X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fgallium%2Fdrivers%2Fllvmpipe%2Flp_rast.c;h=e2c1b6d5cbaccd56d5bbdf36ecb04e246f34b247;hb=c7f5c9a3dc6350252e73b541bb85ab3ed9e64a9c;hp=3a51800c40db022cea66ef763b5780f638abc791;hpb=182c42c8da6edfa66819eef02f4dea310c1f68d7;p=mesa.git diff --git a/src/gallium/drivers/llvmpipe/lp_rast.c b/src/gallium/drivers/llvmpipe/lp_rast.c index 3a51800c40d..e2c1b6d5cba 100644 --- a/src/gallium/drivers/llvmpipe/lp_rast.c +++ b/src/gallium/drivers/llvmpipe/lp_rast.c @@ -28,7 +28,6 @@ #include #include "util/u_memory.h" #include "util/u_math.h" -#include "util/u_cpu_detect.h" #include "util/u_surface.h" #include "lp_scene_queue.h" @@ -42,15 +41,15 @@ #include "lp_scene.h" -/* Begin rasterizing a scene: +/** + * Begin rasterizing a scene. + * Called once per scene by one thread. */ -static boolean +static void lp_rast_begin( struct lp_rasterizer *rast, struct lp_scene *scene ) { const struct pipe_framebuffer_state *fb = &scene->fb; - boolean write_color = fb->nr_cbufs != 0; - boolean write_zstencil = fb->zsbuf != NULL; int i; rast->curr_scene = scene; @@ -58,59 +57,149 @@ lp_rast_begin( struct lp_rasterizer *rast, LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__); rast->state.nr_cbufs = scene->fb.nr_cbufs; - rast->state.write_zstencil = write_zstencil; - rast->state.write_color = write_color; for (i = 0; i < rast->state.nr_cbufs; i++) { struct pipe_surface *cbuf = scene->fb.cbufs[i]; - rast->cbuf[i].map = scene->cbuf_map[i]; rast->cbuf[i].format = cbuf->texture->format; - rast->cbuf[i].width = cbuf->width; - rast->cbuf[i].height = cbuf->height; - rast->cbuf[i].stride = llvmpipe_texture_stride(cbuf->texture, cbuf->level); + rast->cbuf[i].tiles_per_row = align(cbuf->width, TILE_SIZE) / TILE_SIZE; + rast->cbuf[i].blocksize = + util_format_get_blocksize(cbuf->texture->format); + rast->cbuf[i].map = llvmpipe_resource_map(cbuf->texture, + cbuf->face, + cbuf->level, + cbuf->zslice, + LP_TEX_USAGE_READ_WRITE, + LP_TEX_LAYOUT_NONE); } - if (write_zstencil) { + if (fb->zsbuf) { struct pipe_surface *zsbuf = scene->fb.zsbuf; - rast->zsbuf.map = scene->zsbuf_map; - rast->zsbuf.stride = llvmpipe_texture_stride(zsbuf->texture, zsbuf->level); + rast->zsbuf.stride = llvmpipe_resource_stride(zsbuf->texture, zsbuf->level); rast->zsbuf.blocksize = util_format_get_blocksize(zsbuf->texture->format); + + rast->zsbuf.map = llvmpipe_resource_map(zsbuf->texture, + zsbuf->face, + zsbuf->level, + zsbuf->zslice, + LP_TEX_USAGE_READ_WRITE, + LP_TEX_LAYOUT_NONE); + assert(rast->zsbuf.map); } lp_scene_bin_iter_begin( scene ); - - return TRUE; } static void lp_rast_end( struct lp_rasterizer *rast ) { - int i; - - lp_scene_reset( rast->curr_scene ); + struct lp_scene *scene = rast->curr_scene; + unsigned i; - for (i = 0; i < rast->state.nr_cbufs; i++) + /* Unmap color buffers */ + for (i = 0; i < rast->state.nr_cbufs; i++) { + struct pipe_surface *cbuf = scene->fb.cbufs[i]; + llvmpipe_resource_unmap(cbuf->texture, + cbuf->face, + cbuf->level, + cbuf->zslice); rast->cbuf[i].map = NULL; + } + + /* Unmap z/stencil