nkeynes@284: /**
nkeynes@333:  * $Id: pvr2mem.c,v 1.8 2007-01-27 06:21:35 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@325: #include "asic.h"
nkeynes@309: #include <stdio.h>
nkeynes@309: #include <errno.h>
nkeynes@284: 
nkeynes@284: extern char *video_base;
nkeynes@284: 
nkeynes@325: void pvr2_dma_write( sh4addr_t destaddr, char *src, uint32_t count )
nkeynes@325: {
nkeynes@325:     int region;
nkeynes@325: 
nkeynes@325:     switch( destaddr & 0x13800000 ) {
nkeynes@325:     case 0x10000000:
nkeynes@325:     case 0x12000000:
nkeynes@325: 	pvr2_ta_write( src, count );
nkeynes@325: 	break;
nkeynes@325:     case 0x11000000:
nkeynes@325:     case 0x11800000:
nkeynes@325: 	region = MMIO_READ( ASIC, PVRDMARGN1 );
nkeynes@325: 	if( region == 0 ) {
nkeynes@325: 	    pvr2_vram64_write( destaddr, src, count );
nkeynes@325: 	} else {
nkeynes@325: 	    char *dest = mem_get_region(destaddr);
nkeynes@325: 	    memcpy( dest, src, count );
nkeynes@325: 	}
nkeynes@325: 	break;
nkeynes@325:     case 0x10800000:
nkeynes@325:     case 0x12800000:
nkeynes@325: 	pvr2_yuv_write( src, count );
nkeynes@325: 	break;
nkeynes@325:     case 0x13000000:
nkeynes@325:     case 0x13800000:
nkeynes@325: 	region = MMIO_READ( ASIC, PVRDMARGN2 );
nkeynes@325: 	if( region == 0 ) {
nkeynes@325: 	    pvr2_vram64_write( destaddr, src, count );
nkeynes@325: 	} else {
nkeynes@325: 	    char *dest = mem_get_region(destaddr);
nkeynes@325: 	    memcpy( dest, src, count );
nkeynes@325: 	}
nkeynes@325:     }	    
nkeynes@325: }
nkeynes@325: 
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@310:     }    
nkeynes@310: }
nkeynes@310: 
nkeynes@315: 
nkeynes@310: /**
nkeynes@310:  * @param dest Destination image buffer
nkeynes@310:  * @param banks Source data expressed as two bank pointers
nkeynes@310:  * @param offset Offset into banks[0] specifying where the next byte
nkeynes@310:  *  to read is (0..3)
nkeynes@310:  * @param x1,y1 Destination coordinates
nkeynes@310:  * @param width Width of current destination block
nkeynes@315:  * @param stride Total width of image (ie stride) in bytes
nkeynes@315:  */
nkeynes@315: 
nkeynes@315: static void pvr2_vram64_detwiddle_4( uint8_t *dest, uint8_t *banks[2], int offset,
nkeynes@315: 				     int x1, int y1, int width, int stride )
nkeynes@315: {
nkeynes@315:     if( width == 2 ) {
nkeynes@315: 	x1 = x1 >> 1;
nkeynes@315: 	uint8_t t1 = *banks[offset<4?0:1]++;
nkeynes@315: 	uint8_t t2 = *banks[offset<3?0:1]++;
nkeynes@315: 	dest[y1*stride + x1] = (t1 & 0x0F) | (t2<<4);
nkeynes@315: 	dest[(y1+1)*stride + x1] = (t1>>4) | (t2&0xF0);
nkeynes@315:     } else if( width == 4 ) {
nkeynes@315: 	pvr2_vram64_detwiddle_4( dest, banks, offset, x1, y1, 2, stride );
nkeynes@315: 	pvr2_vram64_detwiddle_4( dest, banks, offset+2, x1, y1+2, 2, stride );
nkeynes@315: 	pvr2_vram64_detwiddle_4( dest, banks, offset+4, x1+2, y1, 2, stride );
