XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1 | // Copyright (c) Facebook, Inc. and its affiliates. |
| 2 | // All rights reserved. |
| 3 | // |
| 4 | // Copyright 2019 Google LLC |
| 5 | // |
| 6 | // This source code is licensed under the BSD-style license found in the |
| 7 | // LICENSE file in the root directory of this source tree. |
| 8 | |
| 9 | #pragma once |
| 10 | |
| 11 | #include <stddef.h> |
| 12 | #include <stdint.h> |
| 13 | |
| 14 | #include <xnnpack/params.h> |
| 15 | #include <xnnpack/common.h> |
| 16 | |
| 17 | #ifdef __cplusplus |
| 18 | extern "C" { |
| 19 | #endif |
| 20 | |
| 21 | |
| 22 | #define DECLARE_F32_GEMM_UKERNEL_FUNCTION(fn_name) \ |
| 23 | XNN_INTERNAL void fn_name( \ |
| 24 | size_t mr, \ |
| 25 | size_t nr, \ |
| 26 | size_t k, \ |
| 27 | const float* a, \ |
| 28 | size_t a_stride, \ |
| 29 | const float* w, \ |
| 30 | float* c, \ |
| 31 | size_t cm_stride, \ |
| 32 | size_t cn_stride, \ |
| 33 | const union xnn_f32_output_params* params); |
| 34 | |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame^] | 35 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x2__neon_lane_ld64) |
| 36 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_1x8__neon_lane_ld64) |
| 37 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x8__neon_lane_ld64) |
| 38 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x8__neon_lane_ld128) |
| 39 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_5x8__neon_lane_ld64) |
| 40 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_6x8__neon_lane_ld64) |
| 41 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_6x8__neon_lane_ld128) |
| 42 | |
| 43 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x2__neonfma_lane_ld64) |
| 44 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_1x8__neonfma_lane_ld64) |
| 45 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x8__neonfma_lane_ld64) |
| 46 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x8__neonfma_lane_ld128) |
| 47 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_5x8__neonfma_lane_ld64) |
| 48 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_6x8__neonfma_lane_ld64) |
| 49 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_6x8__neonfma_lane_ld128) |
| 50 | |
| 51 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_1x8__neon_dup_ld64) |
| 52 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x8__neon_dup_ld64) |
| 53 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x8__neon_dup_ld128) |
| 54 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_6x8__neon_dup_ld64) |
| 55 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_6x8__neon_dup_ld128) |
| 56 | |
| 57 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_1x8__neonfma_dup_ld64) |
| 58 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x8__neonfma_dup_ld64) |
| 59 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x8__neonfma_dup_ld128) |
| 60 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_6x8__neonfma_dup_ld64) |
| 61 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_6x8__neonfma_dup_ld128) |
| 62 | |
| 63 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_1x8s4__neon) |
| 64 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x8s4__neon) |
| 65 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_6x8s4__neon) |
| 66 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_8x8s4__neon) |
| 67 | |
| 68 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_1x8s4__neonfma) |
| 69 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x8s4__neonfma) |
| 70 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_6x8s4__neonfma) |
| 71 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_8x8s4__neonfma) |
| 72 | |
| 73 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x8__aarch32_neon_ld64) |
| 74 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x8__aarch32_neon_cortex_a75) |
| 75 | |
Frank Barchard | 21be34f | 2019-10-09 19:32:19 -0700 | [diff] [blame] | 76 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_1x8__aarch64_neonfma_cortex_a53) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 77 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_1x8__aarch64_neonfma_cortex_a57) |
| 78 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_1x8__aarch64_neonfma_cortex_a75) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame^] | 79 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x8__aarch64_neonfma_ld64) |
| 80 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x8__aarch64_neonfma_ld128) |
Frank Barchard | 46fb807 | 2019-10-25 12:54:22 -0700 | [diff] [blame] | 81 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x8__aarch64_neonfma_cortex_a53) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 82 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x8__aarch64_neonfma_cortex_a57) |
| 83 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x8__aarch64_neonfma_cortex_a75) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame^] | 84 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_1x12__aarch64_neonfma_cortex_a53) |
| 85 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x12__aarch64_neonfma_cortex_a53) |
| 86 | |
| 87 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_1x8__sse_load1) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 88 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x8__sse_load1) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame^] | 89 | |
| 90 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_1x8__sse_dup) |
| 91 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x8__sse_dup) |
| 92 | |
