Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 1 | // Copyright 2019 Google LLC |
| 2 | // |
| 3 | // This source code is licensed under the BSD-style license found in the |
| 4 | // LICENSE file in the root directory of this source tree. |
| 5 | |
| 6 | #include <assert.h> |
| 7 | |
| 8 | #include <xnnpack/dwconv.h> |
| 9 | #include <xnnpack/math.h> |
| 10 | |
| 11 | |
| 12 | void xnn_f32_dwconv_spchw_ukernel_5x5p2__scalar( |
Erich Elsen | eda9c11 | 2020-05-11 04:40:25 -0700 | [diff] [blame^] | 13 | size_t input_height, |
| 14 | size_t input_width, |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 15 | const float* input, |
| 16 | const float* weights, |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 17 | const float* zero, |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 18 | float* output, |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 19 | uint32_t padding_top, |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 20 | size_t input_tuple_stride, |
| 21 | size_t output_tuple_stride, |
| 22 | size_t input_width_stride, |
| 23 | size_t output_width_stride, |
Marat Dukhan | f196d01 | 2020-04-15 11:50:03 -0700 | [diff] [blame] | 24 | const union xnn_f32_spchw_params params[restrict XNN_MIN_ELEMENTS(1)]) |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 25 | { |
Erich Elsen | eda9c11 | 2020-05-11 04:40:25 -0700 | [diff] [blame^] | 26 | assert(input_width != 0); |
| 27 | assert(input_height != 0); |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 28 | assert(padding_top == 2); |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 29 | |
Erich Elsen | eda9c11 | 2020-05-11 04:40:25 -0700 | [diff] [blame^] | 30 | const size_t padded_input_height = input_height + padding_top + 2 /* padding_bottom */; |
| 31 | const size_t output_height = padded_input_height - 5 + 1; |
| 32 | |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 33 | const float params_max = params->scalar.max; |
| 34 | const float params_min = params->scalar.min; |
| 35 | |
Erich Elsen | eda9c11 | 2020-05-11 04:40:25 -0700 | [diff] [blame^] | 36 | const size_t input_width_decrement_single = input_width * input_tuple_stride; |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 37 | const size_t input_width_increment_single = input_width_stride - input_width_decrement_single;; |
Erich Elsen | eda9c11 | 2020-05-11 04:40:25 -0700 | [diff] [blame^] | 38 | const size_t output_width_increment_single = output_width_stride - (input_width - 1) * output_tuple_stride; |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 39 | |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 40 | const float* i0 = zero; |
| 41 | const float* i1 = zero;; |
| 42 | const float* i2 = input; |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 43 | const float* i3 = (const float*) ((uintptr_t) i2 + input_width_stride); |
| 44 | const float* i4 = (const float*) ((uintptr_t) i3 + input_width_stride); |
Erich Elsen | eda9c11 | 2020-05-11 04:40:25 -0700 | [diff] [blame^] | 45 | if (input_height <= 2) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 46 | i4 = zero; |
| 47 | } |
Erich Elsen | eda9c11 | 2020-05-11 04:40:25 -0700 | [diff] [blame^] | 48 | if (input_height == 1) { |
| 49 | i3 = zero; |
| 50 | } |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 51 | |
| 52 | float* output0 = output; |
| 53 | |
| 54 | // this almost certainly will use too many scalar registers |
| 55 | // hope the compiler is good at spilling... |
| 56 | const float vw0 = weights[0]; |
| 57 | const float vw1 = weights[1]; |
| 58 | const float vw2 = weights[2]; |
| 59 | const float vw3 = weights[3]; |
| 60 | const float vw4 = weights[4]; |
| 61 | const float vw5 = weights[5]; |
| 62 | const float vw6 = weights[6]; |