nkeynes@315: 	pvr2_vram64_detwiddle_4( dest, banks, offset+6, x1+2, y1+2, 2, stride );
nkeynes@315: 	
nkeynes@315:     } else {
nkeynes@315: 	int subdivide = width >> 1;
nkeynes@315: 	pvr2_vram64_detwiddle_4( dest, banks, offset, x1, y1, subdivide, stride );
nkeynes@315: 	pvr2_vram64_detwiddle_4( dest, banks, offset, x1, y1+subdivide, subdivide, stride );
nkeynes@315: 	pvr2_vram64_detwiddle_4( dest, banks, offset, x1+subdivide, y1, subdivide, stride );
nkeynes@315: 	pvr2_vram64_detwiddle_4( dest, banks, offset, x1+subdivide, y1+subdivide, subdivide, stride );
nkeynes@315:     }
nkeynes@315: }
nkeynes@315: 
nkeynes@315: /**
nkeynes@315:  * @param dest Destination image buffer
nkeynes@315:  * @param banks Source data expressed as two bank pointers
nkeynes@315:  * @param offset Offset into banks[0] specifying where the next byte
nkeynes@315:  *  to read is (0..3)
nkeynes@315:  * @param x1,y1 Destination coordinates
nkeynes@315:  * @param width Width of current destination block
nkeynes@315:  * @param stride Total width of image (ie stride)
nkeynes@310:  */
nkeynes@310: 
nkeynes@310: static void pvr2_vram64_detwiddle_8( uint8_t *dest, uint8_t *banks[2], int offset,
nkeynes@315: 				     int x1, int y1, int width, int stride )
nkeynes@310: {
nkeynes@310:     if( width == 2 ) {
nkeynes@315: 	dest[y1*stride + x1] = *banks[0]++;
nkeynes@315: 	dest[(y1+1)*stride + x1] = *banks[offset<3?0:1]++;
nkeynes@315: 	dest[y1*stride + x1 + 1] = *banks[offset<2?0:1]++;
nkeynes@315: 	dest[(y1+1)*stride + x1 + 1] = *banks[offset==0?0:1]++;
nkeynes@310: 	uint8_t *tmp = banks[0]; /* swap banks */
nkeynes@310: 	banks[0] = banks[1];
nkeynes@310: 	banks[1] = tmp;
nkeynes@310:     } else {
nkeynes@310: 	int subdivide = width >> 1;
nkeynes@315: 	pvr2_vram64_detwiddle_8( dest, banks, offset, x1, y1, subdivide, stride );
nkeynes@315: 	pvr2_vram64_detwiddle_8( dest, banks, offset, x1, y1+subdivide, subdivide, stride );
nkeynes@315: 	pvr2_vram64_detwiddle_8( dest, banks, offset, x1+subdivide, y1, subdivide, stride );
nkeynes@315: 	pvr2_vram64_detwiddle_8( dest, banks, offset, x1+subdivide, y1+subdivide, subdivide, stride );
nkeynes@284:     }
nkeynes@310: }
nkeynes@310: 
nkeynes@310: /**
nkeynes@310:  * @param dest Destination image buffer
nkeynes@310:  * @param banks Source data expressed as two bank pointers
nkeynes@310:  * @param offset Offset into banks[0] specifying where the next word
nkeynes@310:  *  to read is (0 or 1)
nkeynes@310:  * @param x1,y1 Destination coordinates
nkeynes@310:  * @param width Width of current destination block
nkeynes@315:  * @param stride Total width of image (ie stride)
nkeynes@310:  */
nkeynes@310: 
nkeynes@310: static void pvr2_vram64_detwiddle_16( uint16_t *dest, uint16_t *banks[2], int offset,
nkeynes@315: 				      int x1, int y1, int width, int stride )
nkeynes@310: {
nkeynes@310:     if( width == 2 ) {
