scale_neon.cc 21 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31
/*
 *  Copyright 2011 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.
 */

#include "libyuv/basic_types.h"
#include "libyuv/row.h"

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

// This module is for GCC Neon
#if !defined(YUV_DISABLE_ASM) && defined(__ARM_NEON__)

/**
 * NEON downscalers with interpolation.
 *
 * Provided by Fritz Koenig
 *
 */

void ScaleRowDown2_NEON(const uint8* src_ptr, ptrdiff_t /* src_stride */,
                        uint8* dst, int dst_width) {
  asm volatile (
32
  "1:                                          \n"
33 34 35
    // load even pixels into q0, odd into q1
    "vld2.u8    {q0,q1}, [%0]!                 \n"
    "subs       %2, %2, #16                    \n"  // 16 processed per loop
fbarchard@google.com's avatar
fbarchard@google.com committed
36
    "vst1.u8    {q0}, [%1]!                    \n"  // store even pixels
37
    "bgt        1b                             \n"
fbarchard@google.com's avatar
fbarchard@google.com committed
38 39 40 41 42
  : "+r"(src_ptr),          // %0
    "+r"(dst),              // %1
    "+r"(dst_width)         // %2
  :
  : "q0", "q1"              // Clobber List
43 44 45 46 47 48 49 50
  );
}

void ScaleRowDown2Int_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
                           uint8* dst, int dst_width) {
  asm volatile (
    // change the stride to row 2 pointer
    "add        %1, %0                         \n"
51
  "1:                                          \n"
52 53
    "vld1.u8    {q0,q1}, [%0]!                 \n"  // load row 1 and post inc
    "vld1.u8    {q2,q3}, [%1]!                 \n"  // load row 2 and post inc
fbarchard@google.com's avatar
fbarchard@google.com committed
54
    "subs       %3, %3, #16                    \n"  // 16 processed per loop
55 56 57 58 59 60 61 62
    "vpaddl.u8  q0, q0                         \n"  // row 1 add adjacent
    "vpaddl.u8  q1, q1                         \n"
    "vpadal.u8  q0, q2                         \n"  // row 2 add adjacent + row1
    "vpadal.u8  q1, q3                         \n"
    "vrshrn.u16 d0, q0, #2                     \n"  // downshift, round and pack
    "vrshrn.u16 d1, q1, #2                     \n"
    "vst1.u8    {q0}, [%2]!                    \n"
    "bgt        1b                             \n"
fbarchard@google.com's avatar
fbarchard@google.com committed
63 64 65 66 67 68 69
  : "+r"(src_ptr),          // %0
    "+r"(src_stride),       // %1
    "+r"(dst),              // %2
    "+r"(dst_width)         // %3
  :
  : "q0", "q1", "q2", "q3"     // Clobber List
  );
70 71 72 73 74
}

void ScaleRowDown4_NEON(const uint8* src_ptr, ptrdiff_t /* src_stride */,
                        uint8* dst_ptr, int dst_width) {
  asm volatile (
75
  "1:                                          \n"
76
    "vld2.u8    {d0, d1}, [%0]!                \n"
fbarchard@google.com's avatar
fbarchard@google.com committed
77
    "subs       %2, #4                         \n"
78 79 80 81
    "vtrn.u8    d1, d0                         \n"
    "vshrn.u16  d0, q0, #8                     \n"
    "vst1.u32   {d0[1]}, [%1]!                 \n"
    "bgt        1b                             \n"
fbarchard@google.com's avatar
fbarchard@google.com committed
82 83 84 85 86
  : "+r"(src_ptr),          // %0
    "+r"(dst_ptr),          // %1
    "+r"(dst_width)         // %2
  :
  : "q0", "q1", "memory", "cc"
87 88 89 90 91 92 93 94 95
  );
}

