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 -1
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 texcache_entry_index 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;
70 texcache_active_list[i].next = EMPTY_ENTRY;
72 texcache_free_ptr = 0;
73 texcache_ref_counter = 0;
77 * Setup the initial texture ids (must be called after the GL context is
80 void texcache_gl_init( )
83 GLuint texids[MAX_TEXTURES];
85 glGenTextures( MAX_TEXTURES, texids );
86 for( i=0; i<MAX_TEXTURES; i++ ) {
87 texcache_active_list[i].texture_id = texids[i];
92 * Flush all textures from the cache, returning them to the free list.
94 void texcache_flush( )
97 /* clear structures */
98 for( i=0; i<PVR2_RAM_PAGES; i++ ) {
99 texcache_page_lookup[i] = EMPTY_ENTRY;
101 for( i=0; i<MAX_TEXTURES; i++ ) {
102 texcache_free_list[i] = i;
103 texcache_active_list[i].next = EMPTY_ENTRY;
105 texcache_free_ptr = 0;
106 texcache_ref_counter = 0;
110 * Flush all textures and delete. The cache will be non-functional until
111 * the next call to texcache_init(). This would typically be done if
112 * switching GL targets.
114 void texcache_shutdown( )
116 GLuint texids[MAX_TEXTURES];
120 for( i=0; i<MAX_TEXTURES; i++ ) {
121 texids[i] = texcache_active_list[i].texture_id;
123 glDeleteTextures( MAX_TEXTURES, texids );
126 static void texcache_evict( int slot )
128 /* Remove the selected slot from the lookup table */
129 assert( texcache_active_list[slot].texture_addr != -1 );
130 uint32_t evict_page = texcache_active_list[slot].texture_addr >> 12;
131 texcache_entry_index replace_next = texcache_active_list[slot].next;
132 texcache_active_list[slot].texture_addr = -1;
133 texcache_active_list[slot].next = EMPTY_ENTRY; /* Just for safety */
134 if( texcache_page_lookup[evict_page] == slot ) {
135 texcache_page_lookup[evict_page] = replace_next;
137 texcache_entry_index idx = texcache_page_lookup[evict_page];
138 texcache_entry_index next;
140 next = texcache_active_list[idx].next;
142 assert( idx != replace_next );
143 texcache_active_list[idx].next = replace_next;
147 } while( next != EMPTY_ENTRY );
152 * Evict a single texture from the cache.
153 * @return the slot of the evicted texture.
155 static texcache_entry_index texcache_evict_lru( void )
157 /* Full table scan - take over the entry with the lowest lru value */
158 texcache_entry_index slot = 0;
159 int lru_value = texcache_active_list[0].lru_count;
161 for( i=1; i<MAX_TEXTURES; i++ ) {
162 /* FIXME: account for rollover */
163 if( texcache_active_list[i].lru_count < lru_value ) {
165 lru_value = texcache_active_list[i].lru_count;
168 texcache_evict(slot);
174 * Evict all textures contained in the page identified by a texture address.
176 void texcache_invalidate_page( uint32_t texture_addr ) {
177 uint32_t texture_page = texture_addr >> 12;
178 texcache_entry_index idx = texcache_page_lookup[texture_page];
179 if( idx == EMPTY_ENTRY )
181 assert( texcache_free_ptr >= 0 );
183 texcache_entry_t entry = &texcache_active_list[idx];
184 entry->texture_addr = -1;
187 texcache_free_list[texcache_free_ptr] = idx;
189 entry->next = EMPTY_ENTRY;
190 } while( idx != EMPTY_ENTRY );
191 texcache_page_lookup[texture_page] = EMPTY_ENTRY;
195 * Mark all textures that use the palette table as needing a re-read (ie
196 * for when the palette is changed. We could track exactly which ones are
197 * affected, but it's not clear that the extra maintanence overhead is
200 void texcache_invalidate_palette( )
203 for( i=0; i<MAX_TEXTURES; i++ ) {
204 if( texcache_active_list[i].texture_addr != -1 &&
205 PVR2_TEX_IS_PALETTE(texcache_active_list[i].mode) ) {
208 texcache_free_list[texcache_free_ptr] = i;
213 static void decode_pal8_to_32( uint32_t *out, uint8_t *in, int inbytes, uint32_t *pal )
216 for( i=0; i<inbytes; i++ ) {
221 static void decode_pal8_to_16( uint16_t *out, uint8_t *in, int inbytes, uint32_t *pal )
