diff src/video/ataricommon/SDL_ataric2p.S @ 734:0310bb01091f

New C2P conversion routine
author Patrice Mandin <patmandin@gmail.com>
date Wed, 12 Nov 2003 18:45:48 +0000
parents f6ffac90895c
children 93207759e065
line wrap: on
line diff
--- a/src/video/ataricommon/SDL_ataric2p.S	Wed Nov 05 23:11:05 2003 +0000
+++ b/src/video/ataricommon/SDL_ataric2p.S	Wed Nov 12 18:45:48 2003 +0000
@@ -26,70 +26,274 @@
 #endif
 
 /*
- *	Chunky to planar conversion routine
- *	1 byte/pixel -> 4 or 8 bit planes
- *
- *	Patrice Mandin
- *	Xavier Joubert
- */
-
-/* ------------	Function pointers ------------ */
-
-	.globl	_Atari_C2pInit
-	.globl	_Atari_C2pConvert
-
-	.data
-
-	.even
-	.comm	_Atari_C2pInit,4
-	.comm	_Atari_C2pConvert,4
-
-/* ------------	Conversion Table C2P ------------ */
-
-	.globl	_Atari_table_c2p
-
-	.data
-
-	.comm	_Atari_table_c2p,2048
-
-/* ------------	Init table C2P, 8 bits ------------ */
+	Chunky to planar conversion routine
+	1 byte/pixel -> 4 or 8 bit planes
 
-	.globl	_Atari_C2pInit8
-
-	.text
-_Atari_C2pInit8:
-	movel	d2,sp@-
+	Patrice Mandin
+	Xavier Joubert
+	Mikael Kalms
+*/
 
-	lea	_Atari_table_c2p,a0
-	movew	#255,d0
-c2p_initbcl:
-#if defined (__M68020__)
-	lea	a0@(0,d0:w:8),a1
-#else
-	movew	d0,d2
-	lslw	#3,d2
-	lea	a0@(0,d2:w),a1
-#endif
-	moveq	#7,d1
-c2p_initbyte:
-	btst	d1,d0
-	sne	d2
-	negw	d2
-	moveb	d2,a1@(0,d1:w)
-
-	dbra	d1,c2p_initbyte
-
-	dbra	d0,c2p_initbcl
-
-	movel	sp@+,d2
-	rts
+	.globl	_SDL_Atari_C2pConvert
+	.globl	_SDL_Atari_C2pConvert8
+	.globl	_SDL_Atari_C2pConvert4
+	.globl	_SDL_Atari_C2pConvert4_pal
+	.globl	_SDL_Atari_table_c2p
 
