• file
• revisions
• annotate
• diff
• comparison
• raw

media/libvpx/vp8/encoder/arm/armv5te/vp8_packtokens_armv5.asm@b8a032363ba2 (annotated)

media/libvpx/vp8/encoder/arm/armv5te/vp8_packtokens_armv5.asm

Thu, 22 Jan 2015 13:21:57 +0100

author

Michael Schloh von Bennewitz <michael@schloh.com>

date

Thu, 22 Jan 2015 13:21:57 +0100

branch

TOR_BUG_9701

changeset 15

b8a032363ba2

permissions

-rw-r--r--

Incorporate requested changes from Mozilla in review:
https://bugzilla.mozilla.org/show_bug.cgi?id=1123480#c6

michael@0	1	;
michael@0	2	; Copyright (c) 2010 The WebM project authors. All Rights Reserved.
michael@0	3	;
michael@0	4	; Use of this source code is governed by a BSD-style license
michael@0	5	; that can be found in the LICENSE file in the root of the source
michael@0	6	; tree. An additional intellectual property rights grant can be found
michael@0	7	; in the file PATENTS. All contributing project authors may
michael@0	8	; be found in the AUTHORS file in the root of the source tree.
michael@0	9	;
michael@0	10
michael@0	11
michael@0	12	EXPORT |vp8cx_pack_tokens_armv5|
michael@0	13	IMPORT |vp8_validate_buffer_arm|
michael@0	14
michael@0	15	INCLUDE vp8_asm_enc_offsets.asm
michael@0	16
michael@0	17	ARM
michael@0	18	REQUIRE8
michael@0	19	PRESERVE8
michael@0	20
michael@0	21	AREA |.text|, CODE, READONLY
michael@0	22
michael@0	23
michael@0	24	; macro for validating write buffer position
michael@0	25	; needs vp8_writer in r0
michael@0	26	; start shall not be in r1
michael@0	27	MACRO
michael@0	28	VALIDATE_POS $start, $pos
michael@0	29	push {r0-r3, r12, lr} ; rest of regs are preserved by subroutine call
michael@0	30	ldr r2, [r0, #vp8_writer_buffer_end]
michael@0	31	ldr r3, [r0, #vp8_writer_error]
michael@0	32	mov r1, $pos
michael@0	33	mov r0, $start
michael@0	34	bl vp8_validate_buffer_arm
michael@0	35	pop {r0-r3, r12, lr}
michael@0	36	MEND
michael@0	37
michael@0	38
michael@0	39	; r0 vp8_writer *w
michael@0	40	; r1 const TOKENEXTRA *p
michael@0	41	; r2 int xcount
michael@0	42	; r3 vp8_coef_encodings
michael@0	43	; s0 vp8_extra_bits
michael@0	44	; s1 vp8_coef_tree
michael@0	45	|vp8cx_pack_tokens_armv5| PROC
michael@0	46	push {r4-r12, lr}
michael@0	47	sub sp, sp, #16
michael@0	48
michael@0	49	; Add size of xcount * sizeof (TOKENEXTRA) to get stop
michael@0	50	; sizeof (TOKENEXTRA) is 8
michael@0	51	add r2, r1, r2, lsl #3 ; stop = p + xcount*sizeof(TOKENEXTRA)
michael@0	52	str r2, [sp, #0]
michael@0	53	str r3, [sp, #8] ; save vp8_coef_encodings
michael@0	54	ldr r2, [r0, #vp8_writer_lowvalue]
michael@0	55	ldr r5, [r0, #vp8_writer_range]
michael@0	56	ldr r3, [r0, #vp8_writer_count]
michael@0	57	b check_p_lt_stop
michael@0	58
michael@0	59	while_p_lt_stop
michael@0	60	ldrb r6, [r1, #tokenextra_token] ; t
michael@0	61	ldr r4, [sp, #8] ; vp8_coef_encodings
michael@0	62	mov lr, #0
michael@0	63	add r4, r4, r6, lsl #3 ; a = vp8_coef_encodings + t
michael@0	64	ldr r9, [r1, #tokenextra_context_tree] ; pp
michael@0	65
michael@0	66	ldrb r7, [r1, #tokenextra_skip_eob_node]
michael@0	67
michael@0	68	ldr r6, [r4, #vp8_token_value] ; v
michael@0	69	ldr r8, [r4, #vp8_token_len] ; n
michael@0	70
michael@0	71	; vp8 specific skip_eob_node