buffer */ + if (rast->zsbuf.map) { + struct pipe_surface *zsbuf = scene->fb.zsbuf; + llvmpipe_resource_unmap(zsbuf->texture, + zsbuf->face, + zsbuf->level, + zsbuf->zslice); + rast->zsbuf.map = NULL; + } + + lp_scene_reset( rast->curr_scene ); - rast->zsbuf.map = NULL; rast->curr_scene = NULL; + +#ifdef DEBUG + if (0) + debug_printf("Post render scene: tile unswizzle: %u tile swizzle: %u\n", + lp_tile_unswizzle_count, lp_tile_swizzle_count); +#endif } + /** * Begining rasterization of a tile. * \param x window X position of the tile, in pixels * \param y window Y position of the tile, in pixels */ static void -lp_rast_start_tile(struct lp_rasterizer_task *task, +lp_rast_tile_begin(struct lp_rasterizer_task *task, unsigned x, unsigned y) { + struct lp_rasterizer *rast = task->rast; + struct lp_scene *scene = rast->curr_scene; + enum lp_texture_usage usage; + unsigned buf; + LP_DBG(DEBUG_RAST, "%s %d,%d\n", __FUNCTION__, x, y); + assert(x % TILE_SIZE == 0); + assert(y % TILE_SIZE == 0); + task->x = x; task->y = y; + + if (scene->has_color_clear) + usage = LP_TEX_USAGE_WRITE_ALL; + else + usage = LP_TEX_USAGE_READ_WRITE; + + /* get pointers to color tile(s) */ + for (buf = 0; buf < rast->state.nr_cbufs; buf++) { + struct pipe_surface *cbuf = rast->curr_scene->fb.cbufs[buf]; + struct llvmpipe_resource *lpt; + assert(cbuf); + lpt = llvmpipe_resource(cbuf->texture); + task->color_tiles[buf] = llvmpipe_get_texture_tile(lpt, + cbuf->face + cbuf->zslice, + cbuf->level, + usage, + x, y); + assert(task->color_tiles[buf]); + } + + /* get pointer to depth/stencil tile */ + { + struct pipe_surface *zsbuf = rast->curr_scene->fb.zsbuf; + if (zsbuf) { + struct llvmpipe_resource *lpt = llvmpipe_resource(zsbuf->texture); + + if (scene->has_depth_clear) + usage = LP_TEX_USAGE_WRITE_ALL; + else + usage = LP_TEX_USAGE_READ_WRITE; + + /* "prime" the tile: convert data from linear to tiled if necessary + * and update the tile's layout info. + */ + (void) llvmpipe_get_texture_tile(lpt, + zsbuf->face + zsbuf->zslice, + zsbuf->level, + usage, + x, y); + /* Get actual pointer to the tile data. Note that depth/stencil + * data is tiled differently than color data. + */ + task->depth_tile = lp_rast_get_depth_block_pointer(rast, x, y); + + assert(task->depth_tile); + } + else { + task->depth_tile = NULL; + } + } } @@ -124,7 +213,7 @@ lp_rast_clear_color(struct lp_rasterizer_task *task, { struct lp_rasterizer *rast = task->rast; const uint8_t *clear_color = arg.clear_color; - uint8_t **color_tile = task->tile.color; + unsigned i; LP_DBG(DEBUG_RAST, "%s 0x%x,0x%x,0x%x,0x%x\n", __FUNCTION__, @@ -138,7 +227,8 @@ lp_rast_clear_color(struct lp_rasterizer_task *task, clear_color[2] == clear_color[3]) { /* clear to grayscale value {x, x, x, x} */ for (i = 0; i < rast->state.nr_cbufs; i++) { - memset(color_tile[i], clear_color[0], TILE_SIZE * TILE_SIZE * 4); + uint8_t *ptr = task->color_tiles[i]; + memset(ptr, clear_color[0], TILE_SIZE * TILE_SIZE * 4); } } else { @@ -149,8 +239,9 @@ lp_rast_clear_color(struct lp_rasterizer_task *task, */ const unsigned