1 /**************************************************************************
3 * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas.
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 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.
26 **************************************************************************/
29 * Binning code for triangles
32 #include "util/u_math.h"
33 #include "util/u_memory.h"
34 #include "util/u_rect.h"
35 #include "util/u_sse.h"
37 #include "lp_setup_context.h"
39 #include "lp_state_fs.h"
40 #include "lp_state_setup.h"
42 #define NUM_CHANNELS 4
44 #if defined(PIPE_ARCH_SSE)
45 #include <emmintrin.h>
49 subpixel_snap(float a)
51 return util_iround(FIXED_ONE * a);
57 return a * (1.0 / FIXED_ONE);
67 * Alloc space for a new triangle plus the input.a0/dadx/dady arrays
68 * immediately after it.
69 * The memory is allocated from the per-scene pool, not per-tile.
70 * \param tri_size returns number of bytes allocated
71 * \param num_inputs number of fragment shader inputs
72 * \return pointer to triangle space
74 struct lp_rast_triangle *
75 lp_setup_alloc_triangle(struct lp_scene *scene,
80 unsigned input_array_sz = NUM_CHANNELS * (nr_inputs + 1) * sizeof(float);
81 unsigned plane_sz = nr_planes * sizeof(struct lp_rast_plane);
82 struct lp_rast_triangle *tri;
84 *tri_size = (sizeof(struct lp_rast_triangle) +
88 tri = lp_scene_alloc_aligned( scene, *tri_size, 16 );
92 tri->inputs.stride = input_array_sz;
95 char *a = (char *)tri;
96 char *b = (char *)&GET_PLANES(tri)[nr_planes];
97 assert(b - a == *tri_size);
104 lp_setup_print_vertex(struct lp_setup_context *setup,
108 const struct lp_setup_variant_key *key = &setup->setup.variant->key;
111 debug_printf(" wpos (%s[0]) xyzw %f %f %f %f\n",
113 v[0][0], v[0][1], v[0][2], v[0][3]);
115 for (i = 0; i < key->num_inputs; i++) {
116 const float *in = v[key->inputs[i].src_index];
118 debug_printf(" in[%d] (%s[%d]) %s%s%s%s ",
120 name, key->inputs[i].src_index,
121 (key->inputs[i].usage_mask & 0x1) ? "x" : " ",
122 (key->inputs[i].usage_mask & 0x2) ? "y" : " ",
123 (key->inputs[i].usage_mask & 0x4) ? "z" : " ",
124 (key->inputs[i].usage_mask & 0x8) ? "w" : " ");
126 for (j = 0; j < 4; j++)
127 if (key->inputs[i].usage_mask & (1<<j))
128 debug_printf("%.5f ", in[j]);
136 * Print triangle vertex attribs (for debug).
139 lp_setup_print_triangle(struct lp_setup_context *setup,
140 const float (*v0)[4],
141 const float (*v1)[4],
142 const float (*v2)[4])
144 debug_printf("triangle\n");
147 const float ex = v0[0][0] - v2[0][0];
148 const float ey = v0[0][1] - v2[0][1];
149 const float fx = v1[0][0] - v2[0][0];
150 const float fy = v1[0][1] - v2[0][1];
152 /* det = cross(e,f).z */
153 const float det = ex * fy - ey * fx;
155 debug_printf(" - ccw\n");
157 debug_printf(" - cw\n");
159 debug_printf(" - zero area\n");
162 lp_setup_print_vertex(setup, "v0", v0);
163 lp_setup_print_vertex(setup, "v1", v1);
164 lp_setup_print_vertex(setup, "v2", v2);
170 lp_rast_tri_tab[MAX_PLANES+1] = {
171 0, /* should be impossible */
172 LP_RAST_OP_TRIANGLE_1,
173 LP_RAST_OP_TRIANGLE_2,
174 LP_RAST_OP_TRIANGLE_3,
175 LP_RAST_OP_TRIANGLE_4,
176 LP_RAST_OP_TRIANGLE_5,
177 LP_RAST_OP_TRIANGLE_6,
178 LP_RAST_OP_TRIANGLE_7,
179 LP_RAST_OP_TRIANGLE_8
185 * The primitive covers the whole tile- shade whole tile.
