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 | f62bbdc | 2020-08-04 13:59:04 -0700 | [diff] [blame] | 21 | |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 22 | #define DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(fn_name) \ |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 23 | XNN_INTERNAL void fn_name( \ |
| 24 | size_t channels, \ |
| 25 | size_t output_width, \ |
| 26 | const float** input, \ |
| 27 | const float* weights, \ |
| 28 | float* output, \ |
| 29 | size_t input_stride, \ |
| 30 | size_t output_increment, \ |
Frank Barchard | d536072 | 2020-05-17 16:10:36 -0700 | [diff] [blame] | 31 | size_t input_offset, \ |
| 32 | const float* zero, \ |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 33 | const union xnn_f32_default_params* params); |
| 34 | |
| 35 | #define DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(fn_name) \ |
| 36 | XNN_INTERNAL void fn_name( \ |
| 37 | size_t channels, \ |
| 38 | size_t output_width, \ |
| 39 | const float** input, \ |
| 40 | const float* weights, \ |
| 41 | float* output, \ |
| 42 | size_t input_stride, \ |
| 43 | size_t output_increment, \ |
Frank Barchard | d536072 | 2020-05-17 16:10:36 -0700 | [diff] [blame] | 44 | size_t input_offset, \ |
| 45 | const float* zero, \ |
Marat Dukhan | eb09a6b | 2020-04-08 17:34:32 -0700 | [diff] [blame] | 46 | const union xnn_f32_minmax_params* params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 47 | |
Marat Dukhan | f5425ea | 2020-04-24 01:46:00 -0700 | [diff] [blame] | 48 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x4__neon) |
| 49 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x4__neon_acc2) |
| 50 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x4__neon) |
| 51 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2) |
| 52 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma) |
| 53 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma_acc2) |
| 54 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma) |
| 55 | 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] | 56 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x4__sse) |
| 57 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x4__sse_acc2) |
| 58 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x4__sse) |
| 59 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x4__sse_acc2) |
| 60 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x4__avx) |
| 61 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x4__avx_acc2) |
| 62 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up16x4__avx) |
| 63 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up16x4__avx_acc2) |
| 64 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x4__fma3) |
| 65 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x4__fma3_acc2) |
| 66 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up16x4__fma3) |
| 67 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up16x4__fma3_acc2) |
| 68 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up16x4__avx512f) |
| 69 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up16x4__avx512f_acc2) |
| 70 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f) |
| 71 | 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] | 72 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 73 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x9__neon) |
| 74 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2) |
| 75 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x9__neon) |
| 76 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2) |
| 77 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma) |
| 78 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2) |
| 79 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma) |
| 80 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2) |
| 81 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x9__aarch64_neonfma) |
| 82 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x9__aarch64_neonfma_cortex_a55) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 83 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x9__sse) |
| 84 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x9__sse_acc2) |
| 85 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x9__sse) |
| 86 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2) |
| 87 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x9__avx) |
| 88 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x9__avx_acc2) |
| 89 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up16x9__avx) |
| 90 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up16x9__avx_acc2) |
| 91 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x9__fma3) |
| 92 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x9__fma3_acc2) |
| 93 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up16x9__fma3) |
| 94 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up16x9__fma3_acc2) |
| 95 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f) |
| 96 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f_acc2) |
