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 | |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 21 | #define DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(fn_name) \ |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 22 | XNN_INTERNAL void fn_name( \ |
| 23 | size_t channels, \ |
| 24 | size_t output_width, \ |
| 25 | const float** input, \ |
| 26 | const float* weights, \ |
| 27 | float* output, \ |
| 28 | size_t input_stride, \ |
| 29 | size_t output_increment, \ |
Frank Barchard | d536072 | 2020-05-17 16:10:36 -0700 | [diff] [blame] | 30 | size_t input_offset, \ |
| 31 | const float* zero, \ |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 32 | const union xnn_f32_default_params* params); |
| 33 | |
| 34 | #define DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(fn_name) \ |
| 35 | XNN_INTERNAL void fn_name( \ |
| 36 | size_t channels, \ |
| 37 | size_t output_width, \ |
| 38 | const float** input, \ |
| 39 | const float* weights, \ |
| 40 | float* output, \ |
| 41 | size_t input_stride, \ |
| 42 | size_t output_increment, \ |
Frank Barchard | d536072 | 2020-05-17 16:10:36 -0700 | [diff] [blame] | 43 | size_t input_offset, \ |
| 44 | const float* zero, \ |
Marat Dukhan | eb09a6b | 2020-04-08 17:34:32 -0700 | [diff] [blame] | 45 | const union xnn_f32_minmax_params* params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 46 | |
Marat Dukhan | f5425ea | 2020-04-24 01:46:00 -0700 | [diff] [blame] | 47 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x4__neon) |
| 48 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x4__neon_acc2) |
| 49 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x4__neon) |
| 50 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2) |
| 51 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma) |
| 52 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma_acc2) |
| 53 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma) |
| 54 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 55 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x4__psimd) |
| 56 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x4__psimd_acc2) |
| 57 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x4__psimd) |
| 58 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x4__psimd_acc2) |
| 59 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x4__sse) |
| 60 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x4__sse_acc2) |
| 61 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x4__sse) |
| 62 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x4__sse_acc2) |
| 63 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x4__avx) |
| 64 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x4__avx_acc2) |
| 65 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up16x4__avx) |
| 66 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up16x4__avx_acc2) |
| 67 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x4__fma3) |
| 68 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x4__fma3_acc2) |
| 69 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up16x4__fma3) |
| 70 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up16x4__fma3_acc2) |
| 71 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up16x4__avx512f) |
| 72 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up16x4__avx512f_acc2) |
| 73 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f) |
| 74 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f_acc2) |
Marat Dukhan | 5098c3e | 2019-11-07 12:01:19 -0800 | [diff] [blame] | 75 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 76 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x9__neon) |
| 77 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2) |
| 78 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x9__neon) |
| 79 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2) |
| 80 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma) |
| 81 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2) |
| 82 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma) |
| 83 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2) |
| 84 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x9__aarch64_neonfma) |
| 85 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x9__aarch64_neonfma_cortex_a55) |
| 86 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x9__psimd) |
| 87 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x9__psimd_acc2) |
| 88 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x9__psimd) |
| 89 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x9__psimd_acc2) |
| 90 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x9__sse) |
| 91 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x9__sse_acc2) |
| 92 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x9__sse) |
| 93 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2) |
| 94 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x9__avx) |
| 95 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x9__avx_acc2) |
| 96 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up16x9__avx) |
| 97 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up16x9__avx_acc2) |
| 98 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x9__fma3) |
| 99 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x9__fma3_acc2) |
| 100 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up16x9__fma3) |
| 101 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2) |
| 102 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f) |
| 103 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f_acc2) |
| 104 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f) |
| 105 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2) |
Marat Dukhan | 5098c3e | 2019-11-07 12:01:19 -0800 | [diff] [blame] | 106 | |
Marat Dukhan | f5425ea | 2020-04-24 01:46:00 -0700 | [diff] [blame] | 107 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x25__neon) |
