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

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

media/libvpx/vp8/common/x86/dequantize_mmx.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) 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	%include "vpx_ports/x86_abi_support.asm"
michael@0	13
michael@0	14
michael@0	15	;void vp8_dequantize_b_impl_mmx(short *sq, short *dq, short *q)
michael@0	16	global sym(vp8_dequantize_b_impl_mmx) PRIVATE
michael@0	17	sym(vp8_dequantize_b_impl_mmx):
michael@0	18	push rbp
michael@0	19	mov rbp, rsp
michael@0	20	SHADOW_ARGS_TO_STACK 3
michael@0	21	push rsi
michael@0	22	push rdi
michael@0	23	; end prolog
michael@0	24
michael@0	25	mov rsi, arg(0) ;sq
michael@0	26	mov rdi, arg(1) ;dq
michael@0	27	mov rax, arg(2) ;q
michael@0	28
michael@0	29	movq mm1, [rsi]
michael@0	30	pmullw mm1, [rax+0] ; mm4 *= kernel 0 modifiers.
michael@0	31	movq [rdi], mm1
michael@0	32
michael@0	33	movq mm1, [rsi+8]
michael@0	34	pmullw mm1, [rax+8] ; mm4 *= kernel 0 modifiers.
michael@0	35	movq [rdi+8], mm1
michael@0	36
michael@0	37	movq mm1, [rsi+16]
michael@0	38	pmullw mm1, [rax+16] ; mm4 *= kernel 0 modifiers.
michael@0	39	movq [rdi+16], mm1
michael@0	40
michael@0	41	movq mm1, [rsi+24]
michael@0	42	pmullw mm1, [rax+24] ; mm4 *= kernel 0 modifiers.
michael@0	43	movq [rdi+24], mm1
michael@0	44
michael@0	45	; begin epilog
michael@0	46	pop rdi
michael@0	47	pop rsi
michael@0	48	UNSHADOW_ARGS
michael@0	49	pop rbp
michael@0	50	ret
michael@0	51
michael@0	52
michael@0	53	;void dequant_idct_add_mmx(
michael@0	54	;short *input, 0
michael@0	55	;short *dq, 1
michael@0	56	;unsigned char *dest, 2
michael@0	57	;int stride) 3
michael@0	58	global sym(vp8_dequant_idct_add_mmx) PRIVATE
michael@0	59	sym(vp8_dequant_idct_add_mmx):
michael@0	60	push rbp
michael@0	61	mov rbp, rsp
michael@0	62	SHADOW_ARGS_TO_STACK 4
michael@0	63	GET_GOT rbx
michael@0	64	push rdi
michael@0	65	; end prolog
michael@0	66
michael@0	67	mov rax, arg(0) ;input
michael@0	68	mov rdx, arg(1) ;dq
michael@0	69
michael@0	70
michael@0	71	movq mm0, [rax ]
michael@0	72	pmullw mm0, [rdx]
michael@0	73
michael@0	74	movq mm1, [rax +8]
michael@0	75	pmullw mm1, [rdx +8]
michael@0	76
michael@0	77	movq mm2, [rax+16]
michael@0	78	pmullw mm2, [rdx+16]
michael@0	79
michael@0	80	movq mm3, [rax+24]
michael@0	81	pmullw mm3, [rdx+24]
michael@0	82
michael@0	83	mov rdx, arg(2) ;dest
michael@0	84
michael@0	85	pxor mm7, mm7
michael@0	86
michael@0	87
michael@0	88	movq [rax], mm7
michael@0	89	movq [rax+8], mm7
michael@0	90
michael@0	91	movq [rax+16],mm7
michael@0	92	movq [rax+24],mm7
michael@0	93
michael@0	94
michael@0	95	movsxd rdi, dword ptr arg(3) ;stride
michael@0	96
michael@0	97	psubw mm0, mm2 ; b1= 0-2
michael@0	98	paddw mm2, mm2 ;
michael@0	99
michael@0	100	movq mm5, mm1
michael@0	101	paddw mm2, mm0 ; a1 =0+2
michael@0	102
michael@0	103	pmulhw mm5, [GLOBAL(x_s1sqr2)];
michael@0	104	paddw mm5, mm1 ; ip1 * sin(pi/8) * sqrt(2)
michael@0	105
michael@0	106	movq mm7, mm3 ;
michael@0	107	pmulhw mm7, [GLOBAL(x_c1sqr2less1)];
michael@0	108
michael@0	109	paddw mm7, mm3 ; ip3 * cos(pi/8) * sqrt(2)
michael@0	110	psubw mm7, mm5 ; c1
michael@0	111
michael@0	112	movq mm5, mm1
michael@0	113	movq mm4, mm3
michael@0	114
michael@0	115	pmulhw mm5, [GLOBAL(x_c1sqr2less1)]
michael@0	116	paddw mm5, mm1
michael@0	117
michael@0	118	pmulhw mm3, [GLOBAL(x_s1sqr2)]
michael@0	119	paddw mm3, mm4
michael@0	120
michael@0	121	paddw mm3, mm5 ; d1
michael@0	122	movq mm6, mm2 ; a1
michael@0	123
michael@0	124	movq mm4, mm0 ; b1
michael@0	125	paddw mm2, mm3 ;0
michael@0	126
michael@0	127	paddw mm4, mm7 ;1
michael@0	128	psubw mm0, mm7 ;2
michael@0	129
michael@0	130	psubw mm6, mm3 ;3
michael@0	131
michael@0	132	movq mm1, mm2 ; 03 02 01 00