| 97 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f) |
| 98 | 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] | 99 | |
Marat Dukhan | f5425ea | 2020-04-24 01:46:00 -0700 | [diff] [blame] | 100 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x25__neon) |
| 101 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2) |
| 102 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x25__neon) |
| 103 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2) |
| 104 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma) |
| 105 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2) |
| 106 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma) |
| 107 | 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] | 108 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x25__sse) |
| 109 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2) |
| 110 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x25__sse) |
| 111 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2) |
| 112 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x25__avx) |
| 113 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2) |
| 114 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up16x25__avx) |
| 115 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2) |
| 116 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x25__fma3) |
| 117 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2) |
| 118 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up16x25__fma3) |
| 119 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2) |
| 120 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f) |
| 121 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f_acc2) |
| 122 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f) |
| 123 | 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] | 124 | |
Marat Dukhan | b8e7b07 | 2020-06-16 12:34:23 -0700 | [diff] [blame] | 125 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up4x4__wasmsimd) |
| 126 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up8x4__wasmsimd) |
| 127 | |
| 128 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x4__wasmsimd_arm) |
| 129 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x4__wasmsimd_acc2_arm) |
| 130 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_arm) |
| 131 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_acc2_arm) |
| 132 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x4__wasmsimd_x86) |
| 133 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x4__wasmsimd_acc2_x86) |
| 134 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86) |
| 135 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_acc2_x86) |
| 136 | |
| 137 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up4x9__wasmsimd) |
| 138 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up8x9__wasmsimd) |
| 139 | |
| 140 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm) |
| 141 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_acc2_arm) |
| 142 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm) |
| 143 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_acc2_arm) |
| 144 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86) |
| 145 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_acc2_x86) |
| 146 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86) |
| 147 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_acc2_x86) |
| 148 | |
| 149 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up4x25__wasmsimd) |
| 150 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up8x25__wasmsimd) |
| 151 | |
| 152 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm) |
| 153 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_acc2_arm) |
| 154 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm) |
| 155 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_acc2_arm) |
| 156 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86) |
| 157 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_acc2_x86) |
| 158 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86) |
| 159 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_acc2_x86) |
| 160 | |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 161 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up1x4__wasm) |
| 162 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up1x4__wasm_acc2) |
| 163 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up2x4__wasm) |
| 164 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up2x4__wasm_acc2) |
| 165 | |
| 166 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up1x4__wasm) |
| 167 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up1x4__wasm_acc2) |
| 168 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up2x4__wasm) |
| 169 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up2x4__wasm_acc2) |
| 170 | |
| 171 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up1x9__wasm) |
| 172 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up1x9__wasm_acc2) |
| 173 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up2x9__wasm) |
| 174 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up2x9__wasm_acc2) |
