vpx_scale/arm/neon/vp8_vpxyv12_extendframeborders_neon.asm

   1 ;
   2 ;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
   3 ;
   4 ;  Use of this source code is governed by a BSD-style license
   5 ;  that can be found in the LICENSE file in the root of the source
   6 ;  tree. An additional intellectual property rights grant can be found
   7 ;  in the file PATENTS.  All contributing project authors may
   8 ;  be found in the AUTHORS file in the root of the source tree.
   9 ;
  10
  11
  12     EXPORT  |vp8_yv12_extend_frame_borders_neon|
  13     ARM
  14     REQUIRE8
  15     PRESERVE8
  16
  17     INCLUDE asm_com_offsets.asm
  18
  19     AREA ||.text||, CODE, READONLY, ALIGN=2
  20 ;void vp8_yv12_extend_frame_borders_neon (YV12_BUFFER_CONFIG *ybf);
  21 ;Note: this is VP8 function, which has border=32 and 16. Internal y_width and y_height
  22 ; are always multiples of 16.
  23
  24 |vp8_yv12_extend_frame_borders_neon| PROC
  25     push            {r4 - r10, lr}
  26     vpush           {d8 - d15}
  27
  28     ;Not need to load y_width, since: y_width = y_stride - 2*border
  29     ldr             r3, [r0, #yv12_buffer_config_border]
  30     ldr             r1, [r0, #yv12_buffer_config_y_buffer]       ;srcptr1
  31     ldr             r4, [r0, #yv12_buffer_config_y_height]
  32     ldr             lr, [r0, #yv12_buffer_config_y_stride]
  33
  34     cmp             r3, #16
  35     beq             b16_extend_frame_borders
  36
  37 ;=======================
  38 b32_extend_frame_borders
  39 ;border = 32
  40 ;=======================
  41 ;Border copy for Y plane
  42 ;copy the left and right most columns out
  43     sub             r5, r1, r3              ;destptr1
  44     add             r6, r1, lr
  45     sub             r6, r6, r3, lsl #1      ;destptr2
  46     sub             r2, r6, #1              ;srcptr2
  47
  48     ;Do four rows at one time
  49     mov             r12, r4, lsr #2
  50
  51 copy_left_right_y
  52     vld1.8          {d0[], d1[]}, [r1], lr
  53     vld1.8          {d4[], d5[]}, [r2], lr
  54     vld1.8          {d8[], d9[]}, [r1], lr
  55     vld1.8          {d12[], d13[]}, [r2], lr
  56     vld1.8          {d16[], d17[]},  [r1], lr
  57     vld1.8          {d20[], d21[]}, [r2], lr
  58     vld1.8          {d24[], d25[]}, [r1], lr
  59     vld1.8          {d28[], d29[]}, [r2], lr
  60
  61     vmov            q1, q0
  62     vmov            q3, q2
  63     vmov            q5, q4
  64     vmov            q7, q6
  65     vmov            q9, q8
  66     vmov            q11, q10
  67     vmov            q13, q12
  68     vmov            q15, q14
  69
  70     subs            r12, r12, #1
  71
  72     vst1.8          {q0, q1}, [r5], lr
  73     vst1.8          {q2, q3}, [r6], lr
  74     vst1.8          {q4, q5}, [r5], lr
  75     vst1.8          {q6, q7}, [r6], lr
  76     vst1.8          {q8, q9}, [r5], lr
  77     vst1.8          {q10, q11}, [r6], lr
  78     vst1.8          {q12, q13}, [r5], lr
  79     vst1.8          {q14, q15}, [r6], lr
  80
  81     bne             copy_left_right_y
  82
  83 ;Now copy the top and bottom source lines into each line of the respective borders
  84     ldr             r7, [r0, #yv12_buffer_config_y_buffer]       ;srcptr1
  85     mul             r8, r3, lr
  86
  87     mov             r12, lr, lsr #7
  88
  89     sub             r6, r1, r3              ;destptr2
  90     sub             r2, r6, lr              ;srcptr2
  91     sub             r1, r7, r3              ;srcptr1
  92     sub             r5, r1, r8              ;destptr1
  93
  94 copy_top_bottom_y
  95     vld1.8          {q0, q1}, [r1]!
