1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libvpx/vp8/encoder/arm/armv5te/vp8_packtokens_mbrow_armv5.asm Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,352 @@
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_mb_row_tokens_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 +
1.27 + ; macro for validating write buffer position
1.28 + ; needs vp8_writer in r0
1.29 + ; start shall not be in r1
1.30 + MACRO
1.31 + VALIDATE_POS $start, $pos
1.32 + push {r0-r3, r12, lr} ; rest of regs are preserved by subroutine call
1.33 + ldr r2, [r0, #vp8_writer_buffer_end]
1.34 + ldr r3, [r0, #vp8_writer_error]
1.35 + mov r1, $pos
1.36 + mov r0, $start
1.37 + bl vp8_validate_buffer_arm
1.38 + pop {r0-r3, r12, lr}
1.39 + MEND
1.40 +
1.41 +; r0 VP8_COMP *cpi
1.42 +; r1 vp8_writer *w
1.43 +; r2 vp8_coef_encodings
1.44 +; r3 vp8_extra_bits
1.45 +; s0 vp8_coef_tree
1.46 +
1.47 +|vp8cx_pack_mb_row_tokens_armv5| PROC
1.48 + push {r4-r12, lr}
1.49 + sub sp, sp, #24
1.50 +
1.51 + ; Compute address of cpi->common.mb_rows
1.52 + ldr r4, _VP8_COMP_common_
1.53 + ldr r6, _VP8_COMMON_MBrows_
1.54 + add r4, r0, r4
1.55 +
1.56 + ldr r5, [r4, r6] ; load up mb_rows
1.57 +
1.58 + str r2, [sp, #20] ; save vp8_coef_encodings
1.59 + str r5, [sp, #12] ; save mb_rows
1.60 + str r3, [sp, #8] ; save vp8_extra_bits
1.61 +
1.62 + ldr r4, _VP8_COMP_tplist_
1.63 + add r4, r0, r4
1.64 + ldr r7, [r4, #0] ; dereference cpi->tp_list
1.65 +
1.66 + mov r0, r1 ; keep same as other loops
1.67 +
1.68 + ldr r2, [r0, #vp8_writer_lowvalue]
1.69 + ldr r5, [r0, #vp8_writer_range]
1.70 + ldr r3, [r0, #vp8_writer_count]
1.71 +
1.72 +mb_row_loop
1.73 +
1.74 + ldr r1, [r7, #tokenlist_start]
1.75 + ldr r9, [r7, #tokenlist_stop]
1.76 + str r9, [sp, #0] ; save stop for later comparison
1.77 + str r7, [sp, #16] ; tokenlist address for next time
1.78 +
1.79 + b check_p_lt_stop
1.80 +
1.81 + ; actuall work gets done here!
1.82 +
1.83 +while_p_lt_stop
1.84 + ldrb r6, [r1, #tokenextra_token] ; t
1.85 + ldr r4, [sp, #20] ; vp8_coef_encodings
1.86 + mov lr, #0
1.87 + add r4, r4, r6, lsl #3 ; a = vp8_coef_encodings + t
1.88 + ldr r9, [r1, #tokenextra_context_tree] ; pp
1.89 +
1.90 + ldrb r7, [r1, #tokenextra_skip_eob_node]
1.91 +
1.92 + ldr r6, [r4, #vp8_token_value] ; v
1.93 + ldr r8, [r4, #vp8_token_len] ; n
1.94 +
1.95 + ; vp8 specific skip_eob_node
1.96 + cmp r7, #0
1.97 + movne lr, #2 ; i = 2
1.98 + subne r8, r8, #1 ; --n
1.99 +
1.100 + rsb r4, r8, #32 ; 32-n
1.101 + ldr r10, [sp, #64] ; vp8_coef_tree
1.102 +
1.103 + ; v is kept in r12 during the token pack loop
1.104 + lsl r12, r6, r4 ; r12 = v << 32 - n
1.105 +
1.106 +; loop start
1.107 +token_loop
1.108 + ldrb r4, [r9, lr, asr #1] ; pp [i>>1]
1.109 + sub r7, r5, #1 ; range-1
1.110 +
1.111 + ; Decisions are made based on the bit value shifted
1.112 + ; off of v, so set a flag here based on this.
