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

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

media/libvpx/vp8/common/x86/subpixel_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	extern sym(vp8_bilinear_filters_x86_8)
michael@0	14
michael@0	15
michael@0	16	%define BLOCK_HEIGHT_WIDTH 4
michael@0	17	%define vp8_filter_weight 128
michael@0	18	%define VP8_FILTER_SHIFT 7
michael@0	19
michael@0	20
michael@0	21	;void vp8_filter_block1d_h6_mmx
michael@0	22	;(
michael@0	23	; unsigned char *src_ptr,
michael@0	24	; unsigned short *output_ptr,
michael@0	25	; unsigned int src_pixels_per_line,
michael@0	26	; unsigned int pixel_step,
michael@0	27	; unsigned int output_height,
michael@0	28	; unsigned int output_width,
michael@0	29	; short * vp8_filter
michael@0	30	;)
michael@0	31	global sym(vp8_filter_block1d_h6_mmx) PRIVATE
michael@0	32	sym(vp8_filter_block1d_h6_mmx):
michael@0	33	push rbp
michael@0	34	mov rbp, rsp
michael@0	35	SHADOW_ARGS_TO_STACK 7
michael@0	36	GET_GOT rbx
michael@0	37	push rsi
michael@0	38	push rdi
michael@0	39	; end prolog
michael@0	40
michael@0	41	mov rdx, arg(6) ;vp8_filter
michael@0	42
michael@0	43	movq mm1, [rdx + 16] ; do both the negative taps first!!!
michael@0	44	movq mm2, [rdx + 32] ;
michael@0	45	movq mm6, [rdx + 48] ;
michael@0	46	movq mm7, [rdx + 64] ;
michael@0	47
michael@0	48	mov rdi, arg(1) ;output_ptr
michael@0	49	mov rsi, arg(0) ;src_ptr
michael@0	50	movsxd rcx, dword ptr arg(4) ;output_height
michael@0	51	movsxd rax, dword ptr arg(5) ;output_width ; destination pitch?
michael@0	52	pxor mm0, mm0 ; mm0 = 00000000
michael@0	53
michael@0	54	.nextrow:
michael@0	55	movq mm3, [rsi-2] ; mm3 = p-2..p5
michael@0	56	movq mm4, mm3 ; mm4 = p-2..p5
michael@0	57	psrlq mm3, 8 ; mm3 = p-1..p5
michael@0	58	punpcklbw mm3, mm0 ; mm3 = p-1..p2
michael@0	59	pmullw mm3, mm1 ; mm3 *= kernel 1 modifiers.
michael@0	60
michael@0	61	movq mm5, mm4 ; mm5 = p-2..p5
michael@0	62	punpckhbw mm4, mm0 ; mm5 = p2..p5
michael@0	63	pmullw mm4, mm7 ; mm5 *= kernel 4 modifiers
michael@0	64	paddsw mm3, mm4 ; mm3 += mm5
michael@0	65
michael@0	66	movq mm4, mm5 ; mm4 = p-2..p5;
michael@0	67	psrlq mm5, 16 ; mm5 = p0..p5;
michael@0	68	punpcklbw mm5, mm0 ; mm5 = p0..p3
michael@0	69	pmullw mm5, mm2 ; mm5 *= kernel 2 modifiers
michael@0	70	paddsw mm3, mm5 ; mm3 += mm5
michael@0	71
michael@0	72	movq mm5, mm4 ; mm5 = p-2..p5
michael@0	73	psrlq mm4, 24 ; mm4 = p1..p5
michael@0	74	punpcklbw mm4, mm0 ; mm4 = p1..p4
michael@0	75	pmullw mm4, mm6 ; mm5 *= kernel 3 modifiers
michael@0	76	paddsw mm3, mm4 ; mm3 += mm5
michael@0	77
michael@0	78	; do outer positive taps
michael@0	79	movd mm4, [rsi+3]
michael@0	80	punpcklbw mm4, mm0 ; mm5 = p3..p6
michael@0	81	pmullw mm4, [rdx+80] ; mm5 *= kernel 0 modifiers
michael@0	82	paddsw mm3, mm4 ; mm3 += mm5
