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 "os/os_time.h"
37 #include "lp_scene_queue.h"
43 #include "lp_rast_priv.h"
44 #include "gallivm/lp_bld_debug.h"
46 #include "lp_tex_sample.h"
51 const struct lp_rast_state
*jit_state
= NULL
;
52 const struct lp_rasterizer_task
*jit_task
= NULL
;
57 * Begin rasterizing a scene.
58 * Called once per scene by one thread.
61 lp_rast_begin( struct lp_rasterizer
*rast
,
62 struct lp_scene
*scene
)
65 rast
->curr_scene
= scene
;
67 LP_DBG(DEBUG_RAST
, "%s\n", __FUNCTION__
);
69 lp_scene_begin_rasterization( scene
);
70 lp_scene_bin_iter_begin( scene
);
75 lp_rast_end( struct lp_rasterizer
*rast
)
77 lp_scene_end_rasterization( rast
->curr_scene
);
79 rast
->curr_scene
= NULL
;
84 * Begining rasterization of a tile.
85 * \param x window X position of the tile, in pixels
86 * \param y window Y position of the tile, in pixels
89 lp_rast_tile_begin(struct lp_rasterizer_task
*task
,
90 const struct cmd_bin
*bin
)
92 const struct lp_scene
*scene
= task
->scene
;
93 enum lp_texture_usage usage
;
95 LP_DBG(DEBUG_RAST
, "%s %d,%d\n", __FUNCTION__
, bin
->x
, bin
->y
);
98 task
->x
= bin
->x
* TILE_SIZE
;
99 task
->y
= bin
->y
* TILE_SIZE
;
101 /* reset pointers to color tile(s) */
102 memset(task
->color_tiles
, 0, sizeof(task
->color_tiles
));
104 /* get pointer to depth/stencil tile */
106 struct pipe_surface
*zsbuf
= task
->scene
->fb
.zsbuf
;
108 struct llvmpipe_resource
*lpt
= llvmpipe_resource(zsbuf
->texture
);
110 if (scene
->has_depthstencil_clear
)
111 usage
= LP_TEX_USAGE_WRITE_ALL
;
113 usage
= LP_TEX_USAGE_READ_WRITE
;
115 /* "prime" the tile: convert data from linear to tiled if necessary
116 * and update the tile's layout info.
118 (void) llvmpipe_get_texture_tile(lpt
,
119 zsbuf
->u
.tex
.first_layer
,
124 /* Get actual pointer to the tile data. Note that depth/stencil
125 * data is tiled differently than color data.
127 task
->depth_tile
= lp_rast_get_depth_block_pointer(task
,
131 assert(task
->depth_tile
);
134 task
->depth_tile
= NULL
;
141 * Clear the rasterizer's current color tile.
142 * This is a bin command called during bin processing.
145 lp_rast_clear_color(struct lp_rasterizer_task
*task
,
146 const union lp_rast_cmd_arg arg
)
148 const struct lp_scene
*scene
= task
->scene
;
150 if (scene
->fb
.nr_cbufs
) {
154 if (util_format_is_pure_integer(scene
->fb
.cbufs
[0]->format
)) {
156 * We expect int/uint clear values here, though some APIs
157 * might disagree (but in any case util_pack_color()
158 * couldn't handle it)...
160 LP_DBG(DEBUG_RAST
, "%s pure int 0x%x,0x%x,0x%x,0x%x\n", __FUNCTION__
,
161 arg
.clear_color
.ui
[0],
162 arg
.clear_color
.ui
[1],
163 arg
.clear_color
.ui
[2],
164 arg
.clear_color
.ui
[3]);
166 for (i
= 0; i
< scene
->fb
.nr_cbufs
; i
++) {
167 enum pipe_format format
= scene
->fb
.cbufs
[i
]->format
;
169 if (util_format_is_pure_sint(format
)) {
170 util_format_write_4i(format
, arg
.clear_color
.i
, 0, &uc
, 0, 0, 0, 1, 1);
173 assert(util_format_is_pure_uint(format
));
174 util_format_write_4ui(format
, arg
.clear_color
.ui
, 0, &uc
, 0, 0, 0, 1, 1);
177 util_fill_rect(scene
->cbufs
[i
].map
,
178 scene
->fb
.cbufs
[i
]->format
,
179 scene
->cbufs
[i
].stride
,
188 uint8_t clear_color
[4];
190 for (i
= 0; i
< 4; ++i
) {
191 clear_color
[i
] = float_to_ubyte(arg
.clear_color
.f
[i
]);
194 LP_DBG(DEBUG_RAST
, "%s 0x%x,0x%x,0x%x,0x%x\n", __FUNCTION__
,
200 for (i
= 0; i
< scene
->fb
.nr_cbufs
; i
++) {
202 util_pack_color(arg
.clear_color
.f
,
203 scene
->fb
.cbufs
[i
]->format
, &uc
);
205 util_fill_rect(scene
->cbufs
[i
].map
,
206 scene
->fb
.cbufs
[i
]->format
,
207 scene
->cbufs
[i
].stride
,
217 LP_COUNT(nr_color_tile_clear
);
226 * Clear the rasterizer's current z/stencil tile.
