Context Navigation

← Previous Revision
Latest Revision
Next Revision →
Normal
Revision Log

pngvcrd.c@ 202

Last change on this file since 202 was 95, checked in by Rick van der Zwet, 15 years ago
Bad boy, improper move of directory
File size: 143.9 KB

Rev	Line
[95]	1
	2	/* pngvcrd.c - mixed C/assembler version of utilities to read a PNG file
	3	*
	4	* For Intel x86 CPU and Microsoft Visual C++ compiler
	5	*
	6	* Last changed in libpng 1.2.6 - August 15, 2004
	7	* For conditions of distribution and use, see copyright notice in png.h
	8	* Copyright (c) 1998-2004 Glenn Randers-Pehrson
	9	* Copyright (c) 1998, Intel Corporation
	10	*
	11	* Contributed by Nirav Chhatrapati, Intel Corporation, 1998
	12	* Interface to libpng contributed by Gilles Vollant, 1999
	13	*
	14	*
	15	* In png_do_read_interlace() in libpng versions 1.0.3a through 1.0.4d,
	16	* a sign error in the post-MMX cleanup code for each pixel_depth resulted
	17	* in bad pixels at the beginning of some rows of some images, and also
	18	* (due to out-of-range memory reads and writes) caused heap corruption
	19	* when compiled with MSVC 6.0. The error was fixed in version 1.0.4e.
	20	*
	21	* [png_read_filter_row_mmx_avg() bpp == 2 bugfix, GRR 20000916]
	22	*
	23	* [runtime MMX configuration, GRR 20010102]
	24	*
	25	*/
	26
	27	#define PNG_INTERNAL
	28	#include "png.h"
	29
	30	#if defined(PNG_ASSEMBLER_CODE_SUPPORTED) && defined(PNG_USE_PNGVCRD)
	31
	32	static int mmx_supported=2;
	33
	34
	35	int PNGAPI
	36	png_mmx_support(void)
	37	{
	38	int mmx_supported_local = 0;
	39	_asm {
	40	push ebx //CPUID will trash these
	41	push ecx
	42	push edx
	43
	44	pushfd //Save Eflag to stack
	45	pop eax //Get Eflag from stack into eax
	46	mov ecx, eax //Make another copy of Eflag in ecx
	47	xor eax, 0x200000 //Toggle ID bit in Eflag [i.e. bit(21)]
	48	push eax //Save modified Eflag back to stack
	49
	50	popfd //Restored modified value back to Eflag reg
	51	pushfd //Save Eflag to stack
	52	pop eax //Get Eflag from stack
	53	push ecx // save original Eflag to stack
	54	popfd // restore original Eflag
	55	xor eax, ecx //Compare the new Eflag with the original Eflag
	56	jz NOT_SUPPORTED //If the same, CPUID instruction is not supported,
	57	//skip following instructions and jump to
	58	//NOT_SUPPORTED label
	59
	60	xor eax, eax //Set eax to zero
	61
	62	_asm _emit 0x0f //CPUID instruction (two bytes opcode)
	63	_asm _emit 0xa2
	64
	65	cmp eax, 1 //make sure eax return non-zero value
	66	jl NOT_SUPPORTED //If eax is zero, mmx not supported
	67
	68	xor eax, eax //set eax to zero
	69	inc eax //Now increment eax to 1. This instruction is
	70	//faster than the instruction "mov eax, 1"
	71
	72	_asm _emit 0x0f //CPUID instruction
	73	_asm _emit 0xa2
	74
	75	and edx, 0x00800000 //mask out all bits but mmx bit(24)
	76	cmp edx, 0 // 0 = mmx not supported
	77	jz NOT_SUPPORTED // non-zero = Yes, mmx IS supported
	78
	79	mov mmx_supported_local, 1 //set return value to 1
	80
	81	NOT_SUPPORTED:
	82	mov eax, mmx_supported_local //move return value to eax
	83	pop edx //CPUID trashed these
	84	pop ecx
	85	pop ebx
	86	}
	87
	88	//mmx_supported_local=0; // test code for force don't support MMX
	89	//printf("MMX : %u (1=MMX supported)\n",mmx_supported_local);
	90
	91	mmx_supported = mmx_supported_local;
	92	return mmx_supported_local;
	93	}
	94
	95	/* Combines the row recently read in with the previous row.
	96	This routine takes care of alpha and transparency if requested.
	97	This routine also handles the two methods of progressive display
	98	of interlaced images, depending on the mask value.
	99	The mask value describes which pixels are to be combined with
	100	the row. The pattern always repeats every 8 pixels, so just 8
	101	bits are needed. A one indicates the pixel is to be combined; a
	102	zero indicates the pixel is to be skipped. This is in addition
	103	to any alpha or transparency value associated with the pixel. If
	104	you want all pixels to be combined, pass 0xff (255) in mask. */
	105
	106	/* Use this routine for x86 platform - uses faster MMX routine if machine
	107	supports MMX */
	108
	109	void /* PRIVATE */
	110	png_combine_row(png_structp png_ptr, png_bytep row, int mask)
	111	{
	112	#ifdef PNG_USE_LOCAL_ARRAYS
	113	const int png_pass_inc[7] = {8, 8, 4, 4, 2, 2, 1};
	114	#endif
	115
	116	png_debug(1,"in png_combine_row_asm\n");
	117
	118	if (mmx_supported == 2) {
	119	#if !defined(PNG_1_0_X)
	120	/* this should have happened in png_init_mmx_flags() already */
	121	png_warning(png_ptr, "asm_flags may not have been initialized");
	122	#endif
	123	png_mmx_support();
	124	}
	125
	126	if (mask == 0xff)
	127	{
	128	png_memcpy(row, png_ptr->row_buf + 1,
	129	(png_size_t)PNG_ROWBYTES(png_ptr->row_info.pixel_depth,
	130	png_ptr->width));
	131	}
	132	/* GRR: add "else if (mask == 0)" case?
	133	* or does png_combine_row() not even get called in that case? */
	134	else
	135	{
	136	switch (png_ptr->row_info.pixel_depth)
	137	{
	138	case 1:
	139	{
	140	png_bytep sp;
	141	png_bytep dp;
	142	int s_inc, s_start, s_end;
	143	int m;
	144	int shift;
	145	png_uint_32 i;
	146
	147	sp = png_ptr->row_buf + 1;
	148	dp = row;
	149	m = 0x80;
	150	#if defined(PNG_READ_PACKSWAP_SUPPORTED)
	151	if (png_ptr->transformations & PNG_PACKSWAP)
	152	{
	153	s_start = 0;
	154	s_end = 7;
	155	s_inc = 1;
	156	}
	157	else
	158	#endif
	159	{
	160	s_start = 7;
	161	s_end = 0;
	162	s_inc = -1;
	163	}
	164
	165	shift = s_start;
	166
	167	for (i = 0; i < png_ptr->width; i++)
	168	{
	169	if (m & mask)
	170	{
	171	int value;
	172
	173	value = (*sp >> shift) & 0x1;
	174	*dp &= (png_byte)((0x7f7f >> (7 - shift)) & 0xff);
	175	*dp \|= (png_byte)(value << shift);
	176	}
	177
	178	if (shift == s_end)
	179	{
	180	shift = s_start;
	181	sp++;
	182	dp++;
	183	}
	184	else
	185	shift += s_inc;
	186
	187	if (m == 1)
	188	m = 0x80;
	189	else
	190	m >>= 1;
	191	}
	192	break;
	193	}
	194
	195	case 2:
	196	{
	197	png_bytep sp;
	198	png_bytep dp;
	199	int s_start, s_end, s_inc;
	200	int m;
	201	int shift;
	202	png_uint_32 i;
	203	int value;
	204
	205	sp = png_ptr->row_buf + 1;
	206	dp = row;
	207	m = 0x80;
	208	#if defined(PNG_READ_PACKSWAP_SUPPORTED)
	209	if (png_ptr->transformations & PNG_PACKSWAP)
	210	{
	211	s_start = 0;
	212	s_end = 6;
	213	s_inc = 2;
	214	}
	215	else
	216	#endif
	217	{
	218	s_start = 6;
	219	s_end = 0;
	220	s_inc = -2;
	221	}
	222
	223	shift = s_start;
	224
	225	for (i = 0; i < png_ptr->width; i++)
	226	{
	227	if (m & mask)
	228	{
	229	value = (*sp >> shift) & 0x3;
	230	*dp &= (png_byte)((0x3f3f >> (6 - shift)) & 0xff);
	231	*dp \|= (png_byte)(value << shift);
	232	}
	233
	234	if (shift == s_end)
	235	{
	236	shift = s_start;
	237	sp++;
	238	dp++;
	239	}
	240	else
	241	shift += s_inc;
	242	if (m == 1)
	243	m = 0x80;
	244	else
	245	m >>= 1;
	246	}
	247	break;
	248	}
	249
	250	case 4:
	251	{
	252	png_bytep sp;
	253	png_bytep dp;
	254	int s_start, s_end, s_inc;
	255	int m;
	256	int shift;
	257	png_uint_32 i;
	258	int value;
	259
	260	sp = png_ptr->row_buf + 1;
	261	dp = row;
	262	m = 0x80;
	263	#if defined(PNG_READ_PACKSWAP_SUPPORTED)
	264	if (png_ptr->transformations & PNG_PACKSWAP)
	265	{
	266	s_start = 0;
	267	s_end = 4;
	268	s_inc = 4;
	269	}
	270	else
	271	#endif
	272	{
	273	s_start = 4;
	274	s_end = 0;
	275	s_inc = -4;
	276	}
	277	shift = s_start;
	278
	279	for (i = 0; i < png_ptr->width; i++)
	280	{
	281	if (m & mask)
	282	{
	283	value = (*sp >> shift) & 0xf;
	284	*dp &= (png_byte)((0xf0f >> (4 - shift)) & 0xff);
	285	*dp \|= (png_byte)(value << shift);
	286	}
	287
	288	if (shift == s_end)
	289	{
	290	shift = s_start;
	291	sp++;
	292	dp++;
	293	}
	294	else
	295	shift += s_inc;
	296	if (m == 1)
	297	m = 0x80;
	298	else
	299	m >>= 1;
	300	}
	301	break;
	302	}
	303
	304	case 8:
	305	{
	306	png_bytep srcptr;
	307	png_bytep dstptr;
	308	png_uint_32 len;
	309	int m;
	310	int diff, unmask;
	311
	312	__int64 mask0=0x0102040810204080;
	313
	314	#if !defined(PNG_1_0_X)
	315	if ((png_ptr->asm_flags & PNG_ASM_FLAG_MMX_READ_COMBINE_ROW)
	316	/* && mmx_supported */ )
	317	#else
	318	if (mmx_supported)
	319	#endif
	320	{
	321	srcptr = png_ptr->row_buf + 1;
	322	dstptr = row;
	323	m = 0x80;
	324	unmask = ~mask;
	325	len = png_ptr->width &~7; //reduce to multiple of 8
	326	diff = png_ptr->width & 7; //amount lost
	327
	328	_asm
	329	{
	330	movd mm7, unmask //load bit pattern
	331	psubb mm6,mm6 //zero mm6
	332	punpcklbw mm7,mm7
	333	punpcklwd mm7,mm7
	334	punpckldq mm7,mm7 //fill register with 8 masks
	335
	336	movq mm0,mask0
	337
	338	pand mm0,mm7 //nonzero if keep byte
	339	pcmpeqb mm0,mm6 //zeros->1s, v versa
	340
	341	mov ecx,len //load length of line (pixels)
	342	mov esi,srcptr //load source
	343	mov ebx,dstptr //load dest
	344	cmp ecx,0 //lcr
	345	je mainloop8end
	346
	347	mainloop8:
	348	movq mm4,[esi]
	349	pand mm4,mm0
	350	movq mm6,mm0
	351	pandn mm6,[ebx]
	352	por mm4,mm6
	353	movq [ebx],mm4
	354
	355	add esi,8 //inc by 8 bytes processed
	356	add ebx,8
	357	sub ecx,8 //dec by 8 pixels processed
	358
	359	ja mainloop8
	360	mainloop8end:
	361
	362	mov ecx,diff
	363	cmp ecx,0
	364	jz end8
	365
	366	mov edx,mask
	367	sal edx,24 //make low byte the high byte
	368
	369	secondloop8:
	370	sal edx,1 //move high bit to CF
	371	jnc skip8 //if CF = 0
	372	mov al,[esi]
	373	mov [ebx],al
	374	skip8:
	375	inc esi
	376	inc ebx
	377
	378	dec ecx
	379	jnz secondloop8
	380	end8:
	381	emms
	382	}
	383	}
	384	else /* mmx not supported - use modified C routine */
	385	{
	386	register unsigned int incr1, initial_val, final_val;
	387	png_size_t pixel_bytes;
	388	png_uint_32 i;
	389	register int disp = png_pass_inc[png_ptr->pass];
	390	int offset_table[7] = {0, 4, 0, 2, 0, 1, 0};
	391
	392	pixel_bytes = (png_ptr->row_info.pixel_depth >> 3);
	393	srcptr = png_ptr->row_buf + 1 + offset_table[png_ptr->pass]*
	394	pixel_bytes;
	395	dstptr = row + offset_table[png_ptr->pass]*pixel_bytes;
	396	initial_val = offset_table[png_ptr->pass]*pixel_bytes;
	397	final_val = png_ptr->width*pixel_bytes;
	398	incr1 = (disp)*pixel_bytes;
	399	for (i = initial_val; i < final_val; i += incr1)
	400	{
	401	png_memcpy(dstptr, srcptr, pixel_bytes);
	402	srcptr += incr1;
	403	dstptr += incr1;
	404	}
	405	} /* end of else */
	406
	407	break;
	408	} // end 8 bpp
	409
	410	case 16:
	411	{
	412	png_bytep srcptr;
	413	png_bytep dstptr;
	414	png_uint_32 len;
	415	int unmask, diff;
	416	__int64 mask1=0x0101020204040808,
	417	mask0=0x1010202040408080;
	418
	419	#if !defined(PNG_1_0_X)
	420	if ((png_ptr->asm_flags & PNG_ASM_FLAG_MMX_READ_COMBINE_ROW)
	421	/* && mmx_supported */ )
	422	#else
	423	if (mmx_supported)
	424	#endif
	425	{
	426	srcptr = png_ptr->row_buf + 1;
	427	dstptr = row;
	428
	429	unmask = ~mask;
	430	len = (png_ptr->width)&~7;
	431	diff = (png_ptr->width)&7;
	432	_asm
	433	{
	434	movd mm7, unmask //load bit pattern
	435	psubb mm6,mm6 //zero mm6
	436	punpcklbw mm7,mm7
	437	punpcklwd mm7,mm7
	438	punpckldq mm7,mm7 //fill register with 8 masks
	439
	440	movq mm0,mask0
	441	movq mm1,mask1
	442
	443	pand mm0,mm7
	444	pand mm1,mm7
	445
	446	pcmpeqb mm0,mm6
	447	pcmpeqb mm1,mm6
	448
	449	mov ecx,len //load length of line
	450	mov esi,srcptr //load source
	451	mov ebx,dstptr //load dest
	452	cmp ecx,0 //lcr
	453	jz mainloop16end
	454
	455	mainloop16:
	456	movq mm4,[esi]
	457	pand mm4,mm0
	458	movq mm6,mm0
	459	movq mm7,[ebx]
	460	pandn mm6,mm7
	461	por mm4,mm6
	462	movq [ebx],mm4
	463
	464	movq mm5,[esi+8]
	465	pand mm5,mm1
	466	movq mm7,mm1
	467	movq mm6,[ebx+8]
	468	pandn mm7,mm6
	469	por mm5,mm7
	470	movq [ebx+8],mm5
	471
	472	add esi,16 //inc by 16 bytes processed
	473	add ebx,16
	474	sub ecx,8 //dec by 8 pixels processed
	475
	476	ja mainloop16
	477
	478	mainloop16end:
	479	mov ecx,diff
	480	cmp ecx,0
	481	jz end16
	482
	483	mov edx,mask
	484	sal edx,24 //make low byte the high byte
	485	secondloop16:
	486	sal edx,1 //move high bit to CF
	487	jnc skip16 //if CF = 0
	488	mov ax,[esi]
	489	mov [ebx],ax
	490	skip16:
	491	add esi,2
	492	add ebx,2
	493
	494	dec ecx
	495	jnz secondloop16
	496	end16:
	497	emms
	498	}
	499	}
	500	else /* mmx not supported - use modified C routine */
	501	{
	502	register unsigned int incr1, initial_val, final_val;
	503	png_size_t pixel_bytes;
	504	png_uint_32 i;
	505	register int disp = png_pass_inc[png_ptr->pass];
	506	int offset_table[7] = {0, 4, 0, 2, 0, 1, 0};
	507
	508	pixel_bytes = (png_ptr->row_info.pixel_depth >> 3);
	509	srcptr = png_ptr->row_buf + 1 + offset_table[png_ptr->pass]*
	510	pixel_bytes;
	511	dstptr = row + offset_table[png_ptr->pass]*pixel_bytes;
	512	initial_val = offset_table[png_ptr->pass]*pixel_bytes;
	513	final_val = png_ptr->width*pixel_bytes;
	514	incr1 = (disp)*pixel_bytes;
	515	for (i = initial_val; i < final_val; i += incr1)
	516	{
	517	png_memcpy(dstptr, srcptr, pixel_bytes);
	518	srcptr += incr1;
	519	dstptr += incr1;
	520	}
	521	} /* end of else */
	522
	523	break;
	524	} // end 16 bpp
	525
	526	case 24:
	527	{
	528	png_bytep srcptr;
	529	png_bytep dstptr;
	530	png_uint_32 len;
	531	int unmask, diff;
	532
	533	__int64 mask2=0x0101010202020404, //24bpp
	534	mask1=0x0408080810101020,
	535	mask0=0x2020404040808080;
	536
	537	srcptr = png_ptr->row_buf + 1;
	538	dstptr = row;
	539
	540	unmask = ~mask;
	541	len = (png_ptr->width)&~7;
	542	diff = (png_ptr->width)&7;
	543
	544	#if !defined(PNG_1_0_X)
	545	if ((png_ptr->asm_flags & PNG_ASM_FLAG_MMX_READ_COMBINE_ROW)
	546	/* && mmx_supported */ )
	547	#else
	548	if (mmx_supported)
	549	#endif
	550	{
	551	_asm
	552	{
	553	movd mm7, unmask //load bit pattern
	554	psubb mm6,mm6 //zero mm6
	555	punpcklbw mm7,mm7
	556	punpcklwd mm7,mm7
	557	punpckldq mm7,mm7 //fill register with 8 masks
	558
	559	movq mm0,mask0
	560	movq mm1,mask1
	561	movq mm2,mask2
	562
	563	pand mm0,mm7
	564	pand mm1,mm7
	565	pand mm2,mm7
	566
	567	pcmpeqb mm0,mm6
	568	pcmpeqb mm1,mm6
	569	pcmpeqb mm2,mm6
	570
	571	mov ecx,len //load length of line
	572	mov esi,srcptr //load source
	573	mov ebx,dstptr //load dest
	574	cmp ecx,0
	575	jz mainloop24end
	576
	577	mainloop24:
	578	movq mm4,[esi]
	579	pand mm4,mm0
	580	movq mm6,mm0
	581	movq mm7,[ebx]
	582	pandn mm6,mm7
	583	por mm4,mm6
	584	movq [ebx],mm4
	585
	586
	587	movq mm5,[esi+8]
	588	pand mm5,mm1
	589	movq mm7,mm1
	590	movq mm6,[ebx+8]
