1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libvpx/vp8/encoder/arm/armv5te/vp8_packtokens_partitions_armv5.asm Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,471 @@
1.4 +;
1.5 +; Copyright (c) 2010 The WebM project authors. All Rights Reserved.
1.6 +;
1.7 +; Use of this source code is governed by a BSD-style license
1.8 +; that can be found in the LICENSE file in the root of the source
1.9 +; tree. An additional intellectual property rights grant can be found
1.10 +; in the file PATENTS. All contributing project authors may
1.11 +; be found in the AUTHORS file in the root of the source tree.
1.12 +;
1.13 +
1.14 +
1.15 + EXPORT |vp8cx_pack_tokens_into_partitions_armv5|
1.16 + IMPORT |vp8_validate_buffer_arm|
1.17 +
1.18 + INCLUDE vp8_asm_enc_offsets.asm
1.19 +
1.20 + ARM
1.21 + REQUIRE8
1.22 + PRESERVE8
1.23 +
1.24 + AREA |.text|, CODE, READONLY
1.25 +
1.26 + ; macro for validating write buffer position
1.27 + ; needs vp8_writer in r0
1.28 + ; start shall not be in r1
1.29 + MACRO
1.30 + VALIDATE_POS $start, $pos
1.31 + push {r0-r3, r12, lr} ; rest of regs are preserved by subroutine call
1.32 + ldr r2, [r0, #vp8_writer_buffer_end]
1.33 + ldr r3, [r0, #vp8_writer_error]
1.34 + mov r1, $pos
1.35 + mov r0, $start
1.36 + bl vp8_validate_buffer_arm
1.37 + pop {r0-r3, r12, lr}
1.38 + MEND
1.39 +
1.40 +; r0 VP8_COMP *cpi
1.41 +; r1 unsigned char *cx_data
1.42 +; r2 const unsigned char *cx_data_end
1.43 +; r3 int num_part
1.44 +; s0 vp8_coef_encodings
1.45 +; s1 vp8_extra_bits,
1.46 +; s2 const vp8_tree_index *
1.47 +
1.48 +|vp8cx_pack_tokens_into_partitions_armv5| PROC
1.49 + push {r4-r12, lr}
1.50 + sub sp, sp, #40
1.51 +
1.52 + ; Compute address of cpi->common.mb_rows
1.53 + ldr r4, _VP8_COMP_common_
1.54 + ldr r6, _VP8_COMMON_MBrows_
1.55 + add r4, r0, r4
1.56 +
1.57 + ldr r5, [r4, r6] ; load up mb_rows
1.58 +
1.59 + str r5, [sp, #36] ; save mb_rows
1.60 + str r1, [sp, #24] ; save ptr = cx_data
1.61 + str r3, [sp, #20] ; save num_part
1.62 + str r2, [sp, #8] ; save cx_data_end
1.63 +
1.64 + ldr r4, _VP8_COMP_tplist_
1.65 + add r4, r0, r4
1.66 + ldr r7, [r4, #0] ; dereference cpi->tp_list
1.67 + str r7, [sp, #32] ; store start of cpi->tp_list
1.68 +
1.69 + ldr r11, _VP8_COMP_bc_ ; load up vp8_writer out of cpi
1.70 + add r0, r0, r11
1.71 +
1.72 + mov r11, #0
1.73 + str r11, [sp, #28] ; i
1.74 +
1.75 +numparts_loop
1.76 + ldr r2, _vp8_writer_sz_ ; load up sizeof(vp8_writer)
1.77 + add r0, r2 ; bc[i + 1]
1.78 +
1.79 + ldr r10, [sp, #24] ; ptr
1.80 + ldr r5, [sp, #36] ; move mb_rows to the counting section
1.81 + subs r5, r5, r11 ; move start point with each partition
1.82 + ; mb_rows starts at i
1.83 + str r5, [sp, #12]
1.84 +
1.85 + ; Reset all of the VP8 Writer data for each partition that
1.86 + ; is processed.
