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

media/libvpx/vp8/common/x86/mfqe_sse2.asm@ac0c01689b40 (annotated)

media/libvpx/vp8/common/x86/mfqe_sse2.asm

Thu, 15 Jan 2015 15:59:08 +0100

author

Michael Schloh von Bennewitz <michael@schloh.com>

date

Thu, 15 Jan 2015 15:59:08 +0100

branch

TOR_BUG_9701

changeset 10

ac0c01689b40

permissions

-rw-r--r--

Implement a real Private Browsing Mode condition by changing the API/ABI;
This solves Tor bug #9701, complying with disk avoidance documented in
https://www.torproject.org/projects/torbrowser/design/#disk-avoidance.

michael@0	1	;
michael@0	2	; Copyright (c) 2012 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	%include "vpx_ports/x86_abi_support.asm"
michael@0	13
michael@0	14	;void vp8_filter_by_weight16x16_sse2
michael@0	15	;(
michael@0	16	; unsigned char *src,
michael@0	17	; int src_stride,
michael@0	18	; unsigned char *dst,
michael@0	19	; int dst_stride,
michael@0	20	; int src_weight
michael@0	21	;)
michael@0	22	global sym(vp8_filter_by_weight16x16_sse2) PRIVATE
michael@0	23	sym(vp8_filter_by_weight16x16_sse2):
michael@0	24	push rbp
michael@0	25	mov rbp, rsp
michael@0	26	SHADOW_ARGS_TO_STACK 5
michael@0	27	SAVE_XMM 6
michael@0	28	GET_GOT rbx
michael@0	29	push rsi
michael@0	30	push rdi
michael@0	31	; end prolog
michael@0	32
michael@0	33	movd xmm0, arg(4) ; src_weight
michael@0	34	pshuflw xmm0, xmm0, 0x0 ; replicate to all low words
michael@0	35	punpcklqdq xmm0, xmm0 ; replicate to all hi words
michael@0	36
michael@0	37	movdqa xmm1, [GLOBAL(tMFQE)]
michael@0	38	psubw xmm1, xmm0 ; dst_weight
michael@0	39
michael@0	40	mov rax, arg(0) ; src
michael@0	41	mov rsi, arg(1) ; src_stride
michael@0	42	mov rdx, arg(2) ; dst
michael@0	43	mov rdi, arg(3) ; dst_stride
michael@0	44
michael@0	45	mov rcx, 16 ; loop count
michael@0	46	pxor xmm6, xmm6
michael@0	47
michael@0	48	.combine
michael@0	49	movdqa xmm2, [rax]
michael@0	50	movdqa xmm4, [rdx]
michael@0	51	add rax, rsi
michael@0	52
michael@0	53	; src * src_weight
michael@0	54	movdqa xmm3, xmm2
michael@0	55	punpcklbw xmm2, xmm6
michael@0	56	punpckhbw xmm3, xmm6
michael@0	57	pmullw xmm2, xmm0
michael@0	58	pmullw xmm3, xmm0
michael@0	59
michael@0	60	; dst * dst_weight
michael@0	61	movdqa xmm5, xmm4
michael@0	62	punpcklbw xmm4, xmm6
michael@0	63	punpckhbw xmm5, xmm6
michael@0	64	pmullw xmm4, xmm1
michael@0	65	pmullw xmm5, xmm1
michael@0	66
michael@0	67	; sum, round and shift
michael@0	68	paddw xmm2, xmm4
michael@0	69	paddw xmm3, xmm5
michael@0	70	paddw xmm2, [GLOBAL(tMFQE_round)]
michael@0	71	paddw xmm3, [GLOBAL(tMFQE_round)]
michael@0	72	psrlw xmm2, 4
michael@0	73	psrlw xmm3, 4
michael@0	74
michael@0	75	packuswb xmm2, xmm3
michael@0	76	movdqa [rdx], xmm2
michael@0	77	add rdx, rdi
michael@0	78
michael@0	79	dec rcx
michael@0	80	jnz .combine
michael@0	81
michael@0	82	; begin epilog
michael@0	83	pop rdi
michael@0	84	pop rsi
michael@0	85	RESTORE_GOT
michael@0	86	RESTORE_XMM
michael@0	87	UNSHADOW_ARGS
michael@0	88	pop rbp
michael@0	89
michael@0	90	ret
michael@0	91
michael@0	92	;void vp8_filter_by_weight8x8_sse2
