1 /**************************************************************************
2 *
3 * Copyright 2009 VMware, Inc.
4 * All Rights Reserved.
5 *
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:
13 *
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
16 * of the Software.
17 *
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.
25 *
26 **************************************************************************/
27
28 #include <limits.h>
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"
34 #include "util/u_string.h"
35 #include "util/u_thread.h"
36 #include "util/u_memset.h"
37 #include "util/os_time.h"
38
39 #include "lp_scene_queue.h"
40 #include "lp_context.h"
41 #include "lp_debug.h"
42 #include "lp_fence.h"
43 #include "lp_perf.h"
44 #include "lp_query.h"
45 #include "lp_rast.h"
46 #include "lp_rast_priv.h"
47 #include "gallivm/lp_bld_format.h"
48 #include "gallivm/lp_bld_debug.h"
49 #include "lp_scene.h"
50 #include "lp_tex_sample.h"
51
52
53 #ifdef DEBUG
54 int jit_line = 0;
55 const struct lp_rast_state *jit_state = NULL;
56 const struct lp_rasterizer_task *jit_task = NULL;
57 #endif
58
59 const float lp_sample_pos_4x[4][2] = { { 0.375, 0.125 },
60 { 0.875, 0.375 },
61 { 0.125, 0.625 },
62 { 0.625, 0.875 } };
63
64 /**
65 * Begin rasterizing a scene.
66 * Called once per scene by one thread.
67 */
68 static void
lp_rast_begin(struct lp_rasterizer * rast,struct lp_scene * scene)69 lp_rast_begin( struct lp_rasterizer *rast,
70 struct lp_scene *scene )
71 {
72 rast->curr_scene = scene;
73
74 LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
75
76 lp_scene_begin_rasterization( scene );
77 lp_scene_bin_iter_begin( scene );
78 }
79
80
81 static void
lp_rast_end(struct lp_rasterizer * rast)82 lp_rast_end( struct lp_rasterizer *rast )
83 {
84 lp_scene_end_rasterization( rast->curr_scene );
85
86 rast->curr_scene = NULL;
87 }
88
89
90 /**
91 * Beginning rasterization of a tile.
92 * \param x window X position of the tile, in pixels
93 * \param y window Y position of the tile, in pixels
94 */
95 static void
lp_rast_tile_begin(struct lp_rasterizer_task * task,const struct cmd_bin * bin,int x,int y)96 lp_rast_tile_begin(struct lp_rasterizer_task *task,
97 const struct cmd_bin *bin,
98 int x, int y)
99 {
100 unsigned i;
101 struct lp_scene *scene = task->scene;
102
103 LP_DBG(DEBUG_RAST, "%s %d,%d\n", __FUNCTION__, x, y);
104
105 task->bin = bin;
106 task->x = x * TILE_SIZE;
107 task->y = y * TILE_SIZE;
108 task->width = TILE_SIZE + x * TILE_SIZE > task->scene->fb.width ?
109 task->scene->fb.width - x * TILE_SIZE : TILE_SIZE;
110 task->height = TILE_SIZE + y * TILE_SIZE > task->scene->fb.height ?
111 task->scene->fb.height - y * TILE_SIZE : TILE_SIZE;
112
113 task->thread_data.vis_counter = 0;
114 task->thread_data.ps_invocations = 0;
115
116 for (i = 0; i < task->scene->fb.nr_cbufs; i++) {
117 if (task->scene->fb.cbufs[i]) {
118 task->color_tiles[i] = scene->cbufs[i].map +
119 scene->cbufs[i].stride * task->y +
120 scene->cbufs[i].format_bytes * task->x;
121 }
122 }
123 if (task->scene->fb.zsbuf) {
124 task->depth_tile = scene->zsbuf.map +
125 scene->zsbuf.stride * task->y +
126 scene->zsbuf.format_bytes * task->x;
127 }
128 }
129
130
131 /**
132 * Clear the rasterizer's current color tile.
133 * This is a bin command called during bin processing.
134 * Clear commands always clear all bound layers.
135 */
136 static void
lp_rast_clear_color(struct lp_rasterizer_task * task,const union lp_rast_cmd_arg arg)137 lp_rast_clear_color(struct lp_rasterizer_task *task,
138 const union lp_rast_cmd_arg arg)
139 {
140 const struct lp_scene *scene = task->scene;
141 unsigned cbuf = arg.clear_rb->cbuf;
142 union util_color uc;
143 enum pipe_format format;
144
145 /* we never bin clear commands for non-existing buffers */
146 assert(cbuf < scene->fb.nr_cbufs);
147 assert(scene->fb.cbufs[cbuf]);
148
149 format = scene->fb.cbufs[cbuf]->format;
150 uc = arg.clear_rb->color_val;
151
152 /*
153 * this is pretty rough since we have target format (bunch of bytes...) here.
154 * dump it as raw 4 dwords.
155 */
156 LP_DBG(DEBUG_RAST, "%s clear value (target format %d) raw 0x%x,0x%x,0x%x,0x%x\n",
157 __FUNCTION__, format, uc.ui[0], uc.ui[1], uc.ui[2], uc.ui[3]);
158
159 for (unsigned s = 0; s < scene->cbufs[cbuf].nr_samples; s++) {
160 void *map = (char *)scene->cbufs[cbuf].map + scene->cbufs[cbuf].sample_stride * s;
161 util_fill_box(map,
162 format,
163 scene->cbufs[cbuf].stride,
164 scene->cbufs[cbuf].layer_stride,
165 task->x,
166 task->y,
167 0,
168 task->width,
169 task->height,
170 scene->fb_max_layer + 1,
171 &uc);
172 }
173
174 /* this will increase for each rb which probably doesn't mean much */
175 LP_COUNT(nr_color_tile_clear);
176 }
177
178
179 /**
180 * Clear the rasterizer's current z/stencil tile.
181 * This is a bin command called during bin processing.
182 * Clear commands always clear all bound layers.
