scale_neon.cc 26.9 KB
Newer Older
1 2 3 4 5 6
/*
 *  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
7
 *  in the file PATENTS. All contributing project authors may
8 9 10 11 12 13 14 15 16 17
 *  be found in the AUTHORS file in the root of the source tree.
 */

#include "libyuv/row.h"

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

18
// This module is for GCC Neon.
19
#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__)
20

21 22
// NEON downscalers with interpolation.
// Provided by Fritz Koenig
23

24
// Read 32x1 throw away even pixels, and write 16x1.
25
void ScaleRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
26 27
                        uint8* dst, int dst_width) {
  asm volatile (
28
    ".p2align   2                              \n"
29
  "1:                                          \n"
30
    // load even pixels into q0, odd into q1
31
    "vld2.8     {q0, q1}, [%0]!                \n"
32
    "subs       %2, %2, #16                    \n"  // 16 processed per loop
33
    "vst1.8     {q1}, [%1]!                    \n"  // store odd pixels
34
    "bgt        1b                             \n"
fbarchard@google.com's avatar
fbarchard@google.com committed
35 36 37 38 39
  : "+r"(src_ptr),          // %0
    "+r"(dst),              // %1
    "+r"(dst_width)         // %2
  :
  : "q0", "q1"              // Clobber List
40 41 42
  );
}

43
// Read 32x2 average down and write 16x1.
44
void ScaleRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
45 46 47 48
                           uint8* dst, int dst_width) {
  asm volatile (
    // change the stride to row 2 pointer
    "add        %1, %0                         \n"
49
    ".p2align   2                              \n"
50
  "1:                                          \n"
51 52
    "vld1.8     {q0, q1}, [%0]!                \n"  // load row 1 and post inc
    "vld1.8     {q2, q3}, [%1]!                \n"  // load row 2 and post inc
fbarchard@google.com's avatar
fbarchard@google.com committed
53
    "subs       %3, %3, #16                    \n"  // 16 processed per loop
54 55 56 57 58 59
    "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"
60
    "vst1.8     {q0}, [%2]!                    \n"
61
    "bgt        1b                             \n"
fbarchard@google.com's avatar
fbarchard@google.com committed
62 63 64 65 66 67 68
  : "+r"(src_ptr),          // %0
    "+r"(src_stride),       // %1
    "+r"(dst),              // %2
    "+r"(dst_width)         // %3
  :
  : "q0", "q1", "q2", "q3"     // Clobber List
  );
69 70
}

71
void ScaleRowDown4_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
72 73
                        uint8* dst_ptr, int dst_width) {
  asm volatile (
74
    ".p2align   2                              \n"
75
  "1:                                          \n"
76 77 78 79
    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n" // src line 0
    "subs       %2, %2, #8                     \n" // 8 processed per loop
    "vst1.8     {d2}, [%1]!                    \n"
    "bgt        1b                             \n"
fbarchard@google.com's avatar
fbarchard@google.com committed
80 81 82 83 84
  : "+r"(src_ptr),          // %0
    "+r"(dst_ptr),          // %1
    "+r"(dst_width)         // %2
  :
  : "q0", "q1", "memory", "cc"
85 86 87
  );
}

88
void ScaleRowDown4Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
89 90 91 92 93
                           uint8* dst_ptr, int dst_width) {
  asm volatile (
    "add        r4, %0, %3                     \n"
    "add        r5, r4, %3                     \n"
    "add        %3, r5, %3                     \n"
94
    ".p2align   2                              \n"
95
  "1:                                          \n"
96 97 98 99
    "vld1.8     {q0}, [%0]!                    \n"   // load up 16x4
    "vld1.8     {q1}, [r4]!                    \n"
    "vld1.8     {q2}, [r5]!                    \n"
    "vld1.8     {q3}, [%3]!                    \n"
100
    "subs       %2, %2, #4                     \n"
101 102 103 104 105 106 107
    "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"
108
    "vst1.32    {d0[0]}, [%1]!                 \n"
109
    "bgt        1b                             \n"
fbarchard@google.com's avatar
fbarchard@google.com committed
110 111 112 113 114
  : "+r"(src_ptr),          // %0
    "+r"(dst_ptr),          // %1
    "+r"(dst_width)         // %2
  : "r"(src_stride)         // %3
  : "r4", "r5", "q0", "q1", "q2", "q3", "memory", "cc"
115 116 117 118 119 120 121
  );
}

