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"
35 #include "lp_setup_context.h"
38 #define NUM_CHANNELS 4
42 * Compute a0 for a constant-valued coefficient (GL_FLAT shading).
44 static void constant_coef( struct lp_setup_context
*setup
,
45 struct lp_rast_triangle
*tri
,
50 tri
->inputs
.a0
[slot
][i
] = value
;
51 tri
->inputs
.dadx
[slot
][i
] = 0.0f
;
52 tri
->inputs
.dady
[slot
][i
] = 0.0f
;
57 * Compute a0, dadx and dady for a linearly interpolated coefficient,
60 static void linear_coef( struct lp_setup_context
*setup
,
61 struct lp_rast_triangle
*tri
,
70 float a1
= v1
[vert_attr
][i
];
71 float a2
= v2
[vert_attr
][i
];
72 float a3
= v3
[vert_attr
][i
];
76 float dadx
= (da12
* tri
->dy31
- tri
->dy12
* da31
) * oneoverarea
;
77 float dady
= (da31
* tri
->dx12
- tri
->dx31
* da12
) * oneoverarea
;
79 tri
->inputs
.dadx
[slot
][i
] = dadx
;
80 tri
->inputs
.dady
[slot
][i
] = dady
;
82 /* calculate a0 as the value which would be sampled for the
83 * fragment at (0,0), taking into account that we want to sample at
84 * pixel centers, in other words (0.5, 0.5).
86 * this is neat but unfortunately not a good way to do things for
87 * triangles with very large values of dadx or dady as it will
88 * result in the subtraction and re-addition from a0 of a very
89 * large number, which means we'll end up loosing a lot of the
90 * fractional bits and precision from a0. the way to fix this is
91 * to define a0 as the sample at a pixel center somewhere near vmin
92 * instead - i'll switch to this later.
94 tri
->inputs
.a0
[slot
][i
] = (a1
-
95 (dadx
* (v1
[0][0] - setup
->pixel_offset
) +
96 dady
* (v1
[0][1] - setup
->pixel_offset
)));
101 * Compute a0, dadx and dady for a perspective-corrected interpolant,
103 * We basically multiply the vertex value by 1/w before computing
104 * the plane coefficients (a0, dadx, dady).
105 * Later, when we compute the value at a particular fragment position we'll
106 * divide the interpolated value by the interpolated W at that fragment.
108 static void perspective_coef( struct lp_setup_context
*setup
,
109 struct lp_rast_triangle
*tri
,
112 const float (*v1
)[4],
113 const float (*v2
)[4],
114 const float (*v3
)[4],
118 /* premultiply by 1/w (v[0][3] is always 1/w):
120 float a1
= v1
[vert_attr
][i
] * v1
[0][3];
121 float a2
= v2
[vert_attr
][i
] * v2
[0][3];
122 float a3
= v3
[vert_attr
][i
] * v3
[0][3];
123 float da12
= a1
- a2
;
124 float da31
= a3
- a1
;
125 float dadx
= (da12
* tri
->dy31
- tri
->dy12
* da31
) * oneoverarea
;
126 float dady
= (da31
* tri
->dx12
- tri
->dx31
* da12
) * oneoverarea
;
128 tri
->inputs
.dadx
[slot
][i
] = dadx
;
129 tri
->inputs
.dady
[slot
][i
] = dady
;
130 tri
->inputs
.a0
[slot
][i
] = (a1
-
131 (dadx
* (v1
[0][0] - setup
->pixel_offset
) +
132 dady
* (v1
[0][1] - setup
->pixel_offset
)));
137 * Special coefficient setup for gl_FragCoord.
138 * X and Y are trivial
139 * Z and W are copied from position_coef which should have already been computed.
140 * We could do a bit less work if we'd examine gl_FragCoord's swizzle mask.
143 setup_fragcoord_coef(struct lp_setup_context
*setup
,
144 struct lp_rast_triangle
*tri
,
147 const float (*v1
)[4],
148 const float (*v2
)[4],
149 const float (*v3
)[4])
152 tri
->inputs
.a0
[slot
][0] = 0.0;
153 tri
->inputs
.dadx
[slot
][0] = 1.0;
154 tri
->inputs
.dady
[slot
][0] = 0.0;
156 tri
->inputs
.a0
[slot
][1] = 0.0;
157 tri
->inputs
.dadx
[slot
][1] = 0.0;
158 tri
->inputs
.dady
[slot
][1] = 1.0;
160 linear_coef(setup
, tri
, oneoverarea
, slot
, v1
, v2
, v3
, 0, 2);
162 linear_coef(setup
, tri
, oneoverarea
, slot
, v1
, v2
, v3
, 0, 3);
167 * Setup the fragment input attribute with the front-facing value.
