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"
36 #include "lp_setup_context.h"
37 #include "lp_setup_coef.h"
39 #include "lp_state_fs.h"
41 #define NUM_CHANNELS 4
46 subpixel_snap(float a
)
48 return util_iround(FIXED_ONE
* a
);
54 return a
* (1.0 / FIXED_ONE
);
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
71 struct lp_rast_triangle
*
72 lp_setup_alloc_triangle(struct lp_scene
*scene
,
77 unsigned input_array_sz
= NUM_CHANNELS
* (nr_inputs
+ 1) * sizeof(float);
78 struct lp_rast_triangle
*tri
;
79 unsigned tri_bytes
, bytes
;
82 tri_bytes
= align(Offset(struct lp_rast_triangle
, plane
[nr_planes
]), 16);
83 bytes
= tri_bytes
+ (3 * input_array_sz
);
85 tri
= lp_scene_alloc_aligned( scene
, bytes
, 16 );
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
);
100 lp_setup_print_vertex(struct lp_setup_context
*setup
,
106 debug_printf(" wpos (%s[0]) xyzw %f %f %f %f\n",
108 v
[0][0], v
[0][1], v
[0][2], v
[0][3]);
110 for (i
= 0; i
< setup
->fs
.nr_inputs
; i
++) {
111 const float *in
= v
[setup
->fs
.input
[i
].src_index
];
113 debug_printf(" in[%d] (%s[%d]) %s%s%s%s ",
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" : " ");
121 for (j
= 0; j
< 4; j
++)
122 if (setup
->fs
.input
[i
].usage_mask
& (1<<j
))
123 debug_printf("%.5f ", in
[j
]);
131 * Print triangle vertex attribs (for debug).
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])
139 debug_printf("triangle\n");
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];
147 /* det = cross(e,f).z */
148 const float det
= ex
* fy
- ey
* fx
;
150 debug_printf(" - ccw\n");
152 debug_printf(" - cw\n");
154 debug_printf(" - zero area\n");
157 lp_setup_print_vertex(setup
, "v0", v0
);
158 lp_setup_print_vertex(setup
, "v1", v1
);
159 lp_setup_print_vertex(setup
, "v2", v2
);
165 lp_rast_tri_tab
[MAX_PLANES
+1] = {
166 0, /* should be impossible */
167 LP_RAST_OP_TRIANGLE_1
,
168 LP_RAST_OP_TRIANGLE_2
,
169 LP_RAST_OP_TRIANGLE_3
,
170 LP_RAST_OP_TRIANGLE_4
,
171 LP_RAST_OP_TRIANGLE_5
,
172 LP_RAST_OP_TRIANGLE_6
,
173 LP_RAST_OP_TRIANGLE_7
,
174 LP_RAST_OP_TRIANGLE_8
180 * The primitive covers the whole tile- shade whole tile.
182 * \param tx, ty the tile position in tiles, not pixels
185 lp_setup_whole_tile(struct lp_setup_context
*setup
,
186 const struct lp_rast_shader_inputs
*inputs
,
189 struct lp_scene
*scene
= setup
->scene
;
191 LP_COUNT(nr_fully_covered_64
);
193 /* if variant is opaque and scissor doesn't effect the tile */
194 if (inputs
->opaque
) {
195 if (!scene
->fb
.zsbuf
) {
197 * All previous rendering will be overwritten so reset the bin.
199 lp_scene_bin_reset( scene
, tx
, ty
);
202 LP_COUNT(nr_shade_opaque_64
);
203 return lp_scene_bin_cmd_with_state( scene
, tx
, ty
,
205 LP_RAST_OP_SHADE_TILE_OPAQUE
,
206 lp_rast_arg_inputs(inputs
) );
208 LP_COUNT(nr_shade_64
);
209 return lp_scene_bin_cmd_with_state( scene
, tx
, ty
,
211 LP_RAST_OP_SHADE_TILE
,
212 lp_rast_arg_inputs(inputs
) );
218 * Do basic setup for triangle rasterization and determine which
219 * framebuffer tiles are touched. Put the triangle in the scene's
220 * bins for the tiles which we overlap.
