1 /**************************************************************************
3 * Copyright 2009 VMware, Inc.
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sub license, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
21 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
22 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
23 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
24 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
26 **************************************************************************/
29 #include "util/u_memory.h"
30 #include "util/u_math.h"
31 #include "util/u_rect.h"
32 #include "util/u_surface.h"
33 #include "util/u_pack_color.h"
35 #include "lp_scene_queue.h"
41 #include "lp_rast_priv.h"
42 #include "lp_tile_soa.h"
43 #include "gallivm/lp_bld_debug.h"
49 const struct lp_rast_state *jit_state = NULL;
50 const struct lp_rasterizer_task *jit_task = NULL;
55 * Begin rasterizing a scene.
56 * Called once per scene by one thread.
59 lp_rast_begin( struct lp_rasterizer *rast,
60 struct lp_scene *scene )
63 rast->curr_scene = scene;
65 LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
67 lp_scene_begin_rasterization( scene );
68 lp_scene_bin_iter_begin( scene );
73 lp_rast_end( struct lp_rasterizer *rast )
75 lp_scene_end_rasterization( rast->curr_scene );
77 rast->curr_scene = NULL;
81 debug_printf("Post render scene: tile unswizzle: %u tile swizzle: %u\n",
82 lp_tile_unswizzle_count, lp_tile_swizzle_count);
88 * Begining rasterization of a tile.
89 * \param x window X position of the tile, in pixels
90 * \param y window Y position of the tile, in pixels
93 lp_rast_tile_begin(struct lp_rasterizer_task *task,
94 const struct cmd_bin *bin)
96 const struct lp_scene *scene = task->scene;
97 enum lp_texture_usage usage;
99 LP_DBG(DEBUG_RAST, "%s %d,%d\n", __FUNCTION__, bin->x, bin->y);
102 task->x = bin->x * TILE_SIZE;
103 task->y = bin->y * TILE_SIZE;
105 /* reset pointers to color tile(s) */
106 memset(task->color_tiles, 0, sizeof(task->color_tiles));
108 /* get pointer to depth/stencil tile */
110 struct pipe_surface *zsbuf = task->scene->fb.zsbuf;
112 struct llvmpipe_resource *lpt = llvmpipe_resource(zsbuf->texture);
114 if (scene->has_depthstencil_clear)
115 usage = LP_TEX_USAGE_WRITE_ALL;
117 usage = LP_TEX_USAGE_READ_WRITE;
119 /* "prime" the tile: convert data from linear to tiled if necessary
120 * and update the tile's layout info.
122 (void) llvmpipe_get_texture_tile(lpt,
123 zsbuf->u.tex.first_layer,
128 /* Get actual pointer to the tile data. Note that depth/stencil
129 * data is tiled differently than color data.
131 task->depth_tile = lp_rast_get_depth_block_pointer(task,
135 assert(task->depth_tile);
138 task->depth_tile = NULL;
145 * Clear the rasterizer's current color tile.
146 * This is a bin command called during bin processing.
149 lp_rast_clear_color(struct lp_rasterizer_task *task,
150 const union lp_rast_cmd_arg arg)
152 const struct lp_scene *scene = task->scene;
153 const uint8_t *clear_color = arg.clear_color;
157 LP_DBG(DEBUG_RAST, "%s 0x%x,0x%x,0x%x,0x%x\n", __FUNCTION__,
163 if (clear_color[0] == clear_color[1] &&
164 clear_color[1] == clear_color[2] &&
165 clear_color[2] == clear_color[3]) {
166 /* clear to grayscale value {x, x, x, x} */
167 for (i = 0; i < scene->fb.nr_cbufs; i++) {
169 lp_rast_get_color_tile_pointer(task, i, LP_TEX_USAGE_WRITE_ALL);
170 memset(ptr, clear_color[0], TILE_SIZE * TILE_SIZE * 4);
175 * Note: if the swizzled tile layout changes (see TILE_PIXEL) this code
176 * will need to change. It'll be pretty obvious when clearing no longer
179 const unsigned chunk = TILE_SIZE / 4;
180 for (i = 0; i < scene->fb.nr_cbufs; i++) {
182 lp_rast_get_color_tile_pointer(task, i, LP_TEX_USAGE_WRITE_ALL);
185 for (j = 0; j < 4 * TILE_SIZE; j++) {
186 memset(c, clear_color[0], chunk);
188 memset(c, clear_color[1], chunk);
190 memset(c, clear_color[2], chunk);
192 memset(c, clear_color[3], chunk);
198 LP_COUNT(nr_color_tile_clear);
207 * Clear the rasterizer's current z/stencil tile.