| 175 | |
| 176 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up1x9__wasm) |
| 177 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up1x9__wasm_acc2) |
| 178 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up2x9__wasm) |
| 179 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up2x9__wasm_acc2) |
| 180 | |
| 181 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up1x25__wasm) |
| 182 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up1x25__wasm_acc2) |
| 183 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up2x25__wasm) |
| 184 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up2x25__wasm_acc2) |
| 185 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 186 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up1x25__wasm) |
| 187 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up1x25__wasm_acc2) |
| 188 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up2x25__wasm) |
| 189 | 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] | 190 | |
| 191 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up1x4__scalar) |
| 192 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up1x4__scalar_acc2) |
| 193 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up2x4__scalar) |
| 194 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up2x4__scalar_acc2) |
| 195 | |
| 196 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up1x4__scalar) |
| 197 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up1x4__scalar_acc2) |
| 198 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up2x4__scalar) |
| 199 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up2x4__scalar_acc2) |
| 200 | |
| 201 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up1x9__scalar) |
| 202 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up1x9__scalar_acc2) |
| 203 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up2x9__scalar) |
| 204 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up2x9__scalar_acc2) |
| 205 | |
| 206 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up1x9__scalar) |
| 207 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up1x9__scalar_acc2) |
| 208 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up2x9__scalar) |
| 209 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up2x9__scalar_acc2) |
| 210 | |
| 211 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up1x25__scalar) |
| 212 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up1x25__scalar_acc2) |
| 213 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up2x25__scalar) |
| 214 | DECLARE_F32_DWCONV_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_ukernel_up2x25__scalar_acc2) |
| 215 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 216 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up1x25__scalar) |
| 217 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up1x25__scalar_acc2) |
| 218 | DECLARE_F32_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f32_dwconv_minmax_ukernel_up2x25__scalar) |
| 219 | 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] | 220 | |
Marat Dukhan | f62bbdc | 2020-08-04 13:59:04 -0700 | [diff] [blame] | 221 | |
Frank Barchard | 5a599a6 | 2020-06-04 20:12:44 -0700 | [diff] [blame] | 222 | #define DECLARE_F16_DWCONV_UNIPASS_UKERNEL_FUNCTION(fn_name) \ |
| 223 | XNN_INTERNAL void fn_name( \ |
| 224 | size_t channels, \ |
| 225 | size_t output_width, \ |
| 226 | const void** input, \ |
| 227 | const void* weights, \ |
| 228 | void* output, \ |
| 229 | size_t input_stride, \ |
| 230 | size_t output_increment, \ |
| 231 | size_t input_offset, \ |
| 232 | const void* zero, \ |
| 233 | const struct xnn_f16_default_params* params); |
| 234 | |
| 235 | #define DECLARE_F16_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(fn_name) \ |
| 236 | XNN_INTERNAL void fn_name( \ |
| 237 | size_t channels, \ |
| 238 | size_t output_width, \ |
| 239 | const void** input, \ |
| 240 | const void* weights, \ |
| 241 | void* output, \ |
| 242 | size_t input_stride, \ |
| 243 | size_t output_increment, \ |
| 244 | size_t input_offset, \ |
| 245 | const void* zero, \ |
| 246 | const struct xnn_f16_minmax_params* params); |
| 247 | |
| 248 | DECLARE_F16_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f16_dwconv_minmax_ukernel_up8x4__neonfp16arith) |
| 249 | DECLARE_F16_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f16_dwconv_minmax_ukernel_up8x4__neonfp16arith_acc2) |
| 250 | DECLARE_F16_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith) |
| 251 | DECLARE_F16_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith_acc2) |
| 252 | DECLARE_F16_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f16_dwconv_minmax_ukernel_up8x9__neonfp16arith) |
| 253 | DECLARE_F16_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f16_dwconv_minmax_ukernel_up8x9__neonfp16arith_acc2) |
| 254 | DECLARE_F16_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith) |
| 255 | DECLARE_F16_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith_acc2) |
| 256 | DECLARE_F16_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith) |
