vp8/decoder/arm/detokenizearm_v6.asm

   1 ;
   2 ;  Copyright (c) 2010 The VP8 project authors. All Rights Reserved.
   3 ;
   4 ;  Use of this source code is governed by a BSD-style license and patent
   5 ;  grant that can be found in the LICENSE file in the root of the source
   6 ;  tree. All contributing project authors may be found in the AUTHORS
   7 ;  file in the root of the source tree.
   8 ;
   9
  10
  11     EXPORT  |vp8_decode_mb_tokens_v5|
  12
  13     AREA    |.text|, CODE, READONLY  ; name this block of code
  14
  15     INCLUDE vpx_asm_offsets.asm
  16
  17 l_qcoeff    EQU     0
  18 l_i         EQU     4
  19 l_type      EQU     8
  20 l_stop      EQU     12
  21 l_c         EQU     16
  22 l_l_ptr      EQU     20
  23 l_a_ptr      EQU     24
  24 l_bc        EQU     28
  25 l_coef_ptr   EQU     32
  26 l_stacksize EQU     64
  27
  28
  29 ;; constant offsets -- these should be created at build time
  30 c_onyxblock2left_offset      EQU 25
  31 c_onyxblock2above_offset     EQU 50
  32 c_entropy_nodes              EQU 11
  33 c_dct_eob_token              EQU 11
  34
  35 |vp8_decode_mb_tokens_v5| PROC
  36     stmdb       sp!, {r4 - r11, lr}
  37     sub         sp, sp, #l_stacksize
  38     mov         r7, r1
  39     mov         r9, r0                      ;DETOK *detoken
  40
  41     ldr         r1, [r9, #detok_current_bc]
  42     ldr         r0, [r9, #detok_qcoeff_start_ptr]
  43     mov         r11, #0
  44     mov         r3, #0x10
  45
  46     cmp         r7, #1
  47     addeq       r11, r11, #24
  48     addeq       r3, r3, #8
  49     addeq       r0, r0, #3, 24
  50
  51     str         r0, [sp, #l_qcoeff]
  52     str         r11, [sp, #l_i]
  53     str         r7, [sp, #l_type]
  54     str         r3, [sp, #l_stop]
  55     str         r1, [sp, #l_bc]
  56
  57     add         lr, r9, r7, lsl #2
  58
  59     ldr         r2, [r1, #bool_decoder_buffer]
  60     ldr         r3, [r1, #bool_decoder_pos]
  61
  62     ldr         r10, [lr, #detok_coef_probs]
  63     ldr         r5, [r1, #bool_decoder_count]
  64     ldr         r6, [r1, #bool_decoder_range]
  65     ldr         r4, [r1, #bool_decoder_value]
  66     add         r8, r2, r3
  67
  68     str         r10, [sp, #l_coef_ptr]
  69
  70
  71     ;align 4
  72 BLOCK_LOOP
  73     ldr         r3, [r9, #detok_ptr_onyxblock2context_leftabove]
  74     ldr         r2, [r9, #DETOK_A]
  75     ldr         r1, [r9, #DETOK_L]
  76     ldrb        r12, [r3, +r11]                                 ; detoken->ptr_onyxblock2context_leftabove[i]
  77
  78     cmp         r7, #0                                          ; check type
  79     moveq       r7, #1
  80     movne       r7, #0
  81
  82     ldr         r0, [r2, +r12, lsl #2]                          ; a
  83     add         r1, r1, r12, lsl #4
  84     add         r3, r3, r11
  85
  86     ldrb        r2, [r3, #c_onyxblock2above_offset]
  87     ldrb        r3, [r3, #c_onyxblock2left_offset]
  88     mov         lr, #c_entropy_nodes
  89 ;;  ;++
  90
  91     ldr         r2, [r0, +r2, lsl #2]!