michael@0	133	movq mm3, mm4 ; 23 22 21 20
michael@0	134
michael@0	135	punpcklwd mm1, mm0 ; 11 01 10 00
michael@0	136	punpckhwd mm2, mm0 ; 13 03 12 02
michael@0	137
michael@0	138	punpcklwd mm3, mm6 ; 31 21 30 20
michael@0	139	punpckhwd mm4, mm6 ; 33 23 32 22
michael@0	140
michael@0	141	movq mm0, mm1 ; 11 01 10 00
michael@0	142	movq mm5, mm2 ; 13 03 12 02
michael@0	143
michael@0	144	punpckldq mm0, mm3 ; 30 20 10 00
michael@0	145	punpckhdq mm1, mm3 ; 31 21 11 01
michael@0	146
michael@0	147	punpckldq mm2, mm4 ; 32 22 12 02
michael@0	148	punpckhdq mm5, mm4 ; 33 23 13 03
michael@0	149
michael@0	150	movq mm3, mm5 ; 33 23 13 03
michael@0	151
michael@0	152	psubw mm0, mm2 ; b1= 0-2
michael@0	153	paddw mm2, mm2 ;
michael@0	154
michael@0	155	movq mm5, mm1
michael@0	156	paddw mm2, mm0 ; a1 =0+2
michael@0	157
michael@0	158	pmulhw mm5, [GLOBAL(x_s1sqr2)];
michael@0	159	paddw mm5, mm1 ; ip1 * sin(pi/8) * sqrt(2)
michael@0	160
michael@0	161	movq mm7, mm3 ;
michael@0	162	pmulhw mm7, [GLOBAL(x_c1sqr2less1)];
michael@0	163
michael@0	164	paddw mm7, mm3 ; ip3 * cos(pi/8) * sqrt(2)
michael@0	165	psubw mm7, mm5 ; c1
michael@0	166
michael@0	167	movq mm5, mm1
michael@0	168	movq mm4, mm3
michael@0	169
michael@0	170	pmulhw mm5, [GLOBAL(x_c1sqr2less1)]
michael@0	171	paddw mm5, mm1
michael@0	172
michael@0	173	pmulhw mm3, [GLOBAL(x_s1sqr2)]
michael@0	174	paddw mm3, mm4
michael@0	175
michael@0	176	paddw mm3, mm5 ; d1
michael@0	177	paddw mm0, [GLOBAL(fours)]
michael@0	178
michael@0	179	paddw mm2, [GLOBAL(fours)]
michael@0	180	movq mm6, mm2 ; a1
michael@0	181
michael@0	182	movq mm4, mm0 ; b1
michael@0	183	paddw mm2, mm3 ;0
michael@0	184
michael@0	185	paddw mm4, mm7 ;1
michael@0	186	psubw mm0, mm7 ;2
michael@0	187
michael@0	188	psubw mm6, mm3 ;3
michael@0	189	psraw mm2, 3
michael@0	190
michael@0	191	psraw mm0, 3
michael@0	192	psraw mm4, 3
michael@0	193
michael@0	194	psraw mm6, 3
michael@0	195
michael@0	196	movq mm1, mm2 ; 03 02 01 00
michael@0	197	movq mm3, mm4 ; 23 22 21 20
michael@0	198
michael@0	199	punpcklwd mm1, mm0 ; 11 01 10 00
michael@0	200	punpckhwd mm2, mm0 ; 13 03 12 02
michael@0	201
michael@0	202	punpcklwd mm3, mm6 ; 31 21 30 20
michael@0	203	punpckhwd mm4, mm6 ; 33 23 32 22
michael@0	204
michael@0	205	movq mm0, mm1 ; 11 01 10 00
michael@0	206	movq mm5, mm2 ; 13 03 12 02
michael@0	207
michael@0	208	punpckldq mm0, mm3 ; 30 20 10 00
michael@0	209	punpckhdq mm1, mm3 ; 31 21 11 01
michael@0	210
michael@0	211	punpckldq mm2, mm4 ; 32 22 12 02
michael@0	212	punpckhdq mm5, mm4 ; 33 23 13 03
michael@0	213
michael@0	214	pxor mm7, mm7
michael@0	215
michael@0	216	movd mm4, [rdx]
michael@0	217	punpcklbw mm4, mm7
michael@0	218	paddsw mm0, mm4
michael@0	219	packuswb mm0, mm7
michael@0	220	movd [rdx], mm0
michael@0	221
michael@0	222	movd mm4, [rdx+rdi]
michael@0	223	punpcklbw mm4, mm7
michael@0	224	paddsw mm1, mm4
michael@0	225	packuswb mm1, mm7
michael@0	226	movd [rdx+rdi], mm1
michael@0	227
michael@0	228	movd mm4, [rdx+2*rdi]
michael@0	229	punpcklbw mm4, mm7
michael@0	230	paddsw mm2, mm4
michael@0	231	packuswb mm2, mm7
michael@0	232	movd [rdx+rdi*2], mm2
michael@0	233
michael@0	234	add rdx, rdi
michael@0	235
michael@0	236	movd mm4, [rdx+2*rdi]
michael@0	237	punpcklbw mm4, mm7
michael@0	238	paddsw mm5, mm4
michael@0	239	packuswb mm5, mm7
michael@0	240	movd [rdx+rdi*2], mm5
michael@0	241
michael@0	242	; begin epilog
michael@0	243	pop rdi
michael@0	244	RESTORE_GOT
michael@0	245	UNSHADOW_ARGS
michael@0	246	pop rbp
michael@0	247	ret
michael@0	248
michael@0	249	SECTION_RODATA
michael@0	250	align 16
michael@0	251	x_s1sqr2:
michael@0	252	times 4 dw 0x8A8C
michael@0	253	align 16
michael@0	254	x_c1sqr2less1:
michael@0	255	times 4 dw 0x4E7B
michael@0	256	align 16
michael@0	257	fours:
michael@0	258	times 4 dw 0x0004