| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/x32-packw/scalar.c.in |
| 3 |
// Generator: tools/xngen |
| 4 |
// |
| 5 |
// Copyright 2023 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 |
|
| 11 |
#include <assert.h> |
| 12 |
#include <stddef.h> |
| 13 |
#include <stdint.h> |
| 14 |
|
| 15 |
#include "xnnpack/math.h" |
| 16 |
#include "xnnpack/packw.h" |
| 17 |
|
| 18 |
|
| 19 |
|
| 20 |
void xnn_x32_packw_gemm_goi_ukernel_x8__scalar_float_u4( |
| 21 |
size_t g, |
| 22 |
size_t nc, |
| 23 |
size_t kc, |
| 24 |
size_t nr, |
| 25 |
size_t kr, |
| 26 |
size_t sr, |
| 27 |
const uint32_t* weights, |
| 28 |
const uint32_t* bias, |
| 29 |
const void* scale, |
| 30 |
uint32_t* packed_weights, |
| 31 |
size_t extra_bytes, |
| 32 |
const void* params) |
| 33 |
{ |
| 34 |
assert(g != 0); |
| 35 |
assert(nc != 0); |
| 36 |
assert(kc != 0); |
| 37 |
assert(nr == 8); |
| 38 |
assert(kr == 1); |
| 39 |
assert(sr == 1); |
| 40 |
assert(weights != NULL); |
| 41 |
assert(packed_weights != NULL); |
| 42 |
|
| 43 |
float* out = (float*) packed_weights; |
| 44 |
const float* b = (const float*) bias; |
| 45 |
|
| 46 |
do { |
| 47 |
// NC main loop multiple of 8 |
| 48 |
const float* w0 = (const float*) weights; |
| 49 |
size_t n = nc; |
| 50 |
for (;n >= 8; n -= 8) { |
| 51 |
if XNN_LIKELY(b != NULL) { |
| 52 |
out[0] = b[0]; |
| 53 |
out[1] = b[1]; |
| 54 |
out[2] = b[2]; |
| 55 |
out[3] = b[3]; |
| 56 |
out[4] = b[4]; |
| 57 |
out[5] = b[5]; |
| 58 |
out[6] = b[6]; |
| 59 |
out[7] = b[7]; |
| 60 |
b += 8; |
| 61 |
} else { |
| 62 |
out[0] = 0; |
| 63 |
out[1] = 0; |
| 64 |
out[2] = 0; |
| 65 |
out[3] = 0; |
| 66 |
out[4] = 0; |
| 67 |
out[5] = 0; |
| 68 |
out[6] = 0; |
| 69 |
out[7] = 0; |
| 70 |
} |
| 71 |
out += 8; |
| 72 |
|
| 73 |
const float* w1 = w0 + kc; |
| 74 |
const float* w2 = w1 + kc; |
| 75 |
const float* w3 = w2 + kc; |
| 76 |
const float* w4 = w3 + kc; |
| 77 |
const float* w5 = w4 + kc; |
| 78 |
const float* w6 = w5 + kc; |
| 79 |
const float* w7 = w6 + kc; |
| 80 |
|
| 81 |
// KC main loop multiple of 8x4 |
| 82 |
size_t k = kc; |
| 83 |
for (; k >= 4; k -= 4) { |
| 84 |
const float v00 = w0[0]; |
| 85 |
const float v01 = w0[1]; |
| 86 |
const float v02 = w0[2]; |
| 87 |
const float v03 = w0[3]; |
| 88 |
w0 += 4; |
| 89 |
const float v10 = w1[0]; |
| 90 |
const float v11 = w1[1]; |
| 91 |
const float v12 = w1[2]; |
| 92 |
const float v13 = w1[3]; |
| 93 |
w1 += 4; |
| 94 |
const float v20 = w2[0]; |
| 95 |
const float v21 = w2[1]; |
| 96 |
const float v22 = w2[2]; |
| 97 |
const float v23 = w2[3]; |
| 98 |
w2 += 4; |
| 99 |
const float v30 = w3[0]; |
| 100 |
const float v31 = w3[1]; |
| 101 |
const float v32 = w3[2]; |
| 102 |
const float v33 = w3[3]; |
| 103 |
w3 += 4; |
| 104 |
const float v40 = w4[0]; |
| 105 |
const float v41 = w4[1]; |
| 106 |
const float v42 = w4[2]; |
