codenotifier - Theora: revision 15057

Project: Theora

Revision: 15057

Author: xiphmont

Date: 22 Jun 2008 17:07:32

Changes:

Eliminate use of SSE instruction in loop filter MMX code by replacing
current loop filter with one from Derf's new decoder. This required
some amount of code refactoring as Derf's code expects slightly
different input.

Files:
modified: /trunk/theora/lib/enc/pp.c (try)
modified: /trunk/theora/lib/enc/x86_32/dct_decode_mmx.c (try)
modified: /trunk/theora/lib/enc/dsp.h (try)
modified: /trunk/theora/lib/enc/x86_64/dct_decode_mmx.c (try)
modified: /trunk/theora/lib/enc/dct_decode.c (try)
modified: /trunk/theora/lib/enc/codec_internal.h (try)

Diff:

......	@@ -22,156 +22,381 @@
22	22	#if defined(USE_ASM)
23	23
24		static const __attribute__((aligned(8),used)) ogg_int64_t V3= 0x0003000300030003LL;
25		static const __attribute__((aligned(8),used)) ogg_int64_t V804= 0x0804080408040804LL;
	24	static const __attribute__((aligned(8),used)) ogg_int64_t OC_V3=
	25	0x0003000300030003LL;
	26	static const __attribute__((aligned(8),used)) ogg_int64_t OC_V4=
	27	0x0004000400040004LL;
26	28
27		#define OC_LOOP_H_4x4 \
28		"lea (%[ll],%[ll],2),%[s]\n" /* esi = ystride3 / \
29		"movd (%[pp]), %%mm0\n" /* 0 0 0 0 3 2 1 0 */ \
30		"movd (%[pp],%[ll]),%%mm1\n" /* 0 0 0 0 7 6 5 4 */ \
31		"movd (%[pp],%[ll],2),%%mm2\n" /* 0 0 0 0 b a 9 8 */ \
32		"movd (%[pp],%[s]),%%mm3\n" /* 0 0 0 0 f e d c */ \
33		"punpcklbw %%mm1,%%mm0\n" /* mm0 = 7 3 6 2 5 1 4 0 */ \
34		"punpcklbw %%mm3,%%mm2\n" /* mm2 = f b e a d 9 c 8 */ \
35		"movq %%mm0,%%mm1\n" /* mm1 = 7 3 6 2 5 1 4 0 */ \
36		"punpcklwd %%mm2,%%mm1\n" /* mm1 = d 9 5 1 c 8 4 0 */ \
37		"punpckhwd %%mm2,%%mm0\n" /* mm0 = f b 7 3 e a 6 2 */ \
38		"pxor %%mm7,%%mm7\n" \
39		"movq %%mm1,%%mm5\n" /* mm5 = d 9 5 1 c 8 4 0 */ \
40		"punpckhbw %%mm7,%%mm5\n" /* mm5 = 0 d 0 9 0 5 0 1 = pix[1]*/ \
41		"punpcklbw %%mm7,%%mm1\n" /* mm1 = 0 c 0 8 0 4 0 0 = pix[0]*/ \
42		"movq %%mm0,%%mm3\n" /* mm3 = f b 7 3 e a 6 2 */ \
43		"punpckhbw %%mm7,%%mm3\n" /* mm3 = 0 f 0 b 0 7 0 3 = pix[3]*/ \
44		"punpcklbw %%mm7,%%mm0\n" /* mm0 = 0 e 0 a 0 6 0 2 = pix[2]*/ \
45		\
46		"psubw %%mm3,%%mm1\n" /* mm1 = pix[0]-pix[3] mm1 - mm3 */ \
47		"movq %%mm0,%%mm7\n" /* mm7 = pix[2]*/ \
48		"psubw %%mm5,%%mm0\n" /* mm0 = pix[2]-pix[1] mm0 - mm5*/ \
49		"pmullw %[V3],%%mm0\n" /* 3 / \