1.87 + ; start_encode
1.88 +
1.89 + ldr r3, [sp, #8]
1.90 + str r3, [r0, #vp8_writer_buffer_end]
1.91 +
1.92 + mov r2, #0 ; vp8_writer_lowvalue
1.93 + mov r5, #255 ; vp8_writer_range
1.94 + mvn r3, #23 ; vp8_writer_count
1.95 +
1.96 + str r2, [r0, #vp8_writer_pos]
1.97 + str r10, [r0, #vp8_writer_buffer]
1.98 +
1.99 + ble end_partition ; if (mb_rows <= 0) end partition
1.100 +
1.101 +mb_row_loop
1.102 +
1.103 + ldr r1, [r7, #tokenlist_start]
1.104 + ldr r9, [r7, #tokenlist_stop]
1.105 + str r9, [sp, #0] ; save stop for later comparison
1.106 + str r7, [sp, #16] ; tokenlist address for next time
1.107 +
1.108 + b check_p_lt_stop
1.109 +
1.110 + ; actual work gets done here!
1.111 +
1.112 +while_p_lt_stop
1.113 + ldrb r6, [r1, #tokenextra_token] ; t
1.114 + ldr r4, [sp, #80] ; vp8_coef_encodings
1.115 + mov lr, #0
1.116 + add r4, r4, r6, lsl #3 ; a = vp8_coef_encodings + t
1.117 + ldr r9, [r1, #tokenextra_context_tree] ; pp
1.118 +
1.119 + ldrb r7, [r1, #tokenextra_skip_eob_node]
1.120 +
1.121 + ldr r6, [r4, #vp8_token_value] ; v
1.122 + ldr r8, [r4, #vp8_token_len] ; n
1.123 +
1.124 + ; vp8 specific skip_eob_node
1.125 + cmp r7, #0
1.126 + movne lr, #2 ; i = 2
1.127 + subne r8, r8, #1 ; --n
1.128 +
1.129 + rsb r4, r8, #32 ; 32-n
1.130 + ldr r10, [sp, #88] ; vp8_coef_tree
1.131 +
1.132 + ; v is kept in r12 during the token pack loop
1.133 + lsl r12, r6, r4 ; r12 = v << 32 - n
1.134 +
1.135 +; loop start
1.136 +token_loop
1.137 + ldrb r4, [r9, lr, asr #1] ; pp [i>>1]
1.138 + sub r7, r5, #1 ; range-1
1.139 +
1.140 + ; Decisions are made based on the bit value shifted
1.141 + ; off of v, so set a flag here based on this.
1.142 + ; This value is refered to as "bb"
1.143 + lsls r12, r12, #1 ; bb = v >> n
1.144 + mul r6, r4, r7 ; ((range-1) * pp[i>>1]))
1.145 +
1.146 + ; bb can only be 0 or 1. So only execute this statement
1.147 + ; if bb == 1, otherwise it will act like i + 0
1.148 + addcs lr, lr, #1 ; i + bb
1.149 +
1.150 + mov r7, #1
1.151 + ldrsb lr, [r10, lr] ; i = vp8_coef_tree[i+bb]
1.152 + add r4, r7, r6, lsr #8 ; 1 + (((range-1) * pp[i>>1]) >> 8)
1.153 +
1.154 + addcs r2, r2, r4 ; if (bb) lowvalue += split
1.155 + subcs r4, r5, r4 ; if (bb) range = range-split
1.156 +
1.157 + ; Counting the leading zeros is used to normalize range.
