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_mb_row_tokens_armv5|
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_COMP *cpi
23; r1 vp8_writer *w
24; r2 vp8_coef_encodings
25; r3 vp8_extra_bits
26; s0 vp8_coef_tree
27
28|vp8cx_pack_mb_row_tokens_armv5| PROC
29    push    {r4-r11, lr}
30    sub     sp, sp, #24
31
32    ; Compute address of cpi->common.mb_rows
33    ldr     r4, _VP8_COMP_common_
34    ldr     r6, _VP8_COMMON_MBrows_
35    add     r4, r0, r4
36
37    ldr     r5, [r4, r6]                ; load up mb_rows
38
39    str     r2, [sp, #20]               ; save vp8_coef_encodings
40    str     r5, [sp, #12]               ; save mb_rows
41    str     r3, [sp, #8]                ; save vp8_extra_bits
42
43    ldr     r4, _VP8_COMP_tplist_
44    add     r4, r0, r4
45    ldr     r7, [r4, #0]                ; dereference cpi->tp_list
46
47    mov     r0, r1                      ; keep same as other loops
48
49    ldr     r2, [r0, #vp8_writer_lowvalue]
50    ldr     r5, [r0, #vp8_writer_range]
51    ldr     r3, [r0, #vp8_writer_count]
52
53mb_row_loop
54
55    ldr     r1, [r7, #tokenlist_start]
56    ldr     r9, [r7, #tokenlist_stop]
57    str     r9, [sp, #0]                ; save stop for later comparison
58    str     r7, [sp, #16]               ; tokenlist address for next time
59
60    b       check_p_lt_stop
61
62    ; actuall work gets done here!
63
64while_p_lt_stop
65    ldr     r6, [r1, #tokenextra_token] ; t
66    ldr     r4, [sp, #20]               ; vp8_coef_encodings
67    mov     lr, #0
68    add     r4, r4, r6, lsl #3          ; a = vp8_coef_encodings + t
69    ldr     r9, [r1, #tokenextra_context_tree]   ; pp
70
71    ldr     r7, [r1, #tokenextra_skip_eob_node]
72
73    ldr     r6, [r4, #vp8_token_value]  ; v
74    ldr     r8, [r4, #vp8_token_len]    ; n
75
76    ; vp8 specific skip_eob_node
77    cmp     r7, #0
78    movne   lr, #2                      ; i = 2
79    subne   r8, r8, #1                  ; --n
80
81    rsb     r4, r8, #32                 ; 32-n
82    ldr     r10, [sp, #60]              ; vp8_coef_tree
83
84    ; v is kept in r12 during the token pack loop
85    lsl     r12, r6, r4                 ; r12 = v << 32 - n
86
87; loop start
88token_loop
89    ldrb    r4, [r9, lr, asr #1]        ; pp [i>>1]
90    sub     r7, r5, #1                  ; range-1
91
92    ; Decisions are made based on the bit value shifted
93    ; off of v, so set a flag here based on this.
94    ; This value is refered to as "bb"
95    lsls    r12, r12, #1                ; bb = v >> n
96    mul     r4, r4, r7                  ; ((range-1) * pp[i>>1]))
97
98    ; bb can only be 0 or 1.  So only execute this statement
99    ; if bb == 1, otherwise it will act like i + 0
100    addcs   lr, lr, #1                  ; i + bb
101
102    mov     r7, #1
103    ldrsb   lr, [r10, lr]               ; i = vp8_coef_tree[i+bb]
104    add     r4, r7, r4, lsr #8          ; 1 + (((range-1) * pp[i>>1]) >> 8)
105
106    addcs   r2, r2, r4                  ; if  (bb) lowvalue += split
107    subcs   r4, r5, r4                  ; if  (bb) range = range-split
108
109    ; Counting the leading zeros is used to normalize range.
