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
,
93 LP_DBG(DEBUG_RAST
, "%s %d,%d\n", __FUNCTION__
, x
, y
);
96 task
->x
= x
* TILE_SIZE
;
97 task
->y
= y
* TILE_SIZE
;
98 task
->width
= TILE_SIZE
+ x
* TILE_SIZE
> task
->scene
->width_aligned
?
99 task
->scene
->width_aligned
- x
* TILE_SIZE
: TILE_SIZE
;
100 task
->height
= TILE_SIZE
+ y
* TILE_SIZE
> task
->scene
->height_aligned
?
101 task
->scene
->height_aligned
- y
* TILE_SIZE
: TILE_SIZE
;
103 /* reset pointers to color and depth tile(s) */
104 memset(task
->color_tiles
, 0, sizeof(task
->color_tiles
));
105 task
->depth_tile
= NULL
;
110 * Clear the rasterizer's current color tile.
111 * This is a bin command called during bin processing.
114 lp_rast_clear_color(struct lp_rasterizer_task
*task
,
115 const union lp_rast_cmd_arg arg
)
117 const struct lp_scene
*scene
= task
->scene
;
119 if (scene
->fb
.nr_cbufs
) {
123 if (util_format_is_pure_integer(scene
->fb
.cbufs
[0]->format
)) {
125 * We expect int/uint clear values here, though some APIs
126 * might disagree (but in any case util_pack_color()
127 * couldn't handle it)...
129 LP_DBG(DEBUG_RAST
, "%s pure int 0x%x,0x%x,0x%x,0x%x\n", __FUNCTION__
,
130 arg
.clear_color
.ui
[0],
131 arg
.clear_color
.ui
[1],
132 arg
.clear_color
.ui
[2],
133 arg
.clear_color
.ui
[3]);
135 for (i
= 0; i
< scene
->fb
.nr_cbufs
; i
++) {
136 enum pipe_format format
= scene
->fb
.cbufs
[i
]->format
;
138 if (util_format_is_pure_sint(format
)) {
139 util_format_write_4i(format
, arg
.clear_color
.i
, 0, &uc
, 0, 0, 0, 1, 1);
142 assert(util_format_is_pure_uint(format
));
143 util_format_write_4ui(format
, arg
.clear_color
.ui
, 0, &uc
, 0, 0, 0, 1, 1);
146 util_fill_rect(scene
->cbufs
[i
].map
,
147 scene
->fb
.cbufs
[i
]->format
,
148 scene
->cbufs
[i
].stride
,
157 uint8_t clear_color
[4];
159 for (i
= 0; i
< 4; ++i
) {
160 clear_color
[i
] = float_to_ubyte(arg
.clear_color
.f
[i
]);
163 LP_DBG(DEBUG_RAST
, "%s 0x%x,0x%x,0x%x,0x%x\n", __FUNCTION__
,
169 for (i
= 0; i
< scene
->fb
.nr_cbufs
; i
++) {
171 util_pack_color(arg
.clear_color
.f
,
172 scene
->fb
.cbufs
[i
]->format
, &uc
);
174 util_fill_rect(scene
->cbufs
[i
].map
,
175 scene
->fb
.cbufs
[i
]->format
,
176 scene
->cbufs
[i
].stride
,
186 LP_COUNT(nr_color_tile_clear
);
193 * Clear the rasterizer's current z/stencil tile.
194 * This is a bin command called during bin processing.
197 lp_rast_clear_zstencil(struct lp_rasterizer_task
*task
,
198 const union lp_rast_cmd_arg arg
)
200 const struct lp_scene
*scene
= task
->scene
;
201 uint64_t clear_value64
= arg
.clear_zstencil
.value
;
202 uint64_t clear_mask64
= arg
.clear_zstencil
.mask
;
203 uint32_t clear_value
= (uint32_t) clear_value64
;
204 uint32_t clear_mask
= (uint32_t) clear_mask64
;
205 const unsigned height
= task
->height
;
206 const unsigned width
= task
->width
;
207 const unsigned block_size
= scene
->zsbuf
.blocksize
;
208 const unsigned dst_stride
= scene
->zsbuf
.stride
;
212 LP_DBG(DEBUG_RAST
, "%s: value=0x%08x, mask=0x%08x\n",
213 __FUNCTION__
, clear_value
, clear_mask
);
216 * Clear the area of the depth/depth buffer matching this tile.