| 257 | DECLARE_F16_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2) |
| 258 | DECLARE_F16_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith) |
| 259 | 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] | 260 | |
Marat Dukhan | f62bbdc | 2020-08-04 13:59:04 -0700 | [diff] [blame] | 261 | |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 262 | #define DECLARE_QU8_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(fn_name) \ |
| 263 | XNN_INTERNAL void fn_name( \ |
| 264 | size_t channels, \ |
| 265 | size_t output_width, \ |
| 266 | const uint8_t** input, \ |
| 267 | const void* weights, \ |
| 268 | uint8_t* output, \ |
| 269 | size_t input_stride, \ |
| 270 | size_t output_increment, \ |
| 271 | size_t input_offset, \ |
| 272 | const uint8_t* zero, \ |
| 273 | const union xnn_qu8_gemm_params* params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 274 | |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 275 | DECLARE_QU8_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_qu8_dwconv_minmax_ukernel_up1x9__scalar) |
| 276 | DECLARE_QU8_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_qu8_dwconv_minmax_ukernel_up8x9__neon) |
| 277 | DECLARE_QU8_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 278 | |
| 279 | |
Marat Dukhan | f62bbdc | 2020-08-04 13:59:04 -0700 | [diff] [blame] | 280 | #define DECLARE_QS8_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(fn_name) \ |
| 281 | XNN_INTERNAL void fn_name( \ |
| 282 | size_t channels, \ |
| 283 | size_t output_width, \ |
| 284 | const int8_t** input, \ |
| 285 | const void* weights, \ |
| 286 | int8_t* output, \ |
| 287 | size_t input_stride, \ |
| 288 | size_t output_increment, \ |
| 289 | size_t input_offset, \ |
| 290 | const int8_t* zero, \ |
| 291 | const union xnn_qs8_gemm_params* params); |
| 292 | |
Marat Dukhan | 023bcf9 | 2020-08-10 12:40:50 -0700 | [diff] [blame] | 293 | DECLARE_QS8_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_qs8_dwconv_minmax_ukernel_up8x9__neon_mul16) |
| 294 | DECLARE_QS8_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_qs8_dwconv_minmax_ukernel_up16x9__neon_mul16) |
| 295 | DECLARE_QS8_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_qs8_dwconv_minmax_ukernel_up24x9__neon_mul16) |
| 296 | DECLARE_QS8_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_qs8_dwconv_minmax_ukernel_up32x9__neon_mul16) |
| 297 | |
Marat Dukhan | f62bbdc | 2020-08-04 13:59:04 -0700 | [diff] [blame] | 298 | DECLARE_QS8_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16) |
| 299 | DECLARE_QS8_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16) |
| 300 | DECLARE_QS8_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16) |
| 301 | |
| 302 | DECLARE_QS8_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16) |
| 303 | DECLARE_QS8_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16) |
| 304 | DECLARE_QS8_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16) |
| 305 | |
| 306 | DECLARE_QS8_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16) |
| 307 | DECLARE_QS8_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16) |
| 308 | DECLARE_QS8_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16) |
| 309 | |
Marat Dukhan | 4013552 | 2020-08-07 01:21:00 -0700 | [diff] [blame] | 310 | DECLARE_QS8_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul16) |
| 311 | DECLARE_QS8_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul16) |
| 312 | |
Marat Dukhan | d65a152 | 2020-08-04 19:28:18 -0700 | [diff] [blame] | 313 | DECLARE_QS8_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_qs8_dwconv_minmax_ukernel_up8x9__avx2_mul32) |
| 314 | DECLARE_QS8_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul32) |
| 315 | DECLARE_QS8_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_qs8_dwconv_minmax_ukernel_up24x9__avx2_mul32) |
| 316 | DECLARE_QS8_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul32) |
| 317 | |
Marat Dukhan | 2ffc5e6 | 2020-09-06 22:33:38 -0700 | [diff] [blame] | 318 | DECLARE_QS8_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_qs8_dwconv_minmax_ukernel_up16x9__avx512skx_mul32) |
| 319 | DECLARE_QS8_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_qs8_dwconv_minmax_ukernel_up32x9__avx512skx_mul32) |
| 320 | |
Marat Dukhan | cc8f34c | 2020-08-05 16:36:38 -0700 | [diff] [blame] | 321 | DECLARE_QS8_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_qs8_dwconv_minmax_ukernel_up8x9__wasmsimd_mul16) |
| 322 | DECLARE_QS8_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_qs8_dwconv_minmax_ukernel_up16x9__wasmsimd_mul16) |
| 323 | DECLARE_QS8_DWCONV_MINMAX_UNIPASS_UKERNEL_FUNCTION(xnn_qs8_dwconv_minmax_ukernel_up24x9__wasmsimd_mul16) |
| 324 | |
Marat Dukhan | f62bbdc | 2020-08-04 13:59:04 -0700 | [diff] [blame] | 325 | |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 326 | #define DECLARE_F32_DWCONV2D_CHW_MINMAX_UKERNEL_FUNCTION(fn_name) \ |
| 327 | XNN_INTERNAL void fn_name( \ |
| 328 | size_t input_height, \ |
| 329 | size_t input_width, \ |
