rotate_neon64.cc 18.2 KB
Newer Older
1 2 3 4 5 6 7 8 9 10
/*
 *  Copyright 2014 The LibYuv Project Authors. All rights reserved.
 *
 *  Use of this source code is governed by a BSD-style license
 *  that can be found in the LICENSE file in the root of the source
 *  tree. An additional intellectual property rights grant can be found
 *  in the file PATENTS. All contributing project authors may
 *  be found in the AUTHORS file in the root of the source tree.
 */

11
#include "libyuv/rotate_row.h"
12
#include "libyuv/row.h"
13 14 15 16 17 18 19 20

#include "libyuv/basic_types.h"

#ifdef __cplusplus
namespace libyuv {
extern "C" {
#endif

21
// This module is for GCC Neon armv8 64 bit.
22
#if !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
23

24 25
static const uvec8 kVTbl4x4Transpose = {0, 4, 8,  12, 1, 5, 9,  13,
                                        2, 6, 10, 14, 3, 7, 11, 15};
Frank Barchard's avatar
Frank Barchard committed
26

Frank Barchard's avatar
Frank Barchard committed
27
void TransposeWx8_NEON(const uint8_t* src,
Frank Barchard's avatar
Frank Barchard committed
28
                       int src_stride,
Frank Barchard's avatar
Frank Barchard committed
29
                       uint8_t* dst,
Frank Barchard's avatar
Frank Barchard committed
30 31
                       int dst_stride,
                       int width) {
Frank Barchard's avatar
Frank Barchard committed
32
  const uint8_t* src_temp;
33 34 35 36 37 38 39 40
  asm volatile(
      // loops are on blocks of 8. loop will stop when
      // counter gets to or below 0. starting the counter
      // at w-8 allow for this
      "sub         %w3, %w3, #8                     \n"

      // handle 8x8 blocks. this should be the majority of the plane
      "1:                                          \n"
41
      "mov         %0, %1                        \n"
42

43 44 45 46 47 48 49 50
      "ld1        {v0.8b}, [%0], %5              \n"
      "ld1        {v1.8b}, [%0], %5              \n"
      "ld1        {v2.8b}, [%0], %5              \n"
      "ld1        {v3.8b}, [%0], %5              \n"
      "ld1        {v4.8b}, [%0], %5              \n"
      "ld1        {v5.8b}, [%0], %5              \n"
      "ld1        {v6.8b}, [%0], %5              \n"
      "ld1        {v7.8b}, [%0]                  \n"
51

52 53 54 55 56 57 58 59
      "trn2     v16.8b, v0.8b, v1.8b             \n"
      "trn1     v17.8b, v0.8b, v1.8b             \n"
      "trn2     v18.8b, v2.8b, v3.8b             \n"
      "trn1     v19.8b, v2.8b, v3.8b             \n"
      "trn2     v20.8b, v4.8b, v5.8b             \n"
      "trn1     v21.8b, v4.8b, v5.8b             \n"
      "trn2     v22.8b, v6.8b, v7.8b             \n"
      "trn1     v23.8b, v6.8b, v7.8b             \n"
60

61 62 63 64 65 66 67 68
      "trn2     v3.4h, v17.4h, v19.4h            \n"
      "trn1     v1.4h, v17.4h, v19.4h            \n"
      "trn2     v2.4h, v16.4h, v18.4h            \n"
      "trn1     v0.4h, v16.4h, v18.4h            \n"
      "trn2     v7.4h, v21.4h, v23.4h            \n"
      "trn1     v5.4h, v21.4h, v23.4h            \n"
      "trn2     v6.4h, v20.4h, v22.4h            \n"
      "trn1     v4.4h, v20.4h, v22.4h            \n"
69

70 71 72 73 74 75 76 77
      "trn2     v21.2s, v1.2s, v5.2s             \n"
      "trn1     v17.2s, v1.2s, v5.2s             \n"
      "trn2     v20.2s, v0.2s, v4.2s             \n"
      "trn1     v16.2s, v0.2s, v4.2s             \n"
      "trn2     v23.2s, v3.2s, v7.2s             \n"
      "trn1     v19.2s, v3.2s, v7.2s             \n"
      "trn2     v22.2s, v2.2s, v6.2s             \n"
      "trn1     v18.2s, v2.2s, v6.2s             \n"
78