187 * \param tx, ty the tile position in tiles, not pixels
190 lp_setup_whole_tile(struct lp_setup_context *setup,
191 const struct lp_rast_shader_inputs *inputs,
194 struct lp_scene *scene = setup->scene;
196 LP_COUNT(nr_fully_covered_64);
198 /* if variant is opaque and scissor doesn't effect the tile */
199 if (inputs->opaque) {
200 if (!scene->fb.zsbuf) {
202 * All previous rendering will be overwritten so reset the bin.
204 lp_scene_bin_reset( scene, tx, ty );
207 LP_COUNT(nr_shade_opaque_64);
208 return lp_scene_bin_cmd_with_state( scene, tx, ty,
210 LP_RAST_OP_SHADE_TILE_OPAQUE,
211 lp_rast_arg_inputs(inputs) );
213 LP_COUNT(nr_shade_64);
214 return lp_scene_bin_cmd_with_state( scene, tx, ty,
216 LP_RAST_OP_SHADE_TILE,
217 lp_rast_arg_inputs(inputs) );
223 * Do basic setup for triangle rasterization and determine which
224 * framebuffer tiles are touched. Put the triangle in the scene's
225 * bins for the tiles which we overlap.
228 do_triangle_ccw(struct lp_setup_context *setup,
229 const float (*v0)[4],
230 const float (*v1)[4],
231 const float (*v2)[4],
232 boolean frontfacing )
234 struct lp_scene *scene = setup->scene;
235 const struct lp_setup_variant_key *key = &setup->setup.variant->key;
236 struct lp_rast_triangle *tri;
237 struct lp_rast_plane *plane;
245 lp_setup_print_triangle(setup, v0, v1, v2);
247 if (setup->scissor_test) {
254 /* x/y positions in fixed point */
255 x[0] = subpixel_snap(v0[0][0] - setup->pixel_offset);
256 x[1] = subpixel_snap(v1[0][0] - setup->pixel_offset);
257 x[2] = subpixel_snap(v2[0][0] - setup->pixel_offset);
259 y[0] = subpixel_snap(v0[0][1] - setup->pixel_offset);
260 y[1] = subpixel_snap(v1[0][1] - setup->pixel_offset);
261 y[2] = subpixel_snap(v2[0][1] - setup->pixel_offset);
265 /* Bounding rectangle (in pixels) */
267 /* Yes this is necessary to accurately calculate bounding boxes
268 * with the two fill-conventions we support. GL (normally) ends
269 * up needing a bottom-left fill convention, which requires
270 * slightly different rounding.
272 int adj = (setup->pixel_offset != 0) ? 1 : 0;
274 bbox.x0 = (MIN3(x[0], x[1], x[2]) + (FIXED_ONE-1)) >> FIXED_ORDER;
275 bbox.x1 = (MAX3(x[0], x[1], x[2]) + (FIXED_ONE-1)) >> FIXED_ORDER;
276 bbox.y0 = (MIN3(y[0], y[1], y[2]) + (FIXED_ONE-1) + adj) >> FIXED_ORDER;
277 bbox.y1 = (MAX3(y[0], y[1], y[2]) + (FIXED_ONE-1) + adj) >> FIXED_ORDER;
279 /* Inclusive coordinates:
285 if (bbox.x1 < bbox.x0 ||
287 if (0) debug_printf("empty bounding box\n");
288 LP_COUNT(nr_culled_tris);
292 if (!u_rect_test_intersection(&setup->draw_region, &bbox)) {
293 if (0) debug_printf("offscreen\n");
294 LP_COUNT(nr_culled_tris);
298 /* Can safely discard negative regions, but need to keep hold of
299 * information about when the triangle extends past screen
300 * boundaries. See trimmed_box in lp_setup_bin_triangle().
