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

media/libvpx/vp9/common/x86/vp9_postproc_sse2.asm@6474c204b198 (annotated)

media/libvpx/vp9/common/x86/vp9_postproc_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
michael@0	12	%include "vpx_ports/x86_abi_support.asm"
michael@0	13
michael@0	14	;void vp9_post_proc_down_and_across_xmm
michael@0	15	;(
michael@0	16	; unsigned char *src_ptr,
michael@0	17	; unsigned char *dst_ptr,
michael@0	18	; int src_pixels_per_line,
michael@0	19	; int dst_pixels_per_line,
michael@0	20	; int rows,
michael@0	21	; int cols,
michael@0	22	; int flimit
michael@0	23	;)
michael@0	24	global sym(vp9_post_proc_down_and_across_xmm) PRIVATE
michael@0	25	sym(vp9_post_proc_down_and_across_xmm):
michael@0	26	push rbp
michael@0	27	mov rbp, rsp
michael@0	28	SHADOW_ARGS_TO_STACK 7
michael@0	29	SAVE_XMM 7
michael@0	30	GET_GOT rbx
michael@0	31	push rsi
michael@0	32	push rdi
michael@0	33	; end prolog
michael@0	34
michael@0	35	%if ABI_IS_32BIT=1 && CONFIG_PIC=1
michael@0	36	ALIGN_STACK 16, rax
michael@0	37	; move the global rd onto the stack, since we don't have enough registers
michael@0	38	; to do PIC addressing
michael@0	39	movdqa xmm0, [GLOBAL(rd42)]
michael@0	40	sub rsp, 16
michael@0	41	movdqa [rsp], xmm0
michael@0	42	%define RD42 [rsp]
michael@0	43	%else
michael@0	44	%define RD42 [GLOBAL(rd42)]
michael@0	45	%endif
michael@0	46
michael@0	47
michael@0	48	movd xmm2, dword ptr arg(6) ;flimit
michael@0	49	punpcklwd xmm2, xmm2
michael@0	50	punpckldq xmm2, xmm2
michael@0	51	punpcklqdq xmm2, xmm2
michael@0	52
michael@0	53	mov rsi, arg(0) ;src_ptr
michael@0	54	mov rdi, arg(1) ;dst_ptr
michael@0	55
michael@0	56	movsxd rcx, DWORD PTR arg(4) ;rows
michael@0	57	movsxd rax, DWORD PTR arg(2) ;src_pixels_per_line ; destination pitch?
michael@0	58	pxor xmm0, xmm0 ; mm0 = 00000000
michael@0	59
michael@0	60	.nextrow:
michael@0	61
michael@0	62	xor rdx, rdx ; clear out rdx for use as loop counter
michael@0	63	.nextcol:
michael@0	64	movq xmm3, QWORD PTR [rsi] ; mm4 = r0 p0..p7
michael@0	65	punpcklbw xmm3, xmm0 ; mm3 = p0..p3
michael@0	66	movdqa xmm1, xmm3 ; mm1 = p0..p3
michael@0	67	psllw xmm3, 2 ;
michael@0	68
michael@0	69	movq xmm5, QWORD PTR [rsi + rax] ; mm4 = r1 p0..p7
michael@0	70	punpcklbw xmm5, xmm0 ; mm5 = r1 p0..p3
michael@0	71	paddusw xmm3, xmm5 ; mm3 += mm6
michael@0	72
michael@0	73	; thresholding
michael@0	74	movdqa xmm7, xmm1 ; mm7 = r0 p0..p3
michael@0	75	psubusw xmm7, xmm5 ; mm7 = r0 p0..p3 - r1 p0..p3
michael@0	76	psubusw xmm5, xmm1 ; mm5 = r1 p0..p3 - r0 p0..p3
michael@0	77	paddusw xmm7, xmm5 ; mm7 = abs(r0 p0..p3 - r1 p0..p3)
michael@0	78	pcmpgtw xmm7, xmm2
michael@0	79
michael@0	80	movq xmm5, QWORD PTR [rsi + 2*rax] ; mm4 = r2 p0..p7
michael@0	81	punpcklbw xmm5, xmm0 ; mm5 = r2 p0..p3
michael@0	82	paddusw xmm3, xmm5 ; mm3 += mm5
michael@0	83
michael@0	84	; thresholding
