nkeynes@284: /**
nkeynes@285:  * $Id: pvr2mem.c,v 1.2 2007-01-15 10:10:51 nkeynes Exp $
nkeynes@284:  *
nkeynes@284:  * PVR2 (Video) VRAM handling routines (mainly for the 64-bit region)
nkeynes@284:  *
nkeynes@284:  * Copyright (c) 2005 Nathan Keynes.
nkeynes@284:  *
nkeynes@284:  * This program is free software; you can redistribute it and/or modify
nkeynes@284:  * it under the terms of the GNU General Public License as published by
nkeynes@284:  * the Free Software Foundation; either version 2 of the License, or
nkeynes@284:  * (at your option) any later version.
nkeynes@284:  *
nkeynes@284:  * This program is distributed in the hope that it will be useful,
nkeynes@284:  * but WITHOUT ANY WARRANTY; without even the implied warranty of
nkeynes@284:  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
nkeynes@284:  * GNU General Public License for more details.
nkeynes@284:  */
nkeynes@284: #include "pvr2.h"
nkeynes@284: 
nkeynes@284: extern char *video_base;
nkeynes@284: 
nkeynes@284: void pvr2_vram64_write( sh4addr_t destaddr, char *src, uint32_t length )
nkeynes@284: {
nkeynes@284:     int bank_flag = (destaddr & 0x04) >> 2;
nkeynes@284:     uint32_t *banks[2];
nkeynes@284:     uint32_t *dwsrc;
nkeynes@284:     int i;
nkeynes@284: 
nkeynes@284:     destaddr = destaddr & 0x7FFFFF;
nkeynes@284:     if( destaddr + length > 0x800000 ) {
nkeynes@284: 	length = 0x800000 - destaddr;
nkeynes@284:     }
nkeynes@284: 
nkeynes@284:     for( i=destaddr & 0xFFFFF000; i < destaddr + length; i+= PAGE_SIZE ) {
nkeynes@284: 	texcache_invalidate_page( i );
nkeynes@284:     }
nkeynes@284: 
nkeynes@284:     banks[0] = ((uint32_t *)(video_base + ((destaddr & 0x007FFFF8) >>1)));
nkeynes@284:     banks[1] = banks[0] + 0x100000;
nkeynes@284:     if( bank_flag ) 
nkeynes@284: 	banks[0]++;
nkeynes@284:     
nkeynes@284:     /* Handle non-aligned start of source */
nkeynes@284:     if( destaddr & 0x03 ) {
nkeynes@284: 	char *dest = ((char *)banks[bank_flag]) + (destaddr & 0x03);
nkeynes@284: 	for( i= destaddr & 0x03; i < 4 && length > 0; i++, length-- ) {
nkeynes@284: 	    *dest++ = *src++;
nkeynes@284: 	}
nkeynes@284: 	bank_flag = !bank_flag;
nkeynes@284:     }
nkeynes@284: 
nkeynes@284:     dwsrc = (uint32_t *)src;
nkeynes@284:     while( length >= 4 ) {
nkeynes@284: 	*banks[bank_flag]++ = *dwsrc++;
nkeynes@284: 	bank_flag = !bank_flag;
nkeynes@284: 	length -= 4;
nkeynes@284:     }
nkeynes@284:     
nkeynes@284:     /* Handle non-aligned end of source */
nkeynes@284:     if( length ) {
nkeynes@284: 	src = (char *)dwsrc;
nkeynes@284: 	char *dest = (char *)banks[bank_flag];
nkeynes@284: 	while( length-- > 0 ) {
nkeynes@284: 	    *dest++ = *src++;
nkeynes@284: 	}
nkeynes@284:     }  
nkeynes@284: }
nkeynes@284: 
nkeynes@284: /**
nkeynes@284:  * Write an image to 64-bit vram, with a line-stride different from the line-size.
nkeynes@284:  * The destaddr must be 32-bit aligned, and both line_bytes and line_stride_bytes
nkeynes@284:  * must be multiples of 4.