1.113 + ; This value is refered to as "bb"
1.114 + lsls r12, r12, #1 ; bb = v >> n
1.115 + mul r6, r4, r7 ; ((range-1) * pp[i>>1]))
1.116 +
1.117 + ; bb can only be 0 or 1. So only execute this statement
1.118 + ; if bb == 1, otherwise it will act like i + 0
1.119 + addcs lr, lr, #1 ; i + bb
1.120 +
1.121 + mov r7, #1
1.122 + ldrsb lr, [r10, lr] ; i = vp8_coef_tree[i+bb]
1.123 + add r4, r7, r6, lsr #8 ; 1 + (((range-1) * pp[i>>1]) >> 8)
1.124 +
1.125 + addcs r2, r2, r4 ; if (bb) lowvalue += split
1.126 + subcs r4, r5, r4 ; if (bb) range = range-split
1.127 +
1.128 + ; Counting the leading zeros is used to normalize range.
1.129 + clz r6, r4
1.130 + sub r6, r6, #24 ; shift
1.131 +
1.132 + ; Flag is set on the sum of count. This flag is used later
1.133 + ; to determine if count >= 0
1.134 + adds r3, r3, r6 ; count += shift
1.135 + lsl r5, r4, r6 ; range <<= shift
1.136 + bmi token_count_lt_zero ; if(count >= 0)
1.137 +
1.138 + sub r6, r6, r3 ; offset = shift - count
1.139 + sub r4, r6, #1 ; offset-1
1.140 + lsls r4, r2, r4 ; if((lowvalue<<(offset-1)) & 0x80000000 )
1.141 + bpl token_high_bit_not_set
1.142 +
1.143 + ldr r4, [r0, #vp8_writer_pos] ; x
1.144 + sub r4, r4, #1 ; x = w->pos-1
1.145 + b token_zero_while_start
1.146 +token_zero_while_loop
1.147 + mov r10, #0
1.148 + strb r10, [r7, r4] ; w->buffer[x] =(unsigned char)0
1.149 + sub r4, r4, #1 ; x--
1.150 +token_zero_while_start
1.151 + cmp r4, #0
1.152 + ldrge r7, [r0, #vp8_writer_buffer]
1.153 + ldrb r11, [r7, r4]
1.154 + cmpge r11, #0xff
1.155 + beq token_zero_while_loop
1.156 +
1.157 + ldr r7, [r0, #vp8_writer_buffer]
1.158 + ldrb r10, [r7, r4] ; w->buffer[x]
1.159 + add r10, r10, #1
1.160 + strb r10, [r7, r4] ; w->buffer[x] + 1
1.161 +token_high_bit_not_set
1.162 + rsb r4, r6, #24 ; 24-offset
1.163 + ldr r10, [r0, #vp8_writer_buffer]
1.164 + lsr r7, r2, r4 ; lowvalue >> (24-offset)
1.165 + ldr r4, [r0, #vp8_writer_pos] ; w->pos
1.166 + lsl r2, r2, r6 ; lowvalue <<= offset
1.167 + mov r6, r3 ; shift = count
1.168 + add r11, r4, #1 ; w->pos++
1.169 + bic r2, r2, #0xff000000 ; lowvalue &= 0xffffff
1.170 + str r11, [r0, #vp8_writer_pos]
1.171 + sub r3, r3, #8 ; count -= 8
1.172 +
1.173 + VALIDATE_POS r10, r11 ; validate_buffer at pos
1.174 +
1.175 + strb r7, [r10, r4] ; w->buffer[w->pos++]
1.176 +
1.177 + ; r10 is used earlier in the loop, but r10 is used as
1.178 + ; temp variable here. So after r10 is used, reload
1.179 + ; vp8_coef_tree_dcd into r10
1.180 + ldr r10, [sp, #64] ; vp8_coef_tree
1.181 +
1.182 +token_count_lt_zero
1.183 + lsl r2, r2, r6 ; lowvalue <<= shift
1.184 +
1.185 + subs r8, r8, #1 ; --n
1.186 + bne token_loop