michael@0	83
michael@0	84	punpcklbw mm5, mm0 ; mm5 = p-2..p1
michael@0	85	pmullw mm5, [rdx] ; mm5 *= kernel 5 modifiers
michael@0	86	paddsw mm3, mm5 ; mm3 += mm5
michael@0	87
michael@0	88	paddsw mm3, [GLOBAL(rd)] ; mm3 += round value
michael@0	89	psraw mm3, VP8_FILTER_SHIFT ; mm3 /= 128
michael@0	90	packuswb mm3, mm0 ; pack and unpack to saturate
michael@0	91	punpcklbw mm3, mm0 ;
michael@0	92
michael@0	93	movq [rdi], mm3 ; store the results in the destination
michael@0	94
michael@0	95	%if ABI_IS_32BIT
michael@0	96	add rsi, dword ptr arg(2) ;src_pixels_per_line ; next line
michael@0	97	add rdi, rax;
michael@0	98	%else
michael@0	99	movsxd r8, dword ptr arg(2) ;src_pixels_per_line
michael@0	100	add rdi, rax;
michael@0	101
michael@0	102	add rsi, r8 ; next line
michael@0	103	%endif
michael@0	104
michael@0	105	dec rcx ; decrement count
michael@0	106	jnz .nextrow ; next row
michael@0	107
michael@0	108	; begin epilog
michael@0	109	pop rdi
michael@0	110	pop rsi
michael@0	111	RESTORE_GOT
michael@0	112	UNSHADOW_ARGS
michael@0	113	pop rbp
michael@0	114	ret
michael@0	115
michael@0	116
michael@0	117	;void vp8_filter_block1dc_v6_mmx
michael@0	118	;(
michael@0	119	; short *src_ptr,
michael@0	120	; unsigned char *output_ptr,
michael@0	121	; int output_pitch,
michael@0	122	; unsigned int pixels_per_line,
michael@0	123	; unsigned int pixel_step,
michael@0	124	; unsigned int output_height,
michael@0	125	; unsigned int output_width,
michael@0	126	; short * vp8_filter
michael@0	127	;)
michael@0	128	global sym(vp8_filter_block1dc_v6_mmx) PRIVATE
michael@0	129	sym(vp8_filter_block1dc_v6_mmx):
michael@0	130	push rbp
michael@0	131	mov rbp, rsp
michael@0	132	SHADOW_ARGS_TO_STACK 8
michael@0	133	GET_GOT rbx
michael@0	134	push rsi
michael@0	135	push rdi
michael@0	136	; end prolog
michael@0	137
michael@0	138	movq mm5, [GLOBAL(rd)]
michael@0	139	push rbx
michael@0	140	mov rbx, arg(7) ;vp8_filter
michael@0	141	movq mm1, [rbx + 16] ; do both the negative taps first!!!
michael@0	142	movq mm2, [rbx + 32] ;
michael@0	143	movq mm6, [rbx + 48] ;
michael@0	144	movq mm7, [rbx + 64] ;
michael@0	145
michael@0	146	movsxd rdx, dword ptr arg(3) ;pixels_per_line
michael@0	147	mov rdi, arg(1) ;output_ptr
michael@0	148	mov rsi, arg(0) ;src_ptr
michael@0	149	sub rsi, rdx
michael@0	150	sub rsi, rdx
michael@0	151	movsxd rcx, DWORD PTR arg(5) ;output_height
michael@0	152	movsxd rax, DWORD PTR arg(2) ;output_pitch ; destination pitch?
michael@0	153	pxor mm0, mm0 ; mm0 = 00000000
michael@0	154
michael@0	155
michael@0	156	.nextrow_cv:
michael@0	157	movq mm3, [rsi+rdx] ; mm3 = p0..p8 = row -1
michael@0	158	pmullw mm3, mm1 ; mm3 *= kernel 1 modifiers.
michael@0	159
michael@0	160
michael@0	161	movq mm4, [rsi + 4*rdx] ; mm4 = p0..p3 = row 2
michael@0	162	pmullw mm4, mm7 ; mm4 *= kernel 4 modifiers.