chunk = TILE_SIZE / 4; for (i = 0; i < rast->state.nr_cbufs; i++) { - uint8_t *c = color_tile[i]; + uint8_t *c = task->color_tiles[i]; unsigned j; + for (j = 0; j < 4 * TILE_SIZE; j++) { memset(c, clear_color[0], chunk); c += chunk; @@ -161,7 +252,6 @@ lp_rast_clear_color(struct lp_rasterizer_task *task, memset(c, clear_color[3], chunk); c += chunk; } - assert(c - color_tile[i] == TILE_SIZE * TILE_SIZE * 4); } } @@ -178,23 +268,15 @@ lp_rast_clear_zstencil(struct lp_rasterizer_task *task, const union lp_rast_cmd_arg arg) { struct lp_rasterizer *rast = task->rast; - const unsigned tile_x = task->x; - const unsigned tile_y = task->y; const unsigned height = TILE_SIZE / TILE_VECTOR_HEIGHT; const unsigned width = TILE_SIZE * TILE_VECTOR_HEIGHT; - unsigned block_size = rast->zsbuf.blocksize; + const unsigned block_size = rast->zsbuf.blocksize; + const unsigned dst_stride = rast->zsbuf.stride * TILE_VECTOR_HEIGHT; uint8_t *dst; - unsigned dst_stride = rast->zsbuf.stride * TILE_VECTOR_HEIGHT; unsigned i, j; LP_DBG(DEBUG_RAST, "%s 0x%x\n", __FUNCTION__, arg.clear_zstencil); - assert(rast->zsbuf.map); - if (!rast->zsbuf.map) - return; - - LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__); - /* * Clear the aera of the swizzled depth/depth buffer matching this tile, in * stripes of TILE_VECTOR_HEIGHT x TILE_SIZE at a time. @@ -203,7 +285,9 @@ lp_rast_clear_zstencil(struct lp_rasterizer_task *task, * TILE_VECTOR_HEIGHT x TILE_VECTOR_WIDTH pixels have consecutive offsets. */ - dst = lp_rast_depth_pointer(rast, tile_x, tile_y); + dst = task->depth_tile; + + assert(dst == lp_rast_get_depth_block_pointer(rast, task->x, task->y)); switch (block_size) { case 1: @@ -236,32 +320,73 @@ lp_rast_clear_zstencil(struct lp_rasterizer_task *task, * Load tile color from the framebuffer surface. * This is a bin command called during bin processing. */ +#if 0 void lp_rast_load_color(struct lp_rasterizer_task *task, const union lp_rast_cmd_arg arg) { struct lp_rasterizer *rast = task->rast; - const unsigned x = task->x, y = task->y; - unsigned i; + unsigned buf; + enum lp_texture_usage usage; LP_DBG(DEBUG_RAST, "%s at %u, %u\n", __FUNCTION__, x, y); - for (i = 0; i < rast->state.nr_cbufs; i++) { - if (x >= rast->cbuf[i].width || y >= rast->cbuf[i].height) - continue; + if (scene->has_color_clear) + usage = LP_TEX_USAGE_WRITE_ALL; + else + usage = LP_TEX_USAGE_READ_WRITE; - lp_tile_read_4ub(rast->cbuf[i].format, - task->tile.color[i], - rast->cbuf[i].map, - rast->cbuf[i].stride, - x, y, - TILE_SIZE, TILE_SIZE); + /* Get pointers to color tile(s). + * This will convert linear data to tiled if needed. + */ + for (buf = 0; buf < rast->state.nr_cbufs; buf++) { + struct pipe_surface *cbuf = rast->curr_scene->fb.cbufs[buf]; + struct llvmpipe_texture *lpt; + assert(cbuf); + lpt = llvmpipe_texture(cbuf->texture); + task->color_tiles[buf] = llvmpipe_get_texture_tile(lpt, + cbuf->face + cbuf->zslice, + cbuf->level, + usage, + task->x, task->y); + assert(task->color_tiles[buf]); + } +} +#endif - LP_COUNT(nr_color_tile_load); + +/** + * Convert the color