michael@0	85	movdqa xmm6, xmm1 ; mm6 = r0 p0..p3
michael@0	86	psubusw xmm6, xmm5 ; mm6 = r0 p0..p3 - r2 p0..p3
michael@0	87	psubusw xmm5, xmm1 ; mm5 = r2 p0..p3 - r2 p0..p3
michael@0	88	paddusw xmm6, xmm5 ; mm6 = abs(r0 p0..p3 - r2 p0..p3)
michael@0	89	pcmpgtw xmm6, xmm2
michael@0	90	por xmm7, xmm6 ; accumulate thresholds
michael@0	91
michael@0	92
michael@0	93	neg rax
michael@0	94	movq xmm5, QWORD PTR [rsi+2*rax] ; mm4 = r-2 p0..p7
michael@0	95	punpcklbw xmm5, xmm0 ; mm5 = r-2 p0..p3
michael@0	96	paddusw xmm3, xmm5 ; mm3 += mm5
michael@0	97
michael@0	98	; thresholding
michael@0	99	movdqa xmm6, xmm1 ; mm6 = r0 p0..p3
michael@0	100	psubusw xmm6, xmm5 ; mm6 = p0..p3 - r-2 p0..p3
michael@0	101	psubusw xmm5, xmm1 ; mm5 = r-2 p0..p3 - p0..p3
michael@0	102	paddusw xmm6, xmm5 ; mm6 = abs(r0 p0..p3 - r-2 p0..p3)
michael@0	103	pcmpgtw xmm6, xmm2
michael@0	104	por xmm7, xmm6 ; accumulate thresholds
michael@0	105
michael@0	106	movq xmm4, QWORD PTR [rsi+rax] ; mm4 = r-1 p0..p7
michael@0	107	punpcklbw xmm4, xmm0 ; mm4 = r-1 p0..p3
michael@0	108	paddusw xmm3, xmm4 ; mm3 += mm5
michael@0	109
michael@0	110	; thresholding
michael@0	111	movdqa xmm6, xmm1 ; mm6 = r0 p0..p3
michael@0	112	psubusw xmm6, xmm4 ; mm6 = p0..p3 - r-2 p0..p3
michael@0	113	psubusw xmm4, xmm1 ; mm5 = r-1 p0..p3 - p0..p3
michael@0	114	paddusw xmm6, xmm4 ; mm6 = abs(r0 p0..p3 - r-1 p0..p3)
michael@0	115	pcmpgtw xmm6, xmm2
michael@0	116	por xmm7, xmm6 ; accumulate thresholds
michael@0	117
michael@0	118
michael@0	119	paddusw xmm3, RD42 ; mm3 += round value
michael@0	120	psraw xmm3, 3 ; mm3 /= 8
michael@0	121
michael@0	122	pand xmm1, xmm7 ; mm1 select vals > thresh from source
michael@0	123	pandn xmm7, xmm3 ; mm7 select vals < thresh from blurred result
michael@0	124	paddusw xmm1, xmm7 ; combination
michael@0	125
michael@0	126	packuswb xmm1, xmm0 ; pack to bytes
michael@0	127	movq QWORD PTR [rdi], xmm1 ;
michael@0	128
michael@0	129	neg rax ; pitch is positive
michael@0	130	add rsi, 8
michael@0	131	add rdi, 8
michael@0	132
michael@0	133	add rdx, 8
michael@0	134	cmp edx, dword arg(5) ;cols
michael@0	135
michael@0	136	jl .nextcol
michael@0	137
michael@0	138	; done with the all cols, start the across filtering in place
michael@0	139	sub rsi, rdx
michael@0	140	sub rdi, rdx
michael@0	141
michael@0	142	xor rdx, rdx
michael@0	143	movq mm0, QWORD PTR [rdi-8];
michael@0	144
michael@0	145	.acrossnextcol:
michael@0	146	movq xmm7, QWORD PTR [rdi +rdx -2]
michael@0	147	movd xmm4, DWORD PTR [rdi +rdx +6]
michael@0	148
michael@0	149	pslldq xmm4, 8
michael@0	150	por xmm4, xmm7
michael@0	151
michael@0	152	movdqa xmm3, xmm4
michael@0	153	psrldq xmm3, 2
michael@0	154	punpcklbw xmm3, xmm0 ; mm3 = p0..p3
michael@0	155	movdqa xmm1, xmm3 ; mm1 = p0..p3
michael@0	156	psllw xmm3, 2
michael@0	157
michael@0	158
michael@0	159	movdqa xmm5, xmm4