michael@0	93	;(
michael@0	94	; unsigned char *src,
michael@0	95	; int src_stride,
michael@0	96	; unsigned char *dst,
michael@0	97	; int dst_stride,
michael@0	98	; int src_weight
michael@0	99	;)
michael@0	100	global sym(vp8_filter_by_weight8x8_sse2) PRIVATE
michael@0	101	sym(vp8_filter_by_weight8x8_sse2):
michael@0	102	push rbp
michael@0	103	mov rbp, rsp
michael@0	104	SHADOW_ARGS_TO_STACK 5
michael@0	105	GET_GOT rbx
michael@0	106	push rsi
michael@0	107	push rdi
michael@0	108	; end prolog
michael@0	109
michael@0	110	movd xmm0, arg(4) ; src_weight
michael@0	111	pshuflw xmm0, xmm0, 0x0 ; replicate to all low words
michael@0	112	punpcklqdq xmm0, xmm0 ; replicate to all hi words
michael@0	113
michael@0	114	movdqa xmm1, [GLOBAL(tMFQE)]
michael@0	115	psubw xmm1, xmm0 ; dst_weight
michael@0	116
michael@0	117	mov rax, arg(0) ; src
michael@0	118	mov rsi, arg(1) ; src_stride
michael@0	119	mov rdx, arg(2) ; dst
michael@0	120	mov rdi, arg(3) ; dst_stride
michael@0	121
michael@0	122	mov rcx, 8 ; loop count
michael@0	123	pxor xmm4, xmm4
michael@0	124
michael@0	125	.combine
michael@0	126	movq xmm2, [rax]
michael@0	127	movq xmm3, [rdx]
michael@0	128	add rax, rsi
michael@0	129
michael@0	130	; src * src_weight
michael@0	131	punpcklbw xmm2, xmm4
michael@0	132	pmullw xmm2, xmm0
michael@0	133
michael@0	134	; dst * dst_weight
michael@0	135	punpcklbw xmm3, xmm4
michael@0	136	pmullw xmm3, xmm1
michael@0	137
michael@0	138	; sum, round and shift
michael@0	139	paddw xmm2, xmm3
michael@0	140	paddw xmm2, [GLOBAL(tMFQE_round)]
michael@0	141	psrlw xmm2, 4
michael@0	142
michael@0	143	packuswb xmm2, xmm4
michael@0	144	movq [rdx], xmm2
michael@0	145	add rdx, rdi
michael@0	146
michael@0	147	dec rcx
michael@0	148	jnz .combine
michael@0	149
michael@0	150	; begin epilog
michael@0	151	pop rdi
michael@0	152	pop rsi
michael@0	153	RESTORE_GOT
michael@0	154	UNSHADOW_ARGS
michael@0	155	pop rbp
michael@0	156
michael@0	157	ret
michael@0	158
michael@0	159	;void vp8_variance_and_sad_16x16_sse2 | arg
michael@0	160	;(
michael@0	161	; unsigned char *src1, 0
michael@0	162	; int stride1, 1
michael@0	163	; unsigned char *src2, 2
michael@0	164	; int stride2, 3
michael@0	165	; unsigned int *variance, 4
michael@0	166	; unsigned int *sad, 5
michael@0	167	;)
michael@0	168	global sym(vp8_variance_and_sad_16x16_sse2) PRIVATE
michael@0	169	sym(vp8_variance_and_sad_16x16_sse2):
michael@0	170	push rbp
michael@0	171	mov rbp, rsp
michael@0	172	SHADOW_ARGS_TO_STACK 6
michael@0	173	GET_GOT rbx
michael@0	174	push rsi
michael@0	175	push rdi
michael@0	176	; end prolog
michael@0	177
michael@0	178	mov rax, arg(0) ; src1
michael@0	179	mov rcx, arg(1) ; stride1
michael@0	180	mov rdx, arg(2) ; src2
michael@0	181	mov rdi, arg(3) ; stride2
michael@0	182
michael@0	183	mov rsi, 16 ; block height
michael@0	184
michael@0	185	; Prep accumulator registers
michael@0	186	pxor xmm3, xmm3 ; SAD
michael@0	187	pxor xmm4, xmm4 ; sum of src2
michael@0	188	pxor xmm5, xmm5 ; sum of src2^2
michael@0	189
michael@0	190	; Because we're working with the actual output frames