tile from tiled to linear layout. + * This is generally only done when we're flushing the scene just prior to + * SwapBuffers. If we didn't do this here, we'd have to convert the entire + * tiled color buffer to linear layout in the llvmpipe_texture_unmap() + * function. It's better to do it here to take advantage of + * threading/parallelism. + * This is a bin command which is stored in all bins. + */ +void +lp_rast_store_color( struct lp_rasterizer_task *task, + const union lp_rast_cmd_arg arg) +{ + struct lp_rasterizer *rast = task->rast; + struct lp_scene *scene = rast->curr_scene; + unsigned buf; + + for (buf = 0; buf < rast->state.nr_cbufs; buf++) { + struct pipe_surface *cbuf = scene->fb.cbufs[buf]; + const unsigned face = cbuf->face, level = cbuf->level; + struct llvmpipe_resource *lpt = llvmpipe_resource(cbuf->texture); + /* this will convert the tiled data to linear if needed */ + (void) llvmpipe_get_texture_tile_linear(lpt, face, level, + LP_TEX_USAGE_READ, + task->x, task->y); } } +/** + * This is a bin command called during bin processing. + */ void lp_rast_set_state(struct lp_rasterizer_task *task, const union lp_rast_cmd_arg arg) @@ -275,7 +400,6 @@ lp_rast_set_state(struct lp_rasterizer_task *task, } - /** * Run the shader on all blocks in a tile. This is used when a tile is * completely contained inside a triangle. @@ -287,7 +411,6 @@ lp_rast_shade_tile(struct lp_rasterizer_task *task, { struct lp_rasterizer *rast = task->rast; const struct lp_rast_state *state = task->current_state; - struct lp_rast_tile *tile = &task->tile; const struct lp_rast_shader_inputs *inputs = arg.shade_tile; const unsigned tile_x = task->x, tile_y = task->y; unsigned x, y; @@ -299,19 +422,17 @@ lp_rast_shade_tile(struct lp_rasterizer_task *task, for (x = 0; x < TILE_SIZE; x += 4) { uint8_t *color[PIPE_MAX_COLOR_BUFS]; uint32_t *depth; - unsigned block_offset, i; - - /* offset of the 16x16 pixel block within the tile */ - block_offset = ((y / 4) * (16 * 16) + (x / 4) * 16); + unsigned i; /* color buffer */ for (i = 0; i < rast->state.nr_cbufs; i++) - color[i] = tile->color[i] + 4 * block_offset; + color[i] = lp_rast_get_color_block_pointer(task, i, + tile_x + x, tile_y + y); /* depth buffer */ - depth = lp_rast_depth_pointer(rast, tile_x + x, tile_y + y); + depth = lp_rast_get_depth_block_pointer(rast, tile_x + x, tile_y + y); - /* run shader */ + /* run shader on 4x4 block */ state->jit_function[RAST_WHOLE]( &state->jit_context, tile_x + x, tile_y + y, inputs->facing, @@ -330,6 +451,8 @@ lp_rast_shade_tile(struct lp_rasterizer_task *task, /** * Compute shading for a 4x4 block of pixels. * This is a bin command called during bin processing. + * \param x X position of quad in window coords + * \param y Y position of quad in window coords */ void lp_rast_shade_quads( struct lp_rasterizer_task *task, const struct lp_rast_shader_inputs *inputs, @@ -338,12 +461,9 @@ void lp_rast_shade_quads( struct lp_rasterizer_task *task, { const struct lp_rast_state *state = task->current_state; struct lp_rasterizer *rast = task->rast; - struct