1.158 + clz r6, r4
1.159 + sub r6, r6, #24 ; shift
1.160 +
1.161 + ; Flag is set on the sum of count. This flag is used later
1.162 + ; to determine if count >= 0
1.163 + adds r3, r3, r6 ; count += shift
1.164 + lsl r5, r4, r6 ; range <<= shift
1.165 + bmi token_count_lt_zero ; if(count >= 0)
1.166 +
1.167 + sub r6, r6, r3 ; offset = shift - count
1.168 + sub r4, r6, #1 ; offset-1
1.169 + lsls r4, r2, r4 ; if((lowvalue<<(offset-1)) & 0x80000000 )
1.170 + bpl token_high_bit_not_set
1.171 +
1.172 + ldr r4, [r0, #vp8_writer_pos] ; x
1.173 + sub r4, r4, #1 ; x = w->pos-1
1.174 + b token_zero_while_start
1.175 +token_zero_while_loop
1.176 + mov r10, #0
1.177 + strb r10, [r7, r4] ; w->buffer[x] =(unsigned char)0
1.178 + sub r4, r4, #1 ; x--
1.179 +token_zero_while_start
1.180 + cmp r4, #0
1.181 + ldrge r7, [r0, #vp8_writer_buffer]
1.182 + ldrb r11, [r7, r4]
1.183 + cmpge r11, #0xff
1.184 + beq token_zero_while_loop
1.185 +
1.186 + ldr r7, [r0, #vp8_writer_buffer]
1.187 + ldrb r10, [r7, r4] ; w->buffer[x]
1.188 + add r10, r10, #1
1.189 + strb r10, [r7, r4] ; w->buffer[x] + 1
1.190 +token_high_bit_not_set
1.191 + rsb r4, r6, #24 ; 24-offset
1.192 + ldr r10, [r0, #vp8_writer_buffer]
1.193 + lsr r7, r2, r4 ; lowvalue >> (24-offset)
1.194 + ldr r4, [r0, #vp8_writer_pos] ; w->pos
1.195 + lsl r2, r2, r6 ; lowvalue <<= offset
1.196 + mov r6, r3 ; shift = count
1.197 + add r11, r4, #1 ; w->pos++
1.198 + bic r2, r2, #0xff000000 ; lowvalue &= 0xffffff
1.199 + str r11, [r0, #vp8_writer_pos]
1.200 + sub r3, r3, #8 ; count -= 8
1.201 +
1.202 + VALIDATE_POS r10, r11 ; validate_buffer at pos
1.203 +
1.204 + strb r7, [r10, r4] ; w->buffer[w->pos++]
1.205 +
1.206 + ; r10 is used earlier in the loop, but r10 is used as
1.207 + ; temp variable here. So after r10 is used, reload
1.208 + ; vp8_coef_tree_dcd into r10
1.209 + ldr r10, [sp, #88] ; vp8_coef_tree
1.210 +
1.211 +token_count_lt_zero
1.212 + lsl r2, r2, r6 ; lowvalue <<= shift
1.213 +
1.214 + subs r8, r8, #1 ; --n
1.215 + bne token_loop
1.216 +
1.217 + ldrb r6, [r1, #tokenextra_token] ; t
1.218 + ldr r7, [sp, #84] ; vp8_extra_bits
1.219 + ; Add t * sizeof (vp8_extra_bit_struct) to get the desired
1.220 + ; element. Here vp8_extra_bit_struct == 16
1.221 + add r12, r7, r6, lsl #4 ; b = vp8_extra_bits + t
1.222 +
1.223 + ldr r4, [r12, #vp8_extra_bit_struct_base_val]
1.224 + cmp r4, #0
1.225 + beq skip_extra_bits
1.226 +
1.227 +; if( b->base_val)