302 bbox.x0 = MAX2(bbox.x0, 0);
303 bbox.y0 = MAX2(bbox.y0, 0);
305 tri = lp_setup_alloc_triangle(scene,
313 tri->v[0][0] = v0[0][0];
314 tri->v[1][0] = v1[0][0];
315 tri->v[2][0] = v2[0][0];
316 tri->v[0][1] = v0[0][1];
317 tri->v[1][1] = v1[0][1];
318 tri->v[2][1] = v2[0][1];
323 /* Setup parameter interpolants:
325 setup->setup.variant->jit_function( v0,
329 GET_A0(&tri->inputs),
330 GET_DADX(&tri->inputs),
331 GET_DADY(&tri->inputs) );
333 tri->inputs.frontfacing = frontfacing;
334 tri->inputs.disable = FALSE;
335 tri->inputs.opaque = setup->fs.current.variant->opaque;
338 lp_dump_setup_coef(&setup->setup.variant->key,
339 (const float (*)[4])GET_A0(&tri->inputs),
340 (const float (*)[4])GET_DADX(&tri->inputs),
341 (const float (*)[4])GET_DADY(&tri->inputs));
343 plane = GET_PLANES(tri);
345 #if defined(PIPE_ARCH_SSE)
347 __m128i vertx, verty;
348 __m128i shufx, shufy;
349 __m128i dcdx, dcdy, c;
351 __m128i dcdx_neg_mask;
352 __m128i dcdy_neg_mask;
353 __m128i dcdx_zero_mask;
354 __m128i top_left_flag;
355 __m128i c_inc_mask, c_inc;
356 __m128i eo, p0, p1, p2;
357 __m128i zero = _mm_setzero_si128();
359 vertx = _mm_loadu_si128((__m128i *)x); /* vertex x coords */
360 verty = _mm_loadu_si128((__m128i *)y); /* vertex y coords */
362 shufx = _mm_shuffle_epi32(vertx, _MM_SHUFFLE(3,0,2,1));
363 shufy = _mm_shuffle_epi32(verty, _MM_SHUFFLE(3,0,2,1));
365 dcdx = _mm_sub_epi32(verty, shufy);
366 dcdy = _mm_sub_epi32(vertx, shufx);
368 dcdx_neg_mask = _mm_srai_epi32(dcdx, 31);
369 dcdx_zero_mask = _mm_cmpeq_epi32(dcdx, zero);
370 dcdy_neg_mask = _mm_srai_epi32(dcdy, 31);
372 top_left_flag = _mm_set1_epi32((setup->pixel_offset == 0) ? ~0 : 0);
374 c_inc_mask = _mm_or_si128(dcdx_neg_mask,
375 _mm_and_si128(dcdx_zero_mask,
376 _mm_xor_si128(dcdy_neg_mask,
379 c_inc = _mm_srli_epi32(c_inc_mask, 31);
381 c = _mm_sub_epi32(mm_mullo_epi32(dcdx, vertx),
382 mm_mullo_epi32(dcdy, verty));
384 c = _mm_add_epi32(c, c_inc);
386 /* Scale up to match c:
388 dcdx = _mm_slli_epi32(dcdx, FIXED_ORDER);
389 dcdy = _mm_slli_epi32(dcdy, FIXED_ORDER);
391 /* Calculate trivial reject values:
393 eo = _mm_sub_epi32(_mm_andnot_si128(dcdy_neg_mask, dcdy),
394 _mm_and_si128(dcdx_neg_mask, dcdx));
396 /* ei = _mm_sub_epi32(_mm_sub_epi32(dcdy, dcdx), eo); */
398 /* Pointless transpose which gets undone immediately in
401 transpose4_epi32(&c, &dcdx, &dcdy, &eo,
402 &p0, &p1, &p2, &unused);
404 _mm_store_si128((__m128i *)&plane[0], p0);
405 _mm_store_si128((__m128i *)&plane[1], p1);
406 _mm_store_si128((__m128i *)&plane[2], p2);
411 plane[0].dcdy = x[0] - x[1];
412 plane[1].dcdy = x[1] - x[2];
413 plane[2].dcdy = x[2] - x[0];
414 plane[0].dcdx = y[0] - y[1];
415 plane[1].dcdx = y[1] - y[2];
416 plane[2].dcdx = y[2] - y[0];
418 for (i = 0; i < 3; i++) {
419 /* half-edge constants, will be interated over the whole render
422 plane[i].c = plane[i].dcdx * x[i] - plane[i].dcdy * y[i];
424 /* correct for top-left vs. bottom-left fill convention.