nkeynes@315: 	dest[y1*stride + x1] = *banks[0]++;
nkeynes@315: 	dest[(y1+1)*stride + x1] = *banks[offset]++;
nkeynes@315: 	dest[y1*stride + x1 + 1] = *banks[1]++;
nkeynes@315: 	dest[(y1+1)*stride + x1 + 1] = *banks[offset^1]++;
nkeynes@310:     } else {
nkeynes@310: 	int subdivide = width >> 1;
nkeynes@315: 	pvr2_vram64_detwiddle_16( dest, banks, offset, x1, y1, subdivide, stride );
nkeynes@315: 	pvr2_vram64_detwiddle_16( dest, banks, offset, x1, y1+subdivide, subdivide, stride );
nkeynes@315: 	pvr2_vram64_detwiddle_16( dest, banks, offset, x1+subdivide, y1, subdivide, stride );
nkeynes@315: 	pvr2_vram64_detwiddle_16( dest, banks, offset, x1+subdivide, y1+subdivide, subdivide, stride );
nkeynes@310:     }
nkeynes@310: }
nkeynes@310: 
nkeynes@310: /**
nkeynes@315:  * Read an image from 64-bit vram stored as twiddled 4-bit pixels. The 
nkeynes@315:  * image is written out to the destination in detwiddled form.
nkeynes@315:  * @param dest destination buffer, which must be at least width*height/2 in length
nkeynes@315:  * @param srcaddr source address in vram
nkeynes@315:  * @param width image width (must be a power of 2)
nkeynes@315:  * @param height image height (must be a power of 2)
nkeynes@315:  */
nkeynes@315: void pvr2_vram64_read_twiddled_4( char *dest, sh4addr_t srcaddr, uint32_t width, uint32_t height )
nkeynes@315: {
nkeynes@315:     int offset_flag = (srcaddr & 0x07);
nkeynes@315:     uint8_t *banks[2];
nkeynes@315:     uint8_t *wdest = (uint8_t*)dest;
nkeynes@315:     uint32_t stride = width >> 1;
nkeynes@315:     int i,j;
nkeynes@315: 
nkeynes@315:     srcaddr = srcaddr & 0x7FFFF8;
nkeynes@315: 
nkeynes@315:     banks[0] = (uint8_t *)(video_base + (srcaddr>>1));
nkeynes@315:     banks[1] = banks[0] + 0x400000;
nkeynes@315:     if( offset_flag & 0x04 ) { // If source is not 64-bit aligned, swap the banks
nkeynes@315: 	uint8_t *tmp = banks[0];
nkeynes@315: 	banks[0] = banks[1];
nkeynes@315: 	banks[1] = tmp + 4;
nkeynes@315: 	offset_flag &= 0x03;
nkeynes@315:     }
nkeynes@315:     banks[0] += offset_flag;
nkeynes@315: 
nkeynes@315:     if( width > height ) {
nkeynes@315: 	for( i=0; i<width; i+=height ) {
nkeynes@315: 	    pvr2_vram64_detwiddle_4( wdest, banks, offset_flag, i, 0, height, stride );
nkeynes@315: 	}
nkeynes@315:     } else if( height > width ) {
nkeynes@315: 	for( i=0; i<height; i+=width ) {
nkeynes@315: 	    pvr2_vram64_detwiddle_4( wdest, banks, offset_flag, 0, i, width, stride );
nkeynes@315: 	}
nkeynes@315:     } else if( width == 1 ) {
nkeynes@315: 	*wdest = *banks[0];
nkeynes@315:     } else {
nkeynes@315: 	pvr2_vram64_detwiddle_4( wdest, banks, offset_flag, 0, 0, width, stride );
nkeynes@315:     }   
nkeynes@315: }
nkeynes@315: 
nkeynes@315: /**
nkeynes@310:  * Read an image from 64-bit vram stored as twiddled 8-bit pixels. The 
nkeynes@310:  * image is written out to the destination in detwiddled form.
