llvmpipe: allow tri_3_16 at any 4-aligned location within a tile
[mesa.git] / src / gallium / drivers / llvmpipe / lp_setup_tri.c
1 /**************************************************************************
2 *
3 * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas.
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 TUNGSTEN GRAPHICS 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 /*
29 * Binning code for triangles
30 */
31
32 #include "util/u_math.h"
33 #include "util/u_memory.h"
34 #include "util/u_rect.h"
35 #include "lp_perf.h"
36 #include "lp_setup_context.h"
37 #include "lp_setup_coef.h"
38 #include "lp_rast.h"
39 #include "lp_state_fs.h"
40
41 #define NUM_CHANNELS 4
42
43
44
45 static INLINE int
46 subpixel_snap(float a)
47 {
48 return util_iround(FIXED_ONE * a);
49 }
50
51 static INLINE float
52 fixed_to_float(int a)
53 {
54 return a * (1.0 / FIXED_ONE);
55 }
56
57
58
59
60
61
62
63 /**
64 * Alloc space for a new triangle plus the input.a0/dadx/dady arrays
65 * immediately after it.
66 * The memory is allocated from the per-scene pool, not per-tile.
67 * \param tri_size returns number of bytes allocated
68 * \param nr_inputs number of fragment shader inputs
69 * \return pointer to triangle space
70 */
71 struct lp_rast_triangle *
72 lp_setup_alloc_triangle(struct lp_scene *scene,
73 unsigned nr_inputs,
74 unsigned nr_planes,
75 unsigned *tri_size)
76 {
77 unsigned input_array_sz = NUM_CHANNELS * (nr_inputs + 1) * sizeof(float);
78 struct lp_rast_triangle *tri;
79 unsigned tri_bytes, bytes;
80 char *inputs;
81
82 tri_bytes = align(Offset(struct lp_rast_triangle, plane[nr_planes]), 16);
83 bytes = tri_bytes + (3 * input_array_sz);
84
85 tri = lp_scene_alloc_aligned( scene, bytes, 16 );
86
87 if (tri) {
88 inputs = ((char *)tri) + tri_bytes;
89 tri->inputs.a0 = (float (*)[4]) inputs;
90 tri->inputs.dadx = (float (*)[4]) (inputs + input_array_sz);
91 tri->inputs.dady = (float (*)[4]) (inputs + 2 * input_array_sz);
92
93 *tri_size = bytes;
94 }
95
96 return tri;
97 }
98
99 void
100 lp_setup_print_vertex(struct lp_setup_context *setup,
101 const char *name,
102 const float (*v)[4])
103 {
104 int i, j;
105
106 debug_printf(" wpos (%s[0]) xyzw %f %f %f %f\n",
107 name,
108 v[0][0], v[0][1], v[0][2], v[0][3]);
109
110 for (i = 0; i < setup->fs.nr_inputs; i++) {
111 const float *in = v[setup->fs.input[i].src_index];
112
113 debug_printf(" in[%d] (%s[%d]) %s%s%s%s ",
114 i,
115 name, setup->fs.input[i].src_index,
116 (setup->fs.input[i].usage_mask & 0x1) ? "x" : " ",
117 (setup->fs.input[i].usage_mask & 0x2) ? "y" : " ",
118 (setup->fs.input[i].usage_mask & 0x4) ? "z" : " ",
119 (setup->fs.input[i].usage_mask & 0x8) ? "w" : " ");
120
121 for (j = 0; j < 4; j++)
122 if (setup->fs.input[i].usage_mask & (1<<j))
123 debug_printf("%.5f ", in[j]);
124
125 debug_printf("\n");
126 }
127 }
128
129
130 /**
131 * Print triangle vertex attribs (for debug).
