Set cpu_used range to [-16, 16] in real-time mode
[libvpx.git] / vp8 / encoder / arm / armv5te / vp8_packtokens_armv5.asm
blobd939287ff0b65cd93a7e79d955164afd30e1053f
2 ; Copyright (c) 2010 The WebM project authors. All Rights Reserved.
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.
12 EXPORT |vp8cx_pack_tokens_armv5|
14 INCLUDE asm_enc_offsets.asm
16 ARM
17 REQUIRE8
18 PRESERVE8
20 AREA |.text|, CODE, READONLY
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_armv5| PROC
29 push {r4-r11, lr}
31 ; Add size of xcount * sizeof (TOKENEXTRA) to get stop
32 ; sizeof (TOKENEXTRA) is 8
33 sub sp, sp, #12
34 add r2, r1, r2, lsl #3 ; stop = p + xcount*sizeof(TOKENEXTRA)
35 str r2, [sp, #0]
36 str r3, [sp, #8] ; save vp8_coef_encodings
37 ldr r2, [r0, #vp8_writer_lowvalue]
38 ldr r5, [r0, #vp8_writer_range]
39 ldr r3, [r0, #vp8_writer_count]
40 b check_p_lt_stop
42 while_p_lt_stop
43 ldrb r6, [r1, #tokenextra_token] ; t
44 ldr r4, [sp, #8] ; vp8_coef_encodings
45 mov lr, #0
46 add r4, r4, r6, lsl #3 ; a = vp8_coef_encodings + t
47 ldr r9, [r1, #tokenextra_context_tree] ; pp
49 ldrb r7, [r1, #tokenextra_skip_eob_node]
51 ldr r6, [r4, #vp8_token_value] ; v
52 ldr r8, [r4, #vp8_token_len] ; n
54 ; vp8 specific skip_eob_node
55 cmp r7, #0
56 movne lr, #2 ; i = 2
57 subne r8, r8, #1 ; --n
59 rsb r4, r8, #32 ; 32-n
60 ldr r10, [sp, #52] ; vp8_coef_tree
62 ; v is kept in r12 during the token pack loop
63 lsl r12, r6, r4 ; r12 = v << 32 - n
65 ; loop start
66 token_loop
67 ldrb r4, [r9, lr, asr #1] ; pp [i>>1]
68 sub r7, r5, #1 ; range-1
70 ; Decisions are made based on the bit value shifted
71 ; off of v, so set a flag here based on this.
72 ; This value is refered to as "bb"
73 lsls r12, r12, #1 ; bb = v >> n
74 mul r4, r4, r7 ; ((range-1) * pp[i>>1]))
76 ; bb can only be 0 or 1. So only execute this statement
77 ; if bb == 1, otherwise it will act like i + 0
78 addcs lr, lr, #1 ; i + bb
80 mov r7, #1
81 ldrsb lr, [r10, lr] ; i = vp8_coef_tree[i+bb]
82 add r4, r7, r4, lsr #8 ; 1 + (((range-1) * pp[i>>1]) >> 8)
84 addcs r2, r2, r4 ; if (bb) lowvalue += split
85 subcs r4, r5, r4 ; if (bb) range = range-split
87 ; Counting the leading zeros is used to normalize range.