void ScaleRowDown4Int_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
                           uint8* dst_ptr, int dst_width) {
  asm volatile (
    "add        r4, %0, %3                     \n"
    "add        r5, r4, %3                     \n"
    "add        %3, r5, %3                     \n"
96
  "1:                                          \n"
97 98 99 100
    "vld1.u8    {q0}, [%0]!                    \n"   // load up 16x4
    "vld1.u8    {q1}, [r4]!                    \n"
    "vld1.u8    {q2}, [r5]!                    \n"
    "vld1.u8    {q3}, [%3]!                    \n"
fbarchard@google.com's avatar
fbarchard@google.com committed
101
    "subs       %2, #4                         \n"
102 103 104 105 106 107 108 109 110
    "vpaddl.u8  q0, q0                         \n"
    "vpadal.u8  q0, q1                         \n"
    "vpadal.u8  q0, q2                         \n"
    "vpadal.u8  q0, q3                         \n"
    "vpaddl.u16 q0, q0                         \n"
    "vrshrn.u32 d0, q0, #4                     \n"   // divide by 16 w/rounding
    "vmovn.u16  d0, q0                         \n"
    "vst1.u32   {d0[0]}, [%1]!                 \n"
    "bgt        1b                             \n"
fbarchard@google.com's avatar
fbarchard@google.com committed
111 112 113 114 115
  : "+r"(src_ptr),          // %0
    "+r"(dst_ptr),          // %1
    "+r"(dst_width)         // %2
  : "r"(src_stride)         // %3
  : "r4", "r5", "q0", "q1", "q2", "q3", "memory", "cc"
116 117 118 119 120 121 122 123 124 125
  );
}

// Down scale from 4 to 3 pixels. Use the neon multilane read/write
// to load up the every 4th pixel into a 4 different registers.
// Point samples 32 pixels to 24 pixels.
void ScaleRowDown34_NEON(const uint8* src_ptr,
                         ptrdiff_t /* src_stride */,
                         uint8* dst_ptr, int dst_width) {
  asm volatile (
126
  "1:                                          \n"
127
    "vld4.u8      {d0, d1, d2, d3}, [%0]!      \n" // src line 0
fbarchard@google.com's avatar
fbarchard@google.com committed
128
    "subs         %2, #24                      \n"
129 130 131
    "vmov         d2, d3                       \n" // order d0, d1, d2
    "vst3.u8      {d0, d1, d2}, [%1]!          \n"
    "bgt          1b                           \n"
fbarchard@google.com's avatar
fbarchard@google.com committed
132 133 134 135 136
  : "+r"(src_ptr),          // %0
    "+r"(dst_ptr),          // %1
    "+r"(dst_width)         // %2
  :
  : "d0", "d1", "d2", "d3", "memory", "cc"
137 138 139 140 141 142 143 144 145
  );
}

void ScaleRowDown34_0_Int_NEON(const uint8* src_ptr,
                               ptrdiff_t src_stride,
                               uint8* dst_ptr, int dst_width) {
  asm volatile (
    "vmov.u8      d24, #3                      \n"
    "add          %3, %0                       \n"
146
  "1:                                          \n"
147 148
    "vld4.u8      {d0, d1, d2, d3}, [%0]!      \n" // src line 0
    "vld4.u8      {d4, d5, d6, d7}, [%3]!      \n" // src line 1
fbarchard@google.com's avatar
fbarchard@google.com committed
149
    "subs         %2, #24                      \n"
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

    // filter src line 0 with src line 1
    // expand chars to shorts to allow for room
    // when adding lines together
    "vmovl.u8     q8, d4                       \n"
    "vmovl.u8     q9, d5                       \n"
    "vmovl.u8     q10, d6                      \n"
    "vmovl.u8     q11, d7                      \n"

    // 3 * line_0 + line_1
    "vmlal.u8     q8, d0, d24                  \n"
    "vmlal.u8     q9, d1, d24                  \n"
    "vmlal.u8     q10, d2, d24                 \n"
    "vmlal.u8     q11, d3, d24                 \n"

    // (3 * line_0 + line_1) >> 2
    "vqrshrn.u16  d0, q8, #2                   \n"
    "vqrshrn.u16  d1, q9, #2                   \n"
    "vqrshrn.u16  d2, q10, #2                  \n"
    "vqrshrn.u16  d3, q11, #2                  \n"

    // a0 = (src[0] * 3 + s[1] * 1) >> 2
    "vmovl.u8     q8, d1                       \n"
    "vmlal.u8     q8, d0, d24                  \n"
    "vqrshrn.u16  d0, q8, #2                   \n"

    // a1 = (src[1] * 1 + s[2] * 1) >> 1
    "vrhadd.u8    d1, d1, d2                   \n"

    // a2 = (src[2] * 1 + s[3] * 3) >> 2
    "vmovl.u8     q8, d2                       \n"
    "vmlal.u8     q8, d3, d24                  \n"
    "vqrshrn.u16  d2, q8, #2                   \n"

    "vst3.u8      {d0, d1, d2}, [%1]!          \n"