132 */
133 void
134 lp_setup_print_triangle(struct lp_setup_context *setup,
135 const float (*v0)[4],
136 const float (*v1)[4],
137 const float (*v2)[4])
138 {
139 debug_printf("triangle\n");
140
141 {
142 const float ex = v0[0][0] - v2[0][0];
143 const float ey = v0[0][1] - v2[0][1];
144 const float fx = v1[0][0] - v2[0][0];
145 const float fy = v1[0][1] - v2[0][1];
146
147 /* det = cross(e,f).z */
148 const float det = ex * fy - ey * fx;
149 if (det < 0.0f)
150 debug_printf(" - ccw\n");
151 else if (det > 0.0f)
152 debug_printf(" - cw\n");
153 else
154 debug_printf(" - zero area\n");
155 }
156
157 lp_setup_print_vertex(setup, "v0", v0);
158 lp_setup_print_vertex(setup, "v1", v1);
159 lp_setup_print_vertex(setup, "v2", v2);
160 }
161
162
163 static unsigned
164 lp_rast_tri_tab[9] = {
165 0, /* should be impossible */
166 LP_RAST_OP_TRIANGLE_1,
167 LP_RAST_OP_TRIANGLE_2,
168 LP_RAST_OP_TRIANGLE_3,
169 LP_RAST_OP_TRIANGLE_4,
170 LP_RAST_OP_TRIANGLE_5,
171 LP_RAST_OP_TRIANGLE_6,
172 LP_RAST_OP_TRIANGLE_7,
173 LP_RAST_OP_TRIANGLE_8
174 };
175
176
177
178 /**
179 * The primitive covers the whole tile- shade whole tile.
180 *
181 * \param tx, ty the tile position in tiles, not pixels
182 */
183 static boolean
184 lp_setup_whole_tile(struct lp_setup_context *setup,
185 const struct lp_rast_shader_inputs *inputs,
186 int tx, int ty)
187 {
188 struct lp_scene *scene = setup->scene;
189
190 LP_COUNT(nr_fully_covered_64);
191
192 /* if variant is opaque and scissor doesn't effect the tile */
193 if (inputs->opaque) {
194 if (!scene->fb.zsbuf) {
195 /*
196 * All previous rendering will be overwritten so reset the bin.
197 */
198 lp_scene_bin_reset( scene, tx, ty );
199 }
200
201 LP_COUNT(nr_shade_opaque_64);
202 return lp_scene_bin_command( scene, tx, ty,
203 LP_RAST_OP_SHADE_TILE_OPAQUE,
204 lp_rast_arg_inputs(inputs) );
205 } else {
206 LP_COUNT(nr_shade_64);
207 return lp_scene_bin_command( scene, tx, ty,
208 LP_RAST_OP_SHADE_TILE,
209 lp_rast_arg_inputs(inputs) );
210 }
211 }
212
213
214 /**
215 * Do basic setup for triangle rasterization and determine which
216 * framebuffer tiles are touched. Put the triangle in the scene's
217 * bins for the tiles which we overlap.
218 */
219 static boolean
220 do_triangle_ccw(struct lp_setup_context *setup,
221 const float (*v0)[4],
222 const float (*v1)[4],
223 const float (*v2)[4],
224 boolean frontfacing )
225 {
226 struct lp_scene *scene = setup->scene;
227 struct lp_rast_triangle *tri;
228 int x[3];
229 int y[3];
230 int area;
231 struct u_rect bbox;
232 unsigned tri_bytes;
233 int i;
234 int nr_planes = 3;
235
236 if (0)
237 lp_setup_print_triangle(setup, v0, v1, v2);
238
239 if (setup->scissor_test) {
240 nr_planes = 7;
241 }
242 else {
243 nr_planes = 3;
244 }
245
246 /* x/y positions in fixed point */
247 x[0] = subpixel_snap(v0[0][0] - setup->pixel_offset);
248 x[1] = subpixel_snap(v1[0][0] - setup->pixel_offset);
249 x[2] = subpixel_snap(v2[0][0] - setup->pixel_offset);
250 y[0] = subpixel_snap(v0[0][1] - setup->pixel_offset);
251 y[1] = subpixel_snap(v1[0][1] - setup->pixel_offset);
252 y[2] = subpixel_snap(v2[0][1] - setup->pixel_offset);
253
254
255 /* Bounding rectangle (in pixels) */
256 {
257 /* Yes this is necessary to accurately calculate bounding boxes
258 * with the two fill-conventions we support. GL (normally) ends
259 * up needing a bottom-left fill convention, which requires
260 * slightly different rounding.