223 do_triangle_ccw(struct lp_setup_context
*setup
,
224 const float (*v0
)[4],
225 const float (*v1
)[4],
226 const float (*v2
)[4],
227 boolean frontfacing
)
229 struct lp_scene
*scene
= setup
->scene
;
230 struct lp_rast_triangle
*tri
;
239 lp_setup_print_triangle(setup
, v0
, v1
, v2
);
241 if (setup
->scissor_test
) {
248 /* x/y positions in fixed point */
249 x
[0] = subpixel_snap(v0
[0][0] - setup
->pixel_offset
);
250 x
[1] = subpixel_snap(v1
[0][0] - setup
->pixel_offset
);
251 x
[2] = subpixel_snap(v2
[0][0] - setup
->pixel_offset
);
252 y
[0] = subpixel_snap(v0
[0][1] - setup
->pixel_offset
);
253 y
[1] = subpixel_snap(v1
[0][1] - setup
->pixel_offset
);
254 y
[2] = subpixel_snap(v2
[0][1] - setup
->pixel_offset
);
257 /* Bounding rectangle (in pixels) */
259 /* Yes this is necessary to accurately calculate bounding boxes
260 * with the two fill-conventions we support. GL (normally) ends
261 * up needing a bottom-left fill convention, which requires
262 * slightly different rounding.
264 int adj
= (setup
->pixel_offset
!= 0) ? 1 : 0;
266 bbox
.x0
= (MIN3(x
[0], x
[1], x
[2]) + (FIXED_ONE
-1)) >> FIXED_ORDER
;
267 bbox
.x1
= (MAX3(x
[0], x
[1], x
[2]) + (FIXED_ONE
-1)) >> FIXED_ORDER
;
268 bbox
.y0
= (MIN3(y
[0], y
[1], y
[2]) + (FIXED_ONE
-1) + adj
) >> FIXED_ORDER
;
269 bbox
.y1
= (MAX3(y
[0], y
[1], y
[2]) + (FIXED_ONE
-1) + adj
) >> FIXED_ORDER
;
271 /* Inclusive coordinates:
277 if (bbox
.x1
< bbox
.x0
||
279 if (0) debug_printf("empty bounding box\n");
280 LP_COUNT(nr_culled_tris
);
284 if (!u_rect_test_intersection(&setup
->draw_region
, &bbox
)) {
285 if (0) debug_printf("offscreen\n");
286 LP_COUNT(nr_culled_tris
);
290 u_rect_find_intersection(&setup
->draw_region
, &bbox
);
292 tri
= lp_setup_alloc_triangle(scene
,
300 tri
->v
[0][0] = v0
[0][0];
301 tri
->v
[1][0] = v1
[0][0];
302 tri
->v
[2][0] = v2
[0][0];
303 tri
->v
[0][1] = v0
[0][1];
304 tri
->v
[1][1] = v1
[0][1];
305 tri
->v
[2][1] = v2
[0][1];
308 tri
->plane
[0].dcdy
= x
[0] - x
[1];
309 tri
->plane
[1].dcdy
= x
[1] - x
[2];
310 tri
->plane
[2].dcdy
= x
[2] - x
[0];
312 tri
->plane
[0].dcdx
= y
[0] - y
[1];
313 tri
->plane
[1].dcdx
= y
[1] - y
[2];
314 tri
->plane
[2].dcdx
= y
[2] - y
[0];
318 /* Setup parameter interpolants:
320 lp_setup_tri_coef( setup
, &tri
->inputs
, v0
, v1
, v2
, frontfacing
);
322 tri
->inputs
.facing
= frontfacing
? 1.0F
: -1.0F
;
323 tri
->inputs
.disable
= FALSE
;
324 tri
->inputs
.opaque
= setup
->fs
.current
.variant
->opaque
;
327 for (i
= 0; i
< 3; i
++) {
328 struct lp_rast_plane
*plane
= &tri
->plane
[i
];
330 /* half-edge constants, will be interated over the whole render
333 plane
->c
= plane
->dcdx
* x
[i
] - plane
->dcdy
* y
[i
];
335 /* correct for top-left vs. bottom-left fill convention.