  96     vld1.8          {q8, q9}, [r2]!
  97     vld1.8          {q2, q3}, [r1]!
  98     vld1.8          {q10, q11}, [r2]!
  99     vld1.8          {q4, q5}, [r1]!
 100     vld1.8          {q12, q13}, [r2]!
 101     vld1.8          {q6, q7}, [r1]!
 102     vld1.8          {q14, q15}, [r2]!
 103
 104     mov             r7, r3
 105
 106 top_bottom_32
 107     subs            r7, r7, #1
 108
 109     vst1.8          {q0, q1}, [r5]!
 110     vst1.8          {q8, q9}, [r6]!
 111     vst1.8          {q2, q3}, [r5]!
 112     vst1.8          {q10, q11}, [r6]!
 113     vst1.8          {q4, q5}, [r5]!
 114     vst1.8          {q12, q13}, [r6]!
 115     vst1.8          {q6, q7}, [r5]!
 116     vst1.8          {q14, q15}, [r6]!
 117
 118     add             r5, r5, lr
 119     sub             r5, r5, #128
 120     add             r6, r6, lr
 121     sub             r6, r6, #128
 122
 123     bne             top_bottom_32
 124
 125     sub             r5, r1, r8
 126     add             r6, r2, lr
 127
 128     subs            r12, r12, #1
 129     bne             copy_top_bottom_y
 130
 131     mov             r7, lr, lsr #4              ;check to see if extra copy is needed
 132     ands            r7, r7, #0x7
 133     bne             extra_top_bottom_y
 134 end_of_border_copy_y
 135
 136 ;Border copy for U, V planes
 137     ldr             r1, [r0, #yv12_buffer_config_u_buffer]       ;srcptr1
 138     mov             lr, lr, lsr #1              ;uv_stride
 139     mov             r3, r3, lsr #1              ;border
 140     mov             r4, r4, lsr #1              ;uv_height
 141     mov             r8, r8, lsr #2
 142
 143     mov             r10, #2
 144
 145 ;copy the left and right most columns out
 146 border_copy_uv
 147     sub             r5, r1, r3              ;destptr1
 148     add             r6, r1, lr
 149     sub             r6, r6, r3, lsl #1      ;destptr2
 150     sub             r2, r6, #1              ;srcptr2
 151
 152     mov             r7, r1
 153
 154     ;Do eight rows at one time
 155     mov             r12, r4, lsr #3
 156
 157 copy_left_right_uv
 158     vld1.8          {d0[], d1[]}, [r1], lr
 159     vld1.8          {d2[], d3[]}, [r2], lr
 160     vld1.8          {d4[], d5[]}, [r1], lr
 161     vld1.8          {d6[], d7[]}, [r2], lr
 162     vld1.8          {d8[], d9[]},  [r1], lr
 163     vld1.8          {d10[], d11[]}, [r2], lr
 164     vld1.8          {d12[], d13[]}, [r1], lr
 165     vld1.8          {d14[], d15[]}, [r2], lr
 166     vld1.8          {d16[], d17[]}, [r1], lr
 167     vld1.8          {d18[], d19[]}, [r2], lr
 168     vld1.8          {d20[], d21[]}, [r1], lr
 169     vld1.8          {d22[], d23[]}, [r2], lr
 170     vld1.8          {d24[], d25[]},  [r1], lr
 171     vld1.8          {d26[], d27[]}, [r2], lr
 172     vld1.8          {d28[], d29[]}, [r1], lr
 173     vld1.8          {d30[], d31[]}, [r2], lr
 174
 175     subs            r12, r12, #1
 176
 177     vst1.8          {q0}, [r5], lr
 178     vst1.8          {q1}, [r6], lr
 179     vst1.8          {q2}, [r5], lr
 180     vst1.8          {q3}, [r6], lr
 181     vst1.8          {q4}, [r5], lr
 182     vst1.8          {q5}, [r6], lr
 183     vst1.8          {q6}, [r5], lr
 184     vst1.8          {q7}, [r6], lr
 185     vst1.8          {q8}, [r5], lr
 186     vst1.8          {q9}, [r6], lr
 187     vst1.8          {q10}, [r5], lr
 188     vst1.8          {q11}, [r6], lr
 189     vst1.8          {q12}, [r5], lr
 190     vst1.8          {q13}, [r6], lr
 191     vst1.8          {q14}, [r5], lr
 192     vst1.8          {q15}, [r6], lr
 193
 194     bne             copy_left_right_uv
 195
 196 ;Now copy the top and bottom source lines into each line of the respective borders
 197     mov             r12, lr, lsr #6
 198
 199     sub             r6, r1, r3              ;destptr2
 200     sub             r2, r6, lr              ;srcptr2
 201     sub             r1, r7, r3              ;srcptr1
 202     sub             r5, r1, r8              ;destptr1
 203
 204 copy_top_bottom_uv
 205     vld1.8          {q0, q1}, [r1]!