michael@0	160	psrldq xmm5, 3
michael@0	161	punpcklbw xmm5, xmm0 ; mm5 = p1..p4
michael@0	162	paddusw xmm3, xmm5 ; mm3 += mm6
michael@0	163
michael@0	164	; thresholding
michael@0	165	movdqa xmm7, xmm1 ; mm7 = p0..p3
michael@0	166	psubusw xmm7, xmm5 ; mm7 = p0..p3 - p1..p4
michael@0	167	psubusw xmm5, xmm1 ; mm5 = p1..p4 - p0..p3
michael@0	168	paddusw xmm7, xmm5 ; mm7 = abs(p0..p3 - p1..p4)
michael@0	169	pcmpgtw xmm7, xmm2
michael@0	170
michael@0	171	movdqa xmm5, xmm4
michael@0	172	psrldq xmm5, 4
michael@0	173	punpcklbw xmm5, xmm0 ; mm5 = p2..p5
michael@0	174	paddusw xmm3, xmm5 ; mm3 += mm5
michael@0	175
michael@0	176	; thresholding
michael@0	177	movdqa xmm6, xmm1 ; mm6 = p0..p3
michael@0	178	psubusw xmm6, xmm5 ; mm6 = p0..p3 - p1..p4
michael@0	179	psubusw xmm5, xmm1 ; mm5 = p1..p4 - p0..p3
michael@0	180	paddusw xmm6, xmm5 ; mm6 = abs(p0..p3 - p1..p4)
michael@0	181	pcmpgtw xmm6, xmm2
michael@0	182	por xmm7, xmm6 ; accumulate thresholds
michael@0	183
michael@0	184
michael@0	185	movdqa xmm5, xmm4 ; mm5 = p-2..p5
michael@0	186	punpcklbw xmm5, xmm0 ; mm5 = p-2..p1
michael@0	187	paddusw xmm3, xmm5 ; mm3 += mm5
michael@0	188
michael@0	189	; thresholding
michael@0	190	movdqa xmm6, xmm1 ; mm6 = p0..p3
michael@0	191	psubusw xmm6, xmm5 ; mm6 = p0..p3 - p1..p4
michael@0	192	psubusw xmm5, xmm1 ; mm5 = p1..p4 - p0..p3
michael@0	193	paddusw xmm6, xmm5 ; mm6 = abs(p0..p3 - p1..p4)
michael@0	194	pcmpgtw xmm6, xmm2
michael@0	195	por xmm7, xmm6 ; accumulate thresholds
michael@0	196
michael@0	197	psrldq xmm4, 1 ; mm4 = p-1..p5
michael@0	198	punpcklbw xmm4, xmm0 ; mm4 = p-1..p2
michael@0	199	paddusw xmm3, xmm4 ; mm3 += mm5
michael@0	200
michael@0	201	; thresholding
michael@0	202	movdqa xmm6, xmm1 ; mm6 = p0..p3
michael@0	203	psubusw xmm6, xmm4 ; mm6 = p0..p3 - p1..p4
michael@0	204	psubusw xmm4, xmm1 ; mm5 = p1..p4 - p0..p3
michael@0	205	paddusw xmm6, xmm4 ; mm6 = abs(p0..p3 - p1..p4)
michael@0	206	pcmpgtw xmm6, xmm2
michael@0	207	por xmm7, xmm6 ; accumulate thresholds
michael@0	208
michael@0	209	paddusw xmm3, RD42 ; mm3 += round value
michael@0	210	psraw xmm3, 3 ; mm3 /= 8
michael@0	211
michael@0	212	pand xmm1, xmm7 ; mm1 select vals > thresh from source
michael@0	213	pandn xmm7, xmm3 ; mm7 select vals < thresh from blurred result
michael@0	214	paddusw xmm1, xmm7 ; combination
michael@0	215
michael@0	216	packuswb xmm1, xmm0 ; pack to bytes
michael@0	217	movq QWORD PTR [rdi+rdx-8], mm0 ; store previous four bytes
michael@0	218	movdq2q mm0, xmm1
michael@0	219
michael@0	220	add rdx, 8
michael@0	221	cmp edx, dword arg(5) ;cols
michael@0	222	jl .acrossnextcol;
michael@0	223
michael@0	224	; last 8 pixels
michael@0	225	movq QWORD PTR [rdi+rdx-8], mm0
michael@0	226
michael@0	227	; done with this rwo
michael@0	228	add rsi,rax ; next line
michael@0	229	mov eax, dword arg(3) ;dst_pixels_per_line ; destination pitch?
