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

media/libvpx/vp9/common/x86/vp9_intrapred_sse2.asm@6474c204b198 (annotated)

media/libvpx/vp9/common/x86/vp9_intrapred_sse2.asm

Wed, 31 Dec 2014 06:09:35 +0100

author

Michael Schloh von Bennewitz <michael@schloh.com>

date

Wed, 31 Dec 2014 06:09:35 +0100

changeset 0

6474c204b198

permissions

-rw-r--r--

Cloned upstream origin tor-browser at tor-browser-31.3.0esr-4.5-1-build1
revision ID fc1c9ff7c1b2defdbc039f12214767608f46423f for hacking purpose.

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	%include "third_party/x86inc/x86inc.asm"
michael@0	12
michael@0	13	SECTION_RODATA
michael@0	14	pw_4: times 8 dw 4
michael@0	15	pw_8: times 8 dw 8
michael@0	16	pw_16: times 8 dw 16
michael@0	17	pw_32: times 8 dw 32
michael@0	18
michael@0	19	SECTION .text
michael@0	20
michael@0	21	INIT_MMX sse
michael@0	22	cglobal dc_predictor_4x4, 4, 5, 2, dst, stride, above, left, goffset
michael@0	23	GET_GOT goffsetq
michael@0	24
michael@0	25	pxor m1, m1
michael@0	26	movd m0, [aboveq]
michael@0	27	punpckldq m0, [leftq]
michael@0	28	psadbw m0, m1
michael@0	29	paddw m0, [GLOBAL(pw_4)]
michael@0	30	psraw m0, 3
michael@0	31	pshufw m0, m0, 0x0
michael@0	32	packuswb m0, m0
michael@0	33	movd [dstq ], m0
michael@0	34	movd [dstq+strideq], m0
michael@0	35	lea dstq, [dstq+strideq*2]
michael@0	36	movd [dstq ], m0
michael@0	37	movd [dstq+strideq], m0
michael@0	38
michael@0	39	RESTORE_GOT
michael@0	40	RET
michael@0	41
michael@0	42	INIT_MMX sse
michael@0	43	cglobal dc_predictor_8x8, 4, 5, 3, dst, stride, above, left, goffset
michael@0	44	GET_GOT goffsetq
michael@0	45
michael@0	46	pxor m1, m1
michael@0	47	movq m0, [aboveq]
michael@0	48	movq m2, [leftq]
michael@0	49	DEFINE_ARGS dst, stride, stride3
michael@0	50	lea stride3q, [strideq*3]
michael@0	51	psadbw m0, m1
michael@0	52	psadbw m2, m1
michael@0	53	paddw m0, m2
michael@0	54	paddw m0, [GLOBAL(pw_8)]
michael@0	55	psraw m0, 4
michael@0	56	pshufw m0, m0, 0x0
michael@0	57	packuswb m0, m0
michael@0	58	movq [dstq ], m0
michael@0	59	movq [dstq+strideq ], m0
michael@0	60	movq [dstq+strideq*2], m0
michael@0	61	movq [dstq+stride3q ], m0
michael@0	62	lea dstq, [dstq+strideq*4]
michael@0	63	movq [dstq ], m0
michael@0	64	movq [dstq+strideq ], m0
michael@0	65	movq [dstq+strideq*2], m0
michael@0	66	movq [dstq+stride3q ], m0
michael@0	67
michael@0	68	RESTORE_GOT
michael@0	69	RET
michael@0	70
michael@0	71	INIT_XMM sse2
michael@0	72	cglobal dc_predictor_16x16, 4, 5, 3, dst, stride, above, left, goffset
michael@0	73	GET_GOT goffsetq
michael@0	74
michael@0	75	pxor m1, m1
michael@0	76	mova m0, [aboveq]
michael@0	77	mova m2, [leftq]
michael@0	78	DEFINE_ARGS dst, stride, stride3, lines4
michael@0	79	lea stride3q, [strideq*3]
michael@0	80	mov lines4d, 4
michael@0	81	psadbw m0, m1
michael@0	82	psadbw m2, m1
michael@0	83	paddw m0, m2
michael@0	84	movhlps m2, m0
michael@0	85	paddw m0, m2
michael@0	86	paddw m0, [GLOBAL(pw_16)]