183 */
184 static void
lp_rast_clear_zstencil(struct lp_rasterizer_task * task,const union lp_rast_cmd_arg arg)185 lp_rast_clear_zstencil(struct lp_rasterizer_task *task,
186 const union lp_rast_cmd_arg arg)
187 {
188 const struct lp_scene *scene = task->scene;
189 uint64_t clear_value64 = arg.clear_zstencil.value;
190 uint64_t clear_mask64 = arg.clear_zstencil.mask;
191 uint32_t clear_value = (uint32_t) clear_value64;
192 uint32_t clear_mask = (uint32_t) clear_mask64;
193 const unsigned height = task->height;
194 const unsigned width = task->width;
195 const unsigned dst_stride = scene->zsbuf.stride;
196 uint8_t *dst;
197 unsigned i, j;
198 unsigned block_size;
199
200 LP_DBG(DEBUG_RAST, "%s: value=0x%08x, mask=0x%08x\n",
201 __FUNCTION__, clear_value, clear_mask);
202
203 /*
204 * Clear the area of the depth/depth buffer matching this tile.
205 */
206
207 if (scene->fb.zsbuf) {
208 unsigned layer;
209
210 for (unsigned s = 0; s < scene->zsbuf.nr_samples; s++) {
211 uint8_t *dst_layer = task->depth_tile + (s * scene->zsbuf.sample_stride);
212 block_size = util_format_get_blocksize(scene->fb.zsbuf->format);
213
214 clear_value &= clear_mask;
215
216 for (layer = 0; layer <= scene->fb_max_layer; layer++) {
217 dst = dst_layer;
218
219 switch (block_size) {
220 case 1:
221 assert(clear_mask == 0xff);
222 for (i = 0; i < height; i++) {
223 uint8_t *row = (uint8_t *)dst;
224 memset(row, (uint8_t) clear_value, width);
225 dst += dst_stride;
226 }
227 break;
228 case 2:
229 if (clear_mask == 0xffff) {
230 for (i = 0; i < height; i++) {
231 uint16_t *row = (uint16_t *)dst;
232 for (j = 0; j < width; j++)
233 *row++ = (uint16_t) clear_value;
234 dst += dst_stride;
235 }
236 }
237 else {
238 for (i = 0; i < height; i++) {
239 uint16_t *row = (uint16_t *)dst;
240 for (j = 0; j < width; j++) {
241 uint16_t tmp = ~clear_mask & *row;
242 *row++ = clear_value | tmp;
243 }
244 dst += dst_stride;
245 }
246 }
247 break;
248 case 4:
249 if (clear_mask == 0xffffffff) {
250 for (i = 0; i < height; i++) {
251 util_memset32(dst, clear_value, width);
252 dst += dst_stride;
253 }
254 }
255 else {
256 for (i = 0; i < height; i++) {
257 uint32_t *row = (uint32_t *)dst;
258 for (j = 0; j < width; j++) {
259 uint32_t tmp = ~clear_mask & *row;
260 *row++ = clear_value | tmp;
261 }
262 dst += dst_stride;
263 }
264 }
265 break;
266 case 8:
267 clear_value64 &= clear_mask64;
268 if (clear_mask64 == 0xffffffffffULL) {
269 for (i = 0; i < height; i++) {
270 util_memset64(dst, clear_value64, width);
271 dst += dst_stride;
272 }
273 }
274 else {
275 for (i = 0; i < height; i++) {
276 uint64_t *row = (uint64_t *)dst;
277 for (j = 0; j < width; j++) {
278 uint64_t tmp = ~clear_mask64 & *row;
279 *row++ = clear_value64 | tmp;
280 }
281 dst += dst_stride;
282 }
283 }
284 break;
285
286 default:
287 assert(0);
288 break;
289 }
290 dst_layer += scene->zsbuf.layer_stride;
291 }
292 }
293 }
294 }
295
296
297
298 /**
299 * Run the shader on all blocks in a tile. This is used when a tile is
300 * completely contained inside a triangle.
301 * This is a bin command called during bin processing.
302 */
303 static void
lp_rast_shade_tile(struct lp_rasterizer_task * task,const union lp_rast_cmd_arg arg)304 lp_rast_shade_tile(struct lp_rasterizer_task *task,
305 const union lp_rast_cmd_arg arg)
306 {
307 const struct lp_scene *scene = task->scene;
308 const struct lp_rast_shader_inputs *inputs = arg.shade_tile;
309 const struct lp_rast_state *state;
310 struct lp_fragment_shader_variant *variant;
311 const unsigned tile_x = task->x, tile_y = task->y;
312 unsigned x, y;
313
314 if (inputs->disable) {
315 /* This command was partially binned and has been disabled */
316 return;
317 }
318
319 LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
320
321 state = task->state;
322 assert(state);
323 if (!state) {
324 return;
325 }
326 variant = state->variant;
327
328 /* render the whole 64x64 tile in 4x4 chunks */
329 for (y = 0; y < task->height; y += 4){
330 for (x = 0; x < task->width; x += 4) {
331 uint8_t *color[PIPE_MAX_COLOR_BUFS];
332 unsigned stride[PIPE_MAX_COLOR_BUFS];
333 unsigned sample_stride[PIPE_MAX_COLOR_BUFS];
334 uint8_t *depth = NULL;
335 unsigned depth_stride = 0;
336 unsigned depth_sample_stride = 0;
337 unsigned i;
338
339 /* color buffer */
340 for (i = 0; i < scene->fb.nr_cbufs; i++){
341 if (scene->fb.cbufs[i]) {
342 stride[i] = scene->cbufs[i].stride;
343 sample_stride[i] = scene->cbufs[i].sample_stride;
344 color[i] = lp_rast_get_color_block_pointer(task, i, tile_x + x,
345 tile_y + y, inputs->layer + inputs->view_index);
346 }
347 else {
348 stride[i] = 0;
349 sample_stride[i] = 0;
350 color[i] = NULL;
351 }
352 }
353
354 /* depth buffer */
355 if (scene->zsbuf.map) {
356 depth = lp_rast_get_depth_block_pointer(task, tile_x + x,
357 tile_y + y, inputs->layer + inputs->view_index);
358 depth_stride = scene->zsbuf.stride;
359 depth_sample_stride = scene->zsbuf.sample_stride;
360 }
361
362 uint64_t mask = 0;
363 for (unsigned i = 0; i < scene->fb_max_samples; i++)
364 mask |= (uint64_t)(0xffff) << (16 * i);
365
366 /* Propagate non-interpolated raster state. */
367 task->thread_data.raster_state.viewport_index = inputs->viewport_index;
368 task->thread_data.raster_state.view_index = inputs->view_index;
369
370 /* run shader on 4x4 block */
371 BEGIN_JIT_CALL(state, task);
372 variant->jit_function[RAST_WHOLE]( &state->jit_context,
373 tile_x + x, tile_y + y,
374 inputs->frontfacing,
375 GET_A0(inputs),
376 GET_DADX(inputs),
377 GET_DADY(inputs),
378 color,
379 depth,
380 mask,
381 &task->thread_data,
382 stride,
383 depth_stride,
384 sample_stride,
385 depth_sample_stride);
386 END_JIT_CALL();
387 }
388 }
389 }
390
391
392 /**
393 * Run the shader on all blocks in a tile. This is used when a tile is
394 * completely contained inside a triangle, and the shader is opaque.