224 for( i=0; i<inbytes; i++ ) {
225 *out++ = (uint16_t)pal[*in++];
229 static void decode_pal4_to_32( uint32_t *out, uint8_t *in, int inbytes, uint32_t *pal )
232 for( i=0; i<inbytes; i++ ) {
233 *out++ = pal[*in & 0x0F];
234 *out++ = pal[(*in >> 4)];
240 static void decode_pal4_to_16( uint16_t *out, uint8_t *in, int inbytes, uint32_t *pal )
243 for( i=0; i<inbytes; i++ ) {
244 *out++ = (uint16_t)pal[*in & 0x0F];
245 *out++ = (uint16_t)pal[(*in >> 4)];
250 #define VQ_CODEBOOK_SIZE 2048 /* 256 entries * 4 pixels per quad * 2 byte pixels */
253 uint16_t quad[256][4];
256 static void vq_get_codebook( struct vq_codebook *codebook,
259 /* Detwiddle the codebook, for the sake of my own sanity if nothing else */
260 uint16_t *p = (uint16_t *)input;
262 for( i=0; i<256; i++ ) {
263 codebook->quad[i][0] = *p++;
264 codebook->quad[i][2] = *p++;
265 codebook->quad[i][1] = *p++;
266 codebook->quad[i][3] = *p++;
270 static void vq_decode( uint16_t *output, unsigned char *input, int width, int height,
271 struct vq_codebook *codebook ) {
274 uint8_t *c = (uint8_t *)input;
275 for( j=0; j<height; j+=2 ) {
276 for( i=0; i<width; i+=2 ) {
278 output[i + j*width] = codebook->quad[code][0];
279 output[i + 1 + j*width] = codebook->quad[code][1];
280 output[i + (j+1)*width] = codebook->quad[code][2];
281 output[i + 1 + (j+1)*width] = codebook->quad[code][3];
286 static inline uint32_t yuv_to_rgb32( float y, float u, float v )
290 int r = (int)(y + v*1.375);
291 int g = (int)(y - u*0.34375 - v*0.6875);
292 int b = (int)(y + u*1.71875);
293 if( r > 255 ) { r = 255; } else if( r < 0 ) { r = 0; }
294 if( g > 255 ) { g = 255; } else if( g < 0 ) { g = 0; }
295 if( b > 255 ) { b = 255; } else if( b < 0 ) { b = 0; }
296 return 0xFF000000 | (r<<16) | (g<<8) | (b);
301 * Convert raster YUV texture data into RGB32 data - most GL implementations don't
302 * directly support this format unfortunately. The input data is formatted as
303 * 32 bits = 2 horizontal pixels, UYVY. This is currently done rather inefficiently
306 static void yuv_decode( uint32_t *output, uint32_t *input, int width, int height )
310 for( y=0; y<height; y++ ) {
311 for( x=0; x<width; x+=2 ) {
312 float u = (float)(*p & 0xFF);
313 float y0 = (float)( (*p>>8)&0xFF );
314 float v = (float)( (*p>>16)&0xFF );
315 float y1 = (float)( (*p>>24)&0xFF );
316 *output++ = yuv_to_rgb32( y0, u, v );
317 *output++ = yuv_to_rgb32( y1, u, v );
324 * Load texture data from the given address and parameters into the currently
325 * bound OpenGL texture.
327 static void texcache_load_texture( uint32_t texture_addr, int width, int height,
329 int bpp_shift = 1; /* bytes per (output) pixel as a power of 2 */
330 GLint intFormat = GL_RGBA, format, type;
331 int tex_format = mode & PVR2_TEX_FORMAT_MASK;
332 struct vq_codebook codebook;
333 GLint filter = GL_LINEAR;
335 glPixelStorei( GL_UNPACK_ROW_LENGTH, 0 );
336 /* Decode the format parameters */
337 switch( tex_format ) {
338 case PVR2_TEX_FORMAT_IDX4:
339 case PVR2_TEX_FORMAT_IDX8:
340 /* For indexed-colour modes, we need to lookup the palette control
341 * word to determine the de-indexed texture format.
343 switch( MMIO_READ( PVR2, RENDER_PALETTE ) & 0x03 ) {
344 case 0: /* ARGB1555 */
346 type = GL_UNSIGNED_SHORT_1_5_5_5_REV;
351 type = GL_UNSIGNED_SHORT_5_6_5;
353 case 2: /* ARGB4444 */
355 type = GL_UNSIGNED_SHORT_4_4_4_4_REV;
357 case 3: /* ARGB8888 */
359 type = GL_UNSIGNED_BYTE;
363 return; /* Can't happen, but it makes gcc stop complaining */
367 case PVR2_TEX_FORMAT_ARGB1555:
369 type = GL_UNSIGNED_SHORT_1_5_5_5_REV;
371 case PVR2_TEX_FORMAT_RGB565:
374 type = GL_UNSIGNED_SHORT_5_6_5;
376 case PVR2_TEX_FORMAT_ARGB4444:
378 type = GL_UNSIGNED_SHORT_4_4_4_4_REV;
380 case PVR2_TEX_FORMAT_YUV422:
381 /* YUV422 isn't directly supported by most implementations, so decode
382 * it to a (reasonably) standard ARGB32.