219 if (scene
->fb
.zsbuf
) {
221 dst
= lp_rast_get_unswizzled_depth_tile_pointer(task
, LP_TEX_USAGE_READ_WRITE
);
223 clear_value
&= clear_mask
;
225 switch (block_size
) {
227 assert(clear_mask
== 0xff);
228 memset(dst
, (uint8_t) clear_value
, height
* width
);
231 if (clear_mask
== 0xffff) {
232 for (i
= 0; i
< height
; i
++) {
233 uint16_t *row
= (uint16_t *)dst
;
234 for (j
= 0; j
< width
; j
++)
235 *row
++ = (uint16_t) clear_value
;
240 for (i
= 0; i
< height
; i
++) {
241 uint16_t *row
= (uint16_t *)dst
;
242 for (j
= 0; j
< width
; j
++) {
243 uint16_t tmp
= ~clear_mask
& *row
;
244 *row
++ = clear_value
| tmp
;
251 if (clear_mask
== 0xffffffff) {
252 for (i
= 0; i
< height
; i
++) {
253 uint32_t *row
= (uint32_t *)dst
;
254 for (j
= 0; j
< width
; j
++)
255 *row
++ = clear_value
;
260 for (i
= 0; i
< height
; i
++) {
261 uint32_t *row
= (uint32_t *)dst
;
262 for (j
= 0; j
< width
; j
++) {
263 uint32_t tmp
= ~clear_mask
& *row
;
264 *row
++ = clear_value
| tmp
;
271 clear_value64
&= clear_mask64
;
272 if (clear_mask64
== 0xffffffffffULL
) {
273 for (i
= 0; i
< height
; i
++) {
274 uint64_t *row
= (uint64_t *)dst
;
275 for (j
= 0; j
< width
; j
++)
276 *row
++ = clear_value64
;
281 for (i
= 0; i
< height
; i
++) {
282 uint64_t *row
= (uint64_t *)dst
;
283 for (j
= 0; j
< width
; j
++) {
284 uint64_t tmp
= ~clear_mask64
& *row
;
285 *row
++ = clear_value64
| tmp
;
302 * Run the shader on all blocks in a tile. This is used when a tile is
303 * completely contained inside a triangle.
304 * This is a bin command called during bin processing.
307 lp_rast_shade_tile(struct lp_rasterizer_task
*task
,
308 const union lp_rast_cmd_arg arg
)
310 const struct lp_scene
*scene
= task
->scene
;
311 const struct lp_rast_shader_inputs
*inputs
= arg
.shade_tile
;
312 const struct lp_rast_state
*state
;
313 struct lp_fragment_shader_variant
*variant
;
314 const unsigned tile_x
= task
->x
, tile_y
= task
->y
;
317 if (inputs
->disable
) {
318 /* This command was partially binned and has been disabled */
322 LP_DBG(DEBUG_RAST
, "%s\n", __FUNCTION__
);
329 variant
= state
->variant
;
331 /* render the whole 64x64 tile in 4x4 chunks */
332 for (y
= 0; y
< task
->height
; y
+= 4){
333 for (x
= 0; x
< task
->width
; x
+= 4) {
334 uint8_t *color
[PIPE_MAX_COLOR_BUFS
];
335 unsigned stride
[PIPE_MAX_COLOR_BUFS
];
336 uint8_t *depth
= NULL
;
337 unsigned depth_stride
= 0;
341 for (i
= 0; i
< scene
->fb
.nr_cbufs
; i
++){
342 stride
[i
] = scene
->cbufs
[i
].stride
;
344 color
[i
] = lp_rast_get_unswizzled_color_block_pointer(task
, i
, tile_x
+ x
, tile_y
+ y
);
348 if (scene
->zsbuf
.map
) {
349 depth
= lp_rast_get_unswizzled_depth_block_pointer(task
, tile_x
+ x
, tile_y
+ y
);
350 depth_stride
= scene
->zsbuf
.stride
;
354 /* run shader on 4x4 block */
355 BEGIN_JIT_CALL(state
, task
);
356 variant
->jit_function
[RAST_WHOLE
]( &state
->jit_context
,
357 tile_x
+ x
, tile_y
+ y
,
375 * Run the shader on all blocks in a tile. This is used when a tile is
376 * completely contained inside a triangle, and the shader is opaque.
377 * This is a bin command called during bin processing.
