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

media/libvpx/vp9/encoder/x86/vp9_subpel_variance_impl_sse2.asm@6474c204b198 (annotated)

media/libvpx/vp9/encoder/x86/vp9_subpel_variance_impl_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 "vpx_ports/x86_abi_support.asm"
michael@0	12
michael@0	13	;void vp9_half_horiz_vert_variance16x_h_sse2
michael@0	14	;(
michael@0	15	; unsigned char *ref_ptr,
michael@0	16	; int ref_pixels_per_line,
michael@0	17	; unsigned char *src_ptr,
michael@0	18	; int src_pixels_per_line,
michael@0	19	; unsigned int Height,
michael@0	20	; int *sum,
michael@0	21	; unsigned int *sumsquared
michael@0	22	;)
michael@0	23	global sym(vp9_half_horiz_vert_variance16x_h_sse2) PRIVATE
michael@0	24	sym(vp9_half_horiz_vert_variance16x_h_sse2):
michael@0	25	push rbp
michael@0	26	mov rbp, rsp
michael@0	27	SHADOW_ARGS_TO_STACK 7
michael@0	28	SAVE_XMM 7
michael@0	29	GET_GOT rbx
michael@0	30	push rsi
michael@0	31	push rdi
michael@0	32	; end prolog
michael@0	33
michael@0	34	pxor xmm6, xmm6 ; error accumulator
michael@0	35	pxor xmm7, xmm7 ; sse eaccumulator
michael@0	36	mov rsi, arg(0) ;ref_ptr ;
michael@0	37
michael@0	38	mov rdi, arg(2) ;src_ptr ;
michael@0	39	movsxd rcx, dword ptr arg(4) ;Height ;
michael@0	40	movsxd rax, dword ptr arg(1) ;ref_pixels_per_line
michael@0	41	movsxd rdx, dword ptr arg(3) ;src_pixels_per_line
michael@0	42
michael@0	43	pxor xmm0, xmm0 ;
michael@0	44
michael@0	45	movdqu xmm5, XMMWORD PTR [rsi]
michael@0	46	movdqu xmm3, XMMWORD PTR [rsi+1]
michael@0	47	pavgb xmm5, xmm3 ; xmm5 = avg(xmm1,xmm3) horizontal line 1
michael@0	48
michael@0	49	lea rsi, [rsi + rax]
michael@0	50
michael@0	51	.half_horiz_vert_variance16x_h_1:
michael@0	52	movdqu xmm1, XMMWORD PTR [rsi] ;
michael@0	53	movdqu xmm2, XMMWORD PTR [rsi+1] ;
michael@0	54	pavgb xmm1, xmm2 ; xmm1 = avg(xmm1,xmm3) horizontal line i+1
michael@0	55
michael@0	56	pavgb xmm5, xmm1 ; xmm = vertical average of the above
michael@0	57
michael@0	58	movdqa xmm4, xmm5
michael@0	59	punpcklbw xmm5, xmm0 ; xmm5 = words of above
michael@0	60	punpckhbw xmm4, xmm0
michael@0	61
michael@0	62	movq xmm3, QWORD PTR [rdi] ; xmm3 = d0,d1,d2..d7
michael@0	63	punpcklbw xmm3, xmm0 ; xmm3 = words of above
michael@0	64	psubw xmm5, xmm3 ; xmm5 -= xmm3
michael@0	65
michael@0	66	movq xmm3, QWORD PTR [rdi+8]
michael@0	67	punpcklbw xmm3, xmm0
michael@0	68	psubw xmm4, xmm3
michael@0	69
michael@0	70	paddw xmm6, xmm5 ; xmm6 += accumulated column differences
michael@0	71	paddw xmm6, xmm4
michael@0	72	pmaddwd xmm5, xmm5 ; xmm5 *= xmm5
michael@0	73	pmaddwd xmm4, xmm4
michael@0	74	paddd xmm7, xmm5 ; xmm7 += accumulated square column differences
michael@0	75	paddd xmm7, xmm4
michael@0	76