208 * This is a bin command called during bin processing.
211 lp_rast_clear_zstencil(struct lp_rasterizer_task *task,
212 const union lp_rast_cmd_arg arg)
214 const struct lp_scene *scene = task->scene;
215 uint32_t clear_value = arg.clear_zstencil.value;
216 uint32_t clear_mask = arg.clear_zstencil.mask;
217 const unsigned height = TILE_SIZE / TILE_VECTOR_HEIGHT;
218 const unsigned width = TILE_SIZE * TILE_VECTOR_HEIGHT;
219 const unsigned block_size = scene->zsbuf.blocksize;
220 const unsigned dst_stride = scene->zsbuf.stride * TILE_VECTOR_HEIGHT;
224 LP_DBG(DEBUG_RAST, "%s: value=0x%08x, mask=0x%08x\n",
225 __FUNCTION__, clear_value, clear_mask);
228 * Clear the aera of the swizzled depth/depth buffer matching this tile, in
229 * stripes of TILE_VECTOR_HEIGHT x TILE_SIZE at a time.
231 * The swizzled depth format is such that the depths for
232 * TILE_VECTOR_HEIGHT x TILE_VECTOR_WIDTH pixels have consecutive offsets.
235 dst = task->depth_tile;
237 clear_value &= clear_mask;
239 switch (block_size) {
241 assert(clear_mask == 0xff);
242 memset(dst, (uint8_t) clear_value, height * width);
245 if (clear_mask == 0xffff) {
246 for (i = 0; i < height; i++) {
247 uint16_t *row = (uint16_t *)dst;
248 for (j = 0; j < width; j++)
249 *row++ = (uint16_t) clear_value;
254 for (i = 0; i < height; i++) {
255 uint16_t *row = (uint16_t *)dst;
256 for (j = 0; j < width; j++) {
257 uint16_t tmp = ~clear_mask & *row;
258 *row++ = clear_value | tmp;
265 if (clear_mask == 0xffffffff) {
266 for (i = 0; i < height; i++) {
267 uint32_t *row = (uint32_t *)dst;
268 for (j = 0; j < width; j++)
269 *row++ = clear_value;
274 for (i = 0; i < height; i++) {
275 uint32_t *row = (uint32_t *)dst;
276 for (j = 0; j < width; j++) {
277 uint32_t tmp = ~clear_mask & *row;
278 *row++ = clear_value | tmp;
293 * Convert the color tile from tiled to linear layout.
294 * This is generally only done when we're flushing the scene just prior to
295 * SwapBuffers. If we didn't do this here, we'd have to convert the entire
296 * tiled color buffer to linear layout in the llvmpipe_texture_unmap()
297 * function. It's better to do it here to take advantage of
298 * threading/parallelism.
299 * This is a bin command which is stored in all bins.
302 lp_rast_store_linear_color( struct lp_rasterizer_task *task )
304 const struct lp_scene *scene = task->scene;
307 for (buf = 0; buf < scene->fb.nr_cbufs; buf++) {
308 struct pipe_surface *cbuf = scene->fb.cbufs[buf];
309 const unsigned layer = cbuf->u.tex.first_layer;
310 const unsigned level = cbuf->u.tex.level;
311 struct llvmpipe_resource *lpt = llvmpipe_resource(cbuf->texture);
313 if (!task->color_tiles[buf])
316 llvmpipe_unswizzle_cbuf_tile(lpt,
320 task->color_tiles[buf]);
327 * Run the shader on all blocks in a tile. This is used when a tile is
328 * completely contained inside a triangle.
329 * This is a bin command called during bin processing.
332 lp_rast_shade_tile(struct lp_rasterizer_task *task,
333 const union lp_rast_cmd_arg arg)
335 const struct lp_scene *scene = task->scene;
336 const struct lp_rast_shader_inputs *inputs = arg.shade_tile;
337 const struct lp_rast_state *state = task->state;
338 struct lp_fragment_shader_variant *variant = state->variant;
339 const unsigned tile_x = task->x, tile_y = task->y;
342 if (inputs->disable) {
343 /* This command was partially binned and has been disabled */
347 LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
349 /* render the whole 64x64 tile in 4x4 chunks */
350 for (y = 0; y < TILE_SIZE; y += 4){
351 for (x = 0; x < TILE_SIZE; x += 4) {
352 uint8_t *color[PIPE_MAX_COLOR_BUFS];
357 for (i = 0; i < scene->fb.nr_cbufs; i++)
358 color[i] = lp_rast_get_color_block_pointer(task, i,
359 tile_x + x, tile_y + y);
362 depth = lp_rast_get_depth_block_pointer(task, tile_x + x, tile_y + y);
364 /* run shader on 4x4 block */
365 BEGIN_JIT_CALL(state, task);
366 variant->jit_function[RAST_WHOLE]( &state->jit_context,
367 tile_x + x, tile_y + y,
383 * Run the shader on all blocks in a tile. This is used when a tile is
384 * completely contained inside a triangle, and the shader is opaque.