380 lp_rast_shade_tile_opaque(struct lp_rasterizer_task
*task
,
381 const union lp_rast_cmd_arg arg
)
383 LP_DBG(DEBUG_RAST
, "%s\n", __FUNCTION__
);
390 lp_rast_shade_tile(task
, arg
);
395 * Compute shading for a 4x4 block of pixels inside a triangle.
396 * This is a bin command called during bin processing.
397 * \param x X position of quad in window coords
398 * \param y Y position of quad in window coords
401 lp_rast_shade_quads_mask(struct lp_rasterizer_task
*task
,
402 const struct lp_rast_shader_inputs
*inputs
,
403 unsigned x
, unsigned y
,
406 const struct lp_rast_state
*state
= task
->state
;
407 struct lp_fragment_shader_variant
*variant
= state
->variant
;
408 const struct lp_scene
*scene
= task
->scene
;
409 uint8_t *color
[PIPE_MAX_COLOR_BUFS
];
410 unsigned stride
[PIPE_MAX_COLOR_BUFS
];
412 unsigned depth_stride
= 0;
418 assert(x
< scene
->tiles_x
* TILE_SIZE
);
419 assert(y
< scene
->tiles_y
* TILE_SIZE
);
420 assert(x
% TILE_VECTOR_WIDTH
== 0);
421 assert(y
% TILE_VECTOR_HEIGHT
== 0);
423 assert((x
% 4) == 0);
424 assert((y
% 4) == 0);
427 for (i
= 0; i
< scene
->fb
.nr_cbufs
; i
++) {
428 stride
[i
] = scene
->cbufs
[i
].stride
;
430 color
[i
] = lp_rast_get_unswizzled_color_block_pointer(task
, i
, x
, y
);
434 if (scene
->zsbuf
.map
) {
435 depth_stride
= scene
->zsbuf
.stride
;
436 depth
= lp_rast_get_unswizzled_depth_block_pointer(task
, x
, y
);
439 assert(lp_check_alignment(state
->jit_context
.u8_blend_color
, 16));
442 * The rasterizer may produce fragments outside our
443 * allocated 4x4 blocks hence need to filter them out here.
445 if ((x
% TILE_SIZE
) < task
->width
&& (y
% TILE_SIZE
) < task
->height
) {
446 /* run shader on 4x4 block */
447 BEGIN_JIT_CALL(state
, task
);
448 variant
->jit_function
[RAST_EDGE_TEST
](&state
->jit_context
,
467 * Begin a new occlusion query.
468 * This is a bin command put in all bins.
472 lp_rast_begin_query(struct lp_rasterizer_task
*task
,
473 const union lp_rast_cmd_arg arg
)
475 struct llvmpipe_query
*pq
= arg
.query_obj
;
477 assert(task
->query
[pq
->type
] == NULL
);
480 case PIPE_QUERY_OCCLUSION_COUNTER
:
481 task
->thread_data
.vis_counter
= 0;
483 case PIPE_QUERY_PRIMITIVES_GENERATED
:
484 case PIPE_QUERY_PRIMITIVES_EMITTED
:
485 case PIPE_QUERY_SO_STATISTICS
:
486 case PIPE_QUERY_PIPELINE_STATISTICS
:
487 case PIPE_QUERY_SO_OVERFLOW_PREDICATE
:
494 task
->query
[pq
->type
] = pq
;
499 * End the current occlusion query.
500 * This is a bin command put in all bins.
504 lp_rast_end_query(struct lp_rasterizer_task
*task
,
505 const union lp_rast_cmd_arg arg
)
507 struct llvmpipe_query
*pq
= arg
.query_obj
;
508 assert(task
->query
[pq
->type
] == pq
|| pq
->type
== PIPE_QUERY_TIMESTAMP
);
511 case PIPE_QUERY_OCCLUSION_COUNTER
:
512 pq
->count
[task
->thread_index
] += task
->thread_data
.vis_counter
;
514 case PIPE_QUERY_TIMESTAMP
:
515 pq
->count
[task
->thread_index
] = os_time_get_nano();
517 case PIPE_QUERY_PRIMITIVES_GENERATED
:
518 case PIPE_QUERY_PRIMITIVES_EMITTED
:
519 case PIPE_QUERY_SO_STATISTICS
:
520 case PIPE_QUERY_PIPELINE_STATISTICS
:
521 case PIPE_QUERY_SO_OVERFLOW_PREDICATE
:
528 if (task
->query
[pq
->type
] == pq
) {
529 task
->query
[pq
->type
] = NULL
;
535 lp_rast_set_state(struct lp_rasterizer_task
*task
,
536 const union lp_rast_cmd_arg arg
)
538 task
->state
= arg
.state
;
544 * Called when we're done writing to a color tile.