nkeynes@310:  * @param dest destination buffer, which must be at least width*height in length
nkeynes@310:  * @param srcaddr source address in vram
nkeynes@310:  * @param width image width (must be a power of 2)
nkeynes@310:  * @param height image height (must be a power of 2)
nkeynes@310:  */
nkeynes@310: void pvr2_vram64_read_twiddled_8( char *dest, sh4addr_t srcaddr, uint32_t width, uint32_t height )
nkeynes@310: {
nkeynes@310:     int offset_flag = (srcaddr & 0x07);
nkeynes@310:     uint8_t *banks[2];
nkeynes@310:     uint8_t *wdest = (uint8_t*)dest;
nkeynes@310:     int i,j;
nkeynes@310: 
nkeynes@310:     srcaddr = srcaddr & 0x7FFFF8;
nkeynes@310: 
nkeynes@310:     banks[0] = (uint8_t *)(video_base + (srcaddr>>1));
nkeynes@310:     banks[1] = banks[0] + 0x400000;
nkeynes@310:     if( offset_flag & 0x04 ) { // If source is not 64-bit aligned, swap the banks
nkeynes@310: 	uint8_t *tmp = banks[0];
nkeynes@310: 	banks[0] = banks[1];
nkeynes@310: 	banks[1] = tmp + 4;
nkeynes@310: 	offset_flag &= 0x03;
nkeynes@310:     }
nkeynes@310:     banks[0] += offset_flag;
nkeynes@310: 
nkeynes@310:     if( width > height ) {
nkeynes@310: 	for( i=0; i<width; i+=height ) {
nkeynes@310: 	    pvr2_vram64_detwiddle_8( wdest, banks, offset_flag, i, 0, height, width );
nkeynes@310: 	}
nkeynes@310:     } else if( height > width ) {
nkeynes@310: 	for( i=0; i<height; i+=width ) {
nkeynes@310: 	    pvr2_vram64_detwiddle_8( wdest, banks, offset_flag, 0, i, width, width );
nkeynes@310: 	}
nkeynes@310:     } else if( width == 1 ) {
nkeynes@310: 	*wdest = *banks[0];
nkeynes@310:     } else {
nkeynes@310: 	pvr2_vram64_detwiddle_8( wdest, banks, offset_flag, 0, 0, width, width );
nkeynes@310:     }   
nkeynes@310: }
nkeynes@310: 
nkeynes@310: /**
nkeynes@310:  * Read an image from 64-bit vram stored as twiddled 16-bit pixels. The 
nkeynes@310:  * image is written out to the destination in detwiddled form.
nkeynes@310:  * @param dest destination buffer, which must be at least width*height*2 in length
nkeynes@310:  * @param srcaddr source address in vram (must be 16-bit aligned)
nkeynes@310:  * @param width image width (must be a power of 2)
nkeynes@310:  * @param height image height (must be a power of 2)
nkeynes@310:  */
nkeynes@310: void pvr2_vram64_read_twiddled_16( char *dest, sh4addr_t srcaddr, uint32_t width, uint32_t height ) {
nkeynes@310:     int offset_flag = (srcaddr & 0x06) >> 1;
nkeynes@310:     uint16_t *banks[2];
nkeynes@310:     uint16_t *wdest = (uint16_t*)dest;
nkeynes@310:     int i,j;
nkeynes@310: 
nkeynes@310:     srcaddr = srcaddr & 0x7FFFF8;
nkeynes@310: 
nkeynes@310:     banks[0] = (uint16_t *)(video_base + (srcaddr>>1));
nkeynes@310:     banks[1] = banks[0] + 0x200000;
nkeynes@310:     if( offset_flag & 0x02 ) { // If source is not 64-bit aligned, swap the banks