| 108 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2) |
| 109 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x25__neon) |
| 110 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2) |
| 111 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma) |
| 112 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2) |
| 113 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma) |
| 114 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 115 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x25__psimd) |
| 116 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x25__psimd_acc2) |
| 117 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x25__psimd) |
| 118 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x25__psimd_acc2) |
| 119 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x25__sse) |
| 120 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2) |
| 121 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x25__sse) |
| 122 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2) |
| 123 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x25__avx) |
| 124 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2) |
| 125 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up16x25__avx) |
| 126 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2) |
| 127 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x25__fma3) |
| 128 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2) |
| 129 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up16x25__fma3) |
| 130 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2) |
| 131 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f) |
| 132 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f_acc2) |
| 133 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f) |
| 134 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2) |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 135 | |
| 136 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up1x4__wasm) |
| 137 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up1x4__wasm_acc2) |
| 138 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up2x4__wasm) |
| 139 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up2x4__wasm_acc2) |
| 140 | |
| 141 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up1x4__wasm) |
| 142 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up1x4__wasm_acc2) |
| 143 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up2x4__wasm) |
| 144 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up2x4__wasm_acc2) |
| 145 | |
| 146 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up1x9__wasm) |
| 147 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up1x9__wasm_acc2) |
| 148 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up2x9__wasm) |
| 149 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up2x9__wasm_acc2) |
| 150 | |
| 151 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up1x9__wasm) |
| 152 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up1x9__wasm_acc2) |
| 153 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up2x9__wasm) |
| 154 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up2x9__wasm_acc2) |
| 155 | |
| 156 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up1x25__wasm) |
| 157 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up1x25__wasm_acc2) |
| 158 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up2x25__wasm) |
| 159 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up2x25__wasm_acc2) |
| 160 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 161 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up1x25__wasm) |
| 162 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up1x25__wasm_acc2) |
| 163 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up2x25__wasm) |
| 164 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up2x25__wasm_acc2) |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 165 | |
| 166 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up1x4__scalar) |
| 167 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up1x4__scalar_acc2) |
| 168 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up2x4__scalar) |
| 169 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up2x4__scalar_acc2) |
| 170 | |
| 171 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up1x4__scalar) |
| 172 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up1x4__scalar_acc2) |
| 173 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up2x4__scalar) |
| 174 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up2x4__scalar_acc2) |
| 175 | |
| 176 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up1x9__scalar) |
| 177 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up1x9__scalar_acc2) |
| 178 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up2x9__scalar) |
| 179 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up2x9__scalar_acc2) |
| 180 | |
| 181 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up1x9__scalar) |
| 182 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up1x9__scalar_acc2) |
| 183 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up2x9__scalar) |
| 184 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up2x9__scalar_acc2) |
| 185 | |
| 186 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up1x25__scalar) |
| 187 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up1x25__scalar_acc2) |
| 188 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up2x25__scalar) |
| 189 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up2x25__scalar_acc2) |
| 190 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 191 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up1x25__scalar) |
| 192 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up1x25__scalar_acc2) |
| 193 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up2x25__scalar) |
| 194 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up2x25__scalar_acc2) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 195 | |