michael@0	72	cmp r7, #0
michael@0	73	movne lr, #2 ; i = 2
michael@0	74	subne r8, r8, #1 ; --n
michael@0	75
michael@0	76	rsb r4, r8, #32 ; 32-n
michael@0	77	ldr r10, [sp, #60] ; vp8_coef_tree
michael@0	78
michael@0	79	; v is kept in r12 during the token pack loop
michael@0	80	lsl r12, r6, r4 ; r12 = v << 32 - n
michael@0	81
michael@0	82	; loop start
michael@0	83	token_loop
michael@0	84	ldrb r4, [r9, lr, asr #1] ; pp [i>>1]
michael@0	85	sub r7, r5, #1 ; range-1
michael@0	86
michael@0	87	; Decisions are made based on the bit value shifted
michael@0	88	; off of v, so set a flag here based on this.
michael@0	89	; This value is refered to as "bb"
michael@0	90	lsls r12, r12, #1 ; bb = v >> n
michael@0	91	mul r6, r4, r7 ; ((range-1) * pp[i>>1]))
michael@0	92
michael@0	93	; bb can only be 0 or 1. So only execute this statement
michael@0	94	; if bb == 1, otherwise it will act like i + 0
michael@0	95	addcs lr, lr, #1 ; i + bb
michael@0	96
michael@0	97	mov r7, #1
michael@0	98	ldrsb lr, [r10, lr] ; i = vp8_coef_tree[i+bb]
michael@0	99	add r4, r7, r6, lsr #8 ; 1 + (((range-1) * pp[i>>1]) >> 8)
michael@0	100
michael@0	101	addcs r2, r2, r4 ; if (bb) lowvalue += split
michael@0	102	subcs r4, r5, r4 ; if (bb) range = range-split
michael@0	103
michael@0	104	; Counting the leading zeros is used to normalize range.
michael@0	105	clz r6, r4
michael@0	106	sub r6, r6, #24 ; shift
michael@0	107
michael@0	108	; Flag is set on the sum of count. This flag is used later
michael@0	109	; to determine if count >= 0
michael@0	110	adds r3, r3, r6 ; count += shift
michael@0	111	lsl r5, r4, r6 ; range <<= shift
michael@0	112	bmi token_count_lt_zero ; if(count >= 0)
michael@0	113
michael@0	114	sub r6, r6, r3 ; offset = shift - count
michael@0	115	sub r4, r6, #1 ; offset-1
michael@0	116	lsls r4, r2, r4 ; if((lowvalue<<(offset-1)) & 0x80000000 )
michael@0	117	bpl token_high_bit_not_set
michael@0	118
michael@0	119	ldr r4, [r0, #vp8_writer_pos] ; x
michael@0	120	sub r4, r4, #1 ; x = w->pos-1
michael@0	121	b token_zero_while_start
michael@0	122	token_zero_while_loop
michael@0	123	mov r10, #0
michael@0	124	strb r10, [r7, r4] ; w->buffer[x] =(unsigned char)0
michael@0	125	sub r4, r4, #1 ; x--
michael@0	126	token_zero_while_start
michael@0	127	cmp r4, #0
michael@0	128	ldrge r7, [r0, #vp8_writer_buffer]
michael@0	129	ldrb r11, [r7, r4]
michael@0	130	cmpge r11, #0xff
michael@0	131	beq token_zero_while_loop
michael@0	132
michael@0	133	ldr r7, [r0, #vp8_writer_buffer]
michael@0	134	ldrb r10, [r7, r4] ; w->buffer[x]
michael@0	135	add r10, r10, #1
michael@0	136	strb r10, [r7, r4] ; w->buffer[x] + 1
michael@0	137	token_high_bit_not_set
michael@0	138	rsb r4, r6, #24 ; 24-offset
michael@0	139	ldr r10, [r0, #vp8_writer_buffer]
michael@0	140	lsr r7, r2, r4 ; lowvalue >> (24-offset)
michael@0	141	ldr r4, [r0, #vp8_writer_pos] ; w->pos
michael@0	142	lsl r2, r2, r6 ; lowvalue <<= offset
michael@0	143	mov r6, r3 ; shift = count
michael@0	144	add r11, r4, #1 ; w->pos++
michael@0	145	bic r2, r2, #0xff000000 ; lowvalue &= 0xffffff
michael@0	146	str r11, [r0, #vp8_writer_pos]
michael@0	147	sub r3, r3, #8 ; count -= 8
michael@0	148