| 107 |
const float v43 = w4[3]; |
| 108 |
w4 += 4; |
| 109 |
const float v50 = w5[0]; |
| 110 |
const float v51 = w5[1]; |
| 111 |
const float v52 = w5[2]; |
| 112 |
const float v53 = w5[3]; |
| 113 |
w5 += 4; |
| 114 |
const float v60 = w6[0]; |
| 115 |
const float v61 = w6[1]; |
| 116 |
const float v62 = w6[2]; |
| 117 |
const float v63 = w6[3]; |
| 118 |
w6 += 4; |
| 119 |
const float v70 = w7[0]; |
| 120 |
const float v71 = w7[1]; |
| 121 |
const float v72 = w7[2]; |
| 122 |
const float v73 = w7[3]; |
| 123 |
w7 += 4; |
| 124 |
out[0] = v00; |
| 125 |
out[1] = v10; |
| 126 |
out[2] = v20; |
| 127 |
out[3] = v30; |
| 128 |
out[4] = v40; |
| 129 |
out[5] = v50; |
| 130 |
out[6] = v60; |
| 131 |
out[7] = v70; |
| 132 |
out[8] = v01; |
| 133 |
out[9] = v11; |
| 134 |
out[10] = v21; |
| 135 |
out[11] = v31; |
| 136 |
out[12] = v41; |
| 137 |
out[13] = v51; |
| 138 |
out[14] = v61; |
| 139 |
out[15] = v71; |
| 140 |
out[16] = v02; |
| 141 |
out[17] = v12; |
| 142 |
out[18] = v22; |
| 143 |
out[19] = v32; |
| 144 |
out[20] = v42; |
| 145 |
out[21] = v52; |
| 146 |
out[22] = v62; |
| 147 |
out[23] = v72; |
| 148 |
out[24] = v03; |
| 149 |
out[25] = v13; |
| 150 |
out[26] = v23; |
| 151 |
out[27] = v33; |
| 152 |
out[28] = v43; |
| 153 |
out[29] = v53; |
| 154 |
out[30] = v63; |
| 155 |
out[31] = v73; |
| 156 |
out += 32; |
| 157 |
} |
| 158 |
|
| 159 |
// KC remainder |
| 160 |
for (; k != 0; --k) { |
| 161 |
const float v0 = *w0++; |
| 162 |
out[0] = v0; |
| 163 |
const float v1 = *w1++; |
| 164 |
out[1] = v1; |
| 165 |
const float v2 = *w2++; |
| 166 |
out[2] = v2; |
| 167 |
const float v3 = *w3++; |
| 168 |
out[3] = v3; |
| 169 |
const float v4 = *w4++; |
| 170 |
out[4] = v4; |
| 171 |
const float v5 = *w5++; |
| 172 |
out[5] = v5; |
| 173 |
const float v6 = *w6++; |
| 174 |
out[6] = v6; |
| 175 |
const float v7 = *w7++; |
| 176 |
out[7] = v7; |
| 177 |
out += 8; |
| 178 |
} |
| 179 |
out = (float*) ((uintptr_t) out + extra_bytes); |
| 180 |
w0 = w7; |
| 181 |
} |
| 182 |
|
| 183 |
// NC remainder (1..7) |
| 184 |
if XNN_UNLIKELY(n != 0) { |
| 185 |
if XNN_LIKELY(b != NULL) { |
| 186 |
size_t nb = n; |
| 187 |
do { |
| 188 |
*out++ = *b++; |
| 189 |
} while (--nb != 0); |
| 190 |
} else { |
| 191 |
size_t nb = n; |
| 192 |
do { |
| 193 |
*out++ = 0; |
| 194 |
} while (--nb != 0); |
| 195 |
} |
| 196 |
out += (8 - n); |
| 197 |
|
| 198 |
// NR remainder has less than 8 rows so last row is not loaded |
| 199 |
const float* w1 = w0 + kc; |
| 200 |
if XNN_UNPREDICTABLE(n < 2) { |
| 201 |
w1 = w0; |
| 202 |
} |
| 203 |
const float* w2 = w1 + kc; |
| 204 |
if XNN_UNPREDICTABLE(n <= 2) { |
| 205 |
w2 = w1; |
| 206 |
} |
| 207 |
const float* w3 = w2 + kc; |
| 208 |
if XNN_UNPREDICTABLE(n < 4) { |
| 209 |
w3 = w2; |
| 210 |
} |
| 211 |
const float* w4 = w3 + kc; |
| 212 |
if XNN_UNPREDICTABLE(n <= 4) { |