1.187 +
1.188 + ldrb r6, [r1, #tokenextra_token] ; t
1.189 + ldr r7, [sp, #8] ; vp8_extra_bits
1.190 + ; Add t * sizeof (vp8_extra_bit_struct) to get the desired
1.191 + ; element. Here vp8_extra_bit_struct == 16
1.192 + add r12, r7, r6, lsl #4 ; b = vp8_extra_bits + t
1.193 +
1.194 + ldr r4, [r12, #vp8_extra_bit_struct_base_val]
1.195 + cmp r4, #0
1.196 + beq skip_extra_bits
1.197 +
1.198 +; if( b->base_val)
1.199 + ldr r8, [r12, #vp8_extra_bit_struct_len] ; L
1.200 + ldrsh lr, [r1, #tokenextra_extra] ; e = p->Extra
1.201 + cmp r8, #0 ; if( L)
1.202 + beq no_extra_bits
1.203 +
1.204 + ldr r9, [r12, #vp8_extra_bit_struct_prob]
1.205 + asr r7, lr, #1 ; v=e>>1
1.206 +
1.207 + ldr r10, [r12, #vp8_extra_bit_struct_tree]
1.208 + str r10, [sp, #4] ; b->tree
1.209 +
1.210 + rsb r4, r8, #32
1.211 + lsl r12, r7, r4
1.212 +
1.213 + mov lr, #0 ; i = 0
1.214 +
1.215 +extra_bits_loop
1.216 + ldrb r4, [r9, lr, asr #1] ; pp[i>>1]
1.217 + sub r7, r5, #1 ; range-1
1.218 + lsls r12, r12, #1 ; v >> n
1.219 + mul r6, r4, r7 ; (range-1) * pp[i>>1]
1.220 + addcs lr, lr, #1 ; i + bb
1.221 +
1.222 + mov r7, #1
1.223 + ldrsb lr, [r10, lr] ; i = b->tree[i+bb]
1.224 + add r4, r7, r6, lsr #8 ; split = 1 + (((range-1) * pp[i>>1]) >> 8)
1.225 +
1.226 + addcs r2, r2, r4 ; if (bb) lowvalue += split
1.227 + subcs r4, r5, r4 ; if (bb) range = range-split
1.228 +
1.229 + clz r6, r4
1.230 + sub r6, r6, #24
1.231 +
1.232 + adds r3, r3, r6 ; count += shift
1.233 + lsl r5, r4, r6 ; range <<= shift
1.234 + bmi extra_count_lt_zero ; if(count >= 0)
1.235 +
1.236 + sub r6, r6, r3 ; offset= shift - count
1.237 + sub r4, r6, #1 ; offset-1
1.238 + lsls r4, r2, r4 ; if((lowvalue<<(offset-1)) & 0x80000000 )
1.239 + bpl extra_high_bit_not_set
1.240 +
1.241 + ldr r4, [r0, #vp8_writer_pos] ; x
1.242 + sub r4, r4, #1 ; x = w->pos - 1
1.243 + b extra_zero_while_start
1.244 +extra_zero_while_loop
1.245 + mov r10, #0
1.246 + strb r10, [r7, r4] ; w->buffer[x] =(unsigned char)0
1.247 + sub r4, r4, #1 ; x--
1.248 +extra_zero_while_start
1.249 + cmp r4, #0
1.250 + ldrge r7, [r0, #vp8_writer_buffer]
1.251 + ldrb r11, [r7, r4]
1.252 + cmpge r11, #0xff
1.253 + beq extra_zero_while_loop
1.254 +
1.255 + ldr r7, [r0, #vp8_writer_buffer]
1.256 + ldrb r10, [r7, r4]
1.257 + add r10, r10, #1
1.258 + strb r10, [r7, r4]
1.259 +extra_high_bit_not_set
1.260 + rsb r4, r6, #24 ; 24-offset
1.261 + ldr r10, [r0, #vp8_writer_buffer]
1.262 + lsr r7, r2, r4 ; lowvalue >> (24-offset)
1.263 + ldr r4, [r0, #vp8_writer_pos]
1.264 + lsl r2, r2, r6 ; lowvalue <<= offset
1.265 + mov r6, r3 ; shift = count
1.266 + add r11, r4, #1 ; w->pos++
