1 /**************************************************************************
3 * Copyright 2007 VMware, Inc.
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 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.
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"
41 #include "lp_context.h"
45 #define NUM_CHANNELS 4
47 #if defined(PIPE_ARCH_SSE)
48 #include <emmintrin.h>
49 #elif defined(_ARCH_PWR8) && defined(PIPE_ARCH_LITTLE_ENDIAN)
51 #include "util/u_pwr8.h"
55 subpixel_snap(float a
)
57 return util_iround(FIXED_ONE
* a
);
63 return a
* (1.0f
/ FIXED_ONE
);
67 /* Position and area in fixed point coordinates */
68 struct fixed_position
{
80 * Alloc space for a new triangle plus the input.a0/dadx/dady arrays
81 * immediately after it.
82 * The memory is allocated from the per-scene pool, not per-tile.
83 * \param tri_size returns number of bytes allocated
84 * \param num_inputs number of fragment shader inputs
85 * \return pointer to triangle space
87 struct lp_rast_triangle
*
88 lp_setup_alloc_triangle(struct lp_scene
*scene
,
93 unsigned input_array_sz
= NUM_CHANNELS
* (nr_inputs
+ 1) * sizeof(float);
94 unsigned plane_sz
= nr_planes
* sizeof(struct lp_rast_plane
);
95 struct lp_rast_triangle
*tri
;
97 *tri_size
= (sizeof(struct lp_rast_triangle
) +
101 tri
= lp_scene_alloc_aligned( scene
, *tri_size
, 16 );
105 tri
->inputs
.stride
= input_array_sz
;
108 char *a
= (char *)tri
;
109 char *b
= (char *)&GET_PLANES(tri
)[nr_planes
];
110 assert(b
- a
== *tri_size
);
117 lp_setup_print_vertex(struct lp_setup_context
*setup
,
121 const struct lp_setup_variant_key
*key
= &setup
->setup
.variant
->key
;
124 debug_printf(" wpos (%s[0]) xyzw %f %f %f %f\n",
126 v
[0][0], v
[0][1], v
[0][2], v
[0][3]);
128 for (i
= 0; i
< key
->num_inputs
; i
++) {
129 const float *in
= v
[key
->inputs
[i
].src_index
];
131 debug_printf(" in[%d] (%s[%d]) %s%s%s%s ",
133 name
, key
->inputs
[i
].src_index
,
134 (key
->inputs
[i
].usage_mask
& 0x1) ? "x" : " ",
135 (key
->inputs
[i
].usage_mask
& 0x2) ? "y" : " ",
136 (key
->inputs
[i
].usage_mask
& 0x4) ? "z" : " ",
137 (key
->inputs
[i
].usage_mask
& 0x8) ? "w" : " ");
139 for (j
= 0; j
< 4; j
++)
140 if (key
->inputs
[i
].usage_mask
& (1<<j
))
141 debug_printf("%.5f ", in
[j
]);
149 * Print triangle vertex attribs (for debug).
152 lp_setup_print_triangle(struct lp_setup_context
*setup
,
153 const float (*v0
)[4],
154 const float (*v1
)[4],
155 const float (*v2
)[4])
157 debug_printf("triangle\n");
160 const float ex
= v0
[0][0] - v2
[0][0];
161 const float ey
= v0
[0][1] - v2
[0][1];
162 const float fx
= v1
[0][0] - v2
[0][0];
163 const float fy
= v1
[0][1] - v2
[0][1];
165 /* det = cross(e,f).z */
166 const float det
= ex
* fy
- ey
* fx
;
168 debug_printf(" - ccw\n");
170 debug_printf(" - cw\n");
172 debug_printf(" - zero area\n");
175 lp_setup_print_vertex(setup
, "v0", v0
);
176 lp_setup_print_vertex(setup
, "v1", v1
);
177 lp_setup_print_vertex(setup
, "v2", v2
);
183 lp_rast_tri_tab
[MAX_PLANES
+1] = {
184 0, /* should be impossible */
185 LP_RAST_OP_TRIANGLE_1
,
186 LP_RAST_OP_TRIANGLE_2
,
187 LP_RAST_OP_TRIANGLE_3
,
188 LP_RAST_OP_TRIANGLE_4
,
189 LP_RAST_OP_TRIANGLE_5
,
190 LP_RAST_OP_TRIANGLE_6
,
191 LP_RAST_OP_TRIANGLE_7
,
192 LP_RAST_OP_TRIANGLE_8
196 lp_rast_32_tri_tab
[MAX_PLANES
+1] = {
197 0, /* should be impossible */
198 LP_RAST_OP_TRIANGLE_32_1
,
199 LP_RAST_OP_TRIANGLE_32_2
,
200 LP_RAST_OP_TRIANGLE_32_3
,
201 LP_RAST_OP_TRIANGLE_32_4
,
202 LP_RAST_OP_TRIANGLE_32_5
,
203 LP_RAST_OP_TRIANGLE_32_6
,
204 LP_RAST_OP_TRIANGLE_32_7
,
205 LP_RAST_OP_TRIANGLE_32_8
211 * The primitive covers the whole tile- shade whole tile.
