Attachment #192646 for bug #271476






            "movl     "_pDest", %%eax\n\t"
            "addl     "_olddx", %%eax\n\t"

#ifdef USE_VERTICAL_FILTER
            "movq    %%mm0, %%mm1\n\t"
            //      pavgb   mm0, qword ptr[edx]
            V_PAVGB ("%%mm0", "(%%edx)", "%%mm2", _ShiftMask)
            //      movntq  qword ptr[eax+_olddx], mm0
            V_MOVNTQ ("(%%eax)", "%%mm0")
            //      pavgb   mm1, qword ptr[edx+ecx]
            V_PAVGB ("%%mm1", "(%%edx, %%ecx)", "%%mm2", _ShiftMask)
            "addl   "_dst_pitchw", %%eax"
            //      movntq  qword ptr[eax+_olddx], mm1
            V_MOVNTQ ("(%%eax)", "%%mm1")
#else

            //      movntq  qword ptr[eax+_olddx], mm0
                V_MOVNTQ ("(%%eax)", "%%mm0")
#endif

           "addl    $8, "_olddx"\n\t"          // bump offset pointer
           "movl    "_olddx", %%eax\n\t"
           "cmpl    "_Last8", %%eax\n\t"       // done with line?
           "jb      1b\n\t"                    // y
#endif

            "popl %%ebx\n\t"

        : /* no outputs */
        : "m"(pBob),
          "m"(src_pitch2),

          "m"(ONES[0]),
          "m"(UVMask[0]),
          "m"(Max_Mov[0]),
          "m"(YMask[0]),
          "m"(olddx)

        : "eax", "ecx", "edx", "esi", "edi",
          "st", "st(1)", "st(2)", "st(3)", "st(4)", "st(5)", "st(6)", "st(7)",

#undef _UVMask
#undef _Max_Mov
#undef _YMask
#undef _olddx

        // adjust for next line
        pSrc  += src_pitch2;




// -*- c++ -*-

unsigned char*	pDest;
long		olddx;
const unsigned char*	pSrcP;
const unsigned char*	pSrc;
const unsigned char*	pBob;

#define _UVMask     "%15"
#define _Max_Mov    "%16"
#define _YMask      "%17"
#define _olddx      "%18"

	for (y=1; y < FldHeight-1; y++)
	{

             // edi - prev weave pixels, 1 line up
             // esi - next weave pixels, 1 line up

             // Save ebx (-fPIC)
             "pushl %%ebx\n\t"

#ifdef IS_SSE2

             // sse2 code deleted for now

#else
             // simple bob first 8 bytes
             "movl		"_pBob",        %%edx\n\t"
             "movl	"_src_pitch2",  %%ecx\n\t"

#ifdef USE_VERTICAL_FILTER
             "movq	    (%%edx),        %%mm0\n\t"
             "movq	    (%%edx, %%ecx), %%mm1\n\t" //, qword ptr[edx+ecx]
             "movq	    %%mm0,          %%mm2\n\t"
             V_PAVGB ("%%mm2", "%%mm1", "%%mm3", _ShiftMask)		// halfway between
             V_PAVGB ("%%mm0", "%%mm2", "%%mm3", _ShiftMask)		// 1/4 way

             V_MOVNTQ	("(%%edi, %%eax)", "%%mm1") // qword ptr[edi+eax], mm1

             // simple bob last 8 bytes
             "movl		"_Last8", %%esi\n\t"
             "movl		%%esi, "_olddx"\n\t"
             "addl		%%edx, %%esi\n\t" // [edx+_olddx]
             "movq	    (%%esi), %%mm0\n\t"
             "movq	    (%%esi, %%ecx), %%mm1\n\t"    // qword ptr[esi+ecx]
             "movq	    %%mm0, %%mm2\n\t"
             V_PAVGB ("%%mm2", "%%mm1", "%%mm3", _ShiftMask)		// halfway between
             V_PAVGB ("%%mm0", "%%mm2", "%%mm3", _ShiftMask)		// 1/4 way
             V_PAVGB ("%%mm1", "%%mm2", "%%mm3", _ShiftMask)		// 3/4 way
             "addl		"_olddx", %%edi\n\t"						// last 8 bytes of dest
             V_MOVNTQ	("%%edi", "%%mm0")
             V_MOVNTQ	("(%%edi, %%eax)", "%%mm1") // qword ptr[edi+eax], mm1)

#else
             "movq	(%%edx), %%mm0\n\t"
             //		pavgb	mm0, qword ptr[edx+ecx]
             V_PAVGB ("%%mm0", "(%%edx, %%ecx)", "%%mm2", _ShiftMask) // qword ptr[edx+ecx], mm2, ShiftMask)
             "movl		"_pDest", %%edi\n\t"
             V_MOVNTQ	("(%%edi)", "%%mm0")

