AlexIndustrial/mesa
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

llvmpipe: Don't limit the number of tiles in the cache.

Browse Source
This commit is contained in:
José Fonseca
2009-08-21 09:51:44 +01:00
parent 87c383a2e5
commit f10ee9a0a1
3 changed files with 95 additions and 238 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/gallium/drivers/llvmpipe/lp_clear.c
-5
View File
@@ -66,11 +66,6 @@ llvmpipe_clear(struct pipe_context *pipe, unsigned buffers, const float *rgba,
util_pack_color(rgba, ps->format, &cv);
lp_tile_cache_clear(llvmpipe->cbuf_cache[i], rgba, cv);
#if !TILE_CLEAR_OPTIMIZATION
/* non-cached surface */
pipe->surface_fill(pipe, ps, 0, 0, ps->width, ps->height, cv);
#endif
}
}
src/gallium/drivers/llvmpipe/lp_tile_cache.c
+82 -170
View File
@@ -44,60 +44,17 @@
#include "lp_tile_cache.h"
/**
* Return the position in the cache for the tile that contains win pos (x,y).
* We currently use a direct mapped cache so this is like a hack key.
* At some point we should investige something more sophisticated, like
* a LRU replacement policy.
*/
#define CACHE_POS(x, y) \
(((x) + (y) * 5) % NUM_ENTRIES)
/**
* Is the tile at (x,y) in cleared state?
*/
static INLINE uint
is_clear_flag_set(const uint *bitvec, union tile_address addr)
{
int pos, bit;
pos = addr.bits.y * (MAX_WIDTH / TILE_SIZE) + addr.bits.x;
assert(pos / 32 < (MAX_WIDTH / TILE_SIZE) * (MAX_HEIGHT / TILE_SIZE) / 32);
bit = bitvec[pos / 32] & (1 << (pos & 31));
return bit;
}
/**
* Mark the tile at (x,y) as not cleared.
*/
static INLINE void
clear_clear_flag(uint *bitvec, union tile_address addr)
{
int pos;
pos = addr.bits.y * (MAX_WIDTH / TILE_SIZE) + addr.bits.x;
assert(pos / 32 < (MAX_WIDTH / TILE_SIZE) * (MAX_HEIGHT / TILE_SIZE) / 32);
bitvec[pos / 32] &= ~(1 << (pos & 31));
}
struct llvmpipe_tile_cache *
lp_create_tile_cache( struct pipe_screen *screen )
{
struct llvmpipe_tile_cache *tc;
uint pos;
tc = align_malloc( sizeof(struct llvmpipe_tile_cache), 16 );
if (tc) {
memset(tc, 0, sizeof *tc);
tc->screen = screen;
for (pos = 0; pos < NUM_ENTRIES; pos++) {
tc->entries[pos].addr.bits.invalid = 1;
}
tc->last_tile = &tc->entries[0]; /* any tile */
}
tc = CALLOC_STRUCT( llvmpipe_tile_cache );
if(!tc)
return NULL;
tc->screen = screen;
return tc;
}
@@ -106,17 +63,23 @@ void
lp_destroy_tile_cache(struct llvmpipe_tile_cache *tc)
{
struct pipe_screen *screen;
uint pos;
unsigned x, y;
for (pos = 0; pos < NUM_ENTRIES; pos++) {
/*assert(tc->entries[pos].x < 0);*/
for (y = 0; y < MAX_HEIGHT; y += TILE_SIZE) {
for (x = 0; x < MAX_WIDTH; x += TILE_SIZE) {
struct llvmpipe_cached_tile *tile = &tc->entries[y/TILE_SIZE][x/TILE_SIZE];
if(tile->color)
align_free(tile->color);
}
}
if (tc->transfer) {
screen = tc->transfer->texture->screen;
screen->tex_transfer_destroy(tc->transfer);
}
align_free( tc );
FREE( tc );
}
@@ -146,11 +109,23 @@ lp_tile_cache_set_surface(struct llvmpipe_tile_cache *tc,
if (ps) {
struct pipe_screen *screen = ps->texture->screen;
unsigned x, y;
tc->transfer = screen->get_tex_transfer(screen, ps->texture, ps->face,
ps->level, ps->zslice,
PIPE_TRANSFER_READ_WRITE,
0, 0, ps->width, ps->height);
for (y = 0; y < ps->height; y += TILE_SIZE) {
for (x = 0; x < ps->width; x += TILE_SIZE) {
struct llvmpipe_cached_tile *tile = &tc->entries[y/TILE_SIZE][x/TILE_SIZE];
tile->status = LP_TILE_STATUS_UNDEFINED;
if(!tile->color)
tile->color = align_malloc( TILE_SIZE*TILE_SIZE*NUM_CHANNELS, 16 );
}
}
}
}
@@ -206,24 +181,33 @@ clear_tile(struct llvmpipe_cached_tile *tile,
/**
* Actually clear the tiles which were flagged as being in a clear state.
* Flush the tile cache: write all dirty tiles back to the transfer.
* any tiles "flagged" as cleared will be "really" cleared.
