnn_maxpool_ukernel_function) xnn_u8_maxpool_minmax_ukernel_9p8x__scalar_c1, |
| 464 | .mr = 9, |
| 465 | .qr = 8, |
| 466 | }; |
| 467 | xnn_params.u8.clamp = (xnn_univector_ukernel_function) xnn_u8_clamp_ukernel__scalar_x4; |
| 468 | xnn_params.u8.lut32norm = xnn_u8_lut32norm_ukernel__scalar; |
| 469 | xnn_params.u8.rmax = xnn_u8_rmax_ukernel__scalar; |
| 470 | #endif // XNN_NO_U8_OPERATORS |
| 471 | |
| 472 | /**************************** X8 micro-kernels ****************************/ |
| 473 | #ifndef XNN_NO_X8_OPERATORS |
Marat Dukhan | 854fb6b | 2020-06-19 12:33:44 -0700 | [diff] [blame] | 474 | init_flags |= XNN_INIT_FLAG_X8; |
| 475 | |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 476 | xnn_params.x8.lut = xnn_x8_lut_ukernel__scalar; |
| 477 | xnn_params.x8.zip = (struct zip_parameters) { |
| 478 | .x2 = (xnn_zipc_ukernel_function) xnn_x8_zip_x2_ukernel__scalar, |
| 479 | .x3 = (xnn_zipc_ukernel_function) xnn_x8_zip_x3_ukernel__scalar, |
| 480 | .x4 = (xnn_zipc_ukernel_function) xnn_x8_zip_x4_ukernel__scalar, |
| 481 | .xm = (xnn_zipv_ukernel_function) xnn_x8_zip_xm_ukernel__scalar, |
| 482 | }; |
| 483 | #endif // XNN_NO_X8_OPERATORS |
| 484 | |
| 485 | /**************************** F32 micro-kernels ****************************/ |
| 486 | #ifndef XNN_NO_F32_OPERATORS |
Marat Dukhan | 854fb6b | 2020-06-19 12:33:44 -0700 | [diff] [blame] | 487 | init_flags |= XNN_INIT_FLAG_F32; |
| 488 | |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 489 | xnn_params.f32.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x4__scalar); |
| 490 | xnn_params.f32.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x4__scalar); |
| 491 | xnn_params.f32.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x4__scalar); |
| 492 | xnn_params.f32.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x4__scalar); |
Marat Dukhan | 467f636 | 2020-05-22 23:21:55 -0700 | [diff] [blame] | 493 | xnn_params.f32.gemm.relu.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_relu_ukernel_4x4__scalar); |
| 494 | xnn_params.f32.gemm.relu.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_relu_ukernel_4x4__scalar); |
| 495 | xnn_params.f32.gemm.relu.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_relu_ukernel_1x4__scalar); |
| 496 | xnn_params.f32.gemm.relu.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_relu_ukernel_1x4__scalar); |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 497 | xnn_params.f32.gemm.linear.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_4x4__scalar); |
| 498 | xnn_params.f32.gemm.linear.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_4x4__scalar); |
| 499 | xnn_params.f32.gemm.linear.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_1x4__scalar); |
| 500 | xnn_params.f32.gemm.linear.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_1x4__scalar); |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 501 | xnn_params.f32.gemm.mr = 4; |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 502 | xnn_params.f32.gemm.nr = 4; |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 503 | |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 504 | xnn_params.f32.gemm2.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x2__scalar); |
| 505 | xnn_params.f32.gemm2.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x2__scalar), |
| 506 | xnn_params.f32.gemm2.linear.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_4x2__scalar); |
| 507 | xnn_params.f32.gemm2.linear.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_4x2__scalar), |
| 508 | xnn_params.f32.gemm2.mr = 4; |
| 509 | xnn_params.f32.gemm2.nr = 2; |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 510 | |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 511 | xnn_params.f32.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up1x4__scalar_acc2; |
| 512 | xnn_params.f32.dwconv[0].linear.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_ukernel_up1x4__scalar_acc2; |
| 513 | xnn_params.f32.dwconv[0].channel_tile = 1; |
| 514 | xnn_params.f32.dwconv[0].primary_tile = 4; |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 515 | |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 516 | xnn_params.f32.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up1x9__scalar_acc2; |
| 517 | xnn_params.f32.dwconv[1].linear.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_ukernel_up1x9__scalar_acc2; |
| 518 | xnn_params.f32.dwconv[1].channel_tile = 1; |
| 519 | xnn_params.f32.dwconv[1].primary_tile = 9; |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 520 | |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 521 | xnn_params.f32.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up1x25__scalar_acc2; |
| 522 | xnn_params.f32.dwconv[2].linear.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_ukernel_up1x25__scalar_acc2; |
| 523 | xnn_params.f32.dwconv[2].channel_tile = 1; |
| 524 | xnn_params.f32.dwconv[2].primary_tile = 25; |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 525 | |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 526 | xnn_params.f32.avgpool = (struct avgpool_parameters) { |
| 527 | .up = (xnn_avgpool_unipass_ukernel_function) xnn_f32_avgpool_minmax_ukernel_9x__scalar_c1, |
| 528 | .mp = (xnn_avgpool_multipass_ukernel_function) xnn_f32_avgpool_minmax_ukernel_9p8x__scalar_c1, |
| 529 | .mr = 9, |
| 530 | .qr = 8, |
| 531 | }; |
| 532 | xnn_params.f32.pavgpool = (struct pavgpool_parameters) { |
| 533 | .up = (xnn_pavgpool_unipass_ukernel_function) xnn_f32_pavgpool_minmax_ukernel_9x__scalar_c1, |
| 534 | .mp = (xnn_pavgpool_multipass_ukernel_function) xnn_f32_pavgpool_minmax_ukernel_9p8x__scalar_c1, |
| 535 | .mr = 9, |
| 536 | .qr = 8, |
| 537 | }; |
| 538 | xnn_params.f32.gavgpool = (struct gavgpool_parameters) { |
| 539 | .up = (xnn_gavgpool_unipass_ukernel_function) xnn_f32_gavgpool_minmax_ukernel_7x__scalar_c1, |
| 540 | .mp = (xnn_gavgpool_multipass_ukernel_function) xnn_f32_gavgpool_minmax_ukernel_7p7x__scalar_c1, |
| 541 | .mr = 7, |
| 542 | }; |
| 543 | xnn_params.f32.maxpool = (struct maxpool_parameters) { |
| 544 | .ukernel = (xnn_maxpool_ukernel_function) xnn_f32_maxpool_minmax_ukernel_9p8x__scalar_c1, |
| 545 | .mr = 9, |
| 546 | .qr = 8, |
| 547 | }; |
| 548 | xnn_params.f32.argmaxpool[0] = (struct argmaxpool_parameters) { |
| 549 | .up = (xnn_argmaxpool_unipass_ukernel_function) xnn_f32_argmaxpool_ukernel_4x__scalar_c1, |
| 550 | .mr = 4, |
| 551 | }; |
| 552 | xnn_params.f32.argmaxpool[1] = (struct argmaxpool_parameters) { |
| 553 | .up = (xnn_argmaxpool_unipass_ukernel_function) xnn_f32_argmaxpool_ukernel_9x__scalar_c1, |
| 554 | .mr = 9, |
| 555 | }; |
| 556 | xnn_params.f32.argmaxpool[2] = (struct argmaxpool_parameters) { |
| 557 | .mp = (xnn_argmaxpool_multipass_ukernel_function) xnn_f32_argmaxpool_ukernel_9p8x__scalar_c1, |
| 558 | .mr = 9, |
| 559 | .qr = 8, |
| 560 | }; |
| 561 | xnn_params.f32.ibilinear = (struct ibilinear_parameters) { |
| 562 | .ukernel = (xnn_ibilinear_ukernel_function) xnn_f32_ibilinear_ukernel__scalar_c2, |
| 563 | .pixel_tile = 1, |
| 564 | .channel_tile = 2, |
| 565 | }; |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 566 | xnn_params.f32.abs = (xnn_univector_ukernel_function) xnn_f32_vabs_ukernel__scalar_x4; |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 567 | xnn_params.f32.clamp = (xnn_univector_ukernel_function) xnn_f32_clamp_ukernel__scalar_x4; |
| 568 | xnn_params.f32.hswish = (xnn_univector_ukernel_function) xnn_f32_hswish_ukernel__scalar_x4; |
Marat Dukhan | 2881333 | 2020-06-10 18:05:38 -0700 | [diff] [blame] | 569 | xnn_params.f32.lrelu = (xnn_univector_ukernel_function) xnn_f32_vlrelu_ukernel__scalar_x4; |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 570 | xnn_params.f32.neg = (xnn_univector_ukernel_function) xnn_f32_vneg_ukernel__scalar_x4; |
Marat Dukhan | 64e5251 | 2020-06-09 13:41:16 -0700 | [diff] [blame] | 571 | xnn_params.f32.rndne = (xnn_univector_ukernel_function) xnn_f32_vrndne_ukernel__scalar_libm_x1; |
| 572 | xnn_params.f32.rndz = (xnn_univector_ukernel_function) xnn_f32_vrndz_ukernel__scalar_libm_x1; |
| 573 | xnn_params.f32.rndu = (xnn_univector_ukernel_function) xnn_f32_vrndu_ukernel__scalar_libm_x1; |
| 574 | xnn_params.f32.rndd = (xnn_univector_ukernel_function) xnn_f32_vrndd_ukernel__scalar_libm_x1; |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 575 | xnn_params.f32.sigmoid = (xnn_univector_ukernel_function) xnn_f32_sigmoid_ukernel__scalar_lut64_p2_div_x2; |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 576 | xnn_params.f32.sqr = (xnn_univector_ukernel_function) xnn_f32_vsqr_ukernel__scalar_x4; |
Marat Dukhan | 6804bbd | 2020-06-30 19:26:11 -0700 | [diff] [blame] | 577 | xnn_params.f32.sqrt = (xnn_univector_ukernel_function) xnn_f32_vsqrt_ukernel__scalar_sqrt_x1; |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 578 | xnn_params.f32.prelu = (struct prelu_parameters) { |
| 579 | .ukernel = (xnn_prelu_ukernel_function) xnn_f32_prelu_ukernel__scalar_2x4, |
| 580 | .row_tile = 4, |
| 581 | .channel_tile = 4, |
| 582 | }; |
| 583 | xnn_params.f32.raddstoreexpminusmax = xnn_f32_raddstoreexpminusmax_ukernel__scalar_p5_x4_acc2; |
| 584 | xnn_params.f32.rmax = xnn_f32_rmax_ukernel__scalar; |
| 585 | xnn_params.f32.vadd = (struct vbinary_parameters) { |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 586 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vadd_minmax_ukernel__scalar_x8, |
| 587 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_minmax_ukernel__scalar_x8, |
| 588 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_minmax_ukernel__scalar_x8, |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 589 | .element_tile = 8, |
| 590 | }; |
| 591 | xnn_params.f32.vdiv = (struct vbinary_parameters) { |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 592 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdiv_minmax_ukernel__scalar_x2, |
| 593 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdivc_minmax_ukernel__scalar_x2, |
| 594 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrdivc_minmax_ukernel__scalar_x2, |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 595 | .element_tile = 2, |
| 596 | }; |
| 597 | xnn_params.f32.vmax = (struct vbinary_parameters) { |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 598 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmax_ukernel__scalar_x8, |
| 599 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__scalar_x8, |
| 600 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__scalar_x8, |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 601 | .element_tile = 8, |
| 602 | }; |
| 603 | xnn_params.f32.vmin = (struct vbinary_parameters) { |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 604 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmin_ukernel__scalar_x8, |
| 605 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__scalar_x8, |
| 606 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__scalar_x8, |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 607 | .element_tile = 8, |
| 608 | }; |
| 609 | xnn_params.f32.vmul = (struct vbinary_parameters) { |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 610 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmul_minmax_ukernel__scalar_x8, |
| 611 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_minmax_ukernel__scalar_x8, |
| 612 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_minmax_ukernel__scalar_x8, |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 613 | .element_tile = 8, |
| 614 | }; |
| 615 | xnn_params.f32.vsub = (struct vbinary_parameters) { |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 616 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsub_minmax_ukernel__scalar_x8, |
| 617 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsubc_minmax_ukernel__scalar_x8, |
| 618 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrsubc_minmax_ukernel__scalar_x8, |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 619 | .element_tile = 8, |
| 620 | }; |
Marat Dukhan | f739926 | 2020-06-05 10:58:44 -0700 | [diff] [blame] | 621 | xnn_params.f32.vsqrdiff = (struct vbinary_parameters) { |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 622 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiff_ukernel__scalar_x8, |
| 623 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiffc_ukernel__scalar_x8, |
| 624 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiffc_ukernel__scalar_x8, |
Marat Dukhan | f739926 | 2020-06-05 10:58:44 -0700 | [diff] [blame] | 625 | .element_tile = 8, |
| 626 | }; |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 627 | xnn_params.f32.vmulcaddc = (struct vmulcaddc_parameters) { |
Marat Dukhan | 9531e9f | 2020-07-24 15:25:02 -0700 | [diff] [blame] | 628 | .ukernel = (xnn_vmulcaddc_ukernel_function) xnn_f32_vmulcaddc_minmax_ukernel_c1__scalar_2x, |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 629 | .channel_tile = 1, |
| 630 | .row_tile = 2, |
| 631 | }; |
| 632 | #ifndef XNN_NO_NCHW_OPERATORS |
| 633 | xnn_params.f32.spmm = (struct spmm_parameters) { |
| 634 | .ukernel = (xnn_spmm_ukernel_function) xnn_f32_spmm_minmax_ukernel_8x1__scalar, |
| 635 | .mr = 8, |
| 636 | .nr = 1, |
| 637 | }; |
| 638 | xnn_params.f32.spmm2 = (struct spmm_parameters) { |
| 639 | .ukernel = (xnn_spmm_ukernel_function) xnn_f32_spmm_minmax_ukernel_8x2__scalar, |
| 640 | .mr = 8, |
| 641 | .nr = 2, |
| 642 | }; |
| 643 | xnn_params.f32.spmm4 = (struct spmm_parameters) { |
| 644 | .ukernel = (xnn_spmm_ukernel_function) xnn_f32_spmm_minmax_ukernel_8x4__scalar, |
| 645 | .mr = 8, |
| 646 | .nr = 4, |
| 647 | }; |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 648 | xnn_params.f32.conv_hwc2chw_3x3c3s2 = (struct conv_hwc2chw_parameters) { |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 649 | .ukernel_with_symm_padding = |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 650 | (xnn_conv_hwc2chw_ukernel_function) xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__scalar_1x1, |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 651 | .output_channel_tile = 4, |
| 652 | .output_height_tile = 1, |
| 653 | .output_width_tile = 1, |
| 654 | }; |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 655 | xnn_params.f32.dwconv_chw_3x3 = (struct dwconv_chw_parameters) { |
Marat Dukhan | 6f469a5 | 2020-10-21 20:02:52 -0700 | [diff] [blame] | 656 | .ukernel = (xnn_dwconv_chw_ukernel_function) xnn_f32_dwconv_chw_ukernel_3x3p1__scalar_1x1_acc3, |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 657 | .input_width_tile = 1, |
| 658 | .output_width_tile = 1, |
| 659 | .output_height_tile = 1, |
| 660 | }; |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 661 | xnn_params.f32.dwconv_chw_3x3s2 = (struct dwconv_chw_parameters) { |
Marat Dukhan | 6f469a5 | 2020-10-21 20:02:52 -0700 | [diff] [blame] | 662 | .ukernel = (xnn_dwconv_chw_ukernel_function) xnn_f32_dwconv_chw_ukernel_3x3s2p1__scalar_1x1_acc3, |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 663 | .input_width_tile = 1, |
| 664 | .output_width_tile = 1, |
| 665 | .output_height_tile = 1, |
| 666 | }; |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 667 | xnn_params.f32.dwconv_chw_5x5 = (struct dwconv_chw_parameters) { |
Marat Dukhan | 6f469a5 | 2020-10-21 20:02:52 -0700 | [diff] [blame] | 668 | .ukernel = (xnn_dwconv_chw_ukernel_function) xnn_f32_dwconv_chw_ukernel_5x5p2__scalar_1x1_acc5, |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 669 | .input_width_tile = 1, |
| 670 | .output_width_tile = 1, |
| 671 | .output_height_tile = 1, |
| 672 | }; |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 673 | xnn_params.f32.dwconv_chw_5x5s2 = (struct dwconv_chw_parameters) { |
Marat Dukhan | 6f469a5 | 2020-10-21 20:02:52 -0700 | [diff] [blame] | 674 | .ukernel = (xnn_dwconv_chw_ukernel_function) xnn_f32_dwconv_chw_ukernel_5x5s2p2__scalar_1x1_acc5, |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 675 | .input_width_tile = 1, |
| 676 | .output_width_tile = 1, |
| 677 | .output_height_tile = 1, |
| 678 | }; |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 679 | xnn_params.f32.gavgpool_cw = (struct gavgpool_cw_parameters) { |
| 680 | .ukernel = (xnn_gavgpool_cw_ukernel_function) xnn_f32_gavgpool_cw_ukernel__scalar_x1, |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 681 | .channel_tile = 1, |
| 682 | }; |
| 683 | #endif // XNN_NO_NCHW_OPERATORS |
| 684 | #endif // XNN_NO_F32_OPERATORS |
| 685 | |
| 686 | /**************************** X32 micro-kernels ****************************/ |
| 687 | #ifndef XNN_NO_X32_OPERATORS |
Marat Dukhan | 854fb6b | 2020-06-19 12:33:44 -0700 | [diff] [blame] | 688 | init_flags |= XNN_INIT_FLAG_X32; |
| 689 | |
Marat Dukhan | 4662b19 | 2020-05-21 15:52:03 -0700 | [diff] [blame] | 690 | xnn_params.x32.fill = (struct fill_parameters) { |
| 691 | .ukernel = (xnn_fill_ukernel_function) xnn_x32_fill_ukernel__scalar_int, |
| 692 | .row_tile = 1, |
| 693 | }; |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 694 | xnn_params.x32.pad = (struct pad_parameters) { |
Marat Dukhan | 63523d4 | 2020-05-22 17:07:33 -0700 | [diff] [blame] | 695 | .ukernel = (xnn_pad_ukernel_function) xnn_x32_pad_ukernel__scalar_int, |
| 696 | .row_tile = 1, |
Marat Dukhan | 3b745a4 | 2020-05-10 21:43:25 -0700 | [diff] [blame] | 697 | }; |
| 698 | xnn_params.x32.unpool = (xnn_unpool_ukernel_function) xnn_x32_unpool_ukernel__scalar; |
| 699 | xnn_params.x32.zip = (struct zip_parameters) { |
| 700 | .x2 = (xnn_zipc_ukernel_function) xnn_x32_zip_x2_ukernel__scalar, |
| 701 | .x3 = (xnn_zipc_ukernel_function) xnn_x32_zip_x3_ukernel__scalar, |
| 702 | .x4 = (xnn_zipc_ukernel_function) xnn_x32_zip_x4_ukernel__scalar, |
| 703 | .xm = (xnn_zipv_ukernel_function) xnn_x32_zip_xm_ukernel__scalar, |
| 704 | }; |
| 705 | #endif // XNN_NO_X32_OPERATORS |
| 706 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 707 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 708 | #elif XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 709 | |
Marat Dukhan | f28cddf | 2020-08-10 14:05:02 -0700 | [diff] [blame] | 710 | /**************************** QS8 micro-kernels ****************************/ |
| 711 | #ifndef XNN_NO_QS8_OPERATORS |
| 712 | init_flags |= XNN_INIT_FLAG_QS8; |
| 713 | |
Marat Dukhan | 31677ad | 2020-10-13 23:59:31 -0700 | [diff] [blame] | 714 | #if XNN_PLATFORM_IOS |
Frank Barchard | bc0c729 | 2020-10-06 13:36:54 -0700 | [diff] [blame] | 715 | #if XNN_ENABLE_ASSEMBLY |
Marat Dukhan | 31677ad | 2020-10-13 23:59:31 -0700 | [diff] [blame] | 716 | if (cpuinfo_has_arm_neon_dot()) { |
| 717 | xnn_params.qs8.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_4x16c4__aarch64_neondot_ld64); |
| 718 | xnn_params.qs8.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_1x16c4__aarch64_neondot_ld64); |
| 719 | xnn_params.qs8.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot); |
| 720 | xnn_params.qs8.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_1x16c4__neondot); |
| 721 | xnn_params.qs8.gemm.mr = 4; |
| 722 | xnn_params.qs8.gemm.nr = 16; |
| 723 | xnn_params.qs8.gemm.log2_kr = 2; |
| 724 | } else { |
| 725 | xnn_params.qs8.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane); |
| 726 | xnn_params.qs8.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_2x8__neon_mlal_lane); |
| 727 | xnn_params.qs8.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_1x8__neon_mlal_lane); |
| 728 | xnn_params.qs8.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_1x8__neon_mlal_lane); |
| 729 | xnn_params.qs8.gemm.mr = 2; |
| 730 | xnn_params.qs8.gemm.nr = 8; |
Frank Barchard | 1e8590e | 2020-10-12 21:20:46 -0700 | [diff] [blame] | 731 | } |
Marat Dukhan | 31677ad | 2020-10-13 23:59:31 -0700 | [diff] [blame] | 732 | #else // !XNN_ENABLE_ASSEMBLY |
| 733 | if (cpuinfo_has_arm_neon_dot()) { |
| 734 | xnn_params.qs8.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot); |
| 735 | xnn_params.qs8.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_1x16c4__neondot); |
| 736 | xnn_params.qs8.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot); |
| 737 | xnn_params.qs8.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_1x16c4__neondot); |
| 738 | xnn_params.qs8.gemm.mr = 4; |
| 739 | xnn_params.qs8.gemm.nr = 16; |
| 740 | xnn_params.qs8.gemm.log2_kr = 2; |
| 741 | } else { |
| 742 | xnn_params.qs8.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane); |
| 743 | xnn_params.qs8.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_2x8__neon_mlal_lane); |
| 744 | xnn_params.qs8.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_1x8__neon_mlal_lane); |
| 745 | xnn_params.qs8.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_1x8__neon_mlal_lane); |
| 746 | xnn_params.qs8.gemm.mr = 2; |
| 747 | xnn_params.qs8.gemm.nr = 8; |
| 748 | } |
| 749 | #endif // XNN_ENABLE_ASSEMBLY |
| 750 | #else // !XNN_PLATFORM_IOS |
| 751 | #if XNN_ENABLE_ASSEMBLY |
| 752 | if (cpuinfo_has_arm_neon_dot()) { |
| 753 | switch (cpuinfo_get_core(0)->uarch) { |
| 754 | case cpuinfo_uarch_cortex_a55: |
| 755 | xnn_params.qs8.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_4x16c4__aarch64_neondot_cortex_a55); |
| 756 | break; |
| 757 | default: |
| 758 | xnn_params.qs8.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_4x16c4__aarch64_neondot_ld64); |
| 759 | break; |
| 760 | } |
| 761 | xnn_params.qs8.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot); |
| 762 | xnn_params.qs8.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_1x16c4__aarch64_neondot_ld64); |
| 763 | xnn_params.qs8.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_1x16c4__neondot); |
| 764 | xnn_params.qs8.gemm.mr = 4; |
| 765 | xnn_params.qs8.gemm.nr = 16; |
| 766 | xnn_params.qs8.gemm.log2_kr = 2; |
| 767 | } else { |
| 768 | xnn_params.qs8.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane); |
| 769 | xnn_params.qs8.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_2x8__neon_mlal_lane); |
| 770 | xnn_params.qs8.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_1x8__neon_mlal_lane); |
| 771 | xnn_params.qs8.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_1x8__neon_mlal_lane); |
| 772 | xnn_params.qs8.gemm.mr = 2; |
| 773 | xnn_params.qs8.gemm.nr = 8; |
| 774 | } |
| 775 | #if XNN_MAX_UARCH_TYPES > 1 |
| 776 | { |
| 777 | /* Choose micro-kernels for little cores according to micro-kernel specification for the big core */ |
| 778 | const uint32_t mr = xnn_params.qs8.gemm.mr; |