             // simple bob last 8 bytes
             "movl		"_Last8", %%esi\n\t"
             "movl		%%esi, "_olddx"\n\t"
             "addl		%%edx, %%esi\n\t" //esi, [edx+_olddx]
             "movq	    (%%esi), %%mm0\n\t"
             //		pavgb	mm0, qword ptr[esi+ecx]
             V_PAVGB	("%%mm0", "(%%esi, %%ecx)", "%%mm2", _ShiftMask) // qword ptr[esi+ecx], mm2, ShiftMask)
             "addl		"_olddx", %%edi\n\t"
             V_MOVNTQ	("(%%edi)", "%%mm0") // qword ptr[edi+_olddx], mm0)
#endif
             // now loop and get the middle qwords
             "movl		"_pSrc", %%esi\n\t"
             "movl		"_pSrcP", %%edi\n\t"
             "movl		$8, "_olddx"\n\t"				// curr offset into all lines

             "1:\n\t"
             "movl		"_pBobP", %%eax\n\t"
             "addl		$8, %%edi\n\t"
             "addl		$8, %%esi\n\t"
             "addl		$8, %%edx\n\t"
             "addl		"_olddx", %%eax\n\t"

#ifdef USE_STRANGE_BOB
#include "StrangeBob.inc"




        "pxor %%mm6, %%mm6\n\t"
        "pxor %%mm7, %%mm7\n\t"

		"movq    -2(%%edx), %%mm0\n\t"		// value a from top left
		"movq    -4(%%edx, %%ecx), %%mm1\n\t"	// value m from bottom right

		"movq	%%mm0, %%mm3\n\t"
		"psubusb	%%mm1, %%mm3\n\t"

		"pcmpeqb	%%mm3, %%mm4\n\t"			// here ff where abs(a,m) > Thres, else 00


		"movq    -4(%%edx), %%mm0\n\t"		// value j
		"movq    4(%%edx, %%ecx), %%mm1\n\t"	// value n
		"movq	%%mm0, %%mm2\n\t"
		"pavgb	%%mm1, %%mm2\n\t"					// avg(j,n)
        "movq	%%mm0, %%mm3\n\t"

        "por     %%mm0, %%mm7\n\t"

        // k & m
		"movq    2(%%edx), %%mm0\n\t"		// value c from top left
		"movq    4(%%edx, %%ecx), %%mm1\n\t"	// value n from bottom right

		"movq	%%mm0, %%mm3\n\t"
		"psubusb	%%mm1, %%mm3\n\t"

		"pcmpeqb	%%mm3, %%mm4\n\t"			// here ff where abs(c,n) > Thres, else 00


		"movq    4(%%edx), %%mm0\n\t"		// value k
		"movq    -4(%%edx, %%ecx), %%mm1\n\t"	// value m
		"movq	%%mm0, %%mm2\n\t"
		V_PAVGB ("%%mm2", "%%mm1", "%%mm3", _ShiftMask)	// avg(k,m)
        "movq	%%mm0, %%mm3\n\t"



        // c & d
		"movq    (%%edx), %%mm0\n\t"		// value b from top left
		"movq    2(%%edx, %%ecx), %%mm1\n\t"	// value f from bottom right

		"movq	%%mm0, %%mm3\n\t"
		"psubusb	%%mm1, %%mm3\n\t"

		"pcmpeqb %%mm4, %%mm3\n\t"			// now ff where abs(b,f) < Thres, else 00
		"pcmpeqb	%%mm3, %%mm4\n\t"			// here ff where abs(b,f) > Thres, else 00

		"movq    2(%%edx), %%mm0\n\t"		// value c
		"movq    -2(%%edx, %%ecx), %%mm1\n\t"	// value d
		"movq	%%mm0, %%mm2\n\t"
		V_PAVGB ("%%mm2", "%%mm1", "%%mm3", _ShiftMask)	// avg(c,d)
        "movq	%%mm0, %%mm3\n\t"

        "por     %%mm0, %%mm7\n\t"

        // a & f
		"movq    (%%edx), %%mm0\n\t"		// value b from top left
		"movq    -2(%%edx, %%ecx), %%mm1\n\t"	// value d from bottom right

		"movq	%%mm0, %%mm3\n\t"
		"psubusb	%%mm1, %%mm3\n\t"

		"pcmpeqb %%mm4, %%mm3\n\t"			// now ff where abs(b,d) < Thres, else 00
		"pcmpeqb	%%mm3, %%mm4\n\t"			// here ff where abs(b,d) > Thres, else 00

		"movq    -2(%%edx), %%mm0\n\t"		// value a
		"movq    2(%%edx, %%ecx), %%mm1\n\t"	// value f
		"movq	%%mm0, %%mm2\n\t"
		V_PAVGB ("%%mm2", "%%mm1", "%%mm3", _ShiftMask)	// avg(a,f)
        "movq	%%mm0, %%mm3\n\t"

 		"pand	"_YMask", %%mm7\n\t"			// mask out chroma from here

		// b,e
		"movq    (%%edx), %%mm0\n\t"		// value b from top
		"movq    (%%edx, %%ecx), %%mm1\n\t"	// value e from bottom
		"movq	%%mm0, %%mm2\n\t"
		V_PAVGB ("%%mm2", "%%mm1", "%%mm3", _ShiftMask)	// avg(b,e)
        "movq	%%mm0, %%mm3\n\t"

        "por     %%mm0, %%mm7\n\t"

		// bob in any leftovers
		"movq    (%%edx), %%mm0\n\t"		// value b from top
		"movq    (%%edx, %%ecx), %%mm1\n\t"	// value e from bottom


// We will also calc here the max/min values to later limit comb




		// selected for the	smallest of abs(a,f), abs(c,d), or abs(b,e), etc.