| 93 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_1x8s4__sse) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 94 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x8s4__sse) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame^] | 95 | |
| 96 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_1x8__avx_broadcast) |
| 97 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x8__avx_broadcast) |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 98 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_5x8__avx_broadcast) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame^] | 99 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_6x8__avx_broadcast) |
| 100 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_7x8__avx_broadcast) |
| 101 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_1x16__avx_broadcast) |
| 102 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_3x16__avx_broadcast) |
| 103 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x16__avx_broadcast) |
| 104 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_5x16__avx_broadcast) |
| 105 | |
| 106 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_1x8__fma3_broadcast) |
| 107 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x8__fma3_broadcast) |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 108 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_5x8__fma3_broadcast) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame^] | 109 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_6x8__fma3_broadcast) |
| 110 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_7x8__fma3_broadcast) |
| 111 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_8x8__fma3_broadcast) |
| 112 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_1x16__fma3_broadcast) |
| 113 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_3x16__fma3_broadcast) |
| 114 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x16__fma3_broadcast) |
| 115 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_5x16__fma3_broadcast) |
| 116 | |
| 117 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_1x16__avx512f_broadcast) |
| 118 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x16__avx512f_broadcast) |
| 119 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_5x16__avx512f_broadcast) |
Marat Dukhan | 0f349c4 | 2019-11-27 11:58:54 -0800 | [diff] [blame] | 120 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_6x16__avx512f_broadcast) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame^] | 121 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_7x16__avx512f_broadcast) |
| 122 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_8x16__avx512f_broadcast) |
| 123 | |
| 124 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_1x8__psimd_loadsplat) |
| 125 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x8__psimd_loadsplat) |
| 126 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_6x8__psimd_loadsplat) |
| 127 | |
| 128 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_1x8__psimd_splat) |
| 129 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x8__psimd_splat) |
| 130 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_6x8__psimd_splat) |
| 131 | |
| 132 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_1x8s4__psimd) |
| 133 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x8s4__psimd) |
| 134 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_6x8s4__psimd) |
| 135 | |
| 136 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_5x8__aarch64_neonfma_cortex_a75) |
Frank Barchard | a7fb855 | 2019-10-23 17:14:17 -0700 | [diff] [blame] | 137 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_6x8__aarch64_neonfma_cortex_a53) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 138 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_6x8__aarch64_neonfma_cortex_a57) |
| 139 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_6x8__aarch64_neonfma_cortex_a73) |
| 140 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_6x8__aarch64_neonfma_cortex_a75) |
| 141 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_6x8__aarch64_neonfma_ld128) |
| 142 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_6x8__aarch64_neonfma_ld64) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 143 | |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame^] | 144 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x2__wasm) |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 145 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_1x4__wasm) |
| 146 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_2x4__wasm) |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 147 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x4__wasm) |
| 148 | |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame^] | 149 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x2__scalar) |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 150 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_1x4__scalar) |
| 151 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_2x4__scalar) |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 152 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x4__scalar) |
| 153 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 154 | #define DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(fn_name) \ |
| 155 | XNN_INTERNAL void fn_name( \ |
| 156 | size_t mr, \ |
| 157 | size_t nr, \ |
| 158 | size_t k, \ |
| 159 | const float* a, \ |
| 160 | size_t a_stride, \ |
| 161 | const float* w, \ |
| 162 | float* c, \ |
| 163 | size_t cm_stride, \ |
| 164 | size_t cn_stride, \ |
| 165 | const float* acc, \ |
| 166 | const union xnn_f32_output_params* params); |