michael@0	77	movdqa xmm5, xmm1 ; save xmm1 for use on the next row
michael@0	78
michael@0	79	lea rsi, [rsi + rax]
michael@0	80	lea rdi, [rdi + rdx]
michael@0	81
michael@0	82	sub rcx, 1 ;
michael@0	83	jnz .half_horiz_vert_variance16x_h_1 ;
michael@0	84
michael@0	85	pxor xmm1, xmm1
michael@0	86	pxor xmm5, xmm5
michael@0	87
michael@0	88	punpcklwd xmm0, xmm6
michael@0	89	punpckhwd xmm1, xmm6
michael@0	90	psrad xmm0, 16
michael@0	91	psrad xmm1, 16
michael@0	92	paddd xmm0, xmm1
michael@0	93	movdqa xmm1, xmm0
michael@0	94
michael@0	95	movdqa xmm6, xmm7
michael@0	96	punpckldq xmm6, xmm5
michael@0	97	punpckhdq xmm7, xmm5
michael@0	98	paddd xmm6, xmm7
michael@0	99
michael@0	100	punpckldq xmm0, xmm5
michael@0	101	punpckhdq xmm1, xmm5
michael@0	102	paddd xmm0, xmm1
michael@0	103
michael@0	104	movdqa xmm7, xmm6
michael@0	105	movdqa xmm1, xmm0
michael@0	106
michael@0	107	psrldq xmm7, 8
michael@0	108	psrldq xmm1, 8
michael@0	109
michael@0	110	paddd xmm6, xmm7
michael@0	111	paddd xmm0, xmm1
michael@0	112
michael@0	113	mov rsi, arg(5) ;[Sum]
michael@0	114	mov rdi, arg(6) ;[SSE]
michael@0	115
michael@0	116	movd [rsi], xmm0
michael@0	117	movd [rdi], xmm6
michael@0	118
michael@0	119	; begin epilog
michael@0	120	pop rdi
michael@0	121	pop rsi
michael@0	122	RESTORE_GOT
michael@0	123	RESTORE_XMM
michael@0	124	UNSHADOW_ARGS
michael@0	125	pop rbp
michael@0	126	ret
michael@0	127
michael@0	128	;void vp9_half_vert_variance16x_h_sse2
michael@0	129	;(
michael@0	130	; unsigned char *ref_ptr,
michael@0	131	; int ref_pixels_per_line,
michael@0	132	; unsigned char *src_ptr,
michael@0	133	; int src_pixels_per_line,
michael@0	134	; unsigned int Height,
michael@0	135	; int *sum,
michael@0	136	; unsigned int *sumsquared
michael@0	137	;)
michael@0	138	global sym(vp9_half_vert_variance16x_h_sse2) PRIVATE
michael@0	139	sym(vp9_half_vert_variance16x_h_sse2):
michael@0	140	push rbp
michael@0	141	mov rbp, rsp
michael@0	142	SHADOW_ARGS_TO_STACK 7
michael@0	143	SAVE_XMM 7
michael@0	144	GET_GOT rbx
michael@0	145	push rsi
michael@0	146	push rdi
michael@0	147	; end prolog
michael@0	148
michael@0	149	pxor xmm6, xmm6 ; error accumulator
michael@0	150	pxor xmm7, xmm7 ; sse eaccumulator
michael@0	151	mov rsi, arg(0) ;ref_ptr
michael@0	152
michael@0	153	mov rdi, arg(2) ;src_ptr
michael@0	154	movsxd rcx, dword ptr arg(4) ;Height
michael@0	155	movsxd rax, dword ptr arg(1) ;ref_pixels_per_line
michael@0	156	movsxd rdx, dword ptr arg(3) ;src_pixels_per_line
michael@0	157
michael@0	158	movdqu xmm5, XMMWORD PTR [rsi]
michael@0	159	lea rsi, [rsi + rax ]
michael@0	160	pxor xmm0, xmm0
michael@0	161
michael@0	162	.half_vert_variance16x_h_1:
michael@0	163	movdqu xmm3, XMMWORD PTR [rsi]
michael@0	164
michael@0	165	pavgb xmm5, xmm3 ; xmm5 = avg(xmm1,xmm3)