79
      "mov         %0, %2                        \n"
80

81 82 83 84 85 86 87 88
      "st1      {v17.8b}, [%0], %6               \n"
      "st1      {v16.8b}, [%0], %6               \n"
      "st1      {v19.8b}, [%0], %6               \n"
      "st1      {v18.8b}, [%0], %6               \n"
      "st1      {v21.8b}, [%0], %6               \n"
      "st1      {v20.8b}, [%0], %6               \n"
      "st1      {v23.8b}, [%0], %6               \n"
      "st1      {v22.8b}, [%0]                   \n"
89

90 91
      "add         %1, %1, #8                    \n"  // src += 8
      "add         %2, %2, %6, lsl #3            \n"  // dst += 8 * dst_stride
92
      "subs        %w3, %w3, #8                  \n"  // w   -= 8
93
      "b.ge        1b                            \n"
94

95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196
      // add 8 back to counter. if the result is 0 there are
      // no residuals.
      "adds        %w3, %w3, #8                    \n"
      "b.eq        4f                              \n"

      // some residual, so between 1 and 7 lines left to transpose
      "cmp         %w3, #2                          \n"
      "b.lt        3f                              \n"

      "cmp         %w3, #4                          \n"
      "b.lt        2f                              \n"

      // 4x8 block
      "mov         %0, %1                          \n"
      "ld1     {v0.s}[0], [%0], %5                 \n"
      "ld1     {v0.s}[1], [%0], %5                 \n"
      "ld1     {v0.s}[2], [%0], %5                 \n"
      "ld1     {v0.s}[3], [%0], %5                 \n"
      "ld1     {v1.s}[0], [%0], %5                 \n"
      "ld1     {v1.s}[1], [%0], %5                 \n"
      "ld1     {v1.s}[2], [%0], %5                 \n"
      "ld1     {v1.s}[3], [%0]                     \n"

      "mov         %0, %2                          \n"

      "ld1      {v2.16b}, [%4]                     \n"

      "tbl      v3.16b, {v0.16b}, v2.16b           \n"
      "tbl      v0.16b, {v1.16b}, v2.16b           \n"

      // TODO(frkoenig): Rework shuffle above to
      // write out with 4 instead of 8 writes.
      "st1 {v3.s}[0], [%0], %6                     \n"
      "st1 {v3.s}[1], [%0], %6                     \n"
      "st1 {v3.s}[2], [%0], %6                     \n"
      "st1 {v3.s}[3], [%0]                         \n"

      "add         %0, %2, #4                      \n"
      "st1 {v0.s}[0], [%0], %6                     \n"
      "st1 {v0.s}[1], [%0], %6                     \n"
      "st1 {v0.s}[2], [%0], %6                     \n"
      "st1 {v0.s}[3], [%0]                         \n"

      "add         %1, %1, #4                      \n"  // src += 4
      "add         %2, %2, %6, lsl #2              \n"  // dst += 4 * dst_stride
      "subs        %w3, %w3, #4                    \n"  // w   -= 4
      "b.eq        4f                              \n"

      // some residual, check to see if it includes a 2x8 block,
      // or less
      "cmp         %w3, #2                         \n"
      "b.lt        3f                              \n"

      // 2x8 block
      "2:                                          \n"
      "mov         %0, %1                          \n"
      "ld1     {v0.h}[0], [%0], %5                 \n"
      "ld1     {v1.h}[0], [%0], %5                 \n"
      "ld1     {v0.h}[1], [%0], %5                 \n"
      "ld1     {v1.h}[1], [%0], %5                 \n"
      "ld1     {v0.h}[2], [%0], %5                 \n"
      "ld1     {v1.h}[2], [%0], %5                 \n"
      "ld1     {v0.h}[3], [%0], %5                 \n"
      "ld1     {v1.h}[3], [%0]                     \n"

      "trn2    v2.8b, v0.8b, v1.8b                 \n"
      "trn1    v3.8b, v0.8b, v1.8b                 \n"

      "mov         %0, %2                          \n"

      "st1     {v3.8b}, [%0], %6                   \n"
      "st1     {v2.8b}, [%0]                       \n"

      "add         %1, %1, #2                      \n"  // src += 2
      "add         %2, %2, %6, lsl #1              \n"  // dst += 2 * dst_stride
      "subs        %w3, %w3,  #2                   \n"  // w   -= 2
      "b.eq        4f                              \n"