michael@0	230	add rdi,rax ; next destination
michael@0	231	mov eax, dword arg(2) ;src_pixels_per_line ; destination pitch?
michael@0	232
michael@0	233	dec rcx ; decrement count
michael@0	234	jnz .nextrow ; next row
michael@0	235
michael@0	236	%if ABI_IS_32BIT=1 && CONFIG_PIC=1
michael@0	237	add rsp,16
michael@0	238	pop rsp
michael@0	239	%endif
michael@0	240	; begin epilog
michael@0	241	pop rdi
michael@0	242	pop rsi
michael@0	243	RESTORE_GOT
michael@0	244	RESTORE_XMM
michael@0	245	UNSHADOW_ARGS
michael@0	246	pop rbp
michael@0	247	ret
michael@0	248	%undef RD42
michael@0	249
michael@0	250
michael@0	251	;void vp9_mbpost_proc_down_xmm(unsigned char *dst,
michael@0	252	; int pitch, int rows, int cols,int flimit)
michael@0	253	extern sym(vp9_rv)
michael@0	254	global sym(vp9_mbpost_proc_down_xmm) PRIVATE
michael@0	255	sym(vp9_mbpost_proc_down_xmm):
michael@0	256	push rbp
michael@0	257	mov rbp, rsp
michael@0	258	SHADOW_ARGS_TO_STACK 5
michael@0	259	SAVE_XMM 7
michael@0	260	GET_GOT rbx
michael@0	261	push rsi
michael@0	262	push rdi
michael@0	263	; end prolog
michael@0	264
michael@0	265	ALIGN_STACK 16, rax
michael@0	266	sub rsp, 128+16
michael@0	267
michael@0	268	; unsigned char d[16][8] at [rsp]
michael@0	269	; create flimit2 at [rsp+128]
michael@0	270	mov eax, dword ptr arg(4) ;flimit
michael@0	271	mov [rsp+128], eax
michael@0	272	mov [rsp+128+4], eax
michael@0	273	mov [rsp+128+8], eax
michael@0	274	mov [rsp+128+12], eax
michael@0	275	%define flimit4 [rsp+128]
michael@0	276
michael@0	277	%if ABI_IS_32BIT=0
michael@0	278	lea r8, [GLOBAL(sym(vp9_rv))]
michael@0	279	%endif
michael@0	280
michael@0	281	;rows +=8;
michael@0	282	add dword arg(2), 8
michael@0	283
michael@0	284	;for(c=0; c<cols; c+=8)
michael@0	285	.loop_col:
michael@0	286	mov rsi, arg(0) ; s
michael@0	287	pxor xmm0, xmm0 ;
michael@0	288
michael@0	289	movsxd rax, dword ptr arg(1) ;pitch ;
michael@0	290	neg rax ; rax = -pitch
michael@0	291
michael@0	292	lea rsi, [rsi + rax*8]; ; rdi = s[-pitch*8]
michael@0	293	neg rax
michael@0	294
michael@0	295
michael@0	296	pxor xmm5, xmm5
michael@0	297	pxor xmm6, xmm6 ;
michael@0	298
michael@0	299	pxor xmm7, xmm7 ;
michael@0	300	mov rdi, rsi
michael@0	301
michael@0	302	mov rcx, 15 ;
michael@0	303
michael@0	304	.loop_initvar:
michael@0	305	movq xmm1, QWORD PTR [rdi];
michael@0	306	punpcklbw xmm1, xmm0 ;
michael@0	307
michael@0	308	paddw xmm5, xmm1 ;
michael@0	309	pmullw xmm1, xmm1 ;
michael@0	310
michael@0	311	movdqa xmm2, xmm1 ;
michael@0	312	punpcklwd xmm1, xmm0 ;
michael@0	313
michael@0	314	punpckhwd xmm2, xmm0 ;
michael@0	315	paddd xmm6, xmm1 ;
michael@0	316
michael@0	317	paddd xmm7, xmm2 ;
michael@0	318	lea rdi, [rdi+rax] ;
michael@0	319
michael@0	320	dec rcx
michael@0	321	jne .loop_initvar
michael@0	322	;save the var and sum
michael@0	323	xor rdx, rdx