213 * \param tx, ty the tile position in tiles, not pixels
216 lp_setup_whole_tile(struct lp_setup_context
*setup
,
217 const struct lp_rast_shader_inputs
*inputs
,
220 struct lp_scene
*scene
= setup
->scene
;
222 LP_COUNT(nr_fully_covered_64
);
224 /* if variant is opaque and scissor doesn't effect the tile */
225 if (inputs
->opaque
) {
226 /* Several things prevent this optimization from working:
227 * - For layered rendering we can't determine if this covers the same layer
228 * as previous rendering (or in case of clears those actually always cover
229 * all layers so optimization is impossible). Need to use fb_max_layer and
230 * not setup->layer_slot to determine this since even if there's currently
231 * no slot assigned previous rendering could have used one.
232 * - If there were any Begin/End query commands in the scene then those
233 * would get removed which would be very wrong. Furthermore, if queries
234 * were just active we also can't do the optimization since to get
235 * accurate query results we unfortunately need to execute the rendering
238 if (!scene
->fb
.zsbuf
&& scene
->fb_max_layer
== 0 && !scene
->had_queries
) {
240 * All previous rendering will be overwritten so reset the bin.
242 lp_scene_bin_reset( scene
, tx
, ty
);
245 LP_COUNT(nr_shade_opaque_64
);
246 return lp_scene_bin_cmd_with_state( scene
, tx
, ty
,
248 LP_RAST_OP_SHADE_TILE_OPAQUE
,
249 lp_rast_arg_inputs(inputs
) );
251 LP_COUNT(nr_shade_64
);
252 return lp_scene_bin_cmd_with_state( scene
, tx
, ty
,
254 LP_RAST_OP_SHADE_TILE
,
255 lp_rast_arg_inputs(inputs
) );
261 * Do basic setup for triangle rasterization and determine which
262 * framebuffer tiles are touched. Put the triangle in the scene's
263 * bins for the tiles which we overlap.
266 do_triangle_ccw(struct lp_setup_context
*setup
,
267 struct fixed_position
* position
,
268 const float (*v0
)[4],
269 const float (*v1
)[4],
270 const float (*v2
)[4],
271 boolean frontfacing
)
273 struct lp_scene
*scene
= setup
->scene
;
274 const struct lp_setup_variant_key
*key
= &setup
->setup
.variant
->key
;
275 struct lp_rast_triangle
*tri
;
276 struct lp_rast_plane
*plane
;
280 unsigned viewport_index
= 0;
282 const float (*pv
)[4];
284 /* Area should always be positive here */
285 assert(position
->area
> 0);
288 lp_setup_print_triangle(setup
, v0
, v1
, v2
);
290 if (setup
->flatshade_first
) {
296 if (setup
->viewport_index_slot
> 0) {
297 unsigned *udata
= (unsigned*)pv
[setup
->viewport_index_slot
];
298 viewport_index
= lp_clamp_viewport_idx(*udata
);
300 if (setup
->layer_slot
> 0) {
301 layer
= *(unsigned*)pv
[setup
->layer_slot
];
302 layer
= MIN2(layer
, scene
->fb_max_layer
);
305 if (setup
->scissor_test
) {
312 /* Bounding rectangle (in pixels) */
314 /* Yes this is necessary to accurately calculate bounding boxes
315 * with the two fill-conventions we support. GL (normally) ends
316 * up needing a bottom-left fill convention, which requires
317 * slightly different rounding.
