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

media/libvpx/vp9/encoder/x86/vp9_sad_mmx.asm@6474c204b198 (annotated)

media/libvpx/vp9/encoder/x86/vp9_sad_mmx.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
michael@0	12	%include "vpx_ports/x86_abi_support.asm"
michael@0	13
michael@0	14	global sym(vp9_sad16x16_mmx) PRIVATE
michael@0	15	global sym(vp9_sad8x16_mmx) PRIVATE
michael@0	16	global sym(vp9_sad8x8_mmx) PRIVATE
michael@0	17	global sym(vp9_sad4x4_mmx) PRIVATE
michael@0	18	global sym(vp9_sad16x8_mmx) PRIVATE
michael@0	19
michael@0	20	;unsigned int vp9_sad16x16_mmx(
michael@0	21	; unsigned char *src_ptr,
michael@0	22	; int src_stride,
michael@0	23	; unsigned char *ref_ptr,
michael@0	24	; int ref_stride)
michael@0	25	sym(vp9_sad16x16_mmx):
michael@0	26	push rbp
michael@0	27	mov rbp, rsp
michael@0	28	SHADOW_ARGS_TO_STACK 4
michael@0	29	push rsi
michael@0	30	push rdi
michael@0	31	; end prolog
michael@0	32
michael@0	33	mov rsi, arg(0) ;src_ptr
michael@0	34	mov rdi, arg(2) ;ref_ptr
michael@0	35
michael@0	36	movsxd rax, dword ptr arg(1) ;src_stride
michael@0	37	movsxd rdx, dword ptr arg(3) ;ref_stride
michael@0	38
michael@0	39	lea rcx, [rsi+rax*8]
michael@0	40
michael@0	41	lea rcx, [rcx+rax*8]
michael@0	42	pxor mm7, mm7
michael@0	43
michael@0	44	pxor mm6, mm6
michael@0	45
michael@0	46	.x16x16sad_mmx_loop:
michael@0	47
michael@0	48	movq mm0, QWORD PTR [rsi]
michael@0	49	movq mm2, QWORD PTR [rsi+8]
michael@0	50
michael@0	51	movq mm1, QWORD PTR [rdi]
michael@0	52	movq mm3, QWORD PTR [rdi+8]
michael@0	53
michael@0	54	movq mm4, mm0
michael@0	55	movq mm5, mm2
michael@0	56
michael@0	57	psubusb mm0, mm1
michael@0	58	psubusb mm1, mm4
michael@0	59
michael@0	60	psubusb mm2, mm3
michael@0	61	psubusb mm3, mm5
michael@0	62
michael@0	63	por mm0, mm1
michael@0	64	por mm2, mm3
michael@0	65
michael@0	66	movq mm1, mm0
michael@0	67	movq mm3, mm2
michael@0	68
michael@0	69	punpcklbw mm0, mm6
michael@0	70	punpcklbw mm2, mm6
michael@0	71
michael@0	72	punpckhbw mm1, mm6
michael@0	73	punpckhbw mm3, mm6
michael@0	74
michael@0	75	paddw mm0, mm2
michael@0	76	paddw mm1, mm3
michael@0	77
michael@0	78
michael@0	79	lea rsi, [rsi+rax]
michael@0	80	add rdi, rdx
michael@0	81
michael@0	82	paddw mm7, mm0
michael@0	83	paddw mm7, mm1
michael@0	84
michael@0	85	cmp rsi, rcx
michael@0	86	jne .x16x16sad_mmx_loop
michael@0	87
michael@0	88
michael@0	89	movq mm0, mm7
michael@0	90
michael@0	91	punpcklwd mm0, mm6
michael@0	92	punpckhwd mm7, mm6
michael@0	93
michael@0	94	paddw mm0, mm7
michael@0	95	movq mm7, mm0
michael@0	96
michael@0	97
michael@0	98	psrlq mm0, 32
michael@0	99	paddw mm7, mm0
michael@0	100
michael@0	101	movq rax, mm7
michael@0	102
michael@0	103	pop rdi
michael@0	104	pop rsi
michael@0	105	mov rsp, rbp
michael@0	106	; begin epilog
michael@0	107	UNSHADOW_ARGS
michael@0	108	pop rbp