385 * This is a bin command called during bin processing.
388 lp_rast_shade_tile_opaque(struct lp_rasterizer_task *task,
389 const union lp_rast_cmd_arg arg)
391 const struct lp_scene *scene = task->scene;
394 LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
396 /* this will prevent converting the layout from tiled to linear */
397 for (i = 0; i < scene->fb.nr_cbufs; i++) {
398 (void)lp_rast_get_color_tile_pointer(task, i, LP_TEX_USAGE_WRITE_ALL);
401 lp_rast_shade_tile(task, arg);
406 * Compute shading for a 4x4 block of pixels inside a triangle.
407 * This is a bin command called during bin processing.
408 * \param x X position of quad in window coords
409 * \param y Y position of quad in window coords
412 lp_rast_shade_quads_mask(struct lp_rasterizer_task *task,
413 const struct lp_rast_shader_inputs *inputs,
414 unsigned x, unsigned y,
417 const struct lp_rast_state *state = task->state;
418 struct lp_fragment_shader_variant *variant = state->variant;
419 const struct lp_scene *scene = task->scene;
420 uint8_t *color[PIPE_MAX_COLOR_BUFS];
427 assert(x % TILE_VECTOR_WIDTH == 0);
428 assert(y % TILE_VECTOR_HEIGHT == 0);
430 assert((x % 4) == 0);
431 assert((y % 4) == 0);
434 for (i = 0; i < scene->fb.nr_cbufs; i++) {
435 color[i] = lp_rast_get_color_block_pointer(task, i, x, y);
436 assert(lp_check_alignment(color[i], 16));
440 depth = lp_rast_get_depth_block_pointer(task, x, y);
443 assert(lp_check_alignment(state->jit_context.blend_color, 16));
445 /* run shader on 4x4 block */
446 BEGIN_JIT_CALL(state, task);
447 variant->jit_function[RAST_EDGE_TEST](&state->jit_context,
463 * Begin a new occlusion query.
464 * This is a bin command put in all bins.
468 lp_rast_begin_query(struct lp_rasterizer_task *task,
469 const union lp_rast_cmd_arg arg)
471 struct llvmpipe_query *pq = arg.query_obj;
473 assert(task->query == NULL);
474 task->vis_counter = 0;
480 * End the current occlusion query.
481 * This is a bin command put in all bins.
485 lp_rast_end_query(struct lp_rasterizer_task *task,
486 const union lp_rast_cmd_arg arg)
490 task->query->count[task->thread_index] += task->vis_counter;
497 lp_rast_set_state(struct lp_rasterizer_task *task,
498 const union lp_rast_cmd_arg arg)
500 task->state = arg.state;
506 * Set top row and left column of the tile's pixels to white. For debugging.
509 outline_tile(uint8_t *tile)
511 const uint8_t val = 0xff;
514 for (i = 0; i < TILE_SIZE; i++) {
515 TILE_PIXEL(tile, i, 0, 0) = val;
516 TILE_PIXEL(tile, i, 0, 1) = val;
517 TILE_PIXEL(tile, i, 0, 2) = val;
518 TILE_PIXEL(tile, i, 0, 3) = val;
520 TILE_PIXEL(tile, 0, i, 0) = val;
521 TILE_PIXEL(tile, 0, i, 1) = val;
522 TILE_PIXEL(tile, 0, i, 2) = val;
523 TILE_PIXEL(tile, 0, i, 3) = val;
529 * Draw grid of gray lines at 16-pixel intervals across the tile to
530 * show the sub-tile boundaries. For debugging.