lp_rast_tile *tile = &task->tile; uint8_t *color[PIPE_MAX_COLOR_BUFS]; void *depth; unsigned i; - unsigned ix, iy; - int block_offset; assert(state); @@ -354,28 +474,23 @@ void lp_rast_shade_quads( struct lp_rasterizer_task *task, assert((x % 4) == 0); assert((y % 4) == 0); - ix = x % TILE_SIZE; - iy = y % TILE_SIZE; - - /* offset of the 16x16 pixel block within the tile */ - block_offset = ((iy / 4) * (16 * 16) + (ix / 4) * 16); - /* color buffer */ - for (i = 0; i < rast->state.nr_cbufs; i++) - color[i] = tile->color[i] + 4 * block_offset; + for (i = 0; i < rast->state.nr_cbufs; i++) { + color[i] = lp_rast_get_color_block_pointer(task, i, x, y); + assert(lp_check_alignment(color[i], 16)); + } /* depth buffer */ - depth = lp_rast_depth_pointer(rast, x, y); + depth = lp_rast_get_depth_block_pointer(rast, x, y); - assert(lp_check_alignment(tile->color[0], 16)); assert(lp_check_alignment(state->jit_context.blend_color, 16)); assert(lp_check_alignment(inputs->step[0], 16)); assert(lp_check_alignment(inputs->step[1], 16)); assert(lp_check_alignment(inputs->step[2], 16)); - /* run shader */ + /* run shader on 4x4 block */ state->jit_function[RAST_EDGE_TEST]( &state->jit_context, x, y, inputs->facing, @@ -445,39 +560,31 @@ outline_subtiles(uint8_t *tile) /** - * Write the rasterizer's color tile to the framebuffer. + * Called when we're done writing to a color tile. */ static void -lp_rast_store_color(struct lp_rasterizer_task *task) +lp_rast_tile_end(struct lp_rasterizer_task *task) { +#if DEBUG struct lp_rasterizer *rast = task->rast; - const unsigned x = task->x, y = task->y; - unsigned i; - - for (i = 0; i < rast->state.nr_cbufs; i++) { - if (x >= rast->cbuf[i].width) - continue; + unsigned buf; - if (y >= rast->cbuf[i].height) - continue; - - LP_DBG(DEBUG_RAST, "%s [%u] %d,%d\n", __FUNCTION__, - task->thread_index, x, y); + for (buf = 0; buf < rast->state.nr_cbufs; buf++) { + uint8_t *color = lp_rast_get_color_block_pointer(task, buf, + task->x, task->y); if (LP_DEBUG & DEBUG_SHOW_SUBTILES) - outline_subtiles(task->tile.color[i]); + outline_subtiles(color); else if (LP_DEBUG & DEBUG_SHOW_TILES) - outline_tile(task->tile.color[i]); - - lp_tile_write_4ub(rast->cbuf[i].format, - task->tile.color[i], - rast->cbuf[i].map, - rast->cbuf[i].stride, - x, y, - TILE_SIZE, TILE_SIZE); - - LP_COUNT(nr_color_tile_store); + outline_tile(color); } +#else + (void) outline_subtiles; +#endif + + /* debug */ + memset(task->color_tiles, 0, sizeof(task->color_tiles)); + task->depth_tile = NULL; } @@ -491,18 +598,7 @@ lp_rast_fence(struct lp_rasterizer_task *task, const union lp_rast_cmd_arg arg) { struct lp_fence *fence = arg.fence; - - pipe_mutex_lock( fence->mutex ); - - fence->count++; - assert(fence->count <= fence->rank); - - LP_DBG(DEBUG_RAST, "%s count=%u rank=%u\n", __FUNCTION__, - fence->count, fence->rank); - - pipe_condvar_signal( fence->signalled ); - - pipe_mutex_unlock( fence->mutex ); + lp_fence_signal(fence); } @@ -523,7 +619,7 @@ rasterize_bin(struct lp_rasterizer_task *task, struct cmd_block *block; unsigned