Frank Barchard | 5a599a6 | 2020-06-04 20:12:44 -0700 | [diff] [blame] | 196 | #define DECLARE_F16_DWCONV_UNIPASS_UKERNEL_FUNCTION(fn_name) \ |
| 197 | XNN_INTERNAL void fn_name( \ |
| 198 | size_t channels, \ |
| 199 | size_t output_width, \ |
| 200 | const void** input, \ |
| 201 | const void* weights, \ |
| 202 | void* output, \ |
| 203 | size_t input_stride, \ |
| 204 | size_t output_increment, \ |
| 205 | size_t input_offset, \ |
| 206 | const void* zero, \ |
| 207 | const struct xnn_f16_default_params* params); |
| 208 | |
| 209 | #define DECLARE_F16_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(fn_name) \ |
| 210 | XNN_INTERNAL void fn_name( \ |
| 211 | size_t channels, \ |
| 212 | size_t output_width, \ |
| 213 | const void** input, \ |
| 214 | const void* weights, \ |
| 215 | void* output, \ |
| 216 | size_t input_stride, \ |
| 217 | size_t output_increment, \ |
| 218 | size_t input_offset, \ |
| 219 | const void* zero, \ |
| 220 | const struct xnn_f16_minmax_params* params); |
| 221 | |
| 222 | DECLARE_F16_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f16_dwconv_minmax_ukernel_up8x4__neonfp16arith) |
| 223 | DECLARE_F16_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f16_dwconv_minmax_ukernel_up8x4__neonfp16arith_acc2) |
| 224 | DECLARE_F16_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith) |
| 225 | DECLARE_F16_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith_acc2) |
| 226 | DECLARE_F16_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f16_dwconv_minmax_ukernel_up8x9__neonfp16arith) |
| 227 | DECLARE_F16_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f16_dwconv_minmax_ukernel_up8x9__neonfp16arith_acc2) |
| 228 | DECLARE_F16_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith) |
| 229 | DECLARE_F16_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith_acc2) |
| 230 | DECLARE_F16_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith) |
| 231 | DECLARE_F16_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2) |
| 232 | DECLARE_F16_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith) |
| 233 | DECLARE_F16_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 234 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 235 | #define DECLARE_Q8_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(fn_name) \ |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 236 | XNN_INTERNAL void fn_name( \ |
| 237 | size_t channels, \ |
| 238 | size_t output_width, \ |
| 239 | const uint8_t** input, \ |
| 240 | const void* weights, \ |
| 241 | uint8_t* output, \ |
| 242 | size_t input_stride, \ |
| 243 | size_t output_increment, \ |
Frank Barchard | d536072 | 2020-05-17 16:10:36 -0700 | [diff] [blame] | 244 | size_t input_offset, \ |
| 245 | const uint8_t* zero, \ |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 246 | const union xnn_q8_gemm_params* params); |
| 247 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 248 | DECLARE_Q8_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_q8_dwconv_minmax_ukernel_up1x9__scalar) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 249 | DECLARE_Q8_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_q8_dwconv_minmax_ukernel_up8x9__neon) |
| 250 | DECLARE_Q8_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_q8_dwconv_minmax_ukernel_up8x9__sse2) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 251 | |
| 252 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 253 | #define DECLARE_F32_DWCONV_CHW_UKERNEL_FUNCTION(fn_name) \ |
| 254 | XNN_INTERNAL void fn_name( \ |
| 255 | size_t input_height, \ |
| 256 | size_t input_width, \ |
| 257 | const float* input, \ |
| 258 | const float* weights, \ |
| 259 | const float* zero, \ |
| 260 | float* output, \ |
| 261 | uint32_t padding_top, \ |
| 262 | size_t input_tuple_stride, \ |
| 263 | size_t output_tuple_stride, \ |
| 264 | size_t input_height_stride, \ |
| 265 | size_t output_height_stride, \ |
| 266 | const union xnn_f32_chw_params* params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 267 | |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 268 | DECLARE_F32_DWCONV_CHW_UKERNEL_FUNCTION(xnn_f32_dwconv_chw_ukernel_3x3p1__scalar) |
| 269 | DECLARE_F32_DWCONV_CHW_UKERNEL_FUNCTION(xnn_f32_dwconv_chw_ukernel_5x5p2__scalar) |
| 270 | DECLARE_F32_DWCONV_CHW_UKERNEL_FUNCTION(xnn_f32_dwconv_chw_ukernel_3x3s2p1__scalar) |
| 271 | DECLARE_F32_DWCONV_CHW_UKERNEL_FUNCTION(xnn_f32_dwconv_chw_ukernel_5x5s2p2__scalar) |
Erich Elsen | e6214af | 2020-06-10 22:17:22 -0700 | [diff] [blame] | 272 | DECLARE_F32_DWCONV_CHW_UKERNEL_FUNCTION(xnn_f32_dwconv_chw_ukernel_3x3p1__psimd) |
Erich Elsen | fd7a6e3 | 2020-06-11 12:04:44 -0700 | [diff] [blame] | 273 | DECLARE_F32_DWCONV_CHW_UKERNEL_FUNCTION(xnn_f32_dwconv_chw_ukernel_3x3s2p1__psimd) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 274 | DECLARE_F32_DWCONV_CHW_UKERNEL_FUNCTION(xnn_f32_dwconv_chw_ukernel_3x3p1__neonfma) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 275 | DECLARE_F32_DWCONV_CHW_UKERNEL_FUNCTION(xnn_f32_dwconv_chw_ukernel_3x3s2p1__neonfma) |
Erich Elsen | e6214af | 2020-06-10 22:17:22 -0700 | [diff] [blame] | 276 | DECLARE_F32_DWCONV_CHW_UKERNEL_FUNCTION(xnn_f32_dwconv_chw_ukernel_5x5p2__neonfma) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 277 | DECLARE_F32_DWCONV_CHW_UKERNEL_FUNCTION(xnn_f32_dwconv_chw_ukernel_5x5s2p2__neonfma) |
Erich Elsen | e6214af | 2020-06-10 22:17:22 -0700 | [diff] [blame] | 278 | DECLARE_F32_DWCONV_CHW_UKERNEL_FUNCTION(xnn_f32_dwconv_chw_ukernel_3x3p1__sse) |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 279 | DECLARE_F32_DWCONV_CHW_UKERNEL_FUNCTION(xnn_f32_dwconv_chw_ukernel_3x3s2p1__sse) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 280 | |
| 281 | |
| 282 | #ifdef __cplusplus |
Marat Dukhan | 80fc932 | 2019-09-29 21:06:36 -0700 | [diff] [blame] | 283 | } // extern "C" |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 284 | #endif |