50		"paddw %%mm0,%%mm1\n" /* mm1 has f[0] ... f[4]*/ \
51		"paddw %[V804],%%mm1\n" /* add 4 / / add 256 after shift */ \
52		"psraw $3,%%mm1\n" /* >>3 */ \
53		" pextrw $0,%%mm1,%[s]\n" /* In MM1 we have 4 f coefs (16bits) */ \
54		" pextrw $1,%%mm1,%[d]\n" /* now perform MM4 = (_bv+ f) / \
55		" pinsrw $0,(%[bound],%[s],2),%%mm4\n" \
56		" pextrw $2,%%mm1,%[s]\n" \
57		" pinsrw $1,(%[bound],%[d],2),%%mm4\n" \
58		" pextrw $3,%%mm1,%[d]\n" \
59		" pinsrw $2,(%[bound],%[s],2),%%mm4\n" \
60		" pinsrw $3,(%[bound],%[d],2),%%mm4\n" /* new f vals loaded */ \
61		"pxor %%mm0,%%mm0\n" \
62		" paddw %%mm4,%%mm5\n" /(pix[1]+f);/ \
63		" psubw %%mm4,%%mm7\n" /* (pix[2]-f); */ \
64		" packuswb %%mm0,%%mm5\n" /* mm5 = x x x x newpix1 */ \
65		" packuswb %%mm0,%%mm7\n" /* mm7 = x x x x newpix2 */ \
66		" punpcklbw %%mm7,%%mm5\n" /* 2 1 2 1 2 1 2 1 */ \
67		" movd %%mm5,%[d]\n" /* edi = newpix21 */ \
68		" movw %[d],1(%[pp])\n" \
69		" psrlq $32,%%mm5\n" /* why is so big stall here ? */ \
70		" shr $16,%[d]\n" \
71		" movw %[d],1(%[pp],%[ll],1)\n" \
72		" movd %%mm5,%[d]\n" /* eax = newpix21 high part */ \
73		" lea (%[ll],%[ll],2),%[s]\n" \
74		" movw %[d],1(%[pp],%[ll],2)\n" \
75		" shr $16,%[d]\n" \
76		" movw %[d],1(%[pp],%[s])\n"
77
78		static void FilterHoriz__mmx(unsigned char * PixelPtr,
79		ogg_int32_t LineLength,
80		ogg_int16_t *BoundingValuePtr){
	29	static void loop_filter_v(unsigned char *_pix,int _ystride,
	30	const ogg_int16_t *_ll){
81	31	long esi;
82		long edi;
	32	_pix-=_ystride*2;
83	33	__asm__ __volatile__(
84		OC_LOOP_H_4x4
85		: [s]"=&r"(esi),[d]"=&r"(edi)
86		: [pp]"r"(PixelPtr), [ll]"r"((long)LineLength), [bound]"r"(BoundingValuePtr-256), [V3]"m"(V3), [V804]"m"(V804)
87		: "memory"
	34	/mm0=0/
	35	"pxor %%mm0,%%mm0\n\t"
	36	/esi=_ystride3*/
	37	"lea (%[ystride],%[ystride],2),%[s]\n\t"
	38	/mm7=_pix[0...8]/
	39	"movq (%[pix]),%%mm7\n\t"
	40	/mm4=_pix[0...8+_ystride3]*/
	41	"movq (%[pix],%[s]),%%mm4\n\t"
	42	/mm6=_pix[0...8]/
	43	"movq %%mm7,%%mm6\n\t"
	44	/Expand unsigned _pix[0...3] to 16 bits./
	45	"punpcklbw %%mm0,%%mm6\n\t"
	46	"movq %%mm4,%%mm5\n\t"
	47	/Expand unsigned _pix[4...8] to 16 bits./
	48	"punpckhbw %%mm0,%%mm7\n\t"
	49	/Expand other arrays too./
	50	"punpcklbw %%mm0,%%mm4\n\t"
	51	"punpckhbw %%mm0,%%mm5\n\t"
	52	/mm7:mm6=_p[0...8]-_p[0...8+_ystride3]:*/