michael@0	163	paddsw mm3, mm4 ; mm3 += mm4
michael@0	164
michael@0	165	movq mm4, [rsi + 2*rdx] ; mm4 = p0..p3 = row 0
michael@0	166	pmullw mm4, mm2 ; mm4 *= kernel 2 modifiers.
michael@0	167	paddsw mm3, mm4 ; mm3 += mm4
michael@0	168
michael@0	169	movq mm4, [rsi] ; mm4 = p0..p3 = row -2
michael@0	170	pmullw mm4, [rbx] ; mm4 *= kernel 0 modifiers.
michael@0	171	paddsw mm3, mm4 ; mm3 += mm4
michael@0	172
michael@0	173
michael@0	174	add rsi, rdx ; move source forward 1 line to avoid 3 * pitch
michael@0	175	movq mm4, [rsi + 2*rdx] ; mm4 = p0..p3 = row 1
michael@0	176	pmullw mm4, mm6 ; mm4 *= kernel 3 modifiers.
michael@0	177	paddsw mm3, mm4 ; mm3 += mm4
michael@0	178
michael@0	179	movq mm4, [rsi + 4*rdx] ; mm4 = p0..p3 = row 3
michael@0	180	pmullw mm4, [rbx +80] ; mm4 *= kernel 3 modifiers.
michael@0	181	paddsw mm3, mm4 ; mm3 += mm4
michael@0	182
michael@0	183
michael@0	184	paddsw mm3, mm5 ; mm3 += round value
michael@0	185	psraw mm3, VP8_FILTER_SHIFT ; mm3 /= 128
michael@0	186	packuswb mm3, mm0 ; pack and saturate
michael@0	187
michael@0	188	movd [rdi],mm3 ; store the results in the destination
michael@0	189	; the subsequent iterations repeat 3 out of 4 of these reads. Since the
michael@0	190	; recon block should be in cache this shouldn't cost much. Its obviously
michael@0	191	; avoidable!!!.
michael@0	192	lea rdi, [rdi+rax] ;
michael@0	193	dec rcx ; decrement count
michael@0	194	jnz .nextrow_cv ; next row
michael@0	195
michael@0	196	pop rbx
michael@0	197
michael@0	198	; begin epilog
michael@0	199	pop rdi
michael@0	200	pop rsi
michael@0	201	RESTORE_GOT
michael@0	202	UNSHADOW_ARGS
michael@0	203	pop rbp
michael@0	204	ret
michael@0	205
michael@0	206
michael@0	207	;void bilinear_predict8x8_mmx
michael@0	208	;(
michael@0	209	; unsigned char *src_ptr,
michael@0	210	; int src_pixels_per_line,
michael@0	211	; int xoffset,
michael@0	212	; int yoffset,
michael@0	213	; unsigned char *dst_ptr,
michael@0	214	; int dst_pitch
michael@0	215	;)
michael@0	216	global sym(vp8_bilinear_predict8x8_mmx) PRIVATE
michael@0	217	sym(vp8_bilinear_predict8x8_mmx):
michael@0	218	push rbp
michael@0	219	mov rbp, rsp
michael@0	220	SHADOW_ARGS_TO_STACK 6
michael@0	221	GET_GOT rbx
michael@0	222	push rsi
michael@0	223	push rdi
michael@0	224	; end prolog
michael@0	225
michael@0	226	;const short *HFilter = vp8_bilinear_filters_x86_8[xoffset];
michael@0	227	;const short *VFilter = vp8_bilinear_filters_x86_8[yoffset];
michael@0	228
michael@0	229	movsxd rax, dword ptr arg(2) ;xoffset
michael@0	230	mov rdi, arg(4) ;dst_ptr ;
michael@0	231
michael@0	232	shl rax, 5 ; offset * 32
michael@0	233	lea rcx, [GLOBAL(sym(vp8_bilinear_filters_x86_8))]
michael@0	234
michael@0	235	add rax, rcx ; HFilter
michael@0	236	mov rsi, arg(0) ;src_ptr ;
michael@0	237
michael@0	238	movsxd rdx, dword ptr arg(5) ;dst_pitch
michael@0	239	movq mm1, [rax] ;
