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_armv5|
13    IMPORT |vp8_validate_buffer_arm|
14
15    INCLUDE vp8_asm_enc_offsets.asm
16
17    ARM
18    REQUIRE8
19    PRESERVE8
20
21    AREA    |.text|, CODE, READONLY
22
23
24    ; macro for validating write buffer position
25    ; needs vp8_writer in r0
26    ; start shall not be in r1
27    MACRO
28    VALIDATE_POS $start, $pos
29    push {r0-r3, r12, lr}        ; rest of regs are preserved by subroutine call
30    ldr  r2, [r0, #vp8_writer_buffer_end]
31    ldr  r3, [r0, #vp8_writer_error]
32    mov  r1, $pos
33    mov  r0, $start
34    bl   vp8_validate_buffer_arm
35    pop  {r0-r3, r12, lr}
36    MEND
37
38
39; r0 vp8_writer *w
40; r1 const TOKENEXTRA *p
41; r2 int xcount
42; r3 vp8_coef_encodings
43; s0 vp8_extra_bits
44; s1 vp8_coef_tree
45|vp8cx_pack_tokens_armv5| PROC
46    push    {r4-r12, lr}
47    sub     sp, sp, #16
48
49    ; Add size of xcount * sizeof (TOKENEXTRA) to get stop
50    ;  sizeof (TOKENEXTRA) is 8
51    add     r2, r1, r2, lsl #3          ; stop = p + xcount*sizeof(TOKENEXTRA)
52    str     r2, [sp, #0]
53    str     r3, [sp, #8]                ; save vp8_coef_encodings
54    ldr     r2, [r0, #vp8_writer_lowvalue]
55    ldr     r5, [r0, #vp8_writer_range]
56    ldr     r3, [r0, #vp8_writer_count]
57    b       check_p_lt_stop
58
59while_p_lt_stop
60    ldrb    r6, [r1, #tokenextra_token] ; t
61    ldr     r4, [sp, #8]                ; vp8_coef_encodings
62    mov     lr, #0
63    add     r4, r4, r6, lsl #3          ; a = vp8_coef_encodings + t
64    ldr     r9, [r1, #tokenextra_context_tree]   ; pp
65
66    ldrb    r7, [r1, #tokenextra_skip_eob_node]
67
68    ldr     r6, [r4, #vp8_token_value]  ; v
69    ldr     r8, [r4, #vp8_token_len]    ; n
70
71    ; vp8 specific skip_eob_node
72    cmp     r7, #0
73    movne   lr, #2                      ; i = 2
74    subne   r8, r8, #1                  ; --n
75
76    rsb     r4, r8, #32                 ; 32-n
77    ldr     r10, [sp, #60]              ; vp8_coef_tree
78
79    ; v is kept in r12 during the token pack loop
80    lsl     r12, r6, r4                ; r12 = v << 32 - n
81
82; loop start
83token_loop
84    ldrb    r4, [r9, lr, asr #1]        ; pp [i>>1]
85    sub     r7, r5, #1                  ; range-1
86
87    ; Decisions are made based on the bit value shifted
88    ; off of v, so set a flag here based on this.
89    ; This value is refered to as "bb"
90    lsls    r12, r12, #1                ; bb = v >> n
91    mul     r6, r4, r7                  ; ((range-1) * pp[i>>1]))
92
93    ; bb can only be 0 or 1.  So only execute this statement
94    ; if bb == 1, otherwise it will act like i + 0
95    addcs   lr, lr, #1                  ; i + bb
96
97    mov     r7, #1
98    ldrsb   lr, [r10, lr]               ; i = vp8_coef_tree[i+bb]
99    add     r4, r7, r6, lsr #8          ; 1 + (((range-1) * pp[i>>1]) >> 8)
100
101    addcs   r2, r2, r4                  ; if  (bb) lowvalue += split
102    subcs   r4, r5, r4                  ; if  (bb) range = range-split
103
104    ; Counting the leading zeros is used to normalize range.