88 clz r6, r4
89 sub r6, r6, #24 ; shift
91 ; Flag is set on the sum of count. This flag is used later
92 ; to determine if count >= 0
93 adds r3, r3, r6 ; count += shift
94 lsl r5, r4, r6 ; range <<= shift
95 bmi token_count_lt_zero ; if(count >= 0)
97 sub r6, r6, r3 ; offset = shift - count
98 sub r4, r6, #1 ; offset-1
99 lsls r4, r2, r4 ; if((lowvalue<<(offset-1)) & 0x80000000 )
100 bpl token_high_bit_not_set
102 ldr r4, [r0, #vp8_writer_pos] ; x
103 sub r4, r4, #1 ; x = w->pos-1
104 b token_zero_while_start
105 token_zero_while_loop
106 mov r10, #0
107 strb r10, [r7, r4] ; w->buffer[x] =(unsigned char)0
108 sub r4, r4, #1 ; x--
109 token_zero_while_start
110 cmp r4, #0
111 ldrge r7, [r0, #vp8_writer_buffer]
112 ldrb r11, [r7, r4]
113 cmpge r11, #0xff
114 beq token_zero_while_loop
116 ldr r7, [r0, #vp8_writer_buffer]
117 ldrb r10, [r7, r4] ; w->buffer[x]
118 add r10, r10, #1
119 strb r10, [r7, r4] ; w->buffer[x] + 1
120 token_high_bit_not_set
121 rsb r4, r6, #24 ; 24-offset
122 ldr r10, [r0, #vp8_writer_buffer]
123 lsr r7, r2, r4 ; lowvalue >> (24-offset)
124 ldr r4, [r0, #vp8_writer_pos] ; w->pos
125 lsl r2, r2, r6 ; lowvalue <<= offset
126 mov r6, r3 ; shift = count
127 add r11, r4, #1 ; w->pos++
128 bic r2, r2, #0xff000000 ; lowvalue &= 0xffffff
129 str r11, [r0, #vp8_writer_pos]
130 sub r3, r3, #8 ; count -= 8
131 strb r7, [r10, r4] ; w->buffer[w->pos++]
133 ; r10 is used earlier in the loop, but r10 is used as
134 ; temp variable here. So after r10 is used, reload
135 ; vp8_coef_tree_dcd into r10
136 ldr r10, [sp, #52] ; vp8_coef_tree
138 token_count_lt_zero
139 lsl r2, r2, r6 ; lowvalue <<= shift
141 subs r8, r8, #1 ; --n
142 bne token_loop
144 ldrb r6, [r1, #tokenextra_token] ; t
145 ldr r7, [sp, #48] ; vp8_extra_bits
146 ; Add t * sizeof (vp8_extra_bit_struct) to get the desired
147 ; element. Here vp8_extra_bit_struct == 16
148 add r12, r7, r6, lsl #4 ; b = vp8_extra_bits + t
150 ldr r4, [r12, #vp8_extra_bit_struct_base_val]
151 cmp r4, #0
152 beq skip_extra_bits
154 ; if( b->base_val)
155 ldr r8, [r12, #vp8_extra_bit_struct_len] ; L
156 ldrsh lr, [r1, #tokenextra_extra] ; e = p->Extra
157 cmp r8, #0 ; if( L)
158 beq no_extra_bits
160 ldr r9, [r12, #vp8_extra_bit_struct_prob]
161 asr r7, lr, #1 ; v=e>>1
163 ldr r10, [r12, #vp8_extra_bit_struct_tree]
164 str r10, [sp, #4] ; b->tree
166 rsb r4, r8, #32
167 lsl r12, r7, r4
169 mov lr, #0 ; i = 0
171 extra_bits_loop
172 ldrb r4, [r9, lr, asr #1] ; pp[i>>1]
173 sub r7, r5, #1 ; range-1
174 lsls r12, r12, #1 ; v >> n
175 mul r4, r4, r7 ; (range-1) * pp[i>>1]
176 addcs lr, lr, #1 ; i + bb
178 mov r7, #1
179 ldrsb lr, [r10, lr] ; i = b->tree[i+bb]
180 add r4, r7, r4, lsr #8 ; split = 1 + (((range-1) * pp[i>>1]) >> 8)
182 addcs r2, r2, r4 ; if (bb) lowvalue += split
183 subcs r4, r5, r4 ; if (bb) range = range-split
185 clz r6, r4
186 sub r6, r6, #24