michael@0	240
michael@0	241	movq mm2, [rax+16] ;
michael@0	242	movsxd rax, dword ptr arg(3) ;yoffset
michael@0	243
michael@0	244	pxor mm0, mm0 ;
michael@0	245
michael@0	246	shl rax, 5 ; offset*32
michael@0	247	add rax, rcx ; VFilter
michael@0	248
michael@0	249	lea rcx, [rdi+rdx*8] ;
michael@0	250	movsxd rdx, dword ptr arg(1) ;src_pixels_per_line ;
michael@0	251
michael@0	252
michael@0	253
michael@0	254	; get the first horizontal line done ;
michael@0	255	movq mm3, [rsi] ; xx 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14
michael@0	256	movq mm4, mm3 ; make a copy of current line
michael@0	257
michael@0	258	punpcklbw mm3, mm0 ; xx 00 01 02 03 04 05 06
michael@0	259	punpckhbw mm4, mm0 ;
michael@0	260
michael@0	261	pmullw mm3, mm1 ;
michael@0	262	pmullw mm4, mm1 ;
michael@0	263
michael@0	264	movq mm5, [rsi+1] ;
michael@0	265	movq mm6, mm5 ;
michael@0	266
michael@0	267	punpcklbw mm5, mm0 ;
michael@0	268	punpckhbw mm6, mm0 ;
michael@0	269
michael@0	270	pmullw mm5, mm2 ;
michael@0	271	pmullw mm6, mm2 ;
michael@0	272
michael@0	273	paddw mm3, mm5 ;
michael@0	274	paddw mm4, mm6 ;
michael@0	275
michael@0	276	paddw mm3, [GLOBAL(rd)] ; xmm3 += round value
michael@0	277	psraw mm3, VP8_FILTER_SHIFT ; xmm3 /= 128
michael@0	278
michael@0	279	paddw mm4, [GLOBAL(rd)] ;
michael@0	280	psraw mm4, VP8_FILTER_SHIFT ;
michael@0	281
michael@0	282	movq mm7, mm3 ;
michael@0	283	packuswb mm7, mm4 ;
michael@0	284
michael@0	285	add rsi, rdx ; next line
michael@0	286	.next_row_8x8:
michael@0	287	movq mm3, [rsi] ; xx 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14
michael@0	288	movq mm4, mm3 ; make a copy of current line
michael@0	289
michael@0	290	punpcklbw mm3, mm0 ; xx 00 01 02 03 04 05 06
michael@0	291	punpckhbw mm4, mm0 ;
michael@0	292
michael@0	293	pmullw mm3, mm1 ;
michael@0	294	pmullw mm4, mm1 ;
michael@0	295
michael@0	296	movq mm5, [rsi+1] ;
michael@0	297	movq mm6, mm5 ;
michael@0	298
michael@0	299	punpcklbw mm5, mm0 ;
michael@0	300	punpckhbw mm6, mm0 ;
michael@0	301
michael@0	302	pmullw mm5, mm2 ;
michael@0	303	pmullw mm6, mm2 ;
michael@0	304
michael@0	305	paddw mm3, mm5 ;
michael@0	306	paddw mm4, mm6 ;
michael@0	307
michael@0	308	movq mm5, mm7 ;
michael@0	309	movq mm6, mm7 ;
michael@0	310
michael@0	311	punpcklbw mm5, mm0 ;
michael@0	312	punpckhbw mm6, mm0
michael@0	313
michael@0	314	pmullw mm5, [rax] ;
michael@0	315	pmullw mm6, [rax] ;
michael@0	316
michael@0	317	paddw mm3, [GLOBAL(rd)] ; xmm3 += round value
michael@0	318	psraw mm3, VP8_FILTER_SHIFT ; xmm3 /= 128
michael@0	319
michael@0	320	paddw mm4, [GLOBAL(rd)] ;
michael@0	321	psraw mm4, VP8_FILTER_SHIFT ;
michael@0	322
michael@0	323	movq mm7, mm3 ;
michael@0	324	packuswb mm7, mm4 ;
michael@0	325
michael@0	326
michael@0	327	pmullw mm3, [rax+16] ;
michael@0	328	pmullw mm4, [rax+16] ;
michael@0	329
michael@0	330	paddw mm3, mm5 ;