nkeynes@284:  */
nkeynes@284: void pvr2_vram64_write_stride( sh4addr_t destaddr, char *src, uint32_t line_bytes, 
nkeynes@284: 			       uint32_t line_stride_bytes, uint32_t line_count )
nkeynes@284: {
nkeynes@284:     int bank_flag = (destaddr & 0x04) >> 2;
nkeynes@284:     uint32_t *banks[2];
nkeynes@284:     uint32_t *dwsrc;
nkeynes@284:     uint32_t line_gap;
nkeynes@284:     int line_gap_flag;
nkeynes@284:     int i,j;
nkeynes@284: 
nkeynes@284:     destaddr = destaddr & 0x7FFFF8;
nkeynes@284:     i = line_stride_bytes - line_bytes;
nkeynes@284:     line_gap_flag = i & 0x04;
nkeynes@284:     line_gap = i >> 3;
nkeynes@284:     line_bytes >>= 2;
nkeynes@284: 
nkeynes@284:     for( i=destaddr & 0xFFFFF000; i < destaddr + line_stride_bytes*line_count; i+= PAGE_SIZE ) {
nkeynes@284: 	texcache_invalidate_page( i );
nkeynes@284:     }
nkeynes@284: 
nkeynes@284:     banks[0] = (uint32_t *)(video_base + (destaddr >>1));
nkeynes@284:     banks[1] = banks[0] + 0x100000;
nkeynes@284:     if( bank_flag ) 
nkeynes@284: 	banks[0]++;
nkeynes@284:     
nkeynes@284:     dwsrc = (uint32_t *)src;
nkeynes@284:     for( i=0; i<line_count; i++ ) {
nkeynes@284: 	for( j=0; j<line_bytes; j++ ) {
nkeynes@284: 	    *banks[bank_flag]++ = *dwsrc++;
nkeynes@284: 	    bank_flag = !bank_flag;
nkeynes@284: 	}
nkeynes@285: 	banks[0] += line_gap;
nkeynes@285: 	banks[1] += line_gap;
nkeynes@284: 	if( line_gap_flag ) {
nkeynes@285: 	    banks[bank_flag]++;
nkeynes@284: 	    bank_flag = !bank_flag;
nkeynes@284: 	}
nkeynes@284:     }    
nkeynes@284: }
nkeynes@284: 
nkeynes@284: /**
nkeynes@284:  * Read an image from 64-bit vram, with a destination line-stride different from the line-size.
nkeynes@284:  * The srcaddr must be 32-bit aligned, and both line_bytes and line_stride_bytes
nkeynes@284:  * must be multiples of 4. line_stride_bytes must be >= line_bytes.
nkeynes@284:  * This method is used to extract a "stride" texture from vram.
nkeynes@284:  */
nkeynes@284: void pvr2_vram64_read_stride( char *dest, uint32_t dest_line_bytes, sh4addr_t srcaddr,
nkeynes@284: 				   uint32_t src_line_bytes, uint32_t line_count )
nkeynes@284: {
nkeynes@284:     int bank_flag = (srcaddr & 0x04) >> 2;
nkeynes@284:     uint32_t *banks[2];
nkeynes@284:     uint32_t *dwdest;
nkeynes@284:     uint32_t dest_line_gap;
nkeynes@284:     uint32_t src_line_gap;
nkeynes@284:     uint32_t line_bytes;
nkeynes@284:     int src_line_gap_flag;
nkeynes@284:     int i,j;
nkeynes@284: 
nkeynes@284:     srcaddr = srcaddr & 0x7FFFF8;
nkeynes@284:     if( src_line_bytes <= dest_line_bytes ) {
nkeynes@284: 	dest_line_gap = (dest_line_bytes - src_line_bytes) >> 2;
nkeynes@284: 	src_line_gap = 0;
nkeynes@284: 	src_line_gap_flag = 0;
nkeynes@284: 	line_bytes = src_line_bytes >> 2;
nkeynes@284:     } else {
nkeynes@284: 	i = (src_line_bytes - dest_line_bytes);
nkeynes@284: 	src_line_gap_flag = i & 0x04;
nkeynes@284: 	src_line_gap = i >> 3;
nkeynes@284: 	line_bytes = dest_line_bytes >> 2;
nkeynes@284:     }
nkeynes@284: 	