// 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,
122
                         ptrdiff_t src_stride,
123 124
                         uint8* dst_ptr, int dst_width) {
  asm volatile (
125
    ".p2align   2                              \n"
126
  "1:                                          \n"
127 128 129 130 131
    "vld4.8     {d0, d1, d2, d3}, [%0]!      \n" // src line 0
    "subs       %2, %2, #24                  \n"
    "vmov       d2, d3                       \n" // order d0, d1, d2
    "vst3.8     {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
void ScaleRowDown34_0_Box_NEON(const uint8* src_ptr,
141 142 143
                               ptrdiff_t src_stride,
                               uint8* dst_ptr, int dst_width) {
  asm volatile (
144 145 146
    "vmov.u8    d24, #3                        \n"
    "add        %3, %0                         \n"
    ".p2align   2                              \n"
147
  "1:                                          \n"
148 149
    "vld4.8       {d0, d1, d2, d3}, [%0]!      \n" // src line 0
    "vld4.8       {d4, d5, d6, d7}, [%3]!      \n" // src line 1
150
    "subs         %2, %2, #24                  \n"
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

    // 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"

185
    "vst3.8       {d0, d1, d2}, [%1]!          \n"
186 187

    "bgt          1b                           \n"
fbarchard@google.com's avatar
fbarchard@google.com committed
188 189 190 191 192 193
  : "+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"
194 195 196
  );
}

197
void ScaleRowDown34_1_Box_NEON(const uint8* src_ptr,
198 199 200
                               ptrdiff_t src_stride,
                               uint8* dst_ptr, int dst_width) {
  asm volatile (
201 202 203
    "vmov.u8    d24, #3                        \n"
    "add        %3, %0                         \n"
    ".p2align   2                              \n"
204
  "1:                                          \n"
205 206
    "vld4.8       {d0, d1, d2, d3}, [%0]!      \n" // src line 0
    "vld4.8       {d4, d5, d6, d7}, [%3]!      \n" // src line 1
207
    "subs         %2, %2, #24                  \n"
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"

225
    "vst3.8       {d0, d1, d2}, [%1]!          \n"
226
    "bgt          1b                           \n"
fbarchard@google.com's avatar
fbarchard@google.com committed
227 228 229 230 231 232
  : "+r"(src_ptr),          // %0
    "+r"(dst_ptr),          // %1
    "+r"(dst_width),        // %2
    "+r"(src_stride)        // %3
  :
  : "r4", "q0", "q1", "q2", "q3", "d24", "memory", "cc"
233 234 235 236
  );
}

#define HAS_SCALEROWDOWN38_NEON
237
static uvec8 kShuf38 =
238
  { 0, 3, 6, 8, 11, 14, 16, 19, 22, 24, 27, 30, 0, 0, 0, 0 };
239
static uvec8 kShuf38_2 =
240
  { 0, 8, 16, 2, 10, 17, 4, 12, 18, 6, 14, 19, 0, 0, 0, 0 };
241
static vec16 kMult38_Div6 =
242 243
  { 65536 / 12, 65536 / 12, 65536 / 12, 65536 / 12,
    65536 / 12, 65536 / 12, 65536 / 12, 65536 / 12 };
244
static vec16 kMult38_Div9 =
245 246 247 248 249
  { 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,
250
                         ptrdiff_t src_stride,
251 252
                         uint8* dst_ptr, int dst_width) {
  asm volatile (
253 254
    "vld1.8     {q3}, [%3]                     \n"
    ".p2align   2                              \n"
255
  "1:                                          \n"
256 257 258 259 260 261 262
    "vld1.8     {d0, d1, d2, d3}, [%0]!        \n"
    "subs       %2, %2, #12                    \n"
    "vtbl.u8    d4, {d0, d1, d2, d3}, d6       \n"
    "vtbl.u8    d5, {d0, d1, d2, d3}, d7       \n"
    "vst1.8     {d4}, [%1]!                    \n"
    "vst1.32    {d5[0]}, [%1]!                 \n"
    "bgt        1b                             \n"
fbarchard@google.com's avatar
fbarchard@google.com committed
263 264 265 266 267
  : "+r"(src_ptr),          // %0
    "+r"(dst_ptr),          // %1
    "+r"(dst_width)         // %2
  : "r"(&kShuf38)           // %3
  : "d0", "d1", "d2", "d3", "d4", "d5", "memory", "cc"
268 269 270 271
  );
}