	53	"psubw %%mm4,%%mm6\n\t"
	54	"psubw %%mm5,%%mm7\n\t"
	55	/mm5=mm4=_pix[0...8+_ystride]/
	56	"movq (%[pix],%[ystride]),%%mm4\n\t"
	57	/mm1=mm3=mm2=_pix[0..8]+_ystride2]*/
	58	"movq (%[pix],%[ystride],2),%%mm2\n\t"
	59	"movq %%mm4,%%mm5\n\t"
	60	"movq %%mm2,%%mm3\n\t"
	61	"movq %%mm2,%%mm1\n\t"
	62	/Expand these arrays./
	63	"punpckhbw %%mm0,%%mm5\n\t"
	64	"punpcklbw %%mm0,%%mm4\n\t"
	65	"punpckhbw %%mm0,%%mm3\n\t"
	66	"punpcklbw %%mm0,%%mm2\n\t"
	67	/Preload.../
	68	"movq %[OC_V3],%%mm0\n\t"
	69	/mm3:mm2=_pix[0...8+_ystride2]-_pix[0...8+_ystride]*/
	70	"psubw %%mm5,%%mm3\n\t"
	71	"psubw %%mm4,%%mm2\n\t"
	72	/Scale by 3./
	73	"pmullw %%mm0,%%mm3\n\t"
	74	"pmullw %%mm0,%%mm2\n\t"
	75	/Preload.../
	76	"movq %[OC_V4],%%mm0\n\t"
	77	/f=mm3:mm2==_pix[0...8]-_pix[0...8+_ystride3]+
	78	3(_pix[0...8+_ystride2]-_pix[0...8+_ystride])*/
	79	"paddw %%mm7,%%mm3\n\t"
	80	"paddw %%mm6,%%mm2\n\t"
	81	/Add 4./
	82	"paddw %%mm0,%%mm3\n\t"
	83	"paddw %%mm0,%%mm2\n\t"
	84	/"Divide" by 8./
	85	"psraw $3,%%mm3\n\t"
	86	"psraw $3,%%mm2\n\t"
	87	/Now compute lflim of mm3:mm2 cf. Section 7.10 of the sepc./
	88	/Free up mm5./
	89	"packuswb %%mm5,%%mm4\n\t"
	90	/mm0=L L L L/
	91	"movq (%[ll]),%%mm0\n\t"
	92	/if(R_i<-2L\|\|R_i>2L)R_i=0:/
	93	"movq %%mm2,%%mm5\n\t"
	94	"pxor %%mm6,%%mm6\n\t"
	95	"movq %%mm0,%%mm7\n\t"
	96	"psubw %%mm0,%%mm6\n\t"
	97	"psllw $1,%%mm7\n\t"
	98	"psllw $1,%%mm6\n\t"
	99	/mm2==R_3 R_2 R_1 R_0/
	100	/mm5==R_3 R_2 R_1 R_0/
	101	/mm6==-2L -2L -2L -2L/
	102	/mm7==2L 2L 2L 2L/
	103	"pcmpgtw %%mm2,%%mm7\n\t"
	104	"pcmpgtw %%mm6,%%mm5\n\t"
	105	"pand %%mm7,%%mm2\n\t"
	106	"movq %%mm0,%%mm7\n\t"
	107	"pand %%mm5,%%mm2\n\t"
	108	"psllw $1,%%mm7\n\t"
	109	"movq %%mm3,%%mm5\n\t"
	110	/mm3==R_7 R_6 R_5 R_4/
	111	/mm5==R_7 R_6 R_5 R_4/
	112	/mm6==-2L -2L -2L -2L/
	113	/mm7==2L 2L 2L 2L/
	114	"pcmpgtw %%mm3,%%mm7\n\t"
	115	"pcmpgtw %%mm6,%%mm5\n\t"
	116	"pand %%mm7,%%mm3\n\t"
	117	"movq %%mm0,%%mm7\n\t"
	118	"pand %%mm5,%%mm3\n\t"
	119	/*if(R_i<-L)R_i'=R_i+2L;
	120	if(R_i>L)R_i'=R_i-2L;
	121	if(R_i<-L\|\|R_i>L)R_i=-R_i':*/
	122	"psraw $1,%%mm6\n\t"
	123	"movq %%mm2,%%mm5\n\t"
	124	"psllw $1,%%mm7\n\t"
	125	/mm2==R_3 R_2 R_1 R_0/
	126	/mm5==R_3 R_2 R_1 R_0/
	127	/mm6==-L -L -L -L/
	128	/mm0==L L L L/
	129	/mm5=R_i>L?FF:00/
	130	"pcmpgtw %%mm0,%%mm5\n\t"
	131	/mm6=-L>R_i?FF:00/
	132	"pcmpgtw %%mm2,%%mm6\n\t"