337 * note that we're overloading gl_rasterization_rules to mean
338 * both (0.5,0.5) pixel centers *and* bottom-left filling
341 * GL actually has a top-left filling convention, but GL's
342 * notion of "top" differs from gallium's...
344 * Also, sometimes (in FBO cases) GL will render upside down
345 * to its usual method, in which case it will probably want
346 * to use the opposite, top-left convention.
348 if (plane
->dcdx
< 0) {
349 /* both fill conventions want this - adjust for left edges */
352 else if (plane
->dcdx
== 0) {
353 if (setup
->pixel_offset
== 0) {
354 /* correct for top-left fill convention:
356 if (plane
->dcdy
> 0) plane
->c
++;
359 /* correct for bottom-left fill convention:
361 if (plane
->dcdy
< 0) plane
->c
++;
365 plane
->dcdx
*= FIXED_ONE
;
366 plane
->dcdy
*= FIXED_ONE
;
368 /* find trivial reject offsets for each edge for a single-pixel
369 * sized block. These will be scaled up at each recursive level to
370 * match the active blocksize. Scaling in this way works best if
371 * the blocks are square.
374 if (plane
->dcdx
< 0) plane
->eo
-= plane
->dcdx
;
375 if (plane
->dcdy
> 0) plane
->eo
+= plane
->dcdy
;
377 /* Calculate trivial accept offsets from the above.
379 plane
->ei
= plane
->dcdy
- plane
->dcdx
- plane
->eo
;
384 * When rasterizing scissored tris, use the intersection of the
385 * triangle bounding box and the scissor rect to generate the
388 * This permits us to cut off the triangle "tails" that are present
389 * in the intermediate recursive levels caused when two of the
390 * triangles edges don't diverge quickly enough to trivially reject
391 * exterior blocks from the triangle.
393 * It's not really clear if it's worth worrying about these tails,
394 * but since we generate the planes for each scissored tri, it's
395 * free to trim them in this case.
397 * Note that otherwise, the scissor planes only vary in 'C' value,
398 * and even then only on state-changes. Could alternatively store
399 * these planes elsewhere.
401 if (nr_planes
== 7) {
402 tri
->plane
[3].dcdx
= -1;
403 tri
->plane
[3].dcdy
= 0;
404 tri
->plane
[3].c
= 1-bbox
.x0
;
405 tri
->plane
[3].ei
= 0;
406 tri
->plane
[3].eo
= 1;
408 tri
->plane
[4].dcdx
= 1;
409 tri
->plane
[4].dcdy
= 0;
410 tri
->plane
[4].c
= bbox
.x1
+1;
411 tri
->plane
[4].ei
= -1;
412 tri
->plane
[4].eo
= 0;
414 tri
->plane
[5].dcdx
= 0;
415 tri
->plane
[5].dcdy
= 1;
416 tri
->plane
[5].c
= 1-bbox
.y0
;
417 tri
->plane
[5].ei
= 0;
418 tri
->plane
[5].eo
= 1;
420 tri
->plane
[6].dcdx
= 0;
421 tri
->plane
[6].dcdy
= -1;
422 tri
->plane
[6].c
= bbox
.y1
+1;
423 tri
->plane
[6].ei
= -1;
424 tri
->plane
[6].eo
= 0;
427 return lp_setup_bin_triangle( setup
, tri
, &bbox
, nr_planes
);
431 * Round to nearest less or equal power of two of the input.
433 * Undefined if no bit set exists, so code should check against 0 first.