  92     add         r3, r1, r3, lsl #2
  93     str         r3, [sp, #l_l_ptr]
  94     ldr         r3, [r3]
  95
  96     cmp         r2, #0
  97     movne       r2, #1
  98     cmp         r3, #0
  99     addne       r2, r2, #1
 100
 101     str         r0, [sp, #l_a_ptr]
 102     smlabb      r0, r2, lr, r10
 103     mov         r1, #0                                          ; t = 0
 104     str         r7, [sp, #l_c]
 105
 106     ;align 4
 107 COEFF_LOOP
 108     ldr         r3, [r9, #detok_ptr_onyx_coef_bands_x]
 109     ldr         lr, [r9, #detok_onyx_coef_tree_ptr]
 110
 111 ;;the following two lines are used if onyx_coef_bands_x is UINT16
 112 ;;  add         r3, r3, r7, lsl #1
 113 ;;  ldrh        r3, [r3]
 114
 115 ;;the following line is used if onyx_coef_bands_x is UINT8
 116     ldrb        r3, [r7, +r3]
 117
 118
 119 ;;  ;++
 120 ;;  pld         [r8]
 121     ;++
 122     add         r0, r0, r3
 123
 124     ;align 4
 125 get_token_loop
 126     ldrb        r2, [r0, +r1, asr #1]
 127     mov         r3, r6, lsl #8
 128     sub         r3, r3, #256                    ;split = 1 +  (((range-1) * probability) >> 8)
 129     mov         r10, #1
 130
 131     smlawb      r2, r3, r2, r10
 132     ldrb        r12, [r8]                       ;load cx data byte in stall slot
 133     ;++
 134
 135     subs        r3, r4, r2, lsl #24             ;x = value-(split<<24)
 136     addhs       r1, r1, #1                      ;t += 1
 137     movhs       r4, r3                          ;update value
 138     subhs       r2, r6, r2                      ;range = range - split
 139     movlo       r6, r2
 140
 141 ;;; ldrsbhs     r1, [r1, +lr]
 142     ldrsb     r1, [r1, +lr]
 143
 144
 145 ;; use branch for short pipelines ???
 146 ;;  cmp         r2, #0x80
 147 ;;  bcs         |$LN22@decode_mb_to|
 148
 149     clz         r3, r2
 150     sub         r3, r3, #24
 151     subs        r5, r5, r3
 152     mov         r6, r2, lsl r3
 153     mov         r4, r4, lsl r3
 154
 155 ;; use branch for short pipelines ???
 156 ;;  bgt         |$LN22@decode_mb_to|
 157
 158     addle         r5, r5, #8
 159     rsble         r3, r5, #8
 160     addle         r8, r8, #1
 161     orrle         r4, r4, r12, lsl r3
 162
 163 ;;|$LN22@decode_mb_to|
 164
 165     cmp         r1, #0
 166     bgt         get_token_loop
 167
 168     cmn         r1, #c_dct_eob_token             ;if(t == -DCT_EOB_TOKEN)
 169     beq         END_OF_BLOCK
 170
 171     rsb         lr, r1, #0                      ;v = -t;
 172
 173     cmp         lr, #4                          ;if(v > FOUR_TOKEN)
 174     ble         SKIP_EXTRABITS
 175
 176     ldr         r3, [r9, #detok_teb_base_ptr]
 177     mov         r11, #1
 178     add         r7, r3, lr, lsl #4
 179
 180     ldrsh       lr, [r7, #tokenextrabits_min_val];v = teb_ptr->min_val
 181     ldrsh       r0, [r7, #tokenextrabits_length];bits_count = teb_ptr->Length
 182
 183 extrabits_loop
 184     add         r3, r0, r7
 185
 186     ldrb        r2, [r3, #4]
 187     mov         r3, r6, lsl #8
 188     sub         r3, r3, #256                    ;split = 1 +  (((range-1) * probability) >> 8)
 189     mov         r10, #1
 190
 191     smlawb      r2, r3, r2, r10
 192     ldrb        r12, [r8]
 193     ;++
 194
 195     subs        r10, r4, r2, lsl #24            ;x = value-(split<<24)
 196     movhs       r4, r10                         ;update value
 197     subhs       r2, r6, r2                      ;range = range - split
 198     addhs       lr, lr, r11, lsl r0             ;v += ((UINT16)1<<bits_count)
 199     movlo       r6, r2                          ;range = split
 200
 201
 202 ;; use branch for short pipelines ???
 203 ;;  cmp         r2, #0x80
 204 ;;  bcs         |$LN10@decode_mb_to|
 205
 206     clz         r3, r2
 207     sub         r3, r3, #24
 208     subs        r5, r5, r3
 209     mov         r6, r2, lsl r3                  ;range
 210     mov         r4, r4, lsl r3                  ;value
 211
 212     addle       r5, r5, #8
 213     addle       r8, r8, #1
 214     rsble       r3, r5, #8
 215     orrle       r4, r4, r12, lsl r3
 216
 217 ;;|$LN10@decode_mb_to|
 218     subs         r0, r0, #1
 219     bpl         extrabits_loop
 220
 221
 222 SKIP_EXTRABITS
 223     ldr         r11, [sp, #l_qcoeff]
 224     ldr         r0, [sp, #l_coef_ptr]
 225
 226     cmp         r1, #0                          ;check for nonzero token
 227     beq         SKIP_EOB_CHECK              ;if t is zero, we will skip the eob table chec
 228
 229     sub         r3, r6, #1                      ;range - 1
 230     ;++
 231     mov         r3, r3, lsl #7                  ; *= onyx_prob_half  (128)
 232     ;++
 233     mov         r3, r3, lsr #8
 234     add         r2, r3, #1                      ;split
 235
 236     subs        r3, r4, r2, lsl #24             ;x = value-(split<<24)
 237     movhs       r4, r3                          ;update value
 238     subhs       r2, r6, r2                      ;range = range - split
 239     mvnhs       r3, lr
 240     addhs       lr, r3, #1                      ;v = (v ^ -1) + 1
 241     movlo       r6, r2                          ;range = split
 242
 243 ;; use branch for short pipelines ???