michael@0	331	paddw mm4, mm6 ;
michael@0	332
michael@0	333
michael@0	334	paddw mm3, [GLOBAL(rd)] ; xmm3 += round value
michael@0	335	psraw mm3, VP8_FILTER_SHIFT ; xmm3 /= 128
michael@0	336
michael@0	337	paddw mm4, [GLOBAL(rd)] ;
michael@0	338	psraw mm4, VP8_FILTER_SHIFT ;
michael@0	339
michael@0	340	packuswb mm3, mm4
michael@0	341
michael@0	342	movq [rdi], mm3 ; store the results in the destination
michael@0	343
michael@0	344	%if ABI_IS_32BIT
michael@0	345	add rsi, rdx ; next line
michael@0	346	add rdi, dword ptr arg(5) ;dst_pitch ;
michael@0	347	%else
michael@0	348	movsxd r8, dword ptr arg(5) ;dst_pitch
michael@0	349	add rsi, rdx ; next line
michael@0	350	add rdi, r8 ;dst_pitch
michael@0	351	%endif
michael@0	352	cmp rdi, rcx ;
michael@0	353	jne .next_row_8x8
michael@0	354
michael@0	355	; begin epilog
michael@0	356	pop rdi
michael@0	357	pop rsi
michael@0	358	RESTORE_GOT
michael@0	359	UNSHADOW_ARGS
michael@0	360	pop rbp
michael@0	361	ret
michael@0	362
michael@0	363
michael@0	364	;void bilinear_predict8x4_mmx
michael@0	365	;(
michael@0	366	; unsigned char *src_ptr,
michael@0	367	; int src_pixels_per_line,
michael@0	368	; int xoffset,
michael@0	369	; int yoffset,
michael@0	370	; unsigned char *dst_ptr,
michael@0	371	; int dst_pitch
michael@0	372	;)
michael@0	373	global sym(vp8_bilinear_predict8x4_mmx) PRIVATE
michael@0	374	sym(vp8_bilinear_predict8x4_mmx):
michael@0	375	push rbp
michael@0	376	mov rbp, rsp
michael@0	377	SHADOW_ARGS_TO_STACK 6
michael@0	378	GET_GOT rbx
michael@0	379	push rsi
michael@0	380	push rdi
michael@0	381	; end prolog
michael@0	382
michael@0	383	;const short *HFilter = vp8_bilinear_filters_x86_8[xoffset];
michael@0	384	;const short *VFilter = vp8_bilinear_filters_x86_8[yoffset];
michael@0	385
michael@0	386	movsxd rax, dword ptr arg(2) ;xoffset
michael@0	387	mov rdi, arg(4) ;dst_ptr ;
michael@0	388
michael@0	389	lea rcx, [GLOBAL(sym(vp8_bilinear_filters_x86_8))]
michael@0	390	shl rax, 5
michael@0	391
michael@0	392	mov rsi, arg(0) ;src_ptr ;
michael@0	393	add rax, rcx
michael@0	394
michael@0	395	movsxd rdx, dword ptr arg(5) ;dst_pitch
michael@0	396	movq mm1, [rax] ;
michael@0	397
michael@0	398	movq mm2, [rax+16] ;
michael@0	399	movsxd rax, dword ptr arg(3) ;yoffset
michael@0	400
michael@0	401	pxor mm0, mm0 ;
michael@0	402	shl rax, 5
michael@0	403
michael@0	404	add rax, rcx
michael@0	405	lea rcx, [rdi+rdx*4] ;
michael@0	406
michael@0	407	movsxd rdx, dword ptr arg(1) ;src_pixels_per_line ;
michael@0	408
michael@0	409	; get the first horizontal line done ;
michael@0	410	movq mm3, [rsi] ; xx 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14
michael@0	411	movq mm4, mm3 ; make a copy of current line
michael@0	412
michael@0	413	punpcklbw mm3, mm0 ; xx 00 01 02 03 04 05 06
michael@0	414	punpckhbw mm4, mm0 ;
michael@0	415
michael@0	416	pmullw mm3, mm1 ;
michael@0	417	pmullw mm4, mm1 ;
