Home | History | Annotate | Download | only in neon
      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_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_writer *w
     23 ; r1 const TOKENEXTRA *p
     24 ; r2 int xcount
     25 ; r3 vp8_coef_encodings
     26 ; s0 vp8_extra_bits
     27 ; s1 vp8_coef_tree
     28 |vp8cx_pack_tokens_armv7| PROC
     29     push    {r4-r11, lr}
     30 
     31     ; Add size of xcount * sizeof (TOKENEXTRA) to get stop
     32     ;  sizeof (TOKENEXTRA) is 20
     33     add     r2, r2, r2, lsl #2          ; xcount
     34     sub     sp, sp, #12
     35     add     r2, r1, r2, lsl #2          ; stop = p + xcount
     36     str     r2, [sp, #0]
     37     str     r3, [sp, #8]                ; save vp8_coef_encodings
     38     ldr     r2, [r0, #vp8_writer_lowvalue]
     39     ldr     r5, [r0, #vp8_writer_range]
     40     ldr     r3, [r0, #vp8_writer_count]
     41     b       check_p_lt_stop
     42 
     43 while_p_lt_stop
     44     ldr     r6, [r1, #tokenextra_token] ; t
     45     ldr     r4, [sp, #8]                ; vp8_coef_encodings
     46     mov     lr, #0
     47     add     r4, r4, r6, lsl #3          ; a = vp8_coef_encodings + t
     48     ldr     r9, [r1, #tokenextra_context_tree]   ; pp
     49 
     50     ldr     r7, [r1, #tokenextra_skip_eob_node]
     51 
     52     ldr     r6, [r4, #vp8_token_value]  ; v
     53     ldr     r8, [r4, #vp8_token_len]    ; n
     54 
     55     ; vp8 specific skip_eob_node
     56     cmp     r7, #0
     57     movne   lr, #2                      ; i = 2
     58     subne   r8, r8, #1                  ; --n
     59 
     60     ; reverse the stream of bits to be packed.  Normally
     61     ; the most significant bit is peeled off and compared
     62     ; in the form of (v >> --n) & 1.  ARM architecture has
     63     ; the ability to set a flag based on the value of the
     64     ; bit shifted off the bottom of the register.  To make
     65     ; that happen the bitstream is reversed.
     66     rbit    r12, r6
     67     rsb     r4, r8, #32                 ; 32-n
     68     ldr     r10, [sp, #52]              ; vp8_coef_tree
     69 
     70     ; v is kept in r12 during the token pack loop
     71     lsr     r12, r12, r4                ; v >>= 32 - n
     72 
     73 ; loop start
     74 token_loop
     75     ldrb    r4, [r9, lr, asr #1]        ; pp [i>>1]
     76     sub     r7, r5, #1                  ; range-1
     77 
     78     ; Decisions are made based on the bit value shifted
     79     ; off of v, so set a flag here based on this.
     80     ; This value is refered to as "bb"
     81     lsrs    r12, r12, #1                ; bb = v >> n
     82     mul     r4, r4, r7                  ; ((range-1) * pp[i>>1]))
     83 
     84     ; bb can only be 0 or 1.  So only execute this statement
     85     ; if bb == 1, otherwise it will act like i + 0
     86     addcs   lr, lr, #1                  ; i + bb
     87 
     88     mov     r7, #1
     89     ldrsb   lr, [r10, lr]               ; i = vp8_coef_tree[i+bb]
     90     add     r4, r7, r4, lsr #8          ; 1 + (((range-1) * pp[i>>1]) >> 8)
     91 
     92     addcs   r2, r2, r4                  ; if  (bb) lowvalue += split
     93     subcs   r4, r5, r4                  ; if  (bb) range = range-split
     94 
     95     ; Counting the leading zeros is used to normalize range.