 244 ;;  cmp         r2, #0x80
 245 ;;  bcs         |$LN6@decode_mb_to|
 246
 247     clz         r3, r2
 248     sub         r3, r3, #24
 249     subs        r5, r5, r3
 250     mov         r6, r2, lsl r3
 251     mov         r4, r4, lsl r3
 252     ldrleb      r2, [r8], #1
 253     addle       r5, r5, #8
 254     rsble       r3, r5, #8
 255     orrle       r4, r4, r2, lsl r3
 256
 257 ;;|$LN6@decode_mb_to|
 258     add         r0, r0, #0xB
 259
 260     cmn         r1, #1
 261
 262     addlt       r0, r0, #0xB
 263
 264     mvn         r1, #1
 265
 266 SKIP_EOB_CHECK
 267     ldr         r7, [sp, #l_c]
 268     ldr         r3, [r9, #detok_scan]
 269     add         r1, r1, #2
 270     cmp         r7, #(0x10 - 1)                     ;assume one less for now.... increment below
 271
 272     ldr         r3, [r3, +r7, lsl #2]
 273     add         r7, r7, #1
 274     add         r3, r11, r3, lsl #1
 275
 276     str         r7, [sp, #l_c]
 277     strh        lr, [r3]
 278
 279     blt         COEFF_LOOP
 280
 281     sub         r7, r7, #1                          ;if(t != -DCT_EOB_TOKEN) --c
 282
 283 END_OF_BLOCK
 284     ldr         r3, [sp, #l_type]
 285     ldr         r10, [sp, #l_coef_ptr]
 286     ldr         r0, [sp, #l_qcoeff]
 287     ldr         r11, [sp, #l_i]
 288     ldr         r12, [sp, #l_stop]
 289
 290     cmp         r3, #0
 291     moveq       r1, #1
 292     movne       r1, #0
 293     add         r3, r11, r9
 294
 295     cmp         r7, r1
 296     strb        r7, [r3, #detok_eob]
 297
 298     ldr         r7, [sp, #l_l_ptr]
 299     ldr         r2, [sp, #l_a_ptr]
 300     movne       r3, #1
 301     moveq       r3, #0
 302
 303     add         r0, r0, #0x20
 304     add         r11, r11, #1
 305     str         r3, [r7]
 306     str         r3, [r2]
 307     str         r0, [sp, #l_qcoeff]
 308     str         r11, [sp, #l_i]
 309
 310     cmp         r11, r12                            ;i >= stop ?
 311     ldr         r7, [sp, #l_type]
 312     mov         lr, #0xB
 313
 314     blt         BLOCK_LOOP
 315
 316     cmp         r11, #0x19
 317     bne         ln2_decode_mb_to
 318
 319     ldr         r12, [r9, #detok_qcoeff_start_ptr]
 320     ldr         r10, [r9, #detok_coef_probs]
 321     mov         r7, #0
 322     mov         r3, #0x10
 323     str         r12, [sp, #l_qcoeff]
 324     str         r7, [sp, #l_i]
 325     str         r7, [sp, #l_type]
 326     str         r3, [sp, #l_stop]
 327
 328     str         r10, [sp, #l_coef_ptr]
 329
 330     b           BLOCK_LOOP
 331
 332 ln2_decode_mb_to
 333     cmp         r11, #0x10
 334     bne         ln1_decode_mb_to
 335
 336     ldr         r10, [r9, #0x30]
 337
 338     mov         r7, #2
 339     mov         r3, #0x18
 340
 341     str         r7, [sp, #l_type]
 342     str         r3, [sp, #l_stop]
 343
 344     str         r10, [sp, #l_coef_ptr]
 345     b           BLOCK_LOOP
 346
 347 ln1_decode_mb_to
 348     ldr         r2, [sp, #l_bc]
 349     mov         r0, #0
 350     nop
 351
 352     ldr         r3, [r2, #bool_decoder_buffer]
 353     str         r5, [r2, #bool_decoder_count]
 354     str         r4, [r2, #bool_decoder_value]
 355     sub         r3, r8, r3
 356     str         r3, [r2, #bool_decoder_pos]
 357     str         r6, [r2, #bool_decoder_range]
 358
 359     add         sp, sp, #l_stacksize
 360     ldmia       sp!, {r4 - r11, pc}
 361
 362     ENDP  ; |vp8_decode_mb_tokens_v5|
 363
 364     END