      // 1x8 block
      "3:                                          \n"
      "ld1         {v0.b}[0], [%1], %5             \n"
      "ld1         {v0.b}[1], [%1], %5             \n"
      "ld1         {v0.b}[2], [%1], %5             \n"
      "ld1         {v0.b}[3], [%1], %5             \n"
      "ld1         {v0.b}[4], [%1], %5             \n"
      "ld1         {v0.b}[5], [%1], %5             \n"
      "ld1         {v0.b}[6], [%1], %5             \n"
      "ld1         {v0.b}[7], [%1]                 \n"

      "st1         {v0.8b}, [%2]                   \n"

      "4:                                          \n"

      : "=&r"(src_temp),                          // %0
        "+r"(src),                                // %1
        "+r"(dst),                                // %2
        "+r"(width)                               // %3
      : "r"(&kVTbl4x4Transpose),                  // %4
        "r"(static_cast<ptrdiff_t>(src_stride)),  // %5
        "r"(static_cast<ptrdiff_t>(dst_stride))   // %6
      : "memory", "cc", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16",
        "v17", "v18", "v19", "v20", "v21", "v22", "v23");
197 198
}

Frank Barchard's avatar
Frank Barchard committed
199
static const uint8_t kVTbl4x4TransposeDi[32] = {
Frank Barchard's avatar
Frank Barchard committed
200 201
    0, 16, 32, 48, 2, 18, 34, 50, 4, 20, 36, 52, 6, 22, 38, 54,
    1, 17, 33, 49, 3, 19, 35, 51, 5, 21, 37, 53, 7, 23, 39, 55};
202

Frank Barchard's avatar
Frank Barchard committed
203
void TransposeUVWx8_NEON(const uint8_t* src,
Frank Barchard's avatar
Frank Barchard committed
204
                         int src_stride,
Frank Barchard's avatar
Frank Barchard committed
205
                         uint8_t* dst_a,
Frank Barchard's avatar
Frank Barchard committed
206
                         int dst_stride_a,
Frank Barchard's avatar
Frank Barchard committed
207
                         uint8_t* dst_b,
Frank Barchard's avatar
Frank Barchard committed
208
                         int dst_stride_b,
209
                         int width) {
Frank Barchard's avatar
Frank Barchard committed
210
  const uint8_t* src_temp;
211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419
  asm volatile(
      // loops are on blocks of 8. loop will stop when
      // counter gets to or below 0. starting the counter
      // at w-8 allow for this
      "sub       %w4, %w4, #8                    \n"

      // handle 8x8 blocks. this should be the majority of the plane
      "1:                                        \n"
      "mov       %0, %1                          \n"

      "ld1       {v0.16b}, [%0], %5              \n"
      "ld1       {v1.16b}, [%0], %5              \n"
      "ld1       {v2.16b}, [%0], %5              \n"
      "ld1       {v3.16b}, [%0], %5              \n"
      "ld1       {v4.16b}, [%0], %5              \n"
      "ld1       {v5.16b}, [%0], %5              \n"
      "ld1       {v6.16b}, [%0], %5              \n"
      "ld1       {v7.16b}, [%0]                  \n"

      "trn1      v16.16b, v0.16b, v1.16b         \n"
      "trn2      v17.16b, v0.16b, v1.16b         \n"
      "trn1      v18.16b, v2.16b, v3.16b         \n"
      "trn2      v19.16b, v2.16b, v3.16b         \n"
      "trn1      v20.16b, v4.16b, v5.16b         \n"
      "trn2      v21.16b, v4.16b, v5.16b         \n"
      "trn1      v22.16b, v6.16b, v7.16b         \n"
      "trn2      v23.16b, v6.16b, v7.16b         \n"

      "trn1      v0.8h, v16.8h, v18.8h           \n"
      "trn2      v1.8h, v16.8h, v18.8h           \n"
      "trn1      v2.8h, v20.8h, v22.8h           \n"
      "trn2      v3.8h, v20.8h, v22.8h           \n"
      "trn1      v4.8h, v17.8h, v19.8h           \n"
      "trn2      v5.8h, v17.8h, v19.8h           \n"
      "trn1      v6.8h, v21.8h, v23.8h           \n"
      "trn2      v7.8h, v21.8h, v23.8h           \n"