    "bgt          1b                           \n"
fbarchard@google.com's avatar
fbarchard@google.com committed
187 188 189 190 191 192
  : "+r"(src_ptr),          // %0
    "+r"(dst_ptr),          // %1
    "+r"(dst_width),        // %2
    "+r"(src_stride)        // %3
  :
  : "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11", "d24", "memory", "cc"
193 194 195 196 197 198 199 200 201
  );
}

void ScaleRowDown34_1_Int_NEON(const uint8* src_ptr,
                               ptrdiff_t src_stride,
                               uint8* dst_ptr, int dst_width) {
  asm volatile (
    "vmov.u8      d24, #3                      \n"
    "add          %3, %0                       \n"
202
  "1:                                          \n"
203 204
    "vld4.u8      {d0, d1, d2, d3}, [%0]!      \n" // src line 0
    "vld4.u8      {d4, d5, d6, d7}, [%3]!      \n" // src line 1
fbarchard@google.com's avatar
fbarchard@google.com committed
205
    "subs         %2, #24                      \n"
206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224
    // average src line 0 with src line 1
    "vrhadd.u8    q0, q0, q2                   \n"
    "vrhadd.u8    q1, q1, q3                   \n"

    // a0 = (src[0] * 3 + s[1] * 1) >> 2
    "vmovl.u8     q3, d1                       \n"
    "vmlal.u8     q3, d0, d24                  \n"
    "vqrshrn.u16  d0, q3, #2                   \n"

    // a1 = (src[1] * 1 + s[2] * 1) >> 1
    "vrhadd.u8    d1, d1, d2                   \n"

    // a2 = (src[2] * 1 + s[3] * 3) >> 2
    "vmovl.u8     q3, d2                       \n"
    "vmlal.u8     q3, d3, d24                  \n"
    "vqrshrn.u16  d2, q3, #2                   \n"

    "vst3.u8      {d0, d1, d2}, [%1]!          \n"
    "bgt          1b                           \n"
fbarchard@google.com's avatar
fbarchard@google.com committed
225 226 227 228 229 230
  : "+r"(src_ptr),          // %0
    "+r"(dst_ptr),          // %1
    "+r"(dst_width),        // %2
    "+r"(src_stride)        // %3
  :
  : "r4", "q0", "q1", "q2", "q3", "d24", "memory", "cc"
231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251
  );
}

#define HAS_SCALEROWDOWN38_NEON
const uvec8 kShuf38 =
  { 0, 3, 6, 8, 11, 14, 16, 19, 22, 24, 27, 30, 0, 0, 0, 0 };
const uvec8 kShuf38_2 =
  { 0, 8, 16, 2, 10, 17, 4, 12, 18, 6, 14, 19, 0, 0, 0, 0 };
const vec16 kMult38_Div6 =
  { 65536 / 12, 65536 / 12, 65536 / 12, 65536 / 12,
    65536 / 12, 65536 / 12, 65536 / 12, 65536 / 12 };
const vec16 kMult38_Div9 =
  { 65536 / 18, 65536 / 18, 65536 / 18, 65536 / 18,
    65536 / 18, 65536 / 18, 65536 / 18, 65536 / 18 };

// 32 -> 12
void ScaleRowDown38_NEON(const uint8* src_ptr,
                         ptrdiff_t /* src_stride */,
                         uint8* dst_ptr, int dst_width) {
  asm volatile (
    "vld1.u8      {q3}, [%3]                   \n"
252
  "1:                                          \n"
253
    "vld1.u8      {d0, d1, d2, d3}, [%0]!      \n"
fbarchard@google.com's avatar
fbarchard@google.com committed
254
    "subs         %2, #12                      \n"
255 256 257 258 259
    "vtbl.u8      d4, {d0, d1, d2, d3}, d6     \n"
    "vtbl.u8      d5, {d0, d1, d2, d3}, d7     \n"
    "vst1.u8      {d4}, [%1]!                  \n"
    "vst1.u32     {d5[0]}, [%1]!               \n"
    "bgt          1b                           \n"
fbarchard@google.com's avatar
fbarchard@google.com committed
260 261 262 263 264
  : "+r"(src_ptr),          // %0
    "+r"(dst_ptr),          // %1
    "+r"(dst_width)         // %2
  : "r"(&kShuf38)           // %3
  : "d0", "d1", "d2", "d3", "d4", "d5", "memory", "cc"
265 266 267 268 269 270 271 272 273 274 275 276 277
  );
}