	591	pandn mm7,mm6
	592	por mm5,mm7
	593	movq [ebx+8],mm5
	594
	595	movq mm6,[esi+16]
	596	pand mm6,mm2
	597	movq mm4,mm2
	598	movq mm7,[ebx+16]
	599	pandn mm4,mm7
	600	por mm6,mm4
	601	movq [ebx+16],mm6
	602
	603	add esi,24 //inc by 24 bytes processed
	604	add ebx,24
	605	sub ecx,8 //dec by 8 pixels processed
	606
	607	ja mainloop24
	608
	609	mainloop24end:
	610	mov ecx,diff
	611	cmp ecx,0
	612	jz end24
	613
	614	mov edx,mask
	615	sal edx,24 //make low byte the high byte
	616	secondloop24:
	617	sal edx,1 //move high bit to CF
	618	jnc skip24 //if CF = 0
	619	mov ax,[esi]
	620	mov [ebx],ax
	621	xor eax,eax
	622	mov al,[esi+2]
	623	mov [ebx+2],al
	624	skip24:
	625	add esi,3
	626	add ebx,3
	627
	628	dec ecx
	629	jnz secondloop24
	630
	631	end24:
	632	emms
	633	}
	634	}
	635	else /* mmx not supported - use modified C routine */
	636	{
	637	register unsigned int incr1, initial_val, final_val;
	638	png_size_t pixel_bytes;
	639	png_uint_32 i;
	640	register int disp = png_pass_inc[png_ptr->pass];
	641	int offset_table[7] = {0, 4, 0, 2, 0, 1, 0};
	642
	643	pixel_bytes = (png_ptr->row_info.pixel_depth >> 3);
	644	srcptr = png_ptr->row_buf + 1 + offset_table[png_ptr->pass]*
	645	pixel_bytes;
	646	dstptr = row + offset_table[png_ptr->pass]*pixel_bytes;
	647	initial_val = offset_table[png_ptr->pass]*pixel_bytes;
	648	final_val = png_ptr->width*pixel_bytes;
	649	incr1 = (disp)*pixel_bytes;
	650	for (i = initial_val; i < final_val; i += incr1)
	651	{
	652	png_memcpy(dstptr, srcptr, pixel_bytes);
	653	srcptr += incr1;
	654	dstptr += incr1;
	655	}
	656	} /* end of else */
	657
	658	break;
	659	} // end 24 bpp
	660
	661	case 32:
	662	{
	663	png_bytep srcptr;
	664	png_bytep dstptr;
	665	png_uint_32 len;
	666	int unmask, diff;
	667
	668	__int64 mask3=0x0101010102020202, //32bpp
	669	mask2=0x0404040408080808,
	670	mask1=0x1010101020202020,
	671	mask0=0x4040404080808080;
	672
	673	srcptr = png_ptr->row_buf + 1;
	674	dstptr = row;
	675
	676	unmask = ~mask;
	677	len = (png_ptr->width)&~7;
	678	diff = (png_ptr->width)&7;
	679
	680	#if !defined(PNG_1_0_X)
	681	if ((png_ptr->asm_flags & PNG_ASM_FLAG_MMX_READ_COMBINE_ROW)
	682	/* && mmx_supported */ )
	683	#else
	684	if (mmx_supported)
	685	#endif
	686	{
	687	_asm
	688	{
	689	movd mm7, unmask //load bit pattern
	690	psubb mm6,mm6 //zero mm6
	691	punpcklbw mm7,mm7
	692	punpcklwd mm7,mm7
	693	punpckldq mm7,mm7 //fill register with 8 masks
	694
	695	movq mm0,mask0
	696	movq mm1,mask1
	697	movq mm2,mask2
	698	movq mm3,mask3
	699
	700	pand mm0,mm7
	701	pand mm1,mm7
	702	pand mm2,mm7
	703	pand mm3,mm7
	704
	705	pcmpeqb mm0,mm6
	706	pcmpeqb mm1,mm6
	707	pcmpeqb mm2,mm6
	708	pcmpeqb mm3,mm6
	709
	710	mov ecx,len //load length of line
	711	mov esi,srcptr //load source
	712	mov ebx,dstptr //load dest
	713
	714	cmp ecx,0 //lcr
	715	jz mainloop32end
	716
	717	mainloop32:
	718	movq mm4,[esi]
	719	pand mm4,mm0
	720	movq mm6,mm0
	721	movq mm7,[ebx]
	722	pandn mm6,mm7
	723	por mm4,mm6
	724	movq [ebx],mm4
	725
	726	movq mm5,[esi+8]
	727	pand mm5,mm1
	728	movq mm7,mm1
	729	movq mm6,[ebx+8]
	730	pandn mm7,mm6
	731	por mm5,mm7
	732	movq [ebx+8],mm5
	733
	734	movq mm6,[esi+16]
	735	pand mm6,mm2
	736	movq mm4,mm2
	737	movq mm7,[ebx+16]
	738	pandn mm4,mm7
	739	por mm6,mm4
	740	movq [ebx+16],mm6
	741
	742	movq mm7,[esi+24]
	743	pand mm7,mm3
	744	movq mm5,mm3
	745	movq mm4,[ebx+24]
	746	pandn mm5,mm4
	747	por mm7,mm5
	748	movq [ebx+24],mm7
	749
	750	add esi,32 //inc by 32 bytes processed
	751	add ebx,32
	752	sub ecx,8 //dec by 8 pixels processed
	753
	754	ja mainloop32
	755
	756	mainloop32end:
	757	mov ecx,diff
	758	cmp ecx,0
	759	jz end32
	760
	761	mov edx,mask
	762	sal edx,24 //make low byte the high byte
	763	secondloop32:
	764	sal edx,1 //move high bit to CF
	765	jnc skip32 //if CF = 0
	766	mov eax,[esi]
	767	mov [ebx],eax
	768	skip32:
	769	add esi,4
	770	add ebx,4
	771
	772	dec ecx
	773	jnz secondloop32
	774
	775	end32:
	776	emms
	777	}
	778	}
	779	else /* mmx _not supported - Use modified C routine */
	780	{
	781	register unsigned int incr1, initial_val, final_val;
	782	png_size_t pixel_bytes;
	783	png_uint_32 i;
	784	register int disp = png_pass_inc[png_ptr->pass];
	785	int offset_table[7] = {0, 4, 0, 2, 0, 1, 0};
	786
	787	pixel_bytes = (png_ptr->row_info.pixel_depth >> 3);
	788	srcptr = png_ptr->row_buf + 1 + offset_table[png_ptr->pass]*
	789	pixel_bytes;
	790	dstptr = row + offset_table[png_ptr->pass]*pixel_bytes;
	791	initial_val = offset_table[png_ptr->pass]*pixel_bytes;
	792	final_val = png_ptr->width*pixel_bytes;
	793	incr1 = (disp)*pixel_bytes;
	794	for (i = initial_val; i < final_val; i += incr1)
	795	{
	796	png_memcpy(dstptr, srcptr, pixel_bytes);
	797	srcptr += incr1;
	798	dstptr += incr1;
	799	}
	800	} /* end of else */
	801
	802	break;
	803	} // end 32 bpp
	804
	805	case 48:
	806	{
	807	png_bytep srcptr;
	808	png_bytep dstptr;
	809	png_uint_32 len;
	810	int unmask, diff;
	811
	812	__int64 mask5=0x0101010101010202,
	813	mask4=0x0202020204040404,
	814	mask3=0x0404080808080808,
	815	mask2=0x1010101010102020,
	816	mask1=0x2020202040404040,
	817	mask0=0x4040808080808080;
	818
	819	#if !defined(PNG_1_0_X)
	820	if ((png_ptr->asm_flags & PNG_ASM_FLAG_MMX_READ_COMBINE_ROW)
	821	/* && mmx_supported */ )
	822	#else
	823	if (mmx_supported)
	824	#endif
	825	{
	826	srcptr = png_ptr->row_buf + 1;
	827	dstptr = row;
	828
	829	unmask = ~mask;
	830	len = (png_ptr->width)&~7;
	831	diff = (png_ptr->width)&7;
	832	_asm
	833	{
	834	movd mm7, unmask //load bit pattern
	835	psubb mm6,mm6 //zero mm6
	836	punpcklbw mm7,mm7
	837	punpcklwd mm7,mm7
	838	punpckldq mm7,mm7 //fill register with 8 masks
	839
	840	movq mm0,mask0
	841	movq mm1,mask1
	842	movq mm2,mask2
	843	movq mm3,mask3
	844	movq mm4,mask4
	845	movq mm5,mask5
	846
	847	pand mm0,mm7
	848	pand mm1,mm7
	849	pand mm2,mm7
	850	pand mm3,mm7
	851	pand mm4,mm7
	852	pand mm5,mm7
	853
	854	pcmpeqb mm0,mm6
	855	pcmpeqb mm1,mm6
	856	pcmpeqb mm2,mm6
	857	pcmpeqb mm3,mm6
	858	pcmpeqb mm4,mm6
	859	pcmpeqb mm5,mm6
	860
	861	mov ecx,len //load length of line
	862	mov esi,srcptr //load source
	863	mov ebx,dstptr //load dest
	864
	865	cmp ecx,0
	866	jz mainloop48end
	867
	868	mainloop48:
	869	movq mm7,[esi]
	870	pand mm7,mm0
	871	movq mm6,mm0
	872	pandn mm6,[ebx]
	873	por mm7,mm6
	874	movq [ebx],mm7
	875
	876	movq mm6,[esi+8]
	877	pand mm6,mm1
	878	movq mm7,mm1
	879	pandn mm7,[ebx+8]
	880	por mm6,mm7
	881	movq [ebx+8],mm6
	882
	883	movq mm6,[esi+16]
	884	pand mm6,mm2
	885	movq mm7,mm2
	886	pandn mm7,[ebx+16]
	887	por mm6,mm7
	888	movq [ebx+16],mm6
	889
	890	movq mm7,[esi+24]
	891	pand mm7,mm3
	892	movq mm6,mm3
	893	pandn mm6,[ebx+24]
	894	por mm7,mm6
	895	movq [ebx+24],mm7
	896
	897	movq mm6,[esi+32]
	898	pand mm6,mm4
	899	movq mm7,mm4
	900	pandn mm7,[ebx+32]
	901	por mm6,mm7
	902	movq [ebx+32],mm6
	903
	904	movq mm7,[esi+40]
	905	pand mm7,mm5
	906	movq mm6,mm5
	907	pandn mm6,[ebx+40]
	908	por mm7,mm6
	909	movq [ebx+40],mm7
	910
	911	add esi,48 //inc by 32 bytes processed
	912	add ebx,48
	913	sub ecx,8 //dec by 8 pixels processed
	914
	915	ja mainloop48
	916	mainloop48end:
	917
	918	mov ecx,diff
	919	cmp ecx,0
	920	jz end48
	921
	922	mov edx,mask
	923	sal edx,24 //make low byte the high byte
	924
	925	secondloop48:
	926	sal edx,1 //move high bit to CF
	927	jnc skip48 //if CF = 0
	928	mov eax,[esi]
	929	mov [ebx],eax
	930	skip48:
	931	add esi,4
	932	add ebx,4
	933
	934	dec ecx
	935	jnz secondloop48
	936
	937	end48:
	938	emms
	939	}
	940	}
	941	else /* mmx _not supported - Use modified C routine */
	942	{
	943	register unsigned int incr1, initial_val, final_val;
	944	png_size_t pixel_bytes;
	945	png_uint_32 i;
	946	register int disp = png_pass_inc[png_ptr->pass];
	947	int offset_table[7] = {0, 4, 0, 2, 0, 1, 0};
	948
	949	pixel_bytes = (png_ptr->row_info.pixel_depth >> 3);
	950	srcptr = png_ptr->row_buf + 1 + offset_table[png_ptr->pass]*
	951	pixel_bytes;
	952	dstptr = row + offset_table[png_ptr->pass]*pixel_bytes;
	953	initial_val = offset_table[png_ptr->pass]*pixel_bytes;
	954	final_val = png_ptr->width*pixel_bytes;
	955	incr1 = (disp)*pixel_bytes;
	956	for (i = initial_val; i < final_val; i += incr1)
	957	{
	958	png_memcpy(dstptr, srcptr, pixel_bytes);
	959	srcptr += incr1;
	960	dstptr += incr1;
	961	}
	962	} /* end of else */
	963
	964	break;
	965	} // end 48 bpp
	966
	967	default:
	968	{
	969	png_bytep sptr;
	970	png_bytep dp;
	971	png_size_t pixel_bytes;
	972	int offset_table[7] = {0, 4, 0, 2, 0, 1, 0};
	973	unsigned int i;
	974	register int disp = png_pass_inc[png_ptr->pass]; // get the offset
	975	register unsigned int incr1, initial_val, final_val;
	976
	977	pixel_bytes = (png_ptr->row_info.pixel_depth >> 3);
	978	sptr = png_ptr->row_buf + 1 + offset_table[png_ptr->pass]*
	979	pixel_bytes;
	980	dp = row + offset_table[png_ptr->pass]*pixel_bytes;
	981	initial_val = offset_table[png_ptr->pass]*pixel_bytes;
	982	final_val = png_ptr->width*pixel_bytes;
	983	incr1 = (disp)*pixel_bytes;
	984	for (i = initial_val; i < final_val; i += incr1)
	985	{
	986	png_memcpy(dp, sptr, pixel_bytes);
	987	sptr += incr1;
	988	dp += incr1;
	989	}
	990	break;
	991	}
	992	} /* end switch (png_ptr->row_info.pixel_depth) */
	993	} /* end if (non-trivial mask) */
	994
	995	} /* end png_combine_row() */
	996
	997
	998	#if defined(PNG_READ_INTERLACING_SUPPORTED)
	999
	1000	void /* PRIVATE */
	1001	png_do_read_interlace(png_structp png_ptr)
	1002	{
	1003	png_row_infop row_info = &(png_ptr->row_info);
	1004	png_bytep row = png_ptr->row_buf + 1;
	1005	int pass = png_ptr->pass;
	1006	png_uint_32 transformations = png_ptr->transformations;
	1007	#ifdef PNG_USE_LOCAL_ARRAYS
	1008	const int png_pass_inc[7] = {8, 8, 4, 4, 2, 2, 1};
	1009	#endif
	1010
	1011	png_debug(1,"in png_do_read_interlace\n");
	1012
	1013	if (mmx_supported == 2) {
	1014	#if !defined(PNG_1_0_X)
	1015	/* this should have happened in png_init_mmx_flags() already */
	1016	png_warning(png_ptr, "asm_flags may not have been initialized");
	1017	#endif
	1018	png_mmx_support();
	1019	}
	1020
	1021	if (row != NULL && row_info != NULL)
	1022	{
	1023	png_uint_32 final_width;
	1024
	1025	final_width = row_info->width * png_pass_inc[pass];
	1026
	1027	switch (row_info->pixel_depth)
	1028	{
	1029	case 1:
	1030	{
	1031	png_bytep sp, dp;
	1032	int sshift, dshift;
	1033	int s_start, s_end, s_inc;
	1034	png_byte v;
	1035	png_uint_32 i;
	1036	int j;
	1037
	1038	sp = row + (png_size_t)((row_info->width - 1) >> 3);
	1039	dp = row + (png_size_t)((final_width - 1) >> 3);
	1040	#if defined(PNG_READ_PACKSWAP_SUPPORTED)
	1041	if (transformations & PNG_PACKSWAP)
	1042	{
	1043	sshift = (int)((row_info->width + 7) & 7);
	1044	dshift = (int)((final_width + 7) & 7);
	1045	s_start = 7;
	1046	s_end = 0;
	1047	s_inc = -1;
	1048	}
	1049	else
	1050	#endif
	1051	{
	1052	sshift = 7 - (int)((row_info->width + 7) & 7);
	1053	dshift = 7 - (int)((final_width + 7) & 7);
	1054	s_start = 0;
	1055	s_end = 7;
	1056	s_inc = 1;
	1057	}
	1058
	1059	for (i = row_info->width; i; i--)
	1060	{
	1061	v = (png_byte)((*sp >> sshift) & 0x1);
	1062	for (j = 0; j < png_pass_inc[pass]; j++)
	1063	{
	1064	*dp &= (png_byte)((0x7f7f >> (7 - dshift)) & 0xff);
	1065	*dp \|= (png_byte)(v << dshift);
	1066	if (dshift == s_end)
	1067	{
	1068	dshift = s_start;
	1069	dp--;
	1070	}
	1071	else
	1072	dshift += s_inc;
	1073	}
	1074	if (sshift == s_end)
	1075	{
	1076	sshift = s_start;
	1077	sp--;
	1078	}
	1079	else
	1080	sshift += s_inc;
	1081	}
	1082	break;
	1083	}
	1084
	1085	case 2:
	1086	{
	1087	png_bytep sp, dp;
	1088	int sshift, dshift;
	1089	int s_start, s_end, s_inc;
	1090	png_uint_32 i;
	1091
	1092	sp = row + (png_size_t)((row_info->width - 1) >> 2);
	1093	dp = row + (png_size_t)((final_width - 1) >> 2);
	1094	#if defined(PNG_READ_PACKSWAP_SUPPORTED)
	1095	if (transformations & PNG_PACKSWAP)
	1096	{
	1097	sshift = (png_size_t)(((row_info->width + 3) & 3) << 1);
	1098	dshift = (png_size_t)(((final_width + 3) & 3) << 1);
	1099	s_start = 6;
	1100	s_end = 0;
	1101	s_inc = -2;
	1102	}
	1103	else
	1104	#endif
	1105	{
	1106	sshift = (png_size_t)((3 - ((row_info->width + 3) & 3)) << 1);
	1107	dshift = (png_size_t)((3 - ((final_width + 3) & 3)) << 1);
	1108	s_start = 0;
	1109	s_end = 6;
	1110	s_inc = 2;
	1111	}
	1112
	1113	for (i = row_info->width; i; i--)
	1114	{
	1115	png_byte v;
	1116	int j;
	1117
	1118	v = (png_byte)((*sp >> sshift) & 0x3);
	1119	for (j = 0; j < png_pass_inc[pass]; j++)
	1120	{
	1121	*dp &= (png_byte)((0x3f3f >> (6 - dshift)) & 0xff);
	1122	*dp \|= (png_byte)(v << dshift);
	1123	if (dshift == s_end)
	1124	{
	1125	dshift = s_start;
	1126	dp--;
	1127	}
	1128	else
	1129	dshift += s_inc;
	1130	}
	1131	if (sshift == s_end)
	1132	{
	1133	sshift = s_start;
	1134	sp--;
	1135	}
	1136	else
	1137	sshift += s_inc;
	1138	}
	1139	break;
	1140	}
	1141
	1142	case 4:
	1143	{
	1144	png_bytep sp, dp;
	1145	int sshift, dshift;
	1146	int s_start, s_end, s_inc;
	1147	png_uint_32 i;
	1148
	1149	sp = row + (png_size_t)((row_info->width - 1) >> 1);
	1150	dp = row + (png_size_t)((final_width - 1) >> 1);
	1151	#if defined(PNG_READ_PACKSWAP_SUPPORTED)
	1152	if (transformations & PNG_PACKSWAP)
	1153	{
	1154	sshift = (png_size_t)(((row_info->width + 1) & 1) << 2);
	1155	dshift = (png_size_t)(((final_width + 1) & 1) << 2);
	1156	s_start = 4;
	1157	s_end = 0;
	1158	s_inc = -4;
	1159	}
	1160	else
	1161	#endif
	1162	{
	1163	sshift = (png_size_t)((1 - ((row_info->width + 1) & 1)) << 2);
	1164	dshift = (png_size_t)((1 - ((final_width + 1) & 1)) << 2);
	1165	s_start = 0;
	1166	s_end = 4;
	1167	s_inc = 4;
	1168	}
	1169
	1170	for (i = row_info->width; i; i--)
	1171	{
	1172	png_byte v;
	1173	int j;
	1174
	1175	v = (png_byte)((*sp >> sshift) & 0xf);
	1176	for (j = 0; j < png_pass_inc[pass]; j++)
	1177	{
	1178	*dp &= (png_byte)((0xf0f >> (4 - dshift)) & 0xff);
	1179	*dp \|= (png_byte)(v << dshift);
	1180	if (dshift == s_end)
	1181	{
	1182	dshift = s_start;
	1183	dp--;
	1184	}
	1185	else
	1186	dshift += s_inc;
	1187	}
	1188	if (sshift == s_end)
	1189	{
	1190	sshift = s_start;
	1191	sp--;
	1192	}
	1193	else
	1194	sshift += s_inc;
	1195	}
	1196	break;
	1197	}
	1198
	1199	default: // This is the place where the routine is modified
	1200	{
	1201	__int64 const4 = 0x0000000000FFFFFF;
	1202	// __int64 const5 = 0x000000FFFFFF0000; // unused...