105    clz     r6, r4
106    sub     r6, r6, #24                 ; shift
107
108    ; Flag is set on the sum of count.  This flag is used later
109    ; to determine if count >= 0
110    adds    r3, r3, r6                  ; count += shift
111    lsl     r5, r4, r6                  ; range <<= shift
112    bmi     token_count_lt_zero         ; if(count >= 0)
113
114    sub     r6, r6, r3                  ; offset = shift - count
115    sub     r4, r6, #1                  ; offset-1
116    lsls    r4, r2, r4                  ; if((lowvalue<<(offset-1)) & 0x80000000 )
117    bpl     token_high_bit_not_set
118
119    ldr     r4, [r0, #vp8_writer_pos]   ; x
120    sub     r4, r4, #1                  ; x = w->pos-1
121    b       token_zero_while_start
122token_zero_while_loop
123    mov     r10, #0
124    strb    r10, [r7, r4]               ; w->buffer[x] =(unsigned char)0
125    sub     r4, r4, #1                  ; x--
126token_zero_while_start
127    cmp     r4, #0
128    ldrge   r7, [r0, #vp8_writer_buffer]
129    ldrb    r11, [r7, r4]
130    cmpge   r11, #0xff
131    beq     token_zero_while_loop
132
133    ldr     r7, [r0, #vp8_writer_buffer]
134    ldrb    r10, [r7, r4]               ; w->buffer[x]
135    add     r10, r10, #1
136    strb    r10, [r7, r4]               ; w->buffer[x] + 1
137token_high_bit_not_set
138    rsb     r4, r6, #24                 ; 24-offset
139    ldr     r10, [r0, #vp8_writer_buffer]
140    lsr     r7, r2, r4                  ; lowvalue >> (24-offset)
141    ldr     r4, [r0, #vp8_writer_pos]   ; w->pos
142    lsl     r2, r2, r6                  ; lowvalue <<= offset
143    mov     r6, r3                      ; shift = count
144    add     r11, r4, #1                 ; w->pos++
145    bic     r2, r2, #0xff000000         ; lowvalue &= 0xffffff
146    str     r11, [r0, #vp8_writer_pos]
147    sub     r3, r3, #8                  ; count -= 8
148
149    VALIDATE_POS r10, r11               ; validate_buffer at pos
150
151    strb    r7, [r10, r4]               ; w->buffer[w->pos++]
152
153    ; r10 is used earlier in the loop, but r10 is used as
154    ; temp variable here.  So after r10 is used, reload
155    ; vp8_coef_tree_dcd into r10
156    ldr     r10, [sp, #60]              ; vp8_coef_tree
157
158token_count_lt_zero
159    lsl     r2, r2, r6                  ; lowvalue <<= shift
160
161    subs    r8, r8, #1                  ; --n
162    bne     token_loop
163
164    ldrb    r6, [r1, #tokenextra_token] ; t
165    ldr     r7, [sp, #56]               ; vp8_extra_bits
166    ; Add t * sizeof (vp8_extra_bit_struct) to get the desired
167    ;  element.  Here vp8_extra_bit_struct == 16
168    add     r12, r7, r6, lsl #4         ; b = vp8_extra_bits + t
169
170    ldr     r4, [r12, #vp8_extra_bit_struct_base_val]
171    cmp     r4, #0
172    beq     skip_extra_bits
173
174;   if( b->base_val)
175    ldr     r8, [r12, #vp8_extra_bit_struct_len] ; L
176    ldrsh   lr, [r1, #tokenextra_extra] ; e = p->Extra
177    cmp     r8, #0                      ; if( L)
178    beq     no_extra_bits
179
180    ldr     r9, [r12, #vp8_extra_bit_struct_prob]
181    asr     r7, lr, #1                  ; v=e>>1
182
183    ldr     r10, [r12, #vp8_extra_bit_struct_tree]
184    str     r10, [sp, #4]               ; b->tree
185
186    rsb     r4, r8, #32
187    lsl     r12, r7, r4
188
189    mov     lr, #0                      ; i = 0
190
191extra_bits_loop
192    ldrb    r4, [r9, lr, asr #1]            ; pp[i>>1]
193    sub     r7, r5, #1                  ; range-1
194    lsls    r12, r12, #1                ; v >> n
195    mul     r6, r4, r7                  ; (range-1) * pp[i>>1]
196    addcs   lr, lr, #1                  ; i + bb
197
198    mov     r7, #1
199    ldrsb   lr, [r10, lr]               ; i = b->tree[i+bb]
200    add     r4, r7, r6, lsr #8          ; split = 1 +  (((range-1) * pp[i>>1]) >> 8)
201
202    addcs   r2, r2, r4                  ; if  (bb) lowvalue += split
203    subcs   r4, r5, r4                  ; if  (bb) range = range-split
204
205    clz     r6, r4
206    sub     r6, r6, #24
207
208    adds    r3, r3, r6                  ; count += shift