michael@0	324	.loop_row:
michael@0	325	movq xmm1, QWORD PTR [rsi] ; [s-pitch*8]
michael@0	326	movq xmm2, QWORD PTR [rdi] ; [s+pitch*7]
michael@0	327
michael@0	328	punpcklbw xmm1, xmm0
michael@0	329	punpcklbw xmm2, xmm0
michael@0	330
michael@0	331	paddw xmm5, xmm2
michael@0	332	psubw xmm5, xmm1
michael@0	333
michael@0	334	pmullw xmm2, xmm2
michael@0	335	movdqa xmm4, xmm2
michael@0	336
michael@0	337	punpcklwd xmm2, xmm0
michael@0	338	punpckhwd xmm4, xmm0
michael@0	339
michael@0	340	paddd xmm6, xmm2
michael@0	341	paddd xmm7, xmm4
michael@0	342
michael@0	343	pmullw xmm1, xmm1
michael@0	344	movdqa xmm2, xmm1
michael@0	345
michael@0	346	punpcklwd xmm1, xmm0
michael@0	347	psubd xmm6, xmm1
michael@0	348
michael@0	349	punpckhwd xmm2, xmm0
michael@0	350	psubd xmm7, xmm2
michael@0	351
michael@0	352
michael@0	353	movdqa xmm3, xmm6
michael@0	354	pslld xmm3, 4
michael@0	355
michael@0	356	psubd xmm3, xmm6
michael@0	357	movdqa xmm1, xmm5
michael@0	358
michael@0	359	movdqa xmm4, xmm5
michael@0	360	pmullw xmm1, xmm1
michael@0	361
michael@0	362	pmulhw xmm4, xmm4
michael@0	363	movdqa xmm2, xmm1
michael@0	364
michael@0	365	punpcklwd xmm1, xmm4
michael@0	366	punpckhwd xmm2, xmm4
michael@0	367
michael@0	368	movdqa xmm4, xmm7
michael@0	369	pslld xmm4, 4
michael@0	370
michael@0	371	psubd xmm4, xmm7
michael@0	372
michael@0	373	psubd xmm3, xmm1
michael@0	374	psubd xmm4, xmm2
michael@0	375
michael@0	376	psubd xmm3, flimit4
michael@0	377	psubd xmm4, flimit4
michael@0	378
michael@0	379	psrad xmm3, 31
michael@0	380	psrad xmm4, 31
michael@0	381
michael@0	382	packssdw xmm3, xmm4
michael@0	383	packsswb xmm3, xmm0
michael@0	384
michael@0	385	movq xmm1, QWORD PTR [rsi+rax*8]
michael@0	386
michael@0	387	movq xmm2, xmm1
michael@0	388	punpcklbw xmm1, xmm0
michael@0	389
michael@0	390	paddw xmm1, xmm5
michael@0	391	mov rcx, rdx
michael@0	392
michael@0	393	and rcx, 127
michael@0	394	%if ABI_IS_32BIT=1 && CONFIG_PIC=1
michael@0	395	push rax
michael@0	396	lea rax, [GLOBAL(sym(vp9_rv))]
michael@0	397	movdqu xmm4, [rax + rcx*2] ;vp9_rv[rcx*2]
michael@0	398	pop rax
michael@0	399	%elif ABI_IS_32BIT=0
michael@0	400	movdqu xmm4, [r8 + rcx*2] ;vp9_rv[rcx*2]
michael@0	401	%else
michael@0	402	movdqu xmm4, [sym(vp9_rv) + rcx*2]
michael@0	403	%endif
michael@0	404
michael@0	405	paddw xmm1, xmm4
michael@0	406	;paddw xmm1, eight8s
michael@0	407	psraw xmm1, 4
michael@0	408
michael@0	409	packuswb xmm1, xmm0
michael@0	410	pand xmm1, xmm3
michael@0	411
michael@0	412	pandn xmm3, xmm2
michael@0	413	por xmm1, xmm3
michael@0	414
michael@0	415	and rcx, 15
michael@0	416	movq QWORD PTR [rsp + rcx*8], xmm1 ;d[rcx*8]
michael@0	417
michael@0	418	mov rcx, rdx
michael@0	419	sub rcx, 8
michael@0	420
michael@0	421	and rcx, 15
michael@0	422	movq mm0, [rsp + rcx*8] ;d[rcx*8]
michael@0	423
michael@0	424	movq [rsi], mm0