	1203	__int64 const6 = 0x00000000000000FF;
	1204	png_bytep sptr, dp;
	1205	png_uint_32 i;
	1206	png_size_t pixel_bytes;
	1207	int width = row_info->width;
	1208
	1209	pixel_bytes = (row_info->pixel_depth >> 3);
	1210
	1211	sptr = row + (width - 1) * pixel_bytes;
	1212	dp = row + (final_width - 1) * pixel_bytes;
	1213	// New code by Nirav Chhatrapati - Intel Corporation
	1214	// sign fix by GRR
	1215	// NOTE: there is NO MMX code for 48-bit and 64-bit images
	1216
	1217	// use MMX routine if machine supports it
	1218	#if !defined(PNG_1_0_X)
	1219	if ((png_ptr->asm_flags & PNG_ASM_FLAG_MMX_READ_INTERLACE)
	1220	/* && mmx_supported */ )
	1221	#else
	1222	if (mmx_supported)
	1223	#endif
	1224	{
	1225	if (pixel_bytes == 3)
	1226	{
	1227	if (((pass == 0) \|\| (pass == 1)) && width)
	1228	{
	1229	_asm
	1230	{
	1231	mov esi, sptr
	1232	mov edi, dp
	1233	mov ecx, width
	1234	sub edi, 21 // (png_pass_inc[pass] - 1)*pixel_bytes
	1235	loop_pass0:
	1236	movd mm0, [esi] ; X X X X X v2 v1 v0
	1237	pand mm0, const4 ; 0 0 0 0 0 v2 v1 v0
	1238	movq mm1, mm0 ; 0 0 0 0 0 v2 v1 v0
	1239	psllq mm0, 16 ; 0 0 0 v2 v1 v0 0 0
	1240	movq mm2, mm0 ; 0 0 0 v2 v1 v0 0 0
	1241	psllq mm0, 24 ; v2 v1 v0 0 0 0 0 0
	1242	psrlq mm1, 8 ; 0 0 0 0 0 0 v2 v1
	1243	por mm0, mm2 ; v2 v1 v0 v2 v1 v0 0 0
	1244	por mm0, mm1 ; v2 v1 v0 v2 v1 v0 v2 v1
	1245	movq mm3, mm0 ; v2 v1 v0 v2 v1 v0 v2 v1
	1246	psllq mm0, 16 ; v0 v2 v1 v0 v2 v1 0 0
	1247	movq mm4, mm3 ; v2 v1 v0 v2 v1 v0 v2 v1
	1248	punpckhdq mm3, mm0 ; v0 v2 v1 v0 v2 v1 v0 v2
	1249	movq [edi+16] , mm4
	1250	psrlq mm0, 32 ; 0 0 0 0 v0 v2 v1 v0
	1251	movq [edi+8] , mm3
	1252	punpckldq mm0, mm4 ; v1 v0 v2 v1 v0 v2 v1 v0
	1253	sub esi, 3
	1254	movq [edi], mm0
	1255	sub edi, 24
	1256	//sub esi, 3
	1257	dec ecx
	1258	jnz loop_pass0
	1259	EMMS
	1260	}
	1261	}
	1262	else if (((pass == 2) \|\| (pass == 3)) && width)
	1263	{
	1264	_asm
	1265	{
	1266	mov esi, sptr
	1267	mov edi, dp
	1268	mov ecx, width
	1269	sub edi, 9 // (png_pass_inc[pass] - 1)*pixel_bytes
	1270	loop_pass2:
	1271	movd mm0, [esi] ; X X X X X v2 v1 v0
	1272	pand mm0, const4 ; 0 0 0 0 0 v2 v1 v0
	1273	movq mm1, mm0 ; 0 0 0 0 0 v2 v1 v0
	1274	psllq mm0, 16 ; 0 0 0 v2 v1 v0 0 0
	1275	movq mm2, mm0 ; 0 0 0 v2 v1 v0 0 0
	1276	psllq mm0, 24 ; v2 v1 v0 0 0 0 0 0
	1277	psrlq mm1, 8 ; 0 0 0 0 0 0 v2 v1
	1278	por mm0, mm2 ; v2 v1 v0 v2 v1 v0 0 0
	1279	por mm0, mm1 ; v2 v1 v0 v2 v1 v0 v2 v1
	1280	movq [edi+4], mm0 ; move to memory
	1281	psrlq mm0, 16 ; 0 0 v2 v1 v0 v2 v1 v0
	1282	movd [edi], mm0 ; move to memory
	1283	sub esi, 3
	1284	sub edi, 12
	1285	dec ecx
	1286	jnz loop_pass2
	1287	EMMS
	1288	}
	1289	}
	1290	else if (width) /* && ((pass == 4) \|\| (pass == 5)) */
	1291	{
	1292	int width_mmx = ((width >> 1) << 1) - 8;
	1293	if (width_mmx < 0)
	1294	width_mmx = 0;
	1295	width -= width_mmx; // 8 or 9 pix, 24 or 27 bytes
	1296	if (width_mmx)
	1297	{
	1298	_asm
	1299	{
	1300	mov esi, sptr
	1301	mov edi, dp
	1302	mov ecx, width_mmx
	1303	sub esi, 3
	1304	sub edi, 9
	1305	loop_pass4:
	1306	movq mm0, [esi] ; X X v2 v1 v0 v5 v4 v3
	1307	movq mm7, mm0 ; X X v2 v1 v0 v5 v4 v3
	1308	movq mm6, mm0 ; X X v2 v1 v0 v5 v4 v3
	1309	psllq mm0, 24 ; v1 v0 v5 v4 v3 0 0 0
	1310	pand mm7, const4 ; 0 0 0 0 0 v5 v4 v3
	1311	psrlq mm6, 24 ; 0 0 0 X X v2 v1 v0
	1312	por mm0, mm7 ; v1 v0 v5 v4 v3 v5 v4 v3
	1313	movq mm5, mm6 ; 0 0 0 X X v2 v1 v0
	1314	psllq mm6, 8 ; 0 0 X X v2 v1 v0 0
	1315	movq [edi], mm0 ; move quad to memory
	1316	psrlq mm5, 16 ; 0 0 0 0 0 X X v2
	1317	pand mm5, const6 ; 0 0 0 0 0 0 0 v2
	1318	por mm6, mm5 ; 0 0 X X v2 v1 v0 v2
	1319	movd [edi+8], mm6 ; move double to memory
	1320	sub esi, 6
	1321	sub edi, 12
	1322	sub ecx, 2
	1323	jnz loop_pass4
	1324	EMMS
	1325	}
	1326	}
	1327
	1328	sptr -= width_mmx*3;
	1329	dp -= width_mmx*6;
	1330	for (i = width; i; i--)
	1331	{
	1332	png_byte v[8];
	1333	int j;
	1334
	1335	png_memcpy(v, sptr, 3);
	1336	for (j = 0; j < png_pass_inc[pass]; j++)
	1337	{
	1338	png_memcpy(dp, v, 3);
	1339	dp -= 3;
	1340	}
	1341	sptr -= 3;
	1342	}
	1343	}
	1344	} /* end of pixel_bytes == 3 */
	1345
	1346	else if (pixel_bytes == 1)
	1347	{
	1348	if (((pass == 0) \|\| (pass == 1)) && width)
	1349	{
	1350	int width_mmx = ((width >> 2) << 2);
	1351	width -= width_mmx;
	1352	if (width_mmx)
	1353	{
	1354	_asm
	1355	{
	1356	mov esi, sptr
	1357	mov edi, dp
	1358	mov ecx, width_mmx
	1359	sub edi, 31
	1360	sub esi, 3
	1361	loop1_pass0:
	1362	movd mm0, [esi] ; X X X X v0 v1 v2 v3
	1363	movq mm1, mm0 ; X X X X v0 v1 v2 v3
	1364	punpcklbw mm0, mm0 ; v0 v0 v1 v1 v2 v2 v3 v3
	1365	movq mm2, mm0 ; v0 v0 v1 v1 v2 v2 v3 v3
	1366	punpcklwd mm0, mm0 ; v2 v2 v2 v2 v3 v3 v3 v3
	1367	movq mm3, mm0 ; v2 v2 v2 v2 v3 v3 v3 v3
	1368	punpckldq mm0, mm0 ; v3 v3 v3 v3 v3 v3 v3 v3
	1369	punpckhdq mm3, mm3 ; v2 v2 v2 v2 v2 v2 v2 v2
	1370	movq [edi], mm0 ; move to memory v3
	1371	punpckhwd mm2, mm2 ; v0 v0 v0 v0 v1 v1 v1 v1
	1372	movq [edi+8], mm3 ; move to memory v2
	1373	movq mm4, mm2 ; v0 v0 v0 v0 v1 v1 v1 v1
	1374	punpckldq mm2, mm2 ; v1 v1 v1 v1 v1 v1 v1 v1
	1375	punpckhdq mm4, mm4 ; v0 v0 v0 v0 v0 v0 v0 v0
	1376	movq [edi+16], mm2 ; move to memory v1
	1377	movq [edi+24], mm4 ; move to memory v0
	1378	sub esi, 4
	1379	sub edi, 32
	1380	sub ecx, 4
	1381	jnz loop1_pass0
	1382	EMMS
	1383	}
	1384	}
	1385
	1386	sptr -= width_mmx;
	1387	dp -= width_mmx*8;
	1388	for (i = width; i; i--)
	1389	{
	1390	int j;
	1391
	1392	/* I simplified this part in version 1.0.4e
	1393	* here and in several other instances where
	1394	* pixel_bytes == 1 -- GR-P
	1395	*
	1396	* Original code:
	1397	*
	1398	* png_byte v[8];
	1399	* png_memcpy(v, sptr, pixel_bytes);
	1400	* for (j = 0; j < png_pass_inc[pass]; j++)
	1401	* {
	1402	* png_memcpy(dp, v, pixel_bytes);
	1403	* dp -= pixel_bytes;
	1404	* }
	1405	* sptr -= pixel_bytes;
	1406	*
	1407	* Replacement code is in the next three lines:
	1408	*/
	1409
	1410	for (j = 0; j < png_pass_inc[pass]; j++)
	1411	dp-- = sptr;
	1412	sptr--;
	1413	}
	1414	}
	1415	else if (((pass == 2) \|\| (pass == 3)) && width)
	1416	{
	1417	int width_mmx = ((width >> 2) << 2);
	1418	width -= width_mmx;
	1419	if (width_mmx)
	1420	{
	1421	_asm
	1422	{
	1423	mov esi, sptr
	1424	mov edi, dp
	1425	mov ecx, width_mmx
	1426	sub edi, 15
	1427	sub esi, 3
	1428	loop1_pass2:
	1429	movd mm0, [esi] ; X X X X v0 v1 v2 v3
	1430	punpcklbw mm0, mm0 ; v0 v0 v1 v1 v2 v2 v3 v3
	1431	movq mm1, mm0 ; v0 v0 v1 v1 v2 v2 v3 v3
	1432	punpcklwd mm0, mm0 ; v2 v2 v2 v2 v3 v3 v3 v3
	1433	punpckhwd mm1, mm1 ; v0 v0 v0 v0 v1 v1 v1 v1
	1434	movq [edi], mm0 ; move to memory v2 and v3
	1435	sub esi, 4
	1436	movq [edi+8], mm1 ; move to memory v1 and v0
	1437	sub edi, 16
	1438	sub ecx, 4
	1439	jnz loop1_pass2
	1440	EMMS
	1441	}
	1442	}
	1443
	1444	sptr -= width_mmx;
	1445	dp -= width_mmx*4;
	1446	for (i = width; i; i--)
	1447	{
	1448	int j;
	1449
	1450	for (j = 0; j < png_pass_inc[pass]; j++)
	1451	{
	1452	dp-- = sptr;
	1453	}
	1454	sptr --;
	1455	}
	1456	}
	1457	else if (width) /* && ((pass == 4) \|\| (pass == 5))) */
	1458	{
	1459	int width_mmx = ((width >> 3) << 3);
	1460	width -= width_mmx;
	1461	if (width_mmx)
	1462	{
	1463	_asm
	1464	{
	1465	mov esi, sptr
	1466	mov edi, dp
	1467	mov ecx, width_mmx
	1468	sub edi, 15
	1469	sub esi, 7
	1470	loop1_pass4:
	1471	movq mm0, [esi] ; v0 v1 v2 v3 v4 v5 v6 v7
	1472	movq mm1, mm0 ; v0 v1 v2 v3 v4 v5 v6 v7
	1473	punpcklbw mm0, mm0 ; v4 v4 v5 v5 v6 v6 v7 v7
	1474	//movq mm1, mm0 ; v0 v0 v1 v1 v2 v2 v3 v3
	1475	punpckhbw mm1, mm1 ;v0 v0 v1 v1 v2 v2 v3 v3
	1476	movq [edi+8], mm1 ; move to memory v0 v1 v2 and v3
	1477	sub esi, 8
	1478	movq [edi], mm0 ; move to memory v4 v5 v6 and v7
	1479	//sub esi, 4
	1480	sub edi, 16
	1481	sub ecx, 8
	1482	jnz loop1_pass4
	1483	EMMS
	1484	}
	1485	}
	1486
	1487	sptr -= width_mmx;
	1488	dp -= width_mmx*2;
	1489	for (i = width; i; i--)
	1490	{
	1491	int j;
	1492
	1493	for (j = 0; j < png_pass_inc[pass]; j++)
	1494	{
	1495	dp-- = sptr;
	1496	}
	1497	sptr --;
	1498	}
	1499	}
	1500	} /* end of pixel_bytes == 1 */
	1501
	1502	else if (pixel_bytes == 2)
	1503	{
	1504	if (((pass == 0) \|\| (pass == 1)) && width)
	1505	{
	1506	int width_mmx = ((width >> 1) << 1);
	1507	width -= width_mmx;
	1508	if (width_mmx)
	1509	{
	1510	_asm
	1511	{
	1512	mov esi, sptr
	1513	mov edi, dp
	1514	mov ecx, width_mmx
	1515	sub esi, 2
	1516	sub edi, 30
	1517	loop2_pass0:
	1518	movd mm0, [esi] ; X X X X v1 v0 v3 v2
	1519	punpcklwd mm0, mm0 ; v1 v0 v1 v0 v3 v2 v3 v2
	1520	movq mm1, mm0 ; v1 v0 v1 v0 v3 v2 v3 v2
	1521	punpckldq mm0, mm0 ; v3 v2 v3 v2 v3 v2 v3 v2
	1522	punpckhdq mm1, mm1 ; v1 v0 v1 v0 v1 v0 v1 v0
	1523	movq [edi], mm0
	1524	movq [edi + 8], mm0
	1525	movq [edi + 16], mm1
	1526	movq [edi + 24], mm1
	1527	sub esi, 4
	1528	sub edi, 32
	1529	sub ecx, 2
	1530	jnz loop2_pass0
	1531	EMMS
	1532	}
	1533	}
	1534
	1535	sptr -= (width_mmx*2 - 2); // sign fixed
	1536	dp -= (width_mmx*16 - 2); // sign fixed
	1537	for (i = width; i; i--)
	1538	{
	1539	png_byte v[8];
	1540	int j;
	1541	sptr -= 2;
	1542	png_memcpy(v, sptr, 2);
	1543	for (j = 0; j < png_pass_inc[pass]; j++)
	1544	{
	1545	dp -= 2;
	1546	png_memcpy(dp, v, 2);
	1547	}
	1548	}
	1549	}
	1550	else if (((pass == 2) \|\| (pass == 3)) && width)
	1551	{
	1552	int width_mmx = ((width >> 1) << 1) ;
	1553	width -= width_mmx;
	1554	if (width_mmx)
	1555	{
	1556	_asm
	1557	{
	1558	mov esi, sptr
	1559	mov edi, dp
	1560	mov ecx, width_mmx
	1561	sub esi, 2
	1562	sub edi, 14
	1563	loop2_pass2:
	1564	movd mm0, [esi] ; X X X X v1 v0 v3 v2
	1565	punpcklwd mm0, mm0 ; v1 v0 v1 v0 v3 v2 v3 v2
	1566	movq mm1, mm0 ; v1 v0 v1 v0 v3 v2 v3 v2
	1567	punpckldq mm0, mm0 ; v3 v2 v3 v2 v3 v2 v3 v2
	1568	punpckhdq mm1, mm1 ; v1 v0 v1 v0 v1 v0 v1 v0
	1569	movq [edi], mm0
	1570	sub esi, 4
	1571	movq [edi + 8], mm1
	1572	//sub esi, 4
	1573	sub edi, 16
	1574	sub ecx, 2
	1575	jnz loop2_pass2
	1576	EMMS
	1577	}
	1578	}
	1579
	1580	sptr -= (width_mmx*2 - 2); // sign fixed
	1581	dp -= (width_mmx*8 - 2); // sign fixed
	1582	for (i = width; i; i--)
	1583	{
	1584	png_byte v[8];
	1585	int j;
	1586	sptr -= 2;
	1587	png_memcpy(v, sptr, 2);
	1588	for (j = 0; j < png_pass_inc[pass]; j++)
	1589	{
	1590	dp -= 2;
	1591	png_memcpy(dp, v, 2);
	1592	}
	1593	}
	1594	}
	1595	else if (width) // pass == 4 or 5
	1596	{
	1597	int width_mmx = ((width >> 1) << 1) ;
	1598	width -= width_mmx;
	1599	if (width_mmx)
	1600	{
	1601	_asm
	1602	{
	1603	mov esi, sptr
	1604	mov edi, dp
	1605	mov ecx, width_mmx
	1606	sub esi, 2
	1607	sub edi, 6
	1608	loop2_pass4:
	1609	movd mm0, [esi] ; X X X X v1 v0 v3 v2
	1610	punpcklwd mm0, mm0 ; v1 v0 v1 v0 v3 v2 v3 v2
	1611	sub esi, 4
	1612	movq [edi], mm0
	1613	sub edi, 8
	1614	sub ecx, 2
	1615	jnz loop2_pass4
	1616	EMMS
	1617	}
	1618	}
	1619
	1620	sptr -= (width_mmx*2 - 2); // sign fixed
	1621	dp -= (width_mmx*4 - 2); // sign fixed
	1622	for (i = width; i; i--)
	1623	{
	1624	png_byte v[8];
	1625	int j;
	1626	sptr -= 2;
	1627	png_memcpy(v, sptr, 2);
	1628	for (j = 0; j < png_pass_inc[pass]; j++)
	1629	{
	1630	dp -= 2;
	1631	png_memcpy(dp, v, 2);
	1632	}
	1633	}
	1634	}
	1635	} /* end of pixel_bytes == 2 */
	1636
	1637	else if (pixel_bytes == 4)
	1638	{
	1639	if (((pass == 0) \|\| (pass == 1)) && width)
	1640	{
	1641	int width_mmx = ((width >> 1) << 1) ;
	1642	width -= width_mmx;
	1643	if (width_mmx)
	1644	{
	1645	_asm
	1646	{
	1647	mov esi, sptr
	1648	mov edi, dp
	1649	mov ecx, width_mmx
	1650	sub esi, 4
	1651	sub edi, 60
	1652	loop4_pass0:
	1653	movq mm0, [esi] ; v3 v2 v1 v0 v7 v6 v5 v4
	1654	movq mm1, mm0 ; v3 v2 v1 v0 v7 v6 v5 v4
	1655	punpckldq mm0, mm0 ; v7 v6 v5 v4 v7 v6 v5 v4
	1656	punpckhdq mm1, mm1 ; v3 v2 v1 v0 v3 v2 v1 v0
	1657	movq [edi], mm0
	1658	movq [edi + 8], mm0
	1659	movq [edi + 16], mm0
	1660	movq [edi + 24], mm0
	1661	movq [edi+32], mm1
	1662	movq [edi + 40], mm1
	1663	movq [edi+ 48], mm1
	1664	sub esi, 8
	1665	movq [edi + 56], mm1
	1666	sub edi, 64
	1667	sub ecx, 2
	1668	jnz loop4_pass0
	1669	EMMS
	1670	}
	1671	}
	1672
	1673	sptr -= (width_mmx*4 - 4); // sign fixed
	1674	dp -= (width_mmx*32 - 4); // sign fixed
	1675	for (i = width; i; i--)
	1676	{
	1677	png_byte v[8];
	1678	int j;
	1679	sptr -= 4;
	1680	png_memcpy(v, sptr, 4);
	1681	for (j = 0; j < png_pass_inc[pass]; j++)
	1682	{
	1683	dp -= 4;
	1684	png_memcpy(dp, v, 4);
	1685	}
	1686	}
	1687	}
	1688	else if (((pass == 2) \|\| (pass == 3)) && width)
	1689	{
	1690	int width_mmx = ((width >> 1) << 1) ;
	1691	width -= width_mmx;
	1692	if (width_mmx)
	1693	{
	1694	_asm
	1695	{
	1696	mov esi, sptr
	1697	mov edi, dp
	1698	mov ecx, width_mmx
	1699	sub esi, 4
	1700	sub edi, 28
	1701	loop4_pass2:
	1702	movq mm0, [esi] ; v3 v2 v1 v0 v7 v6 v5 v4
	1703	movq mm1, mm0 ; v3 v2 v1 v0 v7 v6 v5 v4
	1704	punpckldq mm0, mm0 ; v7 v6 v5 v4 v7 v6 v5 v4
	1705	punpckhdq mm1, mm1 ; v3 v2 v1 v0 v3 v2 v1 v0
	1706	movq [edi], mm0
	1707	movq [edi + 8], mm0
	1708	movq [edi+16], mm1
	1709	movq [edi + 24], mm1
	1710	sub esi, 8
	1711	sub edi, 32
	1712	sub ecx, 2
	1713	jnz loop4_pass2
	1714	EMMS
	1715	}
	1716	}
	1717
	1718	sptr -= (width_mmx*4 - 4); // sign fixed
	1719	dp -= (width_mmx*16 - 4); // sign fixed
	1720	for (i = width; i; i--)
	1721	{
	1722	png_byte v[8];
	1723	int j;
	1724	sptr -= 4;
	1725	png_memcpy(v, sptr, 4);
	1726	for (j = 0; j < png_pass_inc[pass]; j++)
	1727	{
	1728	dp -= 4;
	1729	png_memcpy(dp, v, 4);
	1730	}
	1731	}
	1732	}
	1733	else if (width) // pass == 4 or 5
	1734	{
	1735	int width_mmx = ((width >> 1) << 1) ;
	1736	width -= width_mmx;
	1737	if (width_mmx)
	1738	{
	1739	_asm
	1740	{
	1741	mov esi, sptr
	1742	mov edi, dp
	1743	mov ecx, width_mmx
	1744	sub esi, 4
	1745	sub edi, 12
	1746	loop4_pass4:
	1747	movq mm0, [esi] ; v3 v2 v1 v0 v7 v6 v5 v4
	1748	movq mm1, mm0 ; v3 v2 v1 v0 v7 v6 v5 v4
	1749	punpckldq mm0, mm0 ; v7 v6 v5 v4 v7 v6 v5 v4
	1750	punpckhdq mm1, mm1 ; v3 v2 v1 v0 v3 v2 v1 v0
	1751	movq [edi], mm0
	1752	sub esi, 8
	1753	movq [edi + 8], mm1
	1754	sub edi, 16
	1755	sub ecx, 2
	1756	jnz loop4_pass4
	1757	EMMS
	1758	}
	1759	}
	1760
	1761	sptr -= (width_mmx*4 - 4); // sign fixed
	1762	dp -= (width_mmx*8 - 4); // sign fixed
	1763	for (i = width; i; i--)
	1764	{
	1765	png_byte v[8];
	1766	int j;
	1767	sptr -= 4;
	1768	png_memcpy(v, sptr, 4);
	1769	for (j = 0; j < png_pass_inc[pass]; j++)
	1770	{
	1771	dp -= 4;
	1772	png_memcpy(dp, v, 4);
	1773	}
	1774	}
	1775	}
	1776
	1777	} /* end of pixel_bytes == 4 */
	1778
	1779	else if (pixel_bytes == 6)
	1780	{
	1781	for (i = width; i; i--)
	1782	{
	1783	png_byte v[8];
	1784	int j;
	1785	png_memcpy(v, sptr, 6);
	1786	for (j = 0; j < png_pass_inc[pass]; j++)
	1787	{
	1788	png_memcpy(dp, v, 6);
	1789	dp -= 6;
	1790	}
