XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1 | // Auto-generated file. Do not edit! |
| 2 | // Template: src/f32-igemm/scalar.c.in |
| 3 | // Generator: tools/xngen |
| 4 | // |
| 5 | // Copyright 2019 Google LLC |
| 6 | // |
| 7 | // This source code is licensed under the BSD-style license found in the |
| 8 | // LICENSE file in the root directory of this source tree. |
| 9 | |
| 10 | #include <assert.h> |
| 11 | |
| 12 | #include <xnnpack/igemm.h> |
| 13 | #include <xnnpack/math.h> |
| 14 | |
| 15 | |
| 16 | void xnn_f32_igemm_ukernel_4x4__scalar( |
| 17 | size_t mr, |
| 18 | size_t nc, |
| 19 | size_t kc, |
| 20 | size_t ks, |
| 21 | const float**restrict a, |
| 22 | const float*restrict w, |
| 23 | float*restrict c, |
| 24 | size_t cm_stride, |
| 25 | size_t cn_stride, |
| 26 | size_t a_offset, |
| 27 | const float* zero, |
| 28 | const union xnn_f32_output_params params[restrict static 1]) |
| 29 | { |
| 30 | assert(mr != 0); |
| 31 | assert(mr <= 4); |
| 32 | assert(nc != 0); |
| 33 | assert(kc != 0); |
| 34 | assert(kc % sizeof(float) == 0); |
| 35 | assert(ks != 0); |
| 36 | assert(ks % (4 * sizeof(void*)) == 0); |
| 37 | assert(a_offset % sizeof(float) == 0); |
Frank Barchard | 7ccaab6 | 2019-11-19 14:19:23 -0800 | [diff] [blame] | 38 | assert(a != NULL); |
| 39 | assert(w != NULL); |
| 40 | assert(c != NULL); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 41 | |
| 42 | float* c0 = c; |
| 43 | float* c1 = (float*) ((uintptr_t) c0 + cm_stride); |
| 44 | if XNN_UNPREDICTABLE(mr < 2) { |
| 45 | c1 = c0; |
| 46 | } |
| 47 | float* c2 = (float*) ((uintptr_t) c1 + cm_stride); |
| 48 | if XNN_UNPREDICTABLE(mr <= 2) { |
| 49 | c2 = c1; |
| 50 | } |
| 51 | float* c3 = (float*) ((uintptr_t) c2 + cm_stride); |
| 52 | if XNN_UNPREDICTABLE(mr != 4) { |
| 53 | c3 = c2; |
| 54 | } |
| 55 | |
| 56 | do { |
| 57 | float vacc00 = w[0]; |
| 58 | float vacc01 = w[1]; |
| 59 | float vacc02 = w[2]; |
| 60 | float vacc03 = w[3]; |
| 61 | float vacc10 = vacc00; |
| 62 | float vacc11 = vacc01; |
| 63 | float vacc12 = vacc02; |
| 64 | float vacc13 = vacc03; |
| 65 | float vacc20 = vacc00; |
| 66 | float vacc21 = vacc01; |
| 67 | float vacc22 = vacc02; |
| 68 | float vacc23 = vacc03; |
| 69 | float vacc30 = vacc00; |
| 70 | float vacc31 = vacc01; |
| 71 | float vacc32 = vacc02; |
| 72 | float vacc33 = vacc03; |
| 73 | w += 4; |
| 74 | |
| 75 | size_t p = ks; |
| 76 | do { |
| 77 | const float* restrict a0 = a[0]; |
Frank Barchard | 7ccaab6 | 2019-11-19 14:19:23 -0800 | [diff] [blame] | 78 | assert(a0 != NULL); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 79 | if XNN_UNPREDICTABLE(a0 != zero) { |
| 80 | a0 = (const float*) ((uintptr_t) a0 + a_offset); |
| 81 | } |
| 82 | const float* restrict a1 = a[1]; |
Frank Barchard | 7ccaab6 | 2019-11-19 14:19:23 -0800 | [diff] [blame] | 83 | assert(a1 != NULL); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 84 | if XNN_UNPREDICTABLE(a1 != zero) { |
| 85 | a1 = (const float*) ((uintptr_t) a1 + a_offset); |
| 86 | } |
| 87 | const float* restrict a2 = a[2]; |
Frank Barchard | 7ccaab6 | 2019-11-19 14:19:23 -0800 | [diff] [blame] | 88 | assert(a2 != NULL); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 89 | if XNN_UNPREDICTABLE(a2 != zero) { |
| 90 | a2 = (const float*) ((uintptr_t) a2 + a_offset); |
| 91 | } |
| 92 | const float* restrict a3 = a[3]; |
Frank Barchard | 7ccaab6 | 2019-11-19 14:19:23 -0800 | [diff] [blame] | 93 | assert(a3 != NULL); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 94 | if XNN_UNPREDICTABLE(a3 != zero) { |
| 95 | a3 = (const float*) ((uintptr_t) a3 + a_offset); |
| 96 | } |
| 97 | a += 4; |
| 98 | |
| 99 | size_t k = kc; |
| 100 | do { |
| 101 | const float va0 = *a0++; |
| 102 | const float va1 = *a1++; |
| 103 | const float va2 = *a2++; |
| 104 | const float va3 = *a3++; |
| 105 | |