*/
static void
lp_tile_cache_flush_clear(struct llvmpipe_tile_cache *tc)
void
lp_flush_tile_cache(struct llvmpipe_tile_cache *tc)
{
struct pipe_transfer *pt = tc->transfer;
struct pipe_screen *screen = pt->texture->screen;
const uint w = tc->transfer->width;
const uint h = tc->transfer->height;
uint x, y;
uint numCleared = 0;
unsigned x, y;
if(!pt)
return;
/* push the tile to all positions marked as clear */
for (y = 0; y < h; y += TILE_SIZE) {
for (x = 0; x < w; x += TILE_SIZE) {
union tile_address addr = tile_address(x, y, 0, 0, 0);
for (y = 0; y < pt->height; y += TILE_SIZE) {
for (x = 0; x < pt->width; x += TILE_SIZE) {
struct llvmpipe_cached_tile *tile = &tc->entries[y/TILE_SIZE][x/TILE_SIZE];
if (is_clear_flag_set(tc->clear_flags, addr)) {
switch(tile->status) {
case LP_TILE_STATUS_UNDEFINED:
break;
case LP_TILE_STATUS_CLEAR: {
/**
* Actually clear the tiles which were flagged as being in a clear state.
*/
struct pipe_screen *screen = pt->texture->screen;
unsigned tw = TILE_SIZE;
unsigned th = TILE_SIZE;
void *dst;
@@ -240,51 +224,16 @@ lp_tile_cache_flush_clear(struct llvmpipe_tile_cache *tc)
x, y, tw, th,
tc->clear_val);
/* do this? */
clear_clear_flag(tc->clear_flags, addr);
screen->transfer_unmap(screen, pt);
break;
}
numCleared++;
case LP_TILE_STATUS_DEFINED:
lp_put_tile_rgba_soa(pt, x, y, tile->color);
break;
}
}
}
#if 0
debug_printf("num cleared: %u\n", numCleared);
#endif
}
/**
* Flush the tile cache: write all dirty tiles back to the transfer.
* any tiles "flagged" as cleared will be "really" cleared.
*/
void
lp_flush_tile_cache(struct llvmpipe_tile_cache *tc)
{
struct pipe_transfer *pt = tc->transfer;
int inuse = 0, pos;
if (pt) {
/* caching a drawing transfer */
for (pos = 0; pos < NUM_ENTRIES; pos++) {
struct llvmpipe_cached_tile *tile = tc->entries + pos;
if (!tile->addr.bits.invalid) {
lp_put_tile_rgba_soa(pt,
tile->addr.bits.x * TILE_SIZE,
tile->addr.bits.y * TILE_SIZE,
tile->color);
tile->addr.bits.invalid = 1; /* mark as empty */
inuse++;
}
}
#if TILE_CLEAR_OPTIMIZATION
lp_tile_cache_flush_clear(tc);
#endif
}
#if 0
debug_printf("flushed tiles in use: %d\n", inuse);
#endif
}
@@ -293,68 +242,34 @@ lp_flush_tile_cache(struct llvmpipe_tile_cache *tc)
* \param x, y position of tile, in pixels
*/
void *
lp_find_cached_tile(struct llvmpipe_tile_cache *tc,
union tile_address addr )
lp_get_cached_tile(struct llvmpipe_tile_cache *tc,
unsigned x, unsigned y )
{
struct llvmpipe_cached_tile *tile = &tc->entries[y/TILE_SIZE][x/TILE_SIZE];
struct pipe_transfer *pt = tc->transfer;
/* cache pos/entry: */
const int pos = CACHE_POS(addr.bits.x,
addr.bits.y);
struct llvmpipe_cached_tile *tile = tc->entries + pos;
switch(tile->status) {
case LP_TILE_STATUS_CLEAR:
/* don't get tile from framebuffer, just clear it */
clear_tile(tile, tc->clear_color);
tile->status = LP_TILE_STATUS_DEFINED;
break;
if (addr.value != tile->addr.value) {
case LP_TILE_STATUS_UNDEFINED:
/* get new tile data from transfer */
lp_get_tile_rgba_soa(pt, x, y, tile->color);
tile->status = LP_TILE_STATUS_DEFINED;
break;
if (tile->addr.bits.invalid == 0) {
/* put dirty tile back in framebuffer */
lp_put_tile_rgba_soa(pt,
tile->addr.bits.x * TILE_SIZE,
tile->addr.bits.y * TILE_SIZE,
tile->color);
}
tile->addr = addr;
if (is_clear_flag_set(tc->clear_flags, addr)) {
/* don't get tile from framebuffer, just clear it */
clear_tile(tile, tc->clear_color);
clear_clear_flag(tc->clear_flags, addr);
}
else {
/* get new tile data from transfer */
lp_get_tile_rgba_soa(pt,
tile->addr.bits.x * TILE_SIZE,
tile->addr.bits.y * TILE_SIZE,
tile->color);
}
case LP_TILE_STATUS_DEFINED:
/* nothing to do */
break;
}
tc->last_tile = tile;
return tile->color;
}
/**
* Given the texture face, level, zslice, x and y values, compute
* the cache entry position/index where we'd hope to find the
* cached texture tile.