547 lp_rast_tile_end(struct lp_rasterizer_task
*task
)
551 for (i
= 0; i
< PIPE_QUERY_TYPES
; ++i
) {
552 if (task
->query
[i
]) {
553 lp_rast_end_query(task
, lp_rast_arg_query(task
->query
[i
]));
558 memset(task
->color_tiles
, 0, sizeof(task
->color_tiles
));
559 task
->depth_tile
= NULL
;
564 static lp_rast_cmd_func dispatch
[LP_RAST_OP_MAX
] =
567 lp_rast_clear_zstencil
,
576 lp_rast_triangle_3_4
,
577 lp_rast_triangle_3_16
,
578 lp_rast_triangle_4_16
,
580 lp_rast_shade_tile_opaque
,
588 do_rasterize_bin(struct lp_rasterizer_task
*task
,
589 const struct cmd_bin
*bin
,
592 const struct cmd_block
*block
;
596 lp_debug_bin(bin
, x
, y
);
598 for (block
= bin
->head
; block
; block
= block
->next
) {
599 for (k
= 0; k
< block
->count
; k
++) {
600 dispatch
[block
->cmd
[k
]]( task
, block
->arg
[k
] );
608 * Rasterize commands for a single bin.
609 * \param x, y position of the bin's tile in the framebuffer
610 * Must be called between lp_rast_begin() and lp_rast_end().
614 rasterize_bin(struct lp_rasterizer_task
*task
,
615 const struct cmd_bin
*bin
, int x
, int y
)
617 lp_rast_tile_begin( task
, bin
, x
, y
);
619 do_rasterize_bin(task
, bin
, x
, y
);
621 lp_rast_tile_end(task
);
626 if (bin
->head
->count
== 1) {
627 if (bin
->head
->cmd
[0] == LP_RAST_OP_SHADE_TILE_OPAQUE
)
628 LP_COUNT(nr_pure_shade_opaque_64
);
629 else if (bin
->head
->cmd
[0] == LP_RAST_OP_SHADE_TILE
)
630 LP_COUNT(nr_pure_shade_64
);
635 /* An empty bin is one that just loads the contents of the tile and
636 * stores them again unchanged. This typically happens when bins have
637 * been flushed for some reason in the middle of a frame, or when
638 * incremental updates are being made to a render target.
640 * Try to avoid doing pointless work in this case.
643 is_empty_bin( const struct cmd_bin
*bin
)
645 return bin
->head
== NULL
;
650 * Rasterize/execute all bins within a scene.
654 rasterize_scene(struct lp_rasterizer_task
*task
,
655 struct lp_scene
*scene
)
659 if (!task
->rast
->no_rast
&& !scene
->discard
) {
660 /* loop over scene bins, rasterize each */
666 while ((bin
= lp_scene_bin_iter_next(scene
, &i
, &j
))) {
667 if (!is_empty_bin( bin
))
668 rasterize_bin(task
, bin
, i
, j
);
675 lp_fence_signal(scene
->fence
);
683 * Called by setup module when it has something for us to render.
686 lp_rast_queue_scene( struct lp_rasterizer
*rast
,
687 struct lp_scene
*scene
)
689 LP_DBG(DEBUG_SETUP
, "%s\n", __FUNCTION__
);
691 if (rast
->num_threads
== 0) {
694 lp_rast_begin( rast
, scene
);
696 rasterize_scene( &rast
->tasks
[0], scene
);
700 rast
->curr_scene
= NULL
;
703 /* threaded rendering! */
706 lp_scene_enqueue( rast
->full_scenes
, scene
);
708 /* signal the threads that there's work to do */
709 for (i
= 0; i
< rast
->num_threads
; i
++) {
710 pipe_semaphore_signal(&rast
->tasks
[i
].work_ready
);
714 LP_DBG(DEBUG_SETUP
, "%s done \n", __FUNCTION__
);
719 lp_rast_finish( struct lp_rasterizer
*rast
)
721 if (rast
->num_threads
== 0) {
727 /* wait for work to complete */
728 for (i
= 0; i
< rast
->num_threads
; i
++) {
729 pipe_semaphore_wait(&rast
->tasks
[i
].work_done
);
736 * This is the thread's main entrypoint.