michael@0	149	VALIDATE_POS r10, r11 ; validate_buffer at pos
michael@0	150
michael@0	151	strb r7, [r10, r4] ; w->buffer[w->pos++]
michael@0	152
michael@0	153	; r10 is used earlier in the loop, but r10 is used as
michael@0	154	; temp variable here. So after r10 is used, reload
michael@0	155	; vp8_coef_tree_dcd into r10
michael@0	156	ldr r10, [sp, #60] ; vp8_coef_tree
michael@0	157
michael@0	158	token_count_lt_zero
michael@0	159	lsl r2, r2, r6 ; lowvalue <<= shift
michael@0	160
michael@0	161	subs r8, r8, #1 ; --n
michael@0	162	bne token_loop
michael@0	163
michael@0	164	ldrb r6, [r1, #tokenextra_token] ; t
michael@0	165	ldr r7, [sp, #56] ; vp8_extra_bits
michael@0	166	; Add t * sizeof (vp8_extra_bit_struct) to get the desired
michael@0	167	; element. Here vp8_extra_bit_struct == 16
michael@0	168	add r12, r7, r6, lsl #4 ; b = vp8_extra_bits + t
michael@0	169
michael@0	170	ldr r4, [r12, #vp8_extra_bit_struct_base_val]
michael@0	171	cmp r4, #0
michael@0	172	beq skip_extra_bits
michael@0	173
michael@0	174	; if( b->base_val)
michael@0	175	ldr r8, [r12, #vp8_extra_bit_struct_len] ; L
michael@0	176	ldrsh lr, [r1, #tokenextra_extra] ; e = p->Extra
michael@0	177	cmp r8, #0 ; if( L)
michael@0	178	beq no_extra_bits
michael@0	179
michael@0	180	ldr r9, [r12, #vp8_extra_bit_struct_prob]
michael@0	181	asr r7, lr, #1 ; v=e>>1
michael@0	182
michael@0	183	ldr r10, [r12, #vp8_extra_bit_struct_tree]
michael@0	184	str r10, [sp, #4] ; b->tree
michael@0	185
michael@0	186	rsb r4, r8, #32
michael@0	187	lsl r12, r7, r4
michael@0	188
michael@0	189	mov lr, #0 ; i = 0
michael@0	190
michael@0	191	extra_bits_loop
michael@0	192	ldrb r4, [r9, lr, asr #1] ; pp[i>>1]
michael@0	193	sub r7, r5, #1 ; range-1
michael@0	194	lsls r12, r12, #1 ; v >> n
michael@0	195	mul r6, r4, r7 ; (range-1) * pp[i>>1]
michael@0	196	addcs lr, lr, #1 ; i + bb
michael@0	197
michael@0	198	mov r7, #1
michael@0	199	ldrsb lr, [r10, lr] ; i = b->tree[i+bb]
michael@0	200	add r4, r7, r6, lsr #8 ; split = 1 + (((range-1) * pp[i>>1]) >> 8)
michael@0	201
michael@0	202	addcs r2, r2, r4 ; if (bb) lowvalue += split
michael@0	203	subcs r4, r5, r4 ; if (bb) range = range-split
michael@0	204
michael@0	205	clz r6, r4
michael@0	206	sub r6, r6, #24
michael@0	207
michael@0	208	adds r3, r3, r6 ; count += shift
michael@0	209	lsl r5, r4, r6 ; range <<= shift
michael@0	210	bmi extra_count_lt_zero ; if(count >= 0)
michael@0	211
michael@0	212	sub r6, r6, r3 ; offset= shift - count
michael@0	213	sub r4, r6, #1 ; offset-1
michael@0	214	lsls r4, r2, r4 ; if((lowvalue<<(offset-1)) & 0x80000000 )
michael@0	215	bpl extra_high_bit_not_set
michael@0	216
michael@0	217	ldr r4, [r0, #vp8_writer_pos] ; x
michael@0	218	sub r4, r4, #1 ; x = w->pos - 1
michael@0	219	b extra_zero_while_start
michael@0	220	extra_zero_while_loop
michael@0	221	mov r10, #0
michael@0	222	strb r10, [r7, r4] ; w->buffer[x] =(unsigned char)0
michael@0	223	sub r4, r4, #1 ; x--
michael@0	224	extra_zero_while_start
michael@0	225	cmp r4, #0
michael@0	226	ldrge r7, [r0, #vp8_writer_buffer]
michael@0	227	ldrb r11, [r7, r4]
michael@0	228	cmpge r11, #0xff
michael@0	229	beq extra_zero_while_loop
michael@0	230
michael@0	231	ldr r7, [r0, #vp8_writer_buffer]