319 int adj
= (setup
->bottom_edge_rule
!= 0) ? 1 : 0;
321 /* Inclusive x0, exclusive x1 */
322 bbox
.x0
= MIN3(position
->x
[0], position
->x
[1], position
->x
[2]) >> FIXED_ORDER
;
323 bbox
.x1
= (MAX3(position
->x
[0], position
->x
[1], position
->x
[2]) - 1) >> FIXED_ORDER
;
325 /* Inclusive / exclusive depending upon adj (bottom-left or top-right) */
326 bbox
.y0
= (MIN3(position
->y
[0], position
->y
[1], position
->y
[2]) + adj
) >> FIXED_ORDER
;
327 bbox
.y1
= (MAX3(position
->y
[0], position
->y
[1], position
->y
[2]) - 1 + adj
) >> FIXED_ORDER
;
330 if (bbox
.x1
< bbox
.x0
||
332 if (0) debug_printf("empty bounding box\n");
333 LP_COUNT(nr_culled_tris
);
337 if (!u_rect_test_intersection(&setup
->draw_regions
[viewport_index
], &bbox
)) {
338 if (0) debug_printf("offscreen\n");
339 LP_COUNT(nr_culled_tris
);
343 /* Can safely discard negative regions, but need to keep hold of
344 * information about when the triangle extends past screen
345 * boundaries. See trimmed_box in lp_setup_bin_triangle().
347 bbox
.x0
= MAX2(bbox
.x0
, 0);
348 bbox
.y0
= MAX2(bbox
.y0
, 0);
350 tri
= lp_setup_alloc_triangle(scene
,
358 tri
->v
[0][0] = v0
[0][0];
359 tri
->v
[1][0] = v1
[0][0];
360 tri
->v
[2][0] = v2
[0][0];
361 tri
->v
[0][1] = v0
[0][1];
362 tri
->v
[1][1] = v1
[0][1];
363 tri
->v
[2][1] = v2
[0][1];
368 /* Setup parameter interpolants:
370 setup
->setup
.variant
->jit_function( v0
,
374 GET_A0(&tri
->inputs
),
375 GET_DADX(&tri
->inputs
),
376 GET_DADY(&tri
->inputs
) );
378 tri
->inputs
.frontfacing
= frontfacing
;
379 tri
->inputs
.disable
= FALSE
;
380 tri
->inputs
.opaque
= setup
->fs
.current
.variant
->opaque
;
381 tri
->inputs
.layer
= layer
;
382 tri
->inputs
.viewport_index
= viewport_index
;
385 lp_dump_setup_coef(&setup
->setup
.variant
->key
,
386 (const float (*)[4])GET_A0(&tri
->inputs
),
387 (const float (*)[4])GET_DADX(&tri
->inputs
),
388 (const float (*)[4])GET_DADY(&tri
->inputs
));
390 plane
= GET_PLANES(tri
);
392 #if defined(PIPE_ARCH_SSE)
394 * XXX this code is effectively disabled for all practical purposes,
395 * as the allowed fb size is tiny if FIXED_ORDER is 8.