		// a,f
		"movq    -2(%%edx), %%mm0\n\t"		// value a from top left
		"movq    2(%%edx, %%ecx), %%mm1\n\t"	// value f from bottom right
		"movq	%%mm0, %%mm6\n\t"
//		pavgb	%%mm6, %%mm1					// avg(a,f), also best so far
		V_PAVGB ("%%mm6", "%%mm1", "%%mm7", _ShiftMask)	// avg(a,f), also best so far

		"por		%%mm1, %%mm7\n\t"					// abs diff, also best so far

		// c,d
		"movq    2(%%edx), %%mm0\n\t"		// value a from top left
		"movq    -2(%%edx, %%ecx), %%mm1\n\t"	// value f from bottom right
		"movq	%%mm0, %%mm2\n\t"
//		pavgb	%%mm2, %%mm1					// avg(c,d)
		V_PAVGB ("%%mm2", "%%mm1", "%%mm3", _ShiftMask)	// avg(c,d)

		"pand	"_YMask", %%mm5\n\t"			// mask out chroma from here also

		// j,n
		"movq    -4(%%edx), %%mm0\n\t"		// value j from top left
		"movq    4(%%edx, %%ecx), %%mm1\n\t"	// value n from bottom right
		"movq	%%mm0, %%mm2\n\t"
//		pavgb	%%mm2, %%mm1					// avg(j,n)
		V_PAVGB ("%%mm2", "%%mm1", "%%mm3", _ShiftMask)	// avg(j,n)

		"por		%%mm1, %%mm7\n\t"			// "

		// k, m
		"movq    4(%%edx), %%mm0\n\t"		// value k from top right
		"movq    -4(%%edx, %%ecx), %%mm1\n\t"	// value n from bottom left
		"movq	%%mm0, %%mm4\n\t"
//		pavgb	%%mm4, %%mm1					// avg(k,m)
		V_PAVGB ("%%mm4", "%%mm1", "%%mm3", _ShiftMask)	// avg(k,m)

		"por		%%mm1, %%mm7\n\t"			// "

		// b,e
		"movq    (%%edx), %%mm0\n\t"		// value b from top
		"movq    (%%edx, %%ecx), %%mm1\n\t"	// value e from bottom

// We will also calc here the max/min values to later limit comb
// so the max excursion will not exceed the Max_Comb constant

Return to bug 271476

Lines 50-80 Link Here

(-)filter/tomsmocomp-old/SearchLoopBottom.inc (-15 / +17 lines)
50		50
51		51
52	"movl "_pDest", %%eax\n\t"	52	"movl "_pDest", %%eax\n\t"
		53	"addl "_olddx", %%eax\n\t"
53		54
54	#ifdef USE_VERTICAL_FILTER	55	#ifdef USE_VERTICAL_FILTER
55	"movq %%mm0, %%mm1\n\t"	56	"movq %%mm0, %%mm1\n\t"
56	// pavgb mm0, qword ptr[ebx]	57	// pavgb mm0, qword ptr[edx]
57	V_PAVGB ("%%mm0", "(%%ebx)", "%%mm2", _ShiftMask)	58	V_PAVGB ("%%mm0", "(%%edx)", "%%mm2", _ShiftMask)
58	// movntq qword ptr[eax+edx], mm0	59	// movntq qword ptr[eax+_olddx], mm0
59	V_MOVNTQ ("(%%eax, %%edx)", "%%mm0")	60	V_MOVNTQ ("(%%eax)", "%%mm0")
60	// pavgb mm1, qword ptr[ebx+ecx]	61	// pavgb mm1, qword ptr[edx+ecx]
61	V_PAVGB ("%%mm1", "(%%ebx, %%ecx)", "%%mm2", _ShiftMask)	62	V_PAVGB ("%%mm1", "(%%edx, %%ecx)", "%%mm2", _ShiftMask)
62	"addl "_dst_pitchw", %%eax"	63	"addl "_dst_pitchw", %%eax"
63	// movntq qword ptr[eax+edx], mm1	64	// movntq qword ptr[eax+_olddx], mm1
64	V_MOVNTQ ("(%%eax, %%edx)", "%%mm1")	65	V_MOVNTQ ("(%%eax)", "%%mm1")
65	#else	66	#else
66		67
67	// movntq qword ptr[eax+edx], mm0	68	// movntq qword ptr[eax+_olddx], mm0
68	V_MOVNTQ ("(%%eax, %%edx)", "%%mm0")	69	V_MOVNTQ ("(%%eax)", "%%mm0")
69	#endif	70	#endif
70		71
71	"leal 8(%%edx), %%edx\n\t" // bump offset pointer	72	"addl $8, "_olddx"\n\t" // bump offset pointer
72	"cmpl "_Last8", %%edx\n\t" // done with line?	73	"movl "_olddx", %%eax\n\t"
		74	"cmpl "_Last8", %%eax\n\t" // done with line?
73	"jb 1b\n\t" // y	75	"jb 1b\n\t" // y
74	#endif	76	#endif
75		77
76	"popl %%ebx\n\t"
77
78	: /* no outputs */	78	: /* no outputs */
79	: "m"(pBob),	79	: "m"(pBob),
80	"m"(src_pitch2),	80	"m"(src_pitch2),
Lines 93-99 Link Here
93	"m"(ONES[0]),	93	"m"(ONES[0]),
94	"m"(UVMask[0]),	94	"m"(UVMask[0]),
95	"m"(Max_Mov[0]),	95	"m"(Max_Mov[0]),
96	"m"(YMask[0])	96	"m"(YMask[0]),
		97	"m"(olddx)
97		98
98	: "eax", "ecx", "edx", "esi", "edi",	99	: "eax", "ecx", "edx", "esi", "edi",
99	"st", "st(1)", "st(2)", "st(3)", "st(4)", "st(5)", "st(6)", "st(7)",	100	"st", "st(1)", "st(2)", "st(3)", "st(4)", "st(5)", "st(6)", "st(7)",
Lines 119-124 Link Here
119	#undef _UVMask	120	#undef _UVMask
120	#undef _Max_Mov	121	#undef _Max_Mov
121	#undef _YMask	122	#undef _YMask
		123	#undef _olddx
122		124
123	// adjust for next line	125	// adjust for next line
124	pSrc += src_pitch2;	126	pSrc += src_pitch2;