395 * This is a bin command called during bin processing.
396 */
397 static void
lp_rast_shade_tile_opaque(struct lp_rasterizer_task * task,const union lp_rast_cmd_arg arg)398 lp_rast_shade_tile_opaque(struct lp_rasterizer_task *task,
399 const union lp_rast_cmd_arg arg)
400 {
401 LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
402
403 assert(task->state);
404 if (!task->state) {
405 return;
406 }
407
408 lp_rast_shade_tile(task, arg);
409 }
410
411
412 /**
413 * Compute shading for a 4x4 block of pixels inside a triangle.
414 * This is a bin command called during bin processing.
415 * \param x X position of quad in window coords
416 * \param y Y position of quad in window coords
417 */
418 void
lp_rast_shade_quads_mask_sample(struct lp_rasterizer_task * task,const struct lp_rast_shader_inputs * inputs,unsigned x,unsigned y,uint64_t mask)419 lp_rast_shade_quads_mask_sample(struct lp_rasterizer_task *task,
420 const struct lp_rast_shader_inputs *inputs,
421 unsigned x, unsigned y,
422 uint64_t mask)
423 {
424 const struct lp_rast_state *state = task->state;
425 struct lp_fragment_shader_variant *variant = state->variant;
426 const struct lp_scene *scene = task->scene;
427 uint8_t *color[PIPE_MAX_COLOR_BUFS];
428 unsigned stride[PIPE_MAX_COLOR_BUFS];
429 unsigned sample_stride[PIPE_MAX_COLOR_BUFS];
430 uint8_t *depth = NULL;
431 unsigned depth_stride = 0;
432 unsigned depth_sample_stride = 0;
433 unsigned i;
434
435 assert(state);
436
437 /* Sanity checks */
438 assert(x < scene->tiles_x * TILE_SIZE);
439 assert(y < scene->tiles_y * TILE_SIZE);
440 assert(x % TILE_VECTOR_WIDTH == 0);
441 assert(y % TILE_VECTOR_HEIGHT == 0);
442
443 assert((x % 4) == 0);
444 assert((y % 4) == 0);
445
446 /* color buffer */
447 for (i = 0; i < scene->fb.nr_cbufs; i++) {
448 if (scene->fb.cbufs[i]) {
449 stride[i] = scene->cbufs[i].stride;
450 sample_stride[i] = scene->cbufs[i].sample_stride;
451 color[i] = lp_rast_get_color_block_pointer(task, i, x, y,
452 inputs->layer + inputs->view_index);
453 }
454 else {
455 stride[i] = 0;
456 sample_stride[i] = 0;
457 color[i] = NULL;
458 }
459 }
460
461 /* depth buffer */
462 if (scene->zsbuf.map) {
463 depth_stride = scene->zsbuf.stride;
464 depth_sample_stride = scene->zsbuf.sample_stride;
465 depth = lp_rast_get_depth_block_pointer(task, x, y, inputs->layer + inputs->view_index);
466 }
467
468 assert(lp_check_alignment(state->jit_context.u8_blend_color, 16));
469
470 /*
471 * The rasterizer may produce fragments outside our
472 * allocated 4x4 blocks hence need to filter them out here.
473 */
474 if ((x % TILE_SIZE) < task->width && (y % TILE_SIZE) < task->height) {
475 /* Propagate non-interpolated raster state. */
476 task->thread_data.raster_state.viewport_index = inputs->viewport_index;
477 task->thread_data.raster_state.view_index = inputs->view_index;
478
479 /* run shader on 4x4 block */
480 BEGIN_JIT_CALL(state, task);
481 variant->jit_function[RAST_EDGE_TEST](&state->jit_context,
482 x, y,
483 inputs->frontfacing,
484 GET_A0(inputs),
485 GET_DADX(inputs),
486 GET_DADY(inputs),
487 color,
488 depth,
489 mask,
490 &task->thread_data,
491 stride,
492 depth_stride,
493 sample_stride,
494 depth_sample_stride);
495 END_JIT_CALL();
496 }
497 }
498
499 void
lp_rast_shade_quads_mask(struct lp_rasterizer_task * task,const struct lp_rast_shader_inputs * inputs,unsigned x,unsigned y,unsigned mask)500 lp_rast_shade_quads_mask(struct lp_rasterizer_task *task,
501 const struct lp_rast_shader_inputs *inputs,
502 unsigned x, unsigned y,
503 unsigned mask)
504 {
505 uint64_t new_mask = 0;
506 for (unsigned i = 0; i < task->scene->fb_max_samples; i++)
507 new_mask |= ((uint64_t)mask) << (16 * i);
508 lp_rast_shade_quads_mask_sample(task, inputs, x, y, new_mask);
509 }
510
511 /**
512 * Directly copy pixels from a texture to the destination color buffer.
513 * This is a bin command called during bin processing.