| 779 | const uint32_t nr = xnn_params.qs8.gemm.nr; |
| 780 | const uint32_t log2_kr = xnn_params.qs8.gemm.log2_kr; |
| 781 | for (size_t i = 1; i < XNN_MAX_UARCH_TYPES; i++) { |
| 782 | const struct cpuinfo_uarch_info* uarch_info = cpuinfo_get_uarch(i); |
| 783 | if (uarch_info == NULL) { |
| 784 | /* No more microarchitectures in the system */ |
| 785 | break; |
| 786 | } |
| 787 | |
| 788 | switch (uarch_info->uarch) { |
| 789 | case cpuinfo_uarch_cortex_a55: |
| 790 | if (mr == 4 && nr == 16 && log2_kr == 2) { |
| 791 | xnn_params.qs8.gemm.minmax.gemm.function[i] = (xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_4x16c4__aarch64_neondot_cortex_a55; |
| 792 | xnn_params.qs8.gemm.minmax.igemm.function[i] = (xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot; |
| 793 | xnn_params.qs8.gemm.minmax.gemm1.function[i] = (xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_1x16c4__aarch64_neondot_ld64; |
| 794 | xnn_params.qs8.gemm.minmax.igemm1.function[i] = (xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_1x16c4__neondot; |
| 795 | } |
| 796 | break; |
| 797 | default: |
| 798 | break; |
| 799 | } |
| 800 | } |
| 801 | } |
| 802 | #endif // XNN_MAX_UARCH_TYPES > 1 |
| 803 | #else // !XNN_ENABLE_ASSEMBLY |
| 804 | if (cpuinfo_has_arm_neon_dot()) { |
| 805 | xnn_params.qs8.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot); |
| 806 | xnn_params.qs8.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_1x16c4__neondot); |
| 807 | xnn_params.qs8.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot); |
| 808 | xnn_params.qs8.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_1x16c4__neondot); |
| 809 | xnn_params.qs8.gemm.mr = 4; |
| 810 | xnn_params.qs8.gemm.nr = 16; |
| 811 | xnn_params.qs8.gemm.log2_kr = 2; |
| 812 | } else { |
| 813 | xnn_params.qs8.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane); |
| 814 | xnn_params.qs8.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_2x8__neon_mlal_lane); |
| 815 | xnn_params.qs8.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_1x8__neon_mlal_lane); |
| 816 | xnn_params.qs8.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_1x8__neon_mlal_lane); |
| 817 | xnn_params.qs8.gemm.mr = 2; |
| 818 | xnn_params.qs8.gemm.nr = 8; |
| 819 | } |
| 820 | #endif // XNN_ENABLE_ASSEMBLY |
| 821 | #endif // XNN_PLATFORM_IOS |
Marat Dukhan | f28cddf | 2020-08-10 14:05:02 -0700 | [diff] [blame] | 822 | |
| 823 | xnn_params.qs8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qs8_dwconv_minmax_ukernel_up8x9__neon_mul16; |
| 824 | xnn_params.qs8.dwconv[0].channel_tile = 8; |
| 825 | xnn_params.qs8.dwconv[0].primary_tile = 9; |
| 826 | |
| 827 | xnn_params.qs8.gavgpool = (struct gavgpool_parameters) { |
| 828 | .up = (xnn_gavgpool_unipass_ukernel_function) xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c8_acc2, |
| 829 | .mp = (xnn_gavgpool_multipass_ukernel_function) xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c8_acc2, |
| 830 | .mr = 7, |
| 831 | }; |
Marat Dukhan | ff20948 | 2020-09-03 14:26:53 -0700 | [diff] [blame] | 832 | |
| 833 | xnn_params.qs8.vadd = (struct vbinary_parameters) { |
| 834 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vadd_minmax_ukernel__neon_ld64_x8, |
| 835 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x8, |
| 836 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x8, |
| 837 | .element_tile = 8, |
| 838 | }; |
Marat Dukhan | f28cddf | 2020-08-10 14:05:02 -0700 | [diff] [blame] | 839 | #endif // XNN_NO_QS8_OPERATORS |
| 840 | |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 841 | /**************************** QU8 micro-kernels ****************************/ |
| 842 | #ifndef XNN_NO_QU8_OPERATORS |
| 843 | init_flags |= XNN_INIT_FLAG_QU8; |
Marat Dukhan | 854fb6b | 2020-06-19 12:33:44 -0700 | [diff] [blame] | 844 | |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 845 | xnn_params.qu8.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_ukernel_8x8__neon); |
| 846 | xnn_params.qu8.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_ukernel_8x8__neon); |
| 847 | xnn_params.qu8.gemm.mr = 8; |
| 848 | xnn_params.qu8.gemm.nr = 8; |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 849 | |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 850 | xnn_params.qu8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qu8_dwconv_minmax_ukernel_up8x9__neon; |
| 851 | xnn_params.qu8.dwconv[0].channel_tile = 8; |
| 852 | xnn_params.qu8.dwconv[0].primary_tile = 9; |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 853 | |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 854 | xnn_params.qu8.avgpool = (struct avgpool_parameters) { |
| 855 | .up = (xnn_avgpool_unipass_ukernel_function) xnn_qu8_avgpool_minmax_ukernel_9x__neon_c8, |
| 856 | .mp = (xnn_avgpool_multipass_ukernel_function) xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 857 | .mr = 9, |
| 858 | .qr = 8, |
| 859 | }; |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 860 | xnn_params.qu8.gavgpool = (struct gavgpool_parameters) { |
| 861 | .up = (xnn_gavgpool_unipass_ukernel_function) xnn_qu8_gavgpool_minmax_ukernel_7x__neon_c8, |
| 862 | .mp = (xnn_gavgpool_multipass_ukernel_function) xnn_qu8_gavgpool_minmax_ukernel_7p7x__neon_c8, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 863 | .mr = 7, |
| 864 | }; |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 865 | xnn_params.qu8.vadd = (xnn_vadd_ukernel_function) xnn_qu8_vadd_minmax_ukernel__neon; |
| 866 | #endif // XNN_NO_QU8_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 867 | |
| 868 | /**************************** U8 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 869 | #ifndef XNN_NO_U8_OPERATORS |
Marat Dukhan | 854fb6b | 2020-06-19 12:33:44 -0700 | [diff] [blame] | 870 | init_flags |= XNN_INIT_FLAG_U8; |
| 871 | |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 872 | xnn_params.u8.maxpool = (struct maxpool_parameters) { |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 873 | .ukernel = (xnn_maxpool_ukernel_function) xnn_u8_maxpool_minmax_ukernel_9p8x__neon_c16, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 874 | .mr = 9, |
| 875 | .qr = 8, |
| 876 | }; |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 877 | xnn_params.u8.clamp = (xnn_univector_ukernel_function) xnn_u8_clamp_ukernel__neon_x64; |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 878 | xnn_params.u8.lut32norm = xnn_u8_lut32norm_ukernel__scalar; |
| 879 | xnn_params.u8.rmax = xnn_u8_rmax_ukernel__neon; |
| 880 | #endif // XNN_NO_U8_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 881 | |
| 882 | /**************************** X8 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 883 | #ifndef XNN_NO_X8_OPERATORS |
Marat Dukhan | 854fb6b | 2020-06-19 12:33:44 -0700 | [diff] [blame] | 884 | init_flags |= XNN_INIT_FLAG_X8; |
| 885 | |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 886 | xnn_params.x8.lut = xnn_x8_lut_ukernel__scalar; |
| 887 | xnn_params.x8.zip = (struct zip_parameters) { |
| 888 | .x2 = (xnn_zipc_ukernel_function) xnn_x8_zip_x2_ukernel__neon, |
| 889 | .x3 = (xnn_zipc_ukernel_function) xnn_x8_zip_x3_ukernel__neon, |
| 890 | .x4 = (xnn_zipc_ukernel_function) xnn_x8_zip_x4_ukernel__neon, |
| 891 | .xm = (xnn_zipv_ukernel_function) xnn_x8_zip_xm_ukernel__neon, |
| 892 | }; |
| 893 | #endif // XNN_NO_X8_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 894 | |
Frank Barchard | 7e2cbb0 | 2020-06-12 01:22:13 -0700 | [diff] [blame] | 895 | /**************************** F16 micro-kernels ****************************/ |
| 896 | #ifndef XNN_NO_F16_OPERATORS |
Marat Dukhan | 8d5d259 | 2020-06-19 12:48:57 -0700 | [diff] [blame] | 897 | if (cpuinfo_has_arm_neon_fp16_arith()) { |
| 898 | init_flags |= XNN_INIT_FLAG_F16; |
Marat Dukhan | 854fb6b | 2020-06-19 12:33:44 -0700 | [diff] [blame] | 899 | |
Frank Barchard | 6b73c4f | 2020-06-26 18:40:40 -0700 | [diff] [blame] | 900 | #if XNN_ENABLE_ASSEMBLY |
| 901 | xnn_params.f16.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_ld32); |
| 902 | xnn_params.f16.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f16_gemm_minmax_ukernel_1x16__aarch64_neonfp16arith_ld32); |
| 903 | #else |
| 904 | xnn_params.f16.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f16_gemm_minmax_ukernel_6x16__neonfp16arith_ld64); |
| 905 | xnn_params.f16.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f16_gemm_minmax_ukernel_1x16__neonfp16arith_ld64); |
| 906 | #endif |
Frank Barchard | 49b4dcc | 2020-06-26 14:07:19 -0700 | [diff] [blame] | 907 | xnn_params.f16.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f16_igemm_minmax_ukernel_6x16__neonfp16arith_ld64); |
Frank Barchard | 49b4dcc | 2020-06-26 14:07:19 -0700 | [diff] [blame] | 908 | xnn_params.f16.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64); |
| 909 | xnn_params.f16.gemm.mr = 6; |
| 910 | xnn_params.f16.gemm.nr = 16; |
| 911 | |
| 912 | xnn_params.f16.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith; |
| 913 | xnn_params.f16.dwconv[0].channel_tile = 16; |
| 914 | xnn_params.f16.dwconv[0].primary_tile = 4; |
| 915 | |
| 916 | xnn_params.f16.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith; |
| 917 | xnn_params.f16.dwconv[1].channel_tile = 16; |
| 918 | xnn_params.f16.dwconv[1].primary_tile = 9; |
| 919 | |
| 920 | xnn_params.f16.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2; |
| 921 | xnn_params.f16.dwconv[2].channel_tile = 8; |
| 922 | xnn_params.f16.dwconv[2].primary_tile = 25; |
| 923 | |
Marat Dukhan | 8d5d259 | 2020-06-19 12:48:57 -0700 | [diff] [blame] | 924 | xnn_params.f16.gavgpool = (struct gavgpool_parameters) { |
| 925 | .up = (xnn_gavgpool_unipass_ukernel_function) xnn_f16_gavgpool_minmax_ukernel_7x__neonfp16arith_c8, |
| 926 | .mp = (xnn_gavgpool_multipass_ukernel_function) xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c8, |
| 927 | .mr = 7, |
| 928 | }; |
Frank Barchard | 01898c0 | 2020-06-23 21:49:50 -0700 | [diff] [blame] | 929 | xnn_params.f16.vadd = (struct vbinary_parameters) { |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 930 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vadd_minmax_ukernel__neonfp16arith_x16, |
| 931 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vaddc_minmax_ukernel__neonfp16arith_x16, |
| 932 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vaddc_minmax_ukernel__neonfp16arith_x16, |
Frank Barchard | 01898c0 | 2020-06-23 21:49:50 -0700 | [diff] [blame] | 933 | .element_tile = 16, |
| 934 | }; |
Frank Barchard | 0ea6a77 | 2020-09-09 15:26:31 -0700 | [diff] [blame] | 935 | xnn_params.f16.vmul = (struct vbinary_parameters) { |
| 936 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vmul_minmax_ukernel__neonfp16arith_x16, |
| 937 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vmulc_minmax_ukernel__neonfp16arith_x16, |
| 938 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vmulc_minmax_ukernel__neonfp16arith_x16, |
| 939 | .element_tile = 16, |
| 940 | }; |
Frank Barchard | 49b4dcc | 2020-06-26 14:07:19 -0700 | [diff] [blame] | 941 | xnn_params.f16.vmulcaddc = (struct vmulcaddc_parameters) { |
Marat Dukhan | 9531e9f | 2020-07-24 15:25:02 -0700 | [diff] [blame] | 942 | .ukernel = (xnn_vmulcaddc_ukernel_function) xnn_f16_vmulcaddc_minmax_ukernel_c8__neonfp16arith_2x, |
Frank Barchard | 49b4dcc | 2020-06-26 14:07:19 -0700 | [diff] [blame] | 943 | .channel_tile = 8, |
| 944 | .row_tile = 2, |
| 945 | }; |
Frank Barchard | a96948e | 2020-09-11 15:34:18 -0700 | [diff] [blame] | 946 | xnn_params.f16.hswish = (xnn_univector_ukernel_function) xnn_f16_hswish_ukernel__neonfp16arith_x16; |
Marat Dukhan | 8d5d259 | 2020-06-19 12:48:57 -0700 | [diff] [blame] | 947 | } |
Frank Barchard | 7e2cbb0 | 2020-06-12 01:22:13 -0700 | [diff] [blame] | 948 | #endif // XNN_NO_F16_OPERATORS |
| 949 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 950 | /**************************** F32 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 951 | #ifndef XNN_NO_F32_OPERATORS |
Marat Dukhan | 854fb6b | 2020-06-19 12:33:44 -0700 | [diff] [blame] | 952 | init_flags |= XNN_INIT_FLAG_F32; |
| 953 | |
Frank Barchard | 0d1052c | 2020-03-23 17:28:13 -0700 | [diff] [blame] | 954 | #if XNN_PLATFORM_IOS |
| 955 | #if XNN_ENABLE_ASSEMBLY |
Frank Barchard | 016e586 | 2020-06-11 12:24:50 -0700 | [diff] [blame] | 956 | xnn_params.f32.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a75); |
| 957 | xnn_params.f32.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a75); |
| 958 | xnn_params.f32.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a75); |
| 959 | xnn_params.f32.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a75); |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 960 | xnn_params.f32.gemm.mr = 6; |
| 961 | xnn_params.f32.gemm.nr = 8; |
Frank Barchard | 0d1052c | 2020-03-23 17:28:13 -0700 | [diff] [blame] | 962 | #else // !XNN_ENABLE_ASSEMBLY |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 963 | xnn_params.f32.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64); |
| 964 | xnn_params.f32.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64); |
| 965 | xnn_params.f32.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__neonfma_lane_ld64); |
| 966 | xnn_params.f32.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__neonfma_lane_ld64); |
| 967 | xnn_params.f32.gemm.mr = 6; |
| 968 | xnn_params.f32.gemm.nr = 8; |
Frank Barchard | 0d1052c | 2020-03-23 17:28:13 -0700 | [diff] [blame] | 969 | #endif // XNN_ENABLE_ASSEMBLY |
| 970 | #else // !XNN_PLATFORM_IOS |
| 971 | #if XNN_ENABLE_ASSEMBLY |
| 972 | switch (cpuinfo_get_core(0)->uarch) { |
| 973 | case cpuinfo_uarch_cortex_a57: |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 974 | xnn_params.f32.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a57); |
| 975 | xnn_params.f32.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a57); |
| 976 | xnn_params.f32.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a57); |
| 977 | xnn_params.f32.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a57); |
| 978 | xnn_params.f32.gemm.mr = 6; |
| 979 | xnn_params.f32.gemm.nr = 8; |
Frank Barchard | 0d1052c | 2020-03-23 17:28:13 -0700 | [diff] [blame] | 980 | break; |
| 981 | case cpuinfo_uarch_cortex_a72: |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 982 | xnn_params.f32.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x8__aarch64_neonfma_cortex_a75); |
| 983 | xnn_params.f32.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x8__aarch64_neonfma_cortex_a75); |
| 984 | xnn_params.f32.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a75); |
| 985 | xnn_params.f32.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a75); |
| 986 | xnn_params.f32.gemm.mr = 4; |
| 987 | xnn_params.f32.gemm.nr = 8; |
Frank Barchard | 0d1052c | 2020-03-23 17:28:13 -0700 | [diff] [blame] | 988 | break; |
| 989 | case cpuinfo_uarch_cortex_a75: |
| 990 | case cpuinfo_uarch_cortex_a76: |
| 991 | case cpuinfo_uarch_exynos_m3: |
| 992 | case cpuinfo_uarch_exynos_m4: |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 993 | xnn_params.f32.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a75); |
| 994 | xnn_params.f32.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a75); |
| 995 | xnn_params.f32.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a75); |
| 996 | xnn_params.f32.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a75); |
| 997 | xnn_params.f32.gemm.mr = 6; |
| 998 | xnn_params.f32.gemm.nr = 8; |
Frank Barchard | 0d1052c | 2020-03-23 17:28:13 -0700 | [diff] [blame] | 999 | break; |
| 1000 | case cpuinfo_uarch_exynos_m1: |
| 1001 | case cpuinfo_uarch_exynos_m2: |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 1002 | xnn_params.f32.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_6x8s4__neonfma); |
| 1003 | xnn_params.f32.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_6x8s4__neonfma); |
| 1004 | xnn_params.f32.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8s4__neonfma); |
| 1005 | xnn_params.f32.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8s4__neonfma); |
| 1006 | xnn_params.f32.gemm.mr = 6; |
| 1007 | xnn_params.f32.gemm.nr = 8; |
| 1008 | xnn_params.f32.gemm.log2_sr = 2; |
Frank Barchard | 0d1052c | 2020-03-23 17:28:13 -0700 | [diff] [blame] | 1009 | break; |
Frank Barchard | 0d1052c | 2020-03-23 17:28:13 -0700 | [diff] [blame] | 1010 | case cpuinfo_uarch_cortex_a53: |
| 1011 | case cpuinfo_uarch_cortex_a55r0: |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 1012 | xnn_params.f32.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a53); |
| 1013 | xnn_params.f32.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a53); |
| 1014 | xnn_params.f32.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a53); |
| 1015 | xnn_params.f32.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a53); |
| 1016 | xnn_params.f32.gemm.mr = 6; |
| 1017 | xnn_params.f32.gemm.nr = 8; |
Frank Barchard | 0d1052c | 2020-03-23 17:28:13 -0700 | [diff] [blame] | 1018 | break; |
| 1019 | case cpuinfo_uarch_cortex_a55: |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 1020 | xnn_params.f32.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a55); |
| 1021 | xnn_params.f32.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a55); |
| 1022 | xnn_params.f32.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a53); |
| 1023 | xnn_params.f32.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a53); |
| 1024 | xnn_params.f32.gemm.mr = 6; |
| 1025 | xnn_params.f32.gemm.nr = 8; |
Frank Barchard | 0d1052c | 2020-03-23 17:28:13 -0700 | [diff] [blame] | 1026 | break; |
| 1027 | case cpuinfo_uarch_cortex_a73: |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 1028 | xnn_params.f32.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a73); |
| 1029 | xnn_params.f32.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a73); |
| 1030 | xnn_params.f32.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a75); |
| 1031 | xnn_params.f32.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a75); |
| 1032 | xnn_params.f32.gemm.mr = 6; |
| 1033 | xnn_params.f32.gemm.nr = 8; |
Frank Barchard | 0d1052c | 2020-03-23 17:28:13 -0700 | [diff] [blame] | 1034 | break; |
| 1035 | default: |
| 1036 | case cpuinfo_uarch_cortex_a77: |
| 1037 | case cpuinfo_uarch_exynos_m5: |
| 1038 | case cpuinfo_uarch_kryo: |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 1039 | xnn_params.f32.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x8__aarch64_neonfma_cortex_a57); |
| 1040 | xnn_params.f32.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x8__aarch64_neonfma_cortex_a57); |
| 1041 | xnn_params.f32.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a57); |
| 1042 | xnn_params.f32.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a57); |
| 1043 | xnn_params.f32.gemm.mr = 4; |
| 1044 | xnn_params.f32.gemm.nr = 8; |
Frank Barchard | 0d1052c | 2020-03-23 17:28:13 -0700 | [diff] [blame] | 1045 | break; |
| 1046 | } |
Marat Dukhan | 05702cf | 2020-03-26 15:41:33 -0700 | [diff] [blame] | 1047 | #if XNN_MAX_UARCH_TYPES > 1 |
| 1048 | { |
| 1049 | /* Choose micro-kernels for little cores according to micro-kernel specification for the big core */ |
| 1050 | const uint32_t mr = xnn_params.f32.gemm.mr; |
| 1051 | const uint32_t nr = xnn_params.f32.gemm.nr; |
| 1052 | const uint32_t log2_sr = xnn_params.f32.gemm.log2_sr; |
| 1053 | for (size_t i = 1; i < XNN_MAX_UARCH_TYPES; i++) { |
| 1054 | const struct cpuinfo_uarch_info* uarch_info = cpuinfo_get_uarch(i); |
| 1055 | if (uarch_info == NULL) { |
| 1056 | /* No more microarchitectures in the system */ |
| 1057 | break; |
| 1058 | } |
| 1059 | |
| 1060 | switch (uarch_info->uarch) { |
| 1061 | case cpuinfo_uarch_cortex_a53: |
| 1062 | case cpuinfo_uarch_cortex_a55r0: |
| 1063 | if (mr == 6 && nr == 8 && log2_sr == 0) { |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 1064 | xnn_params.f32.gemm.minmax.gemm.function[i] = (xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a53; |
| 1065 | xnn_params.f32.gemm.minmax.igemm.function[i] = (xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a53; |
| 1066 | xnn_params.f32.gemm.minmax.gemm1.function[i] = (xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a53; |
| 1067 | xnn_params.f32.gemm.minmax.igemm1.function[i] = (xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a53; |
Marat Dukhan | 05702cf | 2020-03-26 15:41:33 -0700 | [diff] [blame] | 1068 | } else if (mr == 4 && nr == 8 && log2_sr == 0) { |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 1069 | xnn_params.f32.gemm.minmax.gemm.function[i] = (xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x8__aarch64_neonfma_cortex_a53; |
| 1070 | xnn_params.f32.gemm.minmax.igemm.function[i] = (xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x8__aarch64_neonfma_cortex_a53; |
| 1071 | xnn_params.f32.gemm.minmax.gemm1.function[i] = (xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a53; |
| 1072 | xnn_params.f32.gemm.minmax.igemm1.function[i] = (xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a53; |
Marat Dukhan | 05702cf | 2020-03-26 15:41:33 -0700 | [diff] [blame] | 1073 | } |
| 1074 | break; |
| 1075 | case cpuinfo_uarch_cortex_a55: |
| 1076 | if (mr == 6 && nr == 8 && log2_sr == 0) { |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 1077 | xnn_params.f32.gemm.minmax.gemm.function[i] = (xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a55; |
| 1078 | xnn_params.f32.gemm.minmax.igemm.function[i] = (xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a55; |
| 1079 | xnn_params.f32.gemm.minmax.gemm1.function[i] = (xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a53; |
| 1080 | xnn_params.f32.gemm.minmax.igemm1.function[i] = (xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a53; |
Marat Dukhan | 05702cf | 2020-03-26 15:41:33 -0700 | [diff] [blame] | 1081 | } else if (mr == 4 && nr == 8 && log2_sr == 0) { |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 1082 | xnn_params.f32.gemm.minmax.gemm.function[i] = (xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x8__aarch64_neonfma_cortex_a55; |
| 1083 | xnn_params.f32.gemm.minmax.igemm.function[i] = (xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x8__aarch64_neonfma_cortex_a55; |
| 1084 | xnn_params.f32.gemm.minmax.gemm1.function[i] = (xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a53; |
| 1085 | xnn_params.f32.gemm.minmax.igemm1.function[i] = (xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a53; |
Marat Dukhan | 05702cf | 2020-03-26 15:41:33 -0700 | [diff] [blame] | 1086 | } |
| 1087 | break; |
| 1088 | default: |
| 1089 | break; |
| 1090 | } |
| 1091 | } |
| 1092 | } |
| 1093 | #endif // XNN_MAX_UARCH_TYPES > 1 |
Frank Barchard | 0d1052c | 2020-03-23 17:28:13 -0700 | [diff] [blame] | 1094 | #else // !XNN_ENABLE_ASSEMBLY |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 1095 | xnn_params.f32.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64); |
| 1096 | xnn_params.f32.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64); |
| 1097 | xnn_params.f32.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__neonfma_lane_ld64); |
| 1098 | xnn_params.f32.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__neonfma_lane_ld64); |
| 1099 | xnn_params.f32.gemm.mr = 6; |
| 1100 | xnn_params.f32.gemm.nr = 8; |
Marat Dukhan | 31677ad | 2020-10-13 23:59:31 -0700 | [diff] [blame] | 1101 | #endif // XNN_ENABLE_ASSEMBLY |
Frank Barchard | 0d1052c | 2020-03-23 17:28:13 -0700 | [diff] [blame] | 1102 | #endif // XNN_PLATFORM_IOS |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 1103 | xnn_params.f32.gemm2.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x2__neonfma_lane_ld64); |
| 1104 | xnn_params.f32.gemm2.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x2__neonfma_lane_ld64); |
| 1105 | xnn_params.f32.gemm2.mr = 4; |
| 1106 | xnn_params.f32.gemm2.nr = 2; |
| 1107 | |
Marat Dukhan | f5425ea | 2020-04-24 01:46:00 -0700 | [diff] [blame] | 1108 | xnn_params.f32.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma; |
| 1109 | xnn_params.f32.dwconv[0].channel_tile = 8; |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 1110 | xnn_params.f32.dwconv[0].primary_tile = 4; |
| 1111 | |
Frank Barchard | 0d1052c | 2020-03-23 17:28:13 -0700 | [diff] [blame] | 1112 | #if XNN_PLATFORM_IOS |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 1113 | xnn_params.f32.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma; |
| 1114 | xnn_params.f32.dwconv[1].channel_tile = 8; |
| 1115 | xnn_params.f32.dwconv[1].primary_tile = 9; |
Frank Barchard | 0d1052c | 2020-03-23 17:28:13 -0700 | [diff] [blame] | 1116 | #else // !XNN_PLATFORM_IOS |
| 1117 | switch (cpuinfo_get_core(0)->uarch) { |
| 1118 | case cpuinfo_uarch_kryo: |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 1119 | xnn_params.f32.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma; |
| 1120 | xnn_params.f32.dwconv[1].channel_tile = 4; |
| 1121 | xnn_params.f32.dwconv[1].primary_tile = 9; |
Frank Barchard | 0d1052c | 2020-03-23 17:28:13 -0700 | [diff] [blame] | 1122 | break; |
| 1123 | #if XNN_ENABLE_ASSEMBLY |
| 1124 | case cpuinfo_uarch_cortex_a53: |
| 1125 | case cpuinfo_uarch_cortex_a55r0: |
| 1126 | case cpuinfo_uarch_cortex_a55: |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 1127 | xnn_params.f32.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up4x9__aarch64_neonfma_cortex_a55; |
| 1128 | xnn_params.f32.dwconv[1].channel_tile = 4; |
| 1129 | xnn_params.f32.dwconv[1].primary_tile = 9; |
Frank Barchard | 0d1052c | 2020-03-23 17:28:13 -0700 | [diff] [blame] | 1130 | break; |
| 1131 | #endif // XNN_ENABLE_ASSEMBLY |
| 1132 | default: |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 1133 | xnn_params.f32.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma; |
| 1134 | xnn_params.f32.dwconv[1].channel_tile = 8; |
| 1135 | xnn_params.f32.dwconv[1].primary_tile = 9; |
Frank Barchard | 0d1052c | 2020-03-23 17:28:13 -0700 | [diff] [blame] | 1136 | break; |
| 1137 | } |
| 1138 | #endif // XNN_PLATFORM_IOS |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 1139 | |
Marat Dukhan | f5425ea | 2020-04-24 01:46:00 -0700 | [diff] [blame] | 1140 | xnn_params.f32.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2; |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 1141 | xnn_params.f32.dwconv[2].channel_tile = 4; |
| 1142 | xnn_params.f32.dwconv[2].primary_tile = 25; |
| 1143 | |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1144 | xnn_params.f32.avgpool = (struct avgpool_parameters) { |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1145 | .up = (xnn_avgpool_unipass_ukernel_function) xnn_f32_avgpool_minmax_ukernel_9x__neon_c4, |
| 1146 | .mp = (xnn_avgpool_multipass_ukernel_function) xnn_f32_avgpool_minmax_ukernel_9p8x__neon_c4, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1147 | .mr = 9, |
| 1148 | .qr = 8, |
| 1149 | }; |
| 1150 | xnn_params.f32.pavgpool = (struct pavgpool_parameters) { |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1151 | .up = (xnn_pavgpool_unipass_ukernel_function) xnn_f32_pavgpool_minmax_ukernel_9x__neon_c4, |
| 1152 | .mp = (xnn_pavgpool_multipass_ukernel_function) xnn_f32_pavgpool_minmax_ukernel_9p8x__neon_c4, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1153 | .mr = 9, |
| 1154 | .qr = 8, |
| 1155 | }; |
| 1156 | xnn_params.f32.gavgpool = (struct gavgpool_parameters) { |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1157 | .up = (xnn_gavgpool_unipass_ukernel_function) xnn_f32_gavgpool_minmax_ukernel_7x__neon_c4, |
| 1158 | .mp = (xnn_gavgpool_multipass_ukernel_function) xnn_f32_gavgpool_minmax_ukernel_7p7x__neon_c4, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1159 | .mr = 7, |
| 1160 | }; |
| 1161 | xnn_params.f32.maxpool = (struct maxpool_parameters) { |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1162 | .ukernel = (xnn_maxpool_ukernel_function) xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1163 | .mr = 9, |
| 1164 | .qr = 8, |
| 1165 | }; |
| 1166 | xnn_params.f32.argmaxpool[0] = (struct argmaxpool_parameters) { |
Marat Dukhan | ef25c6d | 2020-07-24 00:59:40 -0700 | [diff] [blame] | 1167 | .up = (xnn_argmaxpool_unipass_ukernel_function) xnn_f32_argmaxpool_ukernel_4x__neon_c4, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1168 | .mr = 4, |
| 1169 | }; |
| 1170 | xnn_params.f32.argmaxpool[1] = (struct argmaxpool_parameters) { |
Marat Dukhan | ef25c6d | 2020-07-24 00:59:40 -0700 | [diff] [blame] | 1171 | .up = (xnn_argmaxpool_unipass_ukernel_function) xnn_f32_argmaxpool_ukernel_9x__neon_c4, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1172 | .mr = 9, |
| 1173 | }; |
| 1174 | xnn_params.f32.argmaxpool[2] = (struct argmaxpool_parameters) { |
Marat Dukhan | ef25c6d | 2020-07-24 00:59:40 -0700 | [diff] [blame] | 1175 | .mp = (xnn_argmaxpool_multipass_ukernel_function) xnn_f32_argmaxpool_ukernel_9p8x__neon_c4, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1176 | .mr = 9, |
| 1177 | .qr = 8, |
| 1178 | }; |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 1179 | xnn_params.f32.ibilinear = (struct ibilinear_parameters) { |
| 1180 | .ukernel = (xnn_ibilinear_ukernel_function) xnn_f32_ibilinear_ukernel__neonfma_c8, |
Marat Dukhan | 6972249 | 2019-11-11 19:55:50 -0800 | [diff] [blame] | 1181 | .pixel_tile = 1, |
| 1182 | .channel_tile = 8, |
| 1183 | }; |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 1184 | xnn_params.f32.abs = (xnn_univector_ukernel_function) xnn_f32_vabs_ukernel__neon_x8; |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1185 | xnn_params.f32.clamp = (xnn_univector_ukernel_function) xnn_f32_clamp_ukernel__neon_x8; |
Marat Dukhan | 55dde5b | 2020-07-10 22:48:54 -0700 | [diff] [blame] | 1186 | xnn_params.f32.hswish = (xnn_univector_ukernel_function) xnn_f32_hswish_ukernel__neon_x16; |
Marat Dukhan | 2881333 | 2020-06-10 18:05:38 -0700 | [diff] [blame] | 1187 | xnn_params.f32.lrelu = (xnn_univector_ukernel_function) xnn_f32_vlrelu_ukernel__neon_x8; |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 1188 | xnn_params.f32.neg = (xnn_univector_ukernel_function) xnn_f32_vneg_ukernel__neon_x8; |
Marat Dukhan | 64e5251 | 2020-06-09 13:41:16 -0700 | [diff] [blame] | 1189 | xnn_params.f32.rndne = (xnn_univector_ukernel_function) xnn_f32_vrndne_ukernel__neonv8_x8; |
| 1190 | xnn_params.f32.rndz = (xnn_univector_ukernel_function) xnn_f32_vrndz_ukernel__neonv8_x8; |
| 1191 | xnn_params.f32.rndu = (xnn_univector_ukernel_function) xnn_f32_vrndu_ukernel__neonv8_x8; |
| 1192 | xnn_params.f32.rndd = (xnn_univector_ukernel_function) xnn_f32_vrndd_ukernel__neonv8_x8; |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 1193 | xnn_params.f32.sigmoid = (xnn_univector_ukernel_function) xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x16; |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 1194 | xnn_params.f32.sqr = (xnn_univector_ukernel_function) xnn_f32_vsqr_ukernel__neon_x8; |
Marat Dukhan | 6804bbd | 2020-06-30 19:26:11 -0700 | [diff] [blame] | 1195 | xnn_params.f32.sqrt = (xnn_univector_ukernel_function) xnn_f32_vsqrt_ukernel__neon_sqrt_x4; |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1196 | xnn_params.f32.prelu = (struct prelu_parameters) { |
Marat Dukhan | 69c3f2c | 2019-11-06 12:30:01 -0800 | [diff] [blame] | 1197 | .ukernel = (xnn_prelu_ukernel_function) xnn_f32_prelu_ukernel__neon_2x8, |
| 1198 | .row_tile = 2, |
| 1199 | .channel_tile = 8, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1200 | }; |
Marat Dukhan | 1edc454 | 2020-01-27 12:40:13 -0800 | [diff] [blame] | 1201 | xnn_params.f32.raddstoreexpminusmax = xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16; |
| 1202 | xnn_params.f32.rmax = xnn_f32_rmax_ukernel__neon; |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 1203 | xnn_params.f32.vadd = (struct vbinary_parameters) { |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 1204 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vadd_minmax_ukernel__neon_x8, |
| 1205 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_minmax_ukernel__neon_x8, |
| 1206 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_minmax_ukernel__neon_x8, |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 1207 | .element_tile = 8, |
| 1208 | }; |
Marat Dukhan | 6918050 | 2019-12-06 15:00:31 -0800 | [diff] [blame] | 1209 | xnn_params.f32.vdiv = (struct vbinary_parameters) { |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 1210 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdiv_minmax_ukernel__neon_x8, |
| 1211 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdivc_minmax_ukernel__neon_x8, |
| 1212 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrdivc_minmax_ukernel__neon_x8, |
Marat Dukhan | 6918050 | 2019-12-06 15:00:31 -0800 | [diff] [blame] | 1213 | .element_tile = 8, |
| 1214 | }; |
Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 1215 | xnn_params.f32.vmax = (struct vbinary_parameters) { |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 1216 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmax_ukernel__neon_x8, |
| 1217 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__neon_x8, |
| 1218 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__neon_x8, |
Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 1219 | .element_tile = 8, |
| 1220 | }; |
| 1221 | xnn_params.f32.vmin = (struct vbinary_parameters) { |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 1222 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmin_ukernel__neon_x8, |
| 1223 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__neon_x8, |
| 1224 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__neon_x8, |
Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 1225 | .element_tile = 8, |
| 1226 | }; |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 1227 | xnn_params.f32.vmul = (struct vbinary_parameters) { |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 1228 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmul_minmax_ukernel__neon_x8, |
| 1229 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_minmax_ukernel__neon_x8, |
| 1230 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_minmax_ukernel__neon_x8, |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 1231 | .element_tile = 8, |
| 1232 | }; |
Marat Dukhan | 05f3f6d | 2019-12-03 15:13:53 -0800 | [diff] [blame] | 1233 | xnn_params.f32.vsub = (struct vbinary_parameters) { |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 1234 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsub_minmax_ukernel__neon_x8, |
| 1235 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsubc_minmax_ukernel__neon_x8, |
| 1236 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrsubc_minmax_ukernel__neon_x8, |
Marat Dukhan | 05f3f6d | 2019-12-03 15:13:53 -0800 | [diff] [blame] | 1237 | .element_tile = 8, |
| 1238 | }; |
Marat Dukhan | f739926 | 2020-06-05 10:58:44 -0700 | [diff] [blame] | 1239 | xnn_params.f32.vsqrdiff = (struct vbinary_parameters) { |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 1240 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiff_ukernel__neon_x8, |
| 1241 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiffc_ukernel__neon_x8, |
| 1242 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiffc_ukernel__neon_x8, |
Marat Dukhan | f739926 | 2020-06-05 10:58:44 -0700 | [diff] [blame] | 1243 | .element_tile = 8, |
| 1244 | }; |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1245 | xnn_params.f32.vmulcaddc = (struct vmulcaddc_parameters) { |
Marat Dukhan | 9531e9f | 2020-07-24 15:25:02 -0700 | [diff] [blame] | 1246 | .ukernel = (xnn_vmulcaddc_ukernel_function) xnn_f32_vmulcaddc_minmax_ukernel_c4__neonfma_2x, |
Marat Dukhan | 49e6ee9 | 2019-11-06 15:55:29 -0800 | [diff] [blame] | 1247 | .channel_tile = 4, |
| 1248 | .row_tile = 2, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1249 | }; |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 1250 | #ifndef XNN_NO_NCHW_OPERATORS |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1251 | xnn_params.f32.spmm = (struct spmm_parameters) { |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 1252 | .ukernel = (xnn_spmm_ukernel_function) xnn_f32_spmm_minmax_ukernel_16x1__neonfma_pipelined, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1253 | .mr = 16, |
| 1254 | .nr = 1, |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1255 | }; |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1256 | xnn_params.f32.spmm2 = (struct spmm_parameters) { |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 1257 | .ukernel = (xnn_spmm_ukernel_function) xnn_f32_spmm_minmax_ukernel_16x2__neonfma, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1258 | .mr = 16, |
| 1259 | .nr = 2, |
| 1260 | }; |
| 1261 | xnn_params.f32.spmm4 = (struct spmm_parameters) { |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 1262 | .ukernel = (xnn_spmm_ukernel_function) xnn_f32_spmm_minmax_ukernel_16x4__neonfma, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1263 | .mr = 16, |
| 1264 | .nr = 4, |
| 1265 | }; |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 1266 | xnn_params.f32.conv_hwc2chw_3x3c3s2 = (struct conv_hwc2chw_parameters) { |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1267 | .ukernel_with_symm_padding = |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 1268 | (xnn_conv_hwc2chw_ukernel_function) xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1269 | .output_channel_tile = 4, |
| 1270 | .output_height_tile = 2, |
| 1271 | .output_width_tile = 2, |
| 1272 | }; |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 1273 | xnn_params.f32.dwconv_chw_3x3 = (struct dwconv_chw_parameters) { |
Marat Dukhan | 1c6cad9 | 2020-10-21 15:48:21 -0700 | [diff] [blame] | 1274 | .ukernel = (xnn_dwconv_chw_ukernel_function) xnn_f32_dwconv_chw_ukernel_3x3p1__neonfma_3x4, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1275 | .input_width_tile = 4, |
| 1276 | .output_width_tile = 4, |
| 1277 | .output_height_tile = 3, |
| 1278 | }; |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 1279 | xnn_params.f32.dwconv_chw_3x3s2 = (struct dwconv_chw_parameters) { |
Marat Dukhan | 6f469a5 | 2020-10-21 20:02:52 -0700 | [diff] [blame] | 1280 | .ukernel = (xnn_dwconv_chw_ukernel_function) xnn_f32_dwconv_chw_ukernel_3x3s2p1__neonfma_1x4_acc3, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1281 | .input_width_tile = 4, |
| 1282 | .output_width_tile = 4, |
| 1283 | .output_height_tile = 1, |
| 1284 | }; |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 1285 | xnn_params.f32.dwconv_chw_5x5 = (struct dwconv_chw_parameters) { |
Marat Dukhan | 1c6cad9 | 2020-10-21 15:48:21 -0700 | [diff] [blame] | 1286 | .ukernel = (xnn_dwconv_chw_ukernel_function) xnn_f32_dwconv_chw_ukernel_5x5p2__neonfma_3x4, |
Marat Dukhan | a99918a | 2019-11-15 14:40:12 -0800 | [diff] [blame] | 1287 | .input_width_tile = 4, |
| 1288 | .output_width_tile = 4, |
Erich Elsen | 4ad5115 | 2019-11-19 13:11:53 -0800 | [diff] [blame] | 1289 | .output_height_tile = 3, |
Marat Dukhan | a99918a | 2019-11-15 14:40:12 -0800 | [diff] [blame] | 1290 | }; |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 1291 | xnn_params.f32.dwconv_chw_5x5s2 = (struct dwconv_chw_parameters) { |
Marat Dukhan | 6f469a5 | 2020-10-21 20:02:52 -0700 | [diff] [blame] | 1292 | .ukernel = (xnn_dwconv_chw_ukernel_function) xnn_f32_dwconv_chw_ukernel_5x5s2p2__neonfma_1x4_acc2, |
Marat Dukhan | a99918a | 2019-11-15 14:40:12 -0800 | [diff] [blame] | 1293 | .input_width_tile = 4, |
| 1294 | .output_width_tile = 4, |
| 1295 | .output_height_tile = 1, |
| 1296 | }; |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 1297 | xnn_params.f32.gavgpool_cw = (struct gavgpool_cw_parameters) { |
| 1298 | .ukernel = (xnn_gavgpool_cw_ukernel_function) xnn_f32_gavgpool_cw_ukernel__neon_x4, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1299 | .channel_tile = 4, |
| 1300 | }; |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 1301 | #endif // XNN_NO_NCHW_OPERATORS |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1302 | #endif // XNN_NO_F32_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1303 | |
| 1304 | /**************************** X32 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1305 | #ifndef XNN_NO_X32_OPERATORS |
Marat Dukhan | 854fb6b | 2020-06-19 12:33:44 -0700 | [diff] [blame] | 1306 | init_flags |= XNN_INIT_FLAG_X32; |
| 1307 | |
Marat Dukhan | 4662b19 | 2020-05-21 15:52:03 -0700 | [diff] [blame] | 1308 | xnn_params.x32.fill = (struct fill_parameters) { |
| 1309 | .ukernel = (xnn_fill_ukernel_function) xnn_x32_fill_ukernel__neon, |
| 1310 | .row_tile = 1, |
| 1311 | }; |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1312 | xnn_params.x32.pad = (struct pad_parameters) { |
Marat Dukhan | 63523d4 | 2020-05-22 17:07:33 -0700 | [diff] [blame] | 1313 | .ukernel = (xnn_pad_ukernel_function) xnn_x32_pad_ukernel__neon, |
| 1314 | .row_tile = 1, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1315 | }; |
Marat Dukhan | 57dccd8 | 2020-04-14 00:53:10 -0700 | [diff] [blame] | 1316 | xnn_params.x32.unpool = (xnn_unpool_ukernel_function) xnn_x32_unpool_ukernel__neon; |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1317 | xnn_params.x32.zip = (struct zip_parameters) { |
| 1318 | .x2 = (xnn_zipc_ukernel_function) xnn_x32_zip_x2_ukernel__neon, |
| 1319 | .x3 = (xnn_zipc_ukernel_function) xnn_x32_zip_x3_ukernel__neon, |
| 1320 | .x4 = (xnn_zipc_ukernel_function) xnn_x32_zip_x4_ukernel__neon, |
| 1321 | .xm = (xnn_zipv_ukernel_function) xnn_x32_zip_xm_ukernel__neon, |
| 1322 | }; |
| 1323 | #endif // XNN_NO_X32_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1324 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 1325 | #elif XNN_ARCH_X86 || XNN_ARCH_X86_64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1326 | if (!cpuinfo_has_x86_sse2()) { |
| 1327 | xnn_log_error("XNNPACK initialization failed: SSE2 is not supported"); |
| 1328 | return; |
| 1329 | } |
| 1330 | |
Marat Dukhan | 16f1e1a | 2020-08-04 16:38:22 -0700 | [diff] [blame] | 1331 | /**************************** QS8 micro-kernels ****************************/ |
| 1332 | #ifndef XNN_NO_QS8_OPERATORS |
| 1333 | init_flags |= XNN_INIT_FLAG_QS8; |
| 1334 | |
Marat Dukhan | bb00b1d | 2020-08-10 11:37:23 -0700 | [diff] [blame] | 1335 | if (cpuinfo_has_x86_avx512f() && cpuinfo_has_x86_avx512bw() && cpuinfo_has_x86_avx512dq() && cpuinfo_has_x86_avx512vl()) { |
| 1336 | xnn_params.qs8.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_4x16c8__avx512skx); |
| 1337 | xnn_params.qs8.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx); |
| 1338 | xnn_params.qs8.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_1x16c8__avx512skx); |
| 1339 | xnn_params.qs8.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_1x16c8__avx512skx); |
| 1340 | xnn_params.qs8.gemm.mr = 4; |
| 1341 | xnn_params.qs8.gemm.nr = 16; |
| 1342 | xnn_params.qs8.gemm.log2_kr = 3; |
| 1343 | } else if (cpuinfo_has_x86_xop()) { |
Marat Dukhan | 75215d8 | 2020-08-07 23:08:03 -0700 | [diff] [blame] | 1344 | // XOP should be checked before AVX2: AMD Excavator supports both, but performs better with XOP microkernels |
| 1345 | xnn_params.qs8.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_2x4c8__xop_ld64); |
| 1346 | xnn_params.qs8.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_2x4c8__xop_ld64); |
| 1347 | xnn_params.qs8.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_1x4c8__xop_ld64); |
| 1348 | xnn_params.qs8.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_1x4c8__xop_ld64); |
| 1349 | xnn_params.qs8.gemm.mr = 2; |
| 1350 | xnn_params.qs8.gemm.nr = 4; |
| 1351 | xnn_params.qs8.gemm.log2_kr = 3; |
| 1352 | } else if (cpuinfo_has_x86_avx2()) { |
Marat Dukhan | 16f1e1a | 2020-08-04 16:38:22 -0700 | [diff] [blame] | 1353 | xnn_params.qs8.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_3x8c8__avx2); |
| 1354 | xnn_params.qs8.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_3x8c8__avx2); |
| 1355 | xnn_params.qs8.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_1x8c8__avx2); |
| 1356 | xnn_params.qs8.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_1x8c8__avx2); |
| 1357 | xnn_params.qs8.gemm.mr = 3; |
| 1358 | xnn_params.qs8.gemm.nr = 8; |
| 1359 | xnn_params.qs8.gemm.log2_kr = 3; |
| 1360 | } else if (cpuinfo_has_x86_sse4_1()) { |
| 1361 | xnn_params.qs8.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_3x4c8__sse41_ld64); |
| 1362 | xnn_params.qs8.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_3x4c8__sse41_ld64); |
| 1363 | xnn_params.qs8.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_1x4c8__sse41_ld64); |
| 1364 | xnn_params.qs8.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_1x4c8__sse41_ld64); |
| 1365 | xnn_params.qs8.gemm.mr = 3; |
| 1366 | xnn_params.qs8.gemm.nr = 4; |
| 1367 | xnn_params.qs8.gemm.log2_kr = 3; |
| 1368 | } else if (cpuinfo_has_x86_ssse3()) { |
| 1369 | xnn_params.qs8.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_3x4c8__ssse3_ld64); |
| 1370 | xnn_params.qs8.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_3x4c8__ssse3_ld64); |
| 1371 | xnn_params.qs8.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_1x4c8__ssse3_ld64); |
| 1372 | xnn_params.qs8.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_1x4c8__ssse3_ld64); |
| 1373 | xnn_params.qs8.gemm.mr = 3; |
| 1374 | xnn_params.qs8.gemm.nr = 4; |
| 1375 | xnn_params.qs8.gemm.log2_kr = 3; |
| 1376 | } else { |
Marat Dukhan | 07e5040 | 2020-08-05 17:16:53 -0700 | [diff] [blame] | 1377 | xnn_params.qs8.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld64); |
| 1378 | xnn_params.qs8.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_ld64); |
| 1379 | xnn_params.qs8.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_1x4c8__sse2_ld64); |
| 1380 | xnn_params.qs8.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_1x4c8__sse2_ld64); |
Marat Dukhan | 16f1e1a | 2020-08-04 16:38:22 -0700 | [diff] [blame] | 1381 | xnn_params.qs8.gemm.mr = 3; |
| 1382 | xnn_params.qs8.gemm.nr = 4; |
| 1383 | xnn_params.qs8.gemm.log2_kr = 3; |
| 1384 | } |
| 1385 | |
Marat Dukhan | 2ffc5e6 | 2020-09-06 22:33:38 -0700 | [diff] [blame] | 1386 | if (cpuinfo_has_x86_avx512f() && cpuinfo_has_x86_avx512bw() && cpuinfo_has_x86_avx512dq() && cpuinfo_has_x86_avx512vl()) { |
| 1387 | xnn_params.qs8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qs8_dwconv_minmax_ukernel_up32x9__avx512skx_mul32; |
| 1388 | xnn_params.qs8.dwconv[0].channel_tile = 32; |
| 1389 | } else if (cpuinfo_has_x86_avx2()) { |
Marat Dukhan | d65a152 | 2020-08-04 19:28:18 -0700 | [diff] [blame] | 1390 | xnn_params.qs8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul32; |
| 1391 | xnn_params.qs8.dwconv[0].channel_tile = 16; |
| 1392 | } else if (cpuinfo_has_x86_sse4_1()) { |
Marat Dukhan | 16f1e1a | 2020-08-04 16:38:22 -0700 | [diff] [blame] | 1393 | xnn_params.qs8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16; |
Marat Dukhan | d65a152 | 2020-08-04 19:28:18 -0700 | [diff] [blame] | 1394 | xnn_params.qs8.dwconv[0].channel_tile = 8; |
Marat Dukhan | 16f1e1a | 2020-08-04 16:38:22 -0700 | [diff] [blame] | 1395 | } else if (cpuinfo_has_x86_ssse3()) { |
| 1396 | xnn_params.qs8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16; |
Marat Dukhan | d65a152 | 2020-08-04 19:28:18 -0700 | [diff] [blame] | 1397 | xnn_params.qs8.dwconv[0].channel_tile = 8; |
Marat Dukhan | 16f1e1a | 2020-08-04 16:38:22 -0700 | [diff] [blame] | 1398 | } else if (cpuinfo_has_x86_sse2()) { |
| 1399 | xnn_params.qs8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16; |
Marat Dukhan | d65a152 | 2020-08-04 19:28:18 -0700 | [diff] [blame] | 1400 | xnn_params.qs8.dwconv[0].channel_tile = 8; |
Marat Dukhan | 16f1e1a | 2020-08-04 16:38:22 -0700 | [diff] [blame] | 1401 | } |
Marat Dukhan | 16f1e1a | 2020-08-04 16:38:22 -0700 | [diff] [blame] | 1402 | xnn_params.qs8.dwconv[0].primary_tile = 9; |
Marat Dukhan | 9e0b539 | 2020-08-07 02:29:34 -0700 | [diff] [blame] | 1403 | |
| 1404 | if (cpuinfo_has_x86_sse4_1()) { |
| 1405 | xnn_params.qs8.gavgpool = (struct gavgpool_parameters) { |
| 1406 | .up = (xnn_gavgpool_unipass_ukernel_function) xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2, |
| 1407 | .mp = (xnn_gavgpool_multipass_ukernel_function) xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2, |
| 1408 | .mr = 7, |
| 1409 | }; |
| 1410 | } else if (cpuinfo_has_x86_ssse3()) { |
| 1411 | xnn_params.qs8.gavgpool = (struct gavgpool_parameters) { |
| 1412 | .up = (xnn_gavgpool_unipass_ukernel_function) xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2, |
| 1413 | .mp = (xnn_gavgpool_multipass_ukernel_function) xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2, |
| 1414 | .mr = 7, |
| 1415 | }; |
| 1416 | } else if (cpuinfo_has_x86_sse2()) { |
| 1417 | xnn_params.qs8.gavgpool = (struct gavgpool_parameters) { |
| 1418 | .up = (xnn_gavgpool_unipass_ukernel_function) xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2, |
| 1419 | .mp = (xnn_gavgpool_multipass_ukernel_function) xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2, |
| 1420 | .mr = 7, |
| 1421 | }; |
| 1422 | } |
Marat Dukhan | ff20948 | 2020-09-03 14:26:53 -0700 | [diff] [blame] | 1423 | |
Marat Dukhan | bb9225e | 2020-09-06 22:40:56 -0700 | [diff] [blame] | 1424 | if (cpuinfo_has_x86_xop()) { |
| 1425 | xnn_params.qs8.vadd = (struct vbinary_parameters) { |
| 1426 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x8, |
| 1427 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x8, |
| 1428 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x8, |
| 1429 | .element_tile = 8, |
| 1430 | }; |
| 1431 | } else if (cpuinfo_has_x86_sse4_1()) { |
Marat Dukhan | ff20948 | 2020-09-03 14:26:53 -0700 | [diff] [blame] | 1432 | xnn_params.qs8.vadd = (struct vbinary_parameters) { |
| 1433 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vadd_minmax_ukernel__sse41_mul16_ld64_x8, |
| 1434 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vaddc_minmax_ukernel__sse41_mul16_ld64_x8, |
| 1435 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vaddc_minmax_ukernel__sse41_mul16_ld64_x8, |
| 1436 | .element_tile = 8, |
| 1437 | }; |
| 1438 | } else { |
| 1439 | xnn_params.qs8.vadd = (struct vbinary_parameters) { |
| 1440 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vadd_minmax_ukernel__sse2_mul16_ld64_x8, |
| 1441 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vaddc_minmax_ukernel__sse2_mul16_ld64_x8, |
| 1442 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vaddc_minmax_ukernel__sse2_mul16_ld64_x8, |
| 1443 | .element_tile = 8, |
| 1444 | }; |
| 1445 | } |
Marat Dukhan | 07e5040 | 2020-08-05 17:16:53 -0700 | [diff] [blame] | 1446 | #endif // XNN_NO_QS8_OPERATORS |
Marat Dukhan | 16f1e1a | 2020-08-04 16:38:22 -0700 | [diff] [blame] | 1447 | |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 1448 | /**************************** QU8 micro-kernels ****************************/ |
| 1449 | #ifndef XNN_NO_QU8_OPERATORS |
| 1450 | init_flags |= XNN_INIT_FLAG_QU8; |
Marat Dukhan | 854fb6b | 2020-06-19 12:33:44 -0700 | [diff] [blame] | 1451 | |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 1452 | xnn_params.qu8.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_ukernel_4x4c2__sse2); |
| 1453 | xnn_params.qu8.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_ukernel_4x4c2__sse2); |
| 1454 | xnn_params.qu8.gemm.mr = 4; |
| 1455 | xnn_params.qu8.gemm.nr = 4; |
| 1456 | xnn_params.qu8.gemm.log2_kr = 1; |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 1457 | |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 1458 | xnn_params.qu8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2; |
| 1459 | xnn_params.qu8.dwconv[0].channel_tile = 8; |
| 1460 | xnn_params.qu8.dwconv[0].primary_tile = 9; |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 1461 | |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 1462 | xnn_params.qu8.avgpool = (struct avgpool_parameters) { |
| 1463 | .up = (xnn_avgpool_unipass_ukernel_function) xnn_qu8_avgpool_minmax_ukernel_9x__sse2_c8, |
| 1464 | .mp = (xnn_avgpool_multipass_ukernel_function) xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1465 | .mr = 9, |
| 1466 | .qr = 8, |
| 1467 | }; |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 1468 | xnn_params.qu8.gavgpool = (struct gavgpool_parameters) { |
| 1469 | .up = (xnn_gavgpool_unipass_ukernel_function) xnn_qu8_gavgpool_minmax_ukernel_7x__sse2_c8, |
| 1470 | .mp = (xnn_gavgpool_multipass_ukernel_function) xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1471 | .mr = 7, |
| 1472 | }; |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 1473 | xnn_params.qu8.vadd = (xnn_vadd_ukernel_function) xnn_qu8_vadd_minmax_ukernel__sse2; |
| 1474 | #endif // XNN_NO_QU8_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1475 | |
| 1476 | /**************************** U8 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1477 | #ifndef XNN_NO_U8_OPERATORS |
Marat Dukhan | 854fb6b | 2020-06-19 12:33:44 -0700 | [diff] [blame] | 1478 | init_flags |= XNN_INIT_FLAG_U8; |
| 1479 | |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1480 | xnn_params.u8.maxpool = (struct maxpool_parameters) { |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1481 | .ukernel = (xnn_maxpool_ukernel_function) xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1482 | .mr = 9, |
| 1483 | .qr = 8, |
| 1484 | }; |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1485 | xnn_params.u8.clamp = (xnn_univector_ukernel_function) xnn_u8_clamp_ukernel__sse2_x64; |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1486 | xnn_params.u8.lut32norm = xnn_u8_lut32norm_ukernel__scalar; |
| 1487 | xnn_params.u8.rmax = xnn_u8_rmax_ukernel__sse2; |
| 1488 | #endif // XNN_NO_U8_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1489 | |
| 1490 | /**************************** X8 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1491 | #ifndef XNN_NO_X8_OPERATORS |
Marat Dukhan | 854fb6b | 2020-06-19 12:33:44 -0700 | [diff] [blame] | 1492 | init_flags |= XNN_INIT_FLAG_X8; |
| 1493 | |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1494 | xnn_params.x8.lut = xnn_x8_lut_ukernel__scalar; |
| 1495 | xnn_params.x8.zip = (struct zip_parameters) { |
| 1496 | .x2 = (xnn_zipc_ukernel_function) xnn_x8_zip_x2_ukernel__sse2, |
| 1497 | .x3 = (xnn_zipc_ukernel_function) xnn_x8_zip_x3_ukernel__sse2, |
| 1498 | .x4 = (xnn_zipc_ukernel_function) xnn_x8_zip_x4_ukernel__sse2, |
| 1499 | .xm = (xnn_zipv_ukernel_function) xnn_x8_zip_xm_ukernel__sse2, |
| 1500 | }; |
| 1501 | #endif // XNN_NO_X8_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1502 | |
| 1503 | /**************************** F32 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1504 | #ifndef XNN_NO_F32_OPERATORS |
Marat Dukhan | 854fb6b | 2020-06-19 12:33:44 -0700 | [diff] [blame] | 1505 | init_flags |= XNN_INIT_FLAG_F32; |
| 1506 | |
Marat Dukhan | 0f349c4 | 2019-11-27 11:58:54 -0800 | [diff] [blame] | 1507 | if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx512f()) { |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 1508 | xnn_params.f32.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_7x16__avx512f_broadcast); |
| 1509 | xnn_params.f32.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_7x16__avx512f_broadcast); |
| 1510 | xnn_params.f32.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x16__avx512f_broadcast); |
| 1511 | xnn_params.f32.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x16__avx512f_broadcast); |
| 1512 | xnn_params.f32.gemm.mr = 7; |
| 1513 | xnn_params.f32.gemm.nr = 16; |
Marat Dukhan | 0f349c4 | 2019-11-27 11:58:54 -0800 | [diff] [blame] | 1514 | } else if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_fma3()) { |
Marat Dukhan | 2712132 | 2019-12-09 14:57:40 -0800 | [diff] [blame] | 1515 | switch (cpuinfo_get_core(0)->uarch) { |
| 1516 | case cpuinfo_uarch_zen: |
Marat Dukhan | b3801eb | 2020-03-12 13:41:11 -0700 | [diff] [blame] | 1517 | case cpuinfo_uarch_dhyana: |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 1518 | xnn_params.f32.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x16s4__fma3_broadcast); |
| 1519 | xnn_params.f32.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x16s4__fma3_broadcast); |
| 1520 | xnn_params.f32.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x16s4__fma3_broadcast); |
| 1521 | xnn_params.f32.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x16s4__fma3_broadcast); |
| 1522 | xnn_params.f32.gemm.mr = 4; |
| 1523 | xnn_params.f32.gemm.nr = 16; |
| 1524 | xnn_params.f32.gemm.log2_sr = 2; |
Marat Dukhan | 2712132 | 2019-12-09 14:57:40 -0800 | [diff] [blame] | 1525 | break; |
| 1526 | default: |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 1527 | xnn_params.f32.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_5x16__fma3_broadcast); |
| 1528 | xnn_params.f32.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_5x16__fma3_broadcast); |
| 1529 | xnn_params.f32.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x16__fma3_broadcast); |
| 1530 | xnn_params.f32.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x16__fma3_broadcast); |
| 1531 | xnn_params.f32.gemm.mr = 5; |
| 1532 | xnn_params.f32.gemm.nr = 16; |
Marat Dukhan | 2712132 | 2019-12-09 14:57:40 -0800 | [diff] [blame] | 1533 | break; |
| 1534 | } |
Marat Dukhan | 1025ea3 | 2019-11-21 16:01:08 -0800 | [diff] [blame] | 1535 | } else if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx()) { |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 1536 | xnn_params.f32.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_5x16__avx_broadcast); |
| 1537 | xnn_params.f32.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_5x16__avx_broadcast); |
| 1538 | xnn_params.f32.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x16__avx_broadcast); |
| 1539 | xnn_params.f32.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x16__avx_broadcast); |
| 1540 | xnn_params.f32.gemm.mr = 5; |
| 1541 | xnn_params.f32.gemm.nr = 16; |
Marat Dukhan | 1025ea3 | 2019-11-21 16:01:08 -0800 | [diff] [blame] | 1542 | } else { |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 1543 | xnn_params.f32.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x8__sse_load1); |
| 1544 | xnn_params.f32.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x8__sse_load1); |
| 1545 | xnn_params.f32.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__sse_load1); |
| 1546 | xnn_params.f32.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__sse_load1); |
| 1547 | xnn_params.f32.gemm.mr = 4; |
| 1548 | xnn_params.f32.gemm.nr = 8; |
Marat Dukhan | 1025ea3 | 2019-11-21 16:01:08 -0800 | [diff] [blame] | 1549 | } |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 1550 | xnn_params.f32.gemm2.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x2c4__sse); |
| 1551 | xnn_params.f32.gemm2.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x2c4__sse); |
| 1552 | xnn_params.f32.gemm2.mr = 4; |
| 1553 | xnn_params.f32.gemm2.nr = 2; |
| 1554 | xnn_params.f32.gemm2.log2_kr = 2; |
| 1555 | |
Marat Dukhan | 479f87e | 2019-11-27 15:17:06 -0800 | [diff] [blame] | 1556 | if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx512f()) { |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 1557 | xnn_params.f32.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up16x4__avx512f; |
| 1558 | xnn_params.f32.dwconv[0].channel_tile = 16; |
| 1559 | xnn_params.f32.dwconv[0].primary_tile = 4; |
| 1560 | |
| 1561 | xnn_params.f32.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f; |
| 1562 | xnn_params.f32.dwconv[1].channel_tile = 16; |
| 1563 | xnn_params.f32.dwconv[1].primary_tile = 9; |
| 1564 | |
| 1565 | xnn_params.f32.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f; |
| 1566 | xnn_params.f32.dwconv[2].channel_tile = 16; |
| 1567 | xnn_params.f32.dwconv[2].primary_tile = 25; |
Marat Dukhan | 479f87e | 2019-11-27 15:17:06 -0800 | [diff] [blame] | 1568 | } else if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_fma3()) { |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 1569 | xnn_params.f32.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up16x4__fma3; |
| 1570 | xnn_params.f32.dwconv[0].channel_tile = 16; |
| 1571 | xnn_params.f32.dwconv[0].primary_tile = 4; |
| 1572 | |
| 1573 | xnn_params.f32.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up16x9__fma3; |
| 1574 | xnn_params.f32.dwconv[1].channel_tile = 16; |
| 1575 | xnn_params.f32.dwconv[1].primary_tile = 9; |
| 1576 | |
| 1577 | xnn_params.f32.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up8x25__fma3; |
| 1578 | xnn_params.f32.dwconv[2].channel_tile = 8; |
| 1579 | xnn_params.f32.dwconv[2].primary_tile = 25; |
Marat Dukhan | 17ec5f3 | 2019-11-22 13:34:16 -0800 | [diff] [blame] | 1580 | } else if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx()) { |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 1581 | xnn_params.f32.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up16x4__avx; |
| 1582 | xnn_params.f32.dwconv[0].channel_tile = 16; |
| 1583 | xnn_params.f32.dwconv[0].primary_tile = 4; |
| 1584 | |
| 1585 | xnn_params.f32.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up16x9__avx; |
| 1586 | xnn_params.f32.dwconv[1].channel_tile = 16; |
| 1587 | xnn_params.f32.dwconv[1].primary_tile = 9; |
| 1588 | |
| 1589 | xnn_params.f32.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up8x25__avx; |
| 1590 | xnn_params.f32.dwconv[2].channel_tile = 8; |
| 1591 | xnn_params.f32.dwconv[2].primary_tile = 25; |
Marat Dukhan | 17ec5f3 | 2019-11-22 13:34:16 -0800 | [diff] [blame] | 1592 | } else { |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 1593 | xnn_params.f32.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up8x4__sse; |
| 1594 | xnn_params.f32.dwconv[0].channel_tile = 8; |
| 1595 | xnn_params.f32.dwconv[0].primary_tile = 4; |
| 1596 | |
| 1597 | xnn_params.f32.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up8x9__sse; |
| 1598 | xnn_params.f32.dwconv[1].channel_tile = 8; |
| 1599 | xnn_params.f32.dwconv[1].primary_tile = 9; |
| 1600 | |
| 1601 | xnn_params.f32.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up8x25__sse; |
| 1602 | xnn_params.f32.dwconv[2].channel_tile = 8; |
| 1603 | xnn_params.f32.dwconv[2].primary_tile = 25; |
Marat Dukhan | 17ec5f3 | 2019-11-22 13:34:16 -0800 | [diff] [blame] | 1604 | } |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1605 | xnn_params.f32.avgpool = (struct avgpool_parameters) { |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1606 | .up = (xnn_avgpool_unipass_ukernel_function) xnn_f32_avgpool_minmax_ukernel_9x__sse_c4, |
| 1607 | .mp = (xnn_avgpool_multipass_ukernel_function) xnn_f32_avgpool_minmax_ukernel_9p8x__sse_c4, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1608 | .mr = 9, |
| 1609 | .qr = 8, |
| 1610 | }; |
| 1611 | xnn_params.f32.pavgpool = (struct pavgpool_parameters) { |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1612 | .up = (xnn_pavgpool_unipass_ukernel_function) xnn_f32_pavgpool_minmax_ukernel_9x__sse_c4, |
| 1613 | .mp = (xnn_pavgpool_multipass_ukernel_function) xnn_f32_pavgpool_minmax_ukernel_9p8x__sse_c4, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1614 | .mr = 9, |
| 1615 | .qr = 8, |
| 1616 | }; |
| 1617 | xnn_params.f32.gavgpool = (struct gavgpool_parameters) { |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1618 | .up = (xnn_gavgpool_unipass_ukernel_function) xnn_f32_gavgpool_minmax_ukernel_7x__sse_c4, |
| 1619 | .mp = (xnn_gavgpool_multipass_ukernel_function) xnn_f32_gavgpool_minmax_ukernel_7p7x__sse_c4, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1620 | .mr = 7, |
| 1621 | }; |
| 1622 | xnn_params.f32.maxpool = (struct maxpool_parameters) { |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1623 | .ukernel = (xnn_maxpool_ukernel_function) xnn_f32_maxpool_minmax_ukernel_9p8x__sse_c4, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1624 | .mr = 9, |
| 1625 | .qr = 8, |
| 1626 | }; |
| 1627 | xnn_params.f32.argmaxpool[0] = (struct argmaxpool_parameters) { |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1628 | .up = (xnn_argmaxpool_unipass_ukernel_function) xnn_f32_argmaxpool_ukernel_4x__sse2_c4, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1629 | .mr = 4, |
| 1630 | }; |
| 1631 | xnn_params.f32.argmaxpool[1] = (struct argmaxpool_parameters) { |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1632 | .up = (xnn_argmaxpool_unipass_ukernel_function) xnn_f32_argmaxpool_ukernel_9x__sse2_c4, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1633 | .mr = 9, |
| 1634 | }; |
| 1635 | xnn_params.f32.argmaxpool[2] = (struct argmaxpool_parameters) { |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 1636 | .mp = (xnn_argmaxpool_multipass_ukernel_function) xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1637 | .mr = 9, |
| 1638 | .qr = 8, |
| 1639 | }; |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 1640 | xnn_params.f32.ibilinear = (struct ibilinear_parameters) { |
| 1641 | .ukernel = (xnn_ibilinear_ukernel_function) xnn_f32_ibilinear_ukernel__sse_c8, |
Marat Dukhan | 6972249 | 2019-11-11 19:55:50 -0800 | [diff] [blame] | 1642 | .pixel_tile = 1, |
| 1643 | .channel_tile = 8, |
| 1644 | }; |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 1645 | if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx512f()) { |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 1646 | xnn_params.f32.abs = (xnn_univector_ukernel_function) xnn_f32_vabs_ukernel__avx512f_x16; |
| 1647 | } else if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx()) { |
| 1648 | xnn_params.f32.abs = (xnn_univector_ukernel_function) xnn_f32_vabs_ukernel__avx_x16; |
| 1649 | } else { |
| 1650 | xnn_params.f32.abs = (xnn_univector_ukernel_function) xnn_f32_vabs_ukernel__sse_x8; |
| 1651 | } |
| 1652 | if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx512f()) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1653 | xnn_params.f32.clamp = (xnn_univector_ukernel_function) xnn_f32_clamp_ukernel__avx512f_x16; |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 1654 | } else if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx()) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1655 | xnn_params.f32.clamp = (xnn_univector_ukernel_function) xnn_f32_clamp_ukernel__avx_x16; |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 1656 | } else { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1657 | xnn_params.f32.clamp = (xnn_univector_ukernel_function) xnn_f32_clamp_ukernel__sse_x8; |
Marat Dukhan | e2c3f29 | 2019-11-27 15:40:54 -0800 | [diff] [blame] | 1658 | } |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 1659 | if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx512f()) { |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 1660 | xnn_params.f32.hswish = (xnn_univector_ukernel_function) xnn_f32_hswish_ukernel__avx512f_x16; |
Marat Dukhan | 662faa0 | 2019-12-09 22:48:16 -0800 | [diff] [blame] | 1661 | } else if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_fma3()) { |
| 1662 | xnn_params.f32.hswish = (xnn_univector_ukernel_function) xnn_f32_hswish_ukernel__fma3_x16; |
| 1663 | } else if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx()) { |
| 1664 | xnn_params.f32.hswish = (xnn_univector_ukernel_function) xnn_f32_hswish_ukernel__avx_x16; |
| 1665 | } else { |
| 1666 | xnn_params.f32.hswish = (xnn_univector_ukernel_function) xnn_f32_hswish_ukernel__sse_x8; |
| 1667 | } |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 1668 | if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx512f()) { |
Marat Dukhan | 2881333 | 2020-06-10 18:05:38 -0700 | [diff] [blame] | 1669 | xnn_params.f32.lrelu = (xnn_univector_ukernel_function) xnn_f32_vlrelu_ukernel__avx512f_x16; |
| 1670 | } else if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx()) { |
| 1671 | xnn_params.f32.lrelu = (xnn_univector_ukernel_function) xnn_f32_vlrelu_ukernel__avx_x16; |
Marat Dukhan | 0d3f467 | 2020-06-25 16:42:58 -0700 | [diff] [blame] | 1672 | } else if (cpuinfo_has_x86_sse4_1()) { |
| 1673 | xnn_params.f32.lrelu = (xnn_univector_ukernel_function) xnn_f32_vlrelu_ukernel__sse41_x8; |
Marat Dukhan | 2881333 | 2020-06-10 18:05:38 -0700 | [diff] [blame] | 1674 | } else { |
| 1675 | xnn_params.f32.lrelu = (xnn_univector_ukernel_function) xnn_f32_vlrelu_ukernel__sse_x8; |
| 1676 | } |
| 1677 | if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx512f()) { |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 1678 | xnn_params.f32.neg = (xnn_univector_ukernel_function) xnn_f32_vneg_ukernel__avx512f_x16; |
| 1679 | } else if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx()) { |
| 1680 | xnn_params.f32.neg = (xnn_univector_ukernel_function) xnn_f32_vneg_ukernel__avx_x16; |
| 1681 | } else { |
| 1682 | xnn_params.f32.neg = (xnn_univector_ukernel_function) xnn_f32_vneg_ukernel__sse_x8; |
| 1683 | } |
Marat Dukhan | 64e5251 | 2020-06-09 13:41:16 -0700 | [diff] [blame] | 1684 | if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx512f()) { |
| 1685 | xnn_params.f32.rndne = (xnn_univector_ukernel_function) xnn_f32_vrndne_ukernel__avx512f_x16; |
| 1686 | xnn_params.f32.rndz = (xnn_univector_ukernel_function) xnn_f32_vrndz_ukernel__avx512f_x16; |
| 1687 | xnn_params.f32.rndu = (xnn_univector_ukernel_function) xnn_f32_vrndu_ukernel__avx512f_x16; |
| 1688 | xnn_params.f32.rndd = (xnn_univector_ukernel_function) xnn_f32_vrndd_ukernel__avx512f_x16; |
| 1689 | } else if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx()) { |
| 1690 | xnn_params.f32.rndne = (xnn_univector_ukernel_function) xnn_f32_vrndne_ukernel__avx_x16; |
| 1691 | xnn_params.f32.rndz = (xnn_univector_ukernel_function) xnn_f32_vrndz_ukernel__avx_x16; |
| 1692 | xnn_params.f32.rndu = (xnn_univector_ukernel_function) xnn_f32_vrndu_ukernel__avx_x16; |
| 1693 | xnn_params.f32.rndd = (xnn_univector_ukernel_function) xnn_f32_vrndd_ukernel__avx_x16; |
| 1694 | } else if (cpuinfo_has_x86_sse4_1()) { |
| 1695 | xnn_params.f32.rndne = (xnn_univector_ukernel_function) xnn_f32_vrndne_ukernel__sse41_x8; |
| 1696 | xnn_params.f32.rndz = (xnn_univector_ukernel_function) xnn_f32_vrndz_ukernel__sse41_x8; |
| 1697 | xnn_params.f32.rndu = (xnn_univector_ukernel_function) xnn_f32_vrndu_ukernel__sse41_x8; |
| 1698 | xnn_params.f32.rndd = (xnn_univector_ukernel_function) xnn_f32_vrndd_ukernel__sse41_x8; |
| 1699 | } else { |
| 1700 | xnn_params.f32.rndne = (xnn_univector_ukernel_function) xnn_f32_vrndne_ukernel__sse2_x8; |
| 1701 | xnn_params.f32.rndz = (xnn_univector_ukernel_function) xnn_f32_vrndz_ukernel__sse2_x8; |
| 1702 | xnn_params.f32.rndu = (xnn_univector_ukernel_function) xnn_f32_vrndu_ukernel__sse2_x8; |
| 1703 | xnn_params.f32.rndd = (xnn_univector_ukernel_function) xnn_f32_vrndd_ukernel__sse2_x8; |
| 1704 | } |
Marat Dukhan | d9ca7e6 | 2020-09-23 23:45:29 -0700 | [diff] [blame] | 1705 | if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx512f()) { |
| 1706 | xnn_params.f32.sigmoid = (xnn_univector_ukernel_function) xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x64; |
| 1707 | } else if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx2()) { |
Marat Dukhan | fa0a432 | 2020-01-06 16:14:29 -0800 | [diff] [blame] | 1708 | xnn_params.f32.sigmoid = (xnn_univector_ukernel_function) xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x40; |
T.J. Alumbaugh | dc2b29c | 2020-10-14 13:56:08 -0700 | [diff] [blame] | 1709 | } else if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx()) { |
| 1710 | xnn_params.f32.sigmoid = (xnn_univector_ukernel_function) xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x40; |
Marat Dukhan | 6dd7136 | 2020-09-17 23:11:21 -0700 | [diff] [blame] | 1711 | } else if (cpuinfo_has_x86_sse4_1()) { |
| 1712 | xnn_params.f32.sigmoid = (xnn_univector_ukernel_function) xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x8; |
Marat Dukhan | fa0a432 | 2020-01-06 16:14:29 -0800 | [diff] [blame] | 1713 | } else { |
Marat Dukhan | 6dd7136 | 2020-09-17 23:11:21 -0700 | [diff] [blame] | 1714 | xnn_params.f32.sigmoid = (xnn_univector_ukernel_function) xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x8; |
Marat Dukhan | fa0a432 | 2020-01-06 16:14:29 -0800 | [diff] [blame] | 1715 | } |
Marat Dukhan | 90eca0a | 2020-03-11 00:52:23 -0700 | [diff] [blame] | 1716 | if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx512f()) { |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 1717 | xnn_params.f32.sqr = (xnn_univector_ukernel_function) xnn_f32_vsqr_ukernel__avx512f_x16; |
| 1718 | } else if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx()) { |
| 1719 | xnn_params.f32.sqr = (xnn_univector_ukernel_function) xnn_f32_vsqr_ukernel__avx_x16; |
| 1720 | } else { |
| 1721 | xnn_params.f32.sqr = (xnn_univector_ukernel_function) xnn_f32_vsqr_ukernel__sse_x8; |
| 1722 | } |
Marat Dukhan | 6804bbd | 2020-06-30 19:26:11 -0700 | [diff] [blame] | 1723 | if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx()) { |
| 1724 | xnn_params.f32.sqrt = (xnn_univector_ukernel_function) xnn_f32_vsqrt_ukernel__avx_sqrt_x8; |
| 1725 | } else { |
| 1726 | xnn_params.f32.sqrt = (xnn_univector_ukernel_function) xnn_f32_vsqrt_ukernel__sse_sqrt_x4; |
| 1727 | } |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 1728 | if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx512f()) { |
Marat Dukhan | 90eca0a | 2020-03-11 00:52:23 -0700 | [diff] [blame] | 1729 | xnn_params.f32.prelu = (struct prelu_parameters) { |
| 1730 | .ukernel = (xnn_prelu_ukernel_function) xnn_f32_prelu_ukernel__avx512f_2x16, |
| 1731 | .row_tile = 2, |
| 1732 | .channel_tile = 16, |
| 1733 | }; |
| 1734 | } else if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx()) { |
| 1735 | xnn_params.f32.prelu = (struct prelu_parameters) { |
| 1736 | .ukernel = (xnn_prelu_ukernel_function) xnn_f32_prelu_ukernel__avx_2x16, |
| 1737 | .row_tile = 2, |
| 1738 | .channel_tile = 16, |
| 1739 | }; |
Marat Dukhan | 39b5e94 | 2020-06-24 15:03:48 -0700 | [diff] [blame] | 1740 | } else if (cpuinfo_has_x86_sse4_1()) { |
| 1741 | xnn_params.f32.prelu = (struct prelu_parameters) { |
| 1742 | .ukernel = (xnn_prelu_ukernel_function) xnn_f32_prelu_ukernel__sse41_2x8, |
| 1743 | .row_tile = 2, |
| 1744 | .channel_tile = 8, |
| 1745 | }; |
Marat Dukhan | 90eca0a | 2020-03-11 00:52:23 -0700 | [diff] [blame] | 1746 | } else { |
| 1747 | xnn_params.f32.prelu = (struct prelu_parameters) { |
| 1748 | .ukernel = (xnn_prelu_ukernel_function) xnn_f32_prelu_ukernel__sse2_2x8, |
| 1749 | .row_tile = 2, |
| 1750 | .channel_tile = 8, |
| 1751 | }; |
| 1752 | } |
Marat Dukhan | 1edc454 | 2020-01-27 12:40:13 -0800 | [diff] [blame] | 1753 | xnn_params.f32.raddstoreexpminusmax = xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20_acc2; |
| 1754 | xnn_params.f32.rmax = xnn_f32_rmax_ukernel__sse; |
Marat Dukhan | 9a88efe | 2019-12-10 15:54:24 -0800 | [diff] [blame] | 1755 | if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx512f()) { |
| 1756 | xnn_params.f32.vadd = (struct vbinary_parameters) { |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 1757 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vadd_minmax_ukernel__avx512f_x32, |
| 1758 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_minmax_ukernel__avx512f_x32, |
| 1759 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_minmax_ukernel__avx512f_x32, |
Marat Dukhan | 9a88efe | 2019-12-10 15:54:24 -0800 | [diff] [blame] | 1760 | .element_tile = 32, |
| 1761 | }; |
| 1762 | xnn_params.f32.vdiv = (struct vbinary_parameters) { |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 1763 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdiv_minmax_ukernel__avx512f_x32, |
| 1764 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdivc_minmax_ukernel__avx512f_x32, |
| 1765 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrdivc_minmax_ukernel__avx512f_x32, |
Marat Dukhan | 9a88efe | 2019-12-10 15:54:24 -0800 | [diff] [blame] | 1766 | .element_tile = 32, |
| 1767 | }; |
| 1768 | xnn_params.f32.vmax = (struct vbinary_parameters) { |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 1769 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmax_ukernel__avx512f_x32, |
| 1770 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__avx512f_x32, |
| 1771 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__avx512f_x32, |
Marat Dukhan | 9a88efe | 2019-12-10 15:54:24 -0800 | [diff] [blame] | 1772 | .element_tile = 32, |
| 1773 | }; |
| 1774 | xnn_params.f32.vmin = (struct vbinary_parameters) { |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 1775 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmin_ukernel__avx512f_x32, |
| 1776 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__avx512f_x32, |
| 1777 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__avx512f_x32, |
Marat Dukhan | 9a88efe | 2019-12-10 15:54:24 -0800 | [diff] [blame] | 1778 | .element_tile = 32, |
| 1779 | }; |
| 1780 | xnn_params.f32.vmul = (struct vbinary_parameters) { |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 1781 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmul_minmax_ukernel__avx512f_x32, |
| 1782 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_minmax_ukernel__avx512f_x32, |
| 1783 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_minmax_ukernel__avx512f_x32, |
Marat Dukhan | 9a88efe | 2019-12-10 15:54:24 -0800 | [diff] [blame] | 1784 | .element_tile = 32, |
| 1785 | }; |
| 1786 | xnn_params.f32.vsub = (struct vbinary_parameters) { |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 1787 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsub_minmax_ukernel__avx512f_x32, |
| 1788 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsubc_minmax_ukernel__avx512f_x32, |
| 1789 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrsubc_minmax_ukernel__avx512f_x32, |
Marat Dukhan | 9a88efe | 2019-12-10 15:54:24 -0800 | [diff] [blame] | 1790 | .element_tile = 32, |
| 1791 | }; |
Marat Dukhan | f739926 | 2020-06-05 10:58:44 -0700 | [diff] [blame] | 1792 | xnn_params.f32.vsqrdiff = (struct vbinary_parameters) { |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 1793 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiff_ukernel__avx512f_x32, |
| 1794 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiffc_ukernel__avx512f_x32, |
| 1795 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiffc_ukernel__avx512f_x32, |
Marat Dukhan | f739926 | 2020-06-05 10:58:44 -0700 | [diff] [blame] | 1796 | .element_tile = 32, |
| 1797 | }; |
Marat Dukhan | 9a88efe | 2019-12-10 15:54:24 -0800 | [diff] [blame] | 1798 | } else if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx()) { |
| 1799 | xnn_params.f32.vadd = (struct vbinary_parameters) { |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 1800 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vadd_minmax_ukernel__avx_x16, |
| 1801 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_minmax_ukernel__avx_x16, |
| 1802 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_minmax_ukernel__avx_x16, |
Marat Dukhan | 9a88efe | 2019-12-10 15:54:24 -0800 | [diff] [blame] | 1803 | .element_tile = 16, |
| 1804 | }; |
| 1805 | xnn_params.f32.vdiv = (struct vbinary_parameters) { |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 1806 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdiv_minmax_ukernel__avx_x16, |
| 1807 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdivc_minmax_ukernel__avx_x16, |
| 1808 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrdivc_minmax_ukernel__avx_x16, |
Marat Dukhan | 9a88efe | 2019-12-10 15:54:24 -0800 | [diff] [blame] | 1809 | .element_tile = 16, |
| 1810 | }; |
| 1811 | xnn_params.f32.vmax = (struct vbinary_parameters) { |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 1812 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmax_ukernel__avx_x16, |
| 1813 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__avx_x16, |
| 1814 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__avx_x16, |
Marat Dukhan | 9a88efe | 2019-12-10 15:54:24 -0800 | [diff] [blame] | 1815 | .element_tile = 16, |
| 1816 | }; |
| 1817 | xnn_params.f32.vmin = (struct vbinary_parameters) { |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 1818 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmin_ukernel__avx_x16, |
| 1819 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__avx_x16, |
| 1820 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__avx_x16, |
Marat Dukhan | 9a88efe | 2019-12-10 15:54:24 -0800 | [diff] [blame] | 1821 | .element_tile = 16, |
| 1822 | }; |
| 1823 | xnn_params.f32.vmul = (struct vbinary_parameters) { |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 1824 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmul_minmax_ukernel__avx_x16, |
| 1825 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_minmax_ukernel__avx_x16, |
| 1826 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_minmax_ukernel__avx_x16, |
Marat Dukhan | 9a88efe | 2019-12-10 15:54:24 -0800 | [diff] [blame] | 1827 | .element_tile = 16, |
| 1828 | }; |
| 1829 | xnn_params.f32.vsub = (struct vbinary_parameters) { |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 1830 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsub_minmax_ukernel__avx_x16, |
| 1831 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsubc_minmax_ukernel__avx_x16, |
| 1832 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrsubc_minmax_ukernel__avx_x16, |
Marat Dukhan | 9a88efe | 2019-12-10 15:54:24 -0800 | [diff] [blame] | 1833 | .element_tile = 16, |
| 1834 | }; |
Marat Dukhan | f739926 | 2020-06-05 10:58:44 -0700 | [diff] [blame] | 1835 | xnn_params.f32.vsqrdiff = (struct vbinary_parameters) { |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 1836 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiff_ukernel__avx_x16, |
| 1837 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiffc_ukernel__avx_x16, |
| 1838 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiffc_ukernel__avx_x16, |
Marat Dukhan | f739926 | 2020-06-05 10:58:44 -0700 | [diff] [blame] | 1839 | .element_tile = 16, |
| 1840 | }; |
Marat Dukhan | 9a88efe | 2019-12-10 15:54:24 -0800 | [diff] [blame] | 1841 | } else { |
| 1842 | xnn_params.f32.vadd = (struct vbinary_parameters) { |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 1843 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vadd_minmax_ukernel__sse_x8, |
| 1844 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_minmax_ukernel__sse_x8, |
| 1845 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_minmax_ukernel__sse_x8, |
Marat Dukhan | 9a88efe | 2019-12-10 15:54:24 -0800 | [diff] [blame] | 1846 | .element_tile = 8, |
| 1847 | }; |
| 1848 | xnn_params.f32.vdiv = (struct vbinary_parameters) { |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 1849 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdiv_minmax_ukernel__sse_x8, |
| 1850 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdivc_minmax_ukernel__sse_x8, |
| 1851 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrdivc_minmax_ukernel__sse_x8, |
Marat Dukhan | 9a88efe | 2019-12-10 15:54:24 -0800 | [diff] [blame] | 1852 | .element_tile = 8, |
| 1853 | }; |
| 1854 | xnn_params.f32.vmax = (struct vbinary_parameters) { |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 1855 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmax_ukernel__sse_x8, |
| 1856 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__sse_x8, |
| 1857 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__sse_x8, |
Marat Dukhan | 9a88efe | 2019-12-10 15:54:24 -0800 | [diff] [blame] | 1858 | .element_tile = 8, |
| 1859 | }; |
| 1860 | xnn_params.f32.vmin = (struct vbinary_parameters) { |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 1861 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmin_ukernel__sse_x8, |
| 1862 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__sse_x8, |
| 1863 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__sse_x8, |
Marat Dukhan | 9a88efe | 2019-12-10 15:54:24 -0800 | [diff] [blame] | 1864 | .element_tile = 8, |
| 1865 | }; |
| 1866 | xnn_params.f32.vmul = (struct vbinary_parameters) { |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 1867 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmul_minmax_ukernel__sse_x8, |
| 1868 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_minmax_ukernel__sse_x8, |
| 1869 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_minmax_ukernel__sse_x8, |
Marat Dukhan | 9a88efe | 2019-12-10 15:54:24 -0800 | [diff] [blame] | 1870 | .element_tile = 8, |
| 1871 | }; |
| 1872 | xnn_params.f32.vsub = (struct vbinary_parameters) { |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 1873 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsub_minmax_ukernel__sse_x8, |
| 1874 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsubc_minmax_ukernel__sse_x8, |
| 1875 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrsubc_minmax_ukernel__sse_x8, |
Marat Dukhan | 9a88efe | 2019-12-10 15:54:24 -0800 | [diff] [blame] | 1876 | .element_tile = 8, |
| 1877 | }; |
Marat Dukhan | f739926 | 2020-06-05 10:58:44 -0700 | [diff] [blame] | 1878 | xnn_params.f32.vsqrdiff = (struct vbinary_parameters) { |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 1879 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiff_ukernel__sse_x8, |
| 1880 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiffc_ukernel__sse_x8, |
| 1881 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiffc_ukernel__sse_x8, |
Marat Dukhan | f739926 | 2020-06-05 10:58:44 -0700 | [diff] [blame] | 1882 | .element_tile = 8, |
| 1883 | }; |
Marat Dukhan | 9a88efe | 2019-12-10 15:54:24 -0800 | [diff] [blame] | 1884 | } |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1885 | xnn_params.f32.vmulcaddc = (struct vmulcaddc_parameters) { |
Marat Dukhan | 9531e9f | 2020-07-24 15:25:02 -0700 | [diff] [blame] | 1886 | .ukernel = (xnn_vmulcaddc_ukernel_function) xnn_f32_vmulcaddc_minmax_ukernel_c4__sse_2x, |
Marat Dukhan | 49e6ee9 | 2019-11-06 15:55:29 -0800 | [diff] [blame] | 1887 | .channel_tile = 4, |
| 1888 | .row_tile = 2, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1889 | }; |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 1890 | #ifndef XNN_NO_NCHW_OPERATORS |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1891 | xnn_params.f32.spmm = (struct spmm_parameters) { |
Erich Elsen | 6e80fdc | 2020-06-09 15:35:37 -0700 | [diff] [blame] | 1892 | .ukernel = (xnn_spmm_ukernel_function) xnn_f32_spmm_minmax_ukernel_16x1__sse, |
| 1893 | .mr = 16, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1894 | .nr = 1, |
| 1895 | }; |
Erich Elsen | 5b2e07a | 2020-06-09 03:27:59 -0700 | [diff] [blame] | 1896 | xnn_params.f32.conv_hwc2chw_3x3c3s2 = (struct conv_hwc2chw_parameters) { |
| 1897 | .ukernel_with_symm_padding = |
| 1898 | (xnn_conv_hwc2chw_ukernel_function) xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2, |
| 1899 | .output_channel_tile = 4, |
| 1900 | .output_height_tile = 2, |
| 1901 | .output_width_tile = 2, |
| 1902 | }; |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 1903 | xnn_params.f32.dwconv_chw_3x3 = (struct dwconv_chw_parameters) { |
Marat Dukhan | 6f469a5 | 2020-10-21 20:02:52 -0700 | [diff] [blame] | 1904 | .ukernel = (xnn_dwconv_chw_ukernel_function) xnn_f32_dwconv_chw_ukernel_3x3p1__sse_1x4_acc3, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1905 | .input_width_tile = 4, |
| 1906 | .output_width_tile = 4, |
| 1907 | .output_height_tile = 1, |
| 1908 | }; |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 1909 | xnn_params.f32.dwconv_chw_3x3s2 = (struct dwconv_chw_parameters) { |
Marat Dukhan | 6f469a5 | 2020-10-21 20:02:52 -0700 | [diff] [blame] | 1910 | .ukernel = (xnn_dwconv_chw_ukernel_function) xnn_f32_dwconv_chw_ukernel_3x3s2p1__sse_1x4_acc3, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1911 | .input_width_tile = 4, |
| 1912 | .output_width_tile = 4, |
| 1913 | .output_height_tile = 1, |
| 1914 | }; |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 1915 | xnn_params.f32.gavgpool_cw = (struct gavgpool_cw_parameters) { |
| 1916 | .ukernel = (xnn_gavgpool_cw_ukernel_function) xnn_f32_gavgpool_cw_ukernel__sse_x4, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1917 | .channel_tile = 4, |
| 1918 | }; |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 1919 | #endif // XNN_NO_NCHW_OPERATORS |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1920 | #endif // XNN_NO_F32_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1921 | |
| 1922 | /**************************** X32 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1923 | #ifndef XNN_NO_X32_OPERATORS |
Marat Dukhan | 854fb6b | 2020-06-19 12:33:44 -0700 | [diff] [blame] | 1924 | init_flags |= XNN_INIT_FLAG_X32; |
| 1925 | |
Marat Dukhan | 4662b19 | 2020-05-21 15:52:03 -0700 | [diff] [blame] | 1926 | xnn_params.x32.fill = (struct fill_parameters) { |
| 1927 | .ukernel = (xnn_fill_ukernel_function) xnn_x32_fill_ukernel__sse, |
| 1928 | .row_tile = 1, |
| 1929 | }; |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1930 | xnn_params.x32.pad = (struct pad_parameters) { |
Marat Dukhan | 63523d4 | 2020-05-22 17:07:33 -0700 | [diff] [blame] | 1931 | .ukernel = (xnn_pad_ukernel_function) xnn_x32_pad_ukernel__sse, |
| 1932 | .row_tile = 1, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1933 | }; |
Marat Dukhan | 57dccd8 | 2020-04-14 00:53:10 -0700 | [diff] [blame] | 1934 | xnn_params.x32.unpool = (xnn_unpool_ukernel_function) xnn_x32_unpool_ukernel__sse2; |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1935 | xnn_params.x32.zip = (struct zip_parameters) { |
| 1936 | .x2 = (xnn_zipc_ukernel_function) xnn_x32_zip_x2_ukernel__sse2, |
| 1937 | .x3 = (xnn_zipc_ukernel_function) xnn_x32_zip_x3_ukernel__sse2, |
| 1938 | .x4 = (xnn_zipc_ukernel_function) xnn_x32_zip_x4_ukernel__sse2, |
| 1939 | .xm = (xnn_zipv_ukernel_function) xnn_x32_zip_xm_ukernel__sse2, |
| 1940 | }; |
| 1941 | #endif // XNN_NO_X32_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1942 | |
Marat Dukhan | f42facc | 2020-03-08 15:14:53 -0700 | [diff] [blame] | 1943 | #elif XNN_ARCH_WASMSIMD |
Marat Dukhan | 07e5040 | 2020-08-05 17:16:53 -0700 | [diff] [blame] | 1944 | /**************************** QS8 micro-kernels ****************************/ |
| 1945 | #ifndef XNN_NO_QS8_OPERATORS |
| 1946 | init_flags |= XNN_INIT_FLAG_QS8; |
| 1947 | |
| 1948 | xnn_params.qs8.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_3x4c8__wasmsimd_ld64); |
| 1949 | xnn_params.qs8.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_3x4c8__wasmsimd_ld64); |
| 1950 | xnn_params.qs8.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_ukernel_1x4c8__wasmsimd_ld64); |
| 1951 | xnn_params.qs8.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_ukernel_1x4c8__wasmsimd_ld64); |
| 1952 | xnn_params.qs8.gemm.mr = 3; |
| 1953 | xnn_params.qs8.gemm.nr = 4; |
| 1954 | xnn_params.qs8.gemm.log2_kr = 3; |
| 1955 | |
| 1956 | xnn_params.qs8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qs8_dwconv_minmax_ukernel_up8x9__wasmsimd_mul16; |
| 1957 | xnn_params.qs8.dwconv[0].channel_tile = 8; |
| 1958 | xnn_params.qs8.dwconv[0].primary_tile = 9; |
Marat Dukhan | 9e0b539 | 2020-08-07 02:29:34 -0700 | [diff] [blame] | 1959 | |
| 1960 | xnn_params.qs8.gavgpool = (struct gavgpool_parameters) { |
| 1961 | .up = (xnn_gavgpool_unipass_ukernel_function) xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2, |
| 1962 | .mp = (xnn_gavgpool_multipass_ukernel_function) xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2, |
| 1963 | .mr = 7, |
| 1964 | }; |
Marat Dukhan | ff20948 | 2020-09-03 14:26:53 -0700 | [diff] [blame] | 1965 | |
| 1966 | xnn_params.qs8.vadd = (struct vbinary_parameters) { |
| 1967 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vadd_minmax_ukernel__wasmsimd_x8, |
| 1968 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x8, |
| 1969 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x8, |
| 1970 | .element_tile = 8, |
| 1971 | }; |
Marat Dukhan | 07e5040 | 2020-08-05 17:16:53 -0700 | [diff] [blame] | 1972 | #endif // XNN_NO_QS8_OPERATORS |
| 1973 | |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 1974 | /**************************** QU8 micro-kernels ****************************/ |
| 1975 | #ifndef XNN_NO_QU8_OPERATORS |
| 1976 | init_flags |= XNN_INIT_FLAG_QU8; |
Marat Dukhan | 854fb6b | 2020-06-19 12:33:44 -0700 | [diff] [blame] | 1977 | |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 1978 | xnn_params.qu8.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_ukernel_2x2__scalar); |
| 1979 | xnn_params.qu8.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_ukernel_2x2__scalar); |
| 1980 | xnn_params.qu8.gemm.mr = 2; |
| 1981 | xnn_params.qu8.gemm.nr = 2; |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 1982 | |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 1983 | xnn_params.qu8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qu8_dwconv_minmax_ukernel_up1x9__scalar; |
| 1984 | xnn_params.qu8.dwconv[0].channel_tile = 1; |
| 1985 | xnn_params.qu8.dwconv[0].primary_tile = 9; |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 1986 | |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 1987 | xnn_params.qu8.avgpool = (struct avgpool_parameters) { |
| 1988 | .up = (xnn_avgpool_unipass_ukernel_function) xnn_qu8_avgpool_minmax_ukernel_9x__scalar_c1, |
| 1989 | .mp = (xnn_avgpool_multipass_ukernel_function) xnn_qu8_avgpool_minmax_ukernel_9p8x__scalar_c1, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1990 | .mr = 9, |
| 1991 | .qr = 8, |
| 1992 | }; |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 1993 | xnn_params.qu8.gavgpool = (struct gavgpool_parameters) { |
| 1994 | .up = (xnn_gavgpool_unipass_ukernel_function) xnn_qu8_gavgpool_minmax_ukernel_7x__scalar_c1, |
| 1995 | .mp = (xnn_gavgpool_multipass_ukernel_function) xnn_qu8_gavgpool_minmax_ukernel_7p7x__scalar_c1, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 1996 | .mr = 7, |
| 1997 | }; |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 1998 | xnn_params.qu8.vadd = (xnn_vadd_ukernel_function) xnn_qu8_vadd_minmax_ukernel__scalar; |
| 1999 | #endif // XNN_NO_QU8_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2000 | |
| 2001 | /**************************** U8 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2002 | #ifndef XNN_NO_U8_OPERATORS |
Marat Dukhan | 854fb6b | 2020-06-19 12:33:44 -0700 | [diff] [blame] | 2003 | init_flags |= XNN_INIT_FLAG_U8; |
| 2004 | |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2005 | xnn_params.u8.maxpool = (struct maxpool_parameters) { |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 2006 | .ukernel = (xnn_maxpool_ukernel_function) xnn_u8_maxpool_minmax_ukernel_9p8x__scalar_c1, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2007 | .mr = 9, |
| 2008 | .qr = 8, |
| 2009 | }; |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 2010 | xnn_params.u8.clamp = (xnn_univector_ukernel_function) xnn_u8_clamp_ukernel__scalar_x4; |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2011 | xnn_params.u8.lut32norm = xnn_u8_lut32norm_ukernel__scalar; |
| 2012 | xnn_params.u8.rmax = xnn_u8_rmax_ukernel__scalar; |
| 2013 | #endif // XNN_NO_U8_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2014 | |
| 2015 | /**************************** X8 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2016 | #ifndef XNN_NO_X8_OPERATORS |
Marat Dukhan | 854fb6b | 2020-06-19 12:33:44 -0700 | [diff] [blame] | 2017 | init_flags |= XNN_INIT_FLAG_X8; |
| 2018 | |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2019 | xnn_params.x8.lut = xnn_x8_lut_ukernel__scalar; |
| 2020 | xnn_params.x8.zip = (struct zip_parameters) { |
| 2021 | .x2 = (xnn_zipc_ukernel_function) xnn_x8_zip_x2_ukernel__scalar, |
| 2022 | .x3 = (xnn_zipc_ukernel_function) xnn_x8_zip_x3_ukernel__scalar, |
| 2023 | .x4 = (xnn_zipc_ukernel_function) xnn_x8_zip_x4_ukernel__scalar, |
| 2024 | .xm = (xnn_zipv_ukernel_function) xnn_x8_zip_xm_ukernel__scalar, |
| 2025 | }; |
| 2026 | #endif // XNN_NO_X8_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2027 | |
| 2028 | /**************************** F32 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2029 | #ifndef XNN_NO_F32_OPERATORS |
Marat Dukhan | 854fb6b | 2020-06-19 12:33:44 -0700 | [diff] [blame] | 2030 | init_flags |= XNN_INIT_FLAG_F32; |
| 2031 | |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2032 | if (is_wasm_x86) { |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 2033 | xnn_params.f32.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x8__wasmsimd_splat_x86); |
| 2034 | xnn_params.f32.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x8__wasmsimd_splat_x86); |
| 2035 | xnn_params.f32.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__wasmsimd_splat_x86); |
| 2036 | xnn_params.f32.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__wasmsimd_splat_x86); |
Marat Dukhan | 688f6d8 | 2020-07-14 17:02:11 -0700 | [diff] [blame] | 2037 | xnn_params.f32.gemm.relu.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_relu_ukernel_4x8__wasmsimd_splat); |
| 2038 | xnn_params.f32.gemm.relu.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_relu_ukernel_4x8__wasmsimd_splat); |
| 2039 | xnn_params.f32.gemm.relu.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_relu_ukernel_1x8__wasmsimd_splat); |
| 2040 | xnn_params.f32.gemm.relu.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_relu_ukernel_1x8__wasmsimd_splat); |
Marat Dukhan | 802808c | 2020-06-16 11:01:17 -0700 | [diff] [blame] | 2041 | xnn_params.f32.gemm.linear.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_4x8__wasmsimd_splat); |
| 2042 | xnn_params.f32.gemm.linear.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_4x8__wasmsimd_splat); |
| 2043 | xnn_params.f32.gemm.linear.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_1x8__wasmsimd_splat); |
| 2044 | xnn_params.f32.gemm.linear.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_1x8__wasmsimd_splat); |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 2045 | xnn_params.f32.gemm.mr = 4; |
| 2046 | xnn_params.f32.gemm.nr = 8; |
Marat Dukhan | e39e646 | 2020-07-09 01:33:36 -0700 | [diff] [blame] | 2047 | |
| 2048 | xnn_params.f32.gemm2.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x2c4__wasmsimd_x86); |
| 2049 | xnn_params.f32.gemm2.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x2c4__wasmsimd_x86); |
| 2050 | xnn_params.f32.gemm2.linear.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_4x2c4__wasmsimd); |
| 2051 | xnn_params.f32.gemm2.linear.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_4x2c4__wasmsimd); |
| 2052 | xnn_params.f32.gemm2.mr = 4; |
| 2053 | xnn_params.f32.gemm2.nr = 2; |
| 2054 | xnn_params.f32.gemm2.log2_kr = 2; |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2055 | } else { |
Marat Dukhan | 802808c | 2020-06-16 11:01:17 -0700 | [diff] [blame] | 2056 | xnn_params.f32.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_5x8__wasmsimd_splat_arm); |
| 2057 | xnn_params.f32.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_5x8__wasmsimd_splat_arm); |
| 2058 | xnn_params.f32.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__wasmsimd_splat_arm); |
| 2059 | xnn_params.f32.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__wasmsimd_splat_arm); |
Marat Dukhan | 688f6d8 | 2020-07-14 17:02:11 -0700 | [diff] [blame] | 2060 | xnn_params.f32.gemm.relu.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_relu_ukernel_5x8__wasmsimd_splat); |
| 2061 | xnn_params.f32.gemm.relu.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_relu_ukernel_5x8__wasmsimd_splat); |
| 2062 | xnn_params.f32.gemm.relu.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_relu_ukernel_1x8__wasmsimd_splat); |
| 2063 | xnn_params.f32.gemm.relu.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_relu_ukernel_1x8__wasmsimd_splat); |
Marat Dukhan | 802808c | 2020-06-16 11:01:17 -0700 | [diff] [blame] | 2064 | xnn_params.f32.gemm.linear.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_5x8__wasmsimd_splat); |
| 2065 | xnn_params.f32.gemm.linear.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_5x8__wasmsimd_splat); |
| 2066 | xnn_params.f32.gemm.linear.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_1x8__wasmsimd_splat); |
| 2067 | xnn_params.f32.gemm.linear.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_1x8__wasmsimd_splat); |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 2068 | xnn_params.f32.gemm.mr = 5; |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 2069 | xnn_params.f32.gemm.nr = 8; |
Marat Dukhan | e39e646 | 2020-07-09 01:33:36 -0700 | [diff] [blame] | 2070 | |
| 2071 | xnn_params.f32.gemm2.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x2c4__wasmsimd_arm); |
| 2072 | xnn_params.f32.gemm2.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x2c4__wasmsimd_arm); |
| 2073 | xnn_params.f32.gemm2.linear.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_4x2c4__wasmsimd); |
| 2074 | xnn_params.f32.gemm2.linear.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_4x2c4__wasmsimd); |
| 2075 | xnn_params.f32.gemm2.mr = 4; |
| 2076 | xnn_params.f32.gemm2.nr = 2; |
| 2077 | xnn_params.f32.gemm2.log2_kr = 2; |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2078 | } |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 2079 | |
Marat Dukhan | ac014d7 | 2020-06-16 08:36:47 -0700 | [diff] [blame] | 2080 | if (is_wasm_x86) { |
| 2081 | xnn_params.f32.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86; |
Marat Dukhan | b8e7b07 | 2020-06-16 12:34:23 -0700 | [diff] [blame] | 2082 | xnn_params.f32.dwconv[0].linear.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_ukernel_up8x4__wasmsimd; |
Marat Dukhan | ac014d7 | 2020-06-16 08:36:47 -0700 | [diff] [blame] | 2083 | xnn_params.f32.dwconv[0].channel_tile = 8; |
| 2084 | xnn_params.f32.dwconv[0].primary_tile = 4; |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 2085 | |
Marat Dukhan | ac014d7 | 2020-06-16 08:36:47 -0700 | [diff] [blame] | 2086 | xnn_params.f32.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86; |
Marat Dukhan | b8e7b07 | 2020-06-16 12:34:23 -0700 | [diff] [blame] | 2087 | xnn_params.f32.dwconv[1].linear.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_ukernel_up8x9__wasmsimd; |
Marat Dukhan | ac014d7 | 2020-06-16 08:36:47 -0700 | [diff] [blame] | 2088 | xnn_params.f32.dwconv[1].channel_tile = 8; |
| 2089 | xnn_params.f32.dwconv[1].primary_tile = 9; |
| 2090 | } else { |
| 2091 | xnn_params.f32.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up4x4__wasmsimd_arm; |
Marat Dukhan | b8e7b07 | 2020-06-16 12:34:23 -0700 | [diff] [blame] | 2092 | xnn_params.f32.dwconv[0].linear.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_ukernel_up4x4__wasmsimd; |
Marat Dukhan | ac014d7 | 2020-06-16 08:36:47 -0700 | [diff] [blame] | 2093 | xnn_params.f32.dwconv[0].channel_tile = 4; |
| 2094 | xnn_params.f32.dwconv[0].primary_tile = 4; |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 2095 | |
Marat Dukhan | ac014d7 | 2020-06-16 08:36:47 -0700 | [diff] [blame] | 2096 | xnn_params.f32.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm; |
Marat Dukhan | b8e7b07 | 2020-06-16 12:34:23 -0700 | [diff] [blame] | 2097 | xnn_params.f32.dwconv[1].linear.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_ukernel_up4x9__wasmsimd; |
Marat Dukhan | ac014d7 | 2020-06-16 08:36:47 -0700 | [diff] [blame] | 2098 | xnn_params.f32.dwconv[1].channel_tile = 4; |
| 2099 | xnn_params.f32.dwconv[1].primary_tile = 9; |
| 2100 | } |
| 2101 | |
| 2102 | xnn_params.f32.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm; |
Marat Dukhan | b8e7b07 | 2020-06-16 12:34:23 -0700 | [diff] [blame] | 2103 | xnn_params.f32.dwconv[2].linear.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_ukernel_up4x25__wasmsimd; |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 2104 | xnn_params.f32.dwconv[2].channel_tile = 4; |
| 2105 | xnn_params.f32.dwconv[2].primary_tile = 25; |
| 2106 | |
Marat Dukhan | 3b7432d | 2020-07-16 17:46:32 -0700 | [diff] [blame] | 2107 | if (is_wasm_x86) { |
| 2108 | xnn_params.f32.avgpool = (struct avgpool_parameters) { |
| 2109 | .up = (xnn_avgpool_unipass_ukernel_function) xnn_f32_avgpool_minmax_ukernel_9x__wasmsimd_x86_c4, |
| 2110 | .mp = (xnn_avgpool_multipass_ukernel_function) xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4, |
| 2111 | .mr = 9, |
| 2112 | .qr = 8, |
| 2113 | }; |
Marat Dukhan | 1483c53 | 2020-07-16 18:08:19 -0700 | [diff] [blame] | 2114 | xnn_params.f32.pavgpool = (struct pavgpool_parameters) { |
| 2115 | .up = (xnn_pavgpool_unipass_ukernel_function) xnn_f32_pavgpool_minmax_ukernel_9x__wasmsimd_x86_c4, |
| 2116 | .mp = (xnn_pavgpool_multipass_ukernel_function) xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4, |
| 2117 | .mr = 9, |
| 2118 | .qr = 8, |
| 2119 | }; |
Marat Dukhan | c601680 | 2020-07-16 18:51:28 -0700 | [diff] [blame] | 2120 | xnn_params.f32.gavgpool = (struct gavgpool_parameters) { |
| 2121 | .up = (xnn_gavgpool_unipass_ukernel_function) xnn_f32_gavgpool_minmax_ukernel_7x__wasmsimd_x86_c4, |
| 2122 | .mp = (xnn_gavgpool_multipass_ukernel_function) xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_x86_c4, |
| 2123 | .mr = 7, |
| 2124 | }; |
Marat Dukhan | 3b7432d | 2020-07-16 17:46:32 -0700 | [diff] [blame] | 2125 | } else { |
| 2126 | xnn_params.f32.avgpool = (struct avgpool_parameters) { |
| 2127 | .up = (xnn_avgpool_unipass_ukernel_function) xnn_f32_avgpool_minmax_ukernel_9x__wasmsimd_arm_c4, |
| 2128 | .mp = (xnn_avgpool_multipass_ukernel_function) xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4, |
| 2129 | .mr = 9, |
| 2130 | .qr = 8, |
| 2131 | }; |
Marat Dukhan | 1483c53 | 2020-07-16 18:08:19 -0700 | [diff] [blame] | 2132 | xnn_params.f32.pavgpool = (struct pavgpool_parameters) { |
| 2133 | .up = (xnn_pavgpool_unipass_ukernel_function) xnn_f32_pavgpool_minmax_ukernel_9x__wasmsimd_arm_c4, |
| 2134 | .mp = (xnn_pavgpool_multipass_ukernel_function) xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4, |
| 2135 | .mr = 9, |
| 2136 | .qr = 8, |
| 2137 | }; |
Marat Dukhan | c601680 | 2020-07-16 18:51:28 -0700 | [diff] [blame] | 2138 | xnn_params.f32.gavgpool = (struct gavgpool_parameters) { |
| 2139 | .up = (xnn_gavgpool_unipass_ukernel_function) xnn_f32_gavgpool_minmax_ukernel_7x__wasmsimd_arm_c4, |
| 2140 | .mp = (xnn_gavgpool_multipass_ukernel_function) xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_arm_c4, |
| 2141 | .mr = 7, |
| 2142 | }; |
Marat Dukhan | 3b7432d | 2020-07-16 17:46:32 -0700 | [diff] [blame] | 2143 | } |
Marat Dukhan | f6e2480 | 2020-07-08 22:20:40 -0700 | [diff] [blame] | 2144 | if (is_wasm_x86) { |
| 2145 | xnn_params.f32.maxpool = (struct maxpool_parameters) { |
| 2146 | .ukernel = (xnn_maxpool_ukernel_function) xnn_f32_maxpool_minmax_ukernel_9p8x__wasmsimd_x86_c4, |
| 2147 | .mr = 9, |
| 2148 | .qr = 8, |
| 2149 | }; |
| 2150 | } else { |
| 2151 | xnn_params.f32.maxpool = (struct maxpool_parameters) { |
| 2152 | .ukernel = (xnn_maxpool_ukernel_function) xnn_f32_maxpool_minmax_ukernel_9p8x__wasmsimd_arm_c4, |
| 2153 | .mr = 9, |
| 2154 | .qr = 8, |
| 2155 | }; |
| 2156 | } |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2157 | xnn_params.f32.argmaxpool[0] = (struct argmaxpool_parameters) { |
Marat Dukhan | 40f0552 | 2020-07-16 22:33:12 -0700 | [diff] [blame] | 2158 | .up = (xnn_argmaxpool_unipass_ukernel_function) xnn_f32_argmaxpool_ukernel_4x__wasmsimd_c4, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2159 | .mr = 4, |
| 2160 | }; |
| 2161 | xnn_params.f32.argmaxpool[1] = (struct argmaxpool_parameters) { |
Marat Dukhan | 40f0552 | 2020-07-16 22:33:12 -0700 | [diff] [blame] | 2162 | .up = (xnn_argmaxpool_unipass_ukernel_function) xnn_f32_argmaxpool_ukernel_9x__wasmsimd_c4, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2163 | .mr = 9, |
| 2164 | }; |
| 2165 | xnn_params.f32.argmaxpool[2] = (struct argmaxpool_parameters) { |
Marat Dukhan | 40f0552 | 2020-07-16 22:33:12 -0700 | [diff] [blame] | 2166 | .mp = (xnn_argmaxpool_multipass_ukernel_function) xnn_f32_argmaxpool_ukernel_9p8x__wasmsimd_c4, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2167 | .mr = 9, |
| 2168 | .qr = 8, |
| 2169 | }; |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 2170 | xnn_params.f32.ibilinear = (struct ibilinear_parameters) { |
Marat Dukhan | 00d1d6e | 2020-07-09 01:37:27 -0700 | [diff] [blame] | 2171 | .ukernel = (xnn_ibilinear_ukernel_function) xnn_f32_ibilinear_ukernel__wasmsimd_c8, |
Marat Dukhan | 6972249 | 2019-11-11 19:55:50 -0800 | [diff] [blame] | 2172 | .pixel_tile = 1, |
| 2173 | .channel_tile = 8, |
| 2174 | }; |
Marat Dukhan | 37c8351 | 2020-06-29 13:25:53 -0700 | [diff] [blame] | 2175 | xnn_params.f32.abs = (xnn_univector_ukernel_function) xnn_f32_vabs_ukernel__wasmsimd_x8; |
Marat Dukhan | c303fe6 | 2020-06-26 10:09:25 -0700 | [diff] [blame] | 2176 | if (is_wasm_x86) { |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 2177 | xnn_params.f32.clamp = (xnn_univector_ukernel_function) xnn_f32_clamp_ukernel__wasmsimd_x86_x8; |
Marat Dukhan | c303fe6 | 2020-06-26 10:09:25 -0700 | [diff] [blame] | 2178 | } else { |
Marat Dukhan | 3fa52c8 | 2020-07-08 12:54:33 -0700 | [diff] [blame] | 2179 | xnn_params.f32.clamp = (xnn_univector_ukernel_function) xnn_f32_clamp_ukernel__wasmsimd_arm_x8; |
Marat Dukhan | c303fe6 | 2020-06-26 10:09:25 -0700 | [diff] [blame] | 2180 | } |
Marat Dukhan | 9df9dc6 | 2020-07-10 20:08:49 -0700 | [diff] [blame] | 2181 | xnn_params.f32.hswish = (xnn_univector_ukernel_function) xnn_f32_hswish_ukernel__wasmsimd_x16; |
Marat Dukhan | f4935a2 | 2020-07-16 15:59:10 -0700 | [diff] [blame] | 2182 | if (is_wasm_x86) { |
| 2183 | xnn_params.f32.lrelu = (xnn_univector_ukernel_function) xnn_f32_vlrelu_ukernel__wasmsimd_minmax_x8; |
| 2184 | } else { |
| 2185 | xnn_params.f32.lrelu = (xnn_univector_ukernel_function) xnn_f32_vlrelu_ukernel__wasmsimd_bitselect_x8; |
| 2186 | } |
Marat Dukhan | 37c8351 | 2020-06-29 13:25:53 -0700 | [diff] [blame] | 2187 | xnn_params.f32.neg = (xnn_univector_ukernel_function) xnn_f32_vneg_ukernel__wasmsimd_x8; |
Frank Barchard | 37297a6 | 2020-08-27 12:15:10 -0700 | [diff] [blame] | 2188 | xnn_params.f32.relu = (xnn_univector_ukernel_function) xnn_f32_relu_ukernel__wasmsimd_x16; |
Marat Dukhan | b82b2cd | 2020-07-16 02:23:42 -0700 | [diff] [blame] | 2189 | xnn_params.f32.rndne = (xnn_univector_ukernel_function) xnn_f32_vrndne_ukernel__wasmsimd_addsub_x8; |
| 2190 | if (is_wasm_x86) { |
| 2191 | xnn_params.f32.rndz = (xnn_univector_ukernel_function) xnn_f32_vrndz_ukernel__wasmsimd_addsub_x8; |
| 2192 | } else { |
| 2193 | xnn_params.f32.rndz = (xnn_univector_ukernel_function) xnn_f32_vrndz_ukernel__wasmsimd_cvt_x8; |
| 2194 | } |
| 2195 | xnn_params.f32.rndu = (xnn_univector_ukernel_function) xnn_f32_vrndu_ukernel__wasmsimd_addsub_x8; |
| 2196 | xnn_params.f32.rndd = (xnn_univector_ukernel_function) xnn_f32_vrndd_ukernel__wasmsimd_addsub_x8; |
Marat Dukhan | b3635ed | 2020-07-16 12:36:28 -0700 | [diff] [blame] | 2197 | xnn_params.f32.sigmoid = (xnn_univector_ukernel_function) xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x16; |
Marat Dukhan | 37c8351 | 2020-06-29 13:25:53 -0700 | [diff] [blame] | 2198 | xnn_params.f32.sqr = (xnn_univector_ukernel_function) xnn_f32_vsqr_ukernel__wasmsimd_x8; |
Marat Dukhan | 6804bbd | 2020-06-30 19:26:11 -0700 | [diff] [blame] | 2199 | xnn_params.f32.sqrt = (xnn_univector_ukernel_function) xnn_f32_vsqrt_ukernel__wasmsimd_sqrt_x8; |
Marat Dukhan | 195f8eb | 2020-06-25 12:50:57 -0700 | [diff] [blame] | 2200 | if (is_wasm_x86) { |
| 2201 | xnn_params.f32.prelu = (struct prelu_parameters) { |
Marat Dukhan | 7829928 | 2020-07-15 17:38:06 -0700 | [diff] [blame] | 2202 | .ukernel = (xnn_prelu_ukernel_function) xnn_f32_prelu_ukernel__wasmsimd_minmax_2x8, |
Marat Dukhan | 195f8eb | 2020-06-25 12:50:57 -0700 | [diff] [blame] | 2203 | .row_tile = 2, |
| 2204 | .channel_tile = 8, |
| 2205 | }; |
| 2206 | } else { |
| 2207 | xnn_params.f32.prelu = (struct prelu_parameters) { |
Marat Dukhan | 7829928 | 2020-07-15 17:38:06 -0700 | [diff] [blame] | 2208 | .ukernel = (xnn_prelu_ukernel_function) xnn_f32_prelu_ukernel__wasmsimd_bitselect_2x8, |
Marat Dukhan | 195f8eb | 2020-06-25 12:50:57 -0700 | [diff] [blame] | 2209 | .row_tile = 2, |
| 2210 | .channel_tile = 8, |
| 2211 | }; |
| 2212 | } |
Marat Dukhan | 52238f0 | 2020-07-16 15:30:28 -0700 | [diff] [blame] | 2213 | xnn_params.f32.raddstoreexpminusmax = xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x16_acc2; |
Marat Dukhan | 8c41796 | 2020-07-08 12:27:50 -0700 | [diff] [blame] | 2214 | xnn_params.f32.rmax = xnn_f32_rmax_ukernel__wasmsimd_arm; |
Marat Dukhan | cdc5655 | 2020-06-26 19:49:41 -0700 | [diff] [blame] | 2215 | if (is_wasm_x86) { |
| 2216 | xnn_params.f32.vadd = (struct vbinary_parameters) { |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 2217 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vadd_minmax_ukernel__wasmsimd_x86_x16, |
| 2218 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_minmax_ukernel__wasmsimd_x86_x16, |
| 2219 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_minmax_ukernel__wasmsimd_x86_x16, |
| 2220 | .linear.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vadd_ukernel__wasmsimd_x16, |
| 2221 | .linear.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_ukernel__wasmsimd_x16, |
| 2222 | .linear.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_ukernel__wasmsimd_x16, |
| 2223 | .element_tile = 16, |
Marat Dukhan | cdc5655 | 2020-06-26 19:49:41 -0700 | [diff] [blame] | 2224 | }; |
| 2225 | xnn_params.f32.vdiv = (struct vbinary_parameters) { |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 2226 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdiv_minmax_ukernel__wasmsimd_x86_x4, |
| 2227 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdivc_minmax_ukernel__wasmsimd_x86_x4, |
| 2228 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrdivc_minmax_ukernel__wasmsimd_x86_x4, |
| 2229 | .linear.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdiv_ukernel__wasmsimd_x4, |
| 2230 | .linear.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdivc_ukernel__wasmsimd_x4, |
| 2231 | .linear.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrdivc_ukernel__wasmsimd_x4, |
Marat Dukhan | cdc5655 | 2020-06-26 19:49:41 -0700 | [diff] [blame] | 2232 | .element_tile = 4, |
| 2233 | }; |
| 2234 | xnn_params.f32.vmax = (struct vbinary_parameters) { |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 2235 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmax_ukernel__wasmsimd_x86_x16, |
| 2236 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__wasmsimd_x86_x16, |
| 2237 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__wasmsimd_x86_x16, |
| 2238 | .element_tile = 16, |
Marat Dukhan | cdc5655 | 2020-06-26 19:49:41 -0700 | [diff] [blame] | 2239 | }; |
| 2240 | xnn_params.f32.vmin = (struct vbinary_parameters) { |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 2241 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmin_ukernel__wasmsimd_x86_x16, |
| 2242 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__wasmsimd_x86_x16, |
| 2243 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__wasmsimd_x86_x16, |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 2244 | |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 2245 | .element_tile = 16, |
Marat Dukhan | cdc5655 | 2020-06-26 19:49:41 -0700 | [diff] [blame] | 2246 | }; |
| 2247 | xnn_params.f32.vmul = (struct vbinary_parameters) { |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 2248 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmul_minmax_ukernel__wasmsimd_x86_x16, |
| 2249 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_minmax_ukernel__wasmsimd_x86_x16, |
| 2250 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_minmax_ukernel__wasmsimd_x86_x16, |
| 2251 | .linear.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmul_ukernel__wasmsimd_x16, |
| 2252 | .linear.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_ukernel__wasmsimd_x16, |
| 2253 | .linear.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_ukernel__wasmsimd_x16, |
| 2254 | .element_tile = 16, |
Marat Dukhan | cdc5655 | 2020-06-26 19:49:41 -0700 | [diff] [blame] | 2255 | }; |
| 2256 | xnn_params.f32.vsub = (struct vbinary_parameters) { |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 2257 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsub_minmax_ukernel__wasmsimd_x86_x16, |
| 2258 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsubc_minmax_ukernel__wasmsimd_x86_x16, |
| 2259 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrsubc_minmax_ukernel__wasmsimd_x86_x16, |
| 2260 | .linear.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsub_ukernel__wasmsimd_x16, |
| 2261 | .linear.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsubc_ukernel__wasmsimd_x16, |
| 2262 | .linear.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrsubc_ukernel__wasmsimd_x16, |
| 2263 | .element_tile = 16, |
Marat Dukhan | cdc5655 | 2020-06-26 19:49:41 -0700 | [diff] [blame] | 2264 | }; |
| 2265 | } else { |
| 2266 | xnn_params.f32.vadd = (struct vbinary_parameters) { |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 2267 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vadd_minmax_ukernel__wasmsimd_arm_x16, |
| 2268 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_minmax_ukernel__wasmsimd_arm_x16, |
| 2269 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_minmax_ukernel__wasmsimd_arm_x16, |
| 2270 | .linear.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vadd_ukernel__wasmsimd_x16, |
| 2271 | .linear.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_ukernel__wasmsimd_x16, |
| 2272 | .linear.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_ukernel__wasmsimd_x16, |
| 2273 | .element_tile = 16, |
Marat Dukhan | cdc5655 | 2020-06-26 19:49:41 -0700 | [diff] [blame] | 2274 | }; |
| 2275 | xnn_params.f32.vdiv = (struct vbinary_parameters) { |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 2276 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdiv_minmax_ukernel__wasmsimd_arm_x4, |
| 2277 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdivc_minmax_ukernel__wasmsimd_arm_x4, |
| 2278 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrdivc_minmax_ukernel__wasmsimd_arm_x4, |
| 2279 | .linear.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdiv_ukernel__wasmsimd_x4, |
| 2280 | .linear.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdivc_ukernel__wasmsimd_x4, |
| 2281 | .linear.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrdivc_ukernel__wasmsimd_x4, |
Marat Dukhan | cdc5655 | 2020-06-26 19:49:41 -0700 | [diff] [blame] | 2282 | .element_tile = 4, |
| 2283 | }; |
| 2284 | xnn_params.f32.vmax = (struct vbinary_parameters) { |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 2285 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmax_ukernel__wasmsimd_arm_x16, |
| 2286 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__wasmsimd_arm_x16, |
| 2287 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__wasmsimd_arm_x16, |
| 2288 | .element_tile = 16, |
Marat Dukhan | cdc5655 | 2020-06-26 19:49:41 -0700 | [diff] [blame] | 2289 | }; |
| 2290 | xnn_params.f32.vmin = (struct vbinary_parameters) { |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 2291 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmin_ukernel__wasmsimd_arm_x16, |
| 2292 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__wasmsimd_arm_x16, |
| 2293 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__wasmsimd_arm_x16, |
| 2294 | .element_tile = 16, |
Marat Dukhan | cdc5655 | 2020-06-26 19:49:41 -0700 | [diff] [blame] | 2295 | }; |
| 2296 | xnn_params.f32.vmul = (struct vbinary_parameters) { |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 2297 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmul_minmax_ukernel__wasmsimd_arm_x16, |
| 2298 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_minmax_ukernel__wasmsimd_arm_x16, |
| 2299 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_minmax_ukernel__wasmsimd_arm_x16, |
| 2300 | .linear.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmul_ukernel__wasmsimd_x16, |
| 2301 | .linear.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_ukernel__wasmsimd_x16, |
| 2302 | .linear.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_ukernel__wasmsimd_x16, |
| 2303 | .element_tile = 16, |
Marat Dukhan | cdc5655 | 2020-06-26 19:49:41 -0700 | [diff] [blame] | 2304 | }; |
| 2305 | xnn_params.f32.vsub = (struct vbinary_parameters) { |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 2306 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsub_minmax_ukernel__wasmsimd_arm_x16, |
| 2307 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsubc_minmax_ukernel__wasmsimd_arm_x16, |
| 2308 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrsubc_minmax_ukernel__wasmsimd_arm_x16, |
| 2309 | .linear.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsub_ukernel__wasmsimd_x16, |
| 2310 | .linear.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsubc_ukernel__wasmsimd_x16, |
| 2311 | .linear.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrsubc_ukernel__wasmsimd_x16, |
| 2312 | .element_tile = 16, |
Marat Dukhan | cdc5655 | 2020-06-26 19:49:41 -0700 | [diff] [blame] | 2313 | }; |
| 2314 | } |
Marat Dukhan | f739926 | 2020-06-05 10:58:44 -0700 | [diff] [blame] | 2315 | xnn_params.f32.vsqrdiff = (struct vbinary_parameters) { |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 2316 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiff_ukernel__wasmsimd_x16, |
| 2317 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiffc_ukernel__wasmsimd_x16, |
| 2318 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiffc_ukernel__wasmsimd_x16, |
| 2319 | .element_tile = 16, |
Marat Dukhan | f739926 | 2020-06-05 10:58:44 -0700 | [diff] [blame] | 2320 | }; |
Marat Dukhan | d816f62 | 2020-07-15 10:14:39 -0700 | [diff] [blame] | 2321 | if (is_wasm_x86) { |
| 2322 | xnn_params.f32.vmulcaddc = (struct vmulcaddc_parameters) { |
Marat Dukhan | 9531e9f | 2020-07-24 15:25:02 -0700 | [diff] [blame] | 2323 | .ukernel = (xnn_vmulcaddc_ukernel_function) xnn_f32_vmulcaddc_minmax_ukernel_c4__wasmsimd_x86_2x, |
Marat Dukhan | d816f62 | 2020-07-15 10:14:39 -0700 | [diff] [blame] | 2324 | .channel_tile = 4, |
| 2325 | .row_tile = 2, |
| 2326 | }; |
| 2327 | } else { |
| 2328 | xnn_params.f32.vmulcaddc = (struct vmulcaddc_parameters) { |
Marat Dukhan | 9531e9f | 2020-07-24 15:25:02 -0700 | [diff] [blame] | 2329 | .ukernel = (xnn_vmulcaddc_ukernel_function) xnn_f32_vmulcaddc_minmax_ukernel_c4__wasmsimd_arm_2x, |
Marat Dukhan | d816f62 | 2020-07-15 10:14:39 -0700 | [diff] [blame] | 2330 | .channel_tile = 4, |
| 2331 | .row_tile = 2, |
| 2332 | }; |
| 2333 | } |
Erich Elsen | 6e80fdc | 2020-06-09 15:35:37 -0700 | [diff] [blame] | 2334 | #ifndef XNN_NO_NCHW_OPERATORS |
| 2335 | xnn_params.f32.spmm = (struct spmm_parameters) { |
Frank Barchard | 9e05340 | 2020-10-19 15:29:08 -0700 | [diff] [blame] | 2336 | .ukernel = (xnn_spmm_ukernel_function) xnn_f32_spmm_minmax_ukernel_16x1__wasmsimd_x86, |
Erich Elsen | 6e80fdc | 2020-06-09 15:35:37 -0700 | [diff] [blame] | 2337 | .mr = 16, |
| 2338 | .nr = 1, |
| 2339 | }; |
Erich Elsen | 0a1970e | 2020-06-10 09:24:59 -0700 | [diff] [blame] | 2340 | xnn_params.f32.conv_hwc2chw_3x3c3s2 = (struct conv_hwc2chw_parameters) { |
| 2341 | .ukernel_with_symm_padding = |
| 2342 | (xnn_conv_hwc2chw_ukernel_function) xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__psimd_2x2, |
| 2343 | .output_channel_tile = 4, |
| 2344 | .output_height_tile = 2, |
| 2345 | .output_width_tile = 2, |
| 2346 | }; |
Erich Elsen | 2892889 | 2020-06-12 08:08:19 -0700 | [diff] [blame] | 2347 | xnn_params.f32.dwconv_chw_3x3 = (struct dwconv_chw_parameters) { |
Marat Dukhan | 6f469a5 | 2020-10-21 20:02:52 -0700 | [diff] [blame] | 2348 | .ukernel = (xnn_dwconv_chw_ukernel_function) xnn_f32_dwconv_chw_ukernel_3x3p1__psimd_1x4_acc3, |
Erich Elsen | 2892889 | 2020-06-12 08:08:19 -0700 | [diff] [blame] | 2349 | .input_width_tile = 4, |
| 2350 | .output_width_tile = 4, |
| 2351 | .output_height_tile = 1, |
| 2352 | }; |
| 2353 | xnn_params.f32.dwconv_chw_3x3s2 = (struct dwconv_chw_parameters) { |
Marat Dukhan | 6f469a5 | 2020-10-21 20:02:52 -0700 | [diff] [blame] | 2354 | .ukernel = (xnn_dwconv_chw_ukernel_function) xnn_f32_dwconv_chw_ukernel_3x3s2p1__psimd_1x4_acc3, |
Erich Elsen | 2892889 | 2020-06-12 08:08:19 -0700 | [diff] [blame] | 2355 | .input_width_tile = 4, |
| 2356 | .output_width_tile = 4, |
| 2357 | .output_height_tile = 1, |
| 2358 | }; |
| 2359 | xnn_params.f32.dwconv_chw_5x5 = (struct dwconv_chw_parameters) { |
Marat Dukhan | 1c6cad9 | 2020-10-21 15:48:21 -0700 | [diff] [blame] | 2360 | .ukernel = (xnn_dwconv_chw_ukernel_function) xnn_f32_dwconv_chw_ukernel_5x5p2__psimd_3x4, |
Erich Elsen | 2892889 | 2020-06-12 08:08:19 -0700 | [diff] [blame] | 2361 | .input_width_tile = 4, |
| 2362 | .output_width_tile = 4, |
| 2363 | .output_height_tile = 3, |
| 2364 | }; |
Erich Elsen | 7465a89 | 2020-06-13 14:02:04 -0700 | [diff] [blame] | 2365 | xnn_params.f32.dwconv_chw_5x5s2 = (struct dwconv_chw_parameters) { |
Marat Dukhan | 6f469a5 | 2020-10-21 20:02:52 -0700 | [diff] [blame] | 2366 | .ukernel = (xnn_dwconv_chw_ukernel_function) xnn_f32_dwconv_chw_ukernel_5x5s2p2__psimd_1x4_acc2, |
Erich Elsen | 7465a89 | 2020-06-13 14:02:04 -0700 | [diff] [blame] | 2367 | .input_width_tile = 4, |
| 2368 | .output_width_tile = 4, |
| 2369 | .output_height_tile = 1, |
| 2370 | }; |
Marat Dukhan | c5045bf | 2020-07-27 18:16:35 -0700 | [diff] [blame] | 2371 | if (is_wasm_x86) { |
| 2372 | xnn_params.f32.gavgpool_cw = (struct gavgpool_cw_parameters) { |
| 2373 | .ukernel = (xnn_gavgpool_cw_ukernel_function) xnn_f32_gavgpool_cw_ukernel__wasmsimd_x86_x4, |
| 2374 | .channel_tile = 4, |
| 2375 | }; |
| 2376 | } else { |
| 2377 | xnn_params.f32.gavgpool_cw = (struct gavgpool_cw_parameters) { |
| 2378 | .ukernel = (xnn_gavgpool_cw_ukernel_function) xnn_f32_gavgpool_cw_ukernel__wasmsimd_arm_x4, |
| 2379 | .channel_tile = 4, |
| 2380 | }; |
| 2381 | } |
Erich Elsen | 6e80fdc | 2020-06-09 15:35:37 -0700 | [diff] [blame] | 2382 | #endif // XNN_NO_NCHW_OPERATORS |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2383 | #endif // XNN_NO_F32_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2384 | |
| 2385 | /**************************** X32 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2386 | #ifndef XNN_NO_X32_OPERATORS |
Marat Dukhan | 854fb6b | 2020-06-19 12:33:44 -0700 | [diff] [blame] | 2387 | init_flags |= XNN_INIT_FLAG_X32; |
| 2388 | |
Marat Dukhan | 4662b19 | 2020-05-21 15:52:03 -0700 | [diff] [blame] | 2389 | xnn_params.x32.fill = (struct fill_parameters) { |
Marat Dukhan | 8ee3701 | 2020-07-16 13:17:13 -0700 | [diff] [blame] | 2390 | .ukernel = (xnn_fill_ukernel_function) xnn_x32_fill_ukernel__wasmsimd, |
Marat Dukhan | 4662b19 | 2020-05-21 15:52:03 -0700 | [diff] [blame] | 2391 | .row_tile = 1, |
| 2392 | }; |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2393 | xnn_params.x32.pad = (struct pad_parameters) { |
Marat Dukhan | 9306ae0 | 2020-07-16 15:51:13 -0700 | [diff] [blame] | 2394 | .ukernel = (xnn_pad_ukernel_function) xnn_x32_pad_ukernel__wasmsimd, |
Marat Dukhan | 63523d4 | 2020-05-22 17:07:33 -0700 | [diff] [blame] | 2395 | .row_tile = 1, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2396 | }; |
Marat Dukhan | 9d4bfa2 | 2020-07-16 19:07:04 -0700 | [diff] [blame] | 2397 | xnn_params.x32.unpool = (xnn_unpool_ukernel_function) xnn_x32_unpool_ukernel__wasmsimd; |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2398 | xnn_params.x32.zip = (struct zip_parameters) { |
Marat Dukhan | e3b7876 | 2020-07-16 20:02:58 -0700 | [diff] [blame] | 2399 | .x2 = (xnn_zipc_ukernel_function) xnn_x32_zip_x2_ukernel__wasmsimd, |
| 2400 | .x3 = (xnn_zipc_ukernel_function) xnn_x32_zip_x3_ukernel__wasmsimd, |
| 2401 | .x4 = (xnn_zipc_ukernel_function) xnn_x32_zip_x4_ukernel__wasmsimd, |
| 2402 | .xm = (xnn_zipv_ukernel_function) xnn_x32_zip_xm_ukernel__wasmsimd, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2403 | }; |
| 2404 | #endif // XNN_NO_X32_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2405 | |
Marat Dukhan | a199d49 | 2020-07-24 15:01:25 -0700 | [diff] [blame] | 2406 | #elif XNN_ARCH_WASM |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 2407 | /**************************** QU8 micro-kernels ****************************/ |
| 2408 | #ifndef XNN_NO_QU8_OPERATORS |
| 2409 | init_flags |= XNN_INIT_FLAG_QU8; |
Marat Dukhan | 854fb6b | 2020-06-19 12:33:44 -0700 | [diff] [blame] | 2410 | |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 2411 | xnn_params.qu8.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_ukernel_2x2__scalar); |
| 2412 | xnn_params.qu8.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_ukernel_2x2__scalar); |
| 2413 | xnn_params.qu8.gemm.mr = 2; |
| 2414 | xnn_params.qu8.gemm.nr = 2; |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 2415 | |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 2416 | xnn_params.qu8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qu8_dwconv_minmax_ukernel_up1x9__scalar; |
| 2417 | xnn_params.qu8.dwconv[0].channel_tile = 1; |
| 2418 | xnn_params.qu8.dwconv[0].primary_tile = 9; |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 2419 | |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 2420 | xnn_params.qu8.avgpool = (struct avgpool_parameters) { |
| 2421 | .up = (xnn_avgpool_unipass_ukernel_function) xnn_qu8_avgpool_minmax_ukernel_9x__scalar_c1, |
| 2422 | .mp = (xnn_avgpool_multipass_ukernel_function) xnn_qu8_avgpool_minmax_ukernel_9p8x__scalar_c1, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2423 | .mr = 9, |
| 2424 | .qr = 8, |
| 2425 | }; |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 2426 | xnn_params.qu8.gavgpool = (struct gavgpool_parameters) { |
| 2427 | .up = (xnn_gavgpool_unipass_ukernel_function) xnn_qu8_gavgpool_minmax_ukernel_7x__scalar_c1, |
| 2428 | .mp = (xnn_gavgpool_multipass_ukernel_function) xnn_qu8_gavgpool_minmax_ukernel_7p7x__scalar_c1, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2429 | .mr = 7, |
| 2430 | }; |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 2431 | xnn_params.qu8.vadd = (xnn_vadd_ukernel_function) xnn_qu8_vadd_minmax_ukernel__scalar; |
| 2432 | #endif // XNN_NO_QU8_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2433 | |
| 2434 | /**************************** U8 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2435 | #ifndef XNN_NO_U8_OPERATORS |
Marat Dukhan | 854fb6b | 2020-06-19 12:33:44 -0700 | [diff] [blame] | 2436 | init_flags |= XNN_INIT_FLAG_U8; |
| 2437 | |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2438 | xnn_params.u8.maxpool = (struct maxpool_parameters) { |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 2439 | .ukernel = (xnn_maxpool_ukernel_function) xnn_u8_maxpool_minmax_ukernel_9p8x__scalar_c1, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2440 | .mr = 9, |
| 2441 | .qr = 8, |
| 2442 | }; |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 2443 | xnn_params.u8.clamp = (xnn_univector_ukernel_function) xnn_u8_clamp_ukernel__scalar_x4; |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2444 | xnn_params.u8.lut32norm = xnn_u8_lut32norm_ukernel__scalar; |
| 2445 | xnn_params.u8.rmax = xnn_u8_rmax_ukernel__scalar; |
| 2446 | #endif // XNN_NO_U8_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2447 | |
| 2448 | /**************************** X8 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2449 | #ifndef XNN_NO_X8_OPERATORS |
Marat Dukhan | 854fb6b | 2020-06-19 12:33:44 -0700 | [diff] [blame] | 2450 | init_flags |= XNN_INIT_FLAG_X8; |
| 2451 | |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2452 | xnn_params.x8.lut = xnn_x8_lut_ukernel__scalar; |
| 2453 | xnn_params.x8.zip = (struct zip_parameters) { |
| 2454 | .x2 = (xnn_zipc_ukernel_function) xnn_x8_zip_x2_ukernel__scalar, |
| 2455 | .x3 = (xnn_zipc_ukernel_function) xnn_x8_zip_x3_ukernel__scalar, |
| 2456 | .x4 = (xnn_zipc_ukernel_function) xnn_x8_zip_x4_ukernel__scalar, |
| 2457 | .xm = (xnn_zipv_ukernel_function) xnn_x8_zip_xm_ukernel__scalar, |
| 2458 | }; |
| 2459 | #endif // XNN_NO_X8_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2460 | |
| 2461 | /**************************** F32 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2462 | #ifndef XNN_NO_F32_OPERATORS |
Marat Dukhan | 854fb6b | 2020-06-19 12:33:44 -0700 | [diff] [blame] | 2463 | init_flags |= XNN_INIT_FLAG_F32; |
| 2464 | |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2465 | if (is_wasm_x86) { |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 2466 | xnn_params.f32.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_2x4__scalar); |
| 2467 | xnn_params.f32.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_2x4__scalar); |
| 2468 | xnn_params.f32.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x4__wasm); |
| 2469 | xnn_params.f32.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x4__wasm); |
Marat Dukhan | 467f636 | 2020-05-22 23:21:55 -0700 | [diff] [blame] | 2470 | xnn_params.f32.gemm.relu.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_relu_ukernel_2x4__scalar); |
| 2471 | xnn_params.f32.gemm.relu.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_relu_ukernel_2x4__scalar); |
| 2472 | xnn_params.f32.gemm.relu.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_relu_ukernel_1x4__wasm); |
| 2473 | xnn_params.f32.gemm.relu.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_relu_ukernel_1x4__wasm); |
Marat Dukhan | 869c62d | 2020-04-09 17:17:55 -0700 | [diff] [blame] | 2474 | xnn_params.f32.gemm.linear.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_2x4__scalar); |
| 2475 | xnn_params.f32.gemm.linear.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_2x4__scalar); |
| 2476 | xnn_params.f32.gemm.linear.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_1x4__wasm); |
| 2477 | xnn_params.f32.gemm.linear.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_1x4__wasm); |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 2478 | xnn_params.f32.gemm.mr = 2; |