michael@0	166	movdqa xmm4, xmm5
michael@0	167	punpcklbw xmm5, xmm0
michael@0	168	punpckhbw xmm4, xmm0
michael@0	169
michael@0	170	movq xmm2, QWORD PTR [rdi]
michael@0	171	punpcklbw xmm2, xmm0
michael@0	172	psubw xmm5, xmm2
michael@0	173	movq xmm2, QWORD PTR [rdi+8]
michael@0	174	punpcklbw xmm2, xmm0
michael@0	175	psubw xmm4, xmm2
michael@0	176
michael@0	177	paddw xmm6, xmm5 ; xmm6 += accumulated column differences
michael@0	178	paddw xmm6, xmm4
michael@0	179	pmaddwd xmm5, xmm5 ; xmm5 *= xmm5
michael@0	180	pmaddwd xmm4, xmm4
michael@0	181	paddd xmm7, xmm5 ; xmm7 += accumulated square column differences
michael@0	182	paddd xmm7, xmm4
michael@0	183
michael@0	184	movdqa xmm5, xmm3
michael@0	185
michael@0	186	lea rsi, [rsi + rax]
michael@0	187	lea rdi, [rdi + rdx]
michael@0	188
michael@0	189	sub rcx, 1
michael@0	190	jnz .half_vert_variance16x_h_1
michael@0	191
michael@0	192	pxor xmm1, xmm1
michael@0	193	pxor xmm5, xmm5
michael@0	194
michael@0	195	punpcklwd xmm0, xmm6
michael@0	196	punpckhwd xmm1, xmm6
michael@0	197	psrad xmm0, 16
michael@0	198	psrad xmm1, 16
michael@0	199	paddd xmm0, xmm1
michael@0	200	movdqa xmm1, xmm0
michael@0	201
michael@0	202	movdqa xmm6, xmm7
michael@0	203	punpckldq xmm6, xmm5
michael@0	204	punpckhdq xmm7, xmm5
michael@0	205	paddd xmm6, xmm7
michael@0	206
michael@0	207	punpckldq xmm0, xmm5
michael@0	208	punpckhdq xmm1, xmm5
michael@0	209	paddd xmm0, xmm1
michael@0	210
michael@0	211	movdqa xmm7, xmm6
michael@0	212	movdqa xmm1, xmm0
michael@0	213
michael@0	214	psrldq xmm7, 8
michael@0	215	psrldq xmm1, 8
michael@0	216
michael@0	217	paddd xmm6, xmm7
michael@0	218	paddd xmm0, xmm1
michael@0	219
michael@0	220	mov rsi, arg(5) ;[Sum]
michael@0	221	mov rdi, arg(6) ;[SSE]
michael@0	222
michael@0	223	movd [rsi], xmm0
michael@0	224	movd [rdi], xmm6
michael@0	225
michael@0	226	; begin epilog
michael@0	227	pop rdi
michael@0	228	pop rsi
michael@0	229	RESTORE_GOT
michael@0	230	RESTORE_XMM
michael@0	231	UNSHADOW_ARGS
michael@0	232	pop rbp
michael@0	233	ret
michael@0	234
michael@0	235	;void vp9_half_horiz_variance16x_h_sse2
michael@0	236	;(
michael@0	237	; unsigned char *ref_ptr,
michael@0	238	; int ref_pixels_per_line,
michael@0	239	; unsigned char *src_ptr,
michael@0	240	; int src_pixels_per_line,
michael@0	241	; unsigned int Height,
michael@0	242	; int *sum,
michael@0	243	; unsigned int *sumsquared
michael@0	244	;)
michael@0	245	global sym(vp9_half_horiz_variance16x_h_sse2) PRIVATE
michael@0	246	sym(vp9_half_horiz_variance16x_h_sse2):
michael@0	247	push rbp
michael@0	248	mov rbp, rsp
michael@0	249	SHADOW_ARGS_TO_STACK 7
michael@0	250	SAVE_XMM 7
michael@0	251	GET_GOT rbx
michael@0	252	push rsi
michael@0	253	push rdi
michael@0	254	; end prolog
michael@0	255