514 */
515 static void
lp_rast_blit_tile_to_dest(struct lp_rasterizer_task * task,const union lp_rast_cmd_arg arg)516 lp_rast_blit_tile_to_dest(struct lp_rasterizer_task *task,
517 const union lp_rast_cmd_arg arg)
518 {
519 const struct lp_scene *scene = task->scene;
520 const struct lp_rast_shader_inputs *inputs = arg.shade_tile;
521 const struct lp_rast_state *state = task->state;
522 struct lp_fragment_shader_variant *variant = state->variant;
523 const struct lp_jit_texture *texture = &state->jit_context.textures[0];
524 const uint8_t *src;
525 uint8_t *dst;
526 unsigned src_stride;
527 unsigned dst_stride;
528 struct pipe_surface *cbuf = scene->fb.cbufs[0];
529 const unsigned face_slice = cbuf->u.tex.first_layer;
530 const unsigned level = cbuf->u.tex.level;
531 struct llvmpipe_resource *lpt = llvmpipe_resource(cbuf->texture);
532 int src_x, src_y;
533
534 LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
535
536 if (inputs->disable) {
537 /* This command was partially binned and has been disabled */
538 return;
539 }
540
541 dst = llvmpipe_get_texture_image_address(lpt, face_slice, level);
542
543 if (!dst)
544 return;
545
546 dst_stride = lpt->row_stride[level];
547
548 src = texture->base;
549 src_stride = texture->row_stride[0];
550
551 src_x = util_iround(GET_A0(inputs)[1][0]*texture->width - 0.5f);
552 src_y = util_iround(GET_A0(inputs)[1][1]*texture->height - 0.5f);
553
554 src_x = src_x + task->x;
555 src_y = src_y + task->y;
556
557 if (0) {
558 union util_color uc;
559 uc.ui[0] = 0xff0000ff;
560 util_fill_rect(dst,
561 cbuf->format,
562 dst_stride,
563 task->x,
564 task->y,
565 task->width,
566 task->height,
567 &uc);
568 return;
569 }
570
571 if (src_x >= 0 &&
572 src_y >= 0 &&
573 src_x + task->width <= texture->width &&
574 src_y + task->height <= texture->height) {
575
576 if (variant->shader->kind == LP_FS_KIND_BLIT_RGBA ||
577 (variant->shader->kind == LP_FS_KIND_BLIT_RGB1 &&
578 cbuf->format == PIPE_FORMAT_B8G8R8X8_UNORM)) {
579 util_copy_rect(dst,
580 cbuf->format,
581 dst_stride,
582 task->x, task->y,
583 task->width, task->height,
584 src, src_stride,
585 src_x, src_y);
586 return;
587 }
588
589 if (variant->shader->kind == LP_FS_KIND_BLIT_RGB1) {
590 if (cbuf->format == PIPE_FORMAT_B8G8R8A8_UNORM) {
591 int x, y;
592
593 dst += task->x * 4;
594 src += src_x * 4;
595 dst += task->y * dst_stride;
596 src += src_y * src_stride;
597
598 for (y = 0; y < task->height; ++y) {
599 const uint32_t *src_row = (const uint32_t *)src;
600 uint32_t *dst_row = (uint32_t *)dst;
601
602 for (x = 0; x < task->width; ++x) {
603 *dst_row++ = *src_row++ | 0xff000000;
604 }
605 dst += dst_stride;
606 src += src_stride;
607 }
608
609 return;
610 }
611 }
612
613 }
614
615 /*
616 * Fall back to the jit shaders.
617 */
618
619 lp_rast_shade_tile_opaque(task, arg);
620 }
621
622 static void
lp_rast_blit_tile(struct lp_rasterizer_task * task,const union lp_rast_cmd_arg arg)623 lp_rast_blit_tile(struct lp_rasterizer_task *task,
624 const union lp_rast_cmd_arg arg)
625 {
626 /* This kindof just works, but isn't efficient:
627 */
628 lp_rast_blit_tile_to_dest(task, arg);
629 }
630
631 /**
632 * Begin a new occlusion query.
633 * This is a bin command put in all bins.
634 * Called per thread.
635 */
636 static void
lp_rast_begin_query(struct lp_rasterizer_task * task,const union lp_rast_cmd_arg arg)637 lp_rast_begin_query(struct lp_rasterizer_task *task,
638 const union lp_rast_cmd_arg arg)
639 {
640 struct llvmpipe_query *pq = arg.query_obj;
641
642 switch (pq->type) {
643 case PIPE_QUERY_OCCLUSION_COUNTER:
644 case PIPE_QUERY_OCCLUSION_PREDICATE:
645 case PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE:
646 pq->start[task->thread_index] = task->thread_data.vis_counter;
647 break;
648 case PIPE_QUERY_PIPELINE_STATISTICS:
649 pq->start[task->thread_index] = task->thread_data.ps_invocations;
650 break;
651 case PIPE_QUERY_TIME_ELAPSED:
652 pq->start[task->thread_index] = os_time_get_nano();
653 break;
654 default:
655 assert(0);
656 break;
657 }
658 }
659
660
661 /**
662 * End the current occlusion query.
663 * This is a bin command put in all bins.
664 * Called per thread.
665 */
666 static void
lp_rast_end_query(struct lp_rasterizer_task * task,const union lp_rast_cmd_arg arg)667 lp_rast_end_query(struct lp_rasterizer_task *task,
668 const union lp_rast_cmd_arg arg)
669 {
670 struct llvmpipe_query *pq = arg.query_obj;
671
672 switch (pq->type) {
673 case PIPE_QUERY_OCCLUSION_COUNTER:
674 case PIPE_QUERY_OCCLUSION_PREDICATE:
675 case PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE:
676 pq->end[task->thread_index] +=
677 task->thread_data.vis_counter - pq->start[task->thread_index];
678 pq->start[task->thread_index] = 0;
679 break;
680 case PIPE_QUERY_TIMESTAMP:
681 case PIPE_QUERY_TIME_ELAPSED:
682 pq->end[task->thread_index] = os_time_get_nano();
683 break;
684 case PIPE_QUERY_PIPELINE_STATISTICS:
685 pq->end[task->thread_index] +=
686 task->thread_data.ps_invocations - pq->start[task->thread_index];
687 pq->start[task->thread_index] = 0;
688 break;
689 default:
690 assert(0);
691 break;
692 }
693 }
694
695
696 void
lp_rast_set_state(struct lp_rasterizer_task * task,const union lp_rast_cmd_arg arg)697 lp_rast_set_state(struct lp_rasterizer_task *task,
698 const union lp_rast_cmd_arg arg)
699 {
700 task->state = arg.state;
701 }
702
703
704
705 /**
706 * Called when we're done writing to a color tile.
707 */
708 static void
lp_rast_tile_end(struct lp_rasterizer_task * task)709 lp_rast_tile_end(struct lp_rasterizer_task *task)
710 {
711 unsigned i;
712
713 for (i = 0; i < task->scene->num_active_queries; ++i) {
714 lp_rast_end_query(task, lp_rast_arg_query(task->scene->active_queries[i]));
715 }
716
717 /* debug */
718 memset(task->color_tiles, 0, sizeof(task->color_tiles));
719 task->depth_tile = NULL;
720
721 task->bin = NULL;
722 }
723
724
725
726
727
728
729 /* Currently have two rendering paths only - the general case triangle
730 * path and the super-specialized blit/clear path.