     96     clz     r6, r4
     97     sub     r6, r6, #24                 ; shift
     98 
     99     ; Flag is set on the sum of count.  This flag is used later
    100     ; to determine if count >= 0
    101     adds    r3, r3, r6                  ; count += shift
    102     lsl     r5, r4, r6                  ; range <<= shift
    103     bmi     token_count_lt_zero         ; if(count >= 0)
    104 
    105     sub     r6, r6, r3                  ; offset = shift - count
    106     sub     r4, r6, #1                  ; offset-1
    107     lsls    r4, r2, r4                  ; if((lowvalue<<(offset-1)) & 0x80000000 )
    108     bpl     token_high_bit_not_set
    109 
    110     ldr     r4, [r0, #vp8_writer_pos]   ; x
    111     sub     r4, r4, #1                  ; x = w->pos-1
    112     b       token_zero_while_start
    113 token_zero_while_loop
    114     mov     r10, #0
    115     strb    r10, [r7, r4]               ; w->buffer[x] =(unsigned char)0
    116     sub     r4, r4, #1                  ; x--
    117 token_zero_while_start
    118     cmp     r4, #0
    119     ldrge   r7, [r0, #vp8_writer_buffer]
    120     ldrb    r11, [r7, r4]
    121     cmpge   r11, #0xff
    122     beq     token_zero_while_loop
    123 
    124     ldr     r7, [r0, #vp8_writer_buffer]
    125     ldrb    r10, [r7, r4]               ; w->buffer[x]
    126     add     r10, r10, #1
    127     strb    r10, [r7, r4]               ; w->buffer[x] + 1
    128 token_high_bit_not_set
    129     rsb     r4, r6, #24                 ; 24-offset
    130     ldr     r10, [r0, #vp8_writer_buffer]
    131     lsr     r7, r2, r4                  ; lowvalue >> (24-offset)
    132     ldr     r4, [r0, #vp8_writer_pos]   ; w->pos
    133     lsl     r2, r2, r6                  ; lowvalue <<= offset
    134     mov     r6, r3                      ; shift = count
    135     add     r11, r4, #1                 ; w->pos++
    136     bic     r2, r2, #0xff000000         ; lowvalue &= 0xffffff
    137     str     r11, [r0, #vp8_writer_pos]
    138     sub     r3, r3, #8                  ; count -= 8
    139     strb    r7, [r10, r4]               ; w->buffer[w->pos++]
    140 
    141     ; r10 is used earlier in the loop, but r10 is used as
    142     ; temp variable here.  So after r10 is used, reload
    143     ; vp8_coef_tree_dcd into r10
    144     ldr     r10, [sp, #52]              ; vp8_coef_tree
    145 
    146 token_count_lt_zero
    147     lsl     r2, r2, r6                  ; lowvalue <<= shift
    148 
    149     subs    r8, r8, #1                  ; --n
    150     bne     token_loop
    151 
    152     ldr     r6, [r1, #tokenextra_token] ; t
    153     ldr     r7, [sp, #48]               ; vp8_extra_bits
    154     ; Add t * sizeof (vp8_extra_bit_struct) to get the desired
    155     ;  element.  Here vp8_extra_bit_struct == 20
    156     add     r6, r6, r6, lsl #2          ; b = vp8_extra_bits + t
    157     add     r12, r7, r6, lsl #2         ; b = vp8_extra_bits + t
    158 
    159     ldr     r4, [r12, #vp8_extra_bit_struct_base_val]
    160     cmp     r4, #0
    161     beq     skip_extra_bits
    162 
    163 ;   if( b->base_val)
    164     ldr     r8, [r12, #vp8_extra_bit_struct_len] ; L
    165     ldr     lr, [r1, #tokenextra_extra] ; e = p->Extra
    166     cmp     r8, #0                      ; if( L)
    167     beq     no_extra_bits
    168 
    169     ldr     r9, [r12, #vp8_extra_bit_struct_prob]
    170     asr     r7, lr, #1                  ; v=e>>1
    171 
    172     ldr     r10, [r12, #vp8_extra_bit_struct_tree]
    173     str     r10, [sp, #4]               ; b->tree
    174 
    175     rbit    r12, r7                     ; reverse v
    176     rsb     r4, r8, #32
    177     lsr     r12, r12, r4
    178 
    179     mov     lr, #0                      ; i = 0
    180 
    181 extra_bits_loop
    182     ldrb    r4, [r9, lr, asr #1]            ; pp[i>>1]
    183     sub     r7, r5, #1                  ; range-1
    184     lsrs    r12, r12, #1                ; v >> n
    185     mul     r4, r4, r7                  ; (range-1) * pp[i>>1]
    186     addcs   lr, lr, #1                  ; i + bb
    187 
    188     mov     r7, #1
    189     ldrsb   lr, [r10, lr]               ; i = b->tree[i+bb]
    190     add     r4, r7, r4, lsr #8          ; split = 1 +  (((range-1) * pp[i>>1]) >> 8)
    191 
    192     addcs   r2, r2, r4                  ; if  (bb) lowvalue += split
    193     subcs   r4, r5, r4                  ; if  (bb) range = range-split
    194 
