2 * $Id: texcache.c,v 1.25 2007-02-05 08:52:59 nkeynes Exp $
4 * Texture cache. Responsible for maintaining a working set of OpenGL
8 * Copyright (c) 2005 Nathan Keynes.
10 * This program is free software; you can redistribute it and/or modify
11 * it under the terms of the GNU General Public License as published by
12 * the Free Software Foundation; either version 2 of the License, or
13 * (at your option) any later version.
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18 * GNU General Public License for more details.
22 #include "pvr2/pvr2.h"
24 /** Specifies the maximum number of OpenGL
25 * textures we're willing to have open at a time. If more are
26 * needed, textures will be evicted in LRU order.
28 #define MAX_TEXTURES 256
34 * find entry by texture_addr
36 * move entry to tail of lru list
40 typedef signed short texcache_entry_index;
41 #define EMPTY_ENTRY 0xFF
43 static texcache_entry_index texcache_free_ptr = 0;
44 static GLuint texcache_free_list[MAX_TEXTURES];
46 typedef struct texcache_entry {
47 uint32_t texture_addr;
48 int width, height, mode;
50 texcache_entry_index next;
54 static uint8_t texcache_page_lookup[PVR2_RAM_PAGES];
55 static uint32_t texcache_ref_counter;
56 static struct texcache_entry texcache_active_list[MAX_TEXTURES];
59 * Initialize the texture cache.
64 for( i=0; i<PVR2_RAM_PAGES; i++ ) {
65 texcache_page_lookup[i] = EMPTY_ENTRY;
67 for( i=0; i<MAX_TEXTURES; i++ ) {
68 texcache_free_list[i] = i;
69 texcache_active_list[i].texture_addr = -1;
71 texcache_free_ptr = 0;
72 texcache_ref_counter = 0;
76 * Setup the initial texture ids (must be called after the GL context is
79 void texcache_gl_init( )
82 GLuint texids[MAX_TEXTURES];
84 glGenTextures( MAX_TEXTURES, texids );
85 for( i=0; i<MAX_TEXTURES; i++ ) {
86 texcache_active_list[i].texture_id = texids[i];
91 * Flush all textures from the cache, returning them to the free list.
93 void texcache_flush( )
96 /* clear structures */
97 for( i=0; i<PVR2_RAM_PAGES; i++ ) {
98 texcache_page_lookup[i] = EMPTY_ENTRY;
100 for( i=0; i<MAX_TEXTURES; i++ ) {
101 texcache_free_list[i] = i;
103 texcache_free_ptr = 0;
104 texcache_ref_counter = 0;
108 * Flush all textures and delete. The cache will be non-functional until
109 * the next call to texcache_init(). This would typically be done if
110 * switching GL targets.
112 void texcache_shutdown( )
114 GLuint texids[MAX_TEXTURES];
118 for( i=0; i<MAX_TEXTURES; i++ ) {
119 texids[i] = texcache_active_list[i].texture_id;
121 glDeleteTextures( MAX_TEXTURES, texids );
124 static void texcache_evict( int slot )
126 /* Remove the selected slot from the lookup table */
127 uint32_t evict_page = texcache_active_list[slot].texture_addr >> 12;
128 texcache_entry_index replace_next = texcache_active_list[slot].next;
129 texcache_active_list[slot].texture_addr = -1;
130 texcache_active_list[slot].next = EMPTY_ENTRY; /* Just for safety */
131 if( texcache_page_lookup[evict_page] == slot ) {
132 texcache_page_lookup[evict_page] = replace_next;
134 texcache_entry_index idx = texcache_page_lookup[evict_page];
135 texcache_entry_index next;
137 next = texcache_active_list[idx].next;
139 texcache_active_list[idx].next = replace_next;
143 } while( next != EMPTY_ENTRY );
148 * Evict a single texture from the cache.
149 * @return the slot of the evicted texture.
151 static texcache_entry_index texcache_evict_lru( void )
153 /* Full table scan - take over the entry with the lowest lru value */
154 texcache_entry_index slot = 0;
155 int lru_value = texcache_active_list[0].lru_count;
157 for( i=1; i<MAX_TEXTURES; i++ ) {
158 /* FIXME: account for rollover */
159 if( texcache_active_list[i].lru_count < lru_value ) {
161 lru_value = texcache_active_list[i].lru_count;
164 texcache_evict(slot);
170 * Evict all textures contained in the page identified by a texture address.