michael@0	425	lea rsi, [rsi+rax]
michael@0	426
michael@0	427	lea rdi, [rdi+rax]
michael@0	428	add rdx, 1
michael@0	429
michael@0	430	cmp edx, dword arg(2) ;rows
michael@0	431	jl .loop_row
michael@0	432
michael@0	433	add dword arg(0), 8 ; s += 8
michael@0	434	sub dword arg(3), 8 ; cols -= 8
michael@0	435	cmp dword arg(3), 0
michael@0	436	jg .loop_col
michael@0	437
michael@0	438	add rsp, 128+16
michael@0	439	pop rsp
michael@0	440
michael@0	441	; begin epilog
michael@0	442	pop rdi
michael@0	443	pop rsi
michael@0	444	RESTORE_GOT
michael@0	445	RESTORE_XMM
michael@0	446	UNSHADOW_ARGS
michael@0	447	pop rbp
michael@0	448	ret
michael@0	449	%undef flimit4
michael@0	450
michael@0	451
michael@0	452	;void vp9_mbpost_proc_across_ip_xmm(unsigned char *src,
michael@0	453	; int pitch, int rows, int cols,int flimit)
michael@0	454	global sym(vp9_mbpost_proc_across_ip_xmm) PRIVATE
michael@0	455	sym(vp9_mbpost_proc_across_ip_xmm):
michael@0	456	push rbp
michael@0	457	mov rbp, rsp
michael@0	458	SHADOW_ARGS_TO_STACK 5
michael@0	459	SAVE_XMM 7
michael@0	460	GET_GOT rbx
michael@0	461	push rsi
michael@0	462	push rdi
michael@0	463	; end prolog
michael@0	464
michael@0	465	ALIGN_STACK 16, rax
michael@0	466	sub rsp, 16
michael@0	467
michael@0	468	; create flimit4 at [rsp]
michael@0	469	mov eax, dword ptr arg(4) ;flimit
michael@0	470	mov [rsp], eax
michael@0	471	mov [rsp+4], eax
michael@0	472	mov [rsp+8], eax
michael@0	473	mov [rsp+12], eax
michael@0	474	%define flimit4 [rsp]
michael@0	475
michael@0	476
michael@0	477	;for(r=0;r<rows;r++)
michael@0	478	.ip_row_loop:
michael@0	479
michael@0	480	xor rdx, rdx ;sumsq=0;
michael@0	481	xor rcx, rcx ;sum=0;
michael@0	482	mov rsi, arg(0); s
michael@0	483	mov rdi, -8
michael@0	484	.ip_var_loop:
michael@0	485	;for(i=-8;i<=6;i++)
michael@0	486	;{
michael@0	487	; sumsq += s[i]*s[i];
michael@0	488	; sum += s[i];
michael@0	489	;}
michael@0	490	movzx eax, byte [rsi+rdi]
michael@0	491	add ecx, eax
michael@0	492	mul al
michael@0	493	add edx, eax
michael@0	494	add rdi, 1
michael@0	495	cmp rdi, 6
michael@0	496	jle .ip_var_loop
michael@0	497
michael@0	498
michael@0	499	;mov rax, sumsq
michael@0	500	;movd xmm7, rax
michael@0	501	movd xmm7, edx
michael@0	502
michael@0	503	;mov rax, sum
michael@0	504	;movd xmm6, rax
michael@0	505	movd xmm6, ecx
michael@0	506
michael@0	507	mov rsi, arg(0) ;s
michael@0	508	xor rcx, rcx
michael@0	509
michael@0	510	movsxd rdx, dword arg(3) ;cols
michael@0	511	add rdx, 8
michael@0	512	pxor mm0, mm0
michael@0	513	pxor mm1, mm1
michael@0	514
michael@0	515	pxor xmm0, xmm0
michael@0	516	.nextcol4:
michael@0	517
michael@0	518	movd xmm1, DWORD PTR [rsi+rcx-8] ; -8 -7 -6 -5
michael@0	519	movd xmm2, DWORD PTR [rsi+rcx+7] ; +7 +8 +9 +10
michael@0	520
michael@0	521	punpcklbw xmm1, xmm0 ; expanding