michael@0	418
michael@0	419	movq mm5, [rsi+1] ;
michael@0	420	movq mm6, mm5 ;
michael@0	421
michael@0	422	punpcklbw mm5, mm0 ;
michael@0	423	punpckhbw mm6, mm0 ;
michael@0	424
michael@0	425	pmullw mm5, mm2 ;
michael@0	426	pmullw mm6, mm2 ;
michael@0	427
michael@0	428	paddw mm3, mm5 ;
michael@0	429	paddw mm4, mm6 ;
michael@0	430
michael@0	431	paddw mm3, [GLOBAL(rd)] ; xmm3 += round value
michael@0	432	psraw mm3, VP8_FILTER_SHIFT ; xmm3 /= 128
michael@0	433
michael@0	434	paddw mm4, [GLOBAL(rd)] ;
michael@0	435	psraw mm4, VP8_FILTER_SHIFT ;
michael@0	436
michael@0	437	movq mm7, mm3 ;
michael@0	438	packuswb mm7, mm4 ;
michael@0	439
michael@0	440	add rsi, rdx ; next line
michael@0	441	.next_row_8x4:
michael@0	442	movq mm3, [rsi] ; xx 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14
michael@0	443	movq mm4, mm3 ; make a copy of current line
michael@0	444
michael@0	445	punpcklbw mm3, mm0 ; xx 00 01 02 03 04 05 06
michael@0	446	punpckhbw mm4, mm0 ;
michael@0	447
michael@0	448	pmullw mm3, mm1 ;
michael@0	449	pmullw mm4, mm1 ;
michael@0	450
michael@0	451	movq mm5, [rsi+1] ;
michael@0	452	movq mm6, mm5 ;
michael@0	453
michael@0	454	punpcklbw mm5, mm0 ;
michael@0	455	punpckhbw mm6, mm0 ;
michael@0	456
michael@0	457	pmullw mm5, mm2 ;
michael@0	458	pmullw mm6, mm2 ;
michael@0	459
michael@0	460	paddw mm3, mm5 ;
michael@0	461	paddw mm4, mm6 ;
michael@0	462
michael@0	463	movq mm5, mm7 ;
michael@0	464	movq mm6, mm7 ;
michael@0	465
michael@0	466	punpcklbw mm5, mm0 ;
michael@0	467	punpckhbw mm6, mm0
michael@0	468
michael@0	469	pmullw mm5, [rax] ;
michael@0	470	pmullw mm6, [rax] ;
michael@0	471
michael@0	472	paddw mm3, [GLOBAL(rd)] ; xmm3 += round value
michael@0	473	psraw mm3, VP8_FILTER_SHIFT ; xmm3 /= 128
michael@0	474
michael@0	475	paddw mm4, [GLOBAL(rd)] ;
michael@0	476	psraw mm4, VP8_FILTER_SHIFT ;
michael@0	477
michael@0	478	movq mm7, mm3 ;
michael@0	479	packuswb mm7, mm4 ;
michael@0	480
michael@0	481
michael@0	482	pmullw mm3, [rax+16] ;
michael@0	483	pmullw mm4, [rax+16] ;
michael@0	484
michael@0	485	paddw mm3, mm5 ;
michael@0	486	paddw mm4, mm6 ;
michael@0	487
michael@0	488
michael@0	489	paddw mm3, [GLOBAL(rd)] ; xmm3 += round value
michael@0	490	psraw mm3, VP8_FILTER_SHIFT ; xmm3 /= 128
michael@0	491
michael@0	492	paddw mm4, [GLOBAL(rd)] ;
michael@0	493	psraw mm4, VP8_FILTER_SHIFT ;
michael@0	494
michael@0	495	packuswb mm3, mm4
michael@0	496
michael@0	497	movq [rdi], mm3 ; store the results in the destination
michael@0	498
michael@0	499	%if ABI_IS_32BIT
michael@0	500	add rsi, rdx ; next line
michael@0	501	add rdi, dword ptr arg(5) ;dst_pitch ;
michael@0	502	%else
michael@0	503	movsxd r8, dword ptr arg(5) ;dst_pitch
michael@0	504	add rsi, rdx ; next line
michael@0	505	add rdi, r8
michael@0	506	%endif
michael@0	507	cmp rdi, rcx ;
michael@0	508	jne .next_row_8x4
michael@0	509
michael@0	510	; begin epilog
