Skip to content

P
Paddle-Lite

PaddlePaddle / Paddle-Lite

通知 331
Star 4
Fork 1
P
Paddle-Lite
切换分支/标签
查找文件 普通视图历史永久链接Permalink
gemm_kernel.h 13.0 KB
Newer Older
H
backup  
hjchen2 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 
/* Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */

#pragma once
H
Refine gemm to support arm64  
hjchen2 已提交
17 
#if defined(__ARM_NEON__) || defined(__ARM_NEON)
H
backup  
hjchen2 已提交
18 19 20 21 22 23 24 

#include <arm_neon.h>

namespace paddle_mobile {
namespace operators {
namespace math {
H
Refine gemm to support arm64  
hjchen2 已提交
25 26 
#if __aarch64__
void sgemm_6x16(const float *lhs, const float *rhs, const int k, float *output,
H
backup  
hjchen2 已提交
27 
                const int ldc) {
H
Refine gemm to support arm64  
hjchen2 已提交
28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 
  int kc1 = k;
  int step = 4 * ldc;
  int step1 = 4 * 6;
  asm volatile(
      "dup      v6.4s,     wzr     \n\t"
      "dup      v7.4s,     wzr     \n\t"
      "dup      v8.4s,     wzr     \n\t"
      "dup      v9.4s,     wzr     \n\t"
      "dup      v10.4s,    wzr     \n\t"
      "dup      v11.4s,    wzr     \n\t"
      "dup      v12.4s,    wzr     \n\t"
      "dup      v13.4s,    wzr     \n\t"

      "dup      v14.4s,    wzr     \n\t"
      "dup      v15.4s,    wzr     \n\t"
      "dup      v16.4s,    wzr     \n\t"
      "dup      v17.4s,    wzr     \n\t"
      "dup      v18.4s,    wzr     \n\t"
      "dup      v19.4s,    wzr     \n\t"
      "dup      v20.4s,    wzr     \n\t"
      "dup      v21.4s,    wzr     \n\t"

      "dup      v22.4s,    wzr     \n\t"
      "dup      v23.4s,    wzr     \n\t"
      "dup      v24.4s,    wzr     \n\t"
      "dup      v25.4s,    wzr     \n\t"
      "dup      v26.4s,    wzr     \n\t"
      "dup      v27.4s,    wzr     \n\t"
      "dup      v28.4s,    wzr     \n\t"
      "dup      v29.4s,    wzr     \n\t"

      "subs       %[kc1], %[kc1], #1    \n\t"
      "blt        2f                    \n\t"
      "1:                               \n\t"

      "prfm   pldl1keep,  [%[lhs],  #24]  \n\t"
      "prfm   pldl1keep,  [%[rhs],  #64]  \n\t"

      "ld1      {v0.4s, v1.4s},  [%[lhs]],    %[step1]      \n\t"
      "ld1      {v2.4s, v3.4s, v4.4s, v5.4s}, [%[rhs]], #64 \n\t"

      "fmla     v6.4s,    v2.4s,   v0.s[0]       \n\t"
      "fmla     v7.4s,    v3.4s,   v0.s[0]       \n\t"
      "fmla     v8.4s,    v4.4s,   v0.s[0]       \n\t"
      "fmla     v9.4s,    v5.4s,   v0.s[0]       \n\t"

      "fmla     v10.4s,   v2.4s,   v0.s[1]       \n\t"
      "fmla     v11.4s,   v3.4s,   v0.s[1]       \n\t"
      "fmla     v12.4s,   v4.4s,   v0.s[1]       \n\t"
      "fmla     v13.4s,   v5.4s,   v0.s[1]       \n\t"

      "fmla     v14.4s,   v2.4s,   v0.s[2]       \n\t"
      "fmla     v15.4s,   v3.4s,   v0.s[2]       \n\t"
      "fmla     v16.4s,   v4.4s,   v0.s[2]       \n\t"
      "fmla     v17.4s,   v5.4s,   v0.s[2]       \n\t"

      "fmla     v18.4s,   v2.4s,   v0.s[3]       \n\t"
      "fmla     v19.4s,   v3.4s,   v0.s[3]       \n\t"
      "fmla     v20.4s,   v4.4s,   v0.s[3]       \n\t"
      "fmla     v21.4s,   v5.4s,   v0.s[3]       \n\t"

      "fmla     v22.4s,   v2.4s,   v1.s[0]       \n\t"
      "fmla     v23.4s,   v3.4s,   v1.s[0]       \n\t"
      "fmla     v24.4s,   v4.4s,   v1.s[0]       \n\t"
      "fmla     v25.4s,   v5.4s,   v1.s[0]       \n\t"