michael@0	109	ret
michael@0	110
michael@0	111
michael@0	112	;unsigned int vp9_sad8x16_mmx(
michael@0	113	; unsigned char *src_ptr,
michael@0	114	; int src_stride,
michael@0	115	; unsigned char *ref_ptr,
michael@0	116	; int ref_stride)
michael@0	117	sym(vp9_sad8x16_mmx):
michael@0	118	push rbp
michael@0	119	mov rbp, rsp
michael@0	120	SHADOW_ARGS_TO_STACK 4
michael@0	121	push rsi
michael@0	122	push rdi
michael@0	123	; end prolog
michael@0	124
michael@0	125	mov rsi, arg(0) ;src_ptr
michael@0	126	mov rdi, arg(2) ;ref_ptr
michael@0	127
michael@0	128	movsxd rax, dword ptr arg(1) ;src_stride
michael@0	129	movsxd rdx, dword ptr arg(3) ;ref_stride
michael@0	130
michael@0	131	lea rcx, [rsi+rax*8]
michael@0	132
michael@0	133	lea rcx, [rcx+rax*8]
michael@0	134	pxor mm7, mm7
michael@0	135
michael@0	136	pxor mm6, mm6
michael@0	137
michael@0	138	.x8x16sad_mmx_loop:
michael@0	139
michael@0	140	movq mm0, QWORD PTR [rsi]
michael@0	141	movq mm1, QWORD PTR [rdi]
michael@0	142
michael@0	143	movq mm2, mm0
michael@0	144	psubusb mm0, mm1
michael@0	145
michael@0	146	psubusb mm1, mm2
michael@0	147	por mm0, mm1
michael@0	148
michael@0	149	movq mm2, mm0
michael@0	150	punpcklbw mm0, mm6
michael@0	151
michael@0	152	punpckhbw mm2, mm6
michael@0	153	lea rsi, [rsi+rax]
michael@0	154
michael@0	155	add rdi, rdx
michael@0	156	paddw mm7, mm0
michael@0	157
michael@0	158	paddw mm7, mm2
michael@0	159	cmp rsi, rcx
michael@0	160
michael@0	161	jne .x8x16sad_mmx_loop
michael@0	162
michael@0	163	movq mm0, mm7
michael@0	164	punpcklwd mm0, mm6
michael@0	165
michael@0	166	punpckhwd mm7, mm6
michael@0	167	paddw mm0, mm7
michael@0	168
michael@0	169	movq mm7, mm0
michael@0	170	psrlq mm0, 32
michael@0	171
michael@0	172	paddw mm7, mm0
michael@0	173	movq rax, mm7
michael@0	174
michael@0	175	pop rdi
michael@0	176	pop rsi
michael@0	177	mov rsp, rbp
michael@0	178	; begin epilog
michael@0	179	UNSHADOW_ARGS
michael@0	180	pop rbp
michael@0	181	ret
michael@0	182
michael@0	183
michael@0	184	;unsigned int vp9_sad8x8_mmx(
michael@0	185	; unsigned char *src_ptr,
michael@0	186	; int src_stride,
michael@0	187	; unsigned char *ref_ptr,
michael@0	188	; int ref_stride)
michael@0	189	sym(vp9_sad8x8_mmx):
michael@0	190	push rbp
michael@0	191	mov rbp, rsp
michael@0	192	SHADOW_ARGS_TO_STACK 4
michael@0	193	push rsi
michael@0	194	push rdi
michael@0	195	; end prolog
michael@0	196
michael@0	197	mov rsi, arg(0) ;src_ptr
michael@0	198	mov rdi, arg(2) ;ref_ptr
michael@0	199
michael@0	200	movsxd rax, dword ptr arg(1) ;src_stride
michael@0	201	movsxd rdx, dword ptr arg(3) ;ref_stride
michael@0	202
michael@0	203	lea rcx, [rsi+rax*8]
michael@0	204	pxor mm7, mm7
michael@0	205
michael@0	206	pxor mm6, mm6
michael@0	207
michael@0	208	.x8x8sad_mmx_loop:
michael@0	209
michael@0	210	movq mm0, QWORD PTR [rsi]
michael@0	211	movq mm1, QWORD PTR [rdi]