386 type = GL_UNSIGNED_BYTE;
388 case PVR2_TEX_FORMAT_BUMPMAP:
389 ERROR( "Bumpmap not supported" );
392 ERROR( "Undefined texture format" );
396 if( PVR2_TEX_IS_STRIDE(mode) && tex_format != PVR2_TEX_FORMAT_IDX4 &&
397 tex_format != PVR2_TEX_FORMAT_IDX8 ) {
398 /* Stride textures cannot be mip-mapped, compressed, indexed or twiddled */
399 uint32_t stride = (MMIO_READ( PVR2, RENDER_TEXSIZE ) & 0x003F) << 5;
400 unsigned char data[(width*height) << bpp_shift];
401 if( tex_format == PVR2_TEX_FORMAT_YUV422 ) {
402 unsigned char tmp[(width*height)<<1];
403 pvr2_vram64_read_stride( tmp, width<<1, texture_addr, stride<<1, height );
404 yuv_decode( (uint32_t *)data, (uint32_t *)tmp, width, height );
406 pvr2_vram64_read_stride( data, width<<bpp_shift, texture_addr, stride<<bpp_shift, height );
408 glTexImage2D( GL_TEXTURE_2D, 0, intFormat, width, height, 0, format, type, data );
409 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, filter);
410 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
414 int level=0, last_level = 0, mip_width = width, mip_height = height, src_bytes, dest_bytes;
415 if( PVR2_TEX_IS_MIPMAPPED(mode) ) {
417 for( i=0; 1<<i < width; i++ );
421 filter = GL_LINEAR_MIPMAP_LINEAR;
423 dest_bytes = (mip_width * mip_height) << bpp_shift;
424 src_bytes = dest_bytes; // Modes will change this (below)
426 if( PVR2_TEX_IS_COMPRESSED(mode) ) {
427 uint16_t tmp[VQ_CODEBOOK_SIZE];
428 pvr2_vram64_read( (unsigned char *)tmp, texture_addr, VQ_CODEBOOK_SIZE );
429 texture_addr += VQ_CODEBOOK_SIZE;
430 vq_get_codebook( &codebook, tmp );
433 for( level=last_level; level>= 0; level-- ) {
434 unsigned char data[dest_bytes];
435 /* load data from image, detwiddling/uncompressing as required */
436 if( tex_format == PVR2_TEX_FORMAT_IDX8 ) {
437 src_bytes = (mip_width * mip_height);
438 int bank = (mode >> 25) &0x03;
439 uint32_t *palette = ((uint32_t *)mmio_region_PVR2PAL.mem) + (bank<<8);
440 unsigned char tmp[src_bytes];
441 pvr2_vram64_read_twiddled_8( tmp, texture_addr, mip_width, mip_height );
442 if( bpp_shift == 2 ) {
443 decode_pal8_to_32( (uint32_t *)data, tmp, src_bytes, palette );
445 decode_pal8_to_16( (uint16_t *)data, tmp, src_bytes, palette );
447 } else if( tex_format == PVR2_TEX_FORMAT_IDX4 ) {
448 src_bytes = (mip_width * mip_height) >> 1;
449 int bank = (mode >>21 ) & 0x3F;
450 uint32_t *palette = ((uint32_t *)mmio_region_PVR2PAL.mem) + (bank<<4);
451 unsigned char tmp[src_bytes];
452 pvr2_vram64_read_twiddled_4( tmp, texture_addr, mip_width, mip_height );
453 if( bpp_shift == 2 ) {
454 decode_pal4_to_32( (uint32_t *)data, tmp, src_bytes, palette );
456 decode_pal4_to_16( (uint16_t *)data, tmp, src_bytes, palette );
458 } else if( tex_format == PVR2_TEX_FORMAT_YUV422 ) {
459 src_bytes = ((mip_width*mip_height)<<1);
460 unsigned char tmp[src_bytes];
461 if( PVR2_TEX_IS_TWIDDLED(mode) ) {
462 pvr2_vram64_read_twiddled_16( tmp, texture_addr, mip_width, mip_height );
464 pvr2_vram64_read( tmp, texture_addr, src_bytes );
466 yuv_decode( (uint32_t *)data, (uint32_t *)tmp, mip_width, mip_height );
467 } else if( PVR2_TEX_IS_COMPRESSED(mode) ) {
468 src_bytes = ((mip_width*mip_height) >> 2);
469 unsigned char tmp[src_bytes];
470 if( PVR2_TEX_IS_TWIDDLED(mode) ) {
471 pvr2_vram64_read_twiddled_8( tmp, texture_addr, mip_width>>1, mip_height>>1 );
473 pvr2_vram64_read( tmp, texture_addr, src_bytes );
475 vq_decode( (uint16_t *)data, tmp, mip_width, mip_height, &codebook );
476 } else if( PVR2_TEX_IS_TWIDDLED(mode) ) {
477 pvr2_vram64_read_twiddled_16( data, texture_addr, mip_width, mip_height );
479 pvr2_vram64_read( data, texture_addr, src_bytes );
483 if( level == last_level && level != 0 ) { /* 1x1 stored within a 2x2 */
484 glTexImage2D( GL_TEXTURE_2D, level, intFormat, 1, 1, 0, format, type,
485 data + (3 << bpp_shift) );
486 texture_addr += src_bytes;
488 glTexImage2D( GL_TEXTURE_2D, level, intFormat, mip_width, mip_height, 0, format, type,
490 texture_addr += src_bytes;
498 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, filter);
499 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
503 * Return a texture ID for the texture specified at the supplied address
504 * and given parameters (the same sequence of bytes could in theory have
505 * multiple interpretations). We use the texture address as the primary
506 * index, but allow for multiple instances at each address. The texture
507 * will be bound to the GL_TEXTURE_2D target before being returned.