227 * This is a bin command called during bin processing.
230 lp_rast_clear_zstencil(struct lp_rasterizer_task
*task
,
231 const union lp_rast_cmd_arg arg
)
233 const struct lp_scene
*scene
= task
->scene
;
234 uint32_t clear_value
= arg
.clear_zstencil
.value
;
235 uint32_t clear_mask
= arg
.clear_zstencil
.mask
;
236 const unsigned height
= TILE_SIZE
/ TILE_VECTOR_HEIGHT
;
237 const unsigned width
= TILE_SIZE
* TILE_VECTOR_HEIGHT
;
238 const unsigned block_size
= scene
->zsbuf
.blocksize
;
239 const unsigned dst_stride
= scene
->zsbuf
.stride
* TILE_VECTOR_HEIGHT
;
243 LP_DBG(DEBUG_RAST
, "%s: value=0x%08x, mask=0x%08x\n",
244 __FUNCTION__
, clear_value
, clear_mask
);
247 * Clear the area of the swizzled depth/depth buffer matching this tile, in
248 * stripes of TILE_VECTOR_HEIGHT x TILE_SIZE at a time.
250 * The swizzled depth format is such that the depths for
251 * TILE_VECTOR_HEIGHT x TILE_VECTOR_WIDTH pixels have consecutive offsets.
254 dst
= task
->depth_tile
;
256 clear_value
&= clear_mask
;
258 switch (block_size
) {
260 assert(clear_mask
== 0xff);
261 memset(dst
, (uint8_t) clear_value
, height
* width
);
264 if (clear_mask
== 0xffff) {
265 for (i
= 0; i
< height
; i
++) {
266 uint16_t *row
= (uint16_t *)dst
;
267 for (j
= 0; j
< width
; j
++)
268 *row
++ = (uint16_t) clear_value
;
273 for (i
= 0; i
< height
; i
++) {
274 uint16_t *row
= (uint16_t *)dst
;
275 for (j
= 0; j
< width
; j
++) {
276 uint16_t tmp
= ~clear_mask
& *row
;
277 *row
++ = clear_value
| tmp
;
284 if (clear_mask
== 0xffffffff) {
285 for (i
= 0; i
< height
; i
++) {
286 uint32_t *row
= (uint32_t *)dst
;
287 for (j
= 0; j
< width
; j
++)
288 *row
++ = clear_value
;
293 for (i
= 0; i
< height
; i
++) {
294 uint32_t *row
= (uint32_t *)dst
;
295 for (j
= 0; j
< width
; j
++) {
296 uint32_t tmp
= ~clear_mask
& *row
;
297 *row
++ = clear_value
| tmp
;
312 * Run the shader on all blocks in a tile. This is used when a tile is
313 * completely contained inside a triangle.
314 * This is a bin command called during bin processing.
317 lp_rast_shade_tile(struct lp_rasterizer_task
*task
,
318 const union lp_rast_cmd_arg arg
)
320 const struct lp_scene
*scene
= task
->scene
;
321 const struct lp_rast_shader_inputs
*inputs
= arg
.shade_tile
;
322 const struct lp_rast_state
*state
;
323 struct lp_fragment_shader_variant
*variant
;
324 const unsigned tile_x
= task
->x
, tile_y
= task
->y
;
327 if (inputs
->disable
) {
328 /* This command was partially binned and has been disabled */
332 LP_DBG(DEBUG_RAST
, "%s\n", __FUNCTION__
);
339 variant
= state
->variant
;
341 /* render the whole 64x64 tile in 4x4 chunks */
342 for (y
= 0; y
< TILE_SIZE
; y
+= 4){
343 for (x
= 0; x
< TILE_SIZE
; x
+= 4) {
344 uint8_t *color
[PIPE_MAX_COLOR_BUFS
];
345 unsigned stride
[PIPE_MAX_COLOR_BUFS
];
350 for (i
= 0; i
< scene
->fb
.nr_cbufs
; i
++){
351 stride
[i
] = scene
->cbufs
[i
].stride
;
353 color
[i
] = lp_rast_get_unswizzled_color_block_pointer(task
, i
, tile_x
+ x
, tile_y
+ y
);
357 depth
= lp_rast_get_depth_block_pointer(task
, tile_x
+ x
, tile_y
+ y
);
359 /* run shader on 4x4 block */
360 BEGIN_JIT_CALL(state
, task
);
361 variant
->jit_function
[RAST_WHOLE
]( &state
->jit_context
,
362 tile_x
+ x
, tile_y
+ y
,
379 * Run the shader on all blocks in a tile. This is used when a tile is
380 * completely contained inside a triangle, and the shader is opaque.