michael@0	212
michael@0	213	movq mm2, mm0
michael@0	214	psubusb mm0, mm1
michael@0	215
michael@0	216	psubusb mm1, mm2
michael@0	217	por mm0, mm1
michael@0	218
michael@0	219	movq mm2, mm0
michael@0	220	punpcklbw mm0, mm6
michael@0	221
michael@0	222	punpckhbw mm2, mm6
michael@0	223	paddw mm0, mm2
michael@0	224
michael@0	225	lea rsi, [rsi+rax]
michael@0	226	add rdi, rdx
michael@0	227
michael@0	228	paddw mm7, mm0
michael@0	229	cmp rsi, rcx
michael@0	230
michael@0	231	jne .x8x8sad_mmx_loop
michael@0	232
michael@0	233	movq mm0, mm7
michael@0	234	punpcklwd mm0, mm6
michael@0	235
michael@0	236	punpckhwd mm7, mm6
michael@0	237	paddw mm0, mm7
michael@0	238
michael@0	239	movq mm7, mm0
michael@0	240	psrlq mm0, 32
michael@0	241
michael@0	242	paddw mm7, mm0
michael@0	243	movq rax, mm7
michael@0	244
michael@0	245	pop rdi
michael@0	246	pop rsi
michael@0	247	mov rsp, rbp
michael@0	248	; begin epilog
michael@0	249	UNSHADOW_ARGS
michael@0	250	pop rbp
michael@0	251	ret
michael@0	252
michael@0	253
michael@0	254	;unsigned int vp9_sad4x4_mmx(
michael@0	255	; unsigned char *src_ptr,
michael@0	256	; int src_stride,
michael@0	257	; unsigned char *ref_ptr,
michael@0	258	; int ref_stride)
michael@0	259	sym(vp9_sad4x4_mmx):
michael@0	260	push rbp
michael@0	261	mov rbp, rsp
michael@0	262	SHADOW_ARGS_TO_STACK 4
michael@0	263	push rsi
michael@0	264	push rdi
michael@0	265	; end prolog
michael@0	266
michael@0	267	mov rsi, arg(0) ;src_ptr
michael@0	268	mov rdi, arg(2) ;ref_ptr
michael@0	269
michael@0	270	movsxd rax, dword ptr arg(1) ;src_stride
michael@0	271	movsxd rdx, dword ptr arg(3) ;ref_stride
michael@0	272
michael@0	273	movd mm0, DWORD PTR [rsi]
michael@0	274	movd mm1, DWORD PTR [rdi]
michael@0	275
michael@0	276	movd mm2, DWORD PTR [rsi+rax]
michael@0	277	movd mm3, DWORD PTR [rdi+rdx]
michael@0	278
michael@0	279	punpcklbw mm0, mm2
michael@0	280	punpcklbw mm1, mm3
michael@0	281
michael@0	282	movq mm2, mm0
michael@0	283	psubusb mm0, mm1
michael@0	284
michael@0	285	psubusb mm1, mm2
michael@0	286	por mm0, mm1
michael@0	287
michael@0	288	movq mm2, mm0
michael@0	289	pxor mm3, mm3
michael@0	290
michael@0	291	punpcklbw mm0, mm3
michael@0	292	punpckhbw mm2, mm3
michael@0	293
michael@0	294	paddw mm0, mm2
michael@0	295
michael@0	296	lea rsi, [rsi+rax*2]
michael@0	297	lea rdi, [rdi+rdx*2]
michael@0	298
michael@0	299	movd mm4, DWORD PTR [rsi]
michael@0	300	movd mm5, DWORD PTR [rdi]
michael@0	301
michael@0	302	movd mm6, DWORD PTR [rsi+rax]
michael@0	303	movd mm7, DWORD PTR [rdi+rdx]
michael@0	304
michael@0	305	punpcklbw mm4, mm6
michael@0	306	punpcklbw mm5, mm7
michael@0	307
michael@0	308	movq mm6, mm4
michael@0	309	psubusb mm4, mm5
michael@0	310
michael@0	311	psubusb mm5, mm6
michael@0	312	por mm4, mm5
michael@0	313
michael@0	314	movq mm5, mm4
michael@0	315	punpcklbw mm4, mm3
michael@0	316
michael@0	317	punpckhbw mm5, mm3