 /* ------------	Conversion C2P, 8 bits ------------ */
 
-	.globl	_Atari_C2pConvert8
+	.text
+_SDL_Atari_C2pConvert8:
+	movel	sp@(4),c2p_source
+	movel	sp@(8),c2p_dest
+	movel	sp@(12),c2p_width
+	movel	sp@(16),c2p_height
+	movel	sp@(20),c2p_dblligne
+	movel	sp@(24),c2p_srcpitch
+	movel	sp@(28),c2p_dstpitch
+
+	moveml	d2-d7/a2-a6,sp@-
+
+	movel	c2p_source,c2p_cursrc
+	movel	c2p_dest,c2p_curdst
+	movel	#0x0f0f0f0f,d4
+	movel	#0x00ff00ff,d5
+	movel	#0x55555555,d6
+	movew	c2p_height+2,c2p_row
+	movew	c2p_width+2,d0
+	andw	#-8,d0
+	movew	d0,c2p_rowlen
+
+SDL_Atari_C2p8_rowloop:
+
+	movel	c2p_cursrc,a0
+	movel	c2p_curdst,a1
+
+	movel	a0,a2
+	addw	c2p_rowlen,a2
+
+	movel	a0@+,d0
+	movel	a0@+,d1
+	movel	a0@+,d2
+	movel	a0@+,d3
+/*
+	d0 = a7a6a5a4a3a2a1a0 b7b6b5b4b3b2b1b0 c7c6c5c4c3c2c1c0 d7d6d5d4d3d2d1d0
+	d1 = e7e6e5e4e3e2e1e0 f7f6f5f4f3f2f1f0 g7g6g5g4g3g2g1g0 h7h6h5h4h3h2h1h0
+	d2 = i7i6i5i4i3i2i1i0 j7j6j5j4j3j2j1j0 k7k6k5k4k3k2k1k0 l7l6l5l4l3l2l1l0
+	d3 = m7m6m5m4m3m2m1m0 n7n6n5n4n3n2n1n0 o7o6o5o4o3o2o1o0 p7p6p5p4p3p2p1p0
+*/
+	movel	d1,d7
+	lsrl	#4,d7
+	eorl	d0,d7
+	andl	d4,d7
+	eorl	d7,d0
+	lsll	#4,d7
+	eorl	d7,d1
+
+	movel	d3,d7
+	lsrl	#4,d7
+	eorl	d2,d7
+	andl	d4,d7
+	eorl	d7,d2
+	lsll	#4,d7
+	eorl	d7,d3
+
+	movel	d2,d7
+	lsrl	#8,d7
+	eorl	d0,d7
+ 	andl	d5,d7
+	eorl	d7,d0
+	lsll	#8,d7
+	eorl	d7,d2
+
+	movel	d3,d7
+	lsrl	#8,d7
+	eorl	d1,d7
+ 	andl	d5,d7
+	eorl	d7,d1
+	lsll	#8,d7
+	eorl	d7,d3
+/*
+	d0 = a7a6a5a4e7e6e5e4 i7i6i5i4m7m6m5m4 c7c6c5c4g7g6g5g4 k7k6k5k4o7o6o5o4
+	d1 = a3a2a1a0e3e2e1e0 i3i2i1i0m3m2m1m0 c3c2c1c0g3g2g1g0 k3k2k1k0o3o2o1o0
+	d2 = b7b6b5b4f7f6f5f4 j7j6j5j4n7n6n5n4 d7d6d5d4h7h6h5h4 l7l6l5l4p7p6p5p4
+	d3 = b3b2b1b0f3f2f1f0 j3j2j1j0n3n2n1n0 d3d2d1d0h3h2h1h0 l3l2l1l0p3p2p1p0
+*/
+	bras	SDL_Atari_C2p8_start
+
+SDL_Atari_C2p8_pix16:
+
+	movel	a0@+,d0
+	movel	a0@+,d1
+	movel	a0@+,d2
+	movel	a0@+,d3
+/*
+	d0 = a7a6a5a4a3a2a1a0 b7b6b5b4b3b2b1b0 c7c6c5c4c3c2c1c0 d7d6d5d4d3d2d1d0
+	d1 = e7e6e5e4e3e2e1e0 f7f6f5f4f3f2f1f0 g7g6g5g4g3g2g1g0 h7h6h5h4h3h2h1h0
+	d2 = i7i6i5i4i3i2i1i0 j7j6j5j4j3j2j1j0 k7k6k5k4k3k2k1k0 l7l6l5l4l3l2l1l0
+	d3 = m7m6m5m4m3m2m1m0 n7n6n5n4n3n2n1n0 o7o6o5o4o3o2o1o0 p7p6p5p4p3p2p1p0
+*/
+	movel	d1,d7
+	lsrl	#4,d7
+	movel	a3,a1@+
+	eorl	d0,d7
+	andl	d4,d7
+	eorl	d7,d0
+	lsll	#4,d7
+	eorl	d7,d1
+
+	movel	d3,d7
+	lsrl	#4,d7
+	eorl	d2,d7
+	andl	d4,d7
+	eorl	d7,d2
+	movel	a4,a1@+
+	lsll	#4,d7
+	eorl	d7,d3
+
+	movel	d2,d7
+	lsrl	#8,d7
+	eorl	d0,d7
+	andl	d5,d7
+	eorl	d7,d0
+	movel	a5,a1@+
+	lsll	#8,d7
+	eorl	d7,d2
 