1.228 + ldr r8, [r12, #vp8_extra_bit_struct_len] ; L
1.229 + ldrsh lr, [r1, #tokenextra_extra] ; e = p->Extra
1.230 + cmp r8, #0 ; if( L)
1.231 + beq no_extra_bits
1.232 +
1.233 + ldr r9, [r12, #vp8_extra_bit_struct_prob]
1.234 + asr r7, lr, #1 ; v=e>>1
1.235 +
1.236 + ldr r10, [r12, #vp8_extra_bit_struct_tree]
1.237 + str r10, [sp, #4] ; b->tree
1.238 +
1.239 + rsb r4, r8, #32
1.240 + lsl r12, r7, r4
1.241 +
1.242 + mov lr, #0 ; i = 0
1.243 +
1.244 +extra_bits_loop
1.245 + ldrb r4, [r9, lr, asr #1] ; pp[i>>1]
1.246 + sub r7, r5, #1 ; range-1
1.247 + lsls r12, r12, #1 ; v >> n
1.248 + mul r6, r4, r7 ; (range-1) * pp[i>>1]
1.249 + addcs lr, lr, #1 ; i + bb
1.250 +
1.251 + mov r7, #1
1.252 + ldrsb lr, [r10, lr] ; i = b->tree[i+bb]
1.253 + add r4, r7, r6, lsr #8 ; split = 1 + (((range-1) * pp[i>>1]) >> 8)
1.254 +
1.255 + addcs r2, r2, r4 ; if (bb) lowvalue += split
1.256 + subcs r4, r5, r4 ; if (bb) range = range-split
1.257 +
1.258 + clz r6, r4
1.259 + sub r6, r6, #24
1.260 +
1.261 + adds r3, r3, r6 ; count += shift
1.262 + lsl r5, r4, r6 ; range <<= shift
1.263 + bmi extra_count_lt_zero ; if(count >= 0)
1.264 +
1.265 + sub r6, r6, r3 ; offset= shift - count
1.266 + sub r4, r6, #1 ; offset-1
1.267 + lsls r4, r2, r4 ; if((lowvalue<<(offset-1)) & 0x80000000 )
1.268 + bpl extra_high_bit_not_set
1.269 +
1.270 + ldr r4, [r0, #vp8_writer_pos] ; x
1.271 + sub r4, r4, #1 ; x = w->pos - 1
1.272 + b extra_zero_while_start
1.273 +extra_zero_while_loop
1.274 + mov r10, #0
1.275 + strb r10, [r7, r4] ; w->buffer[x] =(unsigned char)0
1.276 + sub r4, r4, #1 ; x--
1.277 +extra_zero_while_start
1.278 + cmp r4, #0
1.279 + ldrge r7, [r0, #vp8_writer_buffer]
1.280 + ldrb r11, [r7, r4]
1.281 + cmpge r11, #0xff
1.282 + beq extra_zero_while_loop
1.283 +
1.284 + ldr r7, [r0, #vp8_writer_buffer]
1.285 + ldrb r10, [r7, r4]
1.286 + add r10, r10, #1
1.287 + strb r10, [r7, r4]
1.288 +extra_high_bit_not_set
1.289 + rsb r4, r6, #24 ; 24-offset
1.290 + ldr r10, [r0, #vp8_writer_buffer]
1.291 + lsr r7, r2, r4 ; lowvalue >> (24-offset)
1.292 + ldr r4, [r0, #vp8_writer_pos]
1.293 + lsl r2, r2, r6 ; lowvalue <<= offset
1.294 + mov r6, r3 ; shift = count
1.295 + add r11, r4, #1 ; w->pos++
1.296 + bic r2, r2, #0xff000000 ; lowvalue &= 0xffffff
1.297 + str r11, [r0, #vp8_writer_pos]
1.298 + sub r3, r3, #8 ; count -= 8
1.299 +
1.300 + VALIDATE_POS r10, r11 ; validate_buffer at pos
1.301 +
1.302 + strb r7, [r10, r4] ; w->buffer[w->pos++]=(lowvalue >> (24-offset))