michael@0	318	paddw mm4, mm5
michael@0	319
michael@0	320	paddw mm0, mm4
michael@0	321	movq mm1, mm0
michael@0	322
michael@0	323	punpcklwd mm0, mm3
michael@0	324	punpckhwd mm1, mm3
michael@0	325
michael@0	326	paddw mm0, mm1
michael@0	327	movq mm1, mm0
michael@0	328
michael@0	329	psrlq mm0, 32
michael@0	330	paddw mm0, mm1
michael@0	331
michael@0	332	movq rax, mm0
michael@0	333
michael@0	334	pop rdi
michael@0	335	pop rsi
michael@0	336	mov rsp, rbp
michael@0	337	; begin epilog
michael@0	338	UNSHADOW_ARGS
michael@0	339	pop rbp
michael@0	340	ret
michael@0	341
michael@0	342
michael@0	343	;unsigned int vp9_sad16x8_mmx(
michael@0	344	; unsigned char *src_ptr,
michael@0	345	; int src_stride,
michael@0	346	; unsigned char *ref_ptr,
michael@0	347	; int ref_stride)
michael@0	348	sym(vp9_sad16x8_mmx):
michael@0	349	push rbp
michael@0	350	mov rbp, rsp
michael@0	351	SHADOW_ARGS_TO_STACK 4
michael@0	352	push rsi
michael@0	353	push rdi
michael@0	354	; end prolog
michael@0	355
michael@0	356	mov rsi, arg(0) ;src_ptr
michael@0	357	mov rdi, arg(2) ;ref_ptr
michael@0	358
michael@0	359	movsxd rax, dword ptr arg(1) ;src_stride
michael@0	360	movsxd rdx, dword ptr arg(3) ;ref_stride
michael@0	361
michael@0	362	lea rcx, [rsi+rax*8]
michael@0	363	pxor mm7, mm7
michael@0	364
michael@0	365	pxor mm6, mm6
michael@0	366
michael@0	367	.x16x8sad_mmx_loop:
michael@0	368
michael@0	369	movq mm0, [rsi]
michael@0	370	movq mm1, [rdi]
michael@0	371
michael@0	372	movq mm2, [rsi+8]
michael@0	373	movq mm3, [rdi+8]
michael@0	374
michael@0	375	movq mm4, mm0
michael@0	376	movq mm5, mm2
michael@0	377
michael@0	378	psubusb mm0, mm1
michael@0	379	psubusb mm1, mm4
michael@0	380
michael@0	381	psubusb mm2, mm3
michael@0	382	psubusb mm3, mm5
michael@0	383
michael@0	384	por mm0, mm1
michael@0	385	por mm2, mm3
michael@0	386
michael@0	387	movq mm1, mm0
michael@0	388	movq mm3, mm2
michael@0	389
michael@0	390	punpcklbw mm0, mm6
michael@0	391	punpckhbw mm1, mm6
michael@0	392
michael@0	393	punpcklbw mm2, mm6
michael@0	394	punpckhbw mm3, mm6
michael@0	395
michael@0	396
michael@0	397	paddw mm0, mm2
michael@0	398	paddw mm1, mm3
michael@0	399
michael@0	400	paddw mm0, mm1
michael@0	401	lea rsi, [rsi+rax]
michael@0	402
michael@0	403	add rdi, rdx
michael@0	404	paddw mm7, mm0
michael@0	405
michael@0	406	cmp rsi, rcx
michael@0	407	jne .x16x8sad_mmx_loop
michael@0	408
michael@0	409	movq mm0, mm7
michael@0	410	punpcklwd mm0, mm6
michael@0	411
michael@0	412	punpckhwd mm7, mm6
michael@0	413	paddw mm0, mm7
michael@0	414
michael@0	415	movq mm7, mm0
michael@0	416	psrlq mm0, 32
michael@0	417
michael@0	418	paddw mm7, mm0
michael@0	419	movq rax, mm7
michael@0	420
michael@0	421	pop rdi
michael@0	422	pop rsi
michael@0	423	mov rsp, rbp
michael@0	424	; begin epilog
michael@0	425	UNSHADOW_ARGS
michael@0	426	pop rbp
michael@0	427	ret