| 63 | const float vw7 = weights[7]; |
| 64 | const float vw8 = weights[8]; |
| 65 | const float vw9 = weights[9]; |
| 66 | const float vw10 = weights[10]; |
| 67 | const float vw11 = weights[11]; |
| 68 | const float vw12 = weights[12]; |
| 69 | const float vw13 = weights[13]; |
| 70 | const float vw14 = weights[14]; |
| 71 | const float vw15 = weights[15]; |
| 72 | const float vw16 = weights[16]; |
| 73 | const float vw17 = weights[17]; |
| 74 | const float vw18 = weights[18]; |
| 75 | const float vw19 = weights[19]; |
| 76 | const float vw20 = weights[20]; |
| 77 | const float vw21 = weights[21]; |
| 78 | const float vw22 = weights[22]; |
| 79 | const float vw23 = weights[23]; |
| 80 | const float vw24 = weights[24]; |
| 81 | const float vw25 = weights[25]; |
| 82 | |
Erich Elsen | eda9c11 | 2020-05-11 04:40:25 -0700 | [diff] [blame^] | 83 | size_t m = output_height; |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 84 | do { |
| 85 | float vi0x0 = 0.0f; |
| 86 | float vi1x0 = 0.0f; |
| 87 | float vi2x0 = 0.0f; |
| 88 | float vi3x0 = 0.0f; |
| 89 | float vi4x0 = 0.0f; |
| 90 | float vi0x1 = 0.0f; |
| 91 | float vi1x1 = 0.0f; |
| 92 | float vi2x1 = 0.0f; |
| 93 | float vi3x1 = 0.0f; |
| 94 | float vi4x1 = 0.0f; |
| 95 | float vi0x2 = *i0; i0 = (const float*) ((uintptr_t) i0 + input_tuple_stride); |
| 96 | float vi1x2 = *i1; i1 = (const float*) ((uintptr_t) i1 + input_tuple_stride); |
| 97 | float vi2x2 = *i2; i2 = (const float*) ((uintptr_t) i2 + input_tuple_stride); |
| 98 | float vi3x2 = *i3; i3 = (const float*) ((uintptr_t) i3 + input_tuple_stride); |
| 99 | float vi4x2 = *i4; i4 = (const float*) ((uintptr_t) i4 + input_tuple_stride); |
| 100 | |
| 101 | float vi0x3; |
| 102 | float vi1x3; |
| 103 | float vi2x3; |
| 104 | float vi3x3; |
| 105 | float vi4x3; |
Erich Elsen | eda9c11 | 2020-05-11 04:40:25 -0700 | [diff] [blame^] | 106 | if XNN_LIKELY(input_width > 1) { |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 107 | vi0x3 = *i0; i0 = (const float*) ((uintptr_t) i0 + input_tuple_stride); |
| 108 | vi1x3 = *i1; i1 = (const float*) ((uintptr_t) i1 + input_tuple_stride); |
| 109 | vi2x3 = *i2; i2 = (const float*) ((uintptr_t) i2 + input_tuple_stride); |
| 110 | vi3x3 = *i3; i3 = (const float*) ((uintptr_t) i3 + input_tuple_stride); |
| 111 | vi4x3 = *i4; i4 = (const float*) ((uintptr_t) i4 + input_tuple_stride); |
| 112 | } |
| 113 | |
Erich Elsen | eda9c11 | 2020-05-11 04:40:25 -0700 | [diff] [blame^] | 114 | size_t k = input_width; |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 115 | for (; k > 2; k -= 1) { |
| 116 | const float vi0x4 = *i0; i0 = (const float*) ((uintptr_t) i0 + input_tuple_stride); |
| 117 | const float vi1x4 = *i1; i1 = (const float*) ((uintptr_t) i1 + input_tuple_stride); |
| 118 | const float vi2x4 = *i2; i2 = (const float*) ((uintptr_t) i2 + input_tuple_stride); |
| 119 | const float vi3x4 = *i3; i3 = (const float*) ((uintptr_t) i3 + input_tuple_stride); |
| 120 | const float vi4x4 = *i4; i4 = (const float*) ((uintptr_t) i4 + input_tuple_stride); |
| 121 | |
| 122 | const float vrow0_accum = vw1 * vi0x0 + vw2 * vi0x1 + vw3 * vi0x2 + vw4 * vi0x3 + vw5 * vi0x4; |
| 123 | vi0x0 = vi0x1; |
| 124 | vi0x1 = vi0x2; |
| 125 | vi0x2 = vi0x3; |
| 126 | vi0x3 = vi0x4; |
| 127 | const float vrow1_accum = vw6 * vi1x0 + vw7 * vi1x1 + vw8 * vi1x2 + vw9 * vi1x3 + vw10 * vi1x4; |
| 128 | vi1x0 = vi1x1; |
| 129 | vi1x1 = vi1x2; |
| 130 | vi1x2 = vi1x3; |
| 131 | vi1x3 = vi1x4; |
| 132 | const float vrow2_accum = vw11 * vi2x0 + vw12 * vi2x1 + vw13 * vi2x2 + vw14 * vi2x3 + vw15 * vi2x4; |
| 133 | vi2x0 = vi2x1; |
| 134 | vi2x1 = vi2x2; |
| 135 | vi2x2 = vi2x3; |
| 136 | vi2x3 = vi2x4; |
| 137 | const float vrow3_accum = vw16 * vi3x0 + vw17 * vi3x1 + vw18 * vi3x2 + vw19 * vi3x3 + vw20 * vi3x4; |