172 void texcache_invalidate_page( uint32_t texture_addr ) {
173 uint32_t texture_page = texture_addr >> 12;
174 texcache_entry_index idx = texcache_page_lookup[texture_page];
175 if( idx == EMPTY_ENTRY )
177 assert( texcache_free_ptr >= 0 );
179 texcache_entry_t entry = &texcache_active_list[idx];
180 entry->texture_addr = -1;
183 texcache_free_list[texcache_free_ptr] = idx;
185 entry->next = EMPTY_ENTRY;
186 } while( idx != EMPTY_ENTRY );
187 texcache_page_lookup[texture_page] = EMPTY_ENTRY;
191 * Mark all textures that use the palette table as needing a re-read (ie
192 * for when the palette is changed. We could track exactly which ones are
193 * affected, but it's not clear that the extra maintanence overhead is
196 void texcache_invalidate_palette( )
199 for( i=0; i<MAX_TEXTURES; i++ ) {
200 if( texcache_active_list[i].texture_addr != -1 &&
201 PVR2_TEX_IS_PALETTE(texcache_active_list[i].mode) ) {
207 static void decode_pal8_to_32( uint32_t *out, uint8_t *in, int inbytes, uint32_t *pal )
210 for( i=0; i<inbytes; i++ ) {
215 static void decode_pal8_to_16( uint16_t *out, uint8_t *in, int inbytes, uint32_t *pal )
218 for( i=0; i<inbytes; i++ ) {
219 *out++ = (uint16_t)pal[*in++];
223 static void decode_pal4_to_32( uint32_t *out, uint8_t *in, int inbytes, uint32_t *pal )
226 for( i=0; i<inbytes; i++ ) {
227 *out++ = pal[*in & 0x0F];
228 *out++ = pal[(*in >> 4)];
234 static void decode_pal4_to_16( uint16_t *out, uint8_t *in, int inbytes, uint32_t *pal )
237 for( i=0; i<inbytes; i++ ) {
238 *out++ = (uint16_t)pal[*in & 0x0F];
239 *out++ = (uint16_t)pal[(*in >> 4)];
244 #define VQ_CODEBOOK_SIZE 2048 /* 256 entries * 4 pixels per quad * 2 byte pixels */
247 uint16_t quad[256][4];
250 static void vq_get_codebook( struct vq_codebook *codebook,
253 /* Detwiddle the codebook, for the sake of my own sanity if nothing else */
254 uint16_t *p = (uint16_t *)input;
256 for( i=0; i<256; i++ ) {
257 codebook->quad[i][0] = *p++;
258 codebook->quad[i][2] = *p++;
259 codebook->quad[i][1] = *p++;
260 codebook->quad[i][3] = *p++;
264 static void vq_decode( uint16_t *output, char *input, int width, int height,
265 struct vq_codebook *codebook ) {
268 uint8_t *c = (uint8_t *)input;
269 for( j=0; j<height; j+=2 ) {
270 for( i=0; i<width; i+=2 ) {
272 output[i + j*width] = codebook->quad[code][0];
273 output[i + 1 + j*width] = codebook->quad[code][1];
274 output[i + (j+1)*width] = codebook->quad[code][2];
275 output[i + 1 + (j+1)*width] = codebook->quad[code][3];
280 static inline uint32_t yuv_to_rgb32( float y, float u, float v )
284 int r = (int)(y + v*1.375);
285 int g = (int)(y - u*0.34375 - v*0.6875);
286 int b = (int)(y + u*1.71875);
287 if( r > 255 ) { r = 255; } else if( r < 0 ) { r = 0; }
288 if( g > 255 ) { g = 255; } else if( g < 0 ) { g = 0; }
289 if( b > 255 ) { b = 255; } else if( b < 0 ) { b = 0; }
290 return 0xFF000000 | (r<<16) | (g<<8) | (b);
295 * Convert raster YUV texture data into RGB32 data - most GL implementations don't
296 * directly support this format unfortunately. The input data is formatted as
297 * 32 bits = 2 horizontal pixels, UYVY. This is currently done rather inefficiently
300 static void yuv_decode( uint32_t *output, uint32_t *input, int width, int height )
304 for( y=0; y<height; y++ ) {
305 for( x=0; x<width; x+=2 ) {
306 float u = (float)(*p & 0xFF);
307 float y0 = (float)( (*p>>8)&0xFF );
308 float v = (float)( (*p>>16)&0xFF );
309 float y1 = (float)( (*p>>24)&0xFF );
310 *output++ = yuv_to_rgb32( y0, u, v );
311 *output++ = yuv_to_rgb32( y1, u, v );
318 * Load texture data from the given address and parameters into the currently
319 * bound OpenGL texture.