nkeynes@310: 	uint16_t *tmp = banks[0];
nkeynes@310: 	banks[0] = banks[1];
nkeynes@310: 	banks[1] = tmp + 2;
nkeynes@310: 	offset_flag &= 0x01;
nkeynes@310:     }
nkeynes@310:     banks[0] += offset_flag;
nkeynes@310: 	
nkeynes@310: 
nkeynes@310:     if( width > height ) {
nkeynes@310: 	for( i=0; i<width; i+=height ) {
nkeynes@310: 	    pvr2_vram64_detwiddle_16( wdest, banks, offset_flag, i, 0, height, width );
nkeynes@310: 	}
nkeynes@310:     } else if( height > width ) {
nkeynes@310: 	for( i=0; i<height; i+=width ) {
nkeynes@310: 	    pvr2_vram64_detwiddle_16( wdest, banks, offset_flag, 0, i, width, width );
nkeynes@310: 	}
nkeynes@310:     } else if( width == 1 ) {
nkeynes@310: 	*wdest = *banks[0];
nkeynes@310:     } else {
nkeynes@310: 	pvr2_vram64_detwiddle_16( wdest, banks, offset_flag, 0, 0, width, width );
nkeynes@310:     }    
nkeynes@284: }
nkeynes@284: 
nkeynes@333: void pvr2_vram_write_invert( sh4addr_t destaddr, char *src, uint32_t length, uint32_t line_length,
nkeynes@333: 			     uint32_t src_stride )
nkeynes@284: {
nkeynes@284:     char *dest = video_base + (destaddr & 0x007FFFFF);
nkeynes@333:     char *p = src + length - src_stride;
nkeynes@284:     while( p >= src ) {
nkeynes@284: 	memcpy( dest, p, line_length );
nkeynes@333: 	p -= src_stride;
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@309: void pvr2_vram64_dump_file( sh4addr_t addr, uint32_t length, gchar *filename )
nkeynes@309: {
nkeynes@309:     uint32_t tmp[length>>2];
nkeynes@309:     FILE *f = fopen(filename, "wo");
nkeynes@309:     unsigned int i, j;
nkeynes@309: 
nkeynes@309:     if( f == NULL ) {
nkeynes@309: 	ERROR( "Unable to write to dump file '%s' (%s)", filename, strerror(errno) );
nkeynes@309: 	return;
nkeynes@309:     }
nkeynes@309:     pvr2_vram64_read( tmp, addr, length );
nkeynes@309:     fprintf( f, "%08X\n", addr );
nkeynes@309:     for( i =0; i<length>>2; i+=8 ) {
nkeynes@309: 	for( j=i; j<i+8; j++ ) {
nkeynes@309: 	    if( j < length )
nkeynes@309: 		fprintf( f, " %08X", tmp[j] );
nkeynes@309: 	    else
nkeynes@309: 		fprintf( f, "         " );
nkeynes@309: 	}
nkeynes@309: 	fprintf( f, "\n" );
nkeynes@309:     }
nkeynes@309:     fclose(f);
nkeynes@309: }
nkeynes@309: 
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: }
nkeynes@315: 
nkeynes@315: 
nkeynes@315: 
nkeynes@315: /**
nkeynes@315:  * Flush the indicated render buffer back to PVR. Caller is responsible for
nkeynes@315:  * tracking whether there is actually anything in the buffer.
nkeynes@315:  *
nkeynes@315:  * @param buffer A render buffer indicating the address to store to, and the
nkeynes@315:  * format the data needs to be in.
nkeynes@315:  * @param backBuffer TRUE to flush the back buffer, FALSE for 
nkeynes@315:  * the front buffer.