426 * note that we're overloading gl_rasterization_rules to mean
427 * both (0.5,0.5) pixel centers *and* bottom-left filling
430 * GL actually has a top-left filling convention, but GL's
431 * notion of "top" differs from gallium's...
433 * Also, sometimes (in FBO cases) GL will render upside down
434 * to its usual method, in which case it will probably want
435 * to use the opposite, top-left convention.
437 if (plane[i].dcdx < 0) {
438 /* both fill conventions want this - adjust for left edges */
441 else if (plane[i].dcdx == 0) {
442 if (setup->pixel_offset == 0) {
443 /* correct for top-left fill convention:
445 if (plane[i].dcdy > 0) plane[i].c++;
448 /* correct for bottom-left fill convention:
450 if (plane[i].dcdy < 0) plane[i].c++;
454 plane[i].dcdx *= FIXED_ONE;
455 plane[i].dcdy *= FIXED_ONE;
457 /* find trivial reject offsets for each edge for a single-pixel
458 * sized block. These will be scaled up at each recursive level to
459 * match the active blocksize. Scaling in this way works best if
460 * the blocks are square.
463 if (plane[i].dcdx < 0) plane[i].eo -= plane[i].dcdx;
464 if (plane[i].dcdy > 0) plane[i].eo += plane[i].dcdy;
470 debug_printf("p0: %08x/%08x/%08x/%08x\n",
476 debug_printf("p1: %08x/%08x/%08x/%08x\n",
482 debug_printf("p0: %08x/%08x/%08x/%08x\n",
491 * When rasterizing scissored tris, use the intersection of the
492 * triangle bounding box and the scissor rect to generate the
495 * This permits us to cut off the triangle "tails" that are present
496 * in the intermediate recursive levels caused when two of the
497 * triangles edges don't diverge quickly enough to trivially reject
498 * exterior blocks from the triangle.
500 * It's not really clear if it's worth worrying about these tails,
501 * but since we generate the planes for each scissored tri, it's
502 * free to trim them in this case.
504 * Note that otherwise, the scissor planes only vary in 'C' value,
505 * and even then only on state-changes. Could alternatively store
506 * these planes elsewhere.
508 if (nr_planes == 7) {
509 const struct u_rect *scissor = &setup->scissor;
513 plane[3].c = 1-scissor->x0;
518 plane[4].c = scissor->x1+1;
523 plane[5].c = 1-scissor->y0;
528 plane[6].c = scissor->y1+1;
532 return lp_setup_bin_triangle( setup, tri, &bbox, nr_planes );
536 * Round to nearest less or equal power of two of the input.
538 * Undefined if no bit set exists, so code should check against 0 first.
540 static INLINE uint32_t
541 floor_pot(uint32_t n)
543 #if defined(PIPE_CC_GCC) && defined(PIPE_ARCH_X86)
563 lp_setup_bin_triangle( struct lp_setup_context *setup,
564 struct lp_rast_triangle *tri,
565 const struct u_rect *bbox,
568 struct lp_scene *scene = setup->scene;
569 struct u_rect trimmed_box = *bbox;
572 /* What is the largest power-of-two boundary this triangle crosses:
574 int dx = floor_pot((bbox->x0 ^ bbox->x1) |
575 (bbox->y0 ^ bbox->y1));
577 /* The largest dimension of the rasterized area of the triangle
578 * (aligned to a 4x4 grid), rounded down to the nearest power of two:
580 int sz = floor_pot((bbox->x1 - (bbox->x0 & ~3)) |
581 (bbox->y1 - (bbox->y0 & ~3)));
583 /* Now apply scissor, etc to the bounding box. Could do this
584 * earlier, but it confuses the logic for tri-16 and would force
585 * the rasterizer to also respect scissor, etc, just for the rare
586 * cases where a small triangle extends beyond the scissor.