      "trn1      v16.4s, v0.4s, v2.4s            \n"
      "trn2      v17.4s, v0.4s, v2.4s            \n"
      "trn1      v18.4s, v1.4s, v3.4s            \n"
      "trn2      v19.4s, v1.4s, v3.4s            \n"
      "trn1      v20.4s, v4.4s, v6.4s            \n"
      "trn2      v21.4s, v4.4s, v6.4s            \n"
      "trn1      v22.4s, v5.4s, v7.4s            \n"
      "trn2      v23.4s, v5.4s, v7.4s            \n"

      "mov       %0, %2                          \n"

      "st1       {v16.d}[0], [%0], %6            \n"
      "st1       {v18.d}[0], [%0], %6            \n"
      "st1       {v17.d}[0], [%0], %6            \n"
      "st1       {v19.d}[0], [%0], %6            \n"
      "st1       {v16.d}[1], [%0], %6            \n"
      "st1       {v18.d}[1], [%0], %6            \n"
      "st1       {v17.d}[1], [%0], %6            \n"
      "st1       {v19.d}[1], [%0]                \n"

      "mov       %0, %3                          \n"

      "st1       {v20.d}[0], [%0], %7            \n"
      "st1       {v22.d}[0], [%0], %7            \n"
      "st1       {v21.d}[0], [%0], %7            \n"
      "st1       {v23.d}[0], [%0], %7            \n"
      "st1       {v20.d}[1], [%0], %7            \n"
      "st1       {v22.d}[1], [%0], %7            \n"
      "st1       {v21.d}[1], [%0], %7            \n"
      "st1       {v23.d}[1], [%0]                \n"

      "add       %1, %1, #16                     \n"  // src   += 8*2
      "add       %2, %2, %6, lsl #3              \n"  // dst_a += 8 *
                                                      // dst_stride_a
      "add       %3, %3, %7, lsl #3              \n"  // dst_b += 8 *
                                                      // dst_stride_b
      "subs      %w4, %w4,  #8                   \n"  // w     -= 8
      "b.ge      1b                              \n"

      // add 8 back to counter. if the result is 0 there are
      // no residuals.
      "adds      %w4, %w4, #8                    \n"
      "b.eq      4f                              \n"

      // some residual, so between 1 and 7 lines left to transpose
      "cmp       %w4, #2                         \n"
      "b.lt      3f                              \n"

      "cmp       %w4, #4                         \n"
      "b.lt      2f                              \n"

      // TODO(frkoenig): Clean this up
      // 4x8 block
      "mov       %0, %1                          \n"
      "ld1       {v0.8b}, [%0], %5               \n"
      "ld1       {v1.8b}, [%0], %5               \n"
      "ld1       {v2.8b}, [%0], %5               \n"
      "ld1       {v3.8b}, [%0], %5               \n"
      "ld1       {v4.8b}, [%0], %5               \n"
      "ld1       {v5.8b}, [%0], %5               \n"
      "ld1       {v6.8b}, [%0], %5               \n"
      "ld1       {v7.8b}, [%0]                   \n"

      "ld1       {v30.16b}, [%8], #16            \n"
      "ld1       {v31.16b}, [%8]                 \n"

      "tbl       v16.16b, {v0.16b, v1.16b, v2.16b, v3.16b}, v30.16b  \n"
      "tbl       v17.16b, {v0.16b, v1.16b, v2.16b, v3.16b}, v31.16b  \n"
      "tbl       v18.16b, {v4.16b, v5.16b, v6.16b, v7.16b}, v30.16b  \n"
      "tbl       v19.16b, {v4.16b, v5.16b, v6.16b, v7.16b}, v31.16b  \n"

      "mov       %0, %2                          \n"

      "st1       {v16.s}[0],  [%0], %6           \n"
      "st1       {v16.s}[1],  [%0], %6           \n"
      "st1       {v16.s}[2],  [%0], %6           \n"
      "st1       {v16.s}[3],  [%0], %6           \n"

      "add       %0, %2, #4                      \n"
      "st1       {v18.s}[0], [%0], %6            \n"
      "st1       {v18.s}[1], [%0], %6            \n"
      "st1       {v18.s}[2], [%0], %6            \n"
      "st1       {v18.s}[3], [%0]                \n"

      "mov       %0, %3                          \n"