731 */
732 #define TRI ((LP_RAST_FLAGS_TRI <<1)-1) /* general case */
733 #define RECT ((LP_RAST_FLAGS_RECT<<1)-1) /* direct rectangle rasterizer */
734 #define BLIT ((LP_RAST_FLAGS_BLIT<<1)-1) /* write direct-to-dest */
735
736 static const unsigned
737 rast_flags[] = {
738 BLIT, /* clear color */
739 TRI, /* clear zstencil */
740 TRI, /* triangle_1 */
741 TRI, /* triangle_2 */
742 TRI, /* triangle_3 */
743 TRI, /* triangle_4 */
744 TRI, /* triangle_5 */
745 TRI, /* triangle_6 */
746 TRI, /* triangle_7 */
747 TRI, /* triangle_8 */
748 TRI, /* triangle_3_4 */
749 TRI, /* triangle_3_16 */
750 TRI, /* triangle_4_16 */
751 RECT, /* shade_tile */
752 RECT, /* shade_tile_opaque */
753 TRI, /* begin_query */
754 TRI, /* end_query */
755 BLIT, /* set_state, */
756 TRI, /* lp_rast_triangle_32_1 */
757 TRI, /* lp_rast_triangle_32_2 */
758 TRI, /* lp_rast_triangle_32_3 */
759 TRI, /* lp_rast_triangle_32_4 */
760 TRI, /* lp_rast_triangle_32_5 */
761 TRI, /* lp_rast_triangle_32_6 */
762 TRI, /* lp_rast_triangle_32_7 */
763 TRI, /* lp_rast_triangle_32_8 */
764 TRI, /* lp_rast_triangle_32_3_4 */
765 TRI, /* lp_rast_triangle_32_3_16 */
766 TRI, /* lp_rast_triangle_32_4_16 */
767 TRI, /* lp_rast_triangle_ms_1 */
768 TRI, /* lp_rast_triangle_ms_2 */
769 TRI, /* lp_rast_triangle_ms_3 */
770 TRI, /* lp_rast_triangle_ms_4 */
771 TRI, /* lp_rast_triangle_ms_5 */
772 TRI, /* lp_rast_triangle_ms_6 */
773 TRI, /* lp_rast_triangle_ms_7 */
774 TRI, /* lp_rast_triangle_ms_8 */
775 TRI, /* lp_rast_triangle_ms_3_4 */
776 TRI, /* lp_rast_triangle_ms_3_16 */
777 TRI, /* lp_rast_triangle_ms_4_16 */
778
779 RECT, /* rectangle */
780 BLIT, /* blit */
781 };
782
783 /*
784 */
785 static const lp_rast_cmd_func
786 dispatch_blit[] = {
787 lp_rast_clear_color,
788 NULL, /* clear_zstencil */
789 NULL, /* triangle_1 */
790 NULL, /* triangle_2 */
791 NULL, /* triangle_3 */
792 NULL, /* triangle_4 */
793 NULL, /* triangle_5 */
794 NULL, /* triangle_6 */
795 NULL, /* triangle_7 */
796 NULL, /* triangle_8 */
797 NULL, /* triangle_3_4 */
798 NULL, /* triangle_3_16 */
799 NULL, /* triangle_4_16 */
800 NULL, /* shade_tile */
801 NULL, /* shade_tile_opaque */
802 NULL, /* begin_query */
803 NULL, /* end_query */
804 lp_rast_set_state, /* set_state */
805 NULL, /* lp_rast_triangle_32_1 */
806 NULL, /* lp_rast_triangle_32_2 */
807 NULL, /* lp_rast_triangle_32_3 */
808 NULL, /* lp_rast_triangle_32_4 */
809 NULL, /* lp_rast_triangle_32_5 */
810 NULL, /* lp_rast_triangle_32_6 */
811 NULL, /* lp_rast_triangle_32_7 */
812 NULL, /* lp_rast_triangle_32_8 */
813 NULL, /* lp_rast_triangle_32_3_4 */
814 NULL, /* lp_rast_triangle_32_3_16 */
815 NULL, /* lp_rast_triangle_32_4_16 */
816 NULL, /* lp_rast_triangle_ms_1 */
817 NULL, /* lp_rast_triangle_ms_2 */
818 NULL, /* lp_rast_triangle_ms_3 */
819 NULL, /* lp_rast_triangle_ms_4 */
820 NULL, /* lp_rast_triangle_ms_5 */
821 NULL, /* lp_rast_triangle_ms_6 */
822 NULL, /* lp_rast_triangle_ms_7 */
823 NULL, /* lp_rast_triangle_ms_8 */
824 NULL, /* lp_rast_triangle_ms_3_4 */
825 NULL, /* lp_rast_triangle_ms_3_16 */
826 NULL, /* lp_rast_triangle_ms_4_16 */
827
828 NULL, /* rectangle */
829 lp_rast_blit_tile_to_dest,
830 };
831
832
833
834 /* Triangle and general case rasterization: Use the SOA llvm shdaers,
835 * an active swizzled tile for each color buf, etc. Don't blit/clear
836 * directly to destination surface as we know there are swizzled
837 * operations coming.
838 */
839 static const lp_rast_cmd_func
840 dispatch_tri[] = {
841 lp_rast_clear_color,
842 lp_rast_clear_zstencil,
843 lp_rast_triangle_1,
844 lp_rast_triangle_2,
845 lp_rast_triangle_3,
846 lp_rast_triangle_4,
847 lp_rast_triangle_5,
848 lp_rast_triangle_6,
849 lp_rast_triangle_7,
850 lp_rast_triangle_8,
851 lp_rast_triangle_3_4,
852 lp_rast_triangle_3_16,
853 lp_rast_triangle_4_16,
854 lp_rast_shade_tile,
855 lp_rast_shade_tile_opaque,
856 lp_rast_begin_query,
857 lp_rast_end_query,
858 lp_rast_set_state,
859 lp_rast_triangle_32_1,
860 lp_rast_triangle_32_2,
861 lp_rast_triangle_32_3,
862 lp_rast_triangle_32_4,
863 lp_rast_triangle_32_5,
864 lp_rast_triangle_32_6,
865 lp_rast_triangle_32_7,
866 lp_rast_triangle_32_8,
867 lp_rast_triangle_32_3_4,
868 lp_rast_triangle_32_3_16,
869 lp_rast_triangle_32_4_16,
870 lp_rast_triangle_ms_1,
871 lp_rast_triangle_ms_2,
872 lp_rast_triangle_ms_3,
873 lp_rast_triangle_ms_4,
874 lp_rast_triangle_ms_5,
875 lp_rast_triangle_ms_6,
876 lp_rast_triangle_ms_7,
877 lp_rast_triangle_ms_8,
878 lp_rast_triangle_ms_3_4,
879 lp_rast_triangle_ms_3_16,
880 lp_rast_triangle_ms_4_16,
881 lp_rast_rectangle,
882 lp_rast_blit_tile,
883 };
884
885
886 /* Debug rasterization with most fastpaths disabled.