397 if (setup
->fb
.width
<= MAX_FIXED_LENGTH32
&&
398 setup
->fb
.height
<= MAX_FIXED_LENGTH32
&&
399 (bbox
.x1
- bbox
.x0
) <= MAX_FIXED_LENGTH32
&&
400 (bbox
.y1
- bbox
.y0
) <= MAX_FIXED_LENGTH32
) {
401 __m128i vertx
, verty
;
402 __m128i shufx
, shufy
;
403 __m128i dcdx
, dcdy
, c
;
405 __m128i dcdx_neg_mask
;
406 __m128i dcdy_neg_mask
;
407 __m128i dcdx_zero_mask
;
408 __m128i top_left_flag
;
409 __m128i c_inc_mask
, c_inc
;
410 __m128i eo
, p0
, p1
, p2
;
411 __m128i zero
= _mm_setzero_si128();
412 PIPE_ALIGN_VAR(16) int32_t temp_vec
[4];
414 vertx
= _mm_load_si128((__m128i
*)position
->x
); /* vertex x coords */
415 verty
= _mm_load_si128((__m128i
*)position
->y
); /* vertex y coords */
417 shufx
= _mm_shuffle_epi32(vertx
, _MM_SHUFFLE(3,0,2,1));
418 shufy
= _mm_shuffle_epi32(verty
, _MM_SHUFFLE(3,0,2,1));
420 dcdx
= _mm_sub_epi32(verty
, shufy
);
421 dcdy
= _mm_sub_epi32(vertx
, shufx
);
423 dcdx_neg_mask
= _mm_srai_epi32(dcdx
, 31);
424 dcdx_zero_mask
= _mm_cmpeq_epi32(dcdx
, zero
);
425 dcdy_neg_mask
= _mm_srai_epi32(dcdy
, 31);
427 top_left_flag
= _mm_set1_epi32((setup
->bottom_edge_rule
== 0) ? ~0 : 0);
429 c_inc_mask
= _mm_or_si128(dcdx_neg_mask
,
430 _mm_and_si128(dcdx_zero_mask
,
431 _mm_xor_si128(dcdy_neg_mask
,
434 c_inc
= _mm_srli_epi32(c_inc_mask
, 31);
436 c
= _mm_sub_epi32(mm_mullo_epi32(dcdx
, vertx
),
437 mm_mullo_epi32(dcdy
, verty
));
439 c
= _mm_add_epi32(c
, c_inc
);
441 /* Scale up to match c:
443 dcdx
= _mm_slli_epi32(dcdx
, FIXED_ORDER
);
444 dcdy
= _mm_slli_epi32(dcdy
, FIXED_ORDER
);
446 /* Calculate trivial reject values:
448 eo
= _mm_sub_epi32(_mm_andnot_si128(dcdy_neg_mask
, dcdy
),
449 _mm_and_si128(dcdx_neg_mask
, dcdx
));
451 /* ei = _mm_sub_epi32(_mm_sub_epi32(dcdy, dcdx), eo); */
453 /* Pointless transpose which gets undone immediately in
456 transpose4_epi32(&c
, &dcdx
, &dcdy
, &eo
,
457 &p0
, &p1
, &p2
, &unused
);
459 #define STORE_PLANE(plane, vec) do { \
460 _mm_store_si128((__m128i *)&temp_vec, vec); \
461 plane.c = (int64_t)temp_vec[0]; \
462 plane.dcdx = temp_vec[1]; \
463 plane.dcdy = temp_vec[2]; \
464 plane.eo = temp_vec[3]; \
467 STORE_PLANE(plane
[0], p0
);
468 STORE_PLANE(plane
[1], p1
);
469 STORE_PLANE(plane
[2], p2
);
472 #elif defined(_ARCH_PWR8) && defined(PIPE_ARCH_LITTLE_ENDIAN)
474 * XXX this code is effectively disabled for all practical purposes,
475 * as the allowed fb size is tiny if FIXED_ORDER is 8.
477 if (setup
->fb
.width
<= MAX_FIXED_LENGTH32
&&
478 setup
->fb
.height
<= MAX_FIXED_LENGTH32
&&
479 (bbox
.x1
- bbox
.x0
) <= MAX_FIXED_LENGTH32
&&
480 (bbox
.y1
- bbox
.y0
) <= MAX_FIXED_LENGTH32
) {
481 unsigned int bottom_edge
;
482 __m128i vertx
, verty
;
483 __m128i shufx
, shufy
;
484 __m128i dcdx
, dcdy
, c
;
486 __m128i dcdx_neg_mask
;
487 __m128i dcdy_neg_mask
;
488 __m128i dcdx_zero_mask
;
489 __m128i top_left_flag
;
490 __m128i c_inc_mask
, c_inc
;
491 __m128i eo
, p0
, p1
, p2
;
492 __m128i_union vshuf_mask
;
493 __m128i zero
= vec_splats((unsigned char) 0);
494 PIPE_ALIGN_VAR(16) int32_t temp_vec
[4];
496 #ifdef PIPE_ARCH_LITTLE_ENDIAN
497 vshuf_mask
.i
[0] = 0x07060504;
498 vshuf_mask
.i
[1] = 0x0B0A0908;
499 vshuf_mask
.i
[2] = 0x03020100;
500 vshuf_mask
.i
[3] = 0x0F0E0D0C;
502 vshuf_mask