      "fmla     v26.4s,   v2.4s,   v1.s[1]       \n\t"
      "fmla     v27.4s,   v3.4s,   v1.s[1]       \n\t"
      "fmla     v28.4s,   v4.4s,   v1.s[1]       \n\t"
      "fmla     v29.4s,   v5.4s,   v1.s[1]       \n\t"

      "subs       %[kc1], %[kc1], #1      \n\t"
      "bge        1b                      \n\t"
      "2:                                 \n\t"

      "st1      {v6.4s,  v7.4s,  v8.4s,  v9.4s},    [%[c]],   %[step]   \n\t"
      "st1      {v10.4s, v11.4s, v12.4s, v13.4s},   [%[c]],   %[step]   \n\t"
      "st1      {v14.4s, v15.4s, v16.4s, v17.4s},   [%[c]],   %[step]   \n\t"
      "st1      {v18.4s, v19.4s, v20.4s, v21.4s},   [%[c]],   %[step]   \n\t"
      "st1      {v22.4s, v23.4s, v24.4s, v25.4s},   [%[c]],   %[step]   \n\t"
      "st1      {v26.4s, v27.4s, v28.4s, v29.4s},   [%[c]],   %[step]   \n\t"
      : [lhs] "+r"(lhs), [rhs] "+r"(rhs), [c] "+r"(output), [kc1] "+r"(kc1)
      : [step] "r"(step), [step1] "r"(step1)
      : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9",
        "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19",
        "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29");
H
backup  
hjchen2 已提交
114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 
}
#else
void sgemm_6x8(const float *lhs, const float *rhs, const int k, float *output,
               const int ldc) {
  int kc1 = k >> 3;   // k / 8
  int kc2 = k & 0x7;  // k % 8
  int step = sizeof(float) * ldc;
  asm volatile(
      "pld        [%[lhs]]            \n\t"
      "pld        [%[lhs],  #64]      \n\t"
      "pld        [%[rhs]]            \n\t"
      "pld        [%[rhs],  #64]      \n\t"

      "vmov.f32   q4,     #0.0          \n\t"
      "vmov.f32   q5,     #0.0          \n\t"
      "vmov.f32   q6,     #0.0          \n\t"
      "vmov.f32   q7,     #0.0          \n\t"
      "vmov.f32   q8,     #0.0          \n\t"
      "vmov.f32   q9,     #0.0          \n\t"
      "vmov.f32   q10,    #0.0          \n\t"
      "vmov.f32   q11,    #0.0          \n\t"
      "vmov.f32   q12,    #0.0          \n\t"
      "vmov.f32   q13,    #0.0          \n\t"
      "vmov.f32   q14,    #0.0          \n\t"
      "vmov.f32   q15,    #0.0          \n\t"

      "subs       %[kc1], %[kc1], #1    \n\t"
      "blt        2f                    \n\t"
      "1:                               \n\t"

      "pld        [%[lhs], #128]       \n\t"
      "pld        [%[rhs], #128]       \n\t"

      "vld1.32    {d0-d2},  [%[lhs]]!   \n\t"
      "vld1.32    {q2, q3}, [%[rhs]]!   \n\t"

      "vmla.f32   q4,   q2,   d0[0]       \n\t"
      "vmla.f32   q5,   q3,   d0[0]       \n\t"
      "vmla.f32   q6,   q2,   d0[1]       \n\t"
      "vmla.f32   q7,   q3,   d0[1]       \n\t"
      "vmla.f32   q8,   q2,   d1[0]       \n\t"
      "vmla.f32   q9,   q3,   d1[0]       \n\t"
      "vmla.f32   q10,  q2,   d1[1]       \n\t"
      "vmla.f32   q11,  q3,   d1[1]       \n\t"
      "vmla.f32   q12,  q2,   d2[0]       \n\t"
      "vmla.f32   q13,  q3,   d2[0]       \n\t"
      "vmla.f32   q14,  q2,   d2[1]       \n\t"
      "vmla.f32   q15,  q3,   d2[1]       \n\t"

      "vld1.32    {d0-d2},  [%[lhs]]!   \n\t"
      "vld1.32    {q2, q3}, [%[rhs]]!   \n\t"

      "vmla.f32   q4,   q2,   d0[0]       \n\t"
      "vmla.f32   q5,   q3,   d0[0]       \n\t"
      "vmla.f32   q6,   q2,   d0[1]       \n\t"
      "vmla.f32   q7,   q3,   d0[1]       \n\t"
      "vmla.f32   q8,   q2,   d1[0]       \n\t"
      "vmla.f32   q9,   q3,   d1[0]       \n\t"
      "vmla.f32   q10,  q2,   d1[1]       \n\t"
      "vmla.f32   q11,  q3,   d1[1]       \n\t"
      "vmla.f32   q12,  q2,   d2[0]       \n\t"
      "vmla.f32   q13,  q3,   d2[0]       \n\t"
      "vmla.f32   q14,  q2,   d2[1]       \n\t"
      "vmla.f32   q15,  q3,   d2[1]       \n\t"