188 adds r3, r3, r6 ; count += shift
189 lsl r5, r4, r6 ; range <<= shift
190 bmi extra_count_lt_zero ; if(count >= 0)
192 sub r6, r6, r3 ; offset= shift - count
193 sub r4, r6, #1 ; offset-1
194 lsls r4, r2, r4 ; if((lowvalue<<(offset-1)) & 0x80000000 )
195 bpl extra_high_bit_not_set
197 ldr r4, [r0, #vp8_writer_pos] ; x
198 sub r4, r4, #1 ; x = w->pos - 1
199 b extra_zero_while_start
200 extra_zero_while_loop
201 mov r10, #0
202 strb r10, [r7, r4] ; w->buffer[x] =(unsigned char)0
203 sub r4, r4, #1 ; x--
204 extra_zero_while_start
205 cmp r4, #0
206 ldrge r7, [r0, #vp8_writer_buffer]
207 ldrb r11, [r7, r4]
208 cmpge r11, #0xff
209 beq extra_zero_while_loop
211 ldr r7, [r0, #vp8_writer_buffer]
212 ldrb r10, [r7, r4]
213 add r10, r10, #1
214 strb r10, [r7, r4]
215 extra_high_bit_not_set
216 rsb r4, r6, #24 ; 24-offset
217 ldr r10, [r0, #vp8_writer_buffer]
218 lsr r7, r2, r4 ; lowvalue >> (24-offset)
219 ldr r4, [r0, #vp8_writer_pos]
220 lsl r2, r2, r6 ; lowvalue <<= offset
221 mov r6, r3 ; shift = count
222 add r11, r4, #1 ; w->pos++
223 bic r2, r2, #0xff000000 ; lowvalue &= 0xffffff
224 str r11, [r0, #vp8_writer_pos]
225 sub r3, r3, #8 ; count -= 8
226 strb r7, [r10, r4] ; w->buffer[w->pos++]=(lowvalue >> (24-offset))
227 ldr r10, [sp, #4] ; b->tree
228 extra_count_lt_zero
229 lsl r2, r2, r6
231 subs r8, r8, #1 ; --n
232 bne extra_bits_loop ; while (n)
234 no_extra_bits
235 ldr lr, [r1, #4] ; e = p->Extra
236 add r4, r5, #1 ; range + 1
237 tst lr, #1
238 lsr r4, r4, #1 ; split = (range + 1) >> 1
239 addne r2, r2, r4 ; lowvalue += split
240 subne r4, r5, r4 ; range = range-split
241 tst r2, #0x80000000 ; lowvalue & 0x80000000
242 lsl r5, r4, #1 ; range <<= 1
243 beq end_high_bit_not_set
245 ldr r4, [r0, #vp8_writer_pos]
246 mov r7, #0
247 sub r4, r4, #1
248 b end_zero_while_start
249 end_zero_while_loop
250 strb r7, [r6, r4]
251 sub r4, r4, #1 ; x--
252 end_zero_while_start
253 cmp r4, #0
254 ldrge r6, [r0, #vp8_writer_buffer]
255 ldrb r12, [r6, r4]
256 cmpge r12, #0xff
257 beq end_zero_while_loop
259 ldr r6, [r0, #vp8_writer_buffer]
260 ldrb r7, [r6, r4]
261 add r7, r7, #1
262 strb r7, [r6, r4]
263 end_high_bit_not_set
264 adds r3, r3, #1 ; ++count
265 lsl r2, r2, #1 ; lowvalue <<= 1
266 bne end_count_zero
268 ldr r4, [r0, #vp8_writer_pos]
269 mvn r3, #7
270 ldr r7, [r0, #vp8_writer_buffer]
271 lsr r6, r2, #24 ; lowvalue >> 24
272 add r12, r4, #1 ; w->pos++
273 bic r2, r2, #0xff000000 ; lowvalue &= 0xffffff
274 str r12, [r0, #0x10]
275 strb r6, [r7, r4]
276 end_count_zero
277 skip_extra_bits
278 add r1, r1, #TOKENEXTRA_SZ ; ++p
279 check_p_lt_stop
280 ldr r4, [sp, #0] ; stop
281 cmp r1, r4 ; while( p < stop)
282 bcc while_p_lt_stop
284 str r2, [r0, #vp8_writer_lowvalue]
285 str r5, [r0, #vp8_writer_range]
286 str r3, [r0, #vp8_writer_count]
287 add sp, sp, #12
288 pop {r4-r11, pc}
289 ENDP