209    lsl     r5, r4, r6                  ; range <<= shift
210    bmi     extra_count_lt_zero         ; if(count >= 0)
211
212    sub     r6, r6, r3                  ; offset= shift - count
213    sub     r4, r6, #1                  ; offset-1
214    lsls    r4, r2, r4                  ; if((lowvalue<<(offset-1)) & 0x80000000 )
215    bpl     extra_high_bit_not_set
216
217    ldr     r4, [r0, #vp8_writer_pos]   ; x
218    sub     r4, r4, #1                  ; x = w->pos - 1
219    b       extra_zero_while_start
220extra_zero_while_loop
221    mov     r10, #0
222    strb    r10, [r7, r4]               ; w->buffer[x] =(unsigned char)0
223    sub     r4, r4, #1                  ; x--
224extra_zero_while_start
225    cmp     r4, #0
226    ldrge   r7, [r0, #vp8_writer_buffer]
227    ldrb    r11, [r7, r4]
228    cmpge   r11, #0xff
229    beq     extra_zero_while_loop
230
231    ldr     r7, [r0, #vp8_writer_buffer]
232    ldrb    r10, [r7, r4]
233    add     r10, r10, #1
234    strb    r10, [r7, r4]
235extra_high_bit_not_set
236    rsb     r4, r6, #24                 ; 24-offset
237    ldr     r10, [r0, #vp8_writer_buffer]
238    lsr     r7, r2, r4                  ; lowvalue >> (24-offset)
239    ldr     r4, [r0, #vp8_writer_pos]
240    lsl     r2, r2, r6                  ; lowvalue <<= offset
241    mov     r6, r3                      ; shift = count
242    add     r11, r4, #1                 ; w->pos++
243    bic     r2, r2, #0xff000000         ; lowvalue &= 0xffffff
244    str     r11, [r0, #vp8_writer_pos]
245    sub     r3, r3, #8                  ; count -= 8
246
247    VALIDATE_POS r10, r11               ; validate_buffer at pos
248
249    strb    r7, [r10, r4]               ; w->buffer[w->pos++]=(lowvalue >> (24-offset))
250    ldr     r10, [sp, #4]               ; b->tree
251extra_count_lt_zero
252    lsl     r2, r2, r6
253
254    subs    r8, r8, #1                  ; --n
255    bne     extra_bits_loop             ; while (n)
256
257no_extra_bits
258    ldr     lr, [r1, #4]                ; e = p->Extra
259    add     r4, r5, #1                  ; range + 1
260    tst     lr, #1
261    lsr     r4, r4, #1                  ; split = (range + 1) >> 1
262    addne   r2, r2, r4                  ; lowvalue += split
263    subne   r4, r5, r4                  ; range = range-split
264    tst     r2, #0x80000000             ; lowvalue & 0x80000000
265    lsl     r5, r4, #1                  ; range <<= 1
266    beq     end_high_bit_not_set
267
268    ldr     r4, [r0, #vp8_writer_pos]
269    mov     r7, #0
270    sub     r4, r4, #1
271    b       end_zero_while_start
272end_zero_while_loop
273    strb    r7, [r6, r4]
274    sub     r4, r4, #1                  ; x--
275end_zero_while_start
276    cmp     r4, #0
277    ldrge   r6, [r0, #vp8_writer_buffer]
278    ldrb    r12, [r6, r4]
279    cmpge   r12, #0xff
280    beq     end_zero_while_loop
281
282    ldr     r6, [r0, #vp8_writer_buffer]
283    ldrb    r7, [r6, r4]
284    add     r7, r7, #1
285    strb    r7, [r6, r4]
286end_high_bit_not_set
287    adds    r3, r3, #1                  ; ++count
288    lsl     r2, r2, #1                  ; lowvalue  <<= 1
289    bne     end_count_zero
290
291    ldr     r4, [r0, #vp8_writer_pos]
292    mvn     r3, #7
293    ldr     r7, [r0, #vp8_writer_buffer]
294    lsr     r6, r2, #24                 ; lowvalue >> 24
295    add     r12, r4, #1                 ; w->pos++
296    bic     r2, r2, #0xff000000         ; lowvalue &= 0xffffff
297    str     r12, [r0, #vp8_writer_pos]
298
299    VALIDATE_POS r7, r12               ; validate_buffer at pos
300
301    strb    r6, [r7, r4]
302end_count_zero
303skip_extra_bits
304    add     r1, r1, #TOKENEXTRA_SZ      ; ++p
305check_p_lt_stop
306    ldr     r4, [sp, #0]                ; stop
307    cmp     r1, r4                      ; while( p < stop)
308    bcc     while_p_lt_stop
309
310    str     r2, [r0, #vp8_writer_lowvalue]
311    str     r5, [r0, #vp8_writer_range]
312    str     r3, [r0, #vp8_writer_count]
313    add     sp, sp, #16
314    pop     {r4-r12, pc}
315    ENDP
316
317    END
318