k; - lp_rast_start_tile( task, x * TILE_SIZE, y * TILE_SIZE ); + lp_rast_tile_begin( task, x * TILE_SIZE, y * TILE_SIZE ); /* simply execute each of the commands in the block list */ for (block = commands->head; block; block = block->next) { @@ -532,10 +628,7 @@ rasterize_bin(struct lp_rasterizer_task *task, } } - /* Write the rasterizer's tiles to the framebuffer. - */ - if (task->rast->state.write_color) - lp_rast_store_color(task); + lp_rast_tile_end(task); /* Free data for this bin. */ @@ -550,12 +643,12 @@ static struct { const char *name; } cmd_names[] = { - RAST(load_color), RAST(clear_color), RAST(clear_zstencil), RAST(triangle), RAST(shade_tile), RAST(set_state), + RAST(store_color), RAST(fence), }; @@ -608,8 +701,7 @@ is_empty_bin( const struct cmd_bin *bin ) } for (i = 0; i < head->count; i++) - if (head->cmd[i] != lp_rast_load_color && - head->cmd[i] != lp_rast_set_state) { + if (head->cmd[i] != lp_rast_set_state) { return FALSE; } @@ -669,6 +761,9 @@ lp_rast_queue_scene( struct lp_rasterizer *rast, rasterize_scene( &rast->tasks[0], scene ); lp_scene_reset( scene ); + + lp_rast_end( rast ); + rast->curr_scene = NULL; } else { @@ -775,16 +870,6 @@ create_rast_threads(struct lp_rasterizer *rast) { unsigned i; -#ifdef PIPE_OS_WINDOWS - /* Multithreading not supported on windows until conditions and barriers are - * properly implemented. */ - rast->num_threads = 0; -#else - rast->num_threads = util_cpu_caps.nr_cpus; - rast->num_threads = debug_get_num_option("LP_NUM_THREADS", rast->num_threads); - rast->num_threads = MIN2(rast->num_threads, MAX_THREADS); -#endif - /* 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); @@ -797,15 +882,15 @@ create_rast_threads(struct lp_rasterizer *rast) /** - * Create new lp_rasterizer. - * \param empty the queue to put empty scenes on after we've finished - * processing them. + * Create new lp_rasterizer. If num_threads is zero, don't create any + * new threads, do rendering synchronously. + * \param num_threads number of rasterizer threads to create */ struct lp_rasterizer * -lp_rast_create( void ) +lp_rast_create( unsigned num_threads ) { struct lp_rasterizer *rast; - unsigned i, cbuf; + unsigned i; rast = CALLOC_STRUCT(lp_rasterizer); if(!rast) @@ -815,14 +900,12 @@ lp_rast_create( void ) for (i = 0; i < Elements(rast->tasks); i++) { struct lp_rasterizer_task *task = &rast->tasks[i]; - - for (cbuf = 0; cbuf < PIPE_MAX_COLOR_BUFS; cbuf++ ) - task->tile.color[cbuf] = align_malloc(TILE_SIZE * TILE_SIZE * 4, 16); - task->rast = rast; task->thread_index = i; } + rast->num_threads = num_threads; + create_rast_threads(rast); /* for synchronizing rasterization threads */ @@ -836,12 +919,7 @@ lp_rast_create( void ) */ void lp_rast_destroy( struct lp_rasterizer *rast ) { - unsigned i, cbuf; - - for (i = 0; i < Elements(rast->tasks); i++) { - for (cbuf = 0; cbuf < PIPE_MAX_COLOR_BUFS; cbuf++ ) - align_free(rast->tasks[i].tile.color[cbuf]); - } + unsigned i; /* Set exit_flag and signal each thread's work_ready semaphore. * Each thread will be woken up, notice that the exit_flag is set and