	1791	sptr -= 6;
	1792	}
	1793	} /* end of pixel_bytes == 6 */
	1794
	1795	else
	1796	{
	1797	for (i = width; i; i--)
	1798	{
	1799	png_byte v[8];
	1800	int j;
	1801	png_memcpy(v, sptr, pixel_bytes);
	1802	for (j = 0; j < png_pass_inc[pass]; j++)
	1803	{
	1804	png_memcpy(dp, v, pixel_bytes);
	1805	dp -= pixel_bytes;
	1806	}
	1807	sptr-= pixel_bytes;
	1808	}
	1809	}
	1810	} /* end of mmx_supported */
	1811
	1812	else /* MMX not supported: use modified C code - takes advantage
	1813	* of inlining of memcpy for a constant */
	1814	{
	1815	if (pixel_bytes == 1)
	1816	{
	1817	for (i = width; i; i--)
	1818	{
	1819	int j;
	1820	for (j = 0; j < png_pass_inc[pass]; j++)
	1821	dp-- = sptr;
	1822	sptr--;
	1823	}
	1824	}
	1825	else if (pixel_bytes == 3)
	1826	{
	1827	for (i = width; i; i--)
	1828	{
	1829	png_byte v[8];
	1830	int j;
	1831	png_memcpy(v, sptr, pixel_bytes);
	1832	for (j = 0; j < png_pass_inc[pass]; j++)
	1833	{
	1834	png_memcpy(dp, v, pixel_bytes);
	1835	dp -= pixel_bytes;
	1836	}
	1837	sptr -= pixel_bytes;
	1838	}
	1839	}
	1840	else if (pixel_bytes == 2)
	1841	{
	1842	for (i = width; i; i--)
	1843	{
	1844	png_byte v[8];
	1845	int j;
	1846	png_memcpy(v, sptr, pixel_bytes);
	1847	for (j = 0; j < png_pass_inc[pass]; j++)
	1848	{
	1849	png_memcpy(dp, v, pixel_bytes);
	1850	dp -= pixel_bytes;
	1851	}
	1852	sptr -= pixel_bytes;
	1853	}
	1854	}
	1855	else if (pixel_bytes == 4)
	1856	{
	1857	for (i = width; i; i--)
	1858	{
	1859	png_byte v[8];
	1860	int j;
	1861	png_memcpy(v, sptr, pixel_bytes);
	1862	for (j = 0; j < png_pass_inc[pass]; j++)
	1863	{
	1864	png_memcpy(dp, v, pixel_bytes);
	1865	dp -= pixel_bytes;
	1866	}
	1867	sptr -= pixel_bytes;
	1868	}
	1869	}
	1870	else if (pixel_bytes == 6)
	1871	{
	1872	for (i = width; i; i--)
	1873	{
	1874	png_byte v[8];
	1875	int j;
	1876	png_memcpy(v, sptr, pixel_bytes);
	1877	for (j = 0; j < png_pass_inc[pass]; j++)
	1878	{
	1879	png_memcpy(dp, v, pixel_bytes);
	1880	dp -= pixel_bytes;
	1881	}
	1882	sptr -= pixel_bytes;
	1883	}
	1884	}
	1885	else
	1886	{
	1887	for (i = width; i; i--)
	1888	{
	1889	png_byte v[8];
	1890	int j;
	1891	png_memcpy(v, sptr, pixel_bytes);
	1892	for (j = 0; j < png_pass_inc[pass]; j++)
	1893	{
	1894	png_memcpy(dp, v, pixel_bytes);
	1895	dp -= pixel_bytes;
	1896	}
	1897	sptr -= pixel_bytes;
	1898	}
	1899	}
	1900
	1901	} /* end of MMX not supported */
	1902	break;
	1903	}
	1904	} /* end switch (row_info->pixel_depth) */
	1905
	1906	row_info->width = final_width;
	1907
	1908	row_info->rowbytes = PNG_ROWBYTES(row_info->pixel_depth,final_width);
	1909	}
	1910
	1911	}
	1912
	1913	#endif /* PNG_READ_INTERLACING_SUPPORTED */
	1914
	1915
	1916	// These variables are utilized in the functions below. They are declared
	1917	// globally here to ensure alignment on 8-byte boundaries.
	1918
	1919	union uAll {
	1920	__int64 use;
	1921	double align;
	1922	} LBCarryMask = {0x0101010101010101},
	1923	HBClearMask = {0x7f7f7f7f7f7f7f7f},
	1924	ActiveMask, ActiveMask2, ActiveMaskEnd, ShiftBpp, ShiftRem;
	1925
	1926
	1927	// Optimized code for PNG Average filter decoder
	1928	void /* PRIVATE */
	1929	png_read_filter_row_mmx_avg(png_row_infop row_info, png_bytep row
	1930	, png_bytep prev_row)
	1931	{
	1932	int bpp;
	1933	png_uint_32 FullLength;
	1934	png_uint_32 MMXLength;
	1935	//png_uint_32 len;
	1936	int diff;
	1937
	1938	bpp = (row_info->pixel_depth + 7) >> 3; // Get # bytes per pixel
	1939	FullLength = row_info->rowbytes; // # of bytes to filter
	1940	_asm {
	1941	// Init address pointers and offset
	1942	mov edi, row // edi ==> Avg(x)
	1943	xor ebx, ebx // ebx ==> x
	1944	mov edx, edi
	1945	mov esi, prev_row // esi ==> Prior(x)
	1946	sub edx, bpp // edx ==> Raw(x-bpp)
	1947
	1948	xor eax, eax
	1949	// Compute the Raw value for the first bpp bytes
	1950	// Raw(x) = Avg(x) + (Prior(x)/2)
	1951	davgrlp:
	1952	mov al, [esi + ebx] // Load al with Prior(x)
	1953	inc ebx
	1954	shr al, 1 // divide by 2
	1955	add al, [edi+ebx-1] // Add Avg(x); -1 to offset inc ebx
	1956	cmp ebx, bpp
	1957	mov [edi+ebx-1], al // Write back Raw(x);
	1958	// mov does not affect flags; -1 to offset inc ebx
	1959	jb davgrlp
	1960	// get # of bytes to alignment
	1961	mov diff, edi // take start of row
	1962	add diff, ebx // add bpp
	1963	add diff, 0xf // add 7 + 8 to incr past alignment boundary
	1964	and diff, 0xfffffff8 // mask to alignment boundary
	1965	sub diff, edi // subtract from start ==> value ebx at alignment
	1966	jz davggo
	1967	// fix alignment
	1968	// Compute the Raw value for the bytes upto the alignment boundary
	1969	// Raw(x) = Avg(x) + ((Raw(x-bpp) + Prior(x))/2)
	1970	xor ecx, ecx
	1971	davglp1:
	1972	xor eax, eax
	1973	mov cl, [esi + ebx] // load cl with Prior(x)
	1974	mov al, [edx + ebx] // load al with Raw(x-bpp)
	1975	add ax, cx
	1976	inc ebx
	1977	shr ax, 1 // divide by 2
	1978	add al, [edi+ebx-1] // Add Avg(x); -1 to offset inc ebx
	1979	cmp ebx, diff // Check if at alignment boundary
	1980	mov [edi+ebx-1], al // Write back Raw(x);
	1981	// mov does not affect flags; -1 to offset inc ebx
	1982	jb davglp1 // Repeat until at alignment boundary
	1983	davggo:
	1984	mov eax, FullLength
	1985	mov ecx, eax
	1986	sub eax, ebx // subtract alignment fix
	1987	and eax, 0x00000007 // calc bytes over mult of 8
	1988	sub ecx, eax // drop over bytes from original length
	1989	mov MMXLength, ecx
	1990	} // end _asm block
	1991	// Now do the math for the rest of the row
	1992	switch ( bpp )
	1993	{
	1994	case 3:
	1995	{
	1996	ActiveMask.use = 0x0000000000ffffff;
	1997	ShiftBpp.use = 24; // == 3 * 8
	1998	ShiftRem.use = 40; // == 64 - 24
	1999	_asm {
	2000	// Re-init address pointers and offset
	2001	movq mm7, ActiveMask
	2002	mov ebx, diff // ebx ==> x = offset to alignment boundary
	2003	movq mm5, LBCarryMask
	2004	mov edi, row // edi ==> Avg(x)
	2005	movq mm4, HBClearMask
	2006	mov esi, prev_row // esi ==> Prior(x)
	2007	// PRIME the pump (load the first Raw(x-bpp) data set
	2008	movq mm2, [edi + ebx - 8] // Load previous aligned 8 bytes
	2009	// (we correct position in loop below)
	2010	davg3lp:
	2011	movq mm0, [edi + ebx] // Load mm0 with Avg(x)
	2012	// Add (Prev_row/2) to Average
	2013	movq mm3, mm5
	2014	psrlq mm2, ShiftRem // Correct position Raw(x-bpp) data
	2015	movq mm1, [esi + ebx] // Load mm1 with Prior(x)
	2016	movq mm6, mm7
	2017	pand mm3, mm1 // get lsb for each prev_row byte
	2018	psrlq mm1, 1 // divide prev_row bytes by 2
	2019	pand mm1, mm4 // clear invalid bit 7 of each byte
	2020	paddb mm0, mm1 // add (Prev_row/2) to Avg for each byte
	2021	// Add 1st active group (Raw(x-bpp)/2) to Average with LBCarry
	2022	movq mm1, mm3 // now use mm1 for getting LBCarrys
	2023	pand mm1, mm2 // get LBCarrys for each byte where both
	2024	// lsb's were == 1 (Only valid for active group)
	2025	psrlq mm2, 1 // divide raw bytes by 2
	2026	pand mm2, mm4 // clear invalid bit 7 of each byte
	2027	paddb mm2, mm1 // add LBCarrys to (Raw(x-bpp)/2) for each byte
	2028	pand mm2, mm6 // Leave only Active Group 1 bytes to add to Avg
	2029	paddb mm0, mm2 // add (Raw/2) + LBCarrys to Avg for each Active
	2030	// byte
	2031	// Add 2nd active group (Raw(x-bpp)/2) to Average with LBCarry
	2032	psllq mm6, ShiftBpp // shift the mm6 mask to cover bytes 3-5
	2033	movq mm2, mm0 // mov updated Raws to mm2
	2034	psllq mm2, ShiftBpp // shift data to position correctly
	2035	movq mm1, mm3 // now use mm1 for getting LBCarrys
	2036	pand mm1, mm2 // get LBCarrys for each byte where both
	2037	// lsb's were == 1 (Only valid for active group)
	2038	psrlq mm2, 1 // divide raw bytes by 2
	2039	pand mm2, mm4 // clear invalid bit 7 of each byte
	2040	paddb mm2, mm1 // add LBCarrys to (Raw(x-bpp)/2) for each byte
	2041	pand mm2, mm6 // Leave only Active Group 2 bytes to add to Avg
	2042	paddb mm0, mm2 // add (Raw/2) + LBCarrys to Avg for each Active
	2043	// byte
	2044
	2045	// Add 3rd active group (Raw(x-bpp)/2) to Average with LBCarry
	2046	psllq mm6, ShiftBpp // shift the mm6 mask to cover the last two
	2047	// bytes
	2048	movq mm2, mm0 // mov updated Raws to mm2
	2049	psllq mm2, ShiftBpp // shift data to position correctly
	2050	// Data only needs to be shifted once here to
	2051	// get the correct x-bpp offset.
	2052	movq mm1, mm3 // now use mm1 for getting LBCarrys
	2053	pand mm1, mm2 // get LBCarrys for each byte where both
	2054	// lsb's were == 1 (Only valid for active group)
	2055	psrlq mm2, 1 // divide raw bytes by 2
	2056	pand mm2, mm4 // clear invalid bit 7 of each byte
	2057	paddb mm2, mm1 // add LBCarrys to (Raw(x-bpp)/2) for each byte
	2058	pand mm2, mm6 // Leave only Active Group 2 bytes to add to Avg
	2059	add ebx, 8
	2060	paddb mm0, mm2 // add (Raw/2) + LBCarrys to Avg for each Active
	2061	// byte
	2062
	2063	// Now ready to write back to memory
	2064	movq [edi + ebx - 8], mm0
	2065	// Move updated Raw(x) to use as Raw(x-bpp) for next loop
	2066	cmp ebx, MMXLength
	2067	movq mm2, mm0 // mov updated Raw(x) to mm2
	2068	jb davg3lp
	2069	} // end _asm block
	2070	}
	2071	break;
	2072
	2073	case 6:
	2074	case 4:
	2075	case 7:
	2076	case 5:
	2077	{
	2078	ActiveMask.use = 0xffffffffffffffff; // use shift below to clear
	2079	// appropriate inactive bytes
	2080	ShiftBpp.use = bpp << 3;
	2081	ShiftRem.use = 64 - ShiftBpp.use;
	2082	_asm {
	2083	movq mm4, HBClearMask
	2084	// Re-init address pointers and offset
	2085	mov ebx, diff // ebx ==> x = offset to alignment boundary
	2086	// Load ActiveMask and clear all bytes except for 1st active group
	2087	movq mm7, ActiveMask
	2088	mov edi, row // edi ==> Avg(x)
	2089	psrlq mm7, ShiftRem
	2090	mov esi, prev_row // esi ==> Prior(x)
	2091	movq mm6, mm7
	2092	movq mm5, LBCarryMask
	2093	psllq mm6, ShiftBpp // Create mask for 2nd active group
	2094	// PRIME the pump (load the first Raw(x-bpp) data set
	2095	movq mm2, [edi + ebx - 8] // Load previous aligned 8 bytes
	2096	// (we correct position in loop below)
	2097	davg4lp:
	2098	movq mm0, [edi + ebx]
	2099	psrlq mm2, ShiftRem // shift data to position correctly
	2100	movq mm1, [esi + ebx]
	2101	// Add (Prev_row/2) to Average
	2102	movq mm3, mm5
	2103	pand mm3, mm1 // get lsb for each prev_row byte
	2104	psrlq mm1, 1 // divide prev_row bytes by 2
	2105	pand mm1, mm4 // clear invalid bit 7 of each byte
	2106	paddb mm0, mm1 // add (Prev_row/2) to Avg for each byte
	2107	// Add 1st active group (Raw(x-bpp)/2) to Average with LBCarry
	2108	movq mm1, mm3 // now use mm1 for getting LBCarrys
	2109	pand mm1, mm2 // get LBCarrys for each byte where both
	2110	// lsb's were == 1 (Only valid for active group)
	2111	psrlq mm2, 1 // divide raw bytes by 2
	2112	pand mm2, mm4 // clear invalid bit 7 of each byte
	2113	paddb mm2, mm1 // add LBCarrys to (Raw(x-bpp)/2) for each byte
	2114	pand mm2, mm7 // Leave only Active Group 1 bytes to add to Avg
	2115	paddb mm0, mm2 // add (Raw/2) + LBCarrys to Avg for each Active
	2116	// byte
	2117	// Add 2nd active group (Raw(x-bpp)/2) to Average with LBCarry
	2118	movq mm2, mm0 // mov updated Raws to mm2
	2119	psllq mm2, ShiftBpp // shift data to position correctly
	2120	add ebx, 8
	2121	movq mm1, mm3 // now use mm1 for getting LBCarrys
	2122	pand mm1, mm2 // get LBCarrys for each byte where both
	2123	// lsb's were == 1 (Only valid for active group)
	2124	psrlq mm2, 1 // divide raw bytes by 2
	2125	pand mm2, mm4 // clear invalid bit 7 of each byte
	2126	paddb mm2, mm1 // add LBCarrys to (Raw(x-bpp)/2) for each byte
	2127	pand mm2, mm6 // Leave only Active Group 2 bytes to add to Avg
	2128	paddb mm0, mm2 // add (Raw/2) + LBCarrys to Avg for each Active
	2129	// byte
	2130	cmp ebx, MMXLength
	2131	// Now ready to write back to memory
	2132	movq [edi + ebx - 8], mm0
	2133	// Prep Raw(x-bpp) for next loop
	2134	movq mm2, mm0 // mov updated Raws to mm2
	2135	jb davg4lp
	2136	} // end _asm block
	2137	}
	2138	break;
	2139	case 2:
	2140	{
	2141	ActiveMask.use = 0x000000000000ffff;
	2142	ShiftBpp.use = 16; // == 2 * 8 [BUGFIX]
	2143	ShiftRem.use = 48; // == 64 - 16 [BUGFIX]
	2144	_asm {
	2145	// Load ActiveMask
	2146	movq mm7, ActiveMask
	2147	// Re-init address pointers and offset
	2148	mov ebx, diff // ebx ==> x = offset to alignment boundary
	2149	movq mm5, LBCarryMask
	2150	mov edi, row // edi ==> Avg(x)
	2151	movq mm4, HBClearMask
	2152	mov esi, prev_row // esi ==> Prior(x)
	2153	// PRIME the pump (load the first Raw(x-bpp) data set
	2154	movq mm2, [edi + ebx - 8] // Load previous aligned 8 bytes
	2155	// (we correct position in loop below)
	2156	davg2lp:
	2157	movq mm0, [edi + ebx]
	2158	psrlq mm2, ShiftRem // shift data to position correctly [BUGFIX]
	2159	movq mm1, [esi + ebx]
	2160	// Add (Prev_row/2) to Average
	2161	movq mm3, mm5
	2162	pand mm3, mm1 // get lsb for each prev_row byte
	2163	psrlq mm1, 1 // divide prev_row bytes by 2
	2164	pand mm1, mm4 // clear invalid bit 7 of each byte
	2165	movq mm6, mm7
	2166	paddb mm0, mm1 // add (Prev_row/2) to Avg for each byte
	2167	// Add 1st active group (Raw(x-bpp)/2) to Average with LBCarry
	2168	movq mm1, mm3 // now use mm1 for getting LBCarrys
	2169	pand mm1, mm2 // get LBCarrys for each byte where both
	2170	// lsb's were == 1 (Only valid for active group)
	2171	psrlq mm2, 1 // divide raw bytes by 2
	2172	pand mm2, mm4 // clear invalid bit 7 of each byte
	2173	paddb mm2, mm1 // add LBCarrys to (Raw(x-bpp)/2) for each byte
	2174	pand mm2, mm6 // Leave only Active Group 1 bytes to add to Avg
	2175	paddb mm0, mm2 // add (Raw/2) + LBCarrys to Avg for each Active byte
	2176	// Add 2nd active group (Raw(x-bpp)/2) to Average with LBCarry
	2177	psllq mm6, ShiftBpp // shift the mm6 mask to cover bytes 2 & 3
	2178	movq mm2, mm0 // mov updated Raws to mm2
	2179	psllq mm2, ShiftBpp // shift data to position correctly
	2180	movq mm1, mm3 // now use mm1 for getting LBCarrys
	2181	pand mm1, mm2 // get LBCarrys for each byte where both
	2182	// lsb's were == 1 (Only valid for active group)
	2183	psrlq mm2, 1 // divide raw bytes by 2
	2184	pand mm2, mm4 // clear invalid bit 7 of each byte
	2185	paddb mm2, mm1 // add LBCarrys to (Raw(x-bpp)/2) for each byte
	2186	pand mm2, mm6 // Leave only Active Group 2 bytes to add to Avg
	2187	paddb mm0, mm2 // add (Raw/2) + LBCarrys to Avg for each Active byte
	2188
	2189	// Add rdd active group (Raw(x-bpp)/2) to Average with LBCarry
	2190	psllq mm6, ShiftBpp // shift the mm6 mask to cover bytes 4 & 5
	2191	movq mm2, mm0 // mov updated Raws to mm2
	2192	psllq mm2, ShiftBpp // shift data to position correctly
	2193	// Data only needs to be shifted once here to
	2194	// get the correct x-bpp offset.