| 106 | const float vb0 = w[0]; |
| 107 | const float vb1 = w[1]; |
| 108 | const float vb2 = w[2]; |
| 109 | const float vb3 = w[3]; |
| 110 | w += 4; |
| 111 | |
| 112 | vacc00 += va0 * vb0; |
| 113 | vacc01 += va0 * vb1; |
| 114 | vacc02 += va0 * vb2; |
| 115 | vacc03 += va0 * vb3; |
| 116 | vacc10 += va1 * vb0; |
| 117 | vacc11 += va1 * vb1; |
| 118 | vacc12 += va1 * vb2; |
| 119 | vacc13 += va1 * vb3; |
| 120 | vacc20 += va2 * vb0; |
| 121 | vacc21 += va2 * vb1; |
| 122 | vacc22 += va2 * vb2; |
| 123 | vacc23 += va2 * vb3; |
| 124 | vacc30 += va3 * vb0; |
| 125 | vacc31 += va3 * vb1; |
| 126 | vacc32 += va3 * vb2; |
| 127 | vacc33 += va3 * vb3; |
| 128 | |
| 129 | k -= sizeof(float); |
| 130 | } while (k != 0); |
| 131 | p -= 4 * sizeof(void*); |
| 132 | } while (p != 0); |
| 133 | |
| 134 | const float vmin = params->scalar.min; |
| 135 | vacc00 = math_max_f32(vacc00, vmin); |
| 136 | vacc01 = math_max_f32(vacc01, vmin); |
| 137 | vacc02 = math_max_f32(vacc02, vmin); |
| 138 | vacc03 = math_max_f32(vacc03, vmin); |
| 139 | vacc10 = math_max_f32(vacc10, vmin); |
| 140 | vacc11 = math_max_f32(vacc11, vmin); |
| 141 | vacc12 = math_max_f32(vacc12, vmin); |
| 142 | vacc13 = math_max_f32(vacc13, vmin); |
| 143 | vacc20 = math_max_f32(vacc20, vmin); |
| 144 | vacc21 = math_max_f32(vacc21, vmin); |
| 145 | vacc22 = math_max_f32(vacc22, vmin); |
| 146 | vacc23 = math_max_f32(vacc23, vmin); |
| 147 | vacc30 = math_max_f32(vacc30, vmin); |
| 148 | vacc31 = math_max_f32(vacc31, vmin); |
| 149 | vacc32 = math_max_f32(vacc32, vmin); |
| 150 | vacc33 = math_max_f32(vacc33, vmin); |
| 151 | |
| 152 | const float vmax = params->scalar.max; |
| 153 | vacc00 = math_min_f32(vacc00, vmax); |
| 154 | vacc01 = math_min_f32(vacc01, vmax); |
| 155 | vacc02 = math_min_f32(vacc02, vmax); |
| 156 | vacc03 = math_min_f32(vacc03, vmax); |
| 157 | vacc10 = math_min_f32(vacc10, vmax); |
| 158 | vacc11 = math_min_f32(vacc11, vmax); |
| 159 | vacc12 = math_min_f32(vacc12, vmax); |
| 160 | vacc13 = math_min_f32(vacc13, vmax); |
| 161 | vacc20 = math_min_f32(vacc20, vmax); |
| 162 | vacc21 = math_min_f32(vacc21, vmax); |
| 163 | vacc22 = math_min_f32(vacc22, vmax); |
| 164 | vacc23 = math_min_f32(vacc23, vmax); |
| 165 | vacc30 = math_min_f32(vacc30, vmax); |
| 166 | vacc31 = math_min_f32(vacc31, vmax); |
| 167 | vacc32 = math_min_f32(vacc32, vmax); |
| 168 | vacc33 = math_min_f32(vacc33, vmax); |
| 169 | |
| 170 | if XNN_LIKELY(nc >= 4) { |
| 171 | c3[0] = vacc30; |
| 172 | c3[1] = vacc31; |
| 173 | c3[2] = vacc32; |
| 174 | c3[3] = vacc33; |
| 175 | c3 = (float*) ((uintptr_t) c3 + cn_stride); |
| 176 | c2[0] = vacc20; |
| 177 | c2[1] = vacc21; |
| 178 | c2[2] = vacc22; |
| 179 | c2[3] = vacc23; |
| 180 | c2 = (float*) ((uintptr_t) c2 + cn_stride); |
| 181 | c1[0] = vacc10; |
| 182 | c1[1] = vacc11; |
| 183 | c1[2] = vacc12; |
| 184 | c1[3] = vacc13; |
| 185 | c1 = (float*) ((uintptr_t) c1 + cn_stride); |
| 186 | c0[0] = vacc00; |
| 187 | c0[1] = vacc01; |
| 188 | c0[2] = vacc02; |
| 189 | c0[3] = vacc03; |
| 190 | c0 = (float*) ((uintptr_t) c0 + cn_stride); |
| 191 | |
| 192 | a = (const float**restrict) ((uintptr_t) a - ks); |
| 193 | nc -= 4; |
| 194 | } else { |
| 195 | if (nc & 2) { |
| 196 | c3[0] = vacc30; |
| 197 | c3[1] = vacc31; |
| 198 | vacc30 = vacc32; |
| 199 | c3 += 2; |
| 200 | c2[0] = vacc20; |
| 201 | c2[1] = vacc21; |
| 202 | vacc20 = vacc22; |
| 203 | c2 += 2; |
| 204 | c1[0] = vacc10; |
| 205 | c1[1] = vacc11; |
| 206 | vacc10 = vacc12; |
| 207 | c1 += 2; |
| 208 | c0[0] = vacc00; |
| 209 | c0[1] = vacc01; |
| 210 | vacc00 = vacc02; |
| 211 | c0 += 2; |
| 212 | } |
| 213 | if (nc & 1) { |
| 214 | c3[0] = vacc30; |
| 215 | c2[0] = vacc20; |
| 216 | c1[0] = vacc10; |
| 217 | c0[0] = vacc00; |
| 218 | } |
| 219 | |
| 220 | nc = 0; |
| 221 | } |
| 222 | } while (nc != 0); |
| 223 | } |