	133	/mm7=R_i>L?2L:0/
	134	"pand %%mm5,%%mm7\n\t"
	135	/mm2=R_i>L?R_i-2L:R_i/
	136	"psubw %%mm7,%%mm2\n\t"
	137	"movq %%mm0,%%mm7\n\t"
	138	/mm5=-L>R_i\|\|R_i>L/
	139	"por %%mm6,%%mm5\n\t"
	140	"psllw $1,%%mm7\n\t"
	141	/mm7=-L>R_i?2L:0/
	142	"pand %%mm6,%%mm7\n\t"
	143	"pxor %%mm6,%%mm6\n\t"
	144	/mm2=-L>R_i?R_i+2L:R_i/
	145	"paddw %%mm7,%%mm2\n\t"
	146	"psubw %%mm0,%%mm6\n\t"
	147	/mm5=-L>R_i\|\|R_i>L?-R_i':0/
	148	"pand %%mm2,%%mm5\n\t"
	149	"movq %%mm0,%%mm7\n\t"
	150	/mm2=-L>R_i\|\|R_i>L?0:R_i/
	151	"psubw %%mm5,%%mm2\n\t"
	152	"psllw $1,%%mm7\n\t"
	153	/mm2=-L>R_i\|\|R_i>L?-R_i':R_i/
	154	"psubw %%mm5,%%mm2\n\t"
	155	"movq %%mm3,%%mm5\n\t"
	156	/mm3==R_7 R_6 R_5 R_4/
	157	/mm5==R_7 R_6 R_5 R_4/
	158	/mm6==-L -L -L -L/
	159	/mm0==L L L L/
	160	/mm6=-L>R_i?FF:00/
	161	"pcmpgtw %%mm3,%%mm6\n\t"
	162	/mm5=R_i>L?FF:00/
	163	"pcmpgtw %%mm0,%%mm5\n\t"
	164	/mm7=R_i>L?2L:0/
	165	"pand %%mm5,%%mm7\n\t"
	166	/mm2=R_i>L?R_i-2L:R_i/
	167	"psubw %%mm7,%%mm3\n\t"
	168	"psllw $1,%%mm0\n\t"
	169	/mm5=-L>R_i\|\|R_i>L/
	170	"por %%mm6,%%mm5\n\t"
	171	/mm0=-L>R_i?2L:0/
	172	"pand %%mm6,%%mm0\n\t"
	173	/mm3=-L>R_i?R_i+2L:R_i/
	174	"paddw %%mm0,%%mm3\n\t"
	175	/mm5=-L>R_i\|\|R_i>L?-R_i':0/
	176	"pand %%mm3,%%mm5\n\t"
	177	/mm2=-L>R_i\|\|R_i>L?0:R_i/
	178	"psubw %%mm5,%%mm3\n\t"
	179	/mm2=-L>R_i\|\|R_i>L?-R_i':R_i/
	180	"psubw %%mm5,%%mm3\n\t"
	181	/*Unfortunately, there's no unsigned byte+signed byte with unsigned
	182	saturation op code, so we have to promote things back 16 bits.*/
	183	"pxor %%mm0,%%mm0\n\t"
	184	"movq %%mm4,%%mm5\n\t"
	185	"punpcklbw %%mm0,%%mm4\n\t"
	186	"punpckhbw %%mm0,%%mm5\n\t"
	187	"movq %%mm1,%%mm6\n\t"
	188	"punpcklbw %%mm0,%%mm1\n\t"
	189	"punpckhbw %%mm0,%%mm6\n\t"
	190	/_pix[0...8+_ystride]+=R_i/
	191	"paddw %%mm2,%%mm4\n\t"
	192	"paddw %%mm3,%%mm5\n\t"
	193	/_pix[0...8+_ystride2]-=R_i*/
	194	"psubw %%mm2,%%mm1\n\t"
	195	"psubw %%mm3,%%mm6\n\t"
	196	"packuswb %%mm5,%%mm4\n\t"
	197	"packuswb %%mm6,%%mm1\n\t"
	198	/Write it back out./
	199	"movq %%mm4,(%[pix],%[ystride])\n\t"
	200	"movq %%mm1,(%[pix],%[ystride],2)\n\t"
	201	:[s]"=&S"(esi)
	202	:[pix]"r"(_pix),[ystride]"r"((long)_ystride),[ll]"r"(_ll),
	203	[OC_V3]"m"(OC_V3),[OC_V4]"m"(OC_V4)
	204	:"memory"
88	205	);
	206	}
89	207
90		PixelPtr += LineLength*4;
91
	208	/*This code implements the bulk of loop_filter_h().