| 330 | const float* input, \ |
| 331 | const float* weights, \ |
| 332 | const float* zero, \ |
| 333 | float* output, \ |
| 334 | uint32_t padding_top, \ |
Marat Dukhan | 1f29b80 | 2020-05-15 23:46:39 -0700 | [diff] [blame] | 335 | const union xnn_f32_chw_params* params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 336 | |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 337 | DECLARE_F32_DWCONV2D_CHW_MINMAX_UKERNEL_FUNCTION(xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_3x4) |
Marat Dukhan | 98f2eeb | 2020-10-23 23:13:41 -0700 | [diff] [blame^] | 338 | DECLARE_F32_DWCONV2D_CHW_MINMAX_UKERNEL_FUNCTION(xnn_f32_dwconv2d_chw_ukernel_3x3p1__psimd_1x4_acc3) |
| 339 | DECLARE_F32_DWCONV2D_CHW_MINMAX_UKERNEL_FUNCTION(xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc3) |
Marat Dukhan | 470078a | 2020-10-23 22:36:52 -0700 | [diff] [blame] | 340 | DECLARE_F32_DWCONV2D_CHW_MINMAX_UKERNEL_FUNCTION(xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4) |
| 341 | DECLARE_F32_DWCONV2D_CHW_MINMAX_UKERNEL_FUNCTION(xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4) |
| 342 | DECLARE_F32_DWCONV2D_CHW_MINMAX_UKERNEL_FUNCTION(xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_3x4) |
| 343 | DECLARE_F32_DWCONV2D_CHW_MINMAX_UKERNEL_FUNCTION(xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_4x4) |
| 344 | DECLARE_F32_DWCONV2D_CHW_MINMAX_UKERNEL_FUNCTION(xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4) |
| 345 | DECLARE_F32_DWCONV2D_CHW_MINMAX_UKERNEL_FUNCTION(xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4) |
| 346 | DECLARE_F32_DWCONV2D_CHW_MINMAX_UKERNEL_FUNCTION(xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc2) |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 347 | DECLARE_F32_DWCONV2D_CHW_MINMAX_UKERNEL_FUNCTION(xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc3) |
Marat Dukhan | 470078a | 2020-10-23 22:36:52 -0700 | [diff] [blame] | 348 | DECLARE_F32_DWCONV2D_CHW_MINMAX_UKERNEL_FUNCTION(xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc4) |
| 349 | DECLARE_F32_DWCONV2D_CHW_MINMAX_UKERNEL_FUNCTION(xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4_acc2) |
Marat Dukhan | 98f2eeb | 2020-10-23 23:13:41 -0700 | [diff] [blame^] | 350 | DECLARE_F32_DWCONV2D_CHW_MINMAX_UKERNEL_FUNCTION(xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4) |
| 351 | DECLARE_F32_DWCONV2D_CHW_MINMAX_UKERNEL_FUNCTION(xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4) |
| 352 | DECLARE_F32_DWCONV2D_CHW_MINMAX_UKERNEL_FUNCTION(xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_3x4) |
| 353 | DECLARE_F32_DWCONV2D_CHW_MINMAX_UKERNEL_FUNCTION(xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_4x4) |
| 354 | DECLARE_F32_DWCONV2D_CHW_MINMAX_UKERNEL_FUNCTION(xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_5x4) |
| 355 | DECLARE_F32_DWCONV2D_CHW_MINMAX_UKERNEL_FUNCTION(xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4) |
| 356 | DECLARE_F32_DWCONV2D_CHW_MINMAX_UKERNEL_FUNCTION(xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc2) |
| 357 | DECLARE_F32_DWCONV2D_CHW_MINMAX_UKERNEL_FUNCTION(xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc3) |
| 358 | DECLARE_F32_DWCONV2D_CHW_MINMAX_UKERNEL_FUNCTION(xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc4) |
| 359 | DECLARE_F32_DWCONV2D_CHW_MINMAX_UKERNEL_FUNCTION(xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4_acc2) |
| 360 | |
| 361 | DECLARE_F32_DWCONV2D_CHW_MINMAX_UKERNEL_FUNCTION(xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc3) |
| 362 | DECLARE_F32_DWCONV2D_CHW_MINMAX_UKERNEL_FUNCTION(xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__psimd_1x4_acc3) |
| 363 | DECLARE_F32_DWCONV2D_CHW_MINMAX_UKERNEL_FUNCTION(xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1_acc3) |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 364 | DECLARE_F32_DWCONV2D_CHW_MINMAX_UKERNEL_FUNCTION(xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc3) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 365 | |
Marat Dukhan | 98f2eeb | 2020-10-23 23:13:41 -0700 | [diff] [blame^] | 366 | DECLARE_F32_DWCONV2D_CHW_MINMAX_UKERNEL_FUNCTION(xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4) |
| 367 | DECLARE_F32_DWCONV2D_CHW_MINMAX_UKERNEL_FUNCTION(xnn_f32_dwconv2d_chw_ukernel_5x5p2__psimd_3x4) |
| 368 | DECLARE_F32_DWCONV2D_CHW_MINMAX_UKERNEL_FUNCTION(xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5) |
| 369 | |
| 370 | DECLARE_F32_DWCONV2D_CHW_MINMAX_UKERNEL_FUNCTION(xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2) |
| 371 | DECLARE_F32_DWCONV2D_CHW_MINMAX_UKERNEL_FUNCTION(xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__psimd_1x4_acc2) |
| 372 | DECLARE_F32_DWCONV2D_CHW_MINMAX_UKERNEL_FUNCTION(xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc5) |
| 373 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 374 | |
| 375 | #ifdef __cplusplus |
Marat Dukhan | 80fc932 | 2019-09-29 21:06:36 -0700 | [diff] [blame] | 376 | } // extern "C" |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 377 | #endif |