nkeynes@284:     banks[0] = (uint32_t *)(video_base + (srcaddr>>1));
nkeynes@284:     banks[1] = banks[0] + 0x100000;
nkeynes@284:     if( bank_flag )
nkeynes@284: 	banks[0]++;
nkeynes@284:     
nkeynes@284:     dwdest = (uint32_t *)dest;
nkeynes@284:     for( i=0; i<line_count; i++ ) {
nkeynes@284: 	for( j=0; j<line_bytes; j++ ) {
nkeynes@284: 	    *dwdest++ = *banks[bank_flag]++;
nkeynes@284: 	    bank_flag = !bank_flag;
nkeynes@284: 	}
nkeynes@284: 	dwdest += dest_line_gap;
nkeynes@284: 	banks[0] += src_line_gap;
nkeynes@284: 	banks[1] += src_line_gap;
nkeynes@284: 	if( src_line_gap_flag ) {
nkeynes@284: 	    banks[bank_flag]++;
nkeynes@284: 	    bank_flag = !bank_flag;
nkeynes@284: 	}
nkeynes@284:     }
nkeynes@284:     
nkeynes@284: }
nkeynes@284: 
nkeynes@284: void pvr2_vram_write_invert( sh4addr_t destaddr, char *src, uint32_t length, uint32_t line_length )
nkeynes@284: {
nkeynes@284:     char *dest = video_base + (destaddr & 0x007FFFFF);
nkeynes@284:     char *p = src + length - line_length;
nkeynes@284:     while( p >= src ) {
nkeynes@284: 	memcpy( dest, p, line_length );
nkeynes@284: 	p -= line_length;
nkeynes@284: 	dest += line_length;
nkeynes@284:     }
nkeynes@284: }
nkeynes@284: 
nkeynes@284: void pvr2_vram64_read( char *dest, sh4addr_t srcaddr, uint32_t length )
nkeynes@284: {
nkeynes@284:     int bank_flag = (srcaddr & 0x04) >> 2;
nkeynes@284:     uint32_t *banks[2];
nkeynes@284:     uint32_t *dwdest;
nkeynes@284:     int i;
nkeynes@284: 
nkeynes@284:     srcaddr = srcaddr & 0x7FFFFF;
nkeynes@284:     if( srcaddr + length > 0x800000 )
nkeynes@284: 	length = 0x800000 - srcaddr;
nkeynes@284: 
nkeynes@284:     banks[0] = ((uint32_t *)(video_base + ((srcaddr&0x007FFFF8)>>1)));
nkeynes@284:     banks[1] = banks[0] + 0x100000;
nkeynes@284:     if( bank_flag )
nkeynes@284: 	banks[0]++;
nkeynes@284:     
nkeynes@284:     /* Handle non-aligned start of source */
nkeynes@284:     if( srcaddr & 0x03 ) {
nkeynes@284: 	char *src = ((char *)banks[bank_flag]) + (srcaddr & 0x03);
nkeynes@284: 	for( i= srcaddr & 0x03; i < 4 && length > 0; i++, length-- ) {
nkeynes@284: 	    *dest++ = *src++;
nkeynes@284: 	}
nkeynes@284: 	bank_flag = !bank_flag;
nkeynes@284:     }
nkeynes@284: 
nkeynes@284:     dwdest = (uint32_t *)dest;
nkeynes@284:     while( length >= 4 ) {
nkeynes@284: 	*dwdest++ = *banks[bank_flag]++;
nkeynes@284: 	bank_flag = !bank_flag;
nkeynes@284: 	length -= 4;
nkeynes@284:     }
nkeynes@284:     
nkeynes@284:     /* Handle non-aligned end of source */
nkeynes@284:     if( length ) {
nkeynes@284: 	dest = (char *)dwdest;
nkeynes@284: 	char *src = (char *)banks[bank_flag];
nkeynes@284: 	while( length-- > 0 ) {
nkeynes@284: 	    *dest++ = *src++;
nkeynes@284: 	}
nkeynes@284:     }
nkeynes@284: }
nkeynes@284: 
nkeynes@284: void pvr2_vram64_dump( sh4addr_t addr, uint32_t length, FILE *f ) 
nkeynes@284: {
nkeynes@284:     char tmp[length];
nkeynes@284:     pvr2_vram64_read( tmp, addr, length );
nkeynes@284:     fwrite_dump( tmp, length, f );
nkeynes@284: }