168 * \param frontface is the triangle front facing?
170 static void setup_facing_coef( struct lp_setup_context
*setup
,
171 struct lp_rast_triangle
*tri
,
175 /* convert TRUE to 1.0 and FALSE to -1.0 */
176 constant_coef( setup
, tri
, slot
, 2.0f
* frontface
- 1.0f
, 0 );
177 constant_coef( setup
, tri
, slot
, 0.0f
, 1 ); /* wasted */
178 constant_coef( setup
, tri
, slot
, 0.0f
, 2 ); /* wasted */
179 constant_coef( setup
, tri
, slot
, 0.0f
, 3 ); /* wasted */
184 * Compute the tri->coef[] array dadx, dady, a0 values.
186 static void setup_tri_coefficients( struct lp_setup_context
*setup
,
187 struct lp_rast_triangle
*tri
,
189 const float (*v1
)[4],
190 const float (*v2
)[4],
191 const float (*v3
)[4],
196 /* The internal position input is in slot zero:
198 setup_fragcoord_coef(setup
, tri
, oneoverarea
, 0, v1
, v2
, v3
);
200 /* setup interpolation for all the remaining attributes:
202 for (slot
= 0; slot
< setup
->fs
.nr_inputs
; slot
++) {
203 unsigned vert_attr
= setup
->fs
.input
[slot
].src_index
;
206 switch (setup
->fs
.input
[slot
].interp
) {
207 case LP_INTERP_CONSTANT
:
208 for (i
= 0; i
< NUM_CHANNELS
; i
++)
209 constant_coef(setup
, tri
, slot
+1, v3
[vert_attr
][i
], i
);
212 case LP_INTERP_LINEAR
:
213 for (i
= 0; i
< NUM_CHANNELS
; i
++)
214 linear_coef(setup
, tri
, oneoverarea
, slot
+1, v1
, v2
, v3
, vert_attr
, i
);
217 case LP_INTERP_PERSPECTIVE
:
218 for (i
= 0; i
< NUM_CHANNELS
; i
++)
219 perspective_coef(setup
, tri
, oneoverarea
, slot
+1, v1
, v2
, v3
, vert_attr
, i
);
222 case LP_INTERP_POSITION
:
223 /* XXX: fix me - duplicates the values in slot zero.
225 setup_fragcoord_coef(setup
, tri
, oneoverarea
, slot
+1, v1
, v2
, v3
);
228 case LP_INTERP_FACING
:
229 setup_facing_coef(setup
, tri
, slot
+1, frontface
);
240 static INLINE
int subpixel_snap( float a
)
242 return util_iround(FIXED_ONE
* a
- (FIXED_ONE
/ 2));
248 * Alloc space for a new triangle plus the input.a0/dadx/dady arrays
249 * immediately after it.
250 * The memory is allocated from the per-scene pool, not per-tile.
251 * \param tri_size returns number of bytes allocated
252 * \param nr_inputs number of fragment shader inputs
253 * \return pointer to triangle space
255 static INLINE
struct lp_rast_triangle
*
256 alloc_triangle(struct lp_scene
*scene
, unsigned nr_inputs
, unsigned *tri_size
)
258 unsigned input_array_sz
= NUM_CHANNELS
* (nr_inputs
+ 1) * sizeof(float);
259 struct lp_rast_triangle
*tri
;
263 assert(sizeof(*tri
) % 16 == 0);
265 bytes
= sizeof(*tri
) + (3 * input_array_sz
);
267 tri
= lp_scene_alloc_aligned( scene
, bytes
, 16 );
269 inputs
= (char *) (tri
+ 1);
270 tri
->inputs
.a0
= (float (*)[4]) inputs
;
271 tri
->inputs
.dadx
= (float (*)[4]) (inputs
+ input_array_sz
);
272 tri
->inputs
.dady
= (float (*)[4]) (inputs
+ 2 * input_array_sz
);
281 * Print triangle vertex attribs (for debug).
