diff options
author | Brian Paul <[email protected]> | 2009-12-07 18:01:12 -0700 |
---|---|---|
committer | Brian Paul <[email protected]> | 2009-12-07 18:04:54 -0700 |
commit | aab1ceceecbd6449eebce7f5f5b356b1a51552e7 (patch) | |
tree | 5a34c098ac995bff9f1c8bdb167e6c701dcdb726 | |
parent | 87c9ceaea2138e051c48cd8c0fbf5f6658100779 (diff) |
llvmpipe: implement threaded rasterization
The LP_NUM_THREADS env var controls how many threads are created.
The default (and max) is 4, for now.
If LP_NUM_THREADS = 0, threading is not used.
-rw-r--r-- | src/gallium/drivers/llvmpipe/lp_rast.c | 195 | ||||
-rw-r--r-- | src/gallium/drivers/llvmpipe/lp_rast_priv.h | 18 |
2 files changed, 170 insertions, 43 deletions
diff --git a/src/gallium/drivers/llvmpipe/lp_rast.c b/src/gallium/drivers/llvmpipe/lp_rast.c index 37cc28e938e..99f7108b423 100644 --- a/src/gallium/drivers/llvmpipe/lp_rast.c +++ b/src/gallium/drivers/llvmpipe/lp_rast.c @@ -26,6 +26,7 @@ **************************************************************************/ #include "util/u_memory.h" +#include "util/u_math.h" #include "lp_debug.h" #include "lp_state.h" @@ -36,25 +37,6 @@ #include "lp_bin.h" -struct lp_rasterizer *lp_rast_create( struct pipe_screen *screen ) -{ - struct lp_rasterizer *rast; - unsigned i; - - rast = CALLOC_STRUCT(lp_rasterizer); - if(!rast) - return NULL; - - rast->screen = screen; - - for (i = 0; i < Elements(rast->tasks); i++) { - rast->tasks[i].tile.color = align_malloc( TILE_SIZE*TILE_SIZE*4, 16 ); - rast->tasks[i].tile.depth = align_malloc( TILE_SIZE*TILE_SIZE*4, 16 ); - } - - return rast; -} - /** * Begin the rasterization phase. @@ -414,16 +396,25 @@ static void lp_rast_store_color( struct lp_rasterizer *rast, { const unsigned x = rast->tasks[thread_index].x; const unsigned y = rast->tasks[thread_index].y; - unsigned w = TILE_SIZE; - unsigned h = TILE_SIZE; + int w = TILE_SIZE; + int h = TILE_SIZE; if (x + w > rast->width) w -= x + w - rast->width; - if (y + h > rast->height) - h -= y + h - rast->height; + if (y + h > rast->height) { + int h2; + h2 = h - (y + h - rast->height); + assert(h2 <= TILE_SIZE); + h = h2; + } + assert(w >= 0); + assert(h >= 0); + assert(w <= TILE_SIZE); + assert(h <= TILE_SIZE); - LP_DBG(DEBUG_RAST, "%s %d,%d %dx%d\n", __FUNCTION__, x, y, w, h); + LP_DBG(DEBUG_RAST, "%s [%u] %d,%d %dx%d\n", __FUNCTION__, + thread_index, x, y, w, h); lp_tile_write_4ub(rast->cbuf_transfer->format, rast->tasks[thread_index].tile.color, @@ -513,7 +504,7 @@ rasterize_bin( struct lp_rasterizer *rast, /* simply execute each of the commands in the block list */ for (block = commands->head; block; block = block->next) { for (k = 0; k < block->count; k++) { - block->cmd[k]( rast, 0, block->arg[k] ); + block->cmd[k]( rast, thread_index, block->arg[k] ); } } @@ -524,6 +515,41 @@ rasterize_bin( struct lp_rasterizer *rast, /** * Rasterize/execute all bins. */ +static void +rasterize_bins( struct lp_rasterizer *rast, + unsigned thread_index, + struct lp_bins *bins, + const struct pipe_framebuffer_state *fb, + bool write_depth ) +{ + /* loop over tile bins, rasterize each */ +#if 0 + { + unsigned i, j; + for (i = 0; i < bins->tiles_x; i++) { + for (j = 0; j < bins->tiles_y; j++) { + struct cmd_bin *bin = lp_get_bin(bins, i, j); + rasterize_bin( rast, thread_index, + bin, i * TILE_SIZE, j * TILE_SIZE ); + } + } + } +#else + { + struct cmd_bin *bin; + int x, y; + + while ((bin = lp_bin_iter_next(bins, &x, &y))) { + rasterize_bin( rast, thread_index, bin, x * TILE_SIZE, y * TILE_SIZE); + } + } +#endif +} + + +/** + * Called by rasterizer when it has something for us to render. + */ void lp_rasterize_bins( struct lp_rasterizer *rast, struct lp_bins *bins, @@ -539,30 +565,32 @@ lp_rasterize_bins( struct lp_rasterizer *rast, fb->zsbuf != NULL && write_depth, fb->width, fb->height ); - - /* loop over tile bins, rasterize each */ -#if 0 - { - unsigned i, j; - for (i = 0; i < bins->tiles_x; i++) { - for (j = 0; j < bins->tiles_y; j++) { - struct cmd_bin *bin = lp_get_bin(bins, i, j); - rasterize_bin( rast, 0, bin, i * TILE_SIZE, j * TILE_SIZE ); - } - } + + if (rast->num_threads == 0) { + /* no threading */ + lp_bin_iter_begin( bins ); + rasterize_bins( rast, 0, bins, fb, write_depth ); } -#else - { - struct cmd_bin *bin; - int x, y; + else { + /* threaded rendering! */ + unsigned i; + + rast->bins = bins; + rast->fb = fb; + rast->write_depth = write_depth; lp_bin_iter_begin( bins ); - while ((bin = lp_bin_iter_next(bins, &x, &y))) { - rasterize_bin( rast, 0, bin, x * TILE_SIZE, y * TILE_SIZE); + /* signal the threads that there's work to do */ + for (i = 0; i < rast->num_threads; i++) { + pipe_semaphore_signal(&rast->tasks[i].work_ready); + } + + /* wait for work to complete */ + for (i = 0; i < rast->num_threads; i++) { + pipe_semaphore_wait(&rast->tasks[i].work_done); } } -#endif lp_rast_end( rast ); @@ -570,6 +598,87 @@ lp_rasterize_bins( struct lp_rasterizer *rast, } +/** + * This is the thread's main entrypoint. + * It's a simple loop: + * 1. wait for work + * 2. do work + * 3. signal that we're done + */ +static void * +thread_func( void *init_data ) +{ + struct lp_rasterizer_task *task = (struct lp_rasterizer_task *) init_data; + struct lp_rasterizer *rast = task->rast; + int debug = 0; + + while (1) { + /* wait for work */ + if (debug) + debug_printf("thread %d waiting for work\n", task->thread_index); + pipe_semaphore_wait(&task->work_ready); + + /* do work */ + if (debug) + debug_printf("thread %d doing work\n", task->thread_index); + rasterize_bins(rast, task->thread_index, + rast->bins, rast->fb, rast->write_depth); + + /* signal done with work */ + if (debug) + debug_printf("thread %d done working\n", task->thread_index); + pipe_semaphore_signal(&task->work_done); + } + + return NULL; +} + + +/** + * Initialize semaphores and spawn the threads. + */ +static void +create_rast_threads(struct lp_rasterizer *rast) +{ + unsigned i; + + rast->num_threads = debug_get_num_option("LP_NUM_THREADS", MAX_THREADS); + rast->num_threads = MIN2(rast->num_threads, MAX_THREADS); + + /* NOTE: if num_threads is zero, we won't use any threads */ + for (i = 0; i < rast->num_threads; i++) { + pipe_semaphore_init(&rast->tasks[i].work_ready, 0); + pipe_semaphore_init(&rast->tasks[i].work_done, 0); + rast->threads[i] = pipe_thread_create(thread_func, + (void *) &rast->tasks[i]); + } +} + + + +struct lp_rasterizer *lp_rast_create( struct pipe_screen *screen ) +{ + struct lp_rasterizer *rast; + unsigned i; + + rast = CALLOC_STRUCT(lp_rasterizer); + if(!rast) + return NULL; + + rast->screen = screen; + + for (i = 0; i < Elements(rast->tasks); i++) { + rast->tasks[i].tile.color = align_malloc( TILE_SIZE*TILE_SIZE*4, 16 ); + rast->tasks[i].tile.depth = align_malloc( TILE_SIZE*TILE_SIZE*4, 16 ); + rast->tasks[i].rast = rast; + rast->tasks[i].thread_index = i; + } + + create_rast_threads(rast); + + return rast; +} + /* Shutdown: */ diff --git a/src/gallium/drivers/llvmpipe/lp_rast_priv.h b/src/gallium/drivers/llvmpipe/lp_rast_priv.h index 9e7cbd7912e..62f3c877da5 100644 --- a/src/gallium/drivers/llvmpipe/lp_rast_priv.h +++ b/src/gallium/drivers/llvmpipe/lp_rast_priv.h @@ -28,6 +28,7 @@ #ifndef LP_RAST_PRIV_H #define LP_RAST_PRIV_H +#include "pipe/p_thread.h" #include "lp_rast.h" @@ -36,6 +37,7 @@ struct pipe_transfer; struct pipe_screen; +struct lp_rasterizer; /** @@ -69,6 +71,15 @@ struct lp_rasterizer_task } blocks[256]; const struct lp_rast_state *current_state; + + /** "back" pointer */ + struct lp_rasterizer *rast; + + /** "my" index */ + unsigned thread_index; + + pipe_semaphore work_ready; + pipe_semaphore work_done; }; @@ -104,6 +115,13 @@ struct lp_rasterizer /** A task object for each rasterization thread */ struct lp_rasterizer_task tasks[MAX_THREADS]; + + unsigned num_threads; + pipe_thread threads[MAX_THREADS]; + + struct lp_bins *bins; + const struct pipe_framebuffer_state *fb; + boolean write_depth; }; |