435 static INLINE
uint32_t
436 floor_pot(uint32_t n
)
438 #if defined(PIPE_CC_GCC) && defined(PIPE_ARCH_X86)
458 lp_setup_bin_triangle( struct lp_setup_context
*setup
,
459 struct lp_rast_triangle
*tri
,
460 const struct u_rect
*bbox
,
463 struct lp_scene
*scene
= setup
->scene
;
466 /* What is the largest power-of-two boundary this triangle crosses:
468 int dx
= floor_pot((bbox
->x0
^ bbox
->x1
) |
469 (bbox
->y0
^ bbox
->y1
));
471 /* The largest dimension of the rasterized area of the triangle
472 * (aligned to a 4x4 grid), rounded down to the nearest power of two:
474 int sz
= floor_pot((bbox
->x1
- (bbox
->x0
& ~3)) |
475 (bbox
->y1
- (bbox
->y0
& ~3)));
477 /* Determine which tile(s) intersect the triangle's bounding box
481 int ix0
= bbox
->x0
/ TILE_SIZE
;
482 int iy0
= bbox
->y0
/ TILE_SIZE
;
483 int px
= bbox
->x0
& 63 & ~3;
484 int py
= bbox
->y0
& 63 & ~3;
485 int mask
= px
| (py
<< 8);
487 assert(iy0
== bbox
->y1
/ TILE_SIZE
&&
488 ix0
== bbox
->x1
/ TILE_SIZE
);
490 if (nr_planes
== 3) {
493 /* Triangle is contained in a single 4x4 stamp:
495 return lp_scene_bin_cmd_with_state( scene
, ix0
, iy0
,
497 LP_RAST_OP_TRIANGLE_3_4
,
498 lp_rast_arg_triangle(tri
, mask
) );
503 /* Triangle is contained in a single 16x16 block:
505 return lp_scene_bin_cmd_with_state( scene
, ix0
, iy0
,
507 LP_RAST_OP_TRIANGLE_3_16
,
508 lp_rast_arg_triangle(tri
, mask
) );
511 else if (nr_planes
== 4 && sz
< 16)
513 return lp_scene_bin_cmd_with_state(scene
, ix0
, iy0
,
515 LP_RAST_OP_TRIANGLE_4_16
,
516 lp_rast_arg_triangle(tri
, mask
) );
520 /* Triangle is contained in a single tile:
522 return lp_scene_bin_cmd_with_state( scene
, ix0
, iy0
, setup
->fs
.stored
,
523 lp_rast_tri_tab
[nr_planes
],
524 lp_rast_arg_triangle(tri
, (1<<nr_planes
)-1) );
531 int xstep
[MAX_PLANES
];
532 int ystep
[MAX_PLANES
];
535 int ix0
= bbox
->x0
/ TILE_SIZE
;
536 int iy0
= bbox
->y0
/ TILE_SIZE
;
537 int ix1
= bbox
->x1
/ TILE_SIZE
;
538 int iy1
= bbox
->y1
/ TILE_SIZE
;
540 for (i
= 0; i
< nr_planes
; i
++) {
541 c
[i
] = (tri
->plane
[i
].c
+
542 tri
->plane
[i
].dcdy
* iy0
* TILE_SIZE
-
543 tri
->plane
[i
].dcdx
* ix0
* TILE_SIZE
);
545 ei
[i
] = tri
->plane
[i
].ei
<< TILE_ORDER
;
546 eo
[i
] = tri
->plane
[i
].eo
<< TILE_ORDER
;
547 xstep
[i
] = -(tri
->plane
[i
].dcdx
<< TILE_ORDER
);
548 ystep
[i
] = tri
->plane
[i
].dcdy
<< TILE_ORDER
;
553 /* Test tile-sized blocks against the triangle.
554 * Discard blocks fully outside the tri. If the block is fully
555 * contained inside the tri, bin an lp_rast_shade_tile command.
556 * Else, bin a lp_rast_triangle command.