michael@0	191	; we can't depend on any kind of data alignment.
michael@0	192	.accumulate
michael@0	193	movdqa xmm0, [rax] ; src1
michael@0	194	movdqa xmm1, [rdx] ; src2
michael@0	195	add rax, rcx ; src1 + stride1
michael@0	196	add rdx, rdi ; src2 + stride2
michael@0	197
michael@0	198	; SAD(src1, src2)
michael@0	199	psadbw xmm0, xmm1
michael@0	200	paddusw xmm3, xmm0
michael@0	201
michael@0	202	; SUM(src2)
michael@0	203	pxor xmm2, xmm2
michael@0	204	psadbw xmm2, xmm1 ; sum src2 by misusing SAD against 0
michael@0	205	paddusw xmm4, xmm2
michael@0	206
michael@0	207	; pmaddubsw would be ideal if it took two unsigned values. instead,
michael@0	208	; it expects a signed and an unsigned value. so instead we zero extend
michael@0	209	; and operate on words.
michael@0	210	pxor xmm2, xmm2
michael@0	211	movdqa xmm0, xmm1
michael@0	212	punpcklbw xmm0, xmm2
michael@0	213	punpckhbw xmm1, xmm2
michael@0	214	pmaddwd xmm0, xmm0
michael@0	215	pmaddwd xmm1, xmm1
michael@0	216	paddd xmm5, xmm0
michael@0	217	paddd xmm5, xmm1
michael@0	218
michael@0	219	sub rsi, 1
michael@0	220	jnz .accumulate
michael@0	221
michael@0	222	; phaddd only operates on adjacent double words.
michael@0	223	; Finalize SAD and store
michael@0	224	movdqa xmm0, xmm3
michael@0	225	psrldq xmm0, 8
michael@0	226	paddusw xmm0, xmm3
michael@0	227	paddd xmm0, [GLOBAL(t128)]
michael@0	228	psrld xmm0, 8
michael@0	229
michael@0	230	mov rax, arg(5)
michael@0	231	movd [rax], xmm0
michael@0	232
michael@0	233	; Accumulate sum of src2
michael@0	234	movdqa xmm0, xmm4
michael@0	235	psrldq xmm0, 8
michael@0	236	paddusw xmm0, xmm4
michael@0	237	; Square src2. Ignore high value
michael@0	238	pmuludq xmm0, xmm0
michael@0	239	psrld xmm0, 8
michael@0	240
michael@0	241	; phaddw could be used to sum adjacent values but we want
michael@0	242	; all the values summed. promote to doubles, accumulate,
michael@0	243	; shift and sum
michael@0	244	pxor xmm2, xmm2
michael@0	245	movdqa xmm1, xmm5
michael@0	246	punpckldq xmm1, xmm2
michael@0	247	punpckhdq xmm5, xmm2
michael@0	248	paddd xmm1, xmm5
michael@0	249	movdqa xmm2, xmm1
michael@0	250	psrldq xmm1, 8
michael@0	251	paddd xmm1, xmm2
michael@0	252
michael@0	253	psubd xmm1, xmm0
michael@0	254
michael@0	255	; (variance + 128) >> 8
michael@0	256	paddd xmm1, [GLOBAL(t128)]
michael@0	257	psrld xmm1, 8
michael@0	258	mov rax, arg(4)
michael@0	259
michael@0	260	movd [rax], xmm1
michael@0	261
michael@0	262
michael@0	263	; begin epilog
michael@0	264	pop rdi
michael@0	265	pop rsi
michael@0	266	RESTORE_GOT
michael@0	267	UNSHADOW_ARGS
michael@0	268	pop rbp
michael@0	269	ret
michael@0	270
michael@0	271	SECTION_RODATA
michael@0	272	align 16
michael@0	273	t128:
michael@0	274	%ifndef __NASM_VER__
michael@0	275	ddq 128
michael@0	276	%elif CONFIG_BIG_ENDIAN
michael@0	277	dq 0, 128
michael@0	278	%else
michael@0	279	dq 128, 0
michael@0	280	%endif
michael@0	281	align 16
michael@0	282	tMFQE: ; 1 << MFQE_PRECISION
michael@0	283	times 8 dw 0x10
michael@0	284	align 16
michael@0	285	tMFQE_round: ; 1 << (MFQE_PRECISION - 1)
michael@0	286	times 8 dw 0x08
michael@0	287