// 32x3 -> 12x1
void OMITFP ScaleRowDown38_3_Int_NEON(const uint8* src_ptr,
                                      ptrdiff_t src_stride,
                                      uint8* dst_ptr, int dst_width) {
  asm volatile (
    "vld1.u16     {q13}, [%4]                  \n"
    "vld1.u8      {q14}, [%5]                  \n"
    "vld1.u8      {q15}, [%6]                  \n"
    "add          r4, %0, %3, lsl #1           \n"
    "add          %3, %0                       \n"
278
  "1:                                          \n"
279 280 281 282 283 284 285 286

    // d0 = 00 40 01 41 02 42 03 43
    // d1 = 10 50 11 51 12 52 13 53
    // d2 = 20 60 21 61 22 62 23 63
    // d3 = 30 70 31 71 32 72 33 73
    "vld4.u8      {d0, d1, d2, d3}, [%0]!      \n"
    "vld4.u8      {d4, d5, d6, d7}, [%3]!      \n"
    "vld4.u8      {d16, d17, d18, d19}, [r4]!  \n"
fbarchard@google.com's avatar
fbarchard@google.com committed
287
    "subs         %2, #12                      \n"
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

    // Shuffle the input data around to get align the data
    //  so adjacent data can be added. 0,1 - 2,3 - 4,5 - 6,7
    // d0 = 00 10 01 11 02 12 03 13
    // d1 = 40 50 41 51 42 52 43 53
    "vtrn.u8      d0, d1                       \n"
    "vtrn.u8      d4, d5                       \n"
    "vtrn.u8      d16, d17                     \n"

    // d2 = 20 30 21 31 22 32 23 33
    // d3 = 60 70 61 71 62 72 63 73
    "vtrn.u8      d2, d3                       \n"
    "vtrn.u8      d6, d7                       \n"
    "vtrn.u8      d18, d19                     \n"

    // d0 = 00+10 01+11 02+12 03+13
    // d2 = 40+50 41+51 42+52 43+53
    "vpaddl.u8    q0, q0                       \n"
    "vpaddl.u8    q2, q2                       \n"
    "vpaddl.u8    q8, q8                       \n"

    // d3 = 60+70 61+71 62+72 63+73
    "vpaddl.u8    d3, d3                       \n"
    "vpaddl.u8    d7, d7                       \n"
    "vpaddl.u8    d19, d19                     \n"

    // combine source lines
    "vadd.u16     q0, q2                       \n"
    "vadd.u16     q0, q8                       \n"
    "vadd.u16     d4, d3, d7                   \n"
    "vadd.u16     d4, d19                      \n"

    // dst_ptr[3] = (s[6 + st * 0] + s[7 + st * 0]
    //             + s[6 + st * 1] + s[7 + st * 1]
    //             + s[6 + st * 2] + s[7 + st * 2]) / 6
    "vqrdmulh.s16 q2, q2, q13                  \n"
    "vmovn.u16    d4, q2                       \n"

    // Shuffle 2,3 reg around so that 2 can be added to the
    //  0,1 reg and 3 can be added to the 4,5 reg. This
    //  requires expanding from u8 to u16 as the 0,1 and 4,5
    //  registers are already expanded. Then do transposes
    //  to get aligned.
    // q2 = xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33
    "vmovl.u8     q1, d2                       \n"
    "vmovl.u8     q3, d6                       \n"
    "vmovl.u8     q9, d18                      \n"

    // combine source lines
    "vadd.u16     q1, q3                       \n"
    "vadd.u16     q1, q9                       \n"

    // d4 = xx 20 xx 30 xx 22 xx 32
    // d5 = xx 21 xx 31 xx 23 xx 33
    "vtrn.u32     d2, d3                       \n"

    // d4 = xx 20 xx 21 xx 22 xx 23
    // d5 = xx 30 xx 31 xx 32 xx 33
    "vtrn.u16     d2, d3                       \n"

    // 0+1+2, 3+4+5
    "vadd.u16     q0, q1                       \n"

    // Need to divide, but can't downshift as the the value
    //  isn't a power of 2. So multiply by 65536 / n
    //  and take the upper 16 bits.
    "vqrdmulh.s16 q0, q0, q15                  \n"

    // Align for table lookup, vtbl requires registers to
    //  be adjacent
    "vmov.u8      d2, d4                       \n"

    "vtbl.u8      d3, {d0, d1, d2}, d28        \n"
    "vtbl.u8      d4, {d0, d1, d2}, d29        \n"