	209	Data are striped p0 p1 p2 p3 ... p0 p1 p2 p3 ..., so in order to load all
	210	four p0's to one register we must transpose the values in four mmx regs.
	211	When half is done we repeat this for the rest.*/
	212	static void loop_filter_h4(unsigned char *_pix,long _ystride,
	213	const ogg_int16_t *_ll){
	214	long esi;
	215	long edi;
92	216	__asm__ __volatile__(
93		OC_LOOP_H_4x4
94		"emms\n"
95		: [s]"=&r"(esi),[d]"=&r"(edi) \
96		: [pp]"r"(PixelPtr), [ll]"r"((long)LineLength), [bound]"r"(BoundingValuePtr-256), [V3]"m"(V3), [V804]"m"(V804) \
97		: "memory" \
98		);
	217	/x x x x 3 2 1 0/
	218	"movd (%[pix]),%%mm0\n\t"
	219	/esi=_ystride3*/
	220	"lea (%[ystride],%[ystride],2),%[s]\n\t"
	221	/x x x x 7 6 5 4/
	222	"movd (%[pix],%[ystride]),%%mm1\n\t"
	223	/x x x x B A 9 8/
	224	"movd (%[pix],%[ystride],2),%%mm2\n\t"
	225	/x x x x F E D C/
	226	"movd (%[pix],%[s]),%%mm3\n\t"
	227	/mm0=7 3 6 2 5 1 4 0/
	228	"punpcklbw %%mm1,%%mm0\n\t"
	229	/mm2=F B E A D 9 C 8/
	230	"punpcklbw %%mm3,%%mm2\n\t"
	231	/mm1=7 3 6 2 5 1 4 0/
	232	"movq %%mm0,%%mm1\n\t"
	233	/mm0=F B 7 3 E A 6 2/
	234	"punpckhwd %%mm2,%%mm0\n\t"
	235	/mm1=D 9 5 1 C 8 4 0/
	236	"punpcklwd %%mm2,%%mm1\n\t"
	237	"pxor %%mm7,%%mm7\n\t"
	238	/mm5=D 9 5 1 C 8 4 0/
	239	"movq %%mm1,%%mm5\n\t"
	240	/mm1=x C x 8 x 4 x 0==pix[0]/
	241	"punpcklbw %%mm7,%%mm1\n\t"
	242	/mm5=x D x 9 x 5 x 1==pix[1]/
	243	"punpckhbw %%mm7,%%mm5\n\t"
	244	/mm3=F B 7 3 E A 6 2/
	245	"movq %%mm0,%%mm3\n\t"
	246	/mm0=x E x A x 6 x 2==pix[2]/
	247	"punpcklbw %%mm7,%%mm0\n\t"
	248	/mm3=x F x B x 7 x 3==pix[3]/
	249	"punpckhbw %%mm7,%%mm3\n\t"
	250	/mm1=mm1-mm3==pix[0]-pix[3]/
	251	"psubw %%mm3,%%mm1\n\t"
	252	/Save a copy of pix[2] for later./
	253	"movq %%mm0,%%mm4\n\t"
	254	/mm0=mm0-mm5==pix[2]-pix[1]/
	255	"psubw %%mm5,%%mm0\n\t"
	256	/Scale by 3./
	257	"pmullw %[OC_V3],%%mm0\n\t"
	258	/f=mm1==_pix[0]-_pix[3]+ 3(_pix[2]-_pix[1])*/
	259	"paddw %%mm1,%%mm0\n\t"
	260	/Add 4./
	261	"paddw %[OC_V4],%%mm0\n\t"
	262	/"Divide" by 8, producing the residuals R_i./
	263	"psraw $3,%%mm0\n\t"
	264	/Now compute lflim of mm0 cf. Section 7.10 of the sepc./
	265	/mm6=L L L L/
	266	"movq (%[ll]),%%mm6\n\t"
	267	/if(R_i<-2L\|\|R_i>2L)R_i=0:/
	268	"movq %%mm0,%%mm1\n\t"
	269	"pxor %%mm2,%%mm2\n\t"
	270	"movq %%mm6,%%mm3\n\t"
	271	"psubw %%mm6,%%mm2\n\t"
	272	"psllw $1,%%mm3\n\t"