261 */
262 int adj = (setup->pixel_offset != 0) ? 1 : 0;
263
264 bbox.x0 = (MIN3(x[0], x[1], x[2]) + (FIXED_ONE-1)) >> FIXED_ORDER;
265 bbox.x1 = (MAX3(x[0], x[1], x[2]) + (FIXED_ONE-1)) >> FIXED_ORDER;
266 bbox.y0 = (MIN3(y[0], y[1], y[2]) + (FIXED_ONE-1) + adj) >> FIXED_ORDER;
267 bbox.y1 = (MAX3(y[0], y[1], y[2]) + (FIXED_ONE-1) + adj) >> FIXED_ORDER;
268
269 /* Inclusive coordinates:
270 */
271 bbox.x1--;
272 bbox.y1--;
273 }
274
275 if (bbox.x1 < bbox.x0 ||
276 bbox.y1 < bbox.y0) {
277 if (0) debug_printf("empty bounding box\n");
278 LP_COUNT(nr_culled_tris);
279 return TRUE;
280 }
281
282 if (!u_rect_test_intersection(&setup->draw_region, &bbox)) {
283 if (0) debug_printf("offscreen\n");
284 LP_COUNT(nr_culled_tris);
285 return TRUE;
286 }
287
288 u_rect_find_intersection(&setup->draw_region, &bbox);
289
290 tri = lp_setup_alloc_triangle(scene,
291 setup->fs.nr_inputs,
292 nr_planes,
293 &tri_bytes);
294 if (!tri)
295 return FALSE;
296
297 #ifdef DEBUG
298 tri->v[0][0] = v0[0][0];
299 tri->v[1][0] = v1[0][0];
300 tri->v[2][0] = v2[0][0];
301 tri->v[0][1] = v0[0][1];
302 tri->v[1][1] = v1[0][1];
303 tri->v[2][1] = v2[0][1];
304 #endif
305
306 tri->plane[0].dcdy = x[0] - x[1];
307 tri->plane[1].dcdy = x[1] - x[2];
308 tri->plane[2].dcdy = x[2] - x[0];
309
310 tri->plane[0].dcdx = y[0] - y[1];
311 tri->plane[1].dcdx = y[1] - y[2];
312 tri->plane[2].dcdx = y[2] - y[0];
313
314 area = (tri->plane[0].dcdy * tri->plane[2].dcdx -
315 tri->plane[2].dcdy * tri->plane[0].dcdx);
316
317 LP_COUNT(nr_tris);
318
319 /* Cull non-ccw and zero-sized triangles.
320 *
321 * XXX: subject to overflow??
322 */
323 if (area <= 0) {
324 lp_scene_putback_data( scene, tri_bytes );
325 LP_COUNT(nr_culled_tris);
326 return TRUE;
327 }
328
329 /* Setup parameter interpolants:
330 */
331 lp_setup_tri_coef( setup, &tri->inputs, v0, v1, v2, frontfacing );
332
333 tri->inputs.facing = frontfacing ? 1.0F : -1.0F;
334 tri->inputs.disable = FALSE;
335 tri->inputs.opaque = setup->fs.current.variant->opaque;
336 tri->inputs.state = setup->fs.stored;
337
338
339 for (i = 0; i < 3; i++) {
340 struct lp_rast_plane *plane = &tri->plane[i];
341
342 /* half-edge constants, will be interated over the whole render
343 * target.
344 */
345 plane->c = plane->dcdx * x[i] - plane->dcdy * y[i];
346
347 /* correct for top-left vs. bottom-left fill convention.
348 *
349 * note that we're overloading gl_rasterization_rules to mean
350 * both (0.5,0.5) pixel centers *and* bottom-left filling
351 * convention.
352 *
353 * GL actually has a top-left filling convention, but GL's
354 * notion of "top" differs from gallium's...