michael@0	87	psraw m0, 5
michael@0	88	pshuflw m0, m0, 0x0
michael@0	89	punpcklqdq m0, m0
michael@0	90	packuswb m0, m0
michael@0	91	.loop:
michael@0	92	mova [dstq ], m0
michael@0	93	mova [dstq+strideq ], m0
michael@0	94	mova [dstq+strideq*2], m0
michael@0	95	mova [dstq+stride3q ], m0
michael@0	96	lea dstq, [dstq+strideq*4]
michael@0	97	dec lines4d
michael@0	98	jnz .loop
michael@0	99
michael@0	100	RESTORE_GOT
michael@0	101	REP_RET
michael@0	102
michael@0	103	INIT_XMM sse2
michael@0	104	cglobal dc_predictor_32x32, 4, 5, 5, dst, stride, above, left, goffset
michael@0	105	GET_GOT goffsetq
michael@0	106
michael@0	107	pxor m1, m1
michael@0	108	mova m0, [aboveq]
michael@0	109	mova m2, [aboveq+16]
michael@0	110	mova m3, [leftq]
michael@0	111	mova m4, [leftq+16]
michael@0	112	DEFINE_ARGS dst, stride, stride3, lines4
michael@0	113	lea stride3q, [strideq*3]
michael@0	114	mov lines4d, 8
michael@0	115	psadbw m0, m1
michael@0	116	psadbw m2, m1
michael@0	117	psadbw m3, m1
michael@0	118	psadbw m4, m1
michael@0	119	paddw m0, m2
michael@0	120	paddw m0, m3
michael@0	121	paddw m0, m4
michael@0	122	movhlps m2, m0
michael@0	123	paddw m0, m2
michael@0	124	paddw m0, [GLOBAL(pw_32)]
michael@0	125	psraw m0, 6
michael@0	126	pshuflw m0, m0, 0x0
michael@0	127	punpcklqdq m0, m0
michael@0	128	packuswb m0, m0
michael@0	129	.loop:
michael@0	130	mova [dstq ], m0
michael@0	131	mova [dstq +16], m0
michael@0	132	mova [dstq+strideq ], m0
michael@0	133	mova [dstq+strideq +16], m0
michael@0	134	mova [dstq+strideq*2 ], m0
michael@0	135	mova [dstq+strideq*2+16], m0
michael@0	136	mova [dstq+stride3q ], m0
michael@0	137	mova [dstq+stride3q +16], m0
michael@0	138	lea dstq, [dstq+strideq*4]
michael@0	139	dec lines4d
michael@0	140	jnz .loop
michael@0	141
michael@0	142	RESTORE_GOT
michael@0	143	REP_RET
michael@0	144
michael@0	145	INIT_MMX sse
michael@0	146	cglobal v_predictor_4x4, 3, 3, 1, dst, stride, above
michael@0	147	movd m0, [aboveq]
michael@0	148	movd [dstq ], m0
michael@0	149	movd [dstq+strideq], m0
michael@0	150	lea dstq, [dstq+strideq*2]
michael@0	151	movd [dstq ], m0
michael@0	152	movd [dstq+strideq], m0
michael@0	153	RET
michael@0	154
michael@0	155	INIT_MMX sse
michael@0	156	cglobal v_predictor_8x8, 3, 3, 1, dst, stride, above
michael@0	157	movq m0, [aboveq]
michael@0	158	DEFINE_ARGS dst, stride, stride3
michael@0	159	lea stride3q, [strideq*3]
michael@0	160	movq [dstq ], m0
michael@0	161	movq [dstq+strideq ], m0
michael@0	162	movq [dstq+strideq*2], m0
michael@0	163	movq [dstq+stride3q ], m0
michael@0	164	lea dstq, [dstq+strideq*4]
michael@0	165	movq [dstq ], m0
michael@0	166	movq [dstq+strideq ], m0
michael@0	167	movq [dstq+strideq*2], m0
michael@0	168	movq [dstq+stride3q ], m0
michael@0	169	RET
michael@0	170
michael@0	171	INIT_XMM sse2
michael@0	172	cglobal v_predictor_16x16, 3, 4, 1, dst, stride, above
michael@0	173	mova m0, [aboveq]
michael@0	174	DEFINE_ARGS dst, stride, stride3, nlines4
michael@0	175	lea stride3q, [strideq*3]