      "st1       {v17.s}[0], [%0], %7            \n"
      "st1       {v17.s}[1], [%0], %7            \n"
      "st1       {v17.s}[2], [%0], %7            \n"
      "st1       {v17.s}[3], [%0], %7            \n"

      "add       %0, %3, #4                      \n"
      "st1       {v19.s}[0],  [%0], %7           \n"
      "st1       {v19.s}[1],  [%0], %7           \n"
      "st1       {v19.s}[2],  [%0], %7           \n"
      "st1       {v19.s}[3],  [%0]               \n"

      "add       %1, %1, #8                      \n"  // src   += 4 * 2
      "add       %2, %2, %6, lsl #2              \n"  // dst_a += 4 *
                                                      // dst_stride_a
      "add       %3, %3, %7, lsl #2              \n"  // dst_b += 4 *
                                                      // dst_stride_b
      "subs      %w4,  %w4,  #4                  \n"  // w     -= 4
      "b.eq      4f                              \n"

      // some residual, check to see if it includes a 2x8 block,
      // or less
      "cmp       %w4, #2                         \n"
      "b.lt      3f                              \n"

      // 2x8 block
      "2:                                        \n"
      "mov       %0, %1                          \n"
      "ld2       {v0.h, v1.h}[0], [%0], %5       \n"
      "ld2       {v2.h, v3.h}[0], [%0], %5       \n"
      "ld2       {v0.h, v1.h}[1], [%0], %5       \n"
      "ld2       {v2.h, v3.h}[1], [%0], %5       \n"
      "ld2       {v0.h, v1.h}[2], [%0], %5       \n"
      "ld2       {v2.h, v3.h}[2], [%0], %5       \n"
      "ld2       {v0.h, v1.h}[3], [%0], %5       \n"
      "ld2       {v2.h, v3.h}[3], [%0]           \n"

      "trn1      v4.8b, v0.8b, v2.8b             \n"
      "trn2      v5.8b, v0.8b, v2.8b             \n"
      "trn1      v6.8b, v1.8b, v3.8b             \n"
      "trn2      v7.8b, v1.8b, v3.8b             \n"

      "mov       %0, %2                          \n"

      "st1       {v4.d}[0], [%0], %6             \n"
      "st1       {v6.d}[0], [%0]                 \n"

      "mov       %0, %3                          \n"

      "st1       {v5.d}[0], [%0], %7             \n"
      "st1       {v7.d}[0], [%0]                 \n"

      "add       %1, %1, #4                      \n"  // src   += 2 * 2
      "add       %2, %2, %6, lsl #1              \n"  // dst_a += 2 *
                                                      // dst_stride_a
      "add       %3, %3, %7, lsl #1              \n"  // dst_b += 2 *
                                                      // dst_stride_b
      "subs      %w4,  %w4,  #2                  \n"  // w     -= 2
      "b.eq      4f                              \n"

      // 1x8 block
      "3:                                        \n"
      "ld2       {v0.b, v1.b}[0], [%1], %5       \n"
      "ld2       {v0.b, v1.b}[1], [%1], %5       \n"
      "ld2       {v0.b, v1.b}[2], [%1], %5       \n"
      "ld2       {v0.b, v1.b}[3], [%1], %5       \n"
      "ld2       {v0.b, v1.b}[4], [%1], %5       \n"
      "ld2       {v0.b, v1.b}[5], [%1], %5       \n"
      "ld2       {v0.b, v1.b}[6], [%1], %5       \n"
      "ld2       {v0.b, v1.b}[7], [%1]           \n"

      "st1       {v0.d}[0], [%2]                 \n"
      "st1       {v1.d}[0], [%3]                 \n"

      "4:                                        \n"

      : "=&r"(src_temp),                            // %0
        "+r"(src),                                  // %1
        "+r"(dst_a),                                // %2
        "+r"(dst_b),                                // %3
        "+r"(width)                                 // %4
      : "r"(static_cast<ptrdiff_t>(src_stride)),    // %5
        "r"(static_cast<ptrdiff_t>(dst_stride_a)),  // %6
        "r"(static_cast<ptrdiff_t>(dst_stride_b)),  // %7
        "r"(&kVTbl4x4TransposeDi)                   // %8
      : "memory", "cc", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16",
        "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v30", "v31");
420
}
421
#endif  // !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
422 423 424 425 426

#ifdef __cplusplus
}  // extern "C"
}  // namespace libyuv
#endif