michael@0	522	punpcklbw xmm2, xmm0 ; expanding
michael@0	523
michael@0	524	punpcklwd xmm1, xmm0 ; expanding to dwords
michael@0	525	punpcklwd xmm2, xmm0 ; expanding to dwords
michael@0	526
michael@0	527	psubd xmm2, xmm1 ; 7--8 8--7 9--6 10--5
michael@0	528	paddd xmm1, xmm1 ; -8*2 -7*2 -6*2 -5*2
michael@0	529
michael@0	530	paddd xmm1, xmm2 ; 7+-8 8+-7 9+-6 10+-5
michael@0	531	pmaddwd xmm1, xmm2 ; squared of 7+-8 8+-7 9+-6 10+-5
michael@0	532
michael@0	533	paddd xmm6, xmm2
michael@0	534	paddd xmm7, xmm1
michael@0	535
michael@0	536	pshufd xmm6, xmm6, 0 ; duplicate the last ones
michael@0	537	pshufd xmm7, xmm7, 0 ; duplicate the last ones
michael@0	538
michael@0	539	psrldq xmm1, 4 ; 8--7 9--6 10--5 0000
michael@0	540	psrldq xmm2, 4 ; 8--7 9--6 10--5 0000
michael@0	541
michael@0	542	pshufd xmm3, xmm1, 3 ; 0000 8--7 8--7 8--7 squared
michael@0	543	pshufd xmm4, xmm2, 3 ; 0000 8--7 8--7 8--7 squared
michael@0	544
michael@0	545	paddd xmm6, xmm4
michael@0	546	paddd xmm7, xmm3
michael@0	547
michael@0	548	pshufd xmm3, xmm1, 01011111b ; 0000 0000 9--6 9--6 squared
michael@0	549	pshufd xmm4, xmm2, 01011111b ; 0000 0000 9--6 9--6 squared
michael@0	550
michael@0	551	paddd xmm7, xmm3
michael@0	552	paddd xmm6, xmm4
michael@0	553
michael@0	554	pshufd xmm3, xmm1, 10111111b ; 0000 0000 8--7 8--7 squared
michael@0	555	pshufd xmm4, xmm2, 10111111b ; 0000 0000 8--7 8--7 squared
michael@0	556
michael@0	557	paddd xmm7, xmm3
michael@0	558	paddd xmm6, xmm4
michael@0	559
michael@0	560	movdqa xmm3, xmm6
michael@0	561	pmaddwd xmm3, xmm3
michael@0	562
michael@0	563	movdqa xmm5, xmm7
michael@0	564	pslld xmm5, 4
michael@0	565
michael@0	566	psubd xmm5, xmm7
michael@0	567	psubd xmm5, xmm3
michael@0	568
michael@0	569	psubd xmm5, flimit4
michael@0	570	psrad xmm5, 31
michael@0	571
michael@0	572	packssdw xmm5, xmm0
michael@0	573	packsswb xmm5, xmm0
michael@0	574
michael@0	575	movd xmm1, DWORD PTR [rsi+rcx]
michael@0	576	movq xmm2, xmm1
michael@0	577
michael@0	578	punpcklbw xmm1, xmm0
michael@0	579	punpcklwd xmm1, xmm0
michael@0	580
michael@0	581	paddd xmm1, xmm6
michael@0	582	paddd xmm1, [GLOBAL(four8s)]
michael@0	583
michael@0	584	psrad xmm1, 4
michael@0	585	packssdw xmm1, xmm0
michael@0	586
michael@0	587	packuswb xmm1, xmm0
michael@0	588	pand xmm1, xmm5
michael@0	589
michael@0	590	pandn xmm5, xmm2
michael@0	591	por xmm5, xmm1
michael@0	592
michael@0	593	movd [rsi+rcx-8], mm0
michael@0	594	movq mm0, mm1
michael@0	595
michael@0	596	movdq2q mm1, xmm5
michael@0	597	psrldq xmm7, 12
michael@0	598
michael@0	599	psrldq xmm6, 12
michael@0	600	add rcx, 4
michael@0	601
michael@0	602	cmp rcx, rdx
michael@0	603	jl .nextcol4
michael@0	604
michael@0	605	;s+=pitch;
michael@0	606	movsxd rax, dword arg(1)
michael@0	607	add arg(0), rax
michael@0	608
michael@0	609	sub dword arg(2), 1 ;rows-=1