    195     clz     r6, r4
    196     sub     r6, r6, #24
    197 
    198     adds    r3, r3, r6                  ; count += shift
    199     lsl     r5, r4, r6                  ; range <<= shift
    200     bmi     extra_count_lt_zero         ; if(count >= 0)
    201 
    202     sub     r6, r6, r3                  ; offset= shift - count
    203     sub     r4, r6, #1                  ; offset-1
    204     lsls    r4, r2, r4                  ; if((lowvalue<<(offset-1)) & 0x80000000 )
    205     bpl     extra_high_bit_not_set
    206 
    207     ldr     r4, [r0, #vp8_writer_pos]   ; x
    208     sub     r4, r4, #1                  ; x = w->pos - 1
    209     b       extra_zero_while_start
    210 extra_zero_while_loop
    211     mov     r10, #0
    212     strb    r10, [r7, r4]               ; w->buffer[x] =(unsigned char)0
    213     sub     r4, r4, #1                  ; x--
    214 extra_zero_while_start
    215     cmp     r4, #0
    216     ldrge   r7, [r0, #vp8_writer_buffer]
    217     ldrb    r11, [r7, r4]
    218     cmpge   r11, #0xff
    219     beq     extra_zero_while_loop
    220 
    221     ldr     r7, [r0, #vp8_writer_buffer]
    222     ldrb    r10, [r7, r4]
    223     add     r10, r10, #1
    224     strb    r10, [r7, r4]
    225 extra_high_bit_not_set
    226     rsb     r4, r6, #24                 ; 24-offset
    227     ldr     r10, [r0, #vp8_writer_buffer]
    228     lsr     r7, r2, r4                  ; lowvalue >> (24-offset)
    229     ldr     r4, [r0, #vp8_writer_pos]
    230     lsl     r2, r2, r6                  ; lowvalue <<= offset
    231     mov     r6, r3                      ; shift = count
    232     add     r11, r4, #1                 ; w->pos++
    233     bic     r2, r2, #0xff000000         ; lowvalue &= 0xffffff
    234     str     r11, [r0, #vp8_writer_pos]
    235     sub     r3, r3, #8                  ; count -= 8
    236     strb    r7, [r10, r4]               ; w->buffer[w->pos++]=(lowvalue >> (24-offset))
    237     ldr     r10, [sp, #4]               ; b->tree
    238 extra_count_lt_zero
    239     lsl     r2, r2, r6
    240 
    241     subs    r8, r8, #1                  ; --n
    242     bne     extra_bits_loop             ; while (n)
    243 
    244 no_extra_bits
    245     ldr     lr, [r1, #4]                ; e = p->Extra
    246     add     r4, r5, #1                  ; range + 1
    247     tst     lr, #1
    248     lsr     r4, r4, #1                  ; split = (range + 1) >> 1
    249     addne   r2, r2, r4                  ; lowvalue += split
    250     subne   r4, r5, r4                  ; range = range-split
    251     tst     r2, #0x80000000             ; lowvalue & 0x80000000
    252     lsl     r5, r4, #1                  ; range <<= 1
    253     beq     end_high_bit_not_set
    254 
    255     ldr     r4, [r0, #vp8_writer_pos]
    256     mov     r7, #0
    257     sub     r4, r4, #1
    258     b       end_zero_while_start
    259 end_zero_while_loop
    260     strb    r7, [r6, r4]
    261     sub     r4, r4, #1                  ; x--
    262 end_zero_while_start
    263     cmp     r4, #0
    264     ldrge   r6, [r0, #vp8_writer_buffer]
    265     ldrb    r12, [r6, r4]
    266     cmpge   r12, #0xff
    267     beq     end_zero_while_loop
    268 
    269     ldr     r6, [r0, #vp8_writer_buffer]
    270     ldrb    r7, [r6, r4]
    271     add     r7, r7, #1
    272     strb    r7, [r6, r4]
    273 end_high_bit_not_set
    274     adds    r3, r3, #1                  ; ++count
    275     lsl     r2, r2, #1                  ; lowvalue  <<= 1
    276     bne     end_count_zero
    277 
    278     ldr     r4, [r0, #vp8_writer_pos]
    279     mvn     r3, #7
    280     ldr     r7, [r0, #vp8_writer_buffer]
    281     lsr     r6, r2, #24                 ; lowvalue >> 24
    282     add     r12, r4, #1                 ; w->pos++
    283     bic     r2, r2, #0xff000000         ; lowvalue &= 0xffffff
    284     str     r12, [r0, #0x10]
    285     strb    r6, [r7, r4]
    286 end_count_zero
    287 skip_extra_bits
    288     add     r1, r1, #TOKENEXTRA_SZ      ; ++p
    289 check_p_lt_stop
    290     ldr     r4, [sp, #0]                ; stop
    291     cmp     r1, r4                      ; while( p < stop)
    292     bcc     while_p_lt_stop
    293 
    294     str     r2, [r0, #vp8_writer_lowvalue]
    295     str     r5, [r0, #vp8_writer_range]
    296     str     r3, [r0, #vp8_writer_count]
    297     add     sp, sp, #12
    298     pop     {r4-r11, pc}
    299     ENDP
    300 
    301     END
    302