	2195	movq mm1, mm3 // now use mm1 for getting LBCarrys
	2196	pand mm1, mm2 // get LBCarrys for each byte where both
	2197	// lsb's were == 1 (Only valid for active group)
	2198	psrlq mm2, 1 // divide raw bytes by 2
	2199	pand mm2, mm4 // clear invalid bit 7 of each byte
	2200	paddb mm2, mm1 // add LBCarrys to (Raw(x-bpp)/2) for each byte
	2201	pand mm2, mm6 // Leave only Active Group 2 bytes to add to Avg
	2202	paddb mm0, mm2 // add (Raw/2) + LBCarrys to Avg for each Active byte
	2203
	2204	// Add 4th active group (Raw(x-bpp)/2) to Average with LBCarry
	2205	psllq mm6, ShiftBpp // shift the mm6 mask to cover bytes 6 & 7
	2206	movq mm2, mm0 // mov updated Raws to mm2
	2207	psllq mm2, ShiftBpp // shift data to position correctly
	2208	// Data only needs to be shifted once here to
	2209	// get the correct x-bpp offset.
	2210	add ebx, 8
	2211	movq mm1, mm3 // now use mm1 for getting LBCarrys
	2212	pand mm1, mm2 // get LBCarrys for each byte where both
	2213	// lsb's were == 1 (Only valid for active group)
	2214	psrlq mm2, 1 // divide raw bytes by 2
	2215	pand mm2, mm4 // clear invalid bit 7 of each byte
	2216	paddb mm2, mm1 // add LBCarrys to (Raw(x-bpp)/2) for each byte
	2217	pand mm2, mm6 // Leave only Active Group 2 bytes to add to Avg
	2218	paddb mm0, mm2 // add (Raw/2) + LBCarrys to Avg for each Active byte
	2219
	2220	cmp ebx, MMXLength
	2221	// Now ready to write back to memory
	2222	movq [edi + ebx - 8], mm0
	2223	// Prep Raw(x-bpp) for next loop
	2224	movq mm2, mm0 // mov updated Raws to mm2
	2225	jb davg2lp
	2226	} // end _asm block
	2227	}
	2228	break;
	2229
	2230	case 1: // bpp == 1
	2231	{
	2232	_asm {
	2233	// Re-init address pointers and offset
	2234	mov ebx, diff // ebx ==> x = offset to alignment boundary
	2235	mov edi, row // edi ==> Avg(x)
	2236	cmp ebx, FullLength // Test if offset at end of array
	2237	jnb davg1end
	2238	// Do Paeth decode for remaining bytes
	2239	mov esi, prev_row // esi ==> Prior(x)
	2240	mov edx, edi
	2241	xor ecx, ecx // zero ecx before using cl & cx in loop below
	2242	sub edx, bpp // edx ==> Raw(x-bpp)
	2243	davg1lp:
	2244	// Raw(x) = Avg(x) + ((Raw(x-bpp) + Prior(x))/2)
	2245	xor eax, eax
	2246	mov cl, [esi + ebx] // load cl with Prior(x)
	2247	mov al, [edx + ebx] // load al with Raw(x-bpp)
	2248	add ax, cx
	2249	inc ebx
	2250	shr ax, 1 // divide by 2
	2251	add al, [edi+ebx-1] // Add Avg(x); -1 to offset inc ebx
	2252	cmp ebx, FullLength // Check if at end of array
	2253	mov [edi+ebx-1], al // Write back Raw(x);
	2254	// mov does not affect flags; -1 to offset inc ebx
	2255	jb davg1lp
	2256	davg1end:
	2257	} // end _asm block
	2258	}
	2259	return;
	2260
	2261	case 8: // bpp == 8
	2262	{
	2263	_asm {
	2264	// Re-init address pointers and offset
	2265	mov ebx, diff // ebx ==> x = offset to alignment boundary
	2266	movq mm5, LBCarryMask
	2267	mov edi, row // edi ==> Avg(x)
	2268	movq mm4, HBClearMask
	2269	mov esi, prev_row // esi ==> Prior(x)
	2270	// PRIME the pump (load the first Raw(x-bpp) data set
	2271	movq mm2, [edi + ebx - 8] // Load previous aligned 8 bytes
	2272	// (NO NEED to correct position in loop below)
	2273	davg8lp:
	2274	movq mm0, [edi + ebx]
	2275	movq mm3, mm5
	2276	movq mm1, [esi + ebx]
	2277	add ebx, 8
	2278	pand mm3, mm1 // get lsb for each prev_row byte
	2279	psrlq mm1, 1 // divide prev_row bytes by 2
	2280	pand mm3, mm2 // get LBCarrys for each byte where both
	2281	// lsb's were == 1
	2282	psrlq mm2, 1 // divide raw bytes by 2
	2283	pand mm1, mm4 // clear invalid bit 7 of each byte
	2284	paddb mm0, mm3 // add LBCarrys to Avg for each byte
	2285	pand mm2, mm4 // clear invalid bit 7 of each byte
	2286	paddb mm0, mm1 // add (Prev_row/2) to Avg for each byte
	2287	paddb mm0, mm2 // add (Raw/2) to Avg for each byte
	2288	cmp ebx, MMXLength
	2289	movq [edi + ebx - 8], mm0
	2290	movq mm2, mm0 // reuse as Raw(x-bpp)
	2291	jb davg8lp
	2292	} // end _asm block
	2293	}
	2294	break;
	2295	default: // bpp greater than 8
	2296	{
	2297	_asm {
	2298	movq mm5, LBCarryMask
	2299	// Re-init address pointers and offset
	2300	mov ebx, diff // ebx ==> x = offset to alignment boundary
	2301	mov edi, row // edi ==> Avg(x)
	2302	movq mm4, HBClearMask
	2303	mov edx, edi
	2304	mov esi, prev_row // esi ==> Prior(x)
	2305	sub edx, bpp // edx ==> Raw(x-bpp)
	2306	davgAlp:
	2307	movq mm0, [edi + ebx]
	2308	movq mm3, mm5
	2309	movq mm1, [esi + ebx]
	2310	pand mm3, mm1 // get lsb for each prev_row byte
	2311	movq mm2, [edx + ebx]
	2312	psrlq mm1, 1 // divide prev_row bytes by 2
	2313	pand mm3, mm2 // get LBCarrys for each byte where both
	2314	// lsb's were == 1
	2315	psrlq mm2, 1 // divide raw bytes by 2
	2316	pand mm1, mm4 // clear invalid bit 7 of each byte
	2317	paddb mm0, mm3 // add LBCarrys to Avg for each byte
	2318	pand mm2, mm4 // clear invalid bit 7 of each byte
	2319	paddb mm0, mm1 // add (Prev_row/2) to Avg for each byte
	2320	add ebx, 8
	2321	paddb mm0, mm2 // add (Raw/2) to Avg for each byte
	2322	cmp ebx, MMXLength
	2323	movq [edi + ebx - 8], mm0
	2324	jb davgAlp
	2325	} // end _asm block
	2326	}
	2327	break;
	2328	} // end switch ( bpp )
	2329
	2330	_asm {
	2331	// MMX acceleration complete now do clean-up
	2332	// Check if any remaining bytes left to decode
	2333	mov ebx, MMXLength // ebx ==> x = offset bytes remaining after MMX
	2334	mov edi, row // edi ==> Avg(x)
	2335	cmp ebx, FullLength // Test if offset at end of array
	2336	jnb davgend
	2337	// Do Paeth decode for remaining bytes
	2338	mov esi, prev_row // esi ==> Prior(x)
	2339	mov edx, edi
	2340	xor ecx, ecx // zero ecx before using cl & cx in loop below
	2341	sub edx, bpp // edx ==> Raw(x-bpp)
	2342	davglp2:
	2343	// Raw(x) = Avg(x) + ((Raw(x-bpp) + Prior(x))/2)
	2344	xor eax, eax
	2345	mov cl, [esi + ebx] // load cl with Prior(x)
	2346	mov al, [edx + ebx] // load al with Raw(x-bpp)
	2347	add ax, cx
	2348	inc ebx
	2349	shr ax, 1 // divide by 2
	2350	add al, [edi+ebx-1] // Add Avg(x); -1 to offset inc ebx
	2351	cmp ebx, FullLength // Check if at end of array
	2352	mov [edi+ebx-1], al // Write back Raw(x);
	2353	// mov does not affect flags; -1 to offset inc ebx
	2354	jb davglp2
	2355	davgend:
	2356	emms // End MMX instructions; prep for possible FP instrs.
	2357	} // end _asm block
	2358	}
	2359
	2360	// Optimized code for PNG Paeth filter decoder
	2361	void /* PRIVATE */
	2362	png_read_filter_row_mmx_paeth(png_row_infop row_info, png_bytep row,
	2363	png_bytep prev_row)
	2364	{
	2365	png_uint_32 FullLength;
	2366	png_uint_32 MMXLength;
	2367	//png_uint_32 len;
	2368	int bpp;
	2369	int diff;
	2370	//int ptemp;
	2371	int patemp, pbtemp, pctemp;
	2372
	2373	bpp = (row_info->pixel_depth + 7) >> 3; // Get # bytes per pixel
	2374	FullLength = row_info->rowbytes; // # of bytes to filter
	2375	_asm
	2376	{
	2377	xor ebx, ebx // ebx ==> x offset
	2378	mov edi, row
	2379	xor edx, edx // edx ==> x-bpp offset
	2380	mov esi, prev_row
	2381	xor eax, eax
	2382
	2383	// Compute the Raw value for the first bpp bytes
	2384	// Note: the formula works out to be always
	2385	// Paeth(x) = Raw(x) + Prior(x) where x < bpp
	2386	dpthrlp:
	2387	mov al, [edi + ebx]
	2388	add al, [esi + ebx]
	2389	inc ebx
	2390	cmp ebx, bpp
	2391	mov [edi + ebx - 1], al
	2392	jb dpthrlp
	2393	// get # of bytes to alignment
	2394	mov diff, edi // take start of row
	2395	add diff, ebx // add bpp
	2396	xor ecx, ecx
	2397	add diff, 0xf // add 7 + 8 to incr past alignment boundary
	2398	and diff, 0xfffffff8 // mask to alignment boundary
	2399	sub diff, edi // subtract from start ==> value ebx at alignment
	2400	jz dpthgo
	2401	// fix alignment
	2402	dpthlp1:
	2403	xor eax, eax
	2404	// pav = p - a = (a + b - c) - a = b - c
	2405	mov al, [esi + ebx] // load Prior(x) into al
	2406	mov cl, [esi + edx] // load Prior(x-bpp) into cl
	2407	sub eax, ecx // subtract Prior(x-bpp)
	2408	mov patemp, eax // Save pav for later use
	2409	xor eax, eax
	2410	// pbv = p - b = (a + b - c) - b = a - c
	2411	mov al, [edi + edx] // load Raw(x-bpp) into al
	2412	sub eax, ecx // subtract Prior(x-bpp)
	2413	mov ecx, eax
	2414	// pcv = p - c = (a + b - c) -c = (a - c) + (b - c) = pav + pbv
	2415	add eax, patemp // pcv = pav + pbv
	2416	// pc = abs(pcv)
	2417	test eax, 0x80000000
	2418	jz dpthpca
	2419	neg eax // reverse sign of neg values
	2420	dpthpca:
	2421	mov pctemp, eax // save pc for later use
	2422	// pb = abs(pbv)
	2423	test ecx, 0x80000000
	2424	jz dpthpba
	2425	neg ecx // reverse sign of neg values
	2426	dpthpba:
	2427	mov pbtemp, ecx // save pb for later use
	2428	// pa = abs(pav)
	2429	mov eax, patemp
	2430	test eax, 0x80000000
	2431	jz dpthpaa
	2432	neg eax // reverse sign of neg values
	2433	dpthpaa:
	2434	mov patemp, eax // save pa for later use
	2435	// test if pa <= pb
	2436	cmp eax, ecx
	2437	jna dpthabb
	2438	// pa > pb; now test if pb <= pc
	2439	cmp ecx, pctemp
	2440	jna dpthbbc
	2441	// pb > pc; Raw(x) = Paeth(x) + Prior(x-bpp)
	2442	mov cl, [esi + edx] // load Prior(x-bpp) into cl
	2443	jmp dpthpaeth
	2444	dpthbbc:
	2445	// pb <= pc; Raw(x) = Paeth(x) + Prior(x)
	2446	mov cl, [esi + ebx] // load Prior(x) into cl
	2447	jmp dpthpaeth
	2448	dpthabb:
	2449	// pa <= pb; now test if pa <= pc
	2450	cmp eax, pctemp
	2451	jna dpthabc
	2452	// pa > pc; Raw(x) = Paeth(x) + Prior(x-bpp)
	2453	mov cl, [esi + edx] // load Prior(x-bpp) into cl
	2454	jmp dpthpaeth
	2455	dpthabc:
	2456	// pa <= pc; Raw(x) = Paeth(x) + Raw(x-bpp)
	2457	mov cl, [edi + edx] // load Raw(x-bpp) into cl
	2458	dpthpaeth:
	2459	inc ebx
	2460	inc edx
	2461	// Raw(x) = (Paeth(x) + Paeth_Predictor( a, b, c )) mod 256
	2462	add [edi + ebx - 1], cl
	2463	cmp ebx, diff
	2464	jb dpthlp1
	2465	dpthgo:
	2466	mov ecx, FullLength
	2467	mov eax, ecx
	2468	sub eax, ebx // subtract alignment fix
	2469	and eax, 0x00000007 // calc bytes over mult of 8
	2470	sub ecx, eax // drop over bytes from original length
	2471	mov MMXLength, ecx
	2472	} // end _asm block
	2473	// Now do the math for the rest of the row
	2474	switch ( bpp )
	2475	{
	2476	case 3:
	2477	{
	2478	ActiveMask.use = 0x0000000000ffffff;
	2479	ActiveMaskEnd.use = 0xffff000000000000;
	2480	ShiftBpp.use = 24; // == bpp(3) * 8
	2481	ShiftRem.use = 40; // == 64 - 24
	2482	_asm
	2483	{
	2484	mov ebx, diff
	2485	mov edi, row
	2486	mov esi, prev_row
	2487	pxor mm0, mm0
	2488	// PRIME the pump (load the first Raw(x-bpp) data set
	2489	movq mm1, [edi+ebx-8]
	2490	dpth3lp:
	2491	psrlq mm1, ShiftRem // shift last 3 bytes to 1st 3 bytes
	2492	movq mm2, [esi + ebx] // load b=Prior(x)
	2493	punpcklbw mm1, mm0 // Unpack High bytes of a
	2494	movq mm3, [esi+ebx-8] // Prep c=Prior(x-bpp) bytes
	2495	punpcklbw mm2, mm0 // Unpack High bytes of b
	2496	psrlq mm3, ShiftRem // shift last 3 bytes to 1st 3 bytes
	2497	// pav = p - a = (a + b - c) - a = b - c
	2498	movq mm4, mm2
	2499	punpcklbw mm3, mm0 // Unpack High bytes of c
	2500	// pbv = p - b = (a + b - c) - b = a - c
	2501	movq mm5, mm1
	2502	psubw mm4, mm3
	2503	pxor mm7, mm7
	2504	// pcv = p - c = (a + b - c) -c = (a - c) + (b - c) = pav + pbv
	2505	movq mm6, mm4
	2506	psubw mm5, mm3
	2507
	2508	// pa = abs(p-a) = abs(pav)
	2509	// pb = abs(p-b) = abs(pbv)
	2510	// pc = abs(p-c) = abs(pcv)
	2511	pcmpgtw mm0, mm4 // Create mask pav bytes < 0
	2512	paddw mm6, mm5
	2513	pand mm0, mm4 // Only pav bytes < 0 in mm7
	2514	pcmpgtw mm7, mm5 // Create mask pbv bytes < 0
	2515	psubw mm4, mm0
	2516	pand mm7, mm5 // Only pbv bytes < 0 in mm0
	2517	psubw mm4, mm0
	2518	psubw mm5, mm7
	2519	pxor mm0, mm0
	2520	pcmpgtw mm0, mm6 // Create mask pcv bytes < 0
	2521	pand mm0, mm6 // Only pav bytes < 0 in mm7
	2522	psubw mm5, mm7
	2523	psubw mm6, mm0
	2524	// test pa <= pb
	2525	movq mm7, mm4
	2526	psubw mm6, mm0
	2527	pcmpgtw mm7, mm5 // pa > pb?
	2528	movq mm0, mm7
	2529	// use mm7 mask to merge pa & pb
	2530	pand mm5, mm7
	2531	// use mm0 mask copy to merge a & b
	2532	pand mm2, mm0
	2533	pandn mm7, mm4
	2534	pandn mm0, mm1
	2535	paddw mm7, mm5
	2536	paddw mm0, mm2
	2537	// test ((pa <= pb)? pa:pb) <= pc
	2538	pcmpgtw mm7, mm6 // pab > pc?