509 * If the texture has already been bound, return the ID to which it was
510 * bound. Otherwise obtain an unused texture ID and set it up appropriately.
512 GLuint texcache_get_texture( uint32_t texture_addr, int width, int height,
515 uint32_t texture_page = texture_addr >> 12;
516 texcache_entry_index next;
517 texcache_entry_index idx = texcache_page_lookup[texture_page];
518 while( idx != EMPTY_ENTRY ) {
519 texcache_entry_t entry = &texcache_active_list[idx];
520 if( entry->texture_addr == texture_addr &&
521 entry->mode == mode &&
522 entry->width == width &&
523 entry->height == height ) {
524 entry->lru_count = texcache_ref_counter++;
525 glBindTexture( GL_TEXTURE_2D, entry->texture_id );
526 return entry->texture_id;
532 /* Not found - check the free list */
533 texcache_entry_index slot = 0;
535 if( texcache_free_ptr < MAX_TEXTURES ) {
536 slot = texcache_free_list[texcache_free_ptr++];
538 slot = texcache_evict_lru();
541 /* Construct new entry */
542 texcache_active_list[slot].texture_addr = texture_addr;
543 texcache_active_list[slot].width = width;
544 texcache_active_list[slot].height = height;
545 texcache_active_list[slot].mode = mode;
546 texcache_active_list[slot].lru_count = texcache_ref_counter++;
548 /* Add entry to the lookup table */
549 next = texcache_page_lookup[texture_page];
552 fprintf( stderr, "Active list: " );
553 for( i=0; i<MAX_TEXTURES; i++ ) {
554 fprintf( stderr, "%d, ", texcache_active_list[i].next );
556 fprintf( stderr, "\n" );
557 assert( next != slot );
560 assert( next != slot );
561 texcache_active_list[slot].next = next;
562 texcache_page_lookup[texture_page] = slot;
564 /* Construct the GL texture */
565 glBindTexture( GL_TEXTURE_2D, texcache_active_list[slot].texture_id );
566 texcache_load_texture( texture_addr, width, height, mode );
568 texcache_integrity_check();
569 return texcache_active_list[slot].texture_id;
573 * Check the integrity of the texcache. Verifies that every cache slot
574 * appears exactly once on either the free list or one page list. For
575 * active slots, the texture address must also match the page it appears on.
578 void texcache_integrity_check()
581 int slot_found[MAX_TEXTURES];
583 memset( slot_found, 0, sizeof(slot_found) );
585 /* Check entries on the free list */
586 for( i= texcache_free_ptr; i< MAX_TEXTURES; i++ ) {
587 int slot = texcache_free_list[i];
588 assert( slot_found[slot] == 0 );
589 assert( texcache_active_list[slot].next == EMPTY_ENTRY );
590 slot_found[slot] = 1;
593 /* Check entries on the active lists */
594 for( i=0; i< PVR2_RAM_PAGES; i++ ) {
595 int slot = texcache_page_lookup[i];
596 while( slot != EMPTY_ENTRY ) {
597 assert( slot_found[slot] == 0 );
598 assert( (texcache_active_list[slot].texture_addr >> 12) == i );
599 slot_found[slot] = 2;
600 slot = texcache_active_list[slot].next;
604 /* Make sure we didn't miss any entries */
605 for( i=0; i<MAX_TEXTURES; i++ ) {
606 assert( slot_found[i] != 0 );
.