	273	"psllw $1,%%mm2\n\t"
	274	/mm0==R_3 R_2 R_1 R_0/
	275	/mm1==R_3 R_2 R_1 R_0/
	276	/mm2==-2L -2L -2L -2L/
	277	/mm3==2L 2L 2L 2L/
	278	"pcmpgtw %%mm0,%%mm3\n\t"
	279	"pcmpgtw %%mm2,%%mm1\n\t"
	280	"pand %%mm3,%%mm0\n\t"
	281	"pand %%mm1,%%mm0\n\t"
	282	/*if(R_i<-L)R_i'=R_i+2L;
	283	if(R_i>L)R_i'=R_i-2L;
	284	if(R_i<-L\|\|R_i>L)R_i=-R_i':*/
	285	"psraw $1,%%mm2\n\t"
	286	"movq %%mm0,%%mm1\n\t"
	287	"movq %%mm6,%%mm3\n\t"
	288	/mm0==R_3 R_2 R_1 R_0/
	289	/mm1==R_3 R_2 R_1 R_0/
	290	/mm2==-L -L -L -L/
	291	/mm6==L L L L/
	292	/mm2=-L>R_i?FF:00/
	293	"pcmpgtw %%mm0,%%mm2\n\t"
	294	/mm1=R_i>L?FF:00/
	295	"pcmpgtw %%mm6,%%mm1\n\t"
	296	/mm3=2L 2L 2L 2L/
	297	"psllw $1,%%mm3\n\t"
	298	/mm6=2L 2L 2L 2L/
	299	"psllw $1,%%mm6\n\t"
	300	/mm3=R_i>L?2L:0/
	301	"pand %%mm1,%%mm3\n\t"
	302	/mm6=-L>R_i?2L:0/
	303	"pand %%mm2,%%mm6\n\t"
	304	/mm0=R_i>L?R_i-2L:R_i/
	305	"psubw %%mm3,%%mm0\n\t"
	306	/mm1=-L>R_i\|\|R_i>L/
	307	"por %%mm2,%%mm1\n\t"
	308	/mm0=-L>R_i?R_i+2L:R_i/
	309	"paddw %%mm6,%%mm0\n\t"
	310	/mm1=-L>R_i\|\|R_i>L?R_i':0/
	311	"pand %%mm0,%%mm1\n\t"
	312	/mm0=-L>R_i\|\|R_i>L?0:R_i/
	313	"psubw %%mm1,%%mm0\n\t"
	314	/mm0=-L>R_i\|\|R_i>L?-R_i':R_i/
	315	"psubw %%mm1,%%mm0\n\t"
	316	/_pix[1]+=R_i;/
	317	"paddw %%mm0,%%mm5\n\t"
	318	/_pix[2]-=R_i;/
	319	"psubw %%mm0,%%mm4\n\t"
	320	/mm5=x x x x D 9 5 1/
	321	"packuswb %%mm7,%%mm5\n\t"
	322	/mm4=x x x x E A 6 2/
	323	"packuswb %%mm7,%%mm4\n\t"
	324	/mm5=E D A 9 6 5 2 1/
	325	"punpcklbw %%mm4,%%mm5\n\t"
	326	/edi=6 5 2 1/
	327	"movd %%mm5,%%edi\n\t"
	328	"movw %%di,1(%[pix])\n\t"
	329	/Why is there such a big stall here?/
	330	"psrlq $32,%%mm5\n\t"
	331	"shrl $16,%%edi\n\t"
	332	"movw %%di,1(%[pix],%[ystride])\n\t"
	333	/edi=E D A 9/
	334	"movd %%mm5,%%edi\n\t"
	335	"movw %%di,1(%[pix],%[ystride],2)\n\t"
	336	"shrl $16,%%edi\n\t"
	337	"movw %%di,1(%[pix],%[s])\n\t"
	338	:[s]"=&S"(esi),[d]"=&D"(edi),
	339	[pix]"+r"(_pix),[ystride]"+r"(_ystride),[ll]"+r"(_ll)
	340	:[OC_V3]"m"(OC_V3),[OC_V4]"m"(OC_V4)
	341	:"memory"
	342	);
99	343	}
100	344
101		static void FilterVert__mmx(unsigned char * PixelPtr,
102		ogg_int32_t LineLength,
103		ogg_int16_t *BoundingValuePtr){
104		long esi,edi;
105		__asm__ __volatile__(
106		"pxor %%mm0,%%mm0\n" /* mm0 = 0 */
107		"movq (%[pp]),%%mm7\n" /* mm7 = pix[0..7] */