	2539	pxor mm1, mm1
	2540	pand mm3, mm7
	2541	pandn mm7, mm0
	2542	paddw mm7, mm3
	2543	pxor mm0, mm0
	2544	packuswb mm7, mm1
	2545	movq mm3, [esi + ebx] // load c=Prior(x-bpp)
	2546	pand mm7, ActiveMask
	2547	movq mm2, mm3 // load b=Prior(x) step 1
	2548	paddb mm7, [edi + ebx] // add Paeth predictor with Raw(x)
	2549	punpcklbw mm3, mm0 // Unpack High bytes of c
	2550	movq [edi + ebx], mm7 // write back updated value
	2551	movq mm1, mm7 // Now mm1 will be used as Raw(x-bpp)
	2552	// Now do Paeth for 2nd set of bytes (3-5)
	2553	psrlq mm2, ShiftBpp // load b=Prior(x) step 2
	2554	punpcklbw mm1, mm0 // Unpack High bytes of a
	2555	pxor mm7, mm7
	2556	punpcklbw mm2, mm0 // Unpack High bytes of b
	2557	// pbv = p - b = (a + b - c) - b = a - c
	2558	movq mm5, mm1
	2559	// pav = p - a = (a + b - c) - a = b - c
	2560	movq mm4, mm2
	2561	psubw mm5, mm3
	2562	psubw mm4, mm3
	2563	// pcv = p - c = (a + b - c) -c = (a - c) + (b - c) =
	2564	// pav + pbv = pbv + pav
	2565	movq mm6, mm5
	2566	paddw mm6, mm4
	2567
	2568	// pa = abs(p-a) = abs(pav)
	2569	// pb = abs(p-b) = abs(pbv)
	2570	// pc = abs(p-c) = abs(pcv)
	2571	pcmpgtw mm0, mm5 // Create mask pbv bytes < 0
	2572	pcmpgtw mm7, mm4 // Create mask pav bytes < 0
	2573	pand mm0, mm5 // Only pbv bytes < 0 in mm0
	2574	pand mm7, mm4 // Only pav bytes < 0 in mm7
	2575	psubw mm5, mm0
	2576	psubw mm4, mm7
	2577	psubw mm5, mm0
	2578	psubw mm4, mm7
	2579	pxor mm0, mm0
	2580	pcmpgtw mm0, mm6 // Create mask pcv bytes < 0
	2581	pand mm0, mm6 // Only pav bytes < 0 in mm7
	2582	psubw mm6, mm0
	2583	// test pa <= pb
	2584	movq mm7, mm4
	2585	psubw mm6, mm0
	2586	pcmpgtw mm7, mm5 // pa > pb?
	2587	movq mm0, mm7
	2588	// use mm7 mask to merge pa & pb
	2589	pand mm5, mm7
	2590	// use mm0 mask copy to merge a & b
	2591	pand mm2, mm0
	2592	pandn mm7, mm4
	2593	pandn mm0, mm1
	2594	paddw mm7, mm5
	2595	paddw mm0, mm2
	2596	// test ((pa <= pb)? pa:pb) <= pc
	2597	pcmpgtw mm7, mm6 // pab > pc?
	2598	movq mm2, [esi + ebx] // load b=Prior(x)
	2599	pand mm3, mm7
	2600	pandn mm7, mm0
	2601	pxor mm1, mm1
	2602	paddw mm7, mm3
	2603	pxor mm0, mm0
	2604	packuswb mm7, mm1
	2605	movq mm3, mm2 // load c=Prior(x-bpp) step 1
	2606	pand mm7, ActiveMask
	2607	punpckhbw mm2, mm0 // Unpack High bytes of b
	2608	psllq mm7, ShiftBpp // Shift bytes to 2nd group of 3 bytes
	2609	// pav = p - a = (a + b - c) - a = b - c
	2610	movq mm4, mm2
	2611	paddb mm7, [edi + ebx] // add Paeth predictor with Raw(x)
	2612	psllq mm3, ShiftBpp // load c=Prior(x-bpp) step 2
	2613	movq [edi + ebx], mm7 // write back updated value
	2614	movq mm1, mm7
	2615	punpckhbw mm3, mm0 // Unpack High bytes of c
	2616	psllq mm1, ShiftBpp // Shift bytes
	2617	// Now mm1 will be used as Raw(x-bpp)
	2618	// Now do Paeth for 3rd, and final, set of bytes (6-7)
	2619	pxor mm7, mm7
	2620	punpckhbw mm1, mm0 // Unpack High bytes of a
	2621	psubw mm4, mm3
	2622	// pbv = p - b = (a + b - c) - b = a - c
	2623	movq mm5, mm1
	2624	// pcv = p - c = (a + b - c) -c = (a - c) + (b - c) = pav + pbv
	2625	movq mm6, mm4
	2626	psubw mm5, mm3
	2627	pxor mm0, mm0
	2628	paddw mm6, mm5
	2629
	2630	// pa = abs(p-a) = abs(pav)
	2631	// pb = abs(p-b) = abs(pbv)
	2632	// pc = abs(p-c) = abs(pcv)
	2633	pcmpgtw mm0, mm4 // Create mask pav bytes < 0
	2634	pcmpgtw mm7, mm5 // Create mask pbv bytes < 0
	2635	pand mm0, mm4 // Only pav bytes < 0 in mm7
	2636	pand mm7, mm5 // Only pbv bytes < 0 in mm0
	2637	psubw mm4, mm0
	2638	psubw mm5, mm7
	2639	psubw mm4, mm0
	2640	psubw mm5, mm7
	2641	pxor mm0, mm0
	2642	pcmpgtw mm0, mm6 // Create mask pcv bytes < 0
	2643	pand mm0, mm6 // Only pav bytes < 0 in mm7
	2644	psubw mm6, mm0
	2645	// test pa <= pb
	2646	movq mm7, mm4
	2647	psubw mm6, mm0
	2648	pcmpgtw mm7, mm5 // pa > pb?
	2649	movq mm0, mm7
	2650	// use mm0 mask copy to merge a & b
	2651	pand mm2, mm0
	2652	// use mm7 mask to merge pa & pb
	2653	pand mm5, mm7
	2654	pandn mm0, mm1
	2655	pandn mm7, mm4
	2656	paddw mm0, mm2
	2657	paddw mm7, mm5
	2658	// test ((pa <= pb)? pa:pb) <= pc
	2659	pcmpgtw mm7, mm6 // pab > pc?
	2660	pand mm3, mm7
	2661	pandn mm7, mm0
	2662	paddw mm7, mm3
	2663	pxor mm1, mm1
	2664	packuswb mm1, mm7
	2665	// Step ebx to next set of 8 bytes and repeat loop til done
	2666	add ebx, 8
	2667	pand mm1, ActiveMaskEnd
	2668	paddb mm1, [edi + ebx - 8] // add Paeth predictor with Raw(x)
	2669
	2670	cmp ebx, MMXLength
	2671	pxor mm0, mm0 // pxor does not affect flags
	2672	movq [edi + ebx - 8], mm1 // write back updated value
	2673	// mm1 will be used as Raw(x-bpp) next loop
	2674	// mm3 ready to be used as Prior(x-bpp) next loop
	2675	jb dpth3lp
	2676	} // end _asm block
	2677	}
	2678	break;
	2679
	2680	case 6:
	2681	case 7:
	2682	case 5:
	2683	{
	2684	ActiveMask.use = 0x00000000ffffffff;
	2685	ActiveMask2.use = 0xffffffff00000000;
	2686	ShiftBpp.use = bpp << 3; // == bpp * 8
	2687	ShiftRem.use = 64 - ShiftBpp.use;
	2688	_asm
	2689	{
	2690	mov ebx, diff
	2691	mov edi, row
	2692	mov esi, prev_row
	2693	// PRIME the pump (load the first Raw(x-bpp) data set
	2694	movq mm1, [edi+ebx-8]
	2695	pxor mm0, mm0
	2696	dpth6lp:
	2697	// Must shift to position Raw(x-bpp) data
	2698	psrlq mm1, ShiftRem
	2699	// Do first set of 4 bytes
	2700	movq mm3, [esi+ebx-8] // read c=Prior(x-bpp) bytes
	2701	punpcklbw mm1, mm0 // Unpack Low bytes of a
	2702	movq mm2, [esi + ebx] // load b=Prior(x)
	2703	punpcklbw mm2, mm0 // Unpack Low bytes of b
	2704	// Must shift to position Prior(x-bpp) data
	2705	psrlq mm3, ShiftRem
	2706	// pav = p - a = (a + b - c) - a = b - c
	2707	movq mm4, mm2
	2708	punpcklbw mm3, mm0 // Unpack Low bytes of c
	2709	// pbv = p - b = (a + b - c) - b = a - c
	2710	movq mm5, mm1
	2711	psubw mm4, mm3
	2712	pxor mm7, mm7
	2713	// pcv = p - c = (a + b - c) -c = (a - c) + (b - c) = pav + pbv
	2714	movq mm6, mm4
	2715	psubw mm5, mm3
	2716	// pa = abs(p-a) = abs(pav)
	2717	// pb = abs(p-b) = abs(pbv)
	2718	// pc = abs(p-c) = abs(pcv)
	2719	pcmpgtw mm0, mm4 // Create mask pav bytes < 0
	2720	paddw mm6, mm5
	2721	pand mm0, mm4 // Only pav bytes < 0 in mm7
	2722	pcmpgtw mm7, mm5 // Create mask pbv bytes < 0
	2723	psubw mm4, mm0
	2724	pand mm7, mm5 // Only pbv bytes < 0 in mm0
	2725	psubw mm4, mm0
	2726	psubw mm5, mm7
	2727	pxor mm0, mm0
	2728	pcmpgtw mm0, mm6 // Create mask pcv bytes < 0
	2729	pand mm0, mm6 // Only pav bytes < 0 in mm7
	2730	psubw mm5, mm7
	2731	psubw mm6, mm0
	2732	// test pa <= pb
	2733	movq mm7, mm4
	2734	psubw mm6, mm0
	2735	pcmpgtw mm7, mm5 // pa > pb?
	2736	movq mm0, mm7
	2737	// use mm7 mask to merge pa & pb
	2738	pand mm5, mm7
	2739	// use mm0 mask copy to merge a & b
	2740	pand mm2, mm0
	2741	pandn mm7, mm4
	2742	pandn mm0, mm1
	2743	paddw mm7, mm5
	2744	paddw mm0, mm2
	2745	// test ((pa <= pb)? pa:pb) <= pc
	2746	pcmpgtw mm7, mm6 // pab > pc?
	2747	pxor mm1, mm1
	2748	pand mm3, mm7
	2749	pandn mm7, mm0
	2750	paddw mm7, mm3
	2751	pxor mm0, mm0
	2752	packuswb mm7, mm1
	2753	movq mm3, [esi + ebx - 8] // load c=Prior(x-bpp)
	2754	pand mm7, ActiveMask
	2755	psrlq mm3, ShiftRem
	2756	movq mm2, [esi + ebx] // load b=Prior(x) step 1
	2757	paddb mm7, [edi + ebx] // add Paeth predictor with Raw(x)
	2758	movq mm6, mm2
	2759	movq [edi + ebx], mm7 // write back updated value
	2760	movq mm1, [edi+ebx-8]
	2761	psllq mm6, ShiftBpp
	2762	movq mm5, mm7
	2763	psrlq mm1, ShiftRem
	2764	por mm3, mm6
	2765	psllq mm5, ShiftBpp
	2766	punpckhbw mm3, mm0 // Unpack High bytes of c
	2767	por mm1, mm5
	2768	// Do second set of 4 bytes
	2769	punpckhbw mm2, mm0 // Unpack High bytes of b
	2770	punpckhbw mm1, mm0 // Unpack High bytes of a
	2771	// pav = p - a = (a + b - c) - a = b - c
	2772	movq mm4, mm2
	2773	// pbv = p - b = (a + b - c) - b = a - c
	2774	movq mm5, mm1
	2775	psubw mm4, mm3
	2776	pxor mm7, mm7
	2777	// pcv = p - c = (a + b - c) -c = (a - c) + (b - c) = pav + pbv
	2778	movq mm6, mm4
	2779	psubw mm5, mm3
	2780	// pa = abs(p-a) = abs(pav)
	2781	// pb = abs(p-b) = abs(pbv)
	2782	// pc = abs(p-c) = abs(pcv)
	2783	pcmpgtw mm0, mm4 // Create mask pav bytes < 0
	2784	paddw mm6, mm5
	2785	pand mm0, mm4 // Only pav bytes < 0 in mm7
	2786	pcmpgtw mm7, mm5 // Create mask pbv bytes < 0
	2787	psubw mm4, mm0
	2788	pand mm7, mm5 // Only pbv bytes < 0 in mm0
	2789	psubw mm4, mm0
	2790	psubw mm5, mm7
	2791	pxor mm0, mm0
	2792	pcmpgtw mm0, mm6 // Create mask pcv bytes < 0
	2793	pand mm0, mm6 // Only pav bytes < 0 in mm7
	2794	psubw mm5, mm7
	2795	psubw mm6, mm0
	2796	// test pa <= pb
	2797	movq mm7, mm4
	2798	psubw mm6, mm0
	2799	pcmpgtw mm7, mm5 // pa > pb?
	2800	movq mm0, mm7
	2801	// use mm7 mask to merge pa & pb
	2802	pand mm5, mm7
	2803	// use mm0 mask copy to merge a & b
	2804	pand mm2, mm0
	2805	pandn mm7, mm4
	2806	pandn mm0, mm1
	2807	paddw mm7, mm5
	2808	paddw mm0, mm2
	2809	// test ((pa <= pb)? pa:pb) <= pc
	2810	pcmpgtw mm7, mm6 // pab > pc?
	2811	pxor mm1, mm1
	2812	pand mm3, mm7
	2813	pandn mm7, mm0
	2814	pxor mm1, mm1
	2815	paddw mm7, mm3
	2816	pxor mm0, mm0
	2817	// Step ex to next set of 8 bytes and repeat loop til done
	2818	add ebx, 8
	2819	packuswb mm1, mm7
	2820	paddb mm1, [edi + ebx - 8] // add Paeth predictor with Raw(x)
	2821	cmp ebx, MMXLength
	2822	movq [edi + ebx - 8], mm1 // write back updated value
	2823	// mm1 will be used as Raw(x-bpp) next loop
	2824	jb dpth6lp
	2825	} // end _asm block
	2826	}
	2827	break;
	2828
	2829	case 4:
	2830	{
	2831	ActiveMask.use = 0x00000000ffffffff;
	2832	_asm {
	2833	mov ebx, diff
	2834	mov edi, row
	2835	mov esi, prev_row
	2836	pxor mm0, mm0
	2837	// PRIME the pump (load the first Raw(x-bpp) data set
	2838	movq mm1, [edi+ebx-8] // Only time should need to read
	2839	// a=Raw(x-bpp) bytes
	2840	dpth4lp:
	2841	// Do first set of 4 bytes
	2842	movq mm3, [esi+ebx-8] // read c=Prior(x-bpp) bytes
	2843	punpckhbw mm1, mm0 // Unpack Low bytes of a
	2844	movq mm2, [esi + ebx] // load b=Prior(x)
	2845	punpcklbw mm2, mm0 // Unpack High bytes of b
	2846	// pav = p - a = (a + b - c) - a = b - c
	2847	movq mm4, mm2
	2848	punpckhbw mm3, mm0 // Unpack High bytes of c
	2849	// pbv = p - b = (a + b - c) - b = a - c
	2850	movq mm5, mm1
	2851	psubw mm4, mm3
	2852	pxor mm7, mm7
	2853	// pcv = p - c = (a + b - c) -c = (a - c) + (b - c) = pav + pbv
	2854	movq mm6, mm4
	2855	psubw mm5, mm3
	2856	// pa = abs(p-a) = abs(pav)
	2857	// pb = abs(p-b) = abs(pbv)
	2858	// pc = abs(p-c) = abs(pcv)
	2859	pcmpgtw mm0, mm4 // Create mask pav bytes < 0
	2860	paddw mm6, mm5
	2861	pand mm0, mm4 // Only pav bytes < 0 in mm7
	2862	pcmpgtw mm7, mm5 // Create mask pbv bytes < 0
	2863	psubw mm4, mm0
	2864	pand mm7, mm5 // Only pbv bytes < 0 in mm0
	2865	psubw mm4, mm0
	2866	psubw mm5, mm7
	2867	pxor mm0, mm0
	2868	pcmpgtw mm0, mm6 // Create mask pcv bytes < 0
	2869	pand mm0, mm6 // Only pav bytes < 0 in mm7
	2870	psubw mm5, mm7
	2871	psubw mm6, mm0
	2872	// test pa <= pb
	2873	movq mm7, mm4
	2874	psubw mm6, mm0
	2875	pcmpgtw mm7, mm5 // pa > pb?
	2876	movq mm0, mm7
	2877	// use mm7 mask to merge pa & pb
	2878	pand mm5, mm7
	2879	// use mm0 mask copy to merge a & b
	2880	pand mm2, mm0
	2881	pandn mm7, mm4
	2882	pandn mm0, mm1
	2883	paddw mm7, mm5
	2884	paddw mm0, mm2
	2885	// test ((pa <= pb)? pa:pb) <= pc
	2886	pcmpgtw mm7, mm6 // pab > pc?
	2887	pxor mm1, mm1
	2888	pand mm3, mm7
	2889	pandn mm7, mm0
	2890	paddw mm7, mm3
	2891	pxor mm0, mm0
	2892	packuswb mm7, mm1
	2893	movq mm3, [esi + ebx] // load c=Prior(x-bpp)
	2894	pand mm7, ActiveMask
	2895	movq mm2, mm3 // load b=Prior(x) step 1
	2896	paddb mm7, [edi + ebx] // add Paeth predictor with Raw(x)
	2897	punpcklbw mm3, mm0 // Unpack High bytes of c
	2898	movq [edi + ebx], mm7 // write back updated value
	2899	movq mm1, mm7 // Now mm1 will be used as Raw(x-bpp)
	2900	// Do second set of 4 bytes
	2901	punpckhbw mm2, mm0 // Unpack Low bytes of b
	2902	punpcklbw mm1, mm0 // Unpack Low bytes of a
	2903	// pav = p - a = (a + b - c) - a = b - c
	2904	movq mm4, mm2
	2905	// pbv = p - b = (a + b - c) - b = a - c
	2906	movq mm5, mm1
	2907	psubw mm4, mm3
	2908	pxor mm7, mm7
	2909	// pcv = p - c = (a + b - c) -c = (a - c) + (b - c) = pav + pbv
	2910	movq mm6, mm4
	2911	psubw mm5, mm3
	2912	// pa = abs(p-a) = abs(pav)
	2913	// pb = abs(p-b) = abs(pbv)
	2914	// pc = abs(p-c) = abs(pcv)
	2915	pcmpgtw mm0, mm4 // Create mask pav bytes < 0
	2916	paddw mm6, mm5
	2917	pand mm0, mm4 // Only pav bytes < 0 in mm7
	2918	pcmpgtw mm7, mm5 // Create mask pbv bytes < 0
	2919	psubw mm4, mm0
	2920	pand mm7, mm5 // Only pbv bytes < 0 in mm0
	2921	psubw mm4, mm0
	2922	psubw mm5, mm7
	2923	pxor mm0, mm0
	2924	pcmpgtw mm0, mm6 // Create mask pcv bytes < 0
	2925	pand mm0, mm6 // Only pav bytes < 0 in mm7
	2926	psubw mm5, mm7
	2927	psubw mm6, mm0
	2928	// test pa <= pb
	2929	movq mm7, mm4
	2930	psubw mm6, mm0
	2931	pcmpgtw mm7, mm5 // pa > pb?
	2932	movq mm0, mm7
	2933	// use mm7 mask to merge pa & pb
	2934	pand mm5, mm7
	2935	// use mm0 mask copy to merge a & b
	2936	pand mm2, mm0
	2937	pandn mm7, mm4
	2938	pandn mm0, mm1
	2939	paddw mm7, mm5
	2940	paddw mm0, mm2
	2941	// test ((pa <= pb)? pa:pb) <= pc
	2942	pcmpgtw mm7, mm6 // pab > pc?