 206     vld1.8          {q8, q9}, [r2]!
 207     vld1.8          {q2, q3}, [r1]!
 208     vld1.8          {q10, q11}, [r2]!
 209
 210     mov             r7, r3
 211
 212 top_bottom_16
 213     subs            r7, r7, #1
 214
 215     vst1.8          {q0, q1}, [r5]!
 216     vst1.8          {q8, q9}, [r6]!
 217     vst1.8          {q2, q3}, [r5]!
 218     vst1.8          {q10, q11}, [r6]!
 219
 220     add             r5, r5, lr
 221     sub             r5, r5, #64
 222     add             r6, r6, lr
 223     sub             r6, r6, #64
 224
 225     bne             top_bottom_16
 226
 227     sub             r5, r1, r8
 228     add             r6, r2, lr
 229
 230     subs            r12, r12, #1
 231     bne             copy_top_bottom_uv
 232
 233     mov             r7, lr, lsr #3              ;check to see if extra copy is needed
 234     ands            r7, r7, #0x7
 235     bne             extra_top_bottom_uv
 236
 237 end_of_border_copy_uv
 238     subs            r10, r10, #1
 239     ldrne           r1, [r0, #yv12_buffer_config_v_buffer]       ;srcptr1
 240     bne             border_copy_uv
 241
 242     vpop            {d8 - d15}
 243     pop             {r4 - r10, pc}
 244
 245 ;;;;;;;;;;;;;;;;;;;;;;
 246 ;extra copy part for Y
 247 extra_top_bottom_y
 248     vld1.8          {q0}, [r1]!
 249     vld1.8          {q2}, [r2]!
 250
 251     mov             r9, r3, lsr #3
 252
 253 extra_top_bottom_32
 254     subs            r9, r9, #1
 255
 256     vst1.8          {q0}, [r5], lr
 257     vst1.8          {q2}, [r6], lr
 258     vst1.8          {q0}, [r5], lr
 259     vst1.8          {q2}, [r6], lr
 260     vst1.8          {q0}, [r5], lr
 261     vst1.8          {q2}, [r6], lr
 262     vst1.8          {q0}, [r5], lr
 263     vst1.8          {q2}, [r6], lr
 264     vst1.8          {q0}, [r5], lr
 265     vst1.8          {q2}, [r6], lr
 266     vst1.8          {q0}, [r5], lr
 267     vst1.8          {q2}, [r6], lr
 268     vst1.8          {q0}, [r5], lr
 269     vst1.8          {q2}, [r6], lr
 270     vst1.8          {q0}, [r5], lr
 271     vst1.8          {q2}, [r6], lr
 272     bne             extra_top_bottom_32
 273
 274     sub             r5, r1, r8
 275     add             r6, r2, lr
 276     subs            r7, r7, #1
 277     bne             extra_top_bottom_y
 278
 279     b               end_of_border_copy_y
 280
 281 ;extra copy part for UV
 282 extra_top_bottom_uv
 283     vld1.8          {d0}, [r1]!
 284     vld1.8          {d8}, [r2]!