1.303 + ldr r10, [sp, #4] ; b->tree
1.304 +extra_count_lt_zero
1.305 + lsl r2, r2, r6
1.306 +
1.307 + subs r8, r8, #1 ; --n
1.308 + bne extra_bits_loop ; while (n)
1.309 +
1.310 +no_extra_bits
1.311 + ldr lr, [r1, #4] ; e = p->Extra
1.312 + add r4, r5, #1 ; range + 1
1.313 + tst lr, #1
1.314 + lsr r4, r4, #1 ; split = (range + 1) >> 1
1.315 + addne r2, r2, r4 ; lowvalue += split
1.316 + subne r4, r5, r4 ; range = range-split
1.317 + tst r2, #0x80000000 ; lowvalue & 0x80000000
1.318 + lsl r5, r4, #1 ; range <<= 1
1.319 + beq end_high_bit_not_set
1.320 +
1.321 + ldr r4, [r0, #vp8_writer_pos]
1.322 + mov r7, #0
1.323 + sub r4, r4, #1
1.324 + b end_zero_while_start
1.325 +end_zero_while_loop
1.326 + strb r7, [r6, r4]
1.327 + sub r4, r4, #1 ; x--
1.328 +end_zero_while_start
1.329 + cmp r4, #0
1.330 + ldrge r6, [r0, #vp8_writer_buffer]
1.331 + ldrb r12, [r6, r4]
1.332 + cmpge r12, #0xff
1.333 + beq end_zero_while_loop
1.334 +
1.335 + ldr r6, [r0, #vp8_writer_buffer]
1.336 + ldrb r7, [r6, r4]
1.337 + add r7, r7, #1
1.338 + strb r7, [r6, r4]
1.339 +end_high_bit_not_set
1.340 + adds r3, r3, #1 ; ++count
1.341 + lsl r2, r2, #1 ; lowvalue <<= 1
1.342 + bne end_count_zero
1.343 +
1.344 + ldr r4, [r0, #vp8_writer_pos]
1.345 + mvn r3, #7 ; count = -8
1.346 + ldr r7, [r0, #vp8_writer_buffer]
1.347 + lsr r6, r2, #24 ; lowvalue >> 24
1.348 + add r12, r4, #1 ; w->pos++
1.349 + bic r2, r2, #0xff000000 ; lowvalue &= 0xffffff
1.350 + str r12, [r0, #vp8_writer_pos]
1.351 +
1.352 + VALIDATE_POS r7, r12 ; validate_buffer at pos
1.353 +
1.354 + strb r6, [r7, r4]
1.355 +end_count_zero
1.356 +skip_extra_bits
1.357 + add r1, r1, #TOKENEXTRA_SZ ; ++p
1.358 +check_p_lt_stop
1.359 + ldr r4, [sp, #0] ; stop
1.360 + cmp r1, r4 ; while( p < stop)
1.361 + bcc while_p_lt_stop
1.362 +
1.363 + ldr r10, [sp, #20] ; num_parts
1.364 + mov r1, #TOKENLIST_SZ
1.365 + mul r1, r10, r1
1.366 +
1.367 + ldr r6, [sp, #12] ; mb_rows
1.368 + ldr r7, [sp, #16] ; tokenlist address
1.369 + subs r6, r6, r10
1.370 + add r7, r7, r1 ; next element in the array
1.371 + str r6, [sp, #12]
1.372 + bgt mb_row_loop
1.373 +
1.374 +end_partition
1.375 + mov r12, #32
1.376 +
1.377 +stop_encode_loop
1.378 + sub r7, r5, #1 ; range-1
1.379 +
1.380 + mov r4, r7, lsl #7 ; ((range-1) * 128)
1.381 +
1.382 + mov r7, #1
1.383 + add r4, r7, r4, lsr #8 ; 1 + (((range-1) * 128) >> 8)
1.384 +
1.385 + ; Counting the leading zeros is used to normalize range.