-	.text
-_Atari_C2pConvert8:
+	movel	d3,d7
+	lsrl	#8,d7
+	eorl	d1,d7
+	andl	d5,d7
+	eorl	d7,d1
+	movel	a6,a1@+
+	lsll	#8,d7
+	eorl	d7,d3
+/*
+	d0 = a7a6a5a4e7e6e5e4 i7i6i5i4m7m6m5m4 c7c6c5c4g7g6g5g4 k7k6k5k4o7o6o5o4
+	d1 = a3a2a1a0e3e2e1e0 i3i2i1i0m3m2m1m0 c3c2c1c0g3g2g1g0 k3k2k1k0o3o2o1o0
+	d2 = b7b6b5b4f7f6f5f4 j7j6j5j4n7n6n5n4 d7d6d5d4h7h6h5h4 l7l6l5l4p7p6p5p4
+	d3 = b3b2b1b0f3f2f1f0 j3j2j1j0n3n2n1n0 d3d2d1d0h3h2h1h0 l3l2l1l0p3p2p1p0
+*/
+
+SDL_Atari_C2p8_start:
+
+	movel	d2,d7
+	lsrl	#1,d7
+	eorl	d0,d7
+	andl	d6,d7
+	eorl	d7,d0
+	addl	d7,d7
+	eorl	d7,d2
+
+	movel	d3,d7
+	lsrl	#1,d7
+	eorl	d1,d7
+	andl	d6,d7
+	eorl	d7,d1
+	addl	d7,d7
+	eorl	d7,d3
+/*
+	d0 = a7b7a5b5e7f7e5f5 i7j7i5j5m7n7m5n5 c7d7c5d5g7h7g5h5 k7l7k5l5o7p7o5p5
+	d1 = a3b3a1b1e3f3e1f1 i3j3i1j1m3n3m1n1 c3d3c1d1g3h3g1h1 k3l3k1l1o3p3o1p1
+	d2 = a6b6a4b4e6f6e4f4 i6j6i4j4m6n6m4n4 c6d6c4d4g6h6g4h4 k6l6k4l4o6p6o4p4
+	d3 = a2b2a0b0e2f2e0f0 i2j2i0j0m2n2m0n0 c2d2c0d0g2h2g0h0 k2l2k0l0o2p2o0p0
+*/
+	movew	d2,d7
+	movew	d0,d2
+	swap	d2
+	movew	d2,d0
+	movew	d7,d2
+
+	movew	d3,d7
+	movew	d1,d3
+	swap	d3
+	movew	d3,d1
+	movew	d7,d3
+/*
+	d0 = a7b7a5b5e7f7e5f5 i7j7i5j5m7n7m5n5 a6b6a4b4e6f6e4f4 i6j6i4j4m6n6m4n4
+	d1 = a3b3a1b1e3f3e1f1 i3j3i1j1m3n3m1n1 a2b2a0b0e2f2e0f0 i2j2i0j0m2n2m0n0
+	d2 = c7d7c5d5g7h7g5h5 k7l7k5l5o7p7o5p5 c6d6c4d4g6h6g4h4 k6l6k4l4o6p6o4p4
+	d3 = c3d3c1d1g3h3g1h1 k3l3k1l1o3p3o1p1 c2d2c0d0g2h2g0h0 k2l2k0l0o2p2o0p0
+*/
+	movel	d2,d7
+	lsrl	#2,d7
+	eorl	d0,d7
+	andl	#0x33333333,d7
+	eorl	d7,d0
+	lsll	#2,d7
+	eorl	d7,d2
+
+	movel	d3,d7
+	lsrl	#2,d7
+	eorl	d1,d7
+	andl	#0x33333333,d7
+	eorl	d7,d1
+	lsll	#2,d7
+	eorl	d7,d3
+/*
+	d0 = a7b7c7d7e7f7g7h7 i7j7k7l7m7n7o7p7 a6b6c6d6e6f6g6h6 i6j6k6l6m6n6o6p6
+	d1 = a3b3c3d3e3f3g3h3 i3j3k3l3m3n3o3p3 a2b2c2d2e2f2g2h2 i2j2k2l2m2n2o2p2
+	d2 = a5b5c5d5e5f5g5h5 i5j5k5l5m5n5o5p5 a4b4c4d4e4f4g4h4 i4j4k4l4m4n4o4p4
+	d3 = a1b1c1d1e1f1g1h1 i1j1k1l1m1n1o1p1 a0b0c0d0e0f0g0h0 i0j0k0l0m0n0o0p0
+*/
+	swap	d0
+	swap	d1
+	swap	d2
+	swap	d3
+
+	movel	d0,a6
+	movel	d2,a5
+	movel	d1,a4
+	movel	d3,a3
+
+	cmpl	a0,a2
+	bgt	SDL_Atari_C2p8_pix16
+
+	movel	a3,a1@+
+	movel	a4,a1@+
+	movel	a5,a1@+
+	movel	a6,a1@+
+
+	/* Double the line ? */
+
+	movel	c2p_srcpitch,d0
+	movel	c2p_dstpitch,d1
+
+	tstl	c2p_dblligne
+	beqs	SDL_Atari_C2p8_nodblline
+
+	movel	c2p_curdst,a0
+	movel	a0,a1
+	addl	d1,a1
+
+	movew	c2p_width+2,d7
+	lsrw	#4,d7
+	subql	#1,d7
+SDL_Atari_C2p8_dblloop:
+	movel	a0@+,a1@+
+	movel	a0@+,a1@+
+	movel	a0@+,a1@+
+	movel	a0@+,a1@+
+	dbra	d7,SDL_Atari_C2p8_dblloop
+
+	addl	d1,c2p_curdst
+
+SDL_Atari_C2p8_nodblline:
+
+	/* Next line */
+
+	addl	d0,c2p_cursrc
+	addl	d1,c2p_curdst
+
+	subqw	#1,c2p_row
+	bne	SDL_Atari_C2p8_rowloop	
+
+	moveml	sp@+,d2-d7/a2-a6
+	rts
+
+/* ------------	Conversion C2P, 4 bits ------------ */
+
+_SDL_Atari_C2pConvert4:
 	movel	sp@(4),c2p_source
 	movel	sp@(8),c2p_dest
 	movel	sp@(12),c2p_width
@@ -102,138 +306,7 @@
 