737 * It's a simple loop:
740 * 3. signal that we're done
742 static PIPE_THREAD_ROUTINE( thread_function
, init_data
)
744 struct lp_rasterizer_task
*task
= (struct lp_rasterizer_task
*) init_data
;
745 struct lp_rasterizer
*rast
= task
->rast
;
746 boolean debug
= false;
751 debug_printf("thread %d waiting for work\n", task
->thread_index
);
752 pipe_semaphore_wait(&task
->work_ready
);
757 if (task
->thread_index
== 0) {
759 * - get next scene to rasterize
760 * - map the framebuffer surfaces
763 lp_scene_dequeue( rast
->full_scenes
, TRUE
) );
766 /* Wait for all threads to get here so that threads[1+] don't
767 * get a null rast->curr_scene pointer.
769 pipe_barrier_wait( &rast
->barrier
);
773 debug_printf("thread %d doing work\n", task
->thread_index
);
775 rasterize_scene(task
,
778 /* wait for all threads to finish with this scene */
779 pipe_barrier_wait( &rast
->barrier
);
781 /* XXX: shouldn't be necessary:
783 if (task
->thread_index
== 0) {
787 /* signal done with work */
789 debug_printf("thread %d done working\n", task
->thread_index
);
791 pipe_semaphore_signal(&task
->work_done
);
799 * Initialize semaphores and spawn the threads.
802 create_rast_threads(struct lp_rasterizer
*rast
)
806 /* NOTE: if num_threads is zero, we won't use any threads */
807 for (i
= 0; i
< rast
->num_threads
; i
++) {
808 pipe_semaphore_init(&rast
->tasks
[i
].work_ready
, 0);
809 pipe_semaphore_init(&rast
->tasks
[i
].work_done
, 0);
810 rast
->threads
[i
] = pipe_thread_create(thread_function
,
811 (void *) &rast
->tasks
[i
]);
818 * Create new lp_rasterizer. If num_threads is zero, don't create any
819 * new threads, do rendering synchronously.
820 * \param num_threads number of rasterizer threads to create
822 struct lp_rasterizer
*
823 lp_rast_create( unsigned num_threads
)
825 struct lp_rasterizer
*rast
;
828 rast
= CALLOC_STRUCT(lp_rasterizer
);
833 rast
->full_scenes
= lp_scene_queue_create();
834 if (!rast
->full_scenes
) {
838 for (i
= 0; i
< Elements(rast
->tasks
); i
++) {
839 struct lp_rasterizer_task
*task
= &rast
->tasks
[i
];
841 task
->thread_index
= i
;
844 rast
->num_threads
= num_threads
;
846 rast
->no_rast
= debug_get_bool_option("LP_NO_RAST", FALSE
);
848 create_rast_threads(rast
);
850 /* for synchronizing rasterization threads */
851 pipe_barrier_init( &rast
->barrier
, rast
->num_threads
);
853 memset(lp_dummy_tile
, 0, sizeof lp_dummy_tile
);
866 void lp_rast_destroy( struct lp_rasterizer
*rast
)
870 /* Set exit_flag and signal each thread's work_ready semaphore.
871 * Each thread will be woken up, notice that the exit_flag is set and
872 * break out of its main loop. The thread will then exit.
874 rast
->exit_flag
= TRUE
;
875 for (i
= 0; i
< rast
->num_threads
; i
++) {
876 pipe_semaphore_signal(&rast
->tasks
[i
].work_ready
);
879 /* Wait for threads to terminate before cleaning up per-thread data */
880 for (i
= 0; i
< rast
->num_threads
; i
++) {
881 pipe_thread_wait(rast
->threads
[i
]);
884 /* Clean up per-thread data */
885 for (i
= 0; i
< rast
->num_threads
; i
++) {
886 pipe_semaphore_destroy(&rast
->tasks
[i
].work_ready
);
887 pipe_semaphore_destroy(&rast
->tasks
[i
].work_done
);
890 /* for synchronizing rasterization threads */
891 pipe_barrier_destroy( &rast
->barrier
);
893 lp_scene_queue_destroy(rast
->full_scenes
);
899 /** Return number of rasterization threads */
901 lp_rast_get_num_threads( struct lp_rasterizer
*rast
)
903 return rast
->num_threads
;