558 for (y
= iy0
; y
<= iy1
; y
++)
560 boolean in
= FALSE
; /* are we inside the triangle? */
563 for (i
= 0; i
< nr_planes
; i
++)
566 for (x
= ix0
; x
<= ix1
; x
++)
571 for (i
= 0; i
< nr_planes
; i
++) {
572 int planeout
= cx
[i
] + eo
[i
];
573 int planepartial
= cx
[i
] + ei
[i
] - 1;
574 out
|= (planeout
>> 31);
575 partial
|= (planepartial
>> 31) & (1<<i
);
581 break; /* exiting triangle, all done with this row */
582 LP_COUNT(nr_empty_64
);
585 /* Not trivially accepted by at least one plane -
586 * rasterize/shade partial tile
588 int count
= util_bitcount(partial
);
591 if (!lp_scene_bin_cmd_with_state( scene
, x
, y
,
593 lp_rast_tri_tab
[count
],
594 lp_rast_arg_triangle(tri
, partial
) ))
597 LP_COUNT(nr_partially_covered_64
);
600 /* triangle covers the whole tile- shade whole tile */
601 LP_COUNT(nr_fully_covered_64
);
603 if (!lp_setup_whole_tile(setup
, &tri
->inputs
, x
, y
))
607 /* Iterate cx values across the region:
609 for (i
= 0; i
< nr_planes
; i
++)
613 /* Iterate c values down the region:
615 for (i
= 0; i
< nr_planes
; i
++)
623 /* Need to disable any partially binned triangle. This is easier
624 * than trying to locate all the triangle, shade-tile, etc,
625 * commands which may have been binned.
627 tri
->inputs
.disable
= TRUE
;
633 * Try to draw the triangle, restart the scene on failure.
635 static void retry_triangle_ccw( struct lp_setup_context
*setup
,
636 const float (*v0
)[4],
637 const float (*v1
)[4],
638 const float (*v2
)[4],
641 if (!do_triangle_ccw( setup
, v0
, v1
, v2
, front
))
643 if (!lp_setup_flush_and_restart(setup
))
646 if (!do_triangle_ccw( setup
, v0
, v1
, v2
, front
))
652 calc_area(const float (*v0
)[4],
653 const float (*v1
)[4],
654 const float (*v2
)[4])
656 float dx01
= v0
[0][0] - v1
[0][0];
657 float dy01
= v0
[0][1] - v1
[0][1];
658 float dx20
= v2
[0][0] - v0
[0][0];
659 float dy20
= v2
[0][1] - v0
[0][1];
660 return dx01
* dy20
- dx20
* dy01
;
665 * Draw triangle if it's CW, cull otherwise.
667 static void triangle_cw( struct lp_setup_context
*setup
,
668 const float (*v0
)[4],
669 const float (*v1
)[4],
670 const float (*v2
)[4] )
672 float area
= calc_area(v0
, v1
, v2
);
675 retry_triangle_ccw(setup
, v0
, v2
, v1
, !setup
->ccw_is_frontface
);
679 static void triangle_ccw( struct lp_setup_context
*setup
,
680 const float (*v0
)[4],
681 const float (*v1
)[4],
682 const float (*v2
)[4])
684 float area
= calc_area(v0
, v1
, v2
);
687 retry_triangle_ccw(setup
, v0
, v1
, v2
, setup
->ccw_is_frontface
);
691 * Draw triangle whether it's CW or CCW.
693 static void triangle_both( struct lp_setup_context
*setup
,
694 const float (*v0
)[4],
695 const float (*v1
)[4],
696 const float (*v2
)[4] )
698 float area
= calc_area(v0
, v1
, v2
);
701 retry_triangle_ccw( setup
, v0
, v1
, v2
, setup
->ccw_is_frontface
);
702 else if (area
< 0.0f
)
703 retry_triangle_ccw( setup
, v0
, v2
, v1
, !setup
->ccw_is_frontface
);
707 static void triangle_nop( struct lp_setup_context
*setup
,
708 const float (*v0
)[4],
709 const float (*v1
)[4],
710 const float (*v2
)[4] )
716 lp_setup_choose_triangle( struct lp_setup_context
*setup
)
718 switch (setup
->cullmode
) {
720 setup
->triangle
= triangle_both
;
723 setup
->triangle
= setup
->ccw_is_frontface
? triangle_ccw
: triangle_cw
;
725 case PIPE_FACE_FRONT
:
726 setup
->triangle
= setup
->ccw_is_frontface
? triangle_cw
: triangle_ccw
;
729 setup
->triangle
= triangle_nop
;