      "pld        [%[lhs], #128]       \n\t"
      "pld        [%[rhs], #128]       \n\t"

      "vld1.32    {d0-d2},  [%[lhs]]!   \n\t"
      "vld1.32    {q2, q3}, [%[rhs]]!   \n\t"

      "vmla.f32   q4,   q2,   d0[0]       \n\t"
      "vmla.f32   q5,   q3,   d0[0]       \n\t"
      "vmla.f32   q6,   q2,   d0[1]       \n\t"
      "vmla.f32   q7,   q3,   d0[1]       \n\t"
      "vmla.f32   q8,   q2,   d1[0]       \n\t"
      "vmla.f32   q9,   q3,   d1[0]       \n\t"
      "vmla.f32   q10,  q2,   d1[1]       \n\t"
      "vmla.f32   q11,  q3,   d1[1]       \n\t"
      "vmla.f32   q12,  q2,   d2[0]       \n\t"
      "vmla.f32   q13,  q3,   d2[0]       \n\t"
      "vmla.f32   q14,  q2,   d2[1]       \n\t"
      "vmla.f32   q15,  q3,   d2[1]       \n\t"

      "vld1.32    {d0-d2},  [%[lhs]]!   \n\t"
      "vld1.32    {q2, q3}, [%[rhs]]!   \n\t"

      "vmla.f32   q4,   q2,   d0[0]       \n\t"
      "vmla.f32   q5,   q3,   d0[0]       \n\t"
      "vmla.f32   q6,   q2,   d0[1]       \n\t"
      "vmla.f32   q7,   q3,   d0[1]       \n\t"
      "vmla.f32   q8,   q2,   d1[0]       \n\t"
      "vmla.f32   q9,   q3,   d1[0]       \n\t"
      "vmla.f32   q10,  q2,   d1[1]       \n\t"
      "vmla.f32   q11,  q3,   d1[1]       \n\t"
      "vmla.f32   q12,  q2,   d2[0]       \n\t"
      "vmla.f32   q13,  q3,   d2[0]       \n\t"
      "vmla.f32   q14,  q2,   d2[1]       \n\t"
      "vmla.f32   q15,  q3,   d2[1]       \n\t"

      "pld        [%[lhs], #128]       \n\t"
      "pld        [%[rhs], #128]       \n\t"

      "vld1.32    {d0-d2},  [%[lhs]]!   \n\t"
      "vld1.32    {q2, q3}, [%[rhs]]!   \n\t"

      "vmla.f32   q4,   q2,   d0[0]       \n\t"
      "vmla.f32   q5,   q3,   d0[0]       \n\t"
      "vmla.f32   q6,   q2,   d0[1]       \n\t"
      "vmla.f32   q7,   q3,   d0[1]       \n\t"
      "vmla.f32   q8,   q2,   d1[0]       \n\t"
      "vmla.f32   q9,   q3,   d1[0]       \n\t"
      "vmla.f32   q10,  q2,   d1[1]       \n\t"
      "vmla.f32   q11,  q3,   d1[1]       \n\t"
      "vmla.f32   q12,  q2,   d2[0]       \n\t"
      "vmla.f32   q13,  q3,   d2[0]       \n\t"
      "vmla.f32   q14,  q2,   d2[1]       \n\t"
      "vmla.f32   q15,  q3,   d2[1]       \n\t"

      "vld1.32    {d0-d2},  [%[lhs]]!   \n\t"
      "vld1.32    {q2, q3}, [%[rhs]]!   \n\t"

      "vmla.f32   q4,   q2,   d0[0]       \n\t"
      "vmla.f32   q5,   q3,   d0[0]       \n\t"
      "vmla.f32   q6,   q2,   d0[1]       \n\t"
      "vmla.f32   q7,   q3,   d0[1]       \n\t"
      "vmla.f32   q8,   q2,   d1[0]       \n\t"
      "vmla.f32   q9,   q3,   d1[0]       \n\t"
      "vmla.f32   q10,  q2,   d1[1]       \n\t"
      "vmla.f32   q11,  q3,   d1[1]       \n\t"
      "vmla.f32   q12,  q2,   d2[0]       \n\t"
      "vmla.f32   q13,  q3,   d2[0]       \n\t"
      "vmla.f32   q14,  q2,   d2[1]       \n\t"
      "vmla.f32   q15,  q3,   d2[1]       \n\t"