// 32x3 -> 12x1
272
void OMITFP ScaleRowDown38_3_Box_NEON(const uint8* src_ptr,
273 274 275
                                      ptrdiff_t src_stride,
                                      uint8* dst_ptr, int dst_width) {
  asm volatile (
276 277 278 279 280 281
    "vld1.16    {q13}, [%4]                    \n"
    "vld1.8     {q14}, [%5]                    \n"
    "vld1.8     {q15}, [%6]                    \n"
    "add        r4, %0, %3, lsl #1             \n"
    "add        %3, %0                         \n"
    ".p2align   2                              \n"
282
  "1:                                          \n"
283 284 285 286 287

    // 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
288 289 290
    "vld4.8       {d0, d1, d2, d3}, [%0]!      \n"
    "vld4.8       {d4, d5, d6, d7}, [%3]!      \n"
    "vld4.8       {d16, d17, d18, d19}, [r4]!  \n"
291
    "subs         %2, %2, #12                  \n"
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

    // 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"

367 368
    "vst1.8       {d3}, [%1]!                  \n"
    "vst1.32      {d4[0]}, [%1]!               \n"
369
    "bgt          1b                           \n"
fbarchard@google.com's avatar
fbarchard@google.com committed
370 371 372 373 374 375 376 377 378
  : "+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"
379 380 381 382
  );
}

// 32x2 -> 12x1
383
void ScaleRowDown38_2_Box_NEON(const uint8* src_ptr,
384 385 386
                               ptrdiff_t src_stride,
                               uint8* dst_ptr, int dst_width) {
  asm volatile (
387 388 389 390
    "vld1.16    {q13}, [%4]                    \n"
    "vld1.8     {q14}, [%5]                    \n"
    "add        %3, %0                         \n"
    ".p2align   2                              \n"
391
  "1:                                          \n"
392 393 394 395 396

    // 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
397 398
    "vld4.8       {d0, d1, d2, d3}, [%0]!      \n"
    "vld4.8       {d4, d5, d6, d7}, [%3]!      \n"
399
    "subs         %2, %2, #12                  \n"
400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463

    // 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"

464 465
    "vst1.8       {d3}, [%1]!                  \n"
    "vst1.32      {d4[0]}, [%1]!               \n"
466
    "bgt          1b                           \n"
fbarchard@google.com's avatar
fbarchard@google.com committed
467 468 469 470 471 472 473
  : "+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"
474 475 476 477 478 479 480 481 482
  );
}

// 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"
483
    "beq          100f                         \n"
484
    "add          %2, %1                       \n"
485 486
    "cmp          %4, #64                      \n"
    "beq          75f                          \n"
487
    "cmp          %4, #128                     \n"
488 489 490
    "beq          50f                          \n"
    "cmp          %4, #192                     \n"
    "beq          25f                          \n"
491 492 493 494

    "vdup.8       d5, %4                       \n"
    "rsb          %4, #256                     \n"
    "vdup.8       d4, %4                       \n"
495
    // General purpose row blend.
496
  "1:                                          \n"
497 498
    "vld1.8       {q0}, [%1]!                  \n"
    "vld1.8       {q1}, [%2]!                  \n"
499
    "subs         %3, %3, #16                  \n"
500 501 502 503 504 505
    "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"
506
    "vst1.8       {q0}, [%0]!                  \n"
507
    "bgt          1b                           \n"
508
    "b            99f                          \n"
509