Lines 1-6 Link Here

(-)filter/tomsmocomp-old/SearchLoopTop.inc (-18 / +20 lines)
1	// -- c++ --	1	// -- c++ --
2		2
3	unsigned char* pDest;	3	unsigned char* pDest;
		4	long olddx;
4	const unsigned char* pSrcP;	5	const unsigned char* pSrcP;
5	const unsigned char* pSrc;	6	const unsigned char* pSrc;
6	const unsigned char* pBob;	7	const unsigned char* pBob;
Lines 84-89 Link Here
84	#define _UVMask "%15"	85	#define _UVMask "%15"
85	#define _Max_Mov "%16"	86	#define _Max_Mov "%16"
86	#define _YMask "%17"	87	#define _YMask "%17"
		88	#define _olddx "%18"
87		89
88	for (y=1; y < FldHeight-1; y++)	90	for (y=1; y < FldHeight-1; y++)
89	{	91	{
Lines 99-119 Link Here
99	// edi - prev weave pixels, 1 line up	101	// edi - prev weave pixels, 1 line up
100	// esi - next weave pixels, 1 line up	102	// esi - next weave pixels, 1 line up
101		103
102	// Save ebx (-fPIC)
103	"pushl %%ebx\n\t"
104
105	#ifdef IS_SSE2	104	#ifdef IS_SSE2
106		105
107	// sse2 code deleted for now	106	// sse2 code deleted for now
108		107
109	#else	108	#else
110	// simple bob first 8 bytes	109	// simple bob first 8 bytes
111	"movl "_pBob", %%ebx\n\t"	110	"movl "_pBob", %%edx\n\t"
112	"movl "_src_pitch2", %%ecx\n\t"	111	"movl "_src_pitch2", %%ecx\n\t"
113		112
114	#ifdef USE_VERTICAL_FILTER	113	#ifdef USE_VERTICAL_FILTER
115	"movq (%%ebx), %%mm0\n\t"	114	"movq (%%edx), %%mm0\n\t"
116	"movq (%%ebx, %%ecx), %%mm1\n\t" //, qword ptr[ebx+ecx]	115	"movq (%%edx, %%ecx), %%mm1\n\t" //, qword ptr[edx+ecx]
117	"movq %%mm0, %%mm2\n\t"	116	"movq %%mm0, %%mm2\n\t"
118	V_PAVGB ("%%mm2", "%%mm1", "%%mm3", _ShiftMask) // halfway between	117	V_PAVGB ("%%mm2", "%%mm1", "%%mm3", _ShiftMask) // halfway between
119	V_PAVGB ("%%mm0", "%%mm2", "%%mm3", _ShiftMask) // 1/4 way	118	V_PAVGB ("%%mm0", "%%mm2", "%%mm3", _ShiftMask) // 1/4 way
Lines 124-167 Link Here
124	V_MOVNTQ ("(%%edi, %%eax)", "%%mm1") // qword ptr[edi+eax], mm1	123	V_MOVNTQ ("(%%edi, %%eax)", "%%mm1") // qword ptr[edi+eax], mm1
125		124
126	// simple bob last 8 bytes	125	// simple bob last 8 bytes
127	"movl "_Last8", %%edx\n\t"	126	"movl "_Last8", %%esi\n\t"
128	"leal (%%ebx, %%edx), %%esi\n\t" // [ebx+edx]	127	"movl %%esi, "_olddx"\n\t"
		128	"addl %%edx, %%esi\n\t" // [edx+_olddx]
129	"movq (%%esi), %%mm0\n\t"	129	"movq (%%esi), %%mm0\n\t"
130	"movq (%%esi, %%ecx), %%mm1\n\t" // qword ptr[esi+ecx]	130	"movq (%%esi, %%ecx), %%mm1\n\t" // qword ptr[esi+ecx]
131	"movq %%mm0, %%mm2\n\t"	131	"movq %%mm0, %%mm2\n\t"
132	V_PAVGB ("%%mm2", "%%mm1", "%%mm3", _ShiftMask) // halfway between	132	V_PAVGB ("%%mm2", "%%mm1", "%%mm3", _ShiftMask) // halfway between