1.267 + bic r2, r2, #0xff000000 ; lowvalue &= 0xffffff
1.268 + str r11, [r0, #vp8_writer_pos]
1.269 + sub r3, r3, #8 ; count -= 8
1.270 +
1.271 + VALIDATE_POS r10, r11 ; validate_buffer at pos
1.272 +
1.273 + strb r7, [r10, r4] ; w->buffer[w->pos++]=(lowvalue >> (24-offset))
1.274 + ldr r10, [sp, #4] ; b->tree
1.275 +extra_count_lt_zero
1.276 + lsl r2, r2, r6
1.277 +
1.278 + subs r8, r8, #1 ; --n
1.279 + bne extra_bits_loop ; while (n)
1.280 +
1.281 +no_extra_bits
1.282 + ldr lr, [r1, #4] ; e = p->Extra
1.283 + add r4, r5, #1 ; range + 1
1.284 + tst lr, #1
1.285 + lsr r4, r4, #1 ; split = (range + 1) >> 1
1.286 + addne r2, r2, r4 ; lowvalue += split
1.287 + subne r4, r5, r4 ; range = range-split
1.288 + tst r2, #0x80000000 ; lowvalue & 0x80000000
1.289 + lsl r5, r4, #1 ; range <<= 1
1.290 + beq end_high_bit_not_set
1.291 +
1.292 + ldr r4, [r0, #vp8_writer_pos]
1.293 + mov r7, #0
1.294 + sub r4, r4, #1
1.295 + b end_zero_while_start
1.296 +end_zero_while_loop
1.297 + strb r7, [r6, r4]
1.298 + sub r4, r4, #1 ; x--
1.299 +end_zero_while_start
1.300 + cmp r4, #0
1.301 + ldrge r6, [r0, #vp8_writer_buffer]
1.302 + ldrb r12, [r6, r4]
1.303 + cmpge r12, #0xff
1.304 + beq end_zero_while_loop
1.305 +
1.306 + ldr r6, [r0, #vp8_writer_buffer]
1.307 + ldrb r7, [r6, r4]
1.308 + add r7, r7, #1
1.309 + strb r7, [r6, r4]
1.310 +end_high_bit_not_set
1.311 + adds r3, r3, #1 ; ++count
1.312 + lsl r2, r2, #1 ; lowvalue <<= 1
1.313 + bne end_count_zero
1.314 +
1.315 + ldr r4, [r0, #vp8_writer_pos]
1.316 + mvn r3, #7
1.317 + ldr r7, [r0, #vp8_writer_buffer]
1.318 + lsr r6, r2, #24 ; lowvalue >> 24
1.319 + add r12, r4, #1 ; w->pos++
1.320 + bic r2, r2, #0xff000000 ; lowvalue &= 0xffffff
1.321 + str r12, [r0, #vp8_writer_pos]
1.322 +
1.323 + VALIDATE_POS r7, r12 ; validate_buffer at pos
1.324 +
1.325 + strb r6, [r7, r4]
1.326 +end_count_zero
1.327 +skip_extra_bits
1.328 + add r1, r1, #TOKENEXTRA_SZ ; ++p
1.329 +check_p_lt_stop
1.330 + ldr r4, [sp, #0] ; stop
1.331 + cmp r1, r4 ; while( p < stop)
1.332 + bcc while_p_lt_stop
1.333 +
1.334 + ldr r6, [sp, #12] ; mb_rows
1.335 + ldr r7, [sp, #16] ; tokenlist address
1.336 + subs r6, r6, #1
1.337 + add r7, r7, #TOKENLIST_SZ ; next element in the array
1.338 + str r6, [sp, #12]
1.339 + bne mb_row_loop
1.340 +
1.341 + str r2, [r0, #vp8_writer_lowvalue]
1.342 + str r5, [r0, #vp8_writer_range]
1.343 + str r3, [r0, #vp8_writer_count]
1.344 + add sp, sp, #24
1.345 + pop {r4-r12, pc}
1.346 + ENDP
1.347 +
1.348 +_VP8_COMP_common_
1.349 + DCD vp8_comp_common
1.350 +_VP8_COMMON_MBrows_
1.351 + DCD vp8_common_mb_rows
1.352 +_VP8_COMP_tplist_
1.353 + DCD vp8_comp_tplist
1.354 +
1.355 + END