 285
 286     mov             r9, r3, lsr #3
 287
 288 extra_top_bottom_16
 289     subs            r9, r9, #1
 290
 291     vst1.8          {d0}, [r5], lr
 292     vst1.8          {d8}, [r6], lr
 293     vst1.8          {d0}, [r5], lr
 294     vst1.8          {d8}, [r6], lr
 295     vst1.8          {d0}, [r5], lr
 296     vst1.8          {d8}, [r6], lr
 297     vst1.8          {d0}, [r5], lr
 298     vst1.8          {d8}, [r6], lr
 299     vst1.8          {d0}, [r5], lr
 300     vst1.8          {d8}, [r6], lr
 301     vst1.8          {d0}, [r5], lr
 302     vst1.8          {d8}, [r6], lr
 303     vst1.8          {d0}, [r5], lr
 304     vst1.8          {d8}, [r6], lr
 305     vst1.8          {d0}, [r5], lr
 306     vst1.8          {d8}, [r6], lr
 307     bne             extra_top_bottom_16
 308
 309     sub             r5, r1, r8
 310     add             r6, r2, lr
 311     subs            r7, r7, #1
 312     bne             extra_top_bottom_uv
 313
 314     b               end_of_border_copy_uv
 315
 316
 317 ;=======================
 318 b16_extend_frame_borders
 319 ;border = 16
 320 ;=======================
 321 ;Border copy for Y plane
 322 ;copy the left and right most columns out
 323     sub             r5, r1, r3              ;destptr1
 324     add             r6, r1, lr
 325     sub             r6, r6, r3, lsl #1      ;destptr2
 326     sub             r2, r6, #1              ;srcptr2
 327
 328     ;Do four rows at one time
 329     mov             r12, r4, lsr #2
 330
 331 copy_left_right_y_b16
 332     vld1.8          {d0[], d1[]}, [r1], lr
 333     vld1.8          {d4[], d5[]}, [r2], lr
 334     vld1.8          {d8[], d9[]}, [r1], lr
 335     vld1.8          {d12[], d13[]}, [r2], lr
 336     vld1.8          {d16[], d17[]},  [r1], lr
 337     vld1.8          {d20[], d21[]}, [r2], lr
 338     vld1.8          {d24[], d25[]}, [r1], lr
 339     vld1.8          {d28[], d29[]}, [r2], lr
 340
 341     subs            r12, r12, #1
 342
 343     vst1.8          {q0}, [r5], lr
 344     vst1.8          {q2}, [r6], lr
 345     vst1.8          {q4}, [r5], lr
 346     vst1.8          {q6}, [r6], lr
 347     vst1.8          {q8}, [r5], lr
 348     vst1.8          {q10}, [r6], lr
 349     vst1.8          {q12}, [r5], lr
 350     vst1.8          {q14}, [r6], lr
 351
 352     bne             copy_left_right_y_b16
 353
 354 ;Now copy the top and bottom source lines into each line of the respective borders
 355     ldr             r7, [r0, #yv12_buffer_config_y_buffer]       ;srcptr1
 356     mul             r8, r3, lr
 357
 358     mov             r12, lr, lsr #7
 359
 360     sub             r6, r1, r3              ;destptr2
 361     sub             r2, r6, lr              ;srcptr2
 362     sub             r1, r7, r3              ;srcptr1
 363     sub             r5, r1, r8              ;destptr1
 364
 365 copy_top_bottom_y_b16
 366     vld1.8          {q0, q1}, [r1]!
 367     vld1.8          {q8, q9}, [r2]!
 368     vld1.8          {q2, q3}, [r1]!
 369     vld1.8          {q10, q11}, [r2]!
 370     vld1.8          {q4, q5}, [r1]!
 371     vld1.8          {q12, q13}, [r2]!
 372     vld1.8          {q6, q7}, [r1]!
 373     vld1.8          {q14, q15}, [r2]!
 374
 375     mov             r7, r3
 376
 377 top_bottom_16_b16
 378     subs            r7, r7, #1
 379
 380     vst1.8          {q0, q1}, [r5]!
 381     vst1.8          {q8, q9}, [r6]!
 382     vst1.8          {q2, q3}, [r5]!
 383     vst1.8          {q10, q11}, [r6]!
 384     vst1.8          {q4, q5}, [r5]!
 385     vst1.8          {q12, q13}, [r6]!
 386     vst1.8          {q6, q7}, [r5]!