michael@0	511	pop rdi
michael@0	512	pop rsi
michael@0	513	RESTORE_GOT
michael@0	514	UNSHADOW_ARGS
michael@0	515	pop rbp
michael@0	516	ret
michael@0	517
michael@0	518
michael@0	519	;void bilinear_predict4x4_mmx
michael@0	520	;(
michael@0	521	; unsigned char *src_ptr,
michael@0	522	; int src_pixels_per_line,
michael@0	523	; int xoffset,
michael@0	524	; int yoffset,
michael@0	525	; unsigned char *dst_ptr,
michael@0	526	; int dst_pitch
michael@0	527	;)
michael@0	528	global sym(vp8_bilinear_predict4x4_mmx) PRIVATE
michael@0	529	sym(vp8_bilinear_predict4x4_mmx):
michael@0	530	push rbp
michael@0	531	mov rbp, rsp
michael@0	532	SHADOW_ARGS_TO_STACK 6
michael@0	533	GET_GOT rbx
michael@0	534	push rsi
michael@0	535	push rdi
michael@0	536	; end prolog
michael@0	537
michael@0	538	;const short *HFilter = vp8_bilinear_filters_x86_8[xoffset];
michael@0	539	;const short *VFilter = vp8_bilinear_filters_x86_8[yoffset];
michael@0	540
michael@0	541	movsxd rax, dword ptr arg(2) ;xoffset
michael@0	542	mov rdi, arg(4) ;dst_ptr ;
michael@0	543
michael@0	544	lea rcx, [GLOBAL(sym(vp8_bilinear_filters_x86_8))]
michael@0	545	shl rax, 5
michael@0	546
michael@0	547	add rax, rcx ; HFilter
michael@0	548	mov rsi, arg(0) ;src_ptr ;
michael@0	549
michael@0	550	movsxd rdx, dword ptr arg(5) ;ldst_pitch
michael@0	551	movq mm1, [rax] ;
michael@0	552
michael@0	553	movq mm2, [rax+16] ;
michael@0	554	movsxd rax, dword ptr arg(3) ;yoffset
michael@0	555
michael@0	556	pxor mm0, mm0 ;
michael@0	557	shl rax, 5
michael@0	558
michael@0	559	add rax, rcx
michael@0	560	lea rcx, [rdi+rdx*4] ;
michael@0	561
michael@0	562	movsxd rdx, dword ptr arg(1) ;src_pixels_per_line ;
michael@0	563
michael@0	564	; get the first horizontal line done ;
michael@0	565	movd mm3, [rsi] ; xx 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14
michael@0	566	punpcklbw mm3, mm0 ; xx 00 01 02 03 04 05 06
michael@0	567
michael@0	568	pmullw mm3, mm1 ;
michael@0	569	movd mm5, [rsi+1] ;
michael@0	570
michael@0	571	punpcklbw mm5, mm0 ;
michael@0	572	pmullw mm5, mm2 ;
michael@0	573
michael@0	574	paddw mm3, mm5 ;
michael@0	575	paddw mm3, [GLOBAL(rd)] ; xmm3 += round value
michael@0	576
michael@0	577	psraw mm3, VP8_FILTER_SHIFT ; xmm3 /= 128
michael@0	578
michael@0	579	movq mm7, mm3 ;
michael@0	580	packuswb mm7, mm0 ;
michael@0	581
michael@0	582	add rsi, rdx ; next line
michael@0	583	.next_row_4x4:
michael@0	584	movd mm3, [rsi] ; xx 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14
michael@0	585	punpcklbw mm3, mm0 ; xx 00 01 02 03 04 05 06
michael@0	586
michael@0	587	pmullw mm3, mm1 ;
michael@0	588	movd mm5, [rsi+1] ;
michael@0	589
michael@0	590	punpcklbw mm5, mm0 ;
michael@0	591	pmullw mm5, mm2 ;
michael@0	592
michael@0	593	paddw mm3, mm5 ;
michael@0	594
michael@0	595	movq mm5, mm7 ;
michael@0	596	punpcklbw mm5, mm0 ;
michael@0	597
michael@0	598	pmullw mm5, [rax] ;