887 */
888 static const lp_rast_cmd_func
889 dispatch_tri_debug[] =
890 {
891 lp_rast_clear_color,
892 lp_rast_clear_zstencil,
893 lp_rast_triangle_1,
894 lp_rast_triangle_2,
895 lp_rast_triangle_3,
896 lp_rast_triangle_4,
897 lp_rast_triangle_5,
898 lp_rast_triangle_6,
899 lp_rast_triangle_7,
900 lp_rast_triangle_8,
901 lp_rast_triangle_3_4,
902 lp_rast_triangle_3_16,
903 lp_rast_triangle_4_16,
904 lp_rast_shade_tile,
905 lp_rast_shade_tile,
906 lp_rast_begin_query,
907 lp_rast_end_query,
908 lp_rast_set_state,
909 lp_rast_triangle_32_1,
910 lp_rast_triangle_32_2,
911 lp_rast_triangle_32_3,
912 lp_rast_triangle_32_4,
913 lp_rast_triangle_32_5,
914 lp_rast_triangle_32_6,
915 lp_rast_triangle_32_7,
916 lp_rast_triangle_32_8,
917 lp_rast_triangle_32_3_4,
918 lp_rast_triangle_32_3_16,
919 lp_rast_triangle_32_4_16,
920 lp_rast_triangle_ms_1,
921 lp_rast_triangle_ms_2,
922 lp_rast_triangle_ms_3,
923 lp_rast_triangle_ms_4,
924 lp_rast_triangle_ms_5,
925 lp_rast_triangle_ms_6,
926 lp_rast_triangle_ms_7,
927 lp_rast_triangle_ms_8,
928 lp_rast_triangle_ms_3_4,
929 lp_rast_triangle_ms_3_16,
930 lp_rast_triangle_ms_4_16,
931
932 lp_rast_rectangle,
933 lp_rast_shade_tile,
934 };
935
936 struct lp_bin_info
lp_characterize_bin(const struct cmd_bin * bin)937 lp_characterize_bin(const struct cmd_bin *bin)
938 {
939 struct cmd_block *block;
940 struct lp_bin_info info;
941 unsigned andflags = ~0;
942 unsigned k, j = 0;
943
944 STATIC_ASSERT(ARRAY_SIZE(rast_flags) == LP_RAST_OP_MAX);
945
946 for (block = bin->head; block; block = block->next) {
947 for (k = 0; k < block->count; k++, j++) {
948 andflags &= rast_flags[block->cmd[k]];
949 }
950 }
951
952 info.type = andflags;
953 info.count = j;
954
955 return info;
956 }
957
958
959 static void
blit_rasterize_bin(struct lp_rasterizer_task * task,const struct cmd_bin * bin)960 blit_rasterize_bin(struct lp_rasterizer_task *task,
961 const struct cmd_bin *bin)
962 {
963 const struct cmd_block *block;
964 unsigned k;
965
966 STATIC_ASSERT(ARRAY_SIZE(dispatch_blit) == LP_RAST_OP_MAX);
967
968 if (0) debug_printf("%s\n", __FUNCTION__);
969 for (block = bin->head; block; block = block->next) {
970 for (k = 0; k < block->count; k++) {
971 dispatch_blit[block->cmd[k]]( task, block->arg[k] );
972 }
973 }
974 }
975
976 static void
tri_rasterize_bin(struct lp_rasterizer_task * task,const struct cmd_bin * bin,int x,int y)977 tri_rasterize_bin(struct lp_rasterizer_task *task,
978 const struct cmd_bin *bin,
979 int x, int y)
980 {
981 const struct cmd_block *block;
982 unsigned k;
983
984 STATIC_ASSERT(ARRAY_SIZE(dispatch_tri) == LP_RAST_OP_MAX);
985
986 for (block = bin->head; block; block = block->next) {
987 for (k = 0; k < block->count; k++) {
988 dispatch_tri[block->cmd[k]]( task, block->arg[k] );
989 }
990 }
991 }
992
993 static void
debug_rasterize_bin(struct lp_rasterizer_task * task,const struct cmd_bin * bin)994 debug_rasterize_bin(struct lp_rasterizer_task *task,
995 const struct cmd_bin *bin)
996 {
997 const struct cmd_block *block;
998 unsigned k;
999
1000 STATIC_ASSERT(ARRAY_SIZE(dispatch_tri_debug) == LP_RAST_OP_MAX);
1001
1002 for (block = bin->head; block; block = block->next) {
1003 for (k = 0; k < block->count; k++) {
1004 dispatch_tri_debug[block->cmd[k]]( task, block->arg[k] );
1005 }
1006 }
1007 }
1008
1009
1010 /**
1011 * Rasterize commands for a single bin.
1012 * \param x, y position of the bin's tile in the framebuffer
1013 * Must be called between lp_rast_begin() and lp_rast_end().
1014 * Called per thread.
1015 */
1016 static void
rasterize_bin(struct lp_rasterizer_task * task,const struct cmd_bin * bin,int x,int y)1017 rasterize_bin(struct lp_rasterizer_task *task,
1018 const struct cmd_bin *bin, int x, int y )
1019 {
1020 struct lp_bin_info info = lp_characterize_bin(bin);
1021
1022 lp_rast_tile_begin( task, bin, x, y );
1023
1024 if (LP_DEBUG & DEBUG_NO_FASTPATH)
1025 debug_rasterize_bin(task, bin);
1026 else if (info.type & LP_RAST_FLAGS_BLIT)
1027 blit_rasterize_bin(task, bin);
1028 else if (task->scene->permit_linear_rasterizer &&
1029 !(LP_PERF & PERF_NO_RAST_LINEAR) &&
1030 (info.type & LP_RAST_FLAGS_RECT))
1031 lp_linear_rasterize_bin(task, bin);
1032 else
1033 tri_rasterize_bin(task, bin, x, y);
1034
1035 lp_rast_tile_end(task);
1036
1037 #ifdef DEBUG
1038 /* Debug/Perf flags:
1039 */
1040 if (bin->head->count == 1) {
1041 if (bin->head->cmd[0] == LP_RAST_OP_BLIT)
1042 LP_COUNT(nr_pure_blit_64);
1043 else if (bin->head->cmd[0] == LP_RAST_OP_SHADE_TILE_OPAQUE)
1044 LP_COUNT(nr_pure_shade_opaque_64);
1045 else if (bin->head->cmd[0] == LP_RAST_OP_SHADE_TILE)
1046 LP_COUNT(nr_pure_shade_64);
1047 }
1048 #endif
1049 }
1050
1051
1052 /* An empty bin is one that just loads the contents of the tile and
1053 * stores them again unchanged. This typically happens when bins have
1054 * been flushed for some reason in the middle of a frame, or when
1055 * incremental updates are being made to a render target.