michael@0	176	mov nlines4d, 4
michael@0	177	.loop:
michael@0	178	mova [dstq ], m0
michael@0	179	mova [dstq+strideq ], m0
michael@0	180	mova [dstq+strideq*2], m0
michael@0	181	mova [dstq+stride3q ], m0
michael@0	182	lea dstq, [dstq+strideq*4]
michael@0	183	dec nlines4d
michael@0	184	jnz .loop
michael@0	185	REP_RET
michael@0	186
michael@0	187	INIT_XMM sse2
michael@0	188	cglobal v_predictor_32x32, 3, 4, 2, dst, stride, above
michael@0	189	mova m0, [aboveq]
michael@0	190	mova m1, [aboveq+16]
michael@0	191	DEFINE_ARGS dst, stride, stride3, nlines4
michael@0	192	lea stride3q, [strideq*3]
michael@0	193	mov nlines4d, 8
michael@0	194	.loop:
michael@0	195	mova [dstq ], m0
michael@0	196	mova [dstq +16], m1
michael@0	197	mova [dstq+strideq ], m0
michael@0	198	mova [dstq+strideq +16], m1
michael@0	199	mova [dstq+strideq*2 ], m0
michael@0	200	mova [dstq+strideq*2+16], m1
michael@0	201	mova [dstq+stride3q ], m0
michael@0	202	mova [dstq+stride3q +16], m1
michael@0	203	lea dstq, [dstq+strideq*4]
michael@0	204	dec nlines4d
michael@0	205	jnz .loop
michael@0	206	REP_RET
michael@0	207
michael@0	208	INIT_MMX sse
michael@0	209	cglobal tm_predictor_4x4, 4, 4, 4, dst, stride, above, left
michael@0	210	pxor m1, m1
michael@0	211	movd m2, [aboveq-1]
michael@0	212	movd m0, [aboveq]
michael@0	213	punpcklbw m2, m1
michael@0	214	punpcklbw m0, m1
michael@0	215	pshufw m2, m2, 0x0
michael@0	216	DEFINE_ARGS dst, stride, line, left
michael@0	217	mov lineq, -2
michael@0	218	add leftq, 4
michael@0	219	psubw m0, m2
michael@0	220	.loop:
michael@0	221	movd m2, [leftq+lineq*2]
michael@0	222	movd m3, [leftq+lineq*2+1]
michael@0	223	punpcklbw m2, m1
michael@0	224	punpcklbw m3, m1
michael@0	225	pshufw m2, m2, 0x0
michael@0	226	pshufw m3, m3, 0x0
michael@0	227	paddw m2, m0
michael@0	228	paddw m3, m0
michael@0	229	packuswb m2, m2
michael@0	230	packuswb m3, m3
michael@0	231	movd [dstq ], m2
michael@0	232	movd [dstq+strideq], m3
michael@0	233	lea dstq, [dstq+strideq*2]
michael@0	234	inc lineq
michael@0	235	jnz .loop
michael@0	236	REP_RET
michael@0	237
michael@0	238	INIT_XMM sse2
michael@0	239	cglobal tm_predictor_8x8, 4, 4, 4, dst, stride, above, left
michael@0	240	pxor m1, m1
michael@0	241	movd m2, [aboveq-1]
michael@0	242	movq m0, [aboveq]
michael@0	243	punpcklbw m2, m1
michael@0	244	punpcklbw m0, m1
michael@0	245	pshuflw m2, m2, 0x0
michael@0	246	DEFINE_ARGS dst, stride, line, left
michael@0	247	mov lineq, -4
michael@0	248	punpcklqdq m2, m2
michael@0	249	add leftq, 8
michael@0	250	psubw m0, m2
michael@0	251	.loop:
michael@0	252	movd m2, [leftq+lineq*2]
michael@0	253	movd m3, [leftq+lineq*2+1]
michael@0	254	punpcklbw m2, m1
michael@0	255	punpcklbw m3, m1
michael@0	256	pshuflw m2, m2, 0x0
michael@0	257	pshuflw m3, m3, 0x0
michael@0	258	punpcklqdq m2, m2
michael@0	259	punpcklqdq m3, m3
michael@0	260	paddw m2, m0
michael@0	261	paddw m3, m0
michael@0	262	packuswb m2, m3
michael@0	263	movq [dstq ], m2
michael@0	264	movhps [dstq+strideq], m2