	2943	pxor mm1, mm1
	2944	pand mm3, mm7
	2945	pandn mm7, mm0
	2946	pxor mm1, mm1
	2947	paddw mm7, mm3
	2948	pxor mm0, mm0
	2949	// Step ex to next set of 8 bytes and repeat loop til done
	2950	add ebx, 8
	2951	packuswb mm1, mm7
	2952	paddb mm1, [edi + ebx - 8] // add Paeth predictor with Raw(x)
	2953	cmp ebx, MMXLength
	2954	movq [edi + ebx - 8], mm1 // write back updated value
	2955	// mm1 will be used as Raw(x-bpp) next loop
	2956	jb dpth4lp
	2957	} // end _asm block
	2958	}
	2959	break;
	2960	case 8: // bpp == 8
	2961	{
	2962	ActiveMask.use = 0x00000000ffffffff;
	2963	_asm {
	2964	mov ebx, diff
	2965	mov edi, row
	2966	mov esi, prev_row
	2967	pxor mm0, mm0
	2968	// PRIME the pump (load the first Raw(x-bpp) data set
	2969	movq mm1, [edi+ebx-8] // Only time should need to read
	2970	// a=Raw(x-bpp) bytes
	2971	dpth8lp:
	2972	// Do first set of 4 bytes
	2973	movq mm3, [esi+ebx-8] // read c=Prior(x-bpp) bytes
	2974	punpcklbw mm1, mm0 // Unpack Low bytes of a
	2975	movq mm2, [esi + ebx] // load b=Prior(x)
	2976	punpcklbw mm2, mm0 // Unpack Low bytes of b
	2977	// pav = p - a = (a + b - c) - a = b - c
	2978	movq mm4, mm2
	2979	punpcklbw mm3, mm0 // Unpack Low bytes of c
	2980	// pbv = p - b = (a + b - c) - b = a - c
	2981	movq mm5, mm1
	2982	psubw mm4, mm3
	2983	pxor mm7, mm7
	2984	// pcv = p - c = (a + b - c) -c = (a - c) + (b - c) = pav + pbv
	2985	movq mm6, mm4
	2986	psubw mm5, mm3
	2987	// pa = abs(p-a) = abs(pav)
	2988	// pb = abs(p-b) = abs(pbv)
	2989	// pc = abs(p-c) = abs(pcv)
	2990	pcmpgtw mm0, mm4 // Create mask pav bytes < 0
	2991	paddw mm6, mm5
	2992	pand mm0, mm4 // Only pav bytes < 0 in mm7
	2993	pcmpgtw mm7, mm5 // Create mask pbv bytes < 0
	2994	psubw mm4, mm0
	2995	pand mm7, mm5 // Only pbv bytes < 0 in mm0
	2996	psubw mm4, mm0
	2997	psubw mm5, mm7
	2998	pxor mm0, mm0
	2999	pcmpgtw mm0, mm6 // Create mask pcv bytes < 0
	3000	pand mm0, mm6 // Only pav bytes < 0 in mm7
	3001	psubw mm5, mm7
	3002	psubw mm6, mm0
	3003	// test pa <= pb
	3004	movq mm7, mm4
	3005	psubw mm6, mm0
	3006	pcmpgtw mm7, mm5 // pa > pb?
	3007	movq mm0, mm7
	3008	// use mm7 mask to merge pa & pb
	3009	pand mm5, mm7
	3010	// use mm0 mask copy to merge a & b
	3011	pand mm2, mm0
	3012	pandn mm7, mm4
	3013	pandn mm0, mm1
	3014	paddw mm7, mm5
	3015	paddw mm0, mm2
	3016	// test ((pa <= pb)? pa:pb) <= pc
	3017	pcmpgtw mm7, mm6 // pab > pc?
	3018	pxor mm1, mm1
	3019	pand mm3, mm7
	3020	pandn mm7, mm0
	3021	paddw mm7, mm3
	3022	pxor mm0, mm0
	3023	packuswb mm7, mm1
	3024	movq mm3, [esi+ebx-8] // read c=Prior(x-bpp) bytes
	3025	pand mm7, ActiveMask
	3026	movq mm2, [esi + ebx] // load b=Prior(x)
	3027	paddb mm7, [edi + ebx] // add Paeth predictor with Raw(x)
	3028	punpckhbw mm3, mm0 // Unpack High bytes of c
	3029	movq [edi + ebx], mm7 // write back updated value
	3030	movq mm1, [edi+ebx-8] // read a=Raw(x-bpp) bytes
	3031
	3032	// Do second set of 4 bytes
	3033	punpckhbw mm2, mm0 // Unpack High bytes of b
	3034	punpckhbw mm1, mm0 // Unpack High bytes of a
	3035	// pav = p - a = (a + b - c) - a = b - c
	3036	movq mm4, mm2
	3037	// pbv = p - b = (a + b - c) - b = a - c
	3038	movq mm5, mm1
	3039	psubw mm4, mm3
	3040	pxor mm7, mm7
	3041	// pcv = p - c = (a + b - c) -c = (a - c) + (b - c) = pav + pbv
	3042	movq mm6, mm4
	3043	psubw mm5, mm3
	3044	// pa = abs(p-a) = abs(pav)
	3045	// pb = abs(p-b) = abs(pbv)
	3046	// pc = abs(p-c) = abs(pcv)
	3047	pcmpgtw mm0, mm4 // Create mask pav bytes < 0
	3048	paddw mm6, mm5
	3049	pand mm0, mm4 // Only pav bytes < 0 in mm7
	3050	pcmpgtw mm7, mm5 // Create mask pbv bytes < 0
	3051	psubw mm4, mm0
	3052	pand mm7, mm5 // Only pbv bytes < 0 in mm0
	3053	psubw mm4, mm0
	3054	psubw mm5, mm7
	3055	pxor mm0, mm0
	3056	pcmpgtw mm0, mm6 // Create mask pcv bytes < 0
	3057	pand mm0, mm6 // Only pav bytes < 0 in mm7
	3058	psubw mm5, mm7
	3059	psubw mm6, mm0
	3060	// test pa <= pb
	3061	movq mm7, mm4
	3062	psubw mm6, mm0
	3063	pcmpgtw mm7, mm5 // pa > pb?
	3064	movq mm0, mm7
	3065	// use mm7 mask to merge pa & pb
	3066	pand mm5, mm7
	3067	// use mm0 mask copy to merge a & b
	3068	pand mm2, mm0
	3069	pandn mm7, mm4
	3070	pandn mm0, mm1
	3071	paddw mm7, mm5
	3072	paddw mm0, mm2
	3073	// test ((pa <= pb)? pa:pb) <= pc
	3074	pcmpgtw mm7, mm6 // pab > pc?
	3075	pxor mm1, mm1
	3076	pand mm3, mm7
	3077	pandn mm7, mm0
	3078	pxor mm1, mm1
	3079	paddw mm7, mm3
	3080	pxor mm0, mm0
	3081	// Step ex to next set of 8 bytes and repeat loop til done
	3082	add ebx, 8
	3083	packuswb mm1, mm7
	3084	paddb mm1, [edi + ebx - 8] // add Paeth predictor with Raw(x)
	3085	cmp ebx, MMXLength
	3086	movq [edi + ebx - 8], mm1 // write back updated value
	3087	// mm1 will be used as Raw(x-bpp) next loop
	3088	jb dpth8lp
	3089	} // end _asm block
	3090	}
	3091	break;
	3092
	3093	case 1: // bpp = 1
	3094	case 2: // bpp = 2
	3095	default: // bpp > 8
	3096	{
	3097	_asm {
	3098	mov ebx, diff
	3099	cmp ebx, FullLength
	3100	jnb dpthdend
	3101	mov edi, row
	3102	mov esi, prev_row
	3103	// Do Paeth decode for remaining bytes
	3104	mov edx, ebx
	3105	xor ecx, ecx // zero ecx before using cl & cx in loop below
	3106	sub edx, bpp // Set edx = ebx - bpp
	3107	dpthdlp:
	3108	xor eax, eax
	3109	// pav = p - a = (a + b - c) - a = b - c
	3110	mov al, [esi + ebx] // load Prior(x) into al
	3111	mov cl, [esi + edx] // load Prior(x-bpp) into cl
	3112	sub eax, ecx // subtract Prior(x-bpp)
	3113	mov patemp, eax // Save pav for later use
	3114	xor eax, eax
	3115	// pbv = p - b = (a + b - c) - b = a - c
	3116	mov al, [edi + edx] // load Raw(x-bpp) into al
	3117	sub eax, ecx // subtract Prior(x-bpp)
	3118	mov ecx, eax
	3119	// pcv = p - c = (a + b - c) -c = (a - c) + (b - c) = pav + pbv
	3120	add eax, patemp // pcv = pav + pbv
	3121	// pc = abs(pcv)
	3122	test eax, 0x80000000
	3123	jz dpthdpca
	3124	neg eax // reverse sign of neg values
	3125	dpthdpca:
	3126	mov pctemp, eax // save pc for later use
	3127	// pb = abs(pbv)
	3128	test ecx, 0x80000000
	3129	jz dpthdpba
	3130	neg ecx // reverse sign of neg values
	3131	dpthdpba:
	3132	mov pbtemp, ecx // save pb for later use
	3133	// pa = abs(pav)
	3134	mov eax, patemp
	3135	test eax, 0x80000000
	3136	jz dpthdpaa
	3137	neg eax // reverse sign of neg values
	3138	dpthdpaa:
	3139	mov patemp, eax // save pa for later use
	3140	// test if pa <= pb
	3141	cmp eax, ecx
	3142	jna dpthdabb
	3143	// pa > pb; now test if pb <= pc
	3144	cmp ecx, pctemp
	3145	jna dpthdbbc
	3146	// pb > pc; Raw(x) = Paeth(x) + Prior(x-bpp)
	3147	mov cl, [esi + edx] // load Prior(x-bpp) into cl
	3148	jmp dpthdpaeth
	3149	dpthdbbc:
	3150	// pb <= pc; Raw(x) = Paeth(x) + Prior(x)
	3151	mov cl, [esi + ebx] // load Prior(x) into cl
	3152	jmp dpthdpaeth
	3153	dpthdabb:
	3154	// pa <= pb; now test if pa <= pc
	3155	cmp eax, pctemp
	3156	jna dpthdabc
	3157	// pa > pc; Raw(x) = Paeth(x) + Prior(x-bpp)
	3158	mov cl, [esi + edx] // load Prior(x-bpp) into cl
	3159	jmp dpthdpaeth
	3160	dpthdabc:
	3161	// pa <= pc; Raw(x) = Paeth(x) + Raw(x-bpp)
	3162	mov cl, [edi + edx] // load Raw(x-bpp) into cl
	3163	dpthdpaeth:
	3164	inc ebx
	3165	inc edx
	3166	// Raw(x) = (Paeth(x) + Paeth_Predictor( a, b, c )) mod 256
	3167	add [edi + ebx - 1], cl
	3168	cmp ebx, FullLength
	3169	jb dpthdlp
	3170	dpthdend:
	3171	} // end _asm block
	3172	}
	3173	return; // No need to go further with this one
	3174	} // end switch ( bpp )
	3175	_asm
	3176	{
	3177	// MMX acceleration complete now do clean-up
	3178	// Check if any remaining bytes left to decode
	3179	mov ebx, MMXLength
	3180	cmp ebx, FullLength
	3181	jnb dpthend
	3182	mov edi, row
	3183	mov esi, prev_row
	3184	// Do Paeth decode for remaining bytes
	3185	mov edx, ebx
	3186	xor ecx, ecx // zero ecx before using cl & cx in loop below
	3187	sub edx, bpp // Set edx = ebx - bpp
	3188	dpthlp2:
	3189	xor eax, eax
	3190	// pav = p - a = (a + b - c) - a = b - c
	3191	mov al, [esi + ebx] // load Prior(x) into al
	3192	mov cl, [esi + edx] // load Prior(x-bpp) into cl
	3193	sub eax, ecx // subtract Prior(x-bpp)
	3194	mov patemp, eax // Save pav for later use
	3195	xor eax, eax
	3196	// pbv = p - b = (a + b - c) - b = a - c
	3197	mov al, [edi + edx] // load Raw(x-bpp) into al
	3198	sub eax, ecx // subtract Prior(x-bpp)
	3199	mov ecx, eax
	3200	// pcv = p - c = (a + b - c) -c = (a - c) + (b - c) = pav + pbv
	3201	add eax, patemp // pcv = pav + pbv
	3202	// pc = abs(pcv)
	3203	test eax, 0x80000000
	3204	jz dpthpca2
	3205	neg eax // reverse sign of neg values
	3206	dpthpca2:
	3207	mov pctemp, eax // save pc for later use
	3208	// pb = abs(pbv)
	3209	test ecx, 0x80000000
	3210	jz dpthpba2
	3211	neg ecx // reverse sign of neg values
	3212	dpthpba2:
	3213	mov pbtemp, ecx // save pb for later use
	3214	// pa = abs(pav)
	3215	mov eax, patemp
	3216	test eax, 0x80000000
	3217	jz dpthpaa2
	3218	neg eax // reverse sign of neg values
	3219	dpthpaa2:
	3220	mov patemp, eax // save pa for later use
	3221	// test if pa <= pb
	3222	cmp eax, ecx
	3223	jna dpthabb2
	3224	// pa > pb; now test if pb <= pc
	3225	cmp ecx, pctemp
	3226	jna dpthbbc2
	3227	// pb > pc; Raw(x) = Paeth(x) + Prior(x-bpp)
	3228	mov cl, [esi + edx] // load Prior(x-bpp) into cl
	3229	jmp dpthpaeth2
	3230	dpthbbc2:
	3231	// pb <= pc; Raw(x) = Paeth(x) + Prior(x)
	3232	mov cl, [esi + ebx] // load Prior(x) into cl
	3233	jmp dpthpaeth2
	3234	dpthabb2:
	3235	// pa <= pb; now test if pa <= pc
	3236	cmp eax, pctemp
	3237	jna dpthabc2
	3238	// pa > pc; Raw(x) = Paeth(x) + Prior(x-bpp)
	3239	mov cl, [esi + edx] // load Prior(x-bpp) into cl
	3240	jmp dpthpaeth2
	3241	dpthabc2:
	3242	// pa <= pc; Raw(x) = Paeth(x) + Raw(x-bpp)
	3243	mov cl, [edi + edx] // load Raw(x-bpp) into cl
	3244	dpthpaeth2:
	3245	inc ebx
	3246	inc edx
	3247	// Raw(x) = (Paeth(x) + Paeth_Predictor( a, b, c )) mod 256
	3248	add [edi + ebx - 1], cl
	3249	cmp ebx, FullLength
	3250	jb dpthlp2
	3251	dpthend:
	3252	emms // End MMX instructions; prep for possible FP instrs.
	3253	} // end _asm block
	3254	}
	3255
	3256	// Optimized code for PNG Sub filter decoder
	3257	void /* PRIVATE */
	3258	png_read_filter_row_mmx_sub(png_row_infop row_info, png_bytep row)
	3259	{
	3260	//int test;
	3261	int bpp;
	3262	png_uint_32 FullLength;
	3263	png_uint_32 MMXLength;
	3264	int diff;
	3265
	3266	bpp = (row_info->pixel_depth + 7) >> 3; // Get # bytes per pixel
	3267	FullLength = row_info->rowbytes - bpp; // # of bytes to filter
	3268	_asm {
	3269	mov edi, row
	3270	mov esi, edi // lp = row
	3271	add edi, bpp // rp = row + bpp
	3272	xor eax, eax
	3273	// get # of bytes to alignment
	3274	mov diff, edi // take start of row
	3275	add diff, 0xf // add 7 + 8 to incr past
	3276	// alignment boundary
	3277	xor ebx, ebx
	3278	and diff, 0xfffffff8 // mask to alignment boundary
	3279	sub diff, edi // subtract from start ==> value
	3280	// ebx at alignment
	3281	jz dsubgo
	3282	// fix alignment
	3283	dsublp1:
	3284	mov al, [esi+ebx]
	3285	add [edi+ebx], al
	3286	inc ebx
	3287	cmp ebx, diff
	3288	jb dsublp1
	3289	dsubgo:
	3290	mov ecx, FullLength
	3291	mov edx, ecx
	3292	sub edx, ebx // subtract alignment fix
	3293	and edx, 0x00000007 // calc bytes over mult of 8
	3294	sub ecx, edx // drop over bytes from length
	3295	mov MMXLength, ecx
	3296	} // end _asm block
	3297
	3298	// Now do the math for the rest of the row
	3299	switch ( bpp )
	3300	{
	3301	case 3:
	3302	{
	3303	ActiveMask.use = 0x0000ffffff000000;
	3304	ShiftBpp.use = 24; // == 3 * 8
	3305	ShiftRem.use = 40; // == 64 - 24
	3306	_asm {
	3307	mov edi, row
	3308	movq mm7, ActiveMask // Load ActiveMask for 2nd active byte group
	3309	mov esi, edi // lp = row
	3310	add edi, bpp // rp = row + bpp
	3311	movq mm6, mm7
	3312	mov ebx, diff
	3313	psllq mm6, ShiftBpp // Move mask in mm6 to cover 3rd active
	3314	// byte group
	3315	// PRIME the pump (load the first Raw(x-bpp) data set
	3316	movq mm1, [edi+ebx-8]
	3317	dsub3lp:
	3318	psrlq mm1, ShiftRem // Shift data for adding 1st bpp bytes
	3319	// no need for mask; shift clears inactive bytes
	3320	// Add 1st active group
	3321	movq mm0, [edi+ebx]
	3322	paddb mm0, mm1
	3323	// Add 2nd active group
	3324	movq mm1, mm0 // mov updated Raws to mm1
	3325	psllq mm1, ShiftBpp // shift data to position correctly
	3326	pand mm1, mm7 // mask to use only 2nd active group
	3327	paddb mm0, mm1
	3328	// Add 3rd active group
	3329	movq mm1, mm0 // mov updated Raws to mm1
	3330	psllq mm1, ShiftBpp // shift data to position correctly
	3331	pand mm1, mm6 // mask to use only 3rd active group
	3332	add ebx, 8
	3333	paddb mm0, mm1
	3334	cmp ebx, MMXLength
	3335	movq [edi+ebx-8], mm0 // Write updated Raws back to array
	3336	// Prep for doing 1st add at top of loop
	3337	movq mm1, mm0
	3338	jb dsub3lp
	3339	} // end _asm block
	3340	}
	3341	break;
	3342
	3343	case 1:
	3344	{
	3345	// Placed here just in case this is a duplicate of the
	3346	// non-MMX code for the SUB filter in png_read_filter_row below
	3347	//
	3348	// png_bytep rp;
	3349	// png_bytep lp;
	3350	// png_uint_32 i;
	3351	// bpp = (row_info->pixel_depth + 7) >> 3;
	3352	// for (i = (png_uint_32)bpp, rp = row + bpp, lp = row;
	3353	// i < row_info->rowbytes; i++, rp++, lp++)
	3354	// {
	3355	// rp = (png_byte)(((int)(rp) + (int)(*lp)) & 0xff);
	3356	// }
	3357	_asm {
	3358	mov ebx, diff
	3359	mov edi, row
	3360	cmp ebx, FullLength
	3361	jnb dsub1end
	3362	mov esi, edi // lp = row
	3363	xor eax, eax
	3364	add edi, bpp // rp = row + bpp
	3365	dsub1lp:
	3366	mov al, [esi+ebx]
	3367	add [edi+ebx], al
	3368	inc ebx
	3369	cmp ebx, FullLength
	3370	jb dsub1lp
	3371	dsub1end:
	3372	} // end _asm block
	3373	}
	3374	return;
	3375
	3376	case 6:
	3377	case 7:
	3378	case 4:
	3379	case 5:
	3380	{
	3381	ShiftBpp.use = bpp << 3;
	3382	ShiftRem.use = 64 - ShiftBpp.use;
	3383	_asm {
	3384	mov edi, row
	3385	mov ebx, diff
	3386	mov esi, edi // lp = row
	3387	add edi, bpp // rp = row + bpp
	3388	// PRIME the pump (load the first Raw(x-bpp) data set
	3389	movq mm1, [edi+ebx-8]
	3390	dsub4lp:
	3391	psrlq mm1, ShiftRem // Shift data for adding 1st bpp bytes
	3392	// no need for mask; shift clears inactive bytes
	3393	movq mm0, [edi+ebx]
	3394	paddb mm0, mm1
	3395	// Add 2nd active group
	3396	movq mm1, mm0 // mov updated Raws to mm1
	3397	psllq mm1, ShiftBpp // shift data to position correctly
	3398	// there is no need for any mask
	3399	// since shift clears inactive bits/bytes
	3400	add ebx, 8
	3401	paddb mm0, mm1
	3402	cmp ebx, MMXLength
	3403	movq [edi+ebx-8], mm0
	3404	movq mm1, mm0 // Prep for doing 1st add at top of loop
	3405	jb dsub4lp
	3406	} // end _asm block
	3407	}
	3408	break;
	3409
	3410	case 2:
	3411	{
	3412	ActiveMask.use = 0x00000000ffff0000;
	3413	ShiftBpp.use = 16; // == 2 * 8
	3414	ShiftRem.use = 48; // == 64 - 16
	3415	_asm {
	3416	movq mm7, ActiveMask // Load ActiveMask for 2nd active byte group
	3417	mov ebx, diff
	3418	movq mm6, mm7
	3419	mov edi, row
	3420	psllq mm6, ShiftBpp // Move mask in mm6 to cover 3rd active
	3421	// byte group
	3422	mov esi, edi // lp = row
	3423	movq mm5, mm6
	3424	add edi, bpp // rp = row + bpp
	3425	psllq mm5, ShiftBpp // Move mask in mm5 to cover 4th active
	3426	// byte group
	3427	// PRIME the pump (load the first Raw(x-bpp) data set
	3428	movq mm1, [edi+ebx-8]
	3429	dsub2lp:
	3430	// Add 1st active group
	3431	psrlq mm1, ShiftRem // Shift data for adding 1st bpp bytes
	3432	// no need for mask; shift clears inactive
	3433	// bytes
	3434	movq mm0, [edi+ebx]
	3435	paddb mm0, mm1
	3436	// Add 2nd active group
	3437	movq mm1, mm0 // mov updated Raws to mm1
	3438	psllq mm1, ShiftBpp // shift data to position correctly
	3439	pand mm1, mm7 // mask to use only 2nd active group
	3440	paddb mm0, mm1
	3441	// Add 3rd active group
	3442	movq mm1, mm0 // mov updated Raws to mm1
	3443	psllq mm1, ShiftBpp // shift data to position correctly
	3444	pand mm1, mm6 // mask to use only 3rd active group
	3445	paddb mm0, mm1
	3446	// Add 4th active group
	3447	movq mm1, mm0 // mov updated Raws to mm1
	3448	psllq mm1, ShiftBpp // shift data to position correctly
	3449	pand mm1, mm5 // mask to use only 4th active group
	3450	add ebx, 8
	3451	paddb mm0, mm1
	3452	cmp ebx, MMXLength
	3453	movq [edi+ebx-8], mm0 // Write updated Raws back to array
	3454	movq mm1, mm0 // Prep for doing 1st add at top of loop
	3455	jb dsub2lp
	3456	} // end _asm block
	3457	}
	3458	break;
	3459	case 8:
	3460	{
	3461	_asm {
	3462	mov edi, row
	3463	mov ebx, diff
	3464	mov esi, edi // lp = row
	3465	add edi, bpp // rp = row + bpp
	3466	mov ecx, MMXLength
	3467	movq mm7, [edi+ebx-8] // PRIME the pump (load the first
	3468	// Raw(x-bpp) data set
	3469	and ecx, 0x0000003f // calc bytes over mult of 64
	3470	dsub8lp:
	3471	movq mm0, [edi+ebx] // Load Sub(x) for 1st 8 bytes
	3472	paddb mm0, mm7
	3473	movq mm1, [edi+ebx+8] // Load Sub(x) for 2nd 8 bytes
	3474	movq [edi+ebx], mm0 // Write Raw(x) for 1st 8 bytes
	3475	// Now mm0 will be used as Raw(x-bpp) for
	3476	// the 2nd group of 8 bytes. This will be
	3477	// repeated for each group of 8 bytes with
	3478	// the 8th group being used as the Raw(x-bpp)
	3479	// for the 1st group of the next loop.