284 print_triangle(struct lp_setup_context
*setup
,
285 const float (*v1
)[4],
286 const float (*v2
)[4],
287 const float (*v3
)[4])
291 debug_printf("llvmpipe triangle\n");
292 for (i
= 0; i
< setup
->fs
.nr_inputs
; i
++) {
293 debug_printf(" v1[%d]: %f %f %f %f\n", i
,
294 v1
[i
][0], v1
[i
][1], v1
[i
][2], v1
[i
][3]);
296 for (i
= 0; i
< setup
->fs
.nr_inputs
; i
++) {
297 debug_printf(" v2[%d]: %f %f %f %f\n", i
,
298 v2
[i
][0], v2
[i
][1], v2
[i
][2], v2
[i
][3]);
300 for (i
= 0; i
< setup
->fs
.nr_inputs
; i
++) {
301 debug_printf(" v3[%d]: %f %f %f %f\n", i
,
302 v3
[i
][0], v3
[i
][1], v3
[i
][2], v3
[i
][3]);
308 * Do basic setup for triangle rasterization and determine which
309 * framebuffer tiles are touched. Put the triangle in the scene's
310 * bins for the tiles which we overlap.
313 do_triangle_ccw(struct lp_setup_context
*setup
,
314 const float (*v1
)[4],
315 const float (*v2
)[4],
316 const float (*v3
)[4],
317 boolean frontfacing
)
319 /* x/y positions in fixed point */
320 const int x1
= subpixel_snap(v1
[0][0] + 0.5 - setup
->pixel_offset
);
321 const int x2
= subpixel_snap(v2
[0][0] + 0.5 - setup
->pixel_offset
);
322 const int x3
= subpixel_snap(v3
[0][0] + 0.5 - setup
->pixel_offset
);
323 const int y1
= subpixel_snap(v1
[0][1] + 0.5 - setup
->pixel_offset
);
324 const int y2
= subpixel_snap(v2
[0][1] + 0.5 - setup
->pixel_offset
);
325 const int y3
= subpixel_snap(v3
[0][1] + 0.5 - setup
->pixel_offset
);
327 struct lp_scene
*scene
= lp_setup_get_current_scene(setup
);
328 struct lp_rast_triangle
*tri
;
331 int minx
, maxx
, miny
, maxy
;
335 print_triangle(setup
, v1
, v2
, v3
);
337 tri
= alloc_triangle(scene
, setup
->fs
.nr_inputs
, &tri_bytes
);
340 tri
->v
[0][0] = v1
[0][0];
341 tri
->v
[1][0] = v2
[0][0];
342 tri
->v
[2][0] = v3
[0][0];
343 tri
->v
[0][1] = v1
[0][1];
344 tri
->v
[1][1] = v2
[0][1];
345 tri
->v
[2][1] = v3
[0][1];
356 area
= (tri
->dx12
* tri
->dy31
- tri
->dx31
* tri
->dy12
);
360 /* Cull non-ccw and zero-sized triangles.
362 * XXX: subject to overflow??
365 lp_scene_putback_data( scene
, tri_bytes
);
366 LP_COUNT(nr_culled_tris
);
370 /* Bounding rectangle (in pixels) */
371 minx
= (MIN3(x1
, x2
, x3
) + (FIXED_ONE
-1)) >> FIXED_ORDER
;
372 maxx
= (MAX3(x1
, x2
, x3
) + (FIXED_ONE
-1)) >> FIXED_ORDER
;
373 miny
= (MIN3(y1
, y2
, y3
) + (FIXED_ONE
-1)) >> FIXED_ORDER
;
374 maxy
= (MAX3(y1
, y2
, y3
) + (FIXED_ONE
-1)) >> FIXED_ORDER
;
376 if (setup
->scissor_test
) {
377 minx
= MAX2(minx
, setup
->scissor
.current
.minx
);
378 maxx
= MIN2(maxx
, setup
->scissor
.current
.maxx
);
379 miny
= MAX2(miny
, setup
->scissor
.current
.miny
);
380 maxy
= MIN2(maxy
, setup
->scissor
.current
.maxy
);
385 lp_scene_putback_data( scene
, tri_bytes
);
386 LP_COUNT(nr_culled_tris
);
392 oneoverarea
= ((float)FIXED_ONE
) / (float)area
;
394 /* Setup parameter interpolants:
396 setup_tri_coefficients( setup
, tri
, oneoverarea
, v1
, v2
, v3
, frontfacing
);
398 tri
->inputs
.facing
= frontfacing
? 1.0F
: -1.0F
;
400 /* half-edge constants, will be interated over the whole render target.