 	movel	c2p_source,a0
 	movel	c2p_dest,a1
-	lea	_Atari_table_c2p,a2
-	movel	#0x000f0001,d3
-#if defined(__M68020__)
-	moveq	#0,d0
-#endif
-	
-	movel	c2p_height,d7
-	subql	#1,d7
-c2p8_bcly:
-	movel	a0,a4	| Save start address of source
-	movel	a1,a5	| Save start address of dest
-
-	| Conversion
-                 
-	movel	c2p_width,d6
-	lsrw	#4,d6
-	subql	#1,d6
-c2p8_bclx:
-	| Octets 0-7
-	
-	moveq	#0,d1
-	moveq	#0,d2
-	moveq	#7,d5
-c2p8_bcl07:
-#if defined(__M68020__)
-	moveb	a0@+,d0
-	lea	a2@(0,d0:w:8),a3
-#else
-	moveq	#0,d0
-	moveb	a0@+,d0
-	lslw	#3,d0
-	lea	a2@(0,d0:w),a3
-#endif
-	lsll	#1,d1
-	lsll	#1,d2
-	orl	a3@+,d1
-	orl	a3@,d2
-	dbra	d5,c2p8_bcl07
-
-	movepl	d1,a1@(0)
-	movepl	d2,a1@(8)
-	addw	d3,a1
-	swap	d3
-	
-	| Octets 8-15
-
-	moveq	#0,d1
-	moveq	#0,d2
-	moveq	#7,d5
-c2p8_bcl815:
-#if defined(__M68020__)
-	moveb	a0@+,d0
-	lea	a2@(0,d0:w:8),a3
-#else
-	moveq	#0,d0
-	moveb	a0@+,d0
-	lslw	#3,d0
-	lea	a2@(0,d0:w),a3
-#endif
-	lsll	#1,d1
-	lsll	#1,d2
-	orl	a3@+,d1
-	orl	a3@,d2
-	dbra	d5,c2p8_bcl815
-
-	movepl	d1,a1@(0)
-	movepl	d2,a1@(8)
-	addw	d3,a1
-	swap	d3
-
-	dbra	d6,c2p8_bclx
-
-	| Double line ?
-
-	tstl	c2p_dblligne
-	beq	c2p8_nodblligne
-
-	movel	a5,a6			| src line
-	movel	a5,a1			| dest line
-	addl	c2p_dstpitch,a1
-
-	movel	c2p_width,d6
-	lsrw	#2,d6
-	subql	#1,d6
-c2p8_copydbl:
-	movel	a6@+,a1@+
-	dbra	d6,c2p8_copydbl
-
-	addl	c2p_dstpitch,a5
-c2p8_nodblligne:
-
-	| Next line
-
-	movel	a4,a0		
-	addl	c2p_srcpitch,a0
-	movel	a5,a1
-	addl	c2p_dstpitch,a1
-
-	dbra	d7,c2p8_bcly
-
-	moveml	sp@+,d2-d7/a2-a6
-	rts
-
-/* ------------	Init table C2P, 4 bits ------------ */
-
-	.globl	_Atari_C2pInit4
-
-	.text
-_Atari_C2pInit4:
-	/* Nothing to do */
-	/* work is done in convert_c2p_pal */
-	rts
-
-/* ------------	Conversion C2P, 4 bits ------------ */
-
-	.globl	_Atari_C2pConvert4
-
-	.text
-_Atari_C2pConvert4:
-	movel	sp@(4),c2p_source
-	movel	sp@(8),c2p_dest
-	movel	sp@(12),c2p_width
-	movel	sp@(16),c2p_height
-	movel	sp@(20),c2p_dblligne
-	movel	sp@(24),c2p_srcpitch
-	movel	sp@(28),c2p_dstpitch
-
-	moveml	d2-d7/a2-a6,sp@-
-
-	movel	c2p_source,a0
-	movel	c2p_dest,a1
-	lea	_Atari_table_c2p,a2
+	lea	_SDL_Atari_table_c2p,a2
 	movel	#0x00070001,d3
 #if defined(__M68020__)
 	moveq	#0,d0
@@ -328,18 +401,15 @@
 	moveml	sp@+,d2-d7/a2-a6
 	rts
 