110    clz     r6, r4
111    sub     r6, r6, #24                 ; shift
112
113    ; Flag is set on the sum of count.  This flag is used later
114    ; to determine if count >= 0
115    adds    r3, r3, r6                  ; count += shift
116    lsl     r5, r4, r6                  ; range <<= shift
117    bmi     token_count_lt_zero         ; if(count >= 0)
118
119    sub     r6, r6, r3                  ; offset = shift - count
120    sub     r4, r6, #1                  ; offset-1
121    lsls    r4, r2, r4                  ; if((lowvalue<<(offset-1)) & 0x80000000 )
122    bpl     token_high_bit_not_set
123
124    ldr     r4, [r0, #vp8_writer_pos]   ; x
125    sub     r4, r4, #1                  ; x = w->pos-1
126    b       token_zero_while_start
127token_zero_while_loop
128    mov     r10, #0
129    strb    r10, [r7, r4]               ; w->buffer[x] =(unsigned char)0
130    sub     r4, r4, #1                  ; x--
131token_zero_while_start
132    cmp     r4, #0
133    ldrge   r7, [r0, #vp8_writer_buffer]
134    ldrb    r11, [r7, r4]
135    cmpge   r11, #0xff
136    beq     token_zero_while_loop
137
138    ldr     r7, [r0, #vp8_writer_buffer]
139    ldrb    r10, [r7, r4]               ; w->buffer[x]
140    add     r10, r10, #1
141    strb    r10, [r7, r4]               ; w->buffer[x] + 1
142token_high_bit_not_set
143    rsb     r4, r6, #24                 ; 24-offset
144    ldr     r10, [r0, #vp8_writer_buffer]
145    lsr     r7, r2, r4                  ; lowvalue >> (24-offset)
146    ldr     r4, [r0, #vp8_writer_pos]   ; w->pos
147    lsl     r2, r2, r6                  ; lowvalue <<= offset
148    mov     r6, r3                      ; shift = count
149    add     r11, r4, #1                 ; w->pos++
150    bic     r2, r2, #0xff000000         ; lowvalue &= 0xffffff
151    str     r11, [r0, #vp8_writer_pos]
152    sub     r3, r3, #8                  ; count -= 8
153    strb    r7, [r10, r4]               ; w->buffer[w->pos++]
154
155    ; r10 is used earlier in the loop, but r10 is used as
156    ; temp variable here.  So after r10 is used, reload
157    ; vp8_coef_tree_dcd into r10
158    ldr     r10, [sp, #60]              ; vp8_coef_tree
159
160token_count_lt_zero
161    lsl     r2, r2, r6                  ; lowvalue <<= shift
162
163    subs    r8, r8, #1                  ; --n
164    bne     token_loop
165
166    ldr     r6, [r1, #tokenextra_token] ; t
167    ldr     r7, [sp, #8]                ; vp8_extra_bits
168    ; Add t * sizeof (vp8_extra_bit_struct) to get the desired
169    ;  element.  Here vp8_extra_bit_struct == 20
170    add     r6, r6, r6, lsl #2          ; b = vp8_extra_bits + t
171    add     r12, r7, r6, lsl #2         ; b = vp8_extra_bits + t
172
173    ldr     r4, [r12, #vp8_extra_bit_struct_base_val]
174    cmp     r4, #0
175    beq     skip_extra_bits
176
177;   if( b->base_val)
178    ldr     r8, [r12, #vp8_extra_bit_struct_len] ; L
179    ldr     lr, [r1, #tokenextra_extra] ; e = p->Extra
180    cmp     r8, #0                      ; if( L)
181    beq     no_extra_bits
182
183    ldr     r9, [r12, #vp8_extra_bit_struct_prob]
184    asr     r7, lr, #1                  ; v=e>>1
185
186    ldr     r10, [r12, #vp8_extra_bit_struct_tree]
187    str     r10, [sp, #4]               ; b->tree
188
189    rsb     r4, r8, #32
190    lsl     r12, r7, r4
191