| 213 |
w4 = w3; |
| 214 |
} |
| 215 |
const float* w5 = w4 + kc; |
| 216 |
if XNN_UNPREDICTABLE(n < 6) { |
| 217 |
w5 = w4; |
| 218 |
} |
| 219 |
const float* w6 = w5 + kc; |
| 220 |
if XNN_UNPREDICTABLE(n <= 6) { |
| 221 |
w6 = w5; |
| 222 |
} |
| 223 |
|
| 224 |
// KC main loop multiple of 8x4 |
| 225 |
size_t k = kc; |
| 226 |
for (; k >= 4; k -= 4) { |
| 227 |
const float v00 = w0[0]; |
| 228 |
const float v01 = w0[1]; |
| 229 |
const float v02 = w0[2]; |
| 230 |
const float v03 = w0[3]; |
| 231 |
w0 += 4; |
| 232 |
const float v10 = w1[0]; |
| 233 |
const float v11 = w1[1]; |
| 234 |
const float v12 = w1[2]; |
| 235 |
const float v13 = w1[3]; |
| 236 |
w1 += 4; |
| 237 |
const float v20 = w2[0]; |
| 238 |
const float v21 = w2[1]; |
| 239 |
const float v22 = w2[2]; |
| 240 |
const float v23 = w2[3]; |
| 241 |
w2 += 4; |
| 242 |
const float v30 = w3[0]; |
| 243 |
const float v31 = w3[1]; |
| 244 |
const float v32 = w3[2]; |
| 245 |
const float v33 = w3[3]; |
| 246 |
w3 += 4; |
| 247 |
const float v40 = w4[0]; |
| 248 |
const float v41 = w4[1]; |
| 249 |
const float v42 = w4[2]; |
| 250 |
const float v43 = w4[3]; |
| 251 |
w4 += 4; |
| 252 |
const float v50 = w5[0]; |
| 253 |
const float v51 = w5[1]; |
| 254 |
const float v52 = w5[2]; |
| 255 |
const float v53 = w5[3]; |
| 256 |
w5 += 4; |
| 257 |
const float v60 = w6[0]; |
| 258 |
const float v61 = w6[1]; |
| 259 |
const float v62 = w6[2]; |
| 260 |
const float v63 = w6[3]; |
| 261 |
w6 += 4; |
| 262 |
out[0] = v00; |
| 263 |
out[1] = v10; |
| 264 |
out[2] = v20; |
| 265 |
out[3] = v30; |
| 266 |
out[4] = v40; |
| 267 |
out[5] = v50; |
| 268 |
out[6] = v60; |
| 269 |
out[8] = v01; |
| 270 |
out[9] = v11; |
| 271 |
out[10] = v21; |
| 272 |
out[11] = v31; |
| 273 |
out[12] = v41; |
| 274 |
out[13] = v51; |
| 275 |
out[14] = v61; |
| 276 |
out[16] = v02; |
| 277 |
out[17] = v12; |
| 278 |
out[18] = v22; |
| 279 |
out[19] = v32; |
| 280 |
out[20] = v42; |
| 281 |
out[21] = v52; |
| 282 |
out[22] = v62; |
| 283 |
out[24] = v03; |
| 284 |
out[25] = v13; |
| 285 |
out[26] = v23; |
| 286 |
out[27] = v33; |
| 287 |
out[28] = v43; |
| 288 |
out[29] = v53; |
| 289 |
out[30] = v63; |
| 290 |
out += 32; |
| 291 |
} |
| 292 |
|
| 293 |
// KC remainder of 1..3 |
| 294 |
for (; k != 0; --k) { |
| 295 |
const float v0 = *w0++; |
| 296 |
out[0] = v0; |
| 297 |
const float v1 = *w1++; |
| 298 |
out[1] = v1; |
| 299 |
const float v2 = *w2++; |
| 300 |
out[2] = v2; |
| 301 |
const float v3 = *w3++; |
| 302 |
out[3] = v3; |
| 303 |
const float v4 = *w4++; |
| 304 |
out[4] = v4; |
| 305 |
const float v5 = *w5++; |
| 306 |
out[5] = v5; |
| 307 |
const float v6 = *w6++; |
| 308 |
out[6] = v6; |
| 309 |
out += 8; |
| 310 |
} |
| 311 |
out = (float*) ((uintptr_t) out + extra_bytes); |
| 312 |
} |
| 313 |
weights += nc * kc; |
| 314 |
} while (--g != 0); |
| 315 |
} |