-* ------------	Conversion of a light palette in 4 bits ------------ */
+/* ------------	Conversion of a light palette in 4 bits ------------ */
 
-	.globl	_Atari_C2pConvert4_pal
-
-	.text
-_Atari_C2pConvert4_pal:
+_SDL_Atari_C2pConvert4_pal:
 	/* a0 is a 256-word light palette */
 	movel	sp@(4),a0
 
 	moveml	d2-d3,sp@-
 
-	lea	_Atari_table_c2p,a1
+	lea	_SDL_Atari_table_c2p,a1
 	movew	#255,d3
 c2p_pal_initbcl:
 	movew	a0@+,d0
@@ -367,11 +437,17 @@
 	.data
 
 	.even
-	.comm	c2p_source,4
-	.comm	c2p_dest,4
-	.comm	c2p_width,4
-	.comm	c2p_height,4
-	.comm	c2p_dblligne,4
-	.comm	c2p_srcpitch,4
-	.comm	c2p_dstpitch,4
+	.comm	_SDL_Atari_C2pConvert,4
+	.comm	_SDL_Atari_table_c2p,2048
 
+	.comm	c2p_source,4	/* Source framebuffer */
+	.comm	c2p_dest,4	/* Destination framebuffer */
+	.comm	c2p_width,4	/* Width of zone to convert */
+	.comm	c2p_height,4	/* Height of zone to convert */
+	.comm	c2p_dblligne,4	/* Double the lines while converting ? */
+	.comm	c2p_srcpitch,4	/* Source pitch */
+	.comm	c2p_dstpitch,4	/* Destination pitch */
+	.comm	c2p_cursrc,4	/* Current source line */
+	.comm	c2p_curdst,4	/* Current destination line */
+	.comm	c2p_rowlen,2	/* Line length in bytes */
+	.comm	c2p_row,2	/* Current line number */