| 167 | |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame^] | 168 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_1x8__neon_lane_ld64) |
| 169 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_4x8__neon_lane_ld64) |
| 170 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_4x8__neon_lane_ld128) |
| 171 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_5x8__neon_lane_ld64) |
| 172 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_6x8__neon_lane_ld64) |
| 173 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_6x8__neon_lane_ld128) |
| 174 | |
| 175 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_1x8__neonfma_lane_ld64) |
| 176 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_4x8__neonfma_lane_ld64) |
| 177 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_4x8__neonfma_lane_ld128) |
| 178 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_5x8__neonfma_lane_ld64) |
| 179 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_6x8__neonfma_lane_ld64) |
| 180 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_6x8__neonfma_lane_ld128) |
| 181 | |
| 182 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_1x8__neon_dup_ld64) |
| 183 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_4x8__neon_dup_ld64) |
| 184 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_4x8__neon_dup_ld128) |
| 185 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_6x8__neon_dup_ld64) |
| 186 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_6x8__neon_dup_ld128) |
| 187 | |
| 188 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_1x8__neonfma_dup_ld64) |
| 189 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_4x8__neonfma_dup_ld64) |
| 190 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_4x8__neonfma_dup_ld128) |
| 191 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_6x8__neonfma_dup_ld64) |
| 192 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_6x8__neonfma_dup_ld128) |
| 193 | |
| 194 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_1x8s4__neon) |
| 195 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_4x8s4__neon) |
| 196 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_6x8s4__neon) |
| 197 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_8x8s4__neon) |
| 198 | |
| 199 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_1x8s4__neonfma) |
| 200 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_4x8s4__neonfma) |
| 201 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_6x8s4__neonfma) |
| 202 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_8x8s4__neonfma) |
| 203 | |
Frank Barchard | 21be34f | 2019-10-09 19:32:19 -0700 | [diff] [blame] | 204 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_1x8__aarch64_neonfma_cortex_a53) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 205 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_1x8__aarch64_neonfma_cortex_a57) |
| 206 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_1x8__aarch64_neonfma_cortex_a75) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame^] | 207 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_4x8__aarch64_neonfma_ld64) |
| 208 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_4x8__aarch64_neonfma_ld128) |
Frank Barchard | 46fb807 | 2019-10-25 12:54:22 -0700 | [diff] [blame] | 209 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_4x8__aarch64_neonfma_cortex_a53) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 210 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_4x8__aarch64_neonfma_cortex_a57) |
| 211 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_4x8__aarch64_neonfma_cortex_a75) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 212 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_5x8__aarch64_neonfma_cortex_a75) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame^] | 213 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_6x8__aarch64_neonfma_ld64) |
| 214 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_6x8__aarch64_neonfma_ld128) |
Frank Barchard | a7fb855 | 2019-10-23 17:14:17 -0700 | [diff] [blame] | 215 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_6x8__aarch64_neonfma_cortex_a53) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 216 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_6x8__aarch64_neonfma_cortex_a57) |
| 217 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_6x8__aarch64_neonfma_cortex_a73) |
| 218 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_6x8__aarch64_neonfma_cortex_a75) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame^] | 219 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_1x12__aarch64_neonfma_cortex_a53) |
| 220 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_4x12__aarch64_neonfma_cortex_a53) |
| 221 | |
| 222 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_1x8__sse_load1) |
| 223 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_4x8__sse_load1) |
| 224 | |
| 225 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_1x8__sse_dup) |
| 226 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_4x8__sse_dup) |
| 227 | |
| 228 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_1x8s4__sse) |
| 229 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_4x8s4__sse) |
| 230 | |
| 231 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_1x8__avx_broadcast) |
| 232 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_4x8__avx_broadcast) |
| 233 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_5x8__avx_broadcast) |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 234 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_6x8__avx_broadcast) |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 235 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_7x8__avx_broadcast) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame^] | 236 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_1x16__avx_broadcast) |