* This is basically a direct-map cache.
* XXX There's probably lots of ways in which we can improve this.
*/
static INLINE uint
tex_cache_pos( union tile_address addr )
{
uint entry = (addr.bits.x +
addr.bits.y * 9 +
addr.bits.z * 3 +
addr.bits.face +
addr.bits.level * 7);
return entry % NUM_ENTRIES;
}
/**
* When a whole surface is being cleared to a value we can avoid
* fetching tiles above.
@@ -364,24 +279,21 @@ void
lp_tile_cache_clear(struct llvmpipe_tile_cache *tc, const float *rgba,
uint clearValue)
{
unsigned chan;
uint pos;
struct pipe_transfer *pt = tc->transfer;
const unsigned w = pt->width;
const unsigned h = pt->height;
unsigned x, y, chan;
for(chan = 0; chan < 4; ++chan)
tc->clear_color[chan] = float_to_ubyte(rgba[chan]);
tc->clear_val = clearValue;
#if TILE_CLEAR_OPTIMIZATION
/* set flags to indicate all the tiles are cleared */
memset(tc->clear_flags, 255, sizeof(tc->clear_flags));
#else
/* disable the optimization */
memset(tc->clear_flags, 0, sizeof(tc->clear_flags));
#endif
for (pos = 0; pos < NUM_ENTRIES; pos++) {
struct llvmpipe_cached_tile *tile = tc->entries + pos;
tile->addr.bits.invalid = 1;
/* push the tile to all positions marked as clear */
for (y = 0; y < h; y += TILE_SIZE) {
for (x = 0; x < w; x += TILE_SIZE) {
struct llvmpipe_cached_tile *tile = &tc->entries[y/TILE_SIZE][x/TILE_SIZE];
tile->status = LP_TILE_STATUS_CLEAR;
}
}
}
src/gallium/drivers/llvmpipe/lp_tile_cache.h
+13 -63
View File
@@ -28,41 +28,26 @@
#ifndef LP_TILE_CACHE_H
#define LP_TILE_CACHE_H
#define TILE_CLEAR_OPTIMIZATION 1
#include "pipe/p_compiler.h"
#include "lp_tile_soa.h"
struct llvmpipe_context;
struct llvmpipe_tile_cache;
/* If we need to support > 4096, just expand this to be a 64 bit
* union, or consider tiling in Z as well.
*/
union tile_address {
struct {
unsigned x:6; /* 4096 / TILE_SIZE */
unsigned y:6; /* 4096 / TILE_SIZE */
unsigned z:12; /* 4096 -- z not tiled */
unsigned face:3;
unsigned level:4;
unsigned invalid:1;
} bits;
unsigned value;
enum llvmpipe_tile_status
{
LP_TILE_STATUS_UNDEFINED = 0,
LP_TILE_STATUS_CLEAR = 1,
LP_TILE_STATUS_DEFINED = 2
};
struct llvmpipe_cached_tile
{
union tile_address addr;
/** color in SOA format */
uint8_t ALIGN16_ATTRIB color[TILE_SIZE*TILE_SIZE*NUM_CHANNELS];
};
enum llvmpipe_tile_status status;
#define NUM_ENTRIES 50
/** color in SOA format */
uint8_t *color;
};
/** XXX move these */
@@ -77,8 +62,8 @@ struct llvmpipe_tile_cache
struct pipe_transfer *transfer;
void *transfer_map;
struct llvmpipe_cached_tile entries[NUM_ENTRIES];
uint clear_flags[(MAX_WIDTH / TILE_SIZE) * (MAX_HEIGHT / TILE_SIZE) / 32];
struct llvmpipe_cached_tile entries[MAX_WIDTH/TILE_SIZE][MAX_HEIGHT/TILE_SIZE];
uint8_t clear_color[4]; /**< for color bufs */
uint clear_val; /**< for z+stencil, or packed color clear value */
@@ -113,43 +98,8 @@ lp_tile_cache_clear(struct llvmpipe_tile_cache *tc, const float *rgba,
uint clearValue);
extern void *
lp_find_cached_tile(struct llvmpipe_tile_cache *tc,
union tile_address addr );
static INLINE const union tile_address
tile_address( unsigned x,
unsigned y,
unsigned z,
unsigned face,
unsigned level )
{
union tile_address addr;
addr.value = 0;
addr.bits.x = x / TILE_SIZE;
addr.bits.y = y / TILE_SIZE;
addr.bits.z = z;
addr.bits.face = face;
addr.bits.level = level;
return addr;
}
/* Quickly retrieve tile if it matches last lookup.
*/
static INLINE void *
lp_get_cached_tile(struct llvmpipe_tile_cache *tc,
int x, int y )
{
union tile_address addr = tile_address( x, y, 0, 0, 0 );
if (tc->last_tile->addr.value == addr.value)
return &tc->last_tile->color;
return lp_find_cached_tile( tc, addr );
}
lp_get_cached_tile(struct llvmpipe_tile_cache *tc,
unsigned x, unsigned y );
#endif /* LP_TILE_CACHE_H */