1.386 + clz r6, r4
1.387 + sub r6, r6, #24 ; shift
1.388 +
1.389 + ; Flag is set on the sum of count. This flag is used later
1.390 + ; to determine if count >= 0
1.391 + adds r3, r3, r6 ; count += shift
1.392 + lsl r5, r4, r6 ; range <<= shift
1.393 + bmi token_count_lt_zero_se ; if(count >= 0)
1.394 +
1.395 + sub r6, r6, r3 ; offset = shift - count
1.396 + sub r4, r6, #1 ; offset-1
1.397 + lsls r4, r2, r4 ; if((lowvalue<<(offset-1)) & 0x80000000 )
1.398 + bpl token_high_bit_not_set_se
1.399 +
1.400 + ldr r4, [r0, #vp8_writer_pos] ; x
1.401 + sub r4, r4, #1 ; x = w->pos-1
1.402 + b token_zero_while_start_se
1.403 +token_zero_while_loop_se
1.404 + mov r10, #0
1.405 + strb r10, [r7, r4] ; w->buffer[x] =(unsigned char)0
1.406 + sub r4, r4, #1 ; x--
1.407 +token_zero_while_start_se
1.408 + cmp r4, #0
1.409 + ldrge r7, [r0, #vp8_writer_buffer]
1.410 + ldrb r11, [r7, r4]
1.411 + cmpge r11, #0xff
1.412 + beq token_zero_while_loop_se
1.413 +
1.414 + ldr r7, [r0, #vp8_writer_buffer]
1.415 + ldrb r10, [r7, r4] ; w->buffer[x]
1.416 + add r10, r10, #1
1.417 + strb r10, [r7, r4] ; w->buffer[x] + 1
1.418 +token_high_bit_not_set_se
1.419 + rsb r4, r6, #24 ; 24-offset
1.420 + ldr r10, [r0, #vp8_writer_buffer]
1.421 + lsr r7, r2, r4 ; lowvalue >> (24-offset)
1.422 + ldr r4, [r0, #vp8_writer_pos] ; w->pos
1.423 + lsl r2, r2, r6 ; lowvalue <<= offset
1.424 + mov r6, r3 ; shift = count
1.425 + add r11, r4, #1 ; w->pos++
1.426 + bic r2, r2, #0xff000000 ; lowvalue &= 0xffffff
1.427 + str r11, [r0, #vp8_writer_pos]
1.428 + sub r3, r3, #8 ; count -= 8
1.429 +
1.430 + VALIDATE_POS r10, r11 ; validate_buffer at pos
1.431 +
1.432 + strb r7, [r10, r4] ; w->buffer[w->pos++]
1.433 +
1.434 +token_count_lt_zero_se
1.435 + lsl r2, r2, r6 ; lowvalue <<= shift
1.436 +
1.437 + subs r12, r12, #1
1.438 + bne stop_encode_loop
1.439 +
1.440 + ldr r4, [r0, #vp8_writer_pos] ; w->pos
1.441 + ldr r12, [sp, #24] ; ptr
1.442 + add r12, r12, r4 ; ptr += w->pos
1.443 + str r12, [sp, #24]
1.444 +
1.445 + ldr r11, [sp, #28] ; i
1.446 + ldr r10, [sp, #20] ; num_parts
1.447 +
1.448 + add r11, r11, #1 ; i++
1.449 + str r11, [sp, #28]
1.450 +
1.451 + ldr r7, [sp, #32] ; cpi->tp_list[i]
1.452 + mov r1, #TOKENLIST_SZ
1.453 + add r7, r7, r1 ; next element in cpi->tp_list
1.454 + str r7, [sp, #32] ; cpi->tp_list[i+1]
1.455 +
1.456 + cmp r10, r11
1.457 + bgt numparts_loop
1.458 +
1.459 + add sp, sp, #40
1.460 + pop {r4-r12, pc}
1.461 + ENDP
1.462 +
1.463 +_VP8_COMP_common_
1.464 + DCD vp8_comp_common
1.465 +_VP8_COMMON_MBrows_
1.466 + DCD vp8_common_mb_rows
1.467 +_VP8_COMP_tplist_
1.468 + DCD vp8_comp_tplist
1.469 +_VP8_COMP_bc_
1.470 + DCD vp8_comp_bc
1.471 +_vp8_writer_sz_
1.472 + DCD vp8_writer_sz
1.473 +
1.474 + END