402 tri
->c1
= tri
->dy12
* x1
- tri
->dx12
* y1
;
403 tri
->c2
= tri
->dy23
* x2
- tri
->dx23
* y2
;
404 tri
->c3
= tri
->dy31
* x3
- tri
->dx31
* y3
;
406 /* correct for top-left fill convention:
408 if (tri
->dy12
< 0 || (tri
->dy12
== 0 && tri
->dx12
> 0)) tri
->c1
++;
409 if (tri
->dy23
< 0 || (tri
->dy23
== 0 && tri
->dx23
> 0)) tri
->c2
++;
410 if (tri
->dy31
< 0 || (tri
->dy31
== 0 && tri
->dx31
> 0)) tri
->c3
++;
412 tri
->dy12
*= FIXED_ONE
;
413 tri
->dy23
*= FIXED_ONE
;
414 tri
->dy31
*= FIXED_ONE
;
416 tri
->dx12
*= FIXED_ONE
;
417 tri
->dx23
*= FIXED_ONE
;
418 tri
->dx31
*= FIXED_ONE
;
420 /* find trivial reject offsets for each edge for a single-pixel
421 * sized block. These will be scaled up at each recursive level to
422 * match the active blocksize. Scaling in this way works best if
423 * the blocks are square.
426 if (tri
->dy12
< 0) tri
->eo1
-= tri
->dy12
;
427 if (tri
->dx12
> 0) tri
->eo1
+= tri
->dx12
;
430 if (tri
->dy23
< 0) tri
->eo2
-= tri
->dy23
;
431 if (tri
->dx23
> 0) tri
->eo2
+= tri
->dx23
;
434 if (tri
->dy31
< 0) tri
->eo3
-= tri
->dy31
;
435 if (tri
->dx31
> 0) tri
->eo3
+= tri
->dx31
;
437 /* Calculate trivial accept offsets from the above.
439 tri
->ei1
= tri
->dx12
- tri
->dy12
- tri
->eo1
;
440 tri
->ei2
= tri
->dx23
- tri
->dy23
- tri
->eo2
;
441 tri
->ei3
= tri
->dx31
- tri
->dy31
- tri
->eo3
;
443 /* Fill in the inputs.step[][] arrays.
444 * We've manually unrolled some loops here.
447 const int xstep1
= -tri
->dy12
;
448 const int xstep2
= -tri
->dy23
;
449 const int xstep3
= -tri
->dy31
;
450 const int ystep1
= tri
->dx12
;
451 const int ystep2
= tri
->dx23
;
452 const int ystep3
= tri
->dx31
;
454 #define SETUP_STEP(i, x, y) \
456 tri->inputs.step[0][i] = x * xstep1 + y * ystep1; \
457 tri->inputs.step[1][i] = x * xstep2 + y * ystep2; \
458 tri->inputs.step[2][i] = x * xstep3 + y * ystep3; \
473 SETUP_STEP(10, 0, 3);
474 SETUP_STEP(11, 1, 3);
476 SETUP_STEP(12, 2, 2);
477 SETUP_STEP(13, 3, 2);
478 SETUP_STEP(14, 2, 3);
479 SETUP_STEP(15, 3, 3);
484 * All fields of 'tri' are now set. The remaining code here is
485 * concerned with binning.
488 /* Convert to tile coordinates:
490 minx
= minx
/ TILE_SIZE
;
491 miny
= miny
/ TILE_SIZE
;
492 maxx
= maxx
/ TILE_SIZE
;
493 maxy
= maxy
/ TILE_SIZE
;
496 * Clamp to framebuffer size
498 minx
= MAX2(minx
, 0);
499 miny
= MAX2(miny
, 0);
500 maxx
= MIN2(maxx
, scene
->tiles_x
- 1);
501 maxy
= MIN2(maxy
, scene
->tiles_y
- 1);
503 /* Determine which tile(s) intersect the triangle's bounding box
505 if (miny
== maxy
&& minx
== maxx
)
507 /* Triangle is contained in a single tile:
509 lp_scene_bin_command( scene
, minx
, miny
, lp_rast_triangle
,
510 lp_rast_arg_triangle(tri
) );
515 tri
->dx12
* miny
* TILE_SIZE
-
516 tri
->dy12
* minx
* TILE_SIZE
);
518 tri
->dx23
* miny
* TILE_SIZE
-
519 tri
->dy23
* minx
* TILE_SIZE
);
521 tri
->dx31
* miny
* TILE_SIZE
-
522 tri
->dy31
* minx
* TILE_SIZE
);
524 int ei1
= tri
->ei1
<< TILE_ORDER
;
525 int ei2
= tri
->ei2
<< TILE_ORDER
;
526 int ei3
= tri
->ei3
<< TILE_ORDER
;
528 int eo1
= tri
->eo1
<< TILE_ORDER
;
529 int eo2
= tri
->eo2
<< TILE_ORDER
;
530 int eo3
= tri
->eo3
<< TILE_ORDER
;
532 int xstep1
= -(tri
->dy12
<< TILE_ORDER
);
533 int xstep2
= -(tri
->dy23
<< TILE_ORDER
);
534 int xstep3
= -(tri
->dy31
<< TILE_ORDER
);
536 int ystep1
= tri
->dx12
<< TILE_ORDER
;
537 int ystep2
= tri
->dx23
<< TILE_ORDER
;
538 int ystep3
= tri
->dx31
<< TILE_ORDER
;
542 /* Test tile-sized blocks against the triangle.