| 138 | vi3x0 = vi3x1; |
| 139 | vi3x1 = vi3x2; |
| 140 | vi3x2 = vi3x3; |
| 141 | vi3x3 = vi3x4; |
| 142 | const float vrow4_accum = vw21 * vi4x0 + vw22 * vi4x1 + vw23 * vi4x2 + vw24 * vi4x3 + vw25 * vi4x4; |
| 143 | vi4x0 = vi4x1; |
| 144 | vi4x1 = vi4x2; |
| 145 | vi4x2 = vi4x3; |
| 146 | vi4x3 = vi4x4; |
| 147 | |
| 148 | float voutput = (vw0 + vrow0_accum) + (vrow1_accum + vrow2_accum) + (vrow3_accum + vrow4_accum); |
| 149 | |
| 150 | voutput = math_max_f32(voutput, params_min); |
| 151 | voutput = math_min_f32(voutput, params_max); |
| 152 | |
| 153 | *output0 = voutput; output0 = (float*) ((uintptr_t) output0 + output_tuple_stride); |
| 154 | } |
| 155 | if XNN_LIKELY(k > 1) { |
| 156 | const float vrow0_accum = vw1 * vi0x0 + vw2 * vi0x1 + vw3 * vi0x2 + vw4 * vi0x3; |
| 157 | vi0x0 = vi0x1; |
| 158 | vi0x1 = vi0x2; |
| 159 | vi0x2 = vi0x3; |
| 160 | const float vrow1_accum = vw6 * vi1x0 + vw7 * vi1x1 + vw8 * vi1x2 + vw9 * vi1x3; |
| 161 | vi1x0 = vi1x1; |
| 162 | vi1x1 = vi1x2; |
| 163 | vi1x2 = vi1x3; |
| 164 | const float vrow2_accum = vw11 * vi2x0 + vw12 * vi2x1 + vw13 * vi2x2 + vw14 * vi2x3; |
| 165 | vi2x0 = vi2x1; |
| 166 | vi2x1 = vi2x2; |
| 167 | vi2x2 = vi2x3; |
| 168 | const float vrow3_accum = vw16 * vi3x0 + vw17 * vi3x1 + vw18 * vi3x2 + vw19 * vi3x3; |
| 169 | vi3x0 = vi3x1; |
| 170 | vi3x1 = vi3x2; |
| 171 | vi3x2 = vi3x3; |
| 172 | const float vrow4_accum = vw21 * vi4x0 + vw22 * vi4x1 + vw23 * vi4x2 + vw24 * vi4x3; |
| 173 | vi4x0 = vi4x1; |
| 174 | vi4x1 = vi4x2; |
| 175 | vi4x2 = vi4x3; |
| 176 | |
| 177 | float voutput = (vw0 + vrow0_accum) + (vrow1_accum + vrow2_accum) + (vrow3_accum + vrow4_accum); |
| 178 | |
| 179 | voutput = math_max_f32(voutput, params_min); |
| 180 | voutput = math_min_f32(voutput, params_max); |
| 181 | |
| 182 | *output0 = voutput; output0 = (float*) ((uintptr_t) output0 + output_tuple_stride); |
| 183 | k -= 1; |
| 184 | } |
| 185 | assert(k == 1); |
| 186 | { |
| 187 | const float vrow0_accum = vw1 * vi0x0 + vw2 * vi0x1 + vw3 * vi0x2; |
| 188 | const float vrow1_accum = vw6 * vi1x0 + vw7 * vi1x1 + vw8 * vi1x2; |
| 189 | const float vrow2_accum = vw11 * vi2x0 + vw12 * vi2x1 + vw13 * vi2x2; |
| 190 | const float vrow3_accum = vw16 * vi3x0 + vw17 * vi3x1 + vw18 * vi3x2; |
| 191 | const float vrow4_accum = vw21 * vi4x0 + vw22 * vi4x1 + vw23 * vi4x2; |
| 192 | |
| 193 | float voutput = (vw0 + vrow0_accum) + (vrow1_accum + vrow2_accum) + (vrow3_accum + vrow4_accum); |
| 194 | |
| 195 | voutput = math_max_f32(voutput, params_min); |
| 196 | voutput = math_min_f32(voutput, params_max); |
| 197 | |
| 198 | *output0 = voutput;; |
| 199 | } |
| 200 | |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 201 | i0 = (const float*) ((uintptr_t) i1 - input_width_decrement_single); |
| 202 | i1 = (const float*) ((uintptr_t) i2 - input_width_decrement_single); |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 203 | i2 = (const float*) ((uintptr_t) i2 + input_width_increment_single); |
| 204 | i3 = (const float*) ((uintptr_t) i3 + input_width_increment_single); |
| 205 | i4 = (const float*) ((uintptr_t) i4 + input_width_increment_single); |
| 206 | output0 = (float*) ((uintptr_t) output0 + output_width_increment_single); |
| 207 | m -= 1; |
Erich Elsen | eda9c11 | 2020-05-11 04:40:25 -0700 | [diff] [blame^] | 208 | if (m <= 2) { |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 209 | i4 = zero; |
Erich Elsen | eda9c11 | 2020-05-11 04:40:25 -0700 | [diff] [blame^] | 210 | } |
| 211 | if (m == 1) { |
| 212 | i3 = zero; |
Erich Elsen | 4e5db3d | 2020-05-07 08:57:47 -0700 | [diff] [blame] | 213 | } |
Erich Elsen | 38709a6 | 2019-11-08 11:58:45 -0800 | [diff] [blame] | 214 | } while (m > 0); |
| 215 | } |