108		"lea (%[ll],%[ll],2),%[s]\n" /* esi = ystride3 /
109		"movq (%[pp],%[s]),%%mm4\n" /* mm4 = pix[0..7+ystride3] /
110		"movq %%mm7,%%mm6\n" /* mm6 = pix[0..7] */
111		"punpcklbw %%mm0,%%mm6\n" /* expand unsigned pix[0..3] to 16 bits */
112		"movq %%mm4,%%mm5\n"
113		"punpckhbw %%mm0,%%mm7\n" /* expand unsigned pix[4..7] to 16 bits */
114		"punpcklbw %%mm0,%%mm4\n" /* expand other arrays too */
115		"punpckhbw %%mm0,%%mm5\n"
116		"psubw %%mm4,%%mm6\n" /* mm6 = mm6 - mm4 */
117		"psubw %%mm5,%%mm7\n" /* mm7 = mm7 - mm5 */
118		/* mm7:mm6 = _p[0]-_p[ystride3] /
119		"movq (%[pp],%[ll]),%%mm4\n" /* mm4 = pix[0..7+ystride] */
120		"movq %%mm4,%%mm5\n"
121		"movq (%[pp],%[ll],2),%%mm2\n" /* mm2 = pix[0..7+ystride2] /
122		"movq %%mm2,%%mm3\n"
123		"movq %%mm2,%%mm1\n" //ystride*2
124		"punpckhbw %%mm0,%%mm5\n"
125		"punpcklbw %%mm0,%%mm4\n"
126		"punpckhbw %%mm0,%%mm3\n"
127		"punpcklbw %%mm0,%%mm2\n"
128		"psubw %%mm5,%%mm3\n"
129		"psubw %%mm4,%%mm2\n"
130		/* mm3:mm2 = (pix[ystride2]-pix[ystride]); /
131		"pmullw %[V3],%%mm3\n" /* 3 /
132		"pmullw %[V3],%%mm2\n" /* 3 /
133		"paddw %%mm7,%%mm3\n" /* highpart */
134		"paddw %%mm6,%%mm2\n" /* lowpart of pix[0]-pix[ystride3]+3(pix[ystride2]-pix[ystride]); /
135		"paddw %[V804],%%mm3\n" /* add 4 / / add 256 after shift */
136		"paddw %[V804],%%mm2\n" /* add 4 / / add 256 after shift */
137		"psraw $3,%%mm3\n" /* >>3 f coefs high */
138		"psraw $3,%%mm2\n" /* >>3 f coefs low */
139
140		" pextrw $0,%%mm2,%[s]\n" /* In MM3:MM2 we have f coefs (16bits) */
141		" pextrw $1,%%mm2,%[d]\n" /* now perform MM7:MM6 = (_bv+ f) /
142		" pinsrw $0,(%[bound],%[s],2),%%mm6\n"
143		" pinsrw $1,(%[bound],%[d],2),%%mm6\n"
144
145		" pextrw $2,%%mm2,%[s]\n"
146		" pextrw $3,%%mm2,%[d]\n"
147		" pinsrw $2,(%[bound],%[s],2),%%mm6\n"
148		" pinsrw $3,(%[bound],%[d],2),%%mm6\n"
149
150		" pextrw $0,%%mm3,%[s]\n"
151		" pextrw $1,%%mm3,%[d]\n"
152		" pinsrw $0,(%[bound],%[s],2),%%mm7\n"
153		" pinsrw $1,(%[bound],%[d],2),%%mm7\n"
154
155		" pextrw $2,%%mm3,%[s]\n"
156		" pextrw $3,%%mm3,%[d]\n"
157		" pinsrw $2,(%[bound],%[s],2),%%mm7\n"
158		" pinsrw $3,(%[bound],%[d],2),%%mm7\n" //MM7 MM6 f=*(_bv+(f+4>>3));
159
160		"paddw %%mm6,%%mm4\n" /* (pix[ystride]+f); */
161		"paddw %%mm7,%%mm5\n" /* (pix[ystride]+f); */
162		"movq %%mm1,%%mm2\n"
163		"punpcklbw %%mm0,%%mm1\n"