381 * This is a bin command called during bin processing.
384 lp_rast_shade_tile_opaque(struct lp_rasterizer_task
*task
,
385 const union lp_rast_cmd_arg arg
)
387 LP_DBG(DEBUG_RAST
, "%s\n", __FUNCTION__
);
394 lp_rast_shade_tile(task
, arg
);
399 * Compute shading for a 4x4 block of pixels inside a triangle.
400 * This is a bin command called during bin processing.
401 * \param x X position of quad in window coords
402 * \param y Y position of quad in window coords
405 lp_rast_shade_quads_mask(struct lp_rasterizer_task
*task
,
406 const struct lp_rast_shader_inputs
*inputs
,
407 unsigned x
, unsigned y
,
410 const struct lp_rast_state
*state
= task
->state
;
411 struct lp_fragment_shader_variant
*variant
= state
->variant
;
412 const struct lp_scene
*scene
= task
->scene
;
413 uint8_t *color
[PIPE_MAX_COLOR_BUFS
];
414 unsigned stride
[PIPE_MAX_COLOR_BUFS
];
421 assert(x
< scene
->tiles_x
* TILE_SIZE
);
422 assert(y
< scene
->tiles_y
* TILE_SIZE
);
423 assert(x
% TILE_VECTOR_WIDTH
== 0);
424 assert(y
% TILE_VECTOR_HEIGHT
== 0);
426 assert((x
% 4) == 0);
427 assert((y
% 4) == 0);
430 for (i
= 0; i
< scene
->fb
.nr_cbufs
; i
++) {
431 stride
[i
] = scene
->cbufs
[i
].stride
;
433 color
[i
] = lp_rast_get_unswizzled_color_block_pointer(task
, i
, x
, y
);
437 depth
= lp_rast_get_depth_block_pointer(task
, x
, y
);
440 assert(lp_check_alignment(state
->jit_context
.u8_blend_color
, 16));
442 /* run shader on 4x4 block */
443 BEGIN_JIT_CALL(state
, task
);
444 variant
->jit_function
[RAST_EDGE_TEST
](&state
->jit_context
,
461 * Begin a new occlusion query.
462 * This is a bin command put in all bins.
466 lp_rast_begin_query(struct lp_rasterizer_task
*task
,
467 const union lp_rast_cmd_arg arg
)
469 struct llvmpipe_query
*pq
= arg
.query_obj
;
471 assert(task
->query
[pq
->type
] == NULL
);
474 case PIPE_QUERY_OCCLUSION_COUNTER
:
475 task
->thread_data
.vis_counter
= 0;
477 case PIPE_QUERY_PRIMITIVES_GENERATED
:
478 case PIPE_QUERY_PRIMITIVES_EMITTED
:
479 case PIPE_QUERY_SO_STATISTICS
:
480 case PIPE_QUERY_PIPELINE_STATISTICS
:
487 task
->query
[pq
->type
] = pq
;
492 * End the current occlusion query.
493 * This is a bin command put in all bins.
497 lp_rast_end_query(struct lp_rasterizer_task
*task
,
498 const union lp_rast_cmd_arg arg
)
500 struct llvmpipe_query
*pq
= arg
.query_obj
;
501 assert(task
->query
[pq
->type
] == pq
|| pq
->type
== PIPE_QUERY_TIMESTAMP
);
504 case PIPE_QUERY_OCCLUSION_COUNTER
:
505 pq
->count
[task
->thread_index
] += task
->thread_data
.vis_counter
;
507 case PIPE_QUERY_TIMESTAMP
:
508 pq
->count
[task
->thread_index
] = os_time_get_nano();
510 case PIPE_QUERY_PRIMITIVES_GENERATED
:
511 case PIPE_QUERY_PRIMITIVES_EMITTED
:
512 case PIPE_QUERY_SO_STATISTICS
:
513 case PIPE_QUERY_PIPELINE_STATISTICS
:
520 if (task
->query
[pq
->type
] == pq
) {
521 task
->query
[pq
->type
] = NULL
;
527 lp_rast_set_state(struct lp_rasterizer_task
*task
,
528 const union lp_rast_cmd_arg arg
)
530 task
->state
= arg
.state
;
536 * Called when we're done writing to a color tile.