588 u_rect_find_intersection(&setup->draw_region, &trimmed_box);
590 /* Determine which tile(s) intersect the triangle's bounding box
594 int ix0 = bbox->x0 / TILE_SIZE;
595 int iy0 = bbox->y0 / TILE_SIZE;
596 int px = bbox->x0 & 63 & ~3;
597 int py = bbox->y0 & 63 & ~3;
598 int mask = px | (py << 8);
600 assert(iy0 == bbox->y1 / TILE_SIZE &&
601 ix0 == bbox->x1 / TILE_SIZE);
603 if (nr_planes == 3) {
606 /* Triangle is contained in a single 4x4 stamp:
608 return lp_scene_bin_cmd_with_state( scene, ix0, iy0,
610 LP_RAST_OP_TRIANGLE_3_4,
611 lp_rast_arg_triangle(tri, mask) );
616 /* Triangle is contained in a single 16x16 block:
618 return lp_scene_bin_cmd_with_state( scene, ix0, iy0,
620 LP_RAST_OP_TRIANGLE_3_16,
621 lp_rast_arg_triangle(tri, mask) );
624 else if (nr_planes == 4 && sz < 16)
626 return lp_scene_bin_cmd_with_state(scene, ix0, iy0,
628 LP_RAST_OP_TRIANGLE_4_16,
629 lp_rast_arg_triangle(tri, mask) );
633 /* Triangle is contained in a single tile:
635 return lp_scene_bin_cmd_with_state( scene, ix0, iy0, setup->fs.stored,
636 lp_rast_tri_tab[nr_planes],
637 lp_rast_arg_triangle(tri, (1<<nr_planes)-1) );
641 struct lp_rast_plane *plane = GET_PLANES(tri);
646 int xstep[MAX_PLANES];
647 int ystep[MAX_PLANES];
650 int ix0 = trimmed_box.x0 / TILE_SIZE;
651 int iy0 = trimmed_box.y0 / TILE_SIZE;
652 int ix1 = trimmed_box.x1 / TILE_SIZE;
653 int iy1 = trimmed_box.y1 / TILE_SIZE;
655 for (i = 0; i < nr_planes; i++) {
657 plane[i].dcdy * iy0 * TILE_SIZE -
658 plane[i].dcdx * ix0 * TILE_SIZE);
660 ei[i] = (plane[i].dcdy -
662 plane[i].eo) << TILE_ORDER;
664 eo[i] = plane[i].eo << TILE_ORDER;
665 xstep[i] = -(plane[i].dcdx << TILE_ORDER);
666 ystep[i] = plane[i].dcdy << TILE_ORDER;
671 /* Test tile-sized blocks against the triangle.
672 * Discard blocks fully outside the tri. If the block is fully
673 * contained inside the tri, bin an lp_rast_shade_tile command.
674 * Else, bin a lp_rast_triangle command.