355 *
356 * Also, sometimes (in FBO cases) GL will render upside down
357 * to its usual method, in which case it will probably want
358 * to use the opposite, top-left convention.
359 */
360 if (plane->dcdx < 0) {
361 /* both fill conventions want this - adjust for left edges */
362 plane->c++;
363 }
364 else if (plane->dcdx == 0) {
365 if (setup->pixel_offset == 0) {
366 /* correct for top-left fill convention:
367 */
368 if (plane->dcdy > 0) plane->c++;
369 }
370 else {
371 /* correct for bottom-left fill convention:
372 */
373 if (plane->dcdy < 0) plane->c++;
374 }
375 }
376
377 plane->dcdx *= FIXED_ONE;
378 plane->dcdy *= FIXED_ONE;
379
380 /* find trivial reject offsets for each edge for a single-pixel
381 * sized block. These will be scaled up at each recursive level to
382 * match the active blocksize. Scaling in this way works best if
383 * the blocks are square.
384 */
385 plane->eo = 0;
386 if (plane->dcdx < 0) plane->eo -= plane->dcdx;
387 if (plane->dcdy > 0) plane->eo += plane->dcdy;
388
389 /* Calculate trivial accept offsets from the above.
390 */
391 plane->ei = plane->dcdy - plane->dcdx - plane->eo;
392 }
393
394
395 /*
396 * When rasterizing scissored tris, use the intersection of the
397 * triangle bounding box and the scissor rect to generate the
398 * scissor planes.
399 *
400 * This permits us to cut off the triangle "tails" that are present
401 * in the intermediate recursive levels caused when two of the
402 * triangles edges don't diverge quickly enough to trivially reject
403 * exterior blocks from the triangle.
404 *
405 * It's not really clear if it's worth worrying about these tails,
406 * but since we generate the planes for each scissored tri, it's
407 * free to trim them in this case.
408 *
409 * Note that otherwise, the scissor planes only vary in 'C' value,
410 * and even then only on state-changes. Could alternatively store
411 * these planes elsewhere.
412 */
413 if (nr_planes == 7) {
414 tri->plane[3].dcdx = -1;
415 tri->plane[3].dcdy = 0;
416 tri->plane[3].c = 1-bbox.x0;
417 tri->plane[3].ei = 0;
418 tri->plane[3].eo = 1;
419
420 tri->plane[4].dcdx = 1;
421 tri->plane[4].dcdy = 0;
422 tri->plane[4].c = bbox.x1+1;
423 tri->plane[4].ei = -1;
424 tri->plane[4].eo = 0;
425
426 tri->plane[5].dcdx = 0;
427 tri->plane[5].dcdy = 1;
428 tri->plane[5].c = 1-bbox.y0;
429 tri->plane[5].ei = 0;
430 tri->plane[5].eo = 1;
431
432 tri->plane[6].dcdx = 0;
433 tri->plane[6].dcdy = -1;
434 tri->plane[6].c = bbox.y1+1;
435 tri->plane[6].ei = -1;
436 tri->plane[6].eo = 0;
437 }
438
439 return lp_setup_bin_triangle( setup, tri, &bbox, nr_planes );
440 }
441
442 /*
443 * __fls: find last set bit in word
444 * @word: The word to search
445 *
446 * Undefined if no zero exists, so code should check against ~0UL first.