133	V_PAVGB ("%%mm0", "%%mm2", "%%mm3", _ShiftMask) // 1/4 way	133	V_PAVGB ("%%mm0", "%%mm2", "%%mm3", _ShiftMask) // 1/4 way
134	V_PAVGB ("%%mm1", "%%mm2", "%%mm3", _ShiftMask) // 3/4 way	134	V_PAVGB ("%%mm1", "%%mm2", "%%mm3", _ShiftMask) // 3/4 way
135	"addl %%edx, %%edi\n\t" // last 8 bytes of dest	135	"addl "_olddx", %%edi\n\t" // last 8 bytes of dest
136	V_MOVNTQ ("%%edi", "%%mm0")	136	V_MOVNTQ ("%%edi", "%%mm0")
137	V_MOVNTQ ("(%%edi, %%eax)", "%%mm1") // qword ptr[edi+eax], mm1)	137	V_MOVNTQ ("(%%edi, %%eax)", "%%mm1") // qword ptr[edi+eax], mm1)
138		138
139	#else	139	#else
140	"movq (%%ebx), %%mm0\n\t"	140	"movq (%%edx), %%mm0\n\t"
141	// pavgb mm0, qword ptr[ebx+ecx]	141	// pavgb mm0, qword ptr[edx+ecx]
142	V_PAVGB ("%%mm0", "(%%ebx, %%ecx)", "%%mm2", _ShiftMask) // qword ptr[ebx+ecx], mm2, ShiftMask)	142	V_PAVGB ("%%mm0", "(%%edx, %%ecx)", "%%mm2", _ShiftMask) // qword ptr[edx+ecx], mm2, ShiftMask)
143	"movl "_pDest", %%edi\n\t"	143	"movl "_pDest", %%edi\n\t"
144	V_MOVNTQ ("(%%edi)", "%%mm0")	144	V_MOVNTQ ("(%%edi)", "%%mm0")
145		145
146	// simple bob last 8 bytes	146	// simple bob last 8 bytes
147	"movl "_Last8", %%edx\n\t"	147	"movl "_Last8", %%esi\n\t"
148	"leal (%%ebx, %%edx), %%esi\n\t" //esi, [ebx+edx]	148	"movl %%esi, "_olddx"\n\t"
		149	"addl %%edx, %%esi\n\t" //esi, [edx+_olddx]
149	"movq (%%esi), %%mm0\n\t"	150	"movq (%%esi), %%mm0\n\t"
150	// pavgb mm0, qword ptr[esi+ecx]	151	// pavgb mm0, qword ptr[esi+ecx]
151	V_PAVGB ("%%mm0", "(%%esi, %%ecx)", "%%mm2", _ShiftMask) // qword ptr[esi+ecx], mm2, ShiftMask)	152	V_PAVGB ("%%mm0", "(%%esi, %%ecx)", "%%mm2", _ShiftMask) // qword ptr[esi+ecx], mm2, ShiftMask)
152	V_MOVNTQ ("(%%edi, %%edx)", "%%mm0") // qword ptr[edi+edx], mm0)	153	"addl "_olddx", %%edi\n\t"
		154	V_MOVNTQ ("(%%edi)", "%%mm0") // qword ptr[edi+_olddx], mm0)
153	#endif	155	#endif
154	// now loop and get the middle qwords	156	// now loop and get the middle qwords
155	"movl "_pSrc", %%esi\n\t"	157	"movl "_pSrc", %%esi\n\t"
156	"movl "_pSrcP", %%edi\n\t"	158	"movl "_pSrcP", %%edi\n\t"
157	"movl $8, %%edx\n\t" // curr offset into all lines	159	"movl $8, "_olddx"\n\t" // curr offset into all lines
158		160
159	"1:\n\t"	161	"1:\n\t"
160	"movl "_pBobP", %%eax\n\t"	162	"movl "_pBobP", %%eax\n\t"
161	"addl $8, %%edi\n\t"	163	"addl $8, %%edi\n\t"
162	"addl $8, %%esi\n\t"	164	"addl $8, %%esi\n\t"
163	"addl $8, %%ebx\n\t"	165	"addl $8, %%edx\n\t"
164	"addl %%edx, %%eax\n\t"	166	"addl "_olddx", %%eax\n\t"
165		167
166	#ifdef USE_STRANGE_BOB	168	#ifdef USE_STRANGE_BOB
167	#include "StrangeBob.inc"	169	#include "StrangeBob.inc"