676 for (y = iy0; y <= iy1; y++)
678 boolean in = FALSE; /* are we inside the triangle? */
681 for (i = 0; i < nr_planes; i++)
684 for (x = ix0; x <= ix1; x++)
689 for (i = 0; i < nr_planes; i++) {
690 int planeout = cx[i] + eo[i];
691 int planepartial = cx[i] + ei[i] - 1;
692 out |= (planeout >> 31);
693 partial |= (planepartial >> 31) & (1<<i);
699 break; /* exiting triangle, all done with this row */
700 LP_COUNT(nr_empty_64);
703 /* Not trivially accepted by at least one plane -
704 * rasterize/shade partial tile
706 int count = util_bitcount(partial);
709 if (!lp_scene_bin_cmd_with_state( scene, x, y,
711 lp_rast_tri_tab[count],
712 lp_rast_arg_triangle(tri, partial) ))
715 LP_COUNT(nr_partially_covered_64);
718 /* triangle covers the whole tile- shade whole tile */
719 LP_COUNT(nr_fully_covered_64);
721 if (!lp_setup_whole_tile(setup, &tri->inputs, x, y))
725 /* Iterate cx values across the region:
727 for (i = 0; i < nr_planes; i++)
731 /* Iterate c values down the region:
733 for (i = 0; i < nr_planes; i++)
741 /* Need to disable any partially binned triangle. This is easier
742 * than trying to locate all the triangle, shade-tile, etc,
743 * commands which may have been binned.
745 tri->inputs.disable = TRUE;
751 * Try to draw the triangle, restart the scene on failure.
753 static void retry_triangle_ccw( struct lp_setup_context *setup,
754 const float (*v0)[4],
755 const float (*v1)[4],
756 const float (*v2)[4],
759 if (!do_triangle_ccw( setup, v0, v1, v2, front ))
761 if (!lp_setup_flush_and_restart(setup))
764 if (!do_triangle_ccw( setup, v0, v1, v2, front ))
770 calc_area(const float (*v0)[4],
771 const float (*v1)[4],
772 const float (*v2)[4])
774 float dx01 = v0[0][0] - v1[0][0];
775 float dy01 = v0[0][1] - v1[0][1];
776 float dx20 = v2[0][0] - v0[0][0];
777 float dy20 = v2[0][1] - v0[0][1];
778 return dx01 * dy20 - dx20 * dy01;
783 * Draw triangle if it's CW, cull otherwise.
785 static void triangle_cw( struct lp_setup_context *setup,
786 const float (*v0)[4],
787 const float (*v1)[4],
788 const float (*v2)[4] )
790 float area = calc_area(v0, v1, v2);
793 retry_triangle_ccw(setup, v0, v2, v1, !setup->ccw_is_frontface);
797 static void triangle_ccw( struct lp_setup_context *setup,
798 const float (*v0)[4],
799 const float (*v1)[4],
800 const float (*v2)[4])
802 float area = calc_area(v0, v1, v2);
805 retry_triangle_ccw(setup, v0, v1, v2, setup->ccw_is_frontface);
809 * Draw triangle whether it's CW or CCW.
811 static void triangle_both( struct lp_setup_context *setup,
812 const float (*v0)[4],
813 const float (*v1)[4],
814 const float (*v2)[4] )
816 float area = calc_area(v0, v1, v2);
819 assert(!util_is_inf_or_nan(v0[0][0]));
820 assert(!util_is_inf_or_nan(v0[0][1]));
821 assert(!util_is_inf_or_nan(v1[0][0]));
822 assert(!util_is_inf_or_nan(v1[0][1]));
823 assert(!util_is_inf_or_nan(v2[0][0]));
824 assert(!util_is_inf_or_nan(v2[0][1]));
825 assert(!util_is_inf_or_nan(area));
829 retry_triangle_ccw( setup, v0, v1, v2, setup->ccw_is_frontface );
830 else if (area < 0.0f)
831 retry_triangle_ccw( setup, v0, v2, v1, !setup->ccw_is_frontface );
835 static void triangle_nop( struct lp_setup_context *setup,
836 const float (*v0)[4],
837 const float (*v1)[4],
838 const float (*v2)[4] )
844 lp_setup_choose_triangle( struct lp_setup_context *setup )
846 switch (setup->cullmode) {
848 setup->triangle = triangle_both;
851 setup->triangle = setup->ccw_is_frontface ? triangle_ccw : triangle_cw;
853 case PIPE_FACE_FRONT:
854 setup->triangle = setup->ccw_is_frontface ? triangle_cw : triangle_ccw;
857 setup->triangle = triangle_nop;