192    mov     lr, #0                      ; i = 0
193
194extra_bits_loop
195    ldrb    r4, [r9, lr, asr #1]            ; pp[i>>1]
196    sub     r7, r5, #1                  ; range-1
197    lsls    r12, r12, #1                ; v >> n
198    mul     r4, r4, r7                  ; (range-1) * pp[i>>1]
199    addcs   lr, lr, #1                  ; i + bb
200
201    mov     r7, #1
202    ldrsb   lr, [r10, lr]               ; i = b->tree[i+bb]
203    add     r4, r7, r4, lsr #8          ; split = 1 +  (((range-1) * pp[i>>1]) >> 8)
204
205    addcs   r2, r2, r4                  ; if  (bb) lowvalue += split
206    subcs   r4, r5, r4                  ; if  (bb) range = range-split
207
208    clz     r6, r4
209    sub     r6, r6, #24
210
211    adds    r3, r3, r6                  ; count += shift
212    lsl     r5, r4, r6                  ; range <<= shift
213    bmi     extra_count_lt_zero         ; if(count >= 0)
214
215    sub     r6, r6, r3                  ; offset= shift - count
216    sub     r4, r6, #1                  ; offset-1
217    lsls    r4, r2, r4                  ; if((lowvalue<<(offset-1)) & 0x80000000 )
218    bpl     extra_high_bit_not_set
219
220    ldr     r4, [r0, #vp8_writer_pos]   ; x
221    sub     r4, r4, #1                  ; x = w->pos - 1
222    b       extra_zero_while_start
223extra_zero_while_loop
224    mov     r10, #0
225    strb    r10, [r7, r4]               ; w->buffer[x] =(unsigned char)0
226    sub     r4, r4, #1                  ; x--
227extra_zero_while_start
228    cmp     r4, #0
229    ldrge   r7, [r0, #vp8_writer_buffer]
230    ldrb    r11, [r7, r4]
231    cmpge   r11, #0xff
232    beq     extra_zero_while_loop
233
234    ldr     r7, [r0, #vp8_writer_buffer]
235    ldrb    r10, [r7, r4]
236    add     r10, r10, #1
237    strb    r10, [r7, r4]
238extra_high_bit_not_set
239    rsb     r4, r6, #24                 ; 24-offset
240    ldr     r10, [r0, #vp8_writer_buffer]
241    lsr     r7, r2, r4                  ; lowvalue >> (24-offset)
242    ldr     r4, [r0, #vp8_writer_pos]
243    lsl     r2, r2, r6                  ; lowvalue <<= offset
244    mov     r6, r3                      ; shift = count
245    add     r11, r4, #1                 ; w->pos++
246    bic     r2, r2, #0xff000000         ; lowvalue &= 0xffffff
247    str     r11, [r0, #vp8_writer_pos]
248    sub     r3, r3, #8                  ; count -= 8
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, #0x10]
298    strb    r6, [r7, r4]
299end_count_zero
300skip_extra_bits
301    add     r1, r1, #TOKENEXTRA_SZ      ; ++p
302check_p_lt_stop
303    ldr     r4, [sp, #0]                ; stop
304    cmp     r1, r4                      ; while( p < stop)
305    bcc     while_p_lt_stop
306
307    ldr     r6, [sp, #12]               ; mb_rows
308    ldr     r7, [sp, #16]               ; tokenlist address
309    subs    r6, r6, #1
310    add     r7, r7, #TOKENLIST_SZ       ; next element in the array
311    str     r6, [sp, #12]
312    bne     mb_row_loop
313
314    str     r2, [r0, #vp8_writer_lowvalue]
315    str     r5, [r0, #vp8_writer_range]
316    str     r3, [r0, #vp8_writer_count]
317    add     sp, sp, #24
318    pop     {r4-r11, pc}
319    ENDP
320
321_VP8_COMP_common_
322    DCD     vp8_comp_common
323_VP8_COMMON_MBrows_
324    DCD     vp8_common_mb_rows
325_VP8_COMP_tplist_
326    DCD     vp8_comp_tplist
327
328    END
329