Lines 31-38 Link Here

(-)filter/tomsmocomp-old/StrangeBob.inc (-20 / +20 lines)
31	"pxor %%mm6, %%mm6\n\t"	31	"pxor %%mm6, %%mm6\n\t"
32	"pxor %%mm7, %%mm7\n\t"	32	"pxor %%mm7, %%mm7\n\t"
33		33
34	"movq -2(%%ebx), %%mm0\n\t" // value a from top left	34	"movq -2(%%edx), %%mm0\n\t" // value a from top left
35	"movq -4(%%ebx, %%ecx), %%mm1\n\t" // value m from bottom right	35	"movq -4(%%edx, %%ecx), %%mm1\n\t" // value m from bottom right
36		36
37	"movq %%mm0, %%mm3\n\t"	37	"movq %%mm0, %%mm3\n\t"
38	"psubusb %%mm1, %%mm3\n\t"	38	"psubusb %%mm1, %%mm3\n\t"
Lines 45-52 Link Here
45	"pcmpeqb %%mm3, %%mm4\n\t" // here ff where abs(a,m) > Thres, else 00	45	"pcmpeqb %%mm3, %%mm4\n\t" // here ff where abs(a,m) > Thres, else 00
46		46
47		47
48	"movq -4(%%ebx), %%mm0\n\t" // value j	48	"movq -4(%%edx), %%mm0\n\t" // value j
49	"movq 4(%%ebx, %%ecx), %%mm1\n\t" // value n	49	"movq 4(%%edx, %%ecx), %%mm1\n\t" // value n
50	"movq %%mm0, %%mm2\n\t"	50	"movq %%mm0, %%mm2\n\t"
51	"pavgb %%mm1, %%mm2\n\t" // avg(j,n)	51	"pavgb %%mm1, %%mm2\n\t" // avg(j,n)
52	"movq %%mm0, %%mm3\n\t"	52	"movq %%mm0, %%mm3\n\t"
Lines 75-82 Link Here
75	"por %%mm0, %%mm7\n\t"	75	"por %%mm0, %%mm7\n\t"
76		76
77	// k & m	77	// k & m
78	"movq 2(%%ebx), %%mm0\n\t" // value c from top left	78	"movq 2(%%edx), %%mm0\n\t" // value c from top left
79	"movq 4(%%ebx, %%ecx), %%mm1\n\t" // value n from bottom right	79	"movq 4(%%edx, %%ecx), %%mm1\n\t" // value n from bottom right
80		80
81	"movq %%mm0, %%mm3\n\t"	81	"movq %%mm0, %%mm3\n\t"
82	"psubusb %%mm1, %%mm3\n\t"	82	"psubusb %%mm1, %%mm3\n\t"
Lines 89-96 Link Here
89	"pcmpeqb %%mm3, %%mm4\n\t" // here ff where abs(c,n) > Thres, else 00	89	"pcmpeqb %%mm3, %%mm4\n\t" // here ff where abs(c,n) > Thres, else 00
90		90
91		91
92	"movq 4(%%ebx), %%mm0\n\t" // value k	92	"movq 4(%%edx), %%mm0\n\t" // value k
93	"movq -4(%%ebx, %%ecx), %%mm1\n\t" // value m	93	"movq -4(%%edx, %%ecx), %%mm1\n\t" // value m
94	"movq %%mm0, %%mm2\n\t"	94	"movq %%mm0, %%mm2\n\t"
95	V_PAVGB ("%%mm2", "%%mm1", "%%mm3", _ShiftMask) // avg(k,m)	95	V_PAVGB ("%%mm2", "%%mm1", "%%mm3", _ShiftMask) // avg(k,m)
96	"movq %%mm0, %%mm3\n\t"	96	"movq %%mm0, %%mm3\n\t"
Lines 120-127 Link Here
120		120
121		121
122	// c & d	122	// c & d
123	"movq (%%ebx), %%mm0\n\t" // value b from top left	123	"movq (%%edx), %%mm0\n\t" // value b from top left
124	"movq 2(%%ebx, %%ecx), %%mm1\n\t" // value f from bottom right	124	"movq 2(%%edx, %%ecx), %%mm1\n\t" // value f from bottom right
125		125
126	"movq %%mm0, %%mm3\n\t"	126	"movq %%mm0, %%mm3\n\t"
127	"psubusb %%mm1, %%mm3\n\t"	127	"psubusb %%mm1, %%mm3\n\t"
Lines 133-140 Link Here
133	"pcmpeqb %%mm4, %%mm3\n\t" // now ff where abs(b,f) < Thres, else 00	133	"pcmpeqb %%mm4, %%mm3\n\t" // now ff where abs(b,f) < Thres, else 00
134	"pcmpeqb %%mm3, %%mm4\n\t" // here ff where abs(b,f) > Thres, else 00	134	"pcmpeqb %%mm3, %%mm4\n\t" // here ff where abs(b,f) > Thres, else 00