 387     vst1.8          {q14, q15}, [r6]!
 388
 389     add             r5, r5, lr
 390     sub             r5, r5, #128
 391     add             r6, r6, lr
 392     sub             r6, r6, #128
 393
 394     bne             top_bottom_16_b16
 395
 396     sub             r5, r1, r8
 397     add             r6, r2, lr
 398
 399     subs            r12, r12, #1
 400     bne             copy_top_bottom_y_b16
 401
 402     mov             r7, lr, lsr #4              ;check to see if extra copy is needed
 403     ands            r7, r7, #0x7
 404     bne             extra_top_bottom_y_b16
 405 end_of_border_copy_y_b16
 406
 407 ;Border copy for U, V planes
 408     ldr             r1, [r0, #yv12_buffer_config_u_buffer]       ;srcptr1
 409     mov             lr, lr, lsr #1              ;uv_stride
 410     mov             r3, r3, lsr #1              ;border
 411     mov             r4, r4, lsr #1              ;uv_height
 412     mov             r8, r8, lsr #2
 413
 414     mov             r10, #2
 415
 416 ;copy the left and right most columns out
 417 border_copy_uv_b16
 418     sub             r5, r1, r3              ;destptr1
 419     add             r6, r1, lr
 420     sub             r6, r6, r3, lsl #1      ;destptr2
 421     sub             r2, r6, #1              ;srcptr2
 422
 423     mov             r7, r1
 424
 425     ;Do eight rows at one time
 426     mov             r12, r4, lsr #3
 427
 428 copy_left_right_uv_b16
 429     vld1.8          {d0[]}, [r1], lr
 430     vld1.8          {d2[]}, [r2], lr
 431     vld1.8          {d4[]}, [r1], lr
 432     vld1.8          {d6[]}, [r2], lr
 433     vld1.8          {d8[]},  [r1], lr
 434     vld1.8          {d10[]}, [r2], lr
 435     vld1.8          {d12[]}, [r1], lr
 436     vld1.8          {d14[]}, [r2], lr
 437     vld1.8          {d16[]}, [r1], lr
 438     vld1.8          {d18[]}, [r2], lr
 439     vld1.8          {d20[]}, [r1], lr
 440     vld1.8          {d22[]}, [r2], lr
 441     vld1.8          {d24[]},  [r1], lr
 442     vld1.8          {d26[]}, [r2], lr
 443     vld1.8          {d28[]}, [r1], lr
 444     vld1.8          {d30[]}, [r2], lr
 445
 446     subs            r12, r12, #1
 447
 448     vst1.8          {d0}, [r5], lr
 449     vst1.8          {d2}, [r6], lr
 450     vst1.8          {d4}, [r5], lr
 451     vst1.8          {d6}, [r6], lr
 452     vst1.8          {d8}, [r5], lr
 453     vst1.8          {d10}, [r6], lr
 454     vst1.8          {d12}, [r5], lr
 455     vst1.8          {d14}, [r6], lr
 456     vst1.8          {d16}, [r5], lr
 457     vst1.8          {d18}, [r6], lr
 458     vst1.8          {d20}, [r5], lr
 459     vst1.8          {d22}, [r6], lr
 460     vst1.8          {d24}, [r5], lr
 461     vst1.8          {d26}, [r6], lr
 462     vst1.8          {d28}, [r5], lr
 463     vst1.8          {d30}, [r6], lr
 464
 465     bne             copy_left_right_uv_b16
 466
 467 ;Now copy the top and bottom source lines into each line of the respective borders
 468     mov             r12, lr, lsr #6
 469
 470     sub             r6, r1, r3              ;destptr2
 471     sub             r2, r6, lr              ;srcptr2
 472     sub             r1, r7, r3              ;srcptr1
 473     sub             r5, r1, r8              ;destptr1
 474
 475 copy_top_bottom_uv_b16
 476     vld1.8          {q0, q1}, [r1]!
 477     vld1.8          {q8, q9}, [r2]!
 478     vld1.8          {q2, q3}, [r1]!
 479     vld1.8          {q10, q11}, [r2]!
 480
 481     mov             r7, r3
 482
 483 top_bottom_8_b16
 484     subs            r7, r7, #1
 485
 486     vst1.8          {q0, q1}, [r5]!
 487     vst1.8          {q8, q9}, [r6]!
 488     vst1.8          {q2, q3}, [r5]!
 489     vst1.8          {q10, q11}, [r6]!
 490
 491     add             r5, r5, lr
 492     sub             r5, r5, #64
 493     add             r6, r6, lr
 494     sub             r6, r6, #64
 495
 496     bne             top_bottom_8_b16
 497
 498     sub             r5, r1, r8
 499     add             r6, r2, lr
 500
 501     subs            r12, r12, #1
 502     bne             copy_top_bottom_uv_b16
 503
 504     mov             r7, lr, lsr #3              ;check to see if extra copy is needed
 505     ands            r7, r7, #0x7
 506     bne             extra_top_bottom_uv_b16
 507
 508 end_of_border_copy_uv_b16
 509     subs            r10, r10, #1
 510     ldrne           r1, [r0, #yv12_buffer_config_v_buffer]       ;srcptr1
 511     bne             border_copy_uv_b16
 512
 513     vpop            {d8-d15}
 514     pop             {r4 - r10, pc}
 515
 516 ;;;;;;;;;;;;;;;;;;;;;;
 517 ;extra copy part for Y
 518 extra_top_bottom_y_b16
 519     vld1.8          {q0}, [r1]!
 520     vld1.8          {q2}, [r2]!
 521
 522     mov             r9, r3, lsr #3
 523
 524 extra_top_bottom_16_b16
 525     subs            r9, r9, #1
 526
 527     vst1.8          {q0}, [r5], lr
 528     vst1.8          {q2}, [r6], lr
 529     vst1.8          {q0}, [r5], lr
 530     vst1.8          {q2}, [r6], lr
 531     vst1.8          {q0}, [r5], lr
 532     vst1.8          {q2}, [r6], lr
 533     vst1.8          {q0}, [r5], lr
 534     vst1.8          {q2}, [r6], lr
 535     vst1.8          {q0}, [r5], lr
 536     vst1.8          {q2}, [r6], lr
 537     vst1.8          {q0}, [r5], lr
 538     vst1.8          {q2}, [r6], lr
 539     vst1.8          {q0}, [r5], lr
 540     vst1.8          {q2}, [r6], lr
 541     vst1.8          {q0}, [r5], lr
 542     vst1.8          {q2}, [r6], lr
 543     bne             extra_top_bottom_16_b16
 544
 545     sub             r5, r1, r8
 546     add             r6, r2, lr
 547     subs            r7, r7, #1
 548     bne             extra_top_bottom_y_b16
 549
 550     b               end_of_border_copy_y_b16
 551
 552 ;extra copy part for UV
 553 extra_top_bottom_uv_b16
 554     vld1.8          {d0}, [r1]!
 555     vld1.8          {d8}, [r2]!
 556
 557     mov             r9, r3, lsr #3
 558
 559 extra_top_bottom_8_b16
 560     subs            r9, r9, #1
 561
 562     vst1.8          {d0}, [r5], lr
 563     vst1.8          {d8}, [r6], lr
 564     vst1.8          {d0}, [r5], lr
 565     vst1.8          {d8}, [r6], lr
 566     vst1.8          {d0}, [r5], lr
 567     vst1.8          {d8}, [r6], lr
 568     vst1.8          {d0}, [r5], lr
 569     vst1.8          {d8}, [r6], lr
 570     vst1.8          {d0}, [r5], lr
 571     vst1.8          {d8}, [r6], lr
 572     vst1.8          {d0}, [r5], lr
 573     vst1.8          {d8}, [r6], lr
 574     vst1.8          {d0}, [r5], lr
 575     vst1.8          {d8}, [r6], lr
 576     vst1.8          {d0}, [r5], lr
 577     vst1.8          {d8}, [r6], lr
 578     bne             extra_top_bottom_8_b16
 579
 580     sub             r5, r1, r8
 581     add             r6, r2, lr
 582     subs            r7, r7, #1
 583     bne             extra_top_bottom_uv_b16
 584
 585     b               end_of_border_copy_uv_b16
 586
 587     ENDP
 588     END