nkeynes@315:  */
nkeynes@315: void pvr2_render_buffer_copy_to_sh4( pvr2_render_buffer_t buffer, 
nkeynes@315: 				     gboolean backBuffer )
nkeynes@315: {
nkeynes@315:     if( buffer->render_addr == -1 )
nkeynes@315: 	return;
nkeynes@327:     GLenum type = colour_formats[buffer->colour_format].type;
nkeynes@327:     GLenum format = colour_formats[buffer->colour_format].format;
nkeynes@327:     int line_size = buffer->width * colour_formats[buffer->colour_format].bpp;
nkeynes@327:     int size = line_size * buffer->height;
nkeynes@315:     
nkeynes@315:     if( backBuffer ) {
nkeynes@315: 	glFinish();
nkeynes@315: 	glReadBuffer( GL_BACK );
nkeynes@315:     } else {
nkeynes@315: 	glReadBuffer( GL_FRONT );
nkeynes@315:     }
nkeynes@315: 
nkeynes@315:     if( buffer->render_addr & 0xFF000000 == 0x04000000 ) {
nkeynes@315: 	/* Interlaced buffer. Go the double copy... :( */
nkeynes@315: 	char target[size];
nkeynes@315: 	glReadPixels( 0, 0, buffer->width, buffer->height, format, type, target );
nkeynes@315: 	pvr2_vram64_write( buffer->render_addr, target, size );
nkeynes@315:     } else {
nkeynes@315: 	/* Regular buffer */
nkeynes@315: 	char target[size];
nkeynes@315: 	glReadPixels( 0, 0, buffer->width, buffer->height, format, type, target );
nkeynes@333: 	if( (buffer->scale & 0xFFFF) == 0x0800 ) {
nkeynes@333: 	    pvr2_vram_write_invert( buffer->render_addr, target, size, line_size, line_size << 1 );
nkeynes@333: 	} else {
nkeynes@333: 	    pvr2_vram_write_invert( buffer->render_addr, target, size, line_size, line_size );
nkeynes@333: 	}
nkeynes@315:     }
nkeynes@315: }
nkeynes@315: 
nkeynes@315: 
nkeynes@315: /**
nkeynes@315:  * Copy data from PVR ram into the GL render buffer. 
nkeynes@315:  *
nkeynes@315:  * @param buffer A render buffer indicating the address to read from, and the
nkeynes@315:  * format the data is in.
nkeynes@315:  * @param backBuffer TRUE to write the back buffer, FALSE for 
nkeynes@315:  * the front buffer.
nkeynes@315:  */
nkeynes@315: void pvr2_render_buffer_copy_from_sh4( pvr2_render_buffer_t buffer, 
nkeynes@315: 				       gboolean backBuffer )
nkeynes@315: {
nkeynes@315:     if( buffer->render_addr == -1 )
nkeynes@315: 	return;
nkeynes@315: 
nkeynes@327:     GLenum type = colour_formats[buffer->colour_format].type;
nkeynes@327:     GLenum format = colour_formats[buffer->colour_format].format;
nkeynes@327:     int line_size = buffer->width * colour_formats[buffer->colour_format].bpp;
nkeynes@327:     int size = line_size * buffer->height;
nkeynes@327: 
nkeynes@315:     if( backBuffer ) {
nkeynes@315: 	glDrawBuffer( GL_BACK );
nkeynes@315:     } else {
nkeynes@315: 	glDrawBuffer( GL_FRONT );
nkeynes@315:     }
nkeynes@315: 
nkeynes@315:     glRasterPos2i( 0, 0 );
nkeynes@315:     if( buffer->render_addr & 0xFF000000 == 0x04000000 ) {
nkeynes@315: 	/* Interlaced buffer. Go the double copy... :( */
nkeynes@315: 	char target[size];
nkeynes@315: 	pvr2_vram64_read( target, buffer->render_addr, size );
nkeynes@315: 	glDrawPixels( buffer->width, buffer->height, 
nkeynes@315: 		      format, type, target );
nkeynes@315:     } else {
nkeynes@315: 	/* Regular buffer - go direct */
nkeynes@315: 	char *target = mem_get_region( buffer->render_addr );
nkeynes@315: 	glDrawPixels( buffer->width, buffer->height, 
nkeynes@315: 		      format, type, target );
nkeynes@315:     }
nkeynes@315: }