1056 *
1057 * Try to avoid doing pointless work in this case.
1058 */
1059 static boolean
is_empty_bin(const struct cmd_bin * bin)1060 is_empty_bin( const struct cmd_bin *bin )
1061 {
1062 return bin->head == NULL;
1063 }
1064
1065
1066 /**
1067 * Rasterize/execute all bins within a scene.
1068 * Called per thread.
1069 */
1070 static void
rasterize_scene(struct lp_rasterizer_task * task,struct lp_scene * scene)1071 rasterize_scene(struct lp_rasterizer_task *task,
1072 struct lp_scene *scene)
1073 {
1074 task->scene = scene;
1075
1076 /* Clear the cache tags. This should not always be necessary but
1077 simpler for now. */
1078 #if LP_USE_TEXTURE_CACHE
1079 memset(task->thread_data.cache->cache_tags, 0,
1080 sizeof(task->thread_data.cache->cache_tags));
1081 #if LP_BUILD_FORMAT_CACHE_DEBUG
1082 task->thread_data.cache->cache_access_total = 0;
1083 task->thread_data.cache->cache_access_miss = 0;
1084 #endif
1085 #endif
1086
1087 if (!task->rast->no_rast) {
1088 /* loop over scene bins, rasterize each */
1089 {
1090 struct cmd_bin *bin;
1091 int i, j;
1092
1093 assert(scene);
1094 while ((bin = lp_scene_bin_iter_next(scene, &i, &j))) {
1095 if (!is_empty_bin( bin ))
1096 rasterize_bin(task, bin, i, j);
1097 }
1098 }
1099 }
1100
1101
1102 #if LP_BUILD_FORMAT_CACHE_DEBUG
1103 {
1104 uint64_t total, miss;
1105 total = task->thread_data.cache->cache_access_total;
1106 miss = task->thread_data.cache->cache_access_miss;
1107 if (total) {
1108 debug_printf("thread %d cache access %llu miss %llu hit rate %f\n",
1109 task->thread_index, (long long unsigned)total,
1110 (long long unsigned)miss,
1111 (float)(total - miss)/(float)total);
1112 }
1113 }
1114 #endif
1115
1116 if (scene->fence) {
1117 lp_fence_signal(scene->fence);
1118 }
1119
1120 task->scene = NULL;
1121 }
1122
1123
1124 /**
1125 * Called by setup module when it has something for us to render.
1126 */
1127 void
lp_rast_queue_scene(struct lp_rasterizer * rast,struct lp_scene * scene)1128 lp_rast_queue_scene( struct lp_rasterizer *rast,
1129 struct lp_scene *scene)
1130 {
1131 LP_DBG(DEBUG_SETUP, "%s\n", __FUNCTION__);
1132
1133 if (rast->num_threads == 0) {
1134 /* no threading */
1135 unsigned fpstate = util_fpstate_get();
1136
1137 /* Make sure that denorms are treated like zeros. This is
1138 * the behavior required by D3D10. OpenGL doesn't care.
1139 */
1140 util_fpstate_set_denorms_to_zero(fpstate);
1141
1142 lp_rast_begin( rast, scene );
1143
1144 rasterize_scene( &rast->tasks[0], scene );
1145
1146 lp_rast_end( rast );
1147
1148 util_fpstate_set(fpstate);
1149
1150 rast->curr_scene = NULL;
1151 }
1152 else {
1153 /* threaded rendering! */
1154 unsigned i;
1155
1156 lp_scene_enqueue( rast->full_scenes, scene );
1157
1158 /* signal the threads that there's work to do */
1159 for (i = 0; i < rast->num_threads; i++) {
1160 pipe_semaphore_signal(&rast->tasks[i].work_ready);
1161 }
1162 }
1163
1164 LP_DBG(DEBUG_SETUP, "%s done \n", __FUNCTION__);
1165 }
1166
1167
1168 void
lp_rast_finish(struct lp_rasterizer * rast)1169 lp_rast_finish( struct lp_rasterizer *rast )
1170 {
1171 if (rast->num_threads == 0) {
1172 /* nothing to do */
1173 }
1174 else {
1175 int i;
1176
1177 /* wait for work to complete */
1178 for (i = 0; i < rast->num_threads; i++) {
1179 pipe_semaphore_wait(&rast->tasks[i].work_done);
1180 }
1181 }
1182 }
1183
1184
1185 /**
1186 * This is the thread's main entrypoint.
1187 * It's a simple loop:
1188 * 1. wait for work
1189 * 2. do work
1190 * 3. signal that we're done
1191 */
1192 static int
thread_function(void * init_data)1193 thread_function(void *init_data)
1194 {
1195 struct lp_rasterizer_task *task = (struct lp_rasterizer_task *) init_data;
1196 struct lp_rasterizer *rast = task->rast;
1197 boolean debug = false;
1198 char thread_name[16];
1199 unsigned fpstate;
1200
1201 snprintf(thread_name, sizeof thread_name, "llvmpipe-%u", task->thread_index);
1202 u_thread_setname(thread_name);
1203
1204 /* Make sure that denorms are treated like zeros. This is
1205 * the behavior required by D3D10. OpenGL doesn't care.
1206 */
1207 fpstate = util_fpstate_get();
1208 util_fpstate_set_denorms_to_zero(fpstate);
1209
1210 while (1) {
1211 /* wait for work */
1212 if (debug)
1213 debug_printf("thread %d waiting for work\n", task->thread_index);
1214 pipe_semaphore_wait(&task->work_ready);
1215
1216 if (rast->exit_flag)
1217 break;
1218
1219 if (task->thread_index == 0) {
1220 /* thread[0]:
1221 * - get next scene to rasterize
1222 * - map the framebuffer surfaces
1223 */
1224 lp_rast_begin( rast,
1225 lp_scene_dequeue( rast->full_scenes, TRUE ) );
1226 }
1227
1228 /* Wait for all threads to get here so that threads[1+] don't
1229 * get a null rast->curr_scene pointer.