| 237 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_3x16__avx_broadcast) |
| 238 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_4x16__avx_broadcast) |
| 239 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_5x16__avx_broadcast) |
| 240 | |
| 241 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_1x8__fma3_broadcast) |
| 242 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_4x8__fma3_broadcast) |
| 243 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_5x8__fma3_broadcast) |
| 244 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_6x8__fma3_broadcast) |
Marat Dukhan | fda12b8 | 2019-11-21 12:27:59 -0800 | [diff] [blame] | 245 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_7x8__fma3_broadcast) |
| 246 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_8x8__fma3_broadcast) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame^] | 247 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_1x16__fma3_broadcast) |
| 248 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_3x16__fma3_broadcast) |
| 249 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_4x16__fma3_broadcast) |
| 250 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_5x16__fma3_broadcast) |
| 251 | |
| 252 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_1x16__avx512f_broadcast) |
| 253 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_4x16__avx512f_broadcast) |
| 254 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_5x16__avx512f_broadcast) |
| 255 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_6x16__avx512f_broadcast) |
| 256 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_7x16__avx512f_broadcast) |
| 257 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_8x16__avx512f_broadcast) |
| 258 | |
| 259 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_1x8__psimd_loadsplat) |
| 260 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_4x8__psimd_loadsplat) |
| 261 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_6x8__psimd_loadsplat) |
| 262 | |
| 263 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_1x8__psimd_splat) |
| 264 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_4x8__psimd_splat) |
| 265 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_6x8__psimd_splat) |
| 266 | |
| 267 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_1x8s4__psimd) |
| 268 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_4x8s4__psimd) |
| 269 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_6x8s4__psimd) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 270 | |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 271 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_1x4__wasm) |
| 272 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_2x4__wasm) |
| 273 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_4x2__wasm) |
| 274 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_4x4__wasm) |
| 275 | |
| 276 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_1x4__scalar) |
| 277 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_2x4__scalar) |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 278 | DECLARE_F32_GEMMINC_UKERNEL_FUNCTION(xnn_f32_gemminc_ukernel_4x4__scalar) |
| 279 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 280 | #define DECLARE_F16_GEMM_UKERNEL_FUNCTION(fn_name) \ |
| 281 | void fn_name( \ |
| 282 | size_t mr, \ |
| 283 | size_t nr, \ |
| 284 | size_t k, \ |
| 285 | const void* a, \ |
| 286 | size_t a_stride, \ |
| 287 | const void* w, \ |
| 288 | void* c, \ |
| 289 | size_t cm_stride, \ |
| 290 | size_t cn_stride, \ |
| 291 | const struct xnn_f16_output_params* params); |
| 292 | |
| 293 | DECLARE_F16_GEMM_UKERNEL_FUNCTION(xnn_f16_gemm_ukernel_4x8__neonfp16arith_ld64) |
| 294 | DECLARE_F16_GEMM_UKERNEL_FUNCTION(xnn_f16_gemm_ukernel_6x8__neonfp16arith_ld64) |
| 295 | DECLARE_F16_GEMM_UKERNEL_FUNCTION(xnn_f16_gemm_ukernel_8x8__neonfp16arith_ld64) |
| 296 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 297 | #define DECLARE_Q8_GEMM_UKERNEL_FUNCTION(fn_name) \ |
| 298 | XNN_INTERNAL void fn_name( \ |
| 299 | size_t mr, \ |
| 300 | size_t nr, \ |
| 301 | size_t k, \ |
| 302 | const uint8_t* a, \ |
| 303 | size_t a_stride, \ |
| 304 | const void* w, \ |
| 305 | uint8_t* c, \ |
| 306 | size_t cm_stride, \ |
| 307 | size_t cn_stride, \ |
| 308 | const union xnn_q8_gemm_params* params); |
| 309 | |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame^] | 310 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 311 | DECLARE_Q8_GEMM_UKERNEL_FUNCTION(xnn_q8_gemm_ukernel_4x8__neon) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 312 | DECLARE_Q8_GEMM_UKERNEL_FUNCTION(xnn_q8_gemm_ukernel_8x8__neon) |
| 313 | |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame^] | 314 | DECLARE_Q8_GEMM_UKERNEL_FUNCTION(xnn_q8_gemm_ukernel_4x8__aarch32_neon) |
| 315 | |
| 316 | DECLARE_Q8_GEMM_UKERNEL_FUNCTION(xnn_q8_gemm_ukernel_8x8__aarch64_neon) |
| 317 | |
| 318 | DECLARE_Q8_GEMM_UKERNEL_FUNCTION(xnn_q8_gemm_ukernel_2x4c8__sse2) |
| 319 | DECLARE_Q8_GEMM_UKERNEL_FUNCTION(xnn_q8_gemm_ukernel_4x4c2__sse2) |
| 320 | |
| 321 | DECLARE_Q8_GEMM_UKERNEL_FUNCTION(xnn_q8_gemm_ukernel_2x2__scalar) |
| 322 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 323 | #ifdef __cplusplus |
Marat Dukhan | 80fc932 | 2019-09-29 21:06:36 -0700 | [diff] [blame] | 324 | } // extern "C" |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 325 | #endif |