| 2479 | xnn_params.f32.gemm.nr = 4; |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2480 | } else { |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 2481 | xnn_params.f32.gemm.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x4__wasm); |
| 2482 | xnn_params.f32.gemm.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x4__wasm); |
| 2483 | xnn_params.f32.gemm.minmax.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x4__wasm); |
| 2484 | xnn_params.f32.gemm.minmax.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x4__wasm); |
Marat Dukhan | 467f636 | 2020-05-22 23:21:55 -0700 | [diff] [blame] | 2485 | xnn_params.f32.gemm.relu.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_relu_ukernel_4x4__wasm); |
| 2486 | xnn_params.f32.gemm.relu.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_relu_ukernel_4x4__wasm); |
| 2487 | xnn_params.f32.gemm.relu.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_relu_ukernel_1x4__wasm); |
| 2488 | xnn_params.f32.gemm.relu.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_relu_ukernel_1x4__wasm); |
Marat Dukhan | 869c62d | 2020-04-09 17:17:55 -0700 | [diff] [blame] | 2489 | xnn_params.f32.gemm.linear.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_4x4__wasm); |
| 2490 | xnn_params.f32.gemm.linear.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_4x4__wasm); |
| 2491 | xnn_params.f32.gemm.linear.gemm1 = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_1x4__wasm); |
| 2492 | xnn_params.f32.gemm.linear.igemm1 = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_1x4__wasm); |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 2493 | xnn_params.f32.gemm.mr = 4; |
| 2494 | xnn_params.f32.gemm.nr = 4; |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2495 | } |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 2496 | xnn_params.f32.gemm2.minmax.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x2__wasm); |
| 2497 | xnn_params.f32.gemm2.minmax.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x2__wasm), |
Marat Dukhan | 869c62d | 2020-04-09 17:17:55 -0700 | [diff] [blame] | 2498 | xnn_params.f32.gemm2.linear.gemm = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_4x2__wasm); |
| 2499 | xnn_params.f32.gemm2.linear.igemm = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_4x2__wasm), |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 2500 | xnn_params.f32.gemm2.mr = 4; |
| 2501 | xnn_params.f32.gemm2.nr = 2; |
| 2502 | |
| 2503 | xnn_params.f32.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up1x4__wasm_acc2; |
Marat Dukhan | 869c62d | 2020-04-09 17:17:55 -0700 | [diff] [blame] | 2504 | xnn_params.f32.dwconv[0].linear.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_ukernel_up1x4__wasm_acc2; |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 2505 | xnn_params.f32.dwconv[0].channel_tile = 1; |
| 2506 | xnn_params.f32.dwconv[0].primary_tile = 4; |
| 2507 | |
| 2508 | xnn_params.f32.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up1x9__wasm_acc2; |
Marat Dukhan | 869c62d | 2020-04-09 17:17:55 -0700 | [diff] [blame] | 2509 | xnn_params.f32.dwconv[1].linear.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_ukernel_up1x9__wasm_acc2; |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 2510 | xnn_params.f32.dwconv[1].channel_tile = 1; |
| 2511 | xnn_params.f32.dwconv[1].primary_tile = 9; |
| 2512 | |
| 2513 | xnn_params.f32.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up1x25__wasm_acc2; |
Marat Dukhan | 869c62d | 2020-04-09 17:17:55 -0700 | [diff] [blame] | 2514 | xnn_params.f32.dwconv[2].linear.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_ukernel_up1x25__wasm_acc2; |
Marat Dukhan | aefaef3 | 2020-04-09 07:09:34 -0700 | [diff] [blame] | 2515 | xnn_params.f32.dwconv[2].channel_tile = 1; |
| 2516 | xnn_params.f32.dwconv[2].primary_tile = 25; |
| 2517 | |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2518 | xnn_params.f32.avgpool = (struct avgpool_parameters) { |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 2519 | .up = (xnn_avgpool_unipass_ukernel_function) xnn_f32_avgpool_minmax_ukernel_9x__wasm_c1, |
| 2520 | .mp = (xnn_avgpool_multipass_ukernel_function) xnn_f32_avgpool_minmax_ukernel_9p8x__wasm_c1, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2521 | .mr = 9, |
| 2522 | .qr = 8, |
| 2523 | }; |
| 2524 | xnn_params.f32.pavgpool = (struct pavgpool_parameters) { |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 2525 | .up = (xnn_pavgpool_unipass_ukernel_function) xnn_f32_pavgpool_minmax_ukernel_9x__wasm_c1, |
| 2526 | .mp = (xnn_pavgpool_multipass_ukernel_function) xnn_f32_pavgpool_minmax_ukernel_9p8x__wasm_c1, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2527 | .mr = 9, |
| 2528 | .qr = 8, |
| 2529 | }; |
| 2530 | xnn_params.f32.gavgpool = (struct gavgpool_parameters) { |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 2531 | .up = (xnn_gavgpool_unipass_ukernel_function) xnn_f32_gavgpool_minmax_ukernel_7x__wasm_c1, |
| 2532 | .mp = (xnn_gavgpool_multipass_ukernel_function) xnn_f32_gavgpool_minmax_ukernel_7p7x__wasm_c1, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2533 | .mr = 7, |
| 2534 | }; |
| 2535 | xnn_params.f32.maxpool = (struct maxpool_parameters) { |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 2536 | .ukernel = (xnn_maxpool_ukernel_function) xnn_f32_maxpool_minmax_ukernel_9p8x__wasm_c1, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2537 | .mr = 9, |
| 2538 | .qr = 8, |
| 2539 | }; |
| 2540 | xnn_params.f32.argmaxpool[0] = (struct argmaxpool_parameters) { |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 2541 | .up = (xnn_argmaxpool_unipass_ukernel_function) xnn_f32_argmaxpool_ukernel_4x__scalar_c1, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2542 | .mr = 4, |
| 2543 | }; |
| 2544 | xnn_params.f32.argmaxpool[1] = (struct argmaxpool_parameters) { |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 2545 | .up = (xnn_argmaxpool_unipass_ukernel_function) xnn_f32_argmaxpool_ukernel_9x__scalar_c1, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2546 | .mr = 9, |
| 2547 | }; |
| 2548 | xnn_params.f32.argmaxpool[2] = (struct argmaxpool_parameters) { |
Marat Dukhan | 9993660 | 2020-04-11 16:47:01 -0700 | [diff] [blame] | 2549 | .mp = (xnn_argmaxpool_multipass_ukernel_function) xnn_f32_argmaxpool_ukernel_9p8x__scalar_c1, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2550 | .mr = 9, |
| 2551 | .qr = 8, |
| 2552 | }; |
Marat Dukhan | 660fd19 | 2020-03-10 04:55:30 -0700 | [diff] [blame] | 2553 | xnn_params.f32.ibilinear = (struct ibilinear_parameters) { |
| 2554 | .ukernel = (xnn_ibilinear_ukernel_function) xnn_f32_ibilinear_ukernel__scalar_c2, |
Marat Dukhan | 6972249 | 2019-11-11 19:55:50 -0800 | [diff] [blame] | 2555 | .pixel_tile = 1, |
| 2556 | .channel_tile = 2, |
| 2557 | }; |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 2558 | xnn_params.f32.abs = (xnn_univector_ukernel_function) xnn_f32_vabs_ukernel__scalar_x4; |
Marat Dukhan | 5c5fa96 | 2020-03-10 18:38:33 -0700 | [diff] [blame] | 2559 | xnn_params.f32.clamp = (xnn_univector_ukernel_function) xnn_f32_clamp_ukernel__wasm_x4; |
Marat Dukhan | c303fe6 | 2020-06-26 10:09:25 -0700 | [diff] [blame] | 2560 | if (is_wasm_x86) { |
| 2561 | xnn_params.f32.hswish = (xnn_univector_ukernel_function) xnn_f32_hswish_ukernel__scalar_x4; |
| 2562 | } else { |
| 2563 | xnn_params.f32.hswish = (xnn_univector_ukernel_function) xnn_f32_hswish_ukernel__wasm_x4; |
| 2564 | } |
Marat Dukhan | 2881333 | 2020-06-10 18:05:38 -0700 | [diff] [blame] | 2565 | xnn_params.f32.lrelu = (xnn_univector_ukernel_function) xnn_f32_vlrelu_ukernel__scalar_x4; |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 2566 | xnn_params.f32.neg = (xnn_univector_ukernel_function) xnn_f32_vneg_ukernel__scalar_x4; |
Frank Barchard | 62c5e23 | 2020-07-21 17:42:19 -0700 | [diff] [blame] | 2567 | if (is_wasm_x86) { |
Frank Barchard | 37297a6 | 2020-08-27 12:15:10 -0700 | [diff] [blame] | 2568 | xnn_params.f32.relu = (xnn_univector_ukernel_function) xnn_f32_relu_ukernel__scalar_x8; |
Frank Barchard | 62c5e23 | 2020-07-21 17:42:19 -0700 | [diff] [blame] | 2569 | } else { |
Frank Barchard | 37297a6 | 2020-08-27 12:15:10 -0700 | [diff] [blame] | 2570 | xnn_params.f32.relu = (xnn_univector_ukernel_function) xnn_f32_relu_ukernel__wasm_x8; |
Frank Barchard | 62c5e23 | 2020-07-21 17:42:19 -0700 | [diff] [blame] | 2571 | } |
Marat Dukhan | 64e5251 | 2020-06-09 13:41:16 -0700 | [diff] [blame] | 2572 | xnn_params.f32.rndne = (xnn_univector_ukernel_function) xnn_f32_vrndne_ukernel__scalar_libm_x4; |
| 2573 | xnn_params.f32.rndz = (xnn_univector_ukernel_function) xnn_f32_vrndz_ukernel__scalar_libm_x4; |
| 2574 | xnn_params.f32.rndu = (xnn_univector_ukernel_function) xnn_f32_vrndu_ukernel__scalar_libm_x4; |
| 2575 | xnn_params.f32.rndd = (xnn_univector_ukernel_function) xnn_f32_vrndd_ukernel__scalar_libm_x4; |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 2576 | xnn_params.f32.sigmoid = (xnn_univector_ukernel_function) xnn_f32_sigmoid_ukernel__scalar_lut64_p2_div_x2; |
Marat Dukhan | 5020b96 | 2020-06-08 13:30:10 -0700 | [diff] [blame] | 2577 | xnn_params.f32.sqr = (xnn_univector_ukernel_function) xnn_f32_vsqr_ukernel__scalar_x4; |
Marat Dukhan | 6804bbd | 2020-06-30 19:26:11 -0700 | [diff] [blame] | 2578 | xnn_params.f32.sqrt = (xnn_univector_ukernel_function) xnn_f32_vsqrt_ukernel__scalar_sqrt_x1; |
Marat Dukhan | 7c1f808 | 2020-06-25 13:26:20 -0700 | [diff] [blame] | 2579 | if (is_wasm_x86) { |
| 2580 | xnn_params.f32.prelu = (struct prelu_parameters) { |
| 2581 | .ukernel = (xnn_prelu_ukernel_function) xnn_f32_prelu_ukernel__scalar_2x4, |
| 2582 | .row_tile = 2, |
| 2583 | .channel_tile = 4, |
| 2584 | }; |
| 2585 | } else { |
| 2586 | xnn_params.f32.prelu = (struct prelu_parameters) { |
| 2587 | .ukernel = (xnn_prelu_ukernel_function) xnn_f32_prelu_ukernel__wasm_2x4, |
| 2588 | .row_tile = 2, |
| 2589 | .channel_tile = 4, |
| 2590 | }; |
| 2591 | } |
Marat Dukhan | 1edc454 | 2020-01-27 12:40:13 -0800 | [diff] [blame] | 2592 | xnn_params.f32.raddstoreexpminusmax = xnn_f32_raddstoreexpminusmax_ukernel__scalar_p5_x4_acc2; |
| 2593 | xnn_params.f32.rmax = xnn_f32_rmax_ukernel__scalar; |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 2594 | xnn_params.f32.vadd = (struct vbinary_parameters) { |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 2595 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vadd_minmax_ukernel__wasm_x8, |
| 2596 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_minmax_ukernel__wasm_x8, |
| 2597 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_minmax_ukernel__wasm_x8, |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 2598 | .element_tile = 8, |
| 2599 | }; |
Marat Dukhan | 6918050 | 2019-12-06 15:00:31 -0800 | [diff] [blame] | 2600 | xnn_params.f32.vdiv = (struct vbinary_parameters) { |
Frank Barchard | c67dd7f | 2020-07-06 11:23:57 -0700 | [diff] [blame] | 2601 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdiv_minmax_ukernel__wasm_x2, |
| 2602 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdivc_minmax_ukernel__wasm_x2, |
| 2603 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrdivc_minmax_ukernel__wasm_x2, |
Marat Dukhan | 6918050 | 2019-12-06 15:00:31 -0800 | [diff] [blame] | 2604 | .element_tile = 2, |
| 2605 | }; |
Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 2606 | xnn_params.f32.vmax = (struct vbinary_parameters) { |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 2607 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmax_ukernel__wasm_x8, |
| 2608 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__wasm_x8, |
| 2609 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__wasm_x8, |
Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 2610 | .element_tile = 8, |
| 2611 | }; |
| 2612 | xnn_params.f32.vmin = (struct vbinary_parameters) { |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 2613 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmin_ukernel__wasm_x8, |
| 2614 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__wasm_x8, |
| 2615 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__wasm_x8, |
Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 2616 | .element_tile = 8, |
| 2617 | }; |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 2618 | xnn_params.f32.vmul = (struct vbinary_parameters) { |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 2619 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmul_minmax_ukernel__wasm_x8, |
| 2620 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_minmax_ukernel__wasm_x8, |
| 2621 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_minmax_ukernel__wasm_x8, |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 2622 | .element_tile = 8, |
| 2623 | }; |
Marat Dukhan | 05f3f6d | 2019-12-03 15:13:53 -0800 | [diff] [blame] | 2624 | xnn_params.f32.vsub = (struct vbinary_parameters) { |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 2625 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsub_minmax_ukernel__wasm_x8, |
| 2626 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsubc_minmax_ukernel__wasm_x8, |
| 2627 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrsubc_minmax_ukernel__wasm_x8, |
Marat Dukhan | 05f3f6d | 2019-12-03 15:13:53 -0800 | [diff] [blame] | 2628 | .element_tile = 8, |
| 2629 | }; |
Marat Dukhan | f739926 | 2020-06-05 10:58:44 -0700 | [diff] [blame] | 2630 | xnn_params.f32.vsqrdiff = (struct vbinary_parameters) { |
Frank Barchard | 9c7308f | 2020-08-31 17:03:01 -0700 | [diff] [blame] | 2631 | .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiff_ukernel__scalar_x8, |
| 2632 | .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiffc_ukernel__scalar_x8, |
| 2633 | .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiffc_ukernel__scalar_x8, |
Marat Dukhan | f739926 | 2020-06-05 10:58:44 -0700 | [diff] [blame] | 2634 | .element_tile = 8, |
| 2635 | }; |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2636 | xnn_params.f32.vmulcaddc = (struct vmulcaddc_parameters) { |
Marat Dukhan | 9531e9f | 2020-07-24 15:25:02 -0700 | [diff] [blame] | 2637 | .ukernel = (xnn_vmulcaddc_ukernel_function) xnn_f32_vmulcaddc_minmax_ukernel_c1__wasm_2x, |
Marat Dukhan | 49e6ee9 | 2019-11-06 15:55:29 -0800 | [diff] [blame] | 2638 | .channel_tile = 1, |
| 2639 | .row_tile = 2, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2640 | }; |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2641 | #ifndef XNN_NO_NCHW_OPERATORS |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2642 | xnn_params.f32.spmm = (struct spmm_parameters) { |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 2643 | .ukernel = (xnn_spmm_ukernel_function) xnn_f32_spmm_minmax_ukernel_8x1__scalar, |
Marat Dukhan | bff791e | 2019-10-24 11:05:37 -0700 | [diff] [blame] | 2644 | .mr = 8, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2645 | .nr = 1, |
| 2646 | }; |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 2647 | xnn_params.f32.spmm2 = (struct spmm_parameters) { |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 2648 | .ukernel = (xnn_spmm_ukernel_function) xnn_f32_spmm_minmax_ukernel_8x2__scalar, |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 2649 | .mr = 8, |
| 2650 | .nr = 2, |
| 2651 | }; |
| 2652 | xnn_params.f32.spmm4 = (struct spmm_parameters) { |
Marat Dukhan | 355ab43 | 2020-04-09 19:01:52 -0700 | [diff] [blame] | 2653 | .ukernel = (xnn_spmm_ukernel_function) xnn_f32_spmm_minmax_ukernel_8x4__scalar, |
Erich Elsen | c6afd9b | 2019-10-24 16:10:53 -0700 | [diff] [blame] | 2654 | .mr = 8, |
| 2655 | .nr = 4, |
| 2656 | }; |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 2657 | xnn_params.f32.conv_hwc2chw_3x3c3s2 = (struct conv_hwc2chw_parameters) { |
Marat Dukhan | 14fe0b2 | 2019-10-23 21:20:07 -0700 | [diff] [blame] | 2658 | .ukernel_with_symm_padding = |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 2659 | (xnn_conv_hwc2chw_ukernel_function) xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__scalar_1x1, |
Marat Dukhan | 14fe0b2 | 2019-10-23 21:20:07 -0700 | [diff] [blame] | 2660 | .output_channel_tile = 4, |
| 2661 | .output_height_tile = 1, |
| 2662 | .output_width_tile = 1, |
| 2663 | }; |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 2664 | xnn_params.f32.dwconv_chw_3x3 = (struct dwconv_chw_parameters) { |
Marat Dukhan | 6f469a5 | 2020-10-21 20:02:52 -0700 | [diff] [blame] | 2665 | .ukernel = (xnn_dwconv_chw_ukernel_function) xnn_f32_dwconv_chw_ukernel_3x3p1__scalar_1x1_acc3, |
Marat Dukhan | 14fe0b2 | 2019-10-23 21:20:07 -0700 | [diff] [blame] | 2666 | .input_width_tile = 1, |
| 2667 | .output_width_tile = 1, |
| 2668 | .output_height_tile = 1, |
| 2669 | }; |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 2670 | xnn_params.f32.dwconv_chw_3x3s2 = (struct dwconv_chw_parameters) { |
Marat Dukhan | 6f469a5 | 2020-10-21 20:02:52 -0700 | [diff] [blame] | 2671 | .ukernel = (xnn_dwconv_chw_ukernel_function) xnn_f32_dwconv_chw_ukernel_3x3s2p1__scalar_1x1_acc3, |
Marat Dukhan | 14fe0b2 | 2019-10-23 21:20:07 -0700 | [diff] [blame] | 2672 | .input_width_tile = 1, |
| 2673 | .output_width_tile = 1, |
| 2674 | .output_height_tile = 1, |
| 2675 | }; |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 2676 | xnn_params.f32.dwconv_chw_5x5 = (struct dwconv_chw_parameters) { |
Marat Dukhan | 6f469a5 | 2020-10-21 20:02:52 -0700 | [diff] [blame] | 2677 | .ukernel = (xnn_dwconv_chw_ukernel_function) xnn_f32_dwconv_chw_ukernel_5x5p2__scalar_1x1_acc5, |
Marat Dukhan | a99918a | 2019-11-15 14:40:12 -0800 | [diff] [blame] | 2678 | .input_width_tile = 1, |
| 2679 | .output_width_tile = 1, |
| 2680 | .output_height_tile = 1, |
| 2681 | }; |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 2682 | xnn_params.f32.dwconv_chw_5x5s2 = (struct dwconv_chw_parameters) { |
Marat Dukhan | 6f469a5 | 2020-10-21 20:02:52 -0700 | [diff] [blame] | 2683 | .ukernel = (xnn_dwconv_chw_ukernel_function) xnn_f32_dwconv_chw_ukernel_5x5s2p2__scalar_1x1_acc5, |
Marat Dukhan | a99918a | 2019-11-15 14:40:12 -0800 | [diff] [blame] | 2684 | .input_width_tile = 1, |
| 2685 | .output_width_tile = 1, |
| 2686 | .output_height_tile = 1, |
| 2687 | }; |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 2688 | xnn_params.f32.gavgpool_cw = (struct gavgpool_cw_parameters) { |
| 2689 | .ukernel = (xnn_gavgpool_cw_ukernel_function) xnn_f32_gavgpool_cw_ukernel__scalar_x1, |
Marat Dukhan | 14fe0b2 | 2019-10-23 21:20:07 -0700 | [diff] [blame] | 2690 | .channel_tile = 1, |
| 2691 | }; |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 2692 | #endif // XNN_NO_NCHW_OPERATORS |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2693 | #endif // XNN_NO_F32_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2694 | |
| 2695 | /**************************** X32 micro-kernels ****************************/ |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2696 | #ifndef XNN_NO_X32_OPERATORS |
Marat Dukhan | 854fb6b | 2020-06-19 12:33:44 -0700 | [diff] [blame] | 2697 | init_flags |= XNN_INIT_FLAG_X32; |
| 2698 | |
Marat Dukhan | 4662b19 | 2020-05-21 15:52:03 -0700 | [diff] [blame] | 2699 | xnn_params.x32.fill = (struct fill_parameters) { |
| 2700 | .ukernel = (xnn_fill_ukernel_function) xnn_x32_fill_ukernel__scalar_float, |
| 2701 | .row_tile = 1, |
| 2702 | }; |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2703 | xnn_params.x32.pad = (struct pad_parameters) { |
Marat Dukhan | 63523d4 | 2020-05-22 17:07:33 -0700 | [diff] [blame] | 2704 | .ukernel = (xnn_pad_ukernel_function) xnn_x32_pad_ukernel__scalar_float, |
| 2705 | .row_tile = 1, |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 2706 | }; |
| 2707 | xnn_params.x32.unpool = (xnn_unpool_ukernel_function) xnn_x32_unpool_ukernel__scalar; |
| 2708 | xnn_params.x32.zip = (struct zip_parameters) { |
| 2709 | .x2 = (xnn_zipc_ukernel_function) xnn_x32_zip_x2_ukernel__scalar, |
| 2710 | .x3 = (xnn_zipc_ukernel_function) xnn_x32_zip_x3_ukernel__scalar, |
| 2711 | .x4 = (xnn_zipc_ukernel_function) xnn_x32_zip_x4_ukernel__scalar, |
| 2712 | .xm = (xnn_zipv_ukernel_function) xnn_x32_zip_xm_ukernel__scalar, |
| 2713 | }; |
| 2714 | #endif // XNN_NO_X32_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2715 | |
| 2716 | #else |
| 2717 | #error "Unsupported architecture" |
| 2718 | #endif |
Marat Dukhan | 854fb6b | 2020-06-19 12:33:44 -0700 | [diff] [blame] | 2719 | xnn_params.init_flags = init_flags; |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2720 | } |
| 2721 | |
Marat Dukhan | 57133c0 | 2020-04-13 00:54:59 -0700 | [diff] [blame] | 2722 | #ifdef _WIN32 |
| 2723 | static BOOL CALLBACK init_windows(PINIT_ONCE init_once, PVOID parameter, PVOID* context) { |
| 2724 | init(); |
| 2725 | return TRUE; |
| 2726 | } |
| 2727 | #endif |
| 2728 | |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2729 | enum xnn_status xnn_initialize(const struct xnn_allocator* allocator) { |
Marat Dukhan | d343c22 | 2019-10-07 09:22:14 -0700 | [diff] [blame] | 2730 | #ifndef __EMSCRIPTEN__ |
| 2731 | if (!cpuinfo_initialize()) { |
| 2732 | return xnn_status_out_of_memory; |
| 2733 | } |
| 2734 | #endif |
Marat Dukhan | 57133c0 | 2020-04-13 00:54:59 -0700 | [diff] [blame] | 2735 | #ifdef _WIN32 |
| 2736 | InitOnceExecuteOnce(&init_guard, &init_windows, NULL, NULL); |
| 2737 | #else |
| 2738 | pthread_once(&init_guard, &init); |
| 2739 | #endif |
Marat Dukhan | 854fb6b | 2020-06-19 12:33:44 -0700 | [diff] [blame] | 2740 | if ((xnn_params.init_flags & XNN_INIT_FLAG_XNNPACK) != 0) { |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 2741 | if (allocator != NULL) { |
| 2742 | memcpy(&xnn_params.allocator, allocator, sizeof(struct xnn_allocator)); |
| 2743 | } else { |
| 2744 | xnn_params.allocator.allocate = &xnn_allocate; |
| 2745 | xnn_params.allocator.reallocate = &xnn_reallocate; |
| 2746 | xnn_params.allocator.deallocate = &xnn_deallocate; |
| 2747 | xnn_params.allocator.aligned_allocate = &xnn_aligned_allocate; |
| 2748 | xnn_params.allocator.aligned_deallocate = &xnn_aligned_deallocate; |
| 2749 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2750 | return xnn_status_success; |
| 2751 | } else { |
| 2752 | return xnn_status_unsupported_hardware; |
| 2753 | } |
| 2754 | } |
| 2755 | |
| 2756 | enum xnn_status xnn_deinitialize(void) { |
Marat Dukhan | d343c22 | 2019-10-07 09:22:14 -0700 | [diff] [blame] | 2757 | #ifndef __EMSCRIPTEN__ |
| 2758 | cpuinfo_deinitialize(); |
| 2759 | #endif |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 2760 | return xnn_status_success; |
| 2761 | } |