michael@0	256	pxor xmm6, xmm6 ; error accumulator
michael@0	257	pxor xmm7, xmm7 ; sse eaccumulator
michael@0	258	mov rsi, arg(0) ;ref_ptr ;
michael@0	259
michael@0	260	mov rdi, arg(2) ;src_ptr ;
michael@0	261	movsxd rcx, dword ptr arg(4) ;Height ;
michael@0	262	movsxd rax, dword ptr arg(1) ;ref_pixels_per_line
michael@0	263	movsxd rdx, dword ptr arg(3) ;src_pixels_per_line
michael@0	264
michael@0	265	pxor xmm0, xmm0 ;
michael@0	266
michael@0	267	.half_horiz_variance16x_h_1:
michael@0	268	movdqu xmm5, XMMWORD PTR [rsi] ; xmm5 = s0,s1,s2..s15
michael@0	269	movdqu xmm3, XMMWORD PTR [rsi+1] ; xmm3 = s1,s2,s3..s16
michael@0	270
michael@0	271	pavgb xmm5, xmm3 ; xmm5 = avg(xmm1,xmm3)
michael@0	272	movdqa xmm1, xmm5
michael@0	273	punpcklbw xmm5, xmm0 ; xmm5 = words of above
michael@0	274	punpckhbw xmm1, xmm0
michael@0	275
michael@0	276	movq xmm3, QWORD PTR [rdi] ; xmm3 = d0,d1,d2..d7
michael@0	277	punpcklbw xmm3, xmm0 ; xmm3 = words of above
michael@0	278	movq xmm2, QWORD PTR [rdi+8]
michael@0	279	punpcklbw xmm2, xmm0
michael@0	280
michael@0	281	psubw xmm5, xmm3 ; xmm5 -= xmm3
michael@0	282	psubw xmm1, xmm2
michael@0	283	paddw xmm6, xmm5 ; xmm6 += accumulated column differences
michael@0	284	paddw xmm6, xmm1
michael@0	285	pmaddwd xmm5, xmm5 ; xmm5 *= xmm5
michael@0	286	pmaddwd xmm1, xmm1
michael@0	287	paddd xmm7, xmm5 ; xmm7 += accumulated square column differences
michael@0	288	paddd xmm7, xmm1
michael@0	289
michael@0	290	lea rsi, [rsi + rax]
michael@0	291	lea rdi, [rdi + rdx]
michael@0	292
michael@0	293	sub rcx, 1 ;
michael@0	294	jnz .half_horiz_variance16x_h_1 ;
michael@0	295
michael@0	296	pxor xmm1, xmm1
michael@0	297	pxor xmm5, xmm5
michael@0	298
michael@0	299	punpcklwd xmm0, xmm6
michael@0	300	punpckhwd xmm1, xmm6
michael@0	301	psrad xmm0, 16
michael@0	302	psrad xmm1, 16
michael@0	303	paddd xmm0, xmm1
michael@0	304	movdqa xmm1, xmm0
michael@0	305
michael@0	306	movdqa xmm6, xmm7
michael@0	307	punpckldq xmm6, xmm5
michael@0	308	punpckhdq xmm7, xmm5
michael@0	309	paddd xmm6, xmm7
michael@0	310
michael@0	311	punpckldq xmm0, xmm5
michael@0	312	punpckhdq xmm1, xmm5
michael@0	313	paddd xmm0, xmm1
michael@0	314
michael@0	315	movdqa xmm7, xmm6
michael@0	316	movdqa xmm1, xmm0
michael@0	317
michael@0	318	psrldq xmm7, 8
michael@0	319	psrldq xmm1, 8
michael@0	320
michael@0	321	paddd xmm6, xmm7
michael@0	322	paddd xmm0, xmm1
michael@0	323
michael@0	324	mov rsi, arg(5) ;[Sum]
michael@0	325	mov rdi, arg(6) ;[SSE]
michael@0	326
michael@0	327	movd [rsi], xmm0
michael@0	328	movd [rdi], xmm6
michael@0	329
michael@0	330	; begin epilog
michael@0	331	pop rdi
michael@0	332	pop rsi
michael@0	333	RESTORE_GOT
michael@0	334	RESTORE_XMM
michael@0	335	UNSHADOW_ARGS
michael@0	336	pop rbp
michael@0	337	ret