    "vst1.u8      {d3}, [%1]!                  \n"
    "vst1.u32     {d4[0]}, [%1]!               \n"
    "bgt          1b                           \n"
fbarchard@google.com's avatar
fbarchard@google.com committed
366 367 368 369 370 371 372 373 374
  : "+r"(src_ptr),          // %0
    "+r"(dst_ptr),          // %1
    "+r"(dst_width),        // %2
    "+r"(src_stride)        // %3
  : "r"(&kMult38_Div6),     // %4
    "r"(&kShuf38_2),        // %5
    "r"(&kMult38_Div9)      // %6
  : "r4", "q0", "q1", "q2", "q3", "q8", "q9",
    "q13", "q14", "q15", "memory", "cc"
375 376 377 378 379 380 381 382 383 384 385
  );
}

// 32x2 -> 12x1
void ScaleRowDown38_2_Int_NEON(const uint8* src_ptr,
                               ptrdiff_t src_stride,
                               uint8* dst_ptr, int dst_width) {
  asm volatile (
    "vld1.u16     {q13}, [%4]                  \n"
    "vld1.u8      {q14}, [%5]                  \n"
    "add          %3, %0                       \n"
386
  "1:                                          \n"
387 388 389 390 391 392 393

    // d0 = 00 40 01 41 02 42 03 43
    // d1 = 10 50 11 51 12 52 13 53
    // d2 = 20 60 21 61 22 62 23 63
    // d3 = 30 70 31 71 32 72 33 73
    "vld4.u8      {d0, d1, d2, d3}, [%0]!      \n"
    "vld4.u8      {d4, d5, d6, d7}, [%3]!      \n"
fbarchard@google.com's avatar
fbarchard@google.com committed
394
    "subs         %2, #12                      \n"
395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461

    // Shuffle the input data around to get align the data
    //  so adjacent data can be added. 0,1 - 2,3 - 4,5 - 6,7
    // d0 = 00 10 01 11 02 12 03 13
    // d1 = 40 50 41 51 42 52 43 53
    "vtrn.u8      d0, d1                       \n"
    "vtrn.u8      d4, d5                       \n"

    // d2 = 20 30 21 31 22 32 23 33
    // d3 = 60 70 61 71 62 72 63 73
    "vtrn.u8      d2, d3                       \n"
    "vtrn.u8      d6, d7                       \n"

    // d0 = 00+10 01+11 02+12 03+13
    // d2 = 40+50 41+51 42+52 43+53
    "vpaddl.u8    q0, q0                       \n"
    "vpaddl.u8    q2, q2                       \n"

    // d3 = 60+70 61+71 62+72 63+73
    "vpaddl.u8    d3, d3                       \n"
    "vpaddl.u8    d7, d7                       \n"

    // combine source lines
    "vadd.u16     q0, q2                       \n"
    "vadd.u16     d4, d3, d7                   \n"

    // dst_ptr[3] = (s[6] + s[7] + s[6+st] + s[7+st]) / 4
    "vqrshrn.u16  d4, q2, #2                   \n"

    // Shuffle 2,3 reg around so that 2 can be added to the
    //  0,1 reg and 3 can be added to the 4,5 reg. This
    //  requires expanding from u8 to u16 as the 0,1 and 4,5
    //  registers are already expanded. Then do transposes
    //  to get aligned.
    // q2 = xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33
    "vmovl.u8     q1, d2                       \n"
    "vmovl.u8     q3, d6                       \n"

    // combine source lines
    "vadd.u16     q1, q3                       \n"

    // d4 = xx 20 xx 30 xx 22 xx 32
    // d5 = xx 21 xx 31 xx 23 xx 33
    "vtrn.u32     d2, d3                       \n"

    // d4 = xx 20 xx 21 xx 22 xx 23
    // d5 = xx 30 xx 31 xx 32 xx 33
    "vtrn.u16     d2, d3                       \n"

    // 0+1+2, 3+4+5
    "vadd.u16     q0, q1                       \n"

    // Need to divide, but can't downshift as the the value
    //  isn't a power of 2. So multiply by 65536 / n
    //  and take the upper 16 bits.
    "vqrdmulh.s16 q0, q0, q13                  \n"

    // Align for table lookup, vtbl requires registers to
    //  be adjacent
    "vmov.u8      d2, d4                       \n"

    "vtbl.u8      d3, {d0, d1, d2}, d28        \n"
    "vtbl.u8      d4, {d0, d1, d2}, d29        \n"