	3480	paddb mm1, mm0
	3481	movq mm2, [edi+ebx+16] // Load Sub(x) for 3rd 8 bytes
	3482	movq [edi+ebx+8], mm1 // Write Raw(x) for 2nd 8 bytes
	3483	paddb mm2, mm1
	3484	movq mm3, [edi+ebx+24] // Load Sub(x) for 4th 8 bytes
	3485	movq [edi+ebx+16], mm2 // Write Raw(x) for 3rd 8 bytes
	3486	paddb mm3, mm2
	3487	movq mm4, [edi+ebx+32] // Load Sub(x) for 5th 8 bytes
	3488	movq [edi+ebx+24], mm3 // Write Raw(x) for 4th 8 bytes
	3489	paddb mm4, mm3
	3490	movq mm5, [edi+ebx+40] // Load Sub(x) for 6th 8 bytes
	3491	movq [edi+ebx+32], mm4 // Write Raw(x) for 5th 8 bytes
	3492	paddb mm5, mm4
	3493	movq mm6, [edi+ebx+48] // Load Sub(x) for 7th 8 bytes
	3494	movq [edi+ebx+40], mm5 // Write Raw(x) for 6th 8 bytes
	3495	paddb mm6, mm5
	3496	movq mm7, [edi+ebx+56] // Load Sub(x) for 8th 8 bytes
	3497	movq [edi+ebx+48], mm6 // Write Raw(x) for 7th 8 bytes
	3498	add ebx, 64
	3499	paddb mm7, mm6
	3500	cmp ebx, ecx
	3501	movq [edi+ebx-8], mm7 // Write Raw(x) for 8th 8 bytes
	3502	jb dsub8lp
	3503	cmp ebx, MMXLength
	3504	jnb dsub8lt8
	3505	dsub8lpA:
	3506	movq mm0, [edi+ebx]
	3507	add ebx, 8
	3508	paddb mm0, mm7
	3509	cmp ebx, MMXLength
	3510	movq [edi+ebx-8], mm0 // use -8 to offset early add to ebx
	3511	movq mm7, mm0 // Move calculated Raw(x) data to mm1 to
	3512	// be the new Raw(x-bpp) for the next loop
	3513	jb dsub8lpA
	3514	dsub8lt8:
	3515	} // end _asm block
	3516	}
	3517	break;
	3518
	3519	default: // bpp greater than 8 bytes
	3520	{
	3521	_asm {
	3522	mov ebx, diff
	3523	mov edi, row
	3524	mov esi, edi // lp = row
	3525	add edi, bpp // rp = row + bpp
	3526	dsubAlp:
	3527	movq mm0, [edi+ebx]
	3528	movq mm1, [esi+ebx]
	3529	add ebx, 8
	3530	paddb mm0, mm1
	3531	cmp ebx, MMXLength
	3532	movq [edi+ebx-8], mm0 // mov does not affect flags; -8 to offset
	3533	// add ebx
	3534	jb dsubAlp
	3535	} // end _asm block
	3536	}
	3537	break;
	3538
	3539	} // end switch ( bpp )
	3540
	3541	_asm {
	3542	mov ebx, MMXLength
	3543	mov edi, row
	3544	cmp ebx, FullLength
	3545	jnb dsubend
	3546	mov esi, edi // lp = row
	3547	xor eax, eax
	3548	add edi, bpp // rp = row + bpp
	3549	dsublp2:
	3550	mov al, [esi+ebx]
	3551	add [edi+ebx], al
	3552	inc ebx
	3553	cmp ebx, FullLength
	3554	jb dsublp2
	3555	dsubend:
	3556	emms // End MMX instructions; prep for possible FP instrs.
	3557	} // end _asm block
	3558	}
	3559
	3560	// Optimized code for PNG Up filter decoder
	3561	void /* PRIVATE */
	3562	png_read_filter_row_mmx_up(png_row_infop row_info, png_bytep row,
	3563	png_bytep prev_row)
	3564	{
	3565	png_uint_32 len;
	3566	len = row_info->rowbytes; // # of bytes to filter
	3567	_asm {
	3568	mov edi, row
	3569	// get # of bytes to alignment
	3570	mov ecx, edi
	3571	xor ebx, ebx
	3572	add ecx, 0x7
	3573	xor eax, eax
	3574	and ecx, 0xfffffff8
	3575	mov esi, prev_row
	3576	sub ecx, edi
	3577	jz dupgo
	3578	// fix alignment
	3579	duplp1:
	3580	mov al, [edi+ebx]
	3581	add al, [esi+ebx]
	3582	inc ebx
	3583	cmp ebx, ecx
	3584	mov [edi + ebx-1], al // mov does not affect flags; -1 to offset inc ebx
	3585	jb duplp1
	3586	dupgo:
	3587	mov ecx, len
	3588	mov edx, ecx
	3589	sub edx, ebx // subtract alignment fix
	3590	and edx, 0x0000003f // calc bytes over mult of 64
	3591	sub ecx, edx // drop over bytes from length
	3592	// Unrolled loop - use all MMX registers and interleave to reduce
	3593	// number of branch instructions (loops) and reduce partial stalls
	3594	duploop:
	3595	movq mm1, [esi+ebx]
	3596	movq mm0, [edi+ebx]
	3597	movq mm3, [esi+ebx+8]
	3598	paddb mm0, mm1
	3599	movq mm2, [edi+ebx+8]
	3600	movq [edi+ebx], mm0
	3601	paddb mm2, mm3
	3602	movq mm5, [esi+ebx+16]
	3603	movq [edi+ebx+8], mm2
	3604	movq mm4, [edi+ebx+16]
	3605	movq mm7, [esi+ebx+24]
	3606	paddb mm4, mm5
	3607	movq mm6, [edi+ebx+24]
	3608	movq [edi+ebx+16], mm4
	3609	paddb mm6, mm7
	3610	movq mm1, [esi+ebx+32]
	3611	movq [edi+ebx+24], mm6
	3612	movq mm0, [edi+ebx+32]
	3613	movq mm3, [esi+ebx+40]
	3614	paddb mm0, mm1
	3615	movq mm2, [edi+ebx+40]
	3616	movq [edi+ebx+32], mm0
	3617	paddb mm2, mm3
	3618	movq mm5, [esi+ebx+48]
	3619	movq [edi+ebx+40], mm2
	3620	movq mm4, [edi+ebx+48]
	3621	movq mm7, [esi+ebx+56]
	3622	paddb mm4, mm5
	3623	movq mm6, [edi+ebx+56]
	3624	movq [edi+ebx+48], mm4
	3625	add ebx, 64
	3626	paddb mm6, mm7
	3627	cmp ebx, ecx
	3628	movq [edi+ebx-8], mm6 // (+56)movq does not affect flags;
	3629	// -8 to offset add ebx
	3630	jb duploop
	3631
	3632	cmp edx, 0 // Test for bytes over mult of 64
	3633	jz dupend
	3634
	3635
	3636	// 2 lines added by lcreeve at netins.net
	3637	// (mail 11 Jul 98 in png-implement list)
	3638	cmp edx, 8 //test for less than 8 bytes
	3639	jb duplt8
	3640
	3641
	3642	add ecx, edx
	3643	and edx, 0x00000007 // calc bytes over mult of 8
	3644	sub ecx, edx // drop over bytes from length
	3645	jz duplt8
	3646	// Loop using MMX registers mm0 & mm1 to update 8 bytes simultaneously
	3647	duplpA:
	3648	movq mm1, [esi+ebx]
	3649	movq mm0, [edi+ebx]
	3650	add ebx, 8
	3651	paddb mm0, mm1
	3652	cmp ebx, ecx
	3653	movq [edi+ebx-8], mm0 // movq does not affect flags; -8 to offset add ebx
	3654	jb duplpA
	3655	cmp edx, 0 // Test for bytes over mult of 8
	3656	jz dupend
	3657	duplt8:
	3658	xor eax, eax
	3659	add ecx, edx // move over byte count into counter
	3660	// Loop using x86 registers to update remaining bytes
	3661	duplp2:
	3662	mov al, [edi + ebx]
	3663	add al, [esi + ebx]
	3664	inc ebx
	3665	cmp ebx, ecx
	3666	mov [edi + ebx-1], al // mov does not affect flags; -1 to offset inc ebx
	3667	jb duplp2
	3668	dupend:
	3669	// Conversion of filtered row completed
	3670	emms // End MMX instructions; prep for possible FP instrs.
	3671	} // end _asm block
	3672	}
	3673
	3674
	3675	// Optimized png_read_filter_row routines
	3676	void /* PRIVATE */
	3677	png_read_filter_row(png_structp png_ptr, png_row_infop row_info, png_bytep
	3678	row, png_bytep prev_row, int filter)
	3679	{
	3680	#ifdef PNG_DEBUG
	3681	char filnm[10];
	3682	#endif
	3683
	3684	if (mmx_supported == 2) {
	3685	#if !defined(PNG_1_0_X)
	3686	/* this should have happened in png_init_mmx_flags() already */
	3687	png_warning(png_ptr, "asm_flags may not have been initialized");
	3688	#endif
	3689	png_mmx_support();
	3690	}
	3691
	3692	#ifdef PNG_DEBUG
	3693	png_debug(1, "in png_read_filter_row\n");
	3694	switch (filter)
	3695	{
	3696	case 0: sprintf(filnm, "none");
	3697	break;
	3698	#if !defined(PNG_1_0_X)
	3699	case 1: sprintf(filnm, "sub-%s",
	3700	(png_ptr->asm_flags & PNG_ASM_FLAG_MMX_READ_FILTER_SUB)? "MMX" : "x86");
	3701	break;
	3702	case 2: sprintf(filnm, "up-%s",
	3703	(png_ptr->asm_flags & PNG_ASM_FLAG_MMX_READ_FILTER_UP)? "MMX" : "x86");
	3704	break;
	3705	case 3: sprintf(filnm, "avg-%s",
	3706	(png_ptr->asm_flags & PNG_ASM_FLAG_MMX_READ_FILTER_AVG)? "MMX" : "x86");
	3707	break;
	3708	case 4: sprintf(filnm, "Paeth-%s",
	3709	(png_ptr->asm_flags & PNG_ASM_FLAG_MMX_READ_FILTER_PAETH)? "MMX":"x86");
	3710	break;
	3711	#else
	3712	case 1: sprintf(filnm, "sub");
	3713	break;
	3714	case 2: sprintf(filnm, "up");
	3715	break;
	3716	case 3: sprintf(filnm, "avg");
	3717	break;
	3718	case 4: sprintf(filnm, "Paeth");
	3719	break;
	3720	#endif
	3721	default: sprintf(filnm, "unknw");
	3722	break;
	3723	}
	3724	png_debug2(0,"row=%5d, %s, ", png_ptr->row_number, filnm);
	3725	png_debug2(0, "pd=%2d, b=%d, ", (int)row_info->pixel_depth,
	3726	(int)((row_info->pixel_depth + 7) >> 3));
	3727	png_debug1(0,"len=%8d, ", row_info->rowbytes);
	3728	#endif /* PNG_DEBUG */
	3729
	3730	switch (filter)
	3731	{
	3732	case PNG_FILTER_VALUE_NONE:
	3733	break;
	3734
	3735	case PNG_FILTER_VALUE_SUB:
	3736	{
	3737	#if !defined(PNG_1_0_X)
	3738	if ((png_ptr->asm_flags & PNG_ASM_FLAG_MMX_READ_FILTER_SUB) &&
	3739	(row_info->pixel_depth >= png_ptr->mmx_bitdepth_threshold) &&
	3740	(row_info->rowbytes >= png_ptr->mmx_rowbytes_threshold))
	3741	#else
	3742	if (mmx_supported)
	3743	#endif
	3744	{
	3745	png_read_filter_row_mmx_sub(row_info, row);
	3746	}
	3747	else
	3748	{
	3749	png_uint_32 i;
	3750	png_uint_32 istop = row_info->rowbytes;
	3751	png_uint_32 bpp = (row_info->pixel_depth + 7) >> 3;
	3752	png_bytep rp = row + bpp;
	3753	png_bytep lp = row;
	3754
	3755	for (i = bpp; i < istop; i++)
	3756	{
	3757	rp = (png_byte)(((int)(rp) + (int)(*lp++)) & 0xff);
	3758	rp++;
	3759	}
	3760	}
	3761	break;
	3762	}
	3763
	3764	case PNG_FILTER_VALUE_UP:
	3765	{
	3766	#if !defined(PNG_1_0_X)
	3767	if ((png_ptr->asm_flags & PNG_ASM_FLAG_MMX_READ_FILTER_UP) &&
	3768	(row_info->pixel_depth >= png_ptr->mmx_bitdepth_threshold) &&
	3769	(row_info->rowbytes >= png_ptr->mmx_rowbytes_threshold))
	3770	#else
	3771	if (mmx_supported)
	3772	#endif
	3773	{
	3774	png_read_filter_row_mmx_up(row_info, row, prev_row);
	3775	}
	3776	else
	3777	{
	3778	png_uint_32 i;
	3779	png_uint_32 istop = row_info->rowbytes;
	3780	png_bytep rp = row;
	3781	png_bytep pp = prev_row;
	3782
	3783	for (i = 0; i < istop; ++i)
	3784	{
	3785	rp = (png_byte)(((int)(rp) + (int)(*pp++)) & 0xff);
	3786	rp++;
	3787	}
	3788	}
	3789	break;
	3790	}
	3791
	3792	case PNG_FILTER_VALUE_AVG:
	3793	{
	3794	#if !defined(PNG_1_0_X)
	3795	if ((png_ptr->asm_flags & PNG_ASM_FLAG_MMX_READ_FILTER_AVG) &&
	3796	(row_info->pixel_depth >= png_ptr->mmx_bitdepth_threshold) &&
	3797	(row_info->rowbytes >= png_ptr->mmx_rowbytes_threshold))
	3798	#else
	3799	if (mmx_supported)
	3800	#endif
	3801	{
	3802	png_read_filter_row_mmx_avg(row_info, row, prev_row);
	3803	}
	3804	else
	3805	{
	3806	png_uint_32 i;
	3807	png_bytep rp = row;
	3808	png_bytep pp = prev_row;
	3809	png_bytep lp = row;
	3810	png_uint_32 bpp = (row_info->pixel_depth + 7) >> 3;
	3811	png_uint_32 istop = row_info->rowbytes - bpp;
	3812
	3813	for (i = 0; i < bpp; i++)
	3814	{
	3815	rp = (png_byte)(((int)(rp) +
	3816	((int)(*pp++) >> 1)) & 0xff);
	3817	rp++;
	3818	}
	3819
	3820	for (i = 0; i < istop; i++)
	3821	{
	3822	rp = (png_byte)(((int)(rp) +
	3823	((int)(pp++ + lp++) >> 1)) & 0xff);
	3824	rp++;
	3825	}
	3826	}
	3827	break;
	3828	}
	3829
	3830	case PNG_FILTER_VALUE_PAETH:
	3831	{
	3832	#if !defined(PNG_1_0_X)
	3833	if ((png_ptr->asm_flags & PNG_ASM_FLAG_MMX_READ_FILTER_PAETH) &&
	3834	(row_info->pixel_depth >= png_ptr->mmx_bitdepth_threshold) &&
	3835	(row_info->rowbytes >= png_ptr->mmx_rowbytes_threshold))
	3836	#else
	3837	if (mmx_supported)
	3838	#endif
	3839	{
	3840	png_read_filter_row_mmx_paeth(row_info, row, prev_row);
	3841	}
	3842	else
	3843	{
	3844	png_uint_32 i;
	3845	png_bytep rp = row;
	3846	png_bytep pp = prev_row;
	3847	png_bytep lp = row;
	3848	png_bytep cp = prev_row;
	3849	png_uint_32 bpp = (row_info->pixel_depth + 7) >> 3;
	3850	png_uint_32 istop=row_info->rowbytes - bpp;
	3851
	3852	for (i = 0; i < bpp; i++)
	3853	{
	3854	rp = (png_byte)(((int)(rp) + (int)(*pp++)) & 0xff);
	3855	rp++;
	3856	}
	3857
	3858	for (i = 0; i < istop; i++) // use leftover rp,pp
	3859	{
	3860	int a, b, c, pa, pb, pc, p;
	3861
	3862	a = *lp++;
	3863	b = *pp++;
	3864	c = *cp++;
	3865
	3866	p = b - c;
	3867	pc = a - c;
	3868
	3869	#ifdef PNG_USE_ABS
	3870	pa = abs(p);
	3871	pb = abs(pc);
	3872	pc = abs(p + pc);
	3873	#else
	3874	pa = p < 0 ? -p : p;
	3875	pb = pc < 0 ? -pc : pc;
	3876	pc = (p + pc) < 0 ? -(p + pc) : p + pc;
	3877	#endif
	3878
	3879	/*
	3880	if (pa <= pb && pa <= pc)
	3881	p = a;
	3882	else if (pb <= pc)
	3883	p = b;
	3884	else
	3885	p = c;
	3886	*/
	3887
	3888	p = (pa <= pb && pa <=pc) ? a : (pb <= pc) ? b : c;
	3889
	3890	rp = (png_byte)(((int)(rp) + p) & 0xff);
	3891	rp++;
	3892	}
	3893	}
	3894	break;
	3895	}
	3896
	3897	default:
	3898	png_warning(png_ptr, "Ignoring bad row filter type");
	3899	*row=0;
	3900	break;
	3901	}
	3902	}
	3903
	3904	#endif /* PNG_ASSEMBLER_CODE_SUPPORTED && PNG_USE_PNGVCRD */

Note: See TracBrowser for help on using the repository browser.

Context Navigation

source: liacs/MIR2010/SourceCode/cximage/png/pngvcrd.c@ 202

Download in other formats: