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 |vp8cx_pack_tokens_into_partitions_armv7| 13 14 INCLUDE vpx_vp8_enc_asm_offsets.asm 15 16 ARM 17 REQUIRE8 18 PRESERVE8 19 20 AREA |.text|, CODE, READONLY 21 22 ; r0 VP8_COMP *cpi 23 ; r1 unsigned char *cx_data 24 ; r2 int num_part 25 ; r3 *size 26 ; s0 vp8_coef_encodings 27 ; s1 vp8_extra_bits, 28 ; s2 const vp8_tree_index *, 29 30 |vp8cx_pack_tokens_into_partitions_armv7| PROC 31 push {r4-r11, lr} 32 sub sp, sp, #44 33 34 ; Compute address of cpi->common.mb_rows 35 ldr r4, _VP8_COMP_common_ 36 ldr r6, _VP8_COMMON_MBrows_ 37 add r4, r0, r4 38 39 ldr r5, [r4, r6] ; load up mb_rows 40 41 str r5, [sp, #36] ; save mb_rows 42 str r1, [sp, #24] ; save cx_data 43 str r2, [sp, #20] ; save num_part 44 str r3, [sp, #8] ; save *size 45 46 ; *size = 3*(num_part -1 ); 47 sub r2, r2, #1 ; num_part - 1 48 add r2, r2, r2, lsl #1 ; 3*(num_part - 1) 49 str r2, [r3] 50 51 add r2, r2, r1 ; cx_data + *size 52 str r2, [sp, #40] ; ptr 53 54 ldr r4, _VP8_COMP_tplist_ 55 add r4, r0, r4 56 ldr r7, [r4, #0] ; dereference cpi->tp_list 57 str r7, [sp, #32] ; store start of cpi->tp_list 58 59 ldr r11, _VP8_COMP_bc2_ ; load up vp8_writer out of cpi 60 add r0, r0, r11 61 62 mov r11, #0 63 str r11, [sp, #28] ; i 64 65 numparts_loop 66 ldr r10, [sp, #40] ; ptr 67 ldr r5, [sp, #36] ; move mb_rows to the counting section 68 str r5, [sp, #12] 69 70 ; Reset all of the VP8 Writer data for each partition that 71 ; is processed. 72 ; start_encode 73 mov r2, #0 ; vp8_writer_lowvalue 74 mov r5, #255 ; vp8_writer_range 75 mvn r3, #23 ; vp8_writer_count 76 77 str r2, [r0, #vp8_writer_value] 78 str r2, [r0, #vp8_writer_pos] 79 str r10, [r0, #vp8_writer_buffer] 80 81 mb_row_loop 82 83 ldr r1, [r7, #tokenlist_start] 84 ldr r9, [r7, #tokenlist_stop] 85 str r9, [sp, #0] ; save stop for later comparison 86 str r7, [sp, #16] ; tokenlist address for next time 87 88 b check_p_lt_stop 89 90 ; actual work gets done here! 91 92 while_p_lt_stop 93 ldr r6, [r1, #tokenextra_token] ; t 94 ldr r4, [sp, #80] ; vp8_coef_encodings 95 mov lr, #0 96 add r4, r4, r6, lsl #3 ; a = vp8_coef_encodings + t 97 ldr r9, [r1, #tokenextra_context_tree] ; pp 98 99 ldr r7, [r1, #tokenextra_skip_eob_node] 100 101 ldr r6, [r4, #vp8_token_value] ; v 102 ldr r8, [r4, #vp8_token_len] ; n 103 104 ; vp8 specific skip_eob_node 105 cmp r7, #0 106 movne lr, #2 ; i = 2 107 subne r8, r8, #1 ; --n 108 109 ; reverse the stream of bits to be packed. Normally 110 ; the most significant bit is peeled off and compared 111 ; in the form of (v >> --n) & 1. ARM architecture has 112 ; the ability to set a flag based on the value of the 113 ; bit shifted off the bottom of the register. To make 114 ; that happen the bitstream is reversed. 115 rbit r12, r6 116 rsb r4, r8, #32 ; 32-n 117 ldr r10, [sp, #88] ; vp8_coef_tree 118 119 ; v is kept in r12 during the token pack loop 120 lsr r12, r12, r4 ; v >>= 32 - n 121 122 ; loop start 123 token_loop 124 ldrb r4, [r9, lr, asr #1] ; pp [i>>1] 125 sub r7, r5, #1 ; range-1 126 127 ; Decisions are made based on the bit value shifted 128 ; off of v, so set a flag here based on this. 129 ; This value is refered to as "bb" 130 lsrs r12, r12, #1 ; bb = v >> n 131 mul r4, r4, r7 ; ((range-1) * pp[i>>1])) 132 133 ; bb can only be 0 or 1. So only execute this statement 134 ; if bb == 1, otherwise it will act like i + 0 135 addcs lr, lr, #1 ; i + bb 136 137 mov r7, #1 138 ldrsb lr, [r10, lr] ; i = vp8_coef_tree[i+bb] 139 add r4, r7, r4, lsr #8 ; 1 + (((range-1) * pp[i>>1]) >> 8) 140 141 addcs r2, r2, r4 ; if (bb) lowvalue += split 142 subcs r4, r5, r4 ; if (bb) range = range-split 143 144 ; Counting the leading zeros is used to normalize range. 145 clz r6, r4 146 sub r6, r6, #24 ; shift 147 148 ; Flag is set on the sum of count. This flag is used later 149 ; to determine if count >= 0 150 adds r3, r3, r6 ; count += shift 151 lsl r5, r4, r6 ; range <<= shift 152 bmi token_count_lt_zero ; if(count >= 0) 153 154 sub r6, r6, r3 ; offset = shift - count 155 sub r4, r6, #1 ; offset-1 156 lsls r4, r2, r4 ; if((lowvalue<<(offset-1)) & 0x80000000 ) 157 bpl token_high_bit_not_set 158 159 ldr r4, [r0, #vp8_writer_pos] ; x 160 sub r4, r4, #1 ; x = w->pos-1 161 b token_zero_while_start 162 token_zero_while_loop 163 mov r10, #0 164 strb r10, [r7, r4] ; w->buffer[x] =(unsigned char)0 165 sub r4, r4, #1 ; x-- 166 token_zero_while_start 167 cmp r4, #0 168 ldrge r7, [r0, #vp8_writer_buffer] 169 ldrb r11, [r7, r4] 170 cmpge r11, #0xff 171 beq token_zero_while_loop 172 173 ldr r7, [r0, #vp8_writer_buffer] 174 ldrb r10, [r7, r4] ; w->buffer[x] 175 add r10, r10, #1 176 strb r10, [r7, r4] ; w->buffer[x] + 1 177 token_high_bit_not_set 178 rsb r4, r6, #24 ; 24-offset 179 ldr r10, [r0, #vp8_writer_buffer] 180 lsr r7, r2, r4 ; lowvalue >> (24-offset) 181 ldr r4, [r0, #vp8_writer_pos] ; w->pos 182 lsl r2, r2, r6 ; lowvalue <<= offset 183 mov r6, r3 ; shift = count 184 add r11, r4, #1 ; w->pos++ 185 bic r2, r2, #0xff000000 ; lowvalue &= 0xffffff 186 str r11, [r0, #vp8_writer_pos] 187 sub r3, r3, #8 ; count -= 8 188 strb r7, [r10, r4] ; w->buffer[w->pos++] 189 190 ; r10 is used earlier in the loop, but r10 is used as 191 ; temp variable here. So after r10 is used, reload 192 ; vp8_coef_tree_dcd into r10 193 ldr r10, [sp, #88] ; vp8_coef_tree 194 195 token_count_lt_zero 196 lsl r2, r2, r6 ; lowvalue <<= shift 197 198 subs r8, r8, #1 ; --n 199 bne token_loop 200 201 ldr r6, [r1, #tokenextra_token] ; t 202 ldr r7, [sp, #84] ; vp8_extra_bits 203 ; Add t * sizeof (vp8_extra_bit_struct) to get the desired 204 ; element. Here vp8_extra_bit_struct == 20 205 add r6, r6, r6, lsl #2 ; b = vp8_extra_bits + t 206 add r12, r7, r6, lsl #2 ; b = vp8_extra_bits + t 207 208 ldr r4, [r12, #vp8_extra_bit_struct_base_val] 209 cmp r4, #0 210 beq skip_extra_bits 211 212 ; if( b->base_val) 213 ldr r8, [r12, #vp8_extra_bit_struct_len] ; L 214 ldr lr, [r1, #tokenextra_extra] ; e = p->Extra 215 cmp r8, #0 ; if( L) 216 beq no_extra_bits 217 218 ldr r9, [r12, #vp8_extra_bit_struct_prob] 219 asr r7, lr, #1 ; v=e>>1 220 221 ldr r10, [r12, #vp8_extra_bit_struct_tree] 222 str r10, [sp, #4] ; b->tree 223 224 rbit r12, r7 ; reverse v 225 rsb r4, r8, #32 226 lsr r12, r12, r4 227 228 mov lr, #0 ; i = 0 229 230 extra_bits_loop 231 ldrb r4, [r9, lr, asr #1] ; pp[i>>1] 232 sub r7, r5, #1 ; range-1 233 lsrs r12, r12, #1 ; v >> n 234 mul r4, r4, r7 ; (range-1) * pp[i>>1] 235 addcs lr, lr, #1 ; i + bb 236 237 mov r7, #1 238 ldrsb lr, [r10, lr] ; i = b->tree[i+bb] 239 add r4, r7, r4, lsr #8 ; split = 1 + (((range-1) * pp[i>>1]) >> 8) 240 241 addcs r2, r2, r4 ; if (bb) lowvalue += split 242 subcs r4, r5, r4 ; if (bb) range = range-split 243 244 clz r6, r4 245 sub r6, r6, #24 246 247 adds r3, r3, r6 ; count += shift 248 lsl r5, r4, r6 ; range <<= shift 249 bmi extra_count_lt_zero ; if(count >= 0) 250 251 sub r6, r6, r3 ; offset= shift - count 252 sub r4, r6, #1 ; offset-1 253 lsls r4, r2, r4 ; if((lowvalue<<(offset-1)) & 0x80000000 ) 254 bpl extra_high_bit_not_set 255 256 ldr r4, [r0, #vp8_writer_pos] ; x 257 sub r4, r4, #1 ; x = w->pos - 1 258 b extra_zero_while_start 259 extra_zero_while_loop 260 mov r10, #0 261 strb r10, [r7, r4] ; w->buffer[x] =(unsigned char)0 262 sub r4, r4, #1 ; x-- 263 extra_zero_while_start 264 cmp r4, #0 265 ldrge r7, [r0, #vp8_writer_buffer] 266 ldrb r11, [r7, r4] 267 cmpge r11, #0xff 268 beq extra_zero_while_loop 269 270 ldr r7, [r0, #vp8_writer_buffer] 271 ldrb r10, [r7, r4] 272 add r10, r10, #1 273 strb r10, [r7, r4] 274 extra_high_bit_not_set 275 rsb r4, r6, #24 ; 24-offset 276 ldr r10, [r0, #vp8_writer_buffer] 277 lsr r7, r2, r4 ; lowvalue >> (24-offset) 278 