164		"punpckhbw %%mm0,%%mm2\n" //[ystride*2]
165		"psubw %%mm6,%%mm1\n" /* (pix[ystride2]-f); /
166		"psubw %%mm7,%%mm2\n" /* (pix[ystride2]-f); /
167		"packuswb %%mm2,%%mm1\n"
168		"packuswb %%mm5,%%mm4\n"
169		"movq %%mm1,(%[pp],%[ll],2)\n" /* pix[ystride2]= /
170		"movq %%mm4,(%[pp],%[ll])\n" /* pix[ystride]= */
171		"emms\n"
172		: [s]"=&r"(esi),[d]"=&r"(edi)
173		: [pp]"r"(PixelPtr-2*LineLength), [ll]"r"((long)LineLength), [bound]"r"(BoundingValuePtr-256), [V3]"m"(V3), [V804]"m"(V804)
174		: "memory"
175		);
	345	static void loop_filter_h(unsigned char *_pix,int _ystride,
	346	const ogg_int16_t *_ll){
	347	_pix-=2;
	348	loop_filter_h4(_pix,_ystride,_ll);
	349	loop_filter_h4(_pix+(_ystride<<2),_ystride,_ll);
	350	}
	351
	352	static void loop_filter_mmx(PB_INSTANCE *pbi, int FLimit){
	353	int j;
	354	ogg_int16_t __attribute__((aligned(8))) ll[4];
	355	unsigned char *cp = pbi->display_fragments;
	356	ogg_uint32_t *bp = pbi->recon_pixel_index_table;
	357
	358	if ( FLimit == 0 ) return;
	359	ll[0]=ll[1]=ll[2]=ll[3]=FLimit;
	360
	361	for ( j = 0; j < 3 ; j++){
	362	ogg_uint32_t *bp_begin = bp;
	363	ogg_uint32_t *bp_end;
	364	int stride;
	365	int h;
	366
	367	switch(j) {
	368	case 0: /* y */
	369	bp_end = bp + pbi->YPlaneFragments;
	370	h = pbi->HFragments;
	371	stride = pbi->YStride;
	372	break;
	373	default: /* u,v, 4:20 specific */
	374	bp_end = bp + pbi->UVPlaneFragments;
	375	h = pbi->HFragments >> 1;
	376	stride = pbi->UVStride;
	377	break;
	378	}
	379
	380	while(bp<bp_end){
	381	ogg_uint32_t *bp_left = bp;
	382	ogg_uint32_t *bp_right = bp + h;
	383	while(bp<bp_right){
	384	if(cp[0]){
	385	if(bp>bp_left)
	386	loop_filter_h(&pbi->LastFrameRecon[bp[0]],stride,ll);
	387	if(bp_left>bp_begin)
	388	loop_filter_v(&pbi->LastFrameRecon[bp[0]],stride,ll);
	389	if(bp+1<bp_right && !cp[1])
	390	loop_filter_h(&pbi->LastFrameRecon[bp[0]]+8,stride,ll);
	391	if(bp+stride<bp_end && !cp[stride])
	392	loop_filter_v(&pbi->LastFrameRecon[bp[h]]+8,stride,ll);
	393	}
	394	bp++;
	395	cp++;
	396	}
	397	}
	398	}
	399
	400	__asm__ __volatile__("emms\n\t");
176	401	}
177	402
......	@@ -179,6 +404,5 @@
179	404	void dsp_mmx_dct_decode_init(DspFunctions *funcs)
180	405	{
181		funcs->FilterVert = FilterVert__mmx;
182		funcs->FilterHoriz = FilterHoriz__mmx;
	406	funcs->LoopFilter = loop_filter_mmx;
183	407	}
184	408