135		135
136	"movq 2(%%ebx), %%mm0\n\t" // value c	136	"movq 2(%%edx), %%mm0\n\t" // value c
137	"movq -2(%%ebx, %%ecx), %%mm1\n\t" // value d	137	"movq -2(%%edx, %%ecx), %%mm1\n\t" // value d
138	"movq %%mm0, %%mm2\n\t"	138	"movq %%mm0, %%mm2\n\t"
139	V_PAVGB ("%%mm2", "%%mm1", "%%mm3", _ShiftMask) // avg(c,d)	139	V_PAVGB ("%%mm2", "%%mm1", "%%mm3", _ShiftMask) // avg(c,d)
140	"movq %%mm0, %%mm3\n\t"	140	"movq %%mm0, %%mm3\n\t"
Lines 163-170 Link Here
163	"por %%mm0, %%mm7\n\t"	163	"por %%mm0, %%mm7\n\t"
164		164
165	// a & f	165	// a & f
166	"movq (%%ebx), %%mm0\n\t" // value b from top left	166	"movq (%%edx), %%mm0\n\t" // value b from top left
167	"movq -2(%%ebx, %%ecx), %%mm1\n\t" // value d from bottom right	167	"movq -2(%%edx, %%ecx), %%mm1\n\t" // value d from bottom right
168		168
169	"movq %%mm0, %%mm3\n\t"	169	"movq %%mm0, %%mm3\n\t"
170	"psubusb %%mm1, %%mm3\n\t"	170	"psubusb %%mm1, %%mm3\n\t"
Lines 176-183 Link Here
176	"pcmpeqb %%mm4, %%mm3\n\t" // now ff where abs(b,d) < Thres, else 00	176	"pcmpeqb %%mm4, %%mm3\n\t" // now ff where abs(b,d) < Thres, else 00
177	"pcmpeqb %%mm3, %%mm4\n\t" // here ff where abs(b,d) > Thres, else 00	177	"pcmpeqb %%mm3, %%mm4\n\t" // here ff where abs(b,d) > Thres, else 00
178		178
179	"movq -2(%%ebx), %%mm0\n\t" // value a	179	"movq -2(%%edx), %%mm0\n\t" // value a
180	"movq 2(%%ebx, %%ecx), %%mm1\n\t" // value f	180	"movq 2(%%edx, %%ecx), %%mm1\n\t" // value f
181	"movq %%mm0, %%mm2\n\t"	181	"movq %%mm0, %%mm2\n\t"
182	V_PAVGB ("%%mm2", "%%mm1", "%%mm3", _ShiftMask) // avg(a,f)	182	V_PAVGB ("%%mm2", "%%mm1", "%%mm3", _ShiftMask) // avg(a,f)
183	"movq %%mm0, %%mm3\n\t"	183	"movq %%mm0, %%mm3\n\t"
Lines 210-217 Link Here
210	"pand "_YMask", %%mm7\n\t" // mask out chroma from here	210	"pand "_YMask", %%mm7\n\t" // mask out chroma from here
211		211
212	// b,e	212	// b,e
213	"movq (%%ebx), %%mm0\n\t" // value b from top	213	"movq (%%edx), %%mm0\n\t" // value b from top
214	"movq (%%ebx, %%ecx), %%mm1\n\t" // value e from bottom	214	"movq (%%edx, %%ecx), %%mm1\n\t" // value e from bottom
215	"movq %%mm0, %%mm2\n\t"	215	"movq %%mm0, %%mm2\n\t"
216	V_PAVGB ("%%mm2", "%%mm1", "%%mm3", _ShiftMask) // avg(b,e)	216	V_PAVGB ("%%mm2", "%%mm1", "%%mm3", _ShiftMask) // avg(b,e)
217	"movq %%mm0, %%mm3\n\t"	217	"movq %%mm0, %%mm3\n\t"
Lines 238-245 Link Here
238	"por %%mm0, %%mm7\n\t"	238	"por %%mm0, %%mm7\n\t"
239		239
240	// bob in any leftovers	240	// bob in any leftovers
241	"movq (%%ebx), %%mm0\n\t" // value b from top	241	"movq (%%edx), %%mm0\n\t" // value b from top
242	"movq (%%ebx, %%ecx), %%mm1\n\t" // value e from bottom	242	"movq (%%edx, %%ecx), %%mm1\n\t" // value e from bottom
243		243
244		244
245	// We will also calc here the max/min values to later limit comb	245	// We will also calc here the max/min values to later limit comb