michael@0	232	ldrb r10, [r7, r4]
michael@0	233	add r10, r10, #1
michael@0	234	strb r10, [r7, r4]
michael@0	235	extra_high_bit_not_set
michael@0	236	rsb r4, r6, #24 ; 24-offset
michael@0	237	ldr r10, [r0, #vp8_writer_buffer]
michael@0	238	lsr r7, r2, r4 ; lowvalue >> (24-offset)
michael@0	239	ldr r4, [r0, #vp8_writer_pos]
michael@0	240	lsl r2, r2, r6 ; lowvalue <<= offset
michael@0	241	mov r6, r3 ; shift = count
michael@0	242	add r11, r4, #1 ; w->pos++
michael@0	243	bic r2, r2, #0xff000000 ; lowvalue &= 0xffffff
michael@0	244	str r11, [r0, #vp8_writer_pos]
michael@0	245	sub r3, r3, #8 ; count -= 8
michael@0	246
michael@0	247	VALIDATE_POS r10, r11 ; validate_buffer at pos
michael@0	248
michael@0	249	strb r7, [r10, r4] ; w->buffer[w->pos++]=(lowvalue >> (24-offset))
michael@0	250	ldr r10, [sp, #4] ; b->tree
michael@0	251	extra_count_lt_zero
michael@0	252	lsl r2, r2, r6
michael@0	253
michael@0	254	subs r8, r8, #1 ; --n
michael@0	255	bne extra_bits_loop ; while (n)
michael@0	256
michael@0	257	no_extra_bits
michael@0	258	ldr lr, [r1, #4] ; e = p->Extra
michael@0	259	add r4, r5, #1 ; range + 1
michael@0	260	tst lr, #1
michael@0	261	lsr r4, r4, #1 ; split = (range + 1) >> 1
michael@0	262	addne r2, r2, r4 ; lowvalue += split
michael@0	263	subne r4, r5, r4 ; range = range-split
michael@0	264	tst r2, #0x80000000 ; lowvalue & 0x80000000
michael@0	265	lsl r5, r4, #1 ; range <<= 1
michael@0	266	beq end_high_bit_not_set
michael@0	267
michael@0	268	ldr r4, [r0, #vp8_writer_pos]
michael@0	269	mov r7, #0
michael@0	270	sub r4, r4, #1
michael@0	271	b end_zero_while_start
michael@0	272	end_zero_while_loop
michael@0	273	strb r7, [r6, r4]
michael@0	274	sub r4, r4, #1 ; x--
michael@0	275	end_zero_while_start
michael@0	276	cmp r4, #0
michael@0	277	ldrge r6, [r0, #vp8_writer_buffer]
michael@0	278	ldrb r12, [r6, r4]
michael@0	279	cmpge r12, #0xff
michael@0	280	beq end_zero_while_loop
michael@0	281
michael@0	282	ldr r6, [r0, #vp8_writer_buffer]
michael@0	283	ldrb r7, [r6, r4]
michael@0	284	add r7, r7, #1
michael@0	285	strb r7, [r6, r4]
michael@0	286	end_high_bit_not_set
michael@0	287	adds r3, r3, #1 ; ++count
michael@0	288	lsl r2, r2, #1 ; lowvalue <<= 1
michael@0	289	bne end_count_zero
michael@0	290
michael@0	291	ldr r4, [r0, #vp8_writer_pos]
michael@0	292	mvn r3, #7
michael@0	293	ldr r7, [r0, #vp8_writer_buffer]
michael@0	294	lsr r6, r2, #24 ; lowvalue >> 24
michael@0	295	add r12, r4, #1 ; w->pos++
michael@0	296	bic r2, r2, #0xff000000 ; lowvalue &= 0xffffff
michael@0	297	str r12, [r0, #vp8_writer_pos]
michael@0	298
michael@0	299	VALIDATE_POS r7, r12 ; validate_buffer at pos
michael@0	300
michael@0	301	strb r6, [r7, r4]
michael@0	302	end_count_zero
michael@0	303	skip_extra_bits
michael@0	304	add r1, r1, #TOKENEXTRA_SZ ; ++p
michael@0	305	check_p_lt_stop
michael@0	306	ldr r4, [sp, #0] ; stop
michael@0	307	cmp r1, r4 ; while( p < stop)
michael@0	308	bcc while_p_lt_stop
michael@0	309
michael@0	310	str r2, [r0, #vp8_writer_lowvalue]
michael@0	311	str r5, [r0, #vp8_writer_range]
michael@0	312	str r3, [r0, #vp8_writer_count]
michael@0	313	add sp, sp, #16
michael@0	314	pop {r4-r12, pc}
michael@0	315	ENDP
michael@0	316
michael@0	317	END