michael@0	599	paddw mm3, [GLOBAL(rd)] ; xmm3 += round value
michael@0	600
michael@0	601	psraw mm3, VP8_FILTER_SHIFT ; xmm3 /= 128
michael@0	602	movq mm7, mm3 ;
michael@0	603
michael@0	604	packuswb mm7, mm0 ;
michael@0	605
michael@0	606	pmullw mm3, [rax+16] ;
michael@0	607	paddw mm3, mm5 ;
michael@0	608
michael@0	609
michael@0	610	paddw mm3, [GLOBAL(rd)] ; xmm3 += round value
michael@0	611	psraw mm3, VP8_FILTER_SHIFT ; xmm3 /= 128
michael@0	612
michael@0	613	packuswb mm3, mm0
michael@0	614	movd [rdi], mm3 ; store the results in the destination
michael@0	615
michael@0	616	%if ABI_IS_32BIT
michael@0	617	add rsi, rdx ; next line
michael@0	618	add rdi, dword ptr arg(5) ;dst_pitch ;
michael@0	619	%else
michael@0	620	movsxd r8, dword ptr arg(5) ;dst_pitch ;
michael@0	621	add rsi, rdx ; next line
michael@0	622	add rdi, r8
michael@0	623	%endif
michael@0	624
michael@0	625	cmp rdi, rcx ;
michael@0	626	jne .next_row_4x4
michael@0	627
michael@0	628	; begin epilog
michael@0	629	pop rdi
michael@0	630	pop rsi
michael@0	631	RESTORE_GOT
michael@0	632	UNSHADOW_ARGS
michael@0	633	pop rbp
michael@0	634	ret
michael@0	635
michael@0	636
michael@0	637
michael@0	638	SECTION_RODATA
michael@0	639	align 16
michael@0	640	rd:
michael@0	641	times 4 dw 0x40
michael@0	642
michael@0	643	align 16
michael@0	644	global HIDDEN_DATA(sym(vp8_six_tap_mmx))
michael@0	645	sym(vp8_six_tap_mmx):
michael@0	646	times 8 dw 0
michael@0	647	times 8 dw 0
michael@0	648	times 8 dw 128
michael@0	649	times 8 dw 0
michael@0	650	times 8 dw 0
michael@0	651	times 8 dw 0
michael@0	652
michael@0	653	times 8 dw 0
michael@0	654	times 8 dw -6
michael@0	655	times 8 dw 123
michael@0	656	times 8 dw 12
michael@0	657	times 8 dw -1
michael@0	658	times 8 dw 0
michael@0	659
michael@0	660	times 8 dw 2
michael@0	661	times 8 dw -11
michael@0	662	times 8 dw 108
michael@0	663	times 8 dw 36
michael@0	664	times 8 dw -8
michael@0	665	times 8 dw 1
michael@0	666
michael@0	667	times 8 dw 0
michael@0	668	times 8 dw -9
michael@0	669	times 8 dw 93
michael@0	670	times 8 dw 50
michael@0	671	times 8 dw -6
michael@0	672	times 8 dw 0
michael@0	673
michael@0	674	times 8 dw 3
michael@0	675	times 8 dw -16
michael@0	676	times 8 dw 77
michael@0	677	times 8 dw 77
michael@0	678	times 8 dw -16
michael@0	679	times 8 dw 3
michael@0	680
michael@0	681	times 8 dw 0
michael@0	682	times 8 dw -6
michael@0	683	times 8 dw 50
michael@0	684	times 8 dw 93
michael@0	685	times 8 dw -9
michael@0	686	times 8 dw 0
michael@0	687
michael@0	688	times 8 dw 1
michael@0	689	times 8 dw -8
michael@0	690	times 8 dw 36
michael@0	691	times 8 dw 108
michael@0	692	times 8 dw -11
michael@0	693	times 8 dw 2
michael@0	694
michael@0	695	times 8 dw 0
michael@0	696	times 8 dw -1
michael@0	697	times 8 dw 12
michael@0	698	times 8 dw 123
michael@0	699	times 8 dw -6
michael@0	700	times 8 dw 0
michael@0	701
michael@0	702