447 */
448 #if defined(PIPE_ARCH_X86)
449 static inline unsigned fls(unsigned word)
450 {
451 asm("bsr %1,%0"
452 : "=r" (word)
453 : "rm" (word));
454 return word;
455 }
456 #else
457 static inline unsigned fls(unsigned n)
458 {
459 n |= (n >> 1);
460 n |= (n >> 2);
461 n |= (n >> 4);
462 n |= (n >> 8);
463 n |= (n >> 16);
464 return n - (n >> 1);
465 }
466 #endif
467
468
469 boolean
470 lp_setup_bin_triangle( struct lp_setup_context *setup,
471 struct lp_rast_triangle *tri,
472 const struct u_rect *bbox,
473 int nr_planes )
474 {
475 struct lp_scene *scene = setup->scene;
476 int i;
477
478 /* What is the largest power-of-two boundary this triangle crosses:
479 */
480 int dx = 1 << fls((bbox->x0 ^ bbox->x1) |
481 (bbox->y0 ^ bbox->y1));
482
483 /* The largest dimension of the rasterized area of the triangle
484 * (aligned to a 4x4 grid), rounded up to the next power of two:
485 */
486 int sz = 1 << fls((bbox->x1 - (bbox->x0 & ~3)) |
487 (bbox->y1 - (bbox->y0 & ~3)));
488
489 if (nr_planes == 3) {
490 if (sz < 16 && dx < 64)
491 {
492 int mask = (bbox->x0 & 63 & ~3) | ((bbox->y0 & 63 & ~3) << 8);
493
494 /* Triangle is contained in a single 16x16 block:
495 */
496 return lp_scene_bin_command( scene,
497 bbox->x0/64, bbox->y0/64,
498 LP_RAST_OP_TRIANGLE_3_16,
499 lp_rast_arg_triangle(tri, mask) );
500 }
501 }
502
503
504 /* Determine which tile(s) intersect the triangle's bounding box
505 */
506 if (dx < TILE_SIZE)
507 {
508 int ix0 = bbox->x0 / TILE_SIZE;
509 int iy0 = bbox->y0 / TILE_SIZE;
510
511 assert(iy0 == bbox->y1 / TILE_SIZE &&
512 ix0 == bbox->x1 / TILE_SIZE);
513
514 /* Triangle is contained in a single tile:
515 */
516 return lp_scene_bin_command( scene, ix0, iy0,
517 lp_rast_tri_tab[nr_planes],
518 lp_rast_arg_triangle(tri, (1<<nr_planes)-1) );
519 }
520 else
521 {
522 int c[7];
523 int ei[7];
524 int eo[7];
525 int xstep[7];
526 int ystep[7];
527 int x, y;
528
529 int ix0 = bbox->x0 / TILE_SIZE;
530 int iy0 = bbox->y0 / TILE_SIZE;
531 int ix1 = bbox->x1 / TILE_SIZE;
532 int iy1 = bbox->y1 / TILE_SIZE;
533
534 for (i = 0; i < nr_planes; i++) {
535 c[i] = (tri->plane[i].c +
536 tri->plane[i].dcdy * iy0 * TILE_SIZE -
537 tri->plane[i].dcdx * ix0 * TILE_SIZE);
538
539 ei[i] = tri->plane[i].ei << TILE_ORDER;
540 eo[i] = tri->plane[i].eo << TILE_ORDER;
541 xstep[i] = -(tri->plane[i].dcdx << TILE_ORDER);
542 ystep[i] = tri->plane[i].dcdy << TILE_ORDER;
543 }
544
545
546
547 /* Test tile-sized blocks against the triangle.
548 * Discard blocks fully outside the tri. If the block is fully
549 * contained inside the tri, bin an lp_rast_shade_tile command.
550 * Else, bin a lp_rast_triangle command.