.i
[0] = 0x00010203;
503 vshuf_mask
.i
[1] = 0x0C0D0E0F;
504 vshuf_mask
.i
[2] = 0x04050607;
505 vshuf_mask
.i
[3] = 0x08090A0B;
508 /* vertex x coords */
509 vertx
= vec_load_si128((const uint32_t *) position
->x
);
510 /* vertex y coords */
511 verty
= vec_load_si128((const uint32_t *) position
->y
);
513 shufx
= vec_perm (vertx
, vertx
, vshuf_mask
.m128i
);
514 shufy
= vec_perm (verty
, verty
, vshuf_mask
.m128i
);
516 dcdx
= vec_sub_epi32(verty
, shufy
);
517 dcdy
= vec_sub_epi32(vertx
, shufx
);
519 dcdx_neg_mask
= vec_srai_epi32(dcdx
, 31);
520 dcdx_zero_mask
= vec_cmpeq_epi32(dcdx
, zero
);
521 dcdy_neg_mask
= vec_srai_epi32(dcdy
, 31);
523 bottom_edge
= (setup
->bottom_edge_rule
== 0) ? ~0 : 0;
524 top_left_flag
= (__m128i
) vec_splats(bottom_edge
);
526 c_inc_mask
= vec_or(dcdx_neg_mask
,
527 vec_and(dcdx_zero_mask
,
528 vec_xor(dcdy_neg_mask
,
531 c_inc
= vec_srli_epi32(c_inc_mask
, 31);
533 c
= vec_sub_epi32(vec_mullo_epi32(dcdx
, vertx
),
534 vec_mullo_epi32(dcdy
, verty
));
536 c
= vec_add_epi32(c
, c_inc
);
538 /* Scale up to match c:
540 dcdx
= vec_slli_epi32(dcdx
, FIXED_ORDER
);
541 dcdy
= vec_slli_epi32(dcdy
, FIXED_ORDER
);
543 /* Calculate trivial reject values:
545 eo
= vec_sub_epi32(vec_andc(dcdy_neg_mask
, dcdy
),
546 vec_and(dcdx_neg_mask
, dcdx
));
548 /* ei = _mm_sub_epi32(_mm_sub_epi32(dcdy, dcdx), eo); */
550 /* Pointless transpose which gets undone immediately in
553 transpose4_epi32(&c
, &dcdx
, &dcdy
, &eo
,
554 &p0
, &p1
, &p2
, &unused
);
556 #define STORE_PLANE(plane, vec) do { \
557 vec_store_si128((uint32_t *)&temp_vec, vec); \
558 plane.c = (int64_t)temp_vec[0]; \
559 plane.dcdx = temp_vec[1]; \
560 plane.dcdy = temp_vec[2]; \
561 plane.eo = temp_vec[3]; \
564 STORE_PLANE(plane
[0], p0
);
565 STORE_PLANE(plane
[1], p1
);
566 STORE_PLANE(plane
[2], p2
);
572 plane
[0].dcdy
= position
->dx01
;
573 plane
[1].dcdy
= position
->x
[1] - position
->x
[2];
574 plane
[2].dcdy
= position
->dx20
;
575 plane
[0].dcdx
= position
->dy01
;
576 plane
[1].dcdx
= position
->y
[1] - position
->y
[2];
577 plane
[2].dcdx
= position
->dy20
;
579 for (i
= 0; i
< 3; i
++) {
580 /* half-edge constants, will be interated over the whole render
583 plane
[i
].c
= IMUL64(plane
[i
].dcdx
, position
->x
[i
]) -
584 IMUL64(plane
[i
].dcdy
, position
->y
[i
]);
586 /* correct for top-left vs. bottom-left fill convention.
588 if (plane
[i
].dcdx
< 0) {
589 /* both fill conventions want this - adjust for left edges */
592 else if (plane
[i
].dcdx
== 0) {
593 if (setup
->bottom_edge_rule
== 0){
594 /* correct for top-left fill convention:
596 if (plane
[i
].dcdy
> 0) plane
[i
].c
++;
599 /* correct for bottom-left fill convention:
601 if (plane
[i
].dcdy
< 0) plane
[i
].c
++;
605 /* Scale up to match c:
607 assert((plane
[i
].dcdx
<< FIXED_ORDER
) >> FIXED_ORDER
== plane
[i
].dcdx
);
608 assert((plane
[i
].dcdy
<< FIXED_ORDER
) >> FIXED_ORDER
== plane
[i
].dcdy
);
609 plane
[i
].dcdx
<<= FIXED_ORDER
;
610 plane
[i
].dcdy
<<= FIXED_ORDER
;
612 /* find trivial reject offsets for each edge for a single-pixel
613 * sized block. These will be scaled up at each recursive level to
614 * match the active blocksize. Scaling in this way works best if
615 * the blocks are square.