      "pld        [%[lhs], #128]       \n\t"
      "pld        [%[rhs], #128]       \n\t"

      "vld1.32    {d0-d2},  [%[lhs]]!   \n\t"
      "vld1.32    {q2, q3}, [%[rhs]]!   \n\t"

      "vmla.f32   q4,   q2,   d0[0]       \n\t"
      "vmla.f32   q5,   q3,   d0[0]       \n\t"
      "vmla.f32   q6,   q2,   d0[1]       \n\t"
      "vmla.f32   q7,   q3,   d0[1]       \n\t"
      "vmla.f32   q8,   q2,   d1[0]       \n\t"
      "vmla.f32   q9,   q3,   d1[0]       \n\t"
      "vmla.f32   q10,  q2,   d1[1]       \n\t"
      "vmla.f32   q11,  q3,   d1[1]       \n\t"
      "vmla.f32   q12,  q2,   d2[0]       \n\t"
      "vmla.f32   q13,  q3,   d2[0]       \n\t"
      "vmla.f32   q14,  q2,   d2[1]       \n\t"
      "vmla.f32   q15,  q3,   d2[1]       \n\t"

      "vld1.32    {d0-d2},  [%[lhs]]!   \n\t"
      "vld1.32    {q2, q3}, [%[rhs]]!   \n\t"

      "vmla.f32   q4,   q2,   d0[0]       \n\t"
      "vmla.f32   q5,   q3,   d0[0]       \n\t"
      "vmla.f32   q6,   q2,   d0[1]       \n\t"
      "vmla.f32   q7,   q3,   d0[1]       \n\t"
      "vmla.f32   q8,   q2,   d1[0]       \n\t"
      "vmla.f32   q9,   q3,   d1[0]       \n\t"
      "vmla.f32   q10,  q2,   d1[1]       \n\t"
      "vmla.f32   q11,  q3,   d1[1]       \n\t"
      "vmla.f32   q12,  q2,   d2[0]       \n\t"
      "vmla.f32   q13,  q3,   d2[0]       \n\t"
      "vmla.f32   q14,  q2,   d2[1]       \n\t"
      "vmla.f32   q15,  q3,   d2[1]       \n\t"

      "subs       %[kc1], %[kc1], #1      \n\t"
      "bge        1b                      \n\t"
      "2:                                 \n\t"

      "subs       %[kc2], %[kc2], #1      \n\t"
      "blt        4f                      \n\t"
      "3:                                 \n\t"

      "vld1.32    {d0-d2},  [%[lhs]]!   \n\t"
      "vld1.32    {q2, q3}, [%[rhs]]!   \n\t"

      "vmla.f32   q4,   q2,   d0[0]       \n\t"
      "vmla.f32   q5,   q3,   d0[0]       \n\t"
      "vmla.f32   q6,   q2,   d0[1]       \n\t"
      "vmla.f32   q7,   q3,   d0[1]       \n\t"
      "vmla.f32   q8,   q2,   d1[0]       \n\t"
      "vmla.f32   q9,   q3,   d1[0]       \n\t"
      "vmla.f32   q10,  q2,   d1[1]       \n\t"
      "vmla.f32   q11,  q3,   d1[1]       \n\t"
      "vmla.f32   q12,  q2,   d2[0]       \n\t"
      "vmla.f32   q13,  q3,   d2[0]       \n\t"
      "vmla.f32   q14,  q2,   d2[1]       \n\t"
      "vmla.f32   q15,  q3,   d2[1]       \n\t"

      "subs       %[kc2], %[kc2], #1      \n\t"
      "bge        3b                      \n\t"
      "4:                                 \n\t"

      "mov        r5,     %[c]            \n\t"
      "mov        r6,     %[step]         \n\t"
      "vst1.32    {q4, q5},   [r5], r6    \n\t"
      "vst1.32    {q6, q7},   [r5], r6    \n\t"
      "vst1.32    {q8, q9},   [r5], r6    \n\t"
      "vst1.32    {q10, q11}, [r5], r6    \n\t"
      "vst1.32    {q12, q13}, [r5], r6    \n\t"
      "vst1.32    {q14, q15}, [r5]        \n\t"
      :
      : [lhs] "r"(lhs), [rhs] "r"(rhs), [c] "r"(output), [kc1] "r"(kc1),
        [kc2] "r"(kc2), [step] "r"(step)
      : "cc", "memory", "r5", "r6", "q0", "q1", "q2", "q3", "q4", "q5", "q6",
        "q7", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
}
#endif  // __aarch64__

}  // namespace math
}  // namespace operators
}  // namespace paddle_mobile

#endif  // __ARM_NEON__