510 511
    // Blend 25 / 75.
  "25:                                         \n"
512 513
    "vld1.8       {q0}, [%1]!                  \n"
    "vld1.8       {q1}, [%2]!                  \n"
514
    "subs         %3, %3, #16                  \n"
515 516
    "vrhadd.u8    q0, q1                       \n"
    "vrhadd.u8    q0, q1                       \n"
517
    "vst1.8       {q0}, [%0]!                  \n"
518 519
    "bgt          25b                          \n"
    "b            99f                          \n"
520

521 522
    // Blend 50 / 50.
  "50:                                         \n"
523 524
    "vld1.8       {q0}, [%1]!                  \n"
    "vld1.8       {q1}, [%2]!                  \n"
525
    "subs         %3, %3, #16                  \n"
526
    "vrhadd.u8    q0, q1                       \n"
527
    "vst1.8       {q0}, [%0]!                  \n"
528 529 530 531 532
    "bgt          50b                          \n"
    "b            99f                          \n"

    // Blend 75 / 25.
  "75:                                         \n"
533 534
    "vld1.8       {q1}, [%1]!                  \n"
    "vld1.8       {q0}, [%2]!                  \n"
535
    "subs         %3, %3, #16                  \n"
536 537
    "vrhadd.u8    q0, q1                       \n"
    "vrhadd.u8    q0, q1                       \n"
538
    "vst1.8       {q0}, [%0]!                  \n"
539 540 541 542 543
    "bgt          75b                          \n"
    "b            99f                          \n"

    // Blend 100 / 0 - Copy row unchanged.
  "100:                                        \n"
544
    "vld1.8       {q0}, [%1]!                  \n"
545
    "subs         %3, %3, #16                  \n"
546
    "vst1.8       {q0}, [%0]!                  \n"
547 548 549
    "bgt          100b                         \n"

  "99:                                         \n"
550
    "vst1.8       {d1[7]}, [%0]                \n"
fbarchard@google.com's avatar
fbarchard@google.com committed
551 552 553 554 555 556 557
  : "+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"
558 559
  );
}
560

561
void ScaleARGBRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
562 563
                            uint8* dst, int dst_width) {
  asm volatile (
564
    ".p2align   2                              \n"
565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585
  "1:                                          \n"
    // load even pixels into q0, odd into q1
    "vld2.32    {q0, q1}, [%0]!                \n"
    "vld2.32    {q2, q3}, [%0]!                \n"
    "subs       %2, %2, #8                     \n"  // 8 processed per loop
    "vst1.8     {q1}, [%1]!                    \n"  // store odd pixels
    "vst1.8     {q3}, [%1]!                    \n"
    "bgt        1b                             \n"
  : "+r"(src_ptr),          // %0
    "+r"(dst),              // %1
    "+r"(dst_width)         // %2
  :
  : "memory", "cc", "q0", "q1", "q2", "q3"  // Clobber List
  );
}

void ScaleARGBRowDown2Box_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, %1, %0                     \n"
586
    ".p2align   2                              \n"
587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617
  "1:                                          \n"
    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
    "vpaddl.u8  q3, q3                         \n"  // A 16 bytes -> 8 shorts.
    "vld4.8     {d16, d18, d20, d22}, [%1]!    \n"  // load 8 more ARGB pixels.
    "vld4.8     {d17, d19, d21, d23}, [%1]!    \n"  // load last 8 ARGB pixels.
    "vpadal.u8  q0, q8                         \n"  // B 16 bytes -> 8 shorts.
    "vpadal.u8  q1, q9                         \n"  // G 16 bytes -> 8 shorts.
    "vpadal.u8  q2, q10                        \n"  // R 16 bytes -> 8 shorts.
    "vpadal.u8  q3, q11                        \n"  // A 16 bytes -> 8 shorts.
    "vrshrn.u16 d0, q0, #2                     \n"  // downshift, round and pack
    "vrshrn.u16 d1, q1, #2                     \n"
    "vrshrn.u16 d2, q2, #2                     \n"
    "vrshrn.u16 d3, q3, #2                     \n"
    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"
    "bgt        1b                             \n"
  : "+r"(src_ptr),          // %0
    "+r"(src_stride),       // %1
    "+r"(dst),              // %2
    "+r"(dst_width)         // %3
  :
  : "memory", "cc", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11"
  );
}