551 */
552 for (y = iy0; y <= iy1; y++)
553 {
554 boolean in = FALSE; /* are we inside the triangle? */
555 int cx[7];
556
557 for (i = 0; i < nr_planes; i++)
558 cx[i] = c[i];
559
560 for (x = ix0; x <= ix1; x++)
561 {
562 int out = 0;
563 int partial = 0;
564
565 for (i = 0; i < nr_planes; i++) {
566 int planeout = cx[i] + eo[i];
567 int planepartial = cx[i] + ei[i] - 1;
568 out |= (planeout >> 31);
569 partial |= (planepartial >> 31) & (1<<i);
570 }
571
572 if (out) {
573 /* do nothing */
574 if (in)
575 break; /* exiting triangle, all done with this row */
576 LP_COUNT(nr_empty_64);
577 }
578 else if (partial) {
579 /* Not trivially accepted by at least one plane -
580 * rasterize/shade partial tile
581 */
582 int count = util_bitcount(partial);
583 in = TRUE;
584 if (!lp_scene_bin_command( scene, x, y,
585 lp_rast_tri_tab[count],
586 lp_rast_arg_triangle(tri, partial) ))
587 goto fail;
588
589 LP_COUNT(nr_partially_covered_64);
590 }
591 else {
592 /* triangle covers the whole tile- shade whole tile */
593 LP_COUNT(nr_fully_covered_64);
594 in = TRUE;
595 if (!lp_setup_whole_tile(setup, &tri->inputs, x, y))
596 goto fail;
597 }
598
599 /* Iterate cx values across the region:
600 */
601 for (i = 0; i < nr_planes; i++)
602 cx[i] += xstep[i];
603 }
604
605 /* Iterate c values down the region:
606 */
607 for (i = 0; i < nr_planes; i++)
608 c[i] += ystep[i];
609 }
610 }
611
612 return TRUE;
613
614 fail:
615 /* Need to disable any partially binned triangle. This is easier
616 * than trying to locate all the triangle, shade-tile, etc,
617 * commands which may have been binned.
618 */
619 tri->inputs.disable = TRUE;
620 return FALSE;
621 }
622
623
624 /**
625 * Draw triangle if it's CW, cull otherwise.
626 */
627 static void triangle_cw( struct lp_setup_context *setup,
628 const float (*v0)[4],
629 const float (*v1)[4],
630 const float (*v2)[4] )
631 {
632 if (!do_triangle_ccw( setup, v1, v0, v2, !setup->ccw_is_frontface ))
633 {
634 lp_setup_flush_and_restart(setup);
635
636 if (!do_triangle_ccw( setup, v1, v0, v2, !setup->ccw_is_frontface ))
637 assert(0);
638 }
639 }
640
641
642 /**
643 * Draw triangle if it's CCW, cull otherwise.
644 */
645 static void triangle_ccw( struct lp_setup_context *setup,
646 const float (*v0)[4],
647 const float (*v1)[4],
648 const float (*v2)[4] )
649 {
650 if (!do_triangle_ccw( setup, v0, v1, v2, setup->ccw_is_frontface ))
651 {
652 lp_setup_flush_and_restart(setup);
653 if (!do_triangle_ccw( setup, v0, v1, v2, setup->ccw_is_frontface ))
654 assert(0);
655 }
656 }
657
658
659
660 /**
661 * Draw triangle whether it's CW or CCW.
662 */
663 static void triangle_both( struct lp_setup_context *setup,
664 const float (*v0)[4],
665 const float (*v1)[4],
666 const float (*v2)[4] )
667 {
668 /* edge vectors e = v0 - v2, f = v1 - v2 */
669 const float ex = v0[0][0] - v2[0][0];
670 const float ey = v0[0][1] - v2[0][1];
671 const float fx = v1[0][0] - v2[0][0];
672 const float fy = v1[0][1] - v2[0][1];
673
674 /* det = cross(e,f).z */
675 const float det = ex * fy - ey * fx;
676 if (det < 0.0f)
677 triangle_ccw( setup, v0, v1, v2 );
678 else if (det > 0.0f)
679 triangle_cw( setup, v0, v1, v2 );
680 }
681
682
683 static void triangle_nop( struct lp_setup_context *setup,
684 const float (*v0)[4],
685 const float (*v1)[4],
686 const float (*v2)[4] )
687 {
688 }
689
690
691 void
692 lp_setup_choose_triangle( struct lp_setup_context *setup )
693 {
694 switch (setup->cullmode) {
695 case PIPE_FACE_NONE:
696 setup->triangle = triangle_both;
697 break;
698 case PIPE_FACE_BACK:
699 setup->triangle = setup->ccw_is_frontface ? triangle_ccw : triangle_cw;
700 break;
701 case PIPE_FACE_FRONT:
702 setup->triangle = setup->ccw_is_frontface ? triangle_cw : triangle_ccw;
703 break;
704 default:
705 setup->triangle = triangle_nop;
706 break;
707 }
708 }