ldr r4, [r0, #vp8_writer_pos] 279 lsl r2, r2, r6 ; lowvalue <<= offset 280 mov r6, r3 ; shift = count 281 add r11, r4, #1 ; w->pos++ 282 bic r2, r2, #0xff000000 ; lowvalue &= 0xffffff 283 str r11, [r0, #vp8_writer_pos] 284 sub r3, r3, #8 ; count -= 8 285 strb r7, [r10, r4] ; w->buffer[w->pos++]=(lowvalue >> (24-offset)) 286 ldr r10, [sp, #4] ; b->tree 287 extra_count_lt_zero 288 lsl r2, r2, r6 289 290 subs r8, r8, #1 ; --n 291 bne extra_bits_loop ; while (n) 292 293 no_extra_bits 294 ldr lr, [r1, #4] ; e = p->Extra 295 add r4, r5, #1 ; range + 1 296 tst lr, #1 297 lsr r4, r4, #1 ; split = (range + 1) >> 1 298 addne r2, r2, r4 ; lowvalue += split 299 subne r4, r5, r4 ; range = range-split 300 tst r2, #0x80000000 ; lowvalue & 0x80000000 301 lsl r5, r4, #1 ; range <<= 1 302 beq end_high_bit_not_set 303 304 ldr r4, [r0, #vp8_writer_pos] 305 mov r7, #0 306 sub r4, r4, #1 307 b end_zero_while_start 308 end_zero_while_loop 309 strb r7, [r6, r4] 310 sub r4, r4, #1 ; x-- 311 end_zero_while_start 312 cmp r4, #0 313 ldrge r6, [r0, #vp8_writer_buffer] 314 ldrb r12, [r6, r4] 315 cmpge r12, #0xff 316 beq end_zero_while_loop 317 318 ldr r6, [r0, #vp8_writer_buffer] 319 ldrb r7, [r6, r4] 320 add r7, r7, #1 321 strb r7, [r6, r4] 322 end_high_bit_not_set 323 adds r3, r3, #1 ; ++count 324 lsl r2, r2, #1 ; lowvalue <<= 1 325 bne end_count_zero 326 327 ldr r4, [r0, #vp8_writer_pos] 328 mvn r3, #7 329 ldr r7, [r0, #vp8_writer_buffer] 330 lsr r6, r2, #24 ; lowvalue >> 24 331 add r12, r4, #1 ; w->pos++ 332 bic r2, r2, #0xff000000 ; lowvalue &= 0xffffff 333 str r12, [r0, #0x10] 334 strb r6, [r7, r4] 335 end_count_zero 336 skip_extra_bits 337 add r1, r1, #TOKENEXTRA_SZ ; ++p 338 check_p_lt_stop 339 ldr r4, [sp, #0] ; stop 340 cmp r1, r4 ; while( p < stop) 341 bcc while_p_lt_stop 342 343 ldr r10, [sp, #20] ; num_parts 344 mov r1, #TOKENLIST_SZ 345 mul r1, r10, r1 346 347 ldr r6, [sp, #12] ; mb_rows 348 ldr r7, [sp, #16] ; tokenlist address 349 subs r6, r6, r10 350 add r7, r7, r1 ; next element in the array 351 str r6, [sp, #12] 352 bgt mb_row_loop 353 354 mov r12, #32 355 356 stop_encode_loop 357 sub r7, r5, #1 ; range-1 358 359 mov r4, r7, lsl #7 ; ((range-1) * 128) 360 361 mov r7, #1 362 add r4, r7, r4, lsr #8 ; 1 + (((range-1) * 128) >> 8) 363 364 ; Counting the leading zeros is used to normalize range. 