    "vst1.u8      {d3}, [%1]!                  \n"
    "vst1.u32     {d4[0]}, [%1]!               \n"
    "bgt          1b                           \n"
fbarchard@google.com's avatar
fbarchard@google.com committed
462 463 464 465 466 467 468
  : "+r"(src_ptr),       // %0
    "+r"(dst_ptr),       // %1
    "+r"(dst_width),     // %2
    "+r"(src_stride)     // %3
  : "r"(&kMult38_Div6),  // %4
    "r"(&kShuf38_2)      // %5
  : "q0", "q1", "q2", "q3", "q13", "q14", "memory", "cc"
469 470 471 472 473 474 475 476 477
  );
}

// 16x2 -> 16x1
void ScaleFilterRows_NEON(uint8* dst_ptr,
                          const uint8* src_ptr, ptrdiff_t src_stride,
                          int dst_width, int source_y_fraction) {
  asm volatile (
    "cmp          %4, #0                       \n"
478
    "beq          100f                         \n"
479
    "add          %2, %1                       \n"
480 481
    "cmp          %4, #64                      \n"
    "beq          75f                          \n"
482
    "cmp          %4, #128                     \n"
483 484 485
    "beq          50f                          \n"
    "cmp          %4, #192                     \n"
    "beq          25f                          \n"
486 487 488 489

    "vdup.8       d5, %4                       \n"
    "rsb          %4, #256                     \n"
    "vdup.8       d4, %4                       \n"
490
    // General purpose row blend.
491
  "1:                                          \n"
492 493 494 495 496 497 498 499 500 501 502
    "vld1.u8      {q0}, [%1]!                  \n"
    "vld1.u8      {q1}, [%2]!                  \n"
    "subs         %3, #16                      \n"
    "vmull.u8     q13, d0, d4                  \n"
    "vmull.u8     q14, d1, d4                  \n"
    "vmlal.u8     q13, d2, d5                  \n"
    "vmlal.u8     q14, d3, d5                  \n"
    "vrshrn.u16   d0, q13, #8                  \n"
    "vrshrn.u16   d1, q14, #8                  \n"
    "vst1.u8      {q0}, [%0]!                  \n"
    "bgt          1b                           \n"
503
    "b            99f                          \n"
504

505 506
    // Blend 25 / 75.
  "25:                                         \n"
507
    "vld1.u8      {q0}, [%1]!                  \n"
508
    "vld1.u8      {q1}, [%2]!                  \n"
509
    "subs         %3, #16                      \n"
510 511
    "vrhadd.u8    q0, q1                       \n"
    "vrhadd.u8    q0, q1                       \n"
512
    "vst1.u8      {q0}, [%0]!                  \n"
513 514
    "bgt          25b                          \n"
    "b            99f                          \n"
515

516 517
    // Blend 50 / 50.
  "50:                                         \n"
518 519 520 521 522
    "vld1.u8      {q0}, [%1]!                  \n"
    "vld1.u8      {q1}, [%2]!                  \n"
    "subs         %3, #16                      \n"
    "vrhadd.u8    q0, q1                       \n"
    "vst1.u8      {q0}, [%0]!                  \n"
523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544
    "bgt          50b                          \n"
    "b            99f                          \n"

    // Blend 75 / 25.
  "75:                                         \n"
    "vld1.u8      {q1}, [%1]!                  \n"
    "vld1.u8      {q0}, [%2]!                  \n"
    "subs         %3, #16                      \n"
    "vrhadd.u8    q0, q1                       \n"
    "vrhadd.u8    q0, q1                       \n"
    "vst1.u8      {q0}, [%0]!                  \n"
    "bgt          75b                          \n"
    "b            99f                          \n"

    // Blend 100 / 0 - Copy row unchanged.
  "100:                                        \n"
    "vld1.u8      {q0}, [%1]!                  \n"
    "subs         %3, #16                      \n"
    "vst1.u8      {q0}, [%0]!                  \n"
    "bgt          100b                         \n"

  "99:                                         \n"
545
    "vst1.u8      {d1[7]}, [%0]                \n"
fbarchard@google.com's avatar
fbarchard@google.com committed
546 547 548 549 550 551 552
  : "+r"(dst_ptr),          // %0
    "+r"(src_ptr),          // %1
    "+r"(src_stride),       // %2
    "+r"(dst_width),        // %3
    "+r"(source_y_fraction) // %4
  :
  : "q0", "q1", "d4", "d5", "q13", "q14", "memory", "cc"
553 554 555 556 557 558 559 560
  );
}
#endif  // __ARM_NEON__

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