321 static texcache_load_texture( uint32_t texture_addr, int width, int height,
323 int bpp_shift = 1; /* bytes per (output) pixel as a power of 2 */
324 GLint intFormat = GL_RGBA, format, type;
325 int tex_format = mode & PVR2_TEX_FORMAT_MASK;
326 struct vq_codebook codebook;
327 GLint filter = GL_LINEAR;
329 /* Decode the format parameters */
330 switch( tex_format ) {
331 case PVR2_TEX_FORMAT_IDX4:
332 case PVR2_TEX_FORMAT_IDX8:
333 /* For indexed-colour modes, we need to lookup the palette control
334 * word to determine the de-indexed texture format.
336 switch( MMIO_READ( PVR2, RENDER_PALETTE ) & 0x03 ) {
337 case 0: /* ARGB1555 */
339 type = GL_UNSIGNED_SHORT_1_5_5_5_REV;
344 type = GL_UNSIGNED_SHORT_5_6_5;
346 case 2: /* ARGB4444 */
348 type = GL_UNSIGNED_SHORT_4_4_4_4_REV;
350 case 3: /* ARGB8888 */
352 type = GL_UNSIGNED_INT_8_8_8_8_REV;
358 case PVR2_TEX_FORMAT_ARGB1555:
360 type = GL_UNSIGNED_SHORT_1_5_5_5_REV;
362 case PVR2_TEX_FORMAT_RGB565:
365 type = GL_UNSIGNED_SHORT_5_6_5;
367 case PVR2_TEX_FORMAT_ARGB4444:
369 type = GL_UNSIGNED_SHORT_4_4_4_4_REV;
371 case PVR2_TEX_FORMAT_YUV422:
372 /* YUV422 isn't directly supported by most implementations, so decode
373 * it to a (reasonably) standard ARGB32.
377 type = GL_UNSIGNED_INT_8_8_8_8_REV;
379 case PVR2_TEX_FORMAT_BUMPMAP:
380 ERROR( "Bumpmap not supported" );
384 if( PVR2_TEX_IS_STRIDE(mode) && tex_format != PVR2_TEX_FORMAT_IDX4 &&
385 tex_format != PVR2_TEX_FORMAT_IDX8 ) {
386 /* Stride textures cannot be mip-mapped, compressed, indexed or twiddled */
387 uint32_t stride = (MMIO_READ( PVR2, RENDER_TEXSIZE ) & 0x003F) << 5;
388 char data[(width*height) << bpp_shift];
389 if( tex_format == PVR2_TEX_FORMAT_YUV422 ) {
390 char tmp[(width*height)<<1];
391 pvr2_vram64_read_stride( tmp, width<<1, texture_addr, stride<<1, height );
392 yuv_decode( (uint32_t *)data, (uint32_t *)tmp, width, height );
394 pvr2_vram64_read_stride( data, width<<bpp_shift, texture_addr, stride<<bpp_shift, height );
396 glTexImage2D( GL_TEXTURE_2D, 0, intFormat, width, height, 0, format, type, data );
397 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, filter);
398 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
402 int level=0, last_level = 0, mip_width = width, mip_height = height, src_bytes, dest_bytes;
403 if( PVR2_TEX_IS_MIPMAPPED(mode) ) {
405 for( i=0; 1<<i < width; i++ );
409 filter = GL_LINEAR_MIPMAP_LINEAR;
411 dest_bytes = (mip_width * mip_height) << bpp_shift;
412 src_bytes = dest_bytes; // Modes will change this (below)
414 if( PVR2_TEX_IS_COMPRESSED(mode) ) {
415 uint16_t tmp[VQ_CODEBOOK_SIZE];
416 pvr2_vram64_read( (char *)tmp, texture_addr, VQ_CODEBOOK_SIZE );
417 texture_addr += VQ_CODEBOOK_SIZE;
418 vq_get_codebook( &codebook, tmp );
421 for( level=last_level; level>= 0; level-- ) {
422 char data[dest_bytes];
423 /* load data from image, detwiddling/uncompressing as required */
424 if( tex_format == PVR2_TEX_FORMAT_IDX8 ) {
425 src_bytes = (mip_width * mip_height);
426 int bank = (mode >> 25) &0x03;
427 uint32_t *palette = ((uint32_t *)mmio_region_PVR2PAL.mem) + (bank<<8);
429 pvr2_vram64_read_twiddled_8( tmp, texture_addr, mip_width, mip_height );
430 if( bpp_shift == 2 ) {
431 decode_pal8_to_32( (uint32_t *)data, tmp, src_bytes, palette );
433 decode_pal8_to_16( (uint16_t *)data, tmp, src_bytes, palette );
435 } else if( tex_format == PVR2_TEX_FORMAT_IDX4 ) {
436 src_bytes = (mip_width * mip_height) >> 1;
437 int bank = (mode >>21 ) & 0x3F;
438 uint32_t *palette = ((uint32_t *)mmio_region_PVR2PAL.mem) + (bank<<4);