1230 */
1231 util_barrier_wait( &rast->barrier );
1232
1233 /* do work */
1234 if (debug)
1235 debug_printf("thread %d doing work\n", task->thread_index);
1236
1237 rasterize_scene(task,
1238 rast->curr_scene);
1239
1240 /* wait for all threads to finish with this scene */
1241 util_barrier_wait( &rast->barrier );
1242
1243 /* XXX: shouldn't be necessary:
1244 */
1245 if (task->thread_index == 0) {
1246 lp_rast_end( rast );
1247 }
1248
1249 /* signal done with work */
1250 if (debug)
1251 debug_printf("thread %d done working\n", task->thread_index);
1252
1253 pipe_semaphore_signal(&task->work_done);
1254 }
1255
1256 #ifdef _WIN32
1257 pipe_semaphore_signal(&task->work_done);
1258 #endif
1259
1260 return 0;
1261 }
1262
1263
1264 /**
1265 * Initialize semaphores and spawn the threads.
1266 */
1267 static void
create_rast_threads(struct lp_rasterizer * rast)1268 create_rast_threads(struct lp_rasterizer *rast)
1269 {
1270 unsigned i;
1271
1272 /* NOTE: if num_threads is zero, we won't use any threads */
1273 for (i = 0; i < rast->num_threads; i++) {
1274 pipe_semaphore_init(&rast->tasks[i].work_ready, 0);
1275 pipe_semaphore_init(&rast->tasks[i].work_done, 0);
1276 rast->threads[i] = u_thread_create(thread_function,
1277 (void *) &rast->tasks[i]);
1278 if (!rast->threads[i]) {
1279 rast->num_threads = i; /* previous thread is max */
1280 break;
1281 }
1282 }
1283 }
1284
1285
1286
1287 /**
1288 * Create new lp_rasterizer. If num_threads is zero, don't create any
1289 * new threads, do rendering synchronously.
1290 * \param num_threads number of rasterizer threads to create
1291 */
1292 struct lp_rasterizer *
lp_rast_create(unsigned num_threads)1293 lp_rast_create( unsigned num_threads )
1294 {
1295 struct lp_rasterizer *rast;
1296 unsigned i;
1297
1298 rast = CALLOC_STRUCT(lp_rasterizer);
1299 if (!rast) {
1300 goto no_rast;
1301 }
1302
1303 rast->full_scenes = lp_scene_queue_create();
1304 if (!rast->full_scenes) {
1305 goto no_full_scenes;
1306 }
1307
1308 for (i = 0; i < MAX2(1, num_threads); i++) {
1309 struct lp_rasterizer_task *task = &rast->tasks[i];
1310 task->rast = rast;
1311 task->thread_index = i;
1312 task->thread_data.cache = align_malloc(sizeof(struct lp_build_format_cache),
1313 16);
1314 if (!task->thread_data.cache) {
1315 goto no_thread_data_cache;
1316 }
1317 }
1318
1319 rast->num_threads = num_threads;
1320
1321 rast->no_rast = debug_get_bool_option("LP_NO_RAST", FALSE);
1322
1323 create_rast_threads(rast);
1324
1325 /* for synchronizing rasterization threads */
1326 if (rast->num_threads > 0) {
1327 util_barrier_init( &rast->barrier, rast->num_threads );
1328 }
1329
1330 memset(lp_dummy_tile, 0, sizeof lp_dummy_tile);
1331
1332 return rast;
1333
1334 no_thread_data_cache:
1335 for (i = 0; i < MAX2(1, rast->num_threads); i++) {
1336 if (rast->tasks[i].thread_data.cache) {
1337 align_free(rast->tasks[i].thread_data.cache);
1338 }
1339 }
1340
1341 lp_scene_queue_destroy(rast->full_scenes);
1342 no_full_scenes:
1343 FREE(rast);
1344 no_rast:
1345 return NULL;
1346 }
1347
1348
1349 /* Shutdown:
1350 */
lp_rast_destroy(struct lp_rasterizer * rast)1351 void lp_rast_destroy( struct lp_rasterizer *rast )
1352 {
1353 unsigned i;
1354
1355 /* Set exit_flag and signal each thread's work_ready semaphore.
1356 * Each thread will be woken up, notice that the exit_flag is set and
1357 * break out of its main loop. The thread will then exit.
1358 */
1359 rast->exit_flag = TRUE;
1360 for (i = 0; i < rast->num_threads; i++) {
1361 pipe_semaphore_signal(&rast->tasks[i].work_ready);
1362 }
1363
1364 /* Wait for threads to terminate before cleaning up per-thread data.
1365 * We don't actually call pipe_thread_wait to avoid dead lock on Windows
1366 * per https://bugs.freedesktop.org/show_bug.cgi?id=76252 */
1367 for (i = 0; i < rast->num_threads; i++) {
1368 #ifdef _WIN32
1369 /* Threads might already be dead - Windows apparently terminates other threads when
1370 * returning from main.
1371 */
1372 DWORD exit_code = STILL_ACTIVE;
1373 if (GetExitCodeThread(rast->threads[i], &exit_code) && exit_code == STILL_ACTIVE)
1374 pipe_semaphore_wait(&rast->tasks[i].work_done);
1375 #else
1376 thrd_join(rast->threads[i], NULL);
1377 #endif
1378 }
1379
1380 /* Clean up per-thread data */
1381 for (i = 0; i < rast->num_threads; i++) {
1382 pipe_semaphore_destroy(&rast->tasks[i].work_ready);
1383 pipe_semaphore_destroy(&rast->tasks[i].work_done);
1384 }
1385 for (i = 0; i < MAX2(1, rast->num_threads); i++) {
1386 align_free(rast->tasks[i].thread_data.cache);
1387 }
1388
1389 /* for synchronizing rasterization threads */
1390 if (rast->num_threads > 0) {
1391 util_barrier_destroy( &rast->barrier );
1392 }
1393
1394 lp_scene_queue_destroy(rast->full_scenes);
1395
1396 FREE(rast);
1397 }
1398
1399
1400