Lines 14-21 Link Here

(-)filter/tomsmocomp-old/WierdBob.inc (-10 / +10 lines)
14	// selected for the smallest of abs(a,f), abs(c,d), or abs(b,e), etc.	14	// selected for the smallest of abs(a,f), abs(c,d), or abs(b,e), etc.
15		15
16	// a,f	16	// a,f
17	"movq -2(%%ebx), %%mm0\n\t" // value a from top left	17	"movq -2(%%edx), %%mm0\n\t" // value a from top left
18	"movq 2(%%ebx, %%ecx), %%mm1\n\t" // value f from bottom right	18	"movq 2(%%edx, %%ecx), %%mm1\n\t" // value f from bottom right
19	"movq %%mm0, %%mm6\n\t"	19	"movq %%mm0, %%mm6\n\t"
20	// pavgb %%mm6, %%mm1 // avg(a,f), also best so far	20	// pavgb %%mm6, %%mm1 // avg(a,f), also best so far
21	V_PAVGB ("%%mm6", "%%mm1", "%%mm7", _ShiftMask) // avg(a,f), also best so far	21	V_PAVGB ("%%mm6", "%%mm1", "%%mm7", _ShiftMask) // avg(a,f), also best so far
Lines 25-32 Link Here
25	"por %%mm1, %%mm7\n\t" // abs diff, also best so far	25	"por %%mm1, %%mm7\n\t" // abs diff, also best so far
26		26
27	// c,d	27	// c,d
28	"movq 2(%%ebx), %%mm0\n\t" // value a from top left	28	"movq 2(%%edx), %%mm0\n\t" // value a from top left
29	"movq -2(%%ebx, %%ecx), %%mm1\n\t" // value f from bottom right	29	"movq -2(%%edx, %%ecx), %%mm1\n\t" // value f from bottom right
30	"movq %%mm0, %%mm2\n\t"	30	"movq %%mm0, %%mm2\n\t"
31	// pavgb %%mm2, %%mm1 // avg(c,d)	31	// pavgb %%mm2, %%mm1 // avg(c,d)
32	V_PAVGB ("%%mm2", "%%mm1", "%%mm3", _ShiftMask) // avg(c,d)	32	V_PAVGB ("%%mm2", "%%mm1", "%%mm3", _ShiftMask) // avg(c,d)
Lines 53-60 Link Here
53	"pand "_YMask", %%mm5\n\t" // mask out chroma from here also	53	"pand "_YMask", %%mm5\n\t" // mask out chroma from here also
54		54
55	// j,n	55	// j,n
56	"movq -4(%%ebx), %%mm0\n\t" // value j from top left	56	"movq -4(%%edx), %%mm0\n\t" // value j from top left
57	"movq 4(%%ebx, %%ecx), %%mm1\n\t" // value n from bottom right	57	"movq 4(%%edx, %%ecx), %%mm1\n\t" // value n from bottom right
58	"movq %%mm0, %%mm2\n\t"	58	"movq %%mm0, %%mm2\n\t"
59	// pavgb %%mm2, %%mm1 // avg(j,n)	59	// pavgb %%mm2, %%mm1 // avg(j,n)
60	V_PAVGB ("%%mm2", "%%mm1", "%%mm3", _ShiftMask) // avg(j,n)	60	V_PAVGB ("%%mm2", "%%mm1", "%%mm3", _ShiftMask) // avg(j,n)
Lines 79-86 Link Here
79	"por %%mm1, %%mm7\n\t" // "	79	"por %%mm1, %%mm7\n\t" // "
80		80
81	// k, m	81	// k, m
82	"movq 4(%%ebx), %%mm0\n\t" // value k from top right	82	"movq 4(%%edx), %%mm0\n\t" // value k from top right
83	"movq -4(%%ebx, %%ecx), %%mm1\n\t" // value n from bottom left	83	"movq -4(%%edx, %%ecx), %%mm1\n\t" // value n from bottom left
84	"movq %%mm0, %%mm4\n\t"	84	"movq %%mm0, %%mm4\n\t"
85	// pavgb %%mm4, %%mm1 // avg(k,m)	85	// pavgb %%mm4, %%mm1 // avg(k,m)
86	V_PAVGB ("%%mm4", "%%mm1", "%%mm3", _ShiftMask) // avg(k,m)	86	V_PAVGB ("%%mm4", "%%mm1", "%%mm3", _ShiftMask) // avg(k,m)
Lines 108-115 Link Here
108	"por %%mm1, %%mm7\n\t" // "	108	"por %%mm1, %%mm7\n\t" // "
109		109
110	// b,e	110	// b,e
111	"movq (%%ebx), %%mm0\n\t" // value b from top	111	"movq (%%edx), %%mm0\n\t" // value b from top
112	"movq (%%ebx, %%ecx), %%mm1\n\t" // value e from bottom	112	"movq (%%edx, %%ecx), %%mm1\n\t" // value e from bottom
113		113
114	// We will also calc here the max/min values to later limit comb	114	// We will also calc here the max/min values to later limit comb
115	// so the max excursion will not exceed the Max_Comb constant	115	// so the max excursion will not exceed the Max_Comb constant