440 pvr2_vram64_read_twiddled_4( tmp, texture_addr, mip_width, mip_height );
441 if( bpp_shift == 2 ) {
442 decode_pal4_to_32( (uint32_t *)data, tmp, src_bytes, palette );
444 decode_pal4_to_16( (uint16_t *)data, tmp, src_bytes, palette );
446 } else if( tex_format == PVR2_TEX_FORMAT_YUV422 ) {
447 src_bytes = ((mip_width*mip_height)<<1);
449 if( PVR2_TEX_IS_TWIDDLED(mode) ) {
450 pvr2_vram64_read_twiddled_16( tmp, texture_addr, mip_width, mip_height );
452 pvr2_vram64_read( tmp, texture_addr, src_bytes );
454 yuv_decode( (uint32_t *)data, (uint32_t *)tmp, mip_width, mip_height );
455 } else if( PVR2_TEX_IS_COMPRESSED(mode) ) {
456 src_bytes = ((mip_width*mip_height) >> 2);
458 if( PVR2_TEX_IS_TWIDDLED(mode) ) {
459 pvr2_vram64_read_twiddled_8( tmp, texture_addr, mip_width>>1, mip_height>>1 );
461 pvr2_vram64_read( tmp, texture_addr, src_bytes );
463 vq_decode( (uint16_t *)data, tmp, mip_width, mip_height, &codebook );
464 } else if( PVR2_TEX_IS_TWIDDLED(mode) ) {
465 pvr2_vram64_read_twiddled_16( data, texture_addr, mip_width, mip_height );
467 pvr2_vram64_read( data, texture_addr, src_bytes );
471 if( level == last_level && level != 0 ) { /* 1x1 stored within a 2x2 */
472 glTexImage2D( GL_TEXTURE_2D, level, intFormat, 1, 1, 0, format, type,
473 data + (3 << bpp_shift) );
474 texture_addr += src_bytes;
476 glTexImage2D( GL_TEXTURE_2D, level, intFormat, mip_width, mip_height, 0, format, type,
478 texture_addr += src_bytes;
486 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, filter);
487 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
491 * Return a texture ID for the texture specified at the supplied address
492 * and given parameters (the same sequence of bytes could in theory have
493 * multiple interpretations). We use the texture address as the primary
494 * index, but allow for multiple instances at each address. The texture
495 * will be bound to the GL_TEXTURE_2D target before being returned.
497 * If the texture has already been bound, return the ID to which it was
498 * bound. Otherwise obtain an unused texture ID and set it up appropriately.
500 GLuint texcache_get_texture( uint32_t texture_addr, int width, int height,
503 uint32_t texture_page = texture_addr >> 12;
504 texcache_entry_index idx = texcache_page_lookup[texture_page];
505 while( idx != EMPTY_ENTRY ) {
506 texcache_entry_t entry = &texcache_active_list[idx];
507 if( entry->texture_addr == texture_addr &&
508 entry->mode == mode &&
509 entry->width == width &&
510 entry->height == height ) {
511 entry->lru_count = texcache_ref_counter++;
512 glBindTexture( GL_TEXTURE_2D, entry->texture_id );
513 return entry->texture_id;
518 /* Not found - check the free list */
521 if( texcache_free_ptr < MAX_TEXTURES ) {
522 slot = texcache_free_list[texcache_free_ptr++];
524 slot = texcache_evict_lru();
527 /* Construct new entry */
528 texcache_active_list[slot].texture_addr = texture_addr;
529 texcache_active_list[slot].width = width;
530 texcache_active_list[slot].height = height;
531 texcache_active_list[slot].mode = mode;
532 texcache_active_list[slot].lru_count = texcache_ref_counter++;
534 /* Add entry to the lookup table */
535 texcache_active_list[slot].next = texcache_page_lookup[texture_page];
536 texcache_page_lookup[texture_page] = slot;
538 /* Construct the GL texture */
539 glBindTexture( GL_TEXTURE_2D, texcache_active_list[slot].texture_id );
540 texcache_load_texture( texture_addr, width, height, mode );
542 return texcache_active_list[slot].texture_id;
.