618 if (plane
[i
].dcdx
< 0) plane
[i
].eo
-= plane
[i
].dcdx
;
619 if (plane
[i
].dcdy
> 0) plane
[i
].eo
+= plane
[i
].dcdy
;
624 debug_printf("p0: %"PRIx64
"/%08x/%08x/%08x\n",
630 debug_printf("p1: %"PRIx64
"/%08x/%08x/%08x\n",
636 debug_printf("p2: %"PRIx64
"/%08x/%08x/%08x\n",
645 * When rasterizing scissored tris, use the intersection of the
646 * triangle bounding box and the scissor rect to generate the
649 * This permits us to cut off the triangle "tails" that are present
650 * in the intermediate recursive levels caused when two of the
651 * triangles edges don't diverge quickly enough to trivially reject
652 * exterior blocks from the triangle.
654 * It's not really clear if it's worth worrying about these tails,
655 * but since we generate the planes for each scissored tri, it's
656 * free to trim them in this case.
658 * Note that otherwise, the scissor planes only vary in 'C' value,
659 * and even then only on state-changes. Could alternatively store
660 * these planes elsewhere.
662 if (nr_planes
== 7) {
663 const struct u_rect
*scissor
= &setup
->scissors
[viewport_index
];
667 plane
[3].c
= 1-scissor
->x0
;
672 plane
[4].c
= scissor
->x1
+1;
677 plane
[5].c
= 1-scissor
->y0
;
682 plane
[6].c
= scissor
->y1
+1;
686 return lp_setup_bin_triangle(setup
, tri
, &bbox
, nr_planes
, viewport_index
);
690 * Round to nearest less or equal power of two of the input.
692 * Undefined if no bit set exists, so code should check against 0 first.
694 static inline uint32_t
695 floor_pot(uint32_t n
)
697 #if defined(PIPE_CC_GCC) && (defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64))
717 lp_setup_bin_triangle( struct lp_setup_context
*setup
,
718 struct lp_rast_triangle
*tri
,
719 const struct u_rect
*bbox
,
721 unsigned viewport_index
)
723 struct lp_scene
*scene
= setup
->scene
;
724 struct u_rect trimmed_box
= *bbox
;
726 /* What is the largest power-of-two boundary this triangle crosses:
728 int dx
= floor_pot((bbox
->x0
^ bbox
->x1
) |
729 (bbox
->y0
^ bbox
->y1
));
731 /* The largest dimension of the rasterized area of the triangle
732 * (aligned to a 4x4 grid), rounded down to the nearest power of two:
734 int max_sz
= ((bbox
->x1
- (bbox
->x0
& ~3)) |
735 (bbox
->y1
- (bbox
->y0
& ~3)));
736 int sz
= floor_pot(max_sz
);
737 boolean use_32bits
= max_sz
<= MAX_FIXED_LENGTH32
;
739 /* Now apply scissor, etc to the bounding box. Could do this
740 * earlier, but it confuses the logic for tri-16 and would force
741 * the rasterizer to also respect scissor, etc, just for the rare
742 * cases where a small triangle extends beyond the scissor.