365 clz r6, r4 366 sub r6, r6, #24 ; shift 367 368 ; Flag is set on the sum of count. This flag is used later 369 ; to determine if count >= 0 370 adds r3, r3, r6 ; count += shift 371 lsl r5, r4, r6 ; range <<= shift 372 bmi token_count_lt_zero_se ; if(count >= 0) 373 374 sub r6, r6, r3 ; offset = shift - count 375 sub r4, r6, #1 ; offset-1 376 lsls r4, r2, r4 ; if((lowvalue<<(offset-1)) & 0x80000000 ) 377 bpl token_high_bit_not_set_se 378 379 ldr r4, [r0, #vp8_writer_pos] ; x 380 sub r4, r4, #1 ; x = w->pos-1 381 b token_zero_while_start_se 382 token_zero_while_loop_se 383 mov r10, #0 384 strb r10, [r7, r4] ; w->buffer[x] =(unsigned char)0 385 sub r4, r4, #1 ; x-- 386 token_zero_while_start_se 387 cmp r4, #0 388 ldrge r7, [r0, #vp8_writer_buffer] 389 ldrb r11, [r7, r4] 390 cmpge r11, #0xff 391 beq token_zero_while_loop_se 392 393 ldr r7, [r0, #vp8_writer_buffer] 394 ldrb r10, [r7, r4] ; w->buffer[x] 395 add r10, r10, #1 396 strb r10, [r7, r4] ; w->buffer[x] + 1 397 token_high_bit_not_set_se 398 rsb r4, r6, #24 ; 24-offset 399 ldr r10, [r0, #vp8_writer_buffer] 400 lsr r7, r2, r4 ; lowvalue >> (24-offset) 401 ldr r4, [r0, #vp8_writer_pos] ; w->pos 402 lsl r2, r2, r6 ; lowvalue <<= offset 403 mov r6, r3 ; shift = count 404 add r11, r4, #1 ; w->pos++ 405 bic r2, r2, #0xff000000 ; lowvalue &= 0xffffff 406 str r11, [r0, #vp8_writer_pos] 407 sub r3, r3, #8 ; count -= 8 408 strb r7, [r10, r4] ; w->buffer[w->pos++] 409 410 token_count_lt_zero_se 411 lsl r2, r2, r6 ; lowvalue <<= shift 412 413 subs r12, r12, #1 414 bne stop_encode_loop 415 416 ldr r10, [sp, #8] ; *size 417 ldr r11, [r10] 418 ldr r4, [r0, #vp8_writer_pos] ; w->pos 419 add r11, r11, r4 ; *size += w->pos 420 str r11, [r10] 421 422 ldr r9, [sp, #20] ; num_parts 423 sub r9, r9, #1 424 ldr r10, [sp, #28] ; i 425 cmp r10, r9 ; if(i<(num_part - 1)) 426 bge skip_write_partition 427 428 ldr r12, [sp, #40] ; ptr 429 add r12, r12, r4 ; ptr += w->pos 430 str r12, [sp, #40] 431 432 ldr r9, [sp, #24] ; cx_data 433 mov r8, r4, asr #8 434 strb r4, [r9, #0] 435 strb r8, [r9, #1] 436 mov r4, r4, asr #16 437 strb r4, [r9, #2] 438 439 add r9, r9, #3 ; cx_data += 3 440 str r9, [sp, #24] 441 442 skip_write_partition 443 444 ldr r11, [sp, #28] ; i 445 ldr r10, [sp, #20] ; num_parts 446 447 add r11, r11, #1 ; i++ 448 str r11, [sp, #28] 449 450 ldr r7, [sp, #32] ; cpi->tp_list[i] 451 mov r1, #TOKENLIST_SZ 452 add r7, r7, r1 ; next element in cpi->tp_list 453 str r7, [sp, #32] ; cpi->tp_list[i+1] 454 455 cmp r10, r11 456 bgt numparts_loop 457 458 459 add sp, sp, #44 460 pop {r4-r11, pc} 461 ENDP 462 463 _VP8_COMP_common_ 464 DCD vp8_comp_common 465 _VP8_COMMON_MBrows_ 466 DCD vp8_common_mb_rows 467 _VP8_COMP_tplist_ 468 DCD vp8_comp_tplist 469 _VP8_COMP_bc2_ 470 DCD vp8_comp_bc2 471 472 END 473