// Reads 4 pixels at a time.
// Alignment requirement: src_argb 4 byte aligned.
618 619
void ScaleARGBRowDownEven_NEON(const uint8* src_argb,  ptrdiff_t src_stride,
                               int src_stepx, uint8* dst_argb, int dst_width) {
620 621
  asm volatile (
    "mov        r12, %3, lsl #2                \n"
622
    ".p2align   2                              \n"
623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644
  "1:                                          \n"
    "vld1.32    {d0[0]}, [%0], r12             \n"
    "vld1.32    {d0[1]}, [%0], r12             \n"
    "vld1.32    {d1[0]}, [%0], r12             \n"
    "vld1.32    {d1[1]}, [%0], r12             \n"
    "subs       %2, %2, #4                     \n"  // 4 pixels per loop.
    "vst1.8     {q0}, [%1]!                    \n"
    "bgt        1b                             \n"
  : "+r"(src_argb),    // %0
    "+r"(dst_argb),    // %1
    "+r"(dst_width)    // %2
  : "r"(src_stepx)     // %3
  : "memory", "cc", "r12", "q0"
  );
}

// Reads 4 pixels at a time.
// Alignment requirement: src_argb 4 byte aligned.
void ScaleARGBRowDownEvenBox_NEON(const uint8* src_argb, ptrdiff_t src_stride,
                                  int src_stepx,
                                  uint8* dst_argb, int dst_width) {
  asm volatile (
645 646 647
    "mov        r12, %4, lsl #2                \n"
    "add        %1, %1, %0                     \n"
    ".p2align   2                              \n"
648
  "1:                                          \n"
649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664
    "vld1.8     {d0}, [%0], r12                \n"  // Read 4 2x2 blocks -> 2x1
    "vld1.8     {d1}, [%1], r12                \n"
    "vld1.8     {d2}, [%0], r12                \n"
    "vld1.8     {d3}, [%1], r12                \n"
    "vld1.8     {d4}, [%0], r12                \n"
    "vld1.8     {d5}, [%1], r12                \n"
    "vld1.8     {d6}, [%0], r12                \n"
    "vld1.8     {d7}, [%1], r12                \n"
    "vaddl.u8   q0, d0, d1                     \n"
    "vaddl.u8   q1, d2, d3                     \n"
    "vaddl.u8   q2, d4, d5                     \n"
    "vaddl.u8   q3, d6, d7                     \n"
    "vswp.8     d1, d2                         \n"  // ab_cd -> ac_bd
    "vswp.8     d5, d6                         \n"  // ef_gh -> eg_fh
    "vadd.u16   q0, q0, q1                     \n"  // (a+b)_(c+d)
    "vadd.u16   q2, q2, q3                     \n"  // (e+f)_(g+h)
665 666 667 668 669 670 671 672 673 674 675 676 677 678
    "vrshrn.u16 d0, q0, #2                     \n"  // first 2 pixels.
    "vrshrn.u16 d1, q2, #2                     \n"  // next 2 pixels.
    "subs       %3, %3, #4                     \n"  // 4 pixels per loop.
    "vst1.8     {q0}, [%2]!                    \n"
    "bgt        1b                             \n"
  : "+r"(src_argb),    // %0
    "+r"(src_stride),  // %1
    "+r"(dst_argb),    // %2
    "+r"(dst_width)    // %3
  : "r"(src_stepx)     // %4
  : "memory", "cc", "r12", "q0", "q1", "q2", "q3"
  );
}

679 680 681 682 683 684
#endif  // __ARM_NEON__

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