michael@0	610	cmp dword arg(2), 0
michael@0	611	jg .ip_row_loop
michael@0	612
michael@0	613	add rsp, 16
michael@0	614	pop rsp
michael@0	615
michael@0	616	; begin epilog
michael@0	617	pop rdi
michael@0	618	pop rsi
michael@0	619	RESTORE_GOT
michael@0	620	RESTORE_XMM
michael@0	621	UNSHADOW_ARGS
michael@0	622	pop rbp
michael@0	623	ret
michael@0	624	%undef flimit4
michael@0	625
michael@0	626
michael@0	627	;void vp9_plane_add_noise_wmt (unsigned char *start, unsigned char *noise,
michael@0	628	; unsigned char blackclamp[16],
michael@0	629	; unsigned char whiteclamp[16],
michael@0	630	; unsigned char bothclamp[16],
michael@0	631	; unsigned int width, unsigned int height, int pitch)
michael@0	632	extern sym(rand)
michael@0	633	global sym(vp9_plane_add_noise_wmt) PRIVATE
michael@0	634	sym(vp9_plane_add_noise_wmt):
michael@0	635	push rbp
michael@0	636	mov rbp, rsp
michael@0	637	SHADOW_ARGS_TO_STACK 8
michael@0	638	GET_GOT rbx
michael@0	639	push rsi
michael@0	640	push rdi
michael@0	641	; end prolog
michael@0	642
michael@0	643	.addnoise_loop:
michael@0	644	call sym(rand) WRT_PLT
michael@0	645	mov rcx, arg(1) ;noise
michael@0	646	and rax, 0xff
michael@0	647	add rcx, rax
michael@0	648
michael@0	649	; we rely on the fact that the clamping vectors are stored contiguously
michael@0	650	; in black/white/both order. Note that we have to reload this here because
michael@0	651	; rdx could be trashed by rand()
michael@0	652	mov rdx, arg(2) ; blackclamp
michael@0	653
michael@0	654
michael@0	655	mov rdi, rcx
michael@0	656	movsxd rcx, dword arg(5) ;[Width]
michael@0	657	mov rsi, arg(0) ;Pos
michael@0	658	xor rax,rax
michael@0	659
michael@0	660	.addnoise_nextset:
michael@0	661	movdqu xmm1,[rsi+rax] ; get the source
michael@0	662
michael@0	663	psubusb xmm1, [rdx] ;blackclamp ; clamp both sides so we don't outrange adding noise
michael@0	664	paddusb xmm1, [rdx+32] ;bothclamp
michael@0	665	psubusb xmm1, [rdx+16] ;whiteclamp
michael@0	666
michael@0	667	movdqu xmm2,[rdi+rax] ; get the noise for this line
michael@0	668	paddb xmm1,xmm2 ; add it in
michael@0	669	movdqu [rsi+rax],xmm1 ; store the result
michael@0	670
michael@0	671	add rax,16 ; move to the next line
michael@0	672
michael@0	673	cmp rax, rcx
michael@0	674	jl .addnoise_nextset
michael@0	675
michael@0	676	movsxd rax, dword arg(7) ; Pitch
michael@0	677	add arg(0), rax ; Start += Pitch
michael@0	678	sub dword arg(6), 1 ; Height -= 1
michael@0	679	jg .addnoise_loop
michael@0	680
michael@0	681	; begin epilog
michael@0	682	pop rdi
michael@0	683	pop rsi
michael@0	684	RESTORE_GOT
michael@0	685	UNSHADOW_ARGS
michael@0	686	pop rbp
michael@0	687	ret
michael@0	688
michael@0	689
michael@0	690	SECTION_RODATA
michael@0	691	align 16
michael@0	692	rd42:
michael@0	693	times 8 dw 0x04
michael@0	694	four8s:
michael@0	695	times 4 dd 8