539 lp_rast_tile_end(struct lp_rasterizer_task
*task
)
543 for (i
= 0; i
< PIPE_QUERY_TYPES
; ++i
) {
544 if (task
->query
[i
]) {
545 lp_rast_end_query(task
, lp_rast_arg_query(task
->query
[i
]));
550 memset(task
->color_tiles
, 0, sizeof(task
->color_tiles
));
551 task
->depth_tile
= NULL
;
556 static lp_rast_cmd_func dispatch
[LP_RAST_OP_MAX
] =
559 lp_rast_clear_zstencil
,
568 lp_rast_triangle_3_4
,
569 lp_rast_triangle_3_16
,
570 lp_rast_triangle_4_16
,
572 lp_rast_shade_tile_opaque
,
580 do_rasterize_bin(struct lp_rasterizer_task
*task
,
581 const struct cmd_bin
*bin
)
583 const struct cmd_block
*block
;
589 for (block
= bin
->head
; block
; block
= block
->next
) {
590 for (k
= 0; k
< block
->count
; k
++) {
591 dispatch
[block
->cmd
[k
]]( task
, block
->arg
[k
] );
599 * Rasterize commands for a single bin.
600 * \param x, y position of the bin's tile in the framebuffer
601 * Must be called between lp_rast_begin() and lp_rast_end().
605 rasterize_bin(struct lp_rasterizer_task
*task
,
606 const struct cmd_bin
*bin
)
608 lp_rast_tile_begin( task
, bin
);
610 do_rasterize_bin(task
, bin
);
612 lp_rast_tile_end(task
);
617 if (bin
->head
->count
== 1) {
618 if (bin
->head
->cmd
[0] == LP_RAST_OP_SHADE_TILE_OPAQUE
)
619 LP_COUNT(nr_pure_shade_opaque_64
);
620 else if (bin
->head
->cmd
[0] == LP_RAST_OP_SHADE_TILE
)
621 LP_COUNT(nr_pure_shade_64
);
626 /* An empty bin is one that just loads the contents of the tile and
627 * stores them again unchanged. This typically happens when bins have
628 * been flushed for some reason in the middle of a frame, or when
629 * incremental updates are being made to a render target.
631 * Try to avoid doing pointless work in this case.
634 is_empty_bin( const struct cmd_bin
*bin
)
636 return bin
->head
== NULL
;
641 * Rasterize/execute all bins within a scene.
645 rasterize_scene(struct lp_rasterizer_task
*task
,
646 struct lp_scene
*scene
)
650 if (!task
->rast
->no_rast
&& !scene
->discard
) {
651 /* loop over scene bins, rasterize each */
655 for (i
= 0; i
< scene
->tiles_x
; i
++) {
656 for (j
= 0; j
< scene
->tiles_y
; j
++) {
657 struct cmd_bin
*bin
= lp_scene_get_bin(scene
, i
, j
);
658 rasterize_bin(task
, bin
, i
, j
);
667 while ((bin
= lp_scene_bin_iter_next(scene
))) {
668 if (!is_empty_bin( bin
))
669 rasterize_bin(task
, bin
);
677 lp_fence_signal(scene
->fence
);
685 * Called by setup module when it has something for us to render.
688 lp_rast_queue_scene( struct lp_rasterizer
*rast
,
689 struct lp_scene
*scene
)
691 LP_DBG(DEBUG_SETUP
, "%s\n", __FUNCTION__
);
693 if (rast
->num_threads
== 0) {
696 lp_rast_begin( rast
, scene
);
698 rasterize_scene( &rast
->tasks
[0], scene
);
702 rast
->curr_scene
= NULL
;
705 /* threaded rendering! */
708 lp_scene_enqueue( rast
->full_scenes
, scene
);
710 /* signal the threads that there's work to do */
711 for (i
= 0; i
< rast
->num_threads
; i
++) {
712 pipe_semaphore_signal(&rast
->tasks
[i
].work_ready
);
716 LP_DBG(DEBUG_SETUP
, "%s done \n", __FUNCTION__
);
721 lp_rast_finish( struct lp_rasterizer
*rast
)
723 if (rast
->num_threads
== 0) {
729 /* wait for work to complete */
730 for (i
= 0; i
< rast
->num_threads
; i
++) {
731 pipe_semaphore_wait(&rast
->tasks
[i
].work_done
);
738 * This is the thread's main entrypoint.