744 u_rect_find_intersection(&setup
->draw_regions
[viewport_index
],
747 /* Determine which tile(s) intersect the triangle's bounding box
751 int ix0
= bbox
->x0
/ TILE_SIZE
;
752 int iy0
= bbox
->y0
/ TILE_SIZE
;
753 unsigned px
= bbox
->x0
& 63 & ~3;
754 unsigned py
= bbox
->y0
& 63 & ~3;
756 assert(iy0
== bbox
->y1
/ TILE_SIZE
&&
757 ix0
== bbox
->x1
/ TILE_SIZE
);
759 if (nr_planes
== 3) {
762 /* Triangle is contained in a single 4x4 stamp:
764 assert(px
+ 4 <= TILE_SIZE
);
765 assert(py
+ 4 <= TILE_SIZE
);
766 return lp_scene_bin_cmd_with_state( scene
, ix0
, iy0
,
769 LP_RAST_OP_TRIANGLE_32_3_4
:
770 LP_RAST_OP_TRIANGLE_3_4
,
771 lp_rast_arg_triangle_contained(tri
, px
, py
) );
776 /* Triangle is contained in a single 16x16 block:
780 * The 16x16 block is only 4x4 aligned, and can exceed the tile
781 * dimensions if the triangle is 16 pixels in one dimension but 4
782 * in the other. So budge the 16x16 back inside the tile.
784 px
= MIN2(px
, TILE_SIZE
- 16);
785 py
= MIN2(py
, TILE_SIZE
- 16);
787 assert(px
+ 16 <= TILE_SIZE
);
788 assert(py
+ 16 <= TILE_SIZE
);
790 return lp_scene_bin_cmd_with_state( scene
, ix0
, iy0
,
793 LP_RAST_OP_TRIANGLE_32_3_16
:
794 LP_RAST_OP_TRIANGLE_3_16
,
795 lp_rast_arg_triangle_contained(tri
, px
, py
) );
798 else if (nr_planes
== 4 && sz
< 16)
800 px
= MIN2(px
, TILE_SIZE
- 16);
801 py
= MIN2(py
, TILE_SIZE
- 16);
803 assert(px
+ 16 <= TILE_SIZE
);
804 assert(py
+ 16 <= TILE_SIZE
);
806 return lp_scene_bin_cmd_with_state(scene
, ix0
, iy0
,
809 LP_RAST_OP_TRIANGLE_32_4_16
:
810 LP_RAST_OP_TRIANGLE_4_16
,
811 lp_rast_arg_triangle_contained(tri
, px
, py
));
815 /* Triangle is contained in a single tile:
817 return lp_scene_bin_cmd_with_state(
818 scene
, ix0
, iy0
, setup
->fs
.stored
,
819 use_32bits
? lp_rast_32_tri_tab
[nr_planes
] : lp_rast_tri_tab
[nr_planes
],
820 lp_rast_arg_triangle(tri
, (1<<nr_planes
)-1));
824 struct lp_rast_plane
*plane
= GET_PLANES(tri
);
825 int64_t c
[MAX_PLANES
];
826 int64_t ei
[MAX_PLANES
];
828 int64_t eo
[MAX_PLANES
];
829 int64_t xstep
[MAX_PLANES
];
830 int64_t ystep
[MAX_PLANES
];
833 int ix0
= trimmed_box
.x0
/ TILE_SIZE
;
834 int iy0
= trimmed_box
.y0
/ TILE_SIZE
;
835 int ix1
= trimmed_box
.x1
/ TILE_SIZE
;
836 int iy1
= trimmed_box
.y1
/ TILE_SIZE
;
838 for (i
= 0; i
< nr_planes
; i
++) {
840 IMUL64(plane
[i
].dcdy
, iy0
) * TILE_SIZE
-
841 IMUL64(plane
[i
].dcdx
, ix0
) * TILE_SIZE
);
843 ei
[i
] = (plane
[i
].dcdy
-
845 (int64_t)plane
[i
].eo
) << TILE_ORDER
;
847 eo
[i
] = (int64_t)plane
[i
].eo
<< TILE_ORDER
;
848 xstep
[i
] = -(((int64_t)plane
[i
].dcdx
) << TILE_ORDER
);
849 ystep
[i
] = ((int64_t)plane
[i
].dcdy
) << TILE_ORDER
;
854 /* Test tile-sized blocks against the triangle.
855 * Discard blocks fully outside the tri. If the block is fully
856 * contained inside the tri, bin an lp_rast_shade_tile command.
857 * Else, bin a lp_rast_triangle command.
859 for (y
= iy0
; y
<= iy1
; y
++)
861 boolean in
= FALSE
; /* are we inside the triangle? */
862 int64_t cx
[MAX_PLANES
];
864 for (i
= 0; i
< nr_planes
; i
++)
867 for (x
= ix0
; x
<= ix1
; x
++)
872 for (i
= 0; i
< nr_planes
; i
++) {
873 