543 * Discard blocks fully outside the tri. If the block is fully
544 * contained inside the tri, bin an lp_rast_shade_tile command.
545 * Else, bin a lp_rast_triangle command.
547 for (y
= miny
; y
<= maxy
; y
++)
552 boolean in
= FALSE
; /* are we inside the triangle? */
554 for (x
= minx
; x
<= maxx
; x
++)
561 LP_COUNT(nr_empty_64
);
563 break; /* exiting triangle, all done with this row */
565 else if (cx1
+ ei1
> 0 &&
569 /* triangle covers the whole tile- shade whole tile */
570 LP_COUNT(nr_fully_covered_64
);
572 if(setup
->fs
.current
.opaque
) {
573 lp_scene_bin_reset( scene
, x
, y
);
574 lp_scene_bin_command( scene
, x
, y
,
576 lp_rast_arg_state(setup
->fs
.stored
) );
578 lp_scene_bin_command( scene
, x
, y
,
580 lp_rast_arg_inputs(&tri
->inputs
) );
584 /* rasterizer/shade partial tile */
585 LP_COUNT(nr_partially_covered_64
);
587 lp_scene_bin_command( scene
, x
, y
,
589 lp_rast_arg_triangle(tri
) );
592 /* Iterate cx values across the region:
599 /* Iterate c values down the region:
610 * Draw triangle if it's CW, cull otherwise.
612 static void triangle_cw( struct lp_setup_context
*setup
,
613 const float (*v0
)[4],
614 const float (*v1
)[4],
615 const float (*v2
)[4] )
617 do_triangle_ccw( setup
, v1
, v0
, v2
, !setup
->ccw_is_frontface
);
622 * Draw triangle if it's CCW, cull otherwise.
624 static void triangle_ccw( struct lp_setup_context
*setup
,
625 const float (*v0
)[4],
626 const float (*v1
)[4],
627 const float (*v2
)[4] )
629 do_triangle_ccw( setup
, v0
, v1
, v2
, setup
->ccw_is_frontface
);
635 * Draw triangle whether it's CW or CCW.
637 static void triangle_both( struct lp_setup_context
*setup
,
638 const float (*v0
)[4],
639 const float (*v1
)[4],
640 const float (*v2
)[4] )
642 /* edge vectors e = v0 - v2, f = v1 - v2 */
643 const float ex
= v0
[0][0] - v2
[0][0];
644 const float ey
= v0
[0][1] - v2
[0][1];
645 const float fx
= v1
[0][0] - v2
[0][0];
646 const float fy
= v1
[0][1] - v2
[0][1];
648 /* det = cross(e,f).z */
649 if (ex
* fy
- ey
* fx
< 0.0f
)
650 triangle_ccw( setup
, v0
, v1
, v2
);
652 triangle_cw( setup
, v0
, v1
, v2
);
656 static void triangle_nop( struct lp_setup_context
*setup
,
657 const float (*v0
)[4],
658 const float (*v1
)[4],
659 const float (*v2
)[4] )
665 lp_setup_choose_triangle( struct lp_setup_context
*setup
)
667 switch (setup
->cullmode
) {
668 case PIPE_WINDING_NONE
:
669 setup
->triangle
= triangle_both
;
671 case PIPE_WINDING_CCW
:
672 setup
->triangle
= triangle_cw
;
674 case PIPE_WINDING_CW
:
675 setup
->triangle
= triangle_ccw
;
678 setup
->triangle
= triangle_nop
;