739 * It's a simple loop:
742 * 3. signal that we're done
744 static PIPE_THREAD_ROUTINE( thread_function
, init_data
)
746 struct lp_rasterizer_task
*task
= (struct lp_rasterizer_task
*) init_data
;
747 struct lp_rasterizer
*rast
= task
->rast
;
748 boolean debug
= false;
753 debug_printf("thread %d waiting for work\n", task
->thread_index
);
754 pipe_semaphore_wait(&task
->work_ready
);
759 if (task
->thread_index
== 0) {
761 * - get next scene to rasterize
762 * - map the framebuffer surfaces
765 lp_scene_dequeue( rast
->full_scenes
, TRUE
) );
768 /* Wait for all threads to get here so that threads[1+] don't
769 * get a null rast->curr_scene pointer.
771 pipe_barrier_wait( &rast
->barrier
);
775 debug_printf("thread %d doing work\n", task
->thread_index
);
777 rasterize_scene(task
,
780 /* wait for all threads to finish with this scene */
781 pipe_barrier_wait( &rast
->barrier
);
783 /* XXX: shouldn't be necessary:
785 if (task
->thread_index
== 0) {
789 /* signal done with work */
791 debug_printf("thread %d done working\n", task
->thread_index
);
793 pipe_semaphore_signal(&task
->work_done
);
801 * Initialize semaphores and spawn the threads.
804 create_rast_threads(struct lp_rasterizer
*rast
)
808 /* NOTE: if num_threads is zero, we won't use any threads */
809 for (i
= 0; i
< rast
->num_threads
; i
++) {
810 pipe_semaphore_init(&rast
->tasks
[i
].work_ready
, 0);
811 pipe_semaphore_init(&rast
->tasks
[i
].work_done
, 0);
812 rast
->threads
[i
] = pipe_thread_create(thread_function
,
813 (void *) &rast
->tasks
[i
]);
820 * Create new lp_rasterizer. If num_threads is zero, don't create any
821 * new threads, do rendering synchronously.
822 * \param num_threads number of rasterizer threads to create
824 struct lp_rasterizer
*
825 lp_rast_create( unsigned num_threads
)
827 struct lp_rasterizer
*rast
;
830 rast
= CALLOC_STRUCT(lp_rasterizer
);
835 rast
->full_scenes
= lp_scene_queue_create();
836 if (!rast
->full_scenes
) {
840 for (i
= 0; i
< Elements(rast
->tasks
); i
++) {
841 struct lp_rasterizer_task
*task
= &rast
->tasks
[i
];
843 task
->thread_index
= i
;
846 rast
->num_threads
= num_threads
;
848 rast
->no_rast
= debug_get_bool_option("LP_NO_RAST", FALSE
);
850 create_rast_threads(rast
);
852 /* for synchronizing rasterization threads */
853 pipe_barrier_init( &rast
->barrier
, rast
->num_threads
);
855 memset(lp_dummy_tile
, 0, sizeof lp_dummy_tile
);
868 void lp_rast_destroy( struct lp_rasterizer
*rast
)
872 /* Set exit_flag and signal each thread's work_ready semaphore.
873 * Each thread will be woken up, notice that the exit_flag is set and
874 * break out of its main loop. The thread will then exit.
876 rast
->exit_flag
= TRUE
;
877 for (i
= 0; i
< rast
->num_threads
; i
++) {
878 pipe_semaphore_signal(&rast
->tasks
[i
].work_ready
);
881 /* Wait for threads to terminate before cleaning up per-thread data */
882 for (i
= 0; i
< rast
->num_threads
; i
++) {
883 pipe_thread_wait(rast
->threads
[i
]);
886 /* Clean up per-thread data */
887 for (i
= 0; i
< rast
->num_threads
; i
++) {
888 pipe_semaphore_destroy(&rast
->tasks
[i
].work_ready
);
889 pipe_semaphore_destroy(&rast
->tasks
[i
].work_done
);
892 /* for synchronizing rasterization threads */
893 pipe_barrier_destroy( &rast
->barrier
);
895 lp_scene_queue_destroy(rast
->full_scenes
);
901 /** Return number of rasterization threads */
903 lp_rast_get_num_threads( struct lp_rasterizer
*rast
)
905 return rast
->num_threads
;