533 outline_subtiles(uint8_t *tile)
535 const uint8_t val = 0x80;
536 const unsigned step = 16;
539 for (i = 0; i < TILE_SIZE; i += step) {
540 for (j = 0; j < TILE_SIZE; j++) {
541 TILE_PIXEL(tile, i, j, 0) = val;
542 TILE_PIXEL(tile, i, j, 1) = val;
543 TILE_PIXEL(tile, i, j, 2) = val;
544 TILE_PIXEL(tile, i, j, 3) = val;
546 TILE_PIXEL(tile, j, i, 0) = val;
547 TILE_PIXEL(tile, j, i, 1) = val;
548 TILE_PIXEL(tile, j, i, 2) = val;
549 TILE_PIXEL(tile, j, i, 3) = val;
559 * Called when we're done writing to a color tile.
562 lp_rast_tile_end(struct lp_rasterizer_task *task)
565 if (LP_DEBUG & (DEBUG_SHOW_SUBTILES | DEBUG_SHOW_TILES)) {
566 const struct lp_scene *scene = task->scene;
569 for (buf = 0; buf < scene->fb.nr_cbufs; buf++) {
570 uint8_t *color = lp_rast_get_color_block_pointer(task, buf,
573 if (LP_DEBUG & DEBUG_SHOW_SUBTILES)
574 outline_subtiles(color);
575 else if (LP_DEBUG & DEBUG_SHOW_TILES)
580 (void) outline_subtiles;
583 lp_rast_store_linear_color(task);
586 union lp_rast_cmd_arg dummy = {0};
587 lp_rast_end_query(task, dummy);
591 memset(task->color_tiles, 0, sizeof(task->color_tiles));
592 task->depth_tile = NULL;
597 static lp_rast_cmd_func dispatch[LP_RAST_OP_MAX] =
600 lp_rast_clear_zstencil,
609 lp_rast_triangle_3_4,
610 lp_rast_triangle_3_16,
611 lp_rast_triangle_4_16,
613 lp_rast_shade_tile_opaque,
621 do_rasterize_bin(struct lp_rasterizer_task *task,
622 const struct cmd_bin *bin)
624 const struct cmd_block *block;
630 for (block = bin->head; block; block = block->next) {
631 for (k = 0; k < block->count; k++) {
632 dispatch[block->cmd[k]]( task, block->arg[k] );
640 * Rasterize commands for a single bin.
641 * \param x, y position of the bin's tile in the framebuffer
642 * Must be called between lp_rast_begin() and lp_rast_end().
646 rasterize_bin(struct lp_rasterizer_task *task,
647 const struct cmd_bin *bin )
649 lp_rast_tile_begin( task, bin );
651 do_rasterize_bin(task, bin);
653 lp_rast_tile_end(task);
658 if (bin->head->count == 1) {
659 if (bin->head->cmd[0] == LP_RAST_OP_SHADE_TILE_OPAQUE)
660 LP_COUNT(nr_pure_shade_opaque_64);
661 else if (bin->head->cmd[0] == LP_RAST_OP_SHADE_TILE)
662 LP_COUNT(nr_pure_shade_64);
667 /* An empty bin is one that just loads the contents of the tile and
668 * stores them again unchanged. This typically happens when bins have
669 * been flushed for some reason in the middle of a frame, or when
670 * incremental updates are being made to a render target.
672 * Try to avoid doing pointless work in this case.
675 is_empty_bin( const struct cmd_bin *bin )
677 return bin->head == NULL;
682 * Rasterize/execute all bins within a scene.
686 rasterize_scene(struct lp_rasterizer_task *task,
687 struct lp_scene *scene)
690 /* loop over scene bins, rasterize each */
694 for (i = 0; i < scene->tiles_x; i++) {
695 for (j = 0; j < scene->tiles_y; j++) {
696 struct cmd_bin *bin = lp_scene_get_bin(scene, i, j);
697 rasterize_bin(task, bin, i, j);
706 while ((bin = lp_scene_bin_iter_next(scene))) {
707 if (!is_empty_bin( bin ))
708 rasterize_bin(task, bin);
714 lp_fence_signal(scene->fence);
722 * Called by setup module when it has something for us to render.
725 lp_rast_queue_scene( struct lp_rasterizer *rast,
726 struct lp_scene *scene)
728 LP_DBG(DEBUG_SETUP, "%s\n", __FUNCTION__);
730 if (rast->num_threads == 0) {
733 lp_rast_begin( rast, scene );
735 rasterize_scene( &rast->tasks[0], scene );
739 rast->curr_scene = NULL;
742 /* threaded rendering! */
745 lp_scene_enqueue( rast->full_scenes, scene );
747 /* signal the threads that there's work to do */
748 for (i = 0; i < rast->num_threads; i++) {
749 pipe_semaphore_signal(&rast->tasks[i].work_ready);
753 LP_DBG(DEBUG_SETUP, "%s done \n", __FUNCTION__);
758 lp_rast_finish( struct lp_rasterizer *rast )
760 if (rast->num_threads == 0) {
766 /* wait for work to complete */
767 for (i = 0; i < rast->num_threads; i++) {
768 pipe_semaphore_wait(&rast->tasks[i].work_done);
775 * This is the thread's main entrypoint.
776 * It's a simple loop:
779 * 3. signal that we're done
781 static PIPE_THREAD_ROUTINE( thread_func, init_data )
783 struct lp_rasterizer_task *task = (struct lp_rasterizer_task *) init_data;
784 struct lp_rasterizer *rast = task->rast;
785 boolean debug = false;
790 debug_printf("thread %d waiting for work\n", task->thread_index);
791 pipe_semaphore_wait(&task->work_ready);
796 if (task->thread_index == 0) {
798 * - get next scene to rasterize
799 * - map the framebuffer surfaces
802 lp_scene_dequeue( rast->full_scenes, TRUE ) );
805 /* Wait for all threads to get here so that threads[1+] don't
806 * get a null rast->curr_scene pointer.
808 pipe_barrier_wait( &rast->barrier );
812 debug_printf("thread %d doing work\n", task->thread_index);
814 rasterize_scene(task,
817 /* wait for all threads to finish with this scene */
818 pipe_barrier_wait( &rast->barrier );
820 /* XXX: shouldn't be necessary:
822 if (task->thread_index == 0) {
826 /* signal done with work */
828 debug_printf("thread %d done working\n", task->thread_index);
830 pipe_semaphore_signal(&task->work_done);
838 * Initialize semaphores and spawn the threads.
841 create_rast_threads(struct lp_rasterizer *rast)
845 /* NOTE: if num_threads is zero, we won't use any threads */
846 for (i = 0; i < rast->num_threads; i++) {
847 pipe_semaphore_init(&rast->tasks[i].work_ready, 0);
848 pipe_semaphore_init(&rast->tasks[i].work_done, 0);
849 rast->threads[i] = pipe_thread_create(thread_func,
850 (void *) &rast->tasks[i]);
857 * Create new lp_rasterizer. If num_threads is zero, don't create any
858 * new threads, do rendering synchronously.
859 * \param num_threads number of rasterizer threads to create
861 struct lp_rasterizer *
862 lp_rast_create( unsigned num_threads )
864 struct lp_rasterizer *rast;
867 rast = CALLOC_STRUCT(lp_rasterizer);
871 rast->full_scenes = lp_scene_queue_create();
873 for (i = 0; i < Elements(rast->tasks); i++) {
874 struct lp_rasterizer_task *task = &rast->tasks[i];
876 task->thread_index = i;
879 rast->num_threads = num_threads;
881 create_rast_threads(rast);
883 /* for synchronizing rasterization threads */
884 pipe_barrier_init( &rast->barrier, rast->num_threads );
886 memset(lp_swizzled_cbuf, 0, sizeof lp_swizzled_cbuf);
888 memset(lp_dummy_tile, 0, sizeof lp_dummy_tile);
896 void lp_rast_destroy( struct lp_rasterizer *rast )
900 /* Set exit_flag and signal each thread's work_ready semaphore.
901 * Each thread will be woken up, notice that the exit_flag is set and
902 * break out of its main loop. The thread will then exit.
904 rast->exit_flag = TRUE;
905 for (i = 0; i < rast->num_threads; i++) {
906 pipe_semaphore_signal(&rast->tasks[i].work_ready);
909 /* Wait for threads to terminate before cleaning up per-thread data */
910 for (i = 0; i < rast->num_threads; i++) {
911 pipe_thread_wait(rast->threads[i]);
914 /* Clean up per-thread data */
915 for (i = 0; i < rast->num_threads; i++) {
916 pipe_semaphore_destroy(&rast->tasks[i].work_ready);
917 pipe_semaphore_destroy(&rast->tasks[i].work_done);
920 /* for synchronizing rasterization threads */
921 pipe_barrier_destroy( &rast->barrier );
923 lp_scene_queue_destroy(rast->full_scenes);
929 /** Return number of rasterization threads */
931 lp_rast_get_num_threads( struct lp_rasterizer *rast )
933 return rast->num_threads;