michael@0	265	lea dstq, [dstq+strideq*2]
michael@0	266	inc lineq
michael@0	267	jnz .loop
michael@0	268	REP_RET
michael@0	269
michael@0	270	INIT_XMM sse2
michael@0	271	cglobal tm_predictor_16x16, 4, 4, 7, dst, stride, above, left
michael@0	272	pxor m1, m1
michael@0	273	movd m2, [aboveq-1]
michael@0	274	mova m0, [aboveq]
michael@0	275	punpcklbw m2, m1
michael@0	276	punpckhbw m4, m0, m1
michael@0	277	punpcklbw m0, m1
michael@0	278	pshuflw m2, m2, 0x0
michael@0	279	DEFINE_ARGS dst, stride, line, left
michael@0	280	mov lineq, -8
michael@0	281	punpcklqdq m2, m2
michael@0	282	add leftq, 16
michael@0	283	psubw m0, m2
michael@0	284	psubw m4, m2
michael@0	285	.loop:
michael@0	286	movd m2, [leftq+lineq*2]
michael@0	287	movd m3, [leftq+lineq*2+1]
michael@0	288	punpcklbw m2, m1
michael@0	289	punpcklbw m3, m1
michael@0	290	pshuflw m2, m2, 0x0
michael@0	291	pshuflw m3, m3, 0x0
michael@0	292	punpcklqdq m2, m2
michael@0	293	punpcklqdq m3, m3
michael@0	294	paddw m5, m2, m0
michael@0	295	paddw m6, m3, m0
michael@0	296	paddw m2, m4
michael@0	297	paddw m3, m4
michael@0	298	packuswb m5, m2
michael@0	299	packuswb m6, m3
michael@0	300	mova [dstq ], m5
michael@0	301	mova [dstq+strideq], m6
michael@0	302	lea dstq, [dstq+strideq*2]
michael@0	303	inc lineq
michael@0	304	jnz .loop
michael@0	305	REP_RET
michael@0	306
michael@0	307	%if ARCH_X86_64
michael@0	308	INIT_XMM sse2
michael@0	309	cglobal tm_predictor_32x32, 4, 4, 10, dst, stride, above, left
michael@0	310	pxor m1, m1
michael@0	311	movd m2, [aboveq-1]
michael@0	312	mova m0, [aboveq]
michael@0	313	mova m4, [aboveq+16]
michael@0	314	punpcklbw m2, m1
michael@0	315	punpckhbw m3, m0, m1
michael@0	316	punpckhbw m5, m4, m1
michael@0	317	punpcklbw m0, m1
michael@0	318	punpcklbw m4, m1
michael@0	319	pshuflw m2, m2, 0x0
michael@0	320	DEFINE_ARGS dst, stride, line, left
michael@0	321	mov lineq, -16
michael@0	322	punpcklqdq m2, m2
michael@0	323	add leftq, 32
michael@0	324	psubw m0, m2
michael@0	325	psubw m3, m2
michael@0	326	psubw m4, m2
michael@0	327	psubw m5, m2
michael@0	328	.loop:
michael@0	329	movd m2, [leftq+lineq*2]
michael@0	330	movd m6, [leftq+lineq*2+1]
michael@0	331	punpcklbw m2, m1
michael@0	332	punpcklbw m6, m1
michael@0	333	pshuflw m2, m2, 0x0
michael@0	334	pshuflw m6, m6, 0x0
michael@0	335	punpcklqdq m2, m2
michael@0	336	punpcklqdq m6, m6
michael@0	337	paddw m7, m2, m0
michael@0	338	paddw m8, m2, m3
michael@0	339	paddw m9, m2, m4
michael@0	340	paddw m2, m5
michael@0	341	packuswb m7, m8
michael@0	342	packuswb m9, m2
michael@0	343	paddw m2, m6, m0
michael@0	344	paddw m8, m6, m3
michael@0	345	mova [dstq ], m7
michael@0	346	paddw m7, m6, m4
michael@0	347	paddw m6, m5
michael@0	348	mova [dstq +16], m9
michael@0	349	packuswb m2, m8
michael@0	350	packuswb m7, m6
michael@0	351	mova [dstq+strideq ], m2
michael@0	352	mova [dstq+strideq+16], m7
michael@0	353	lea dstq, [dstq+strideq*2]
michael@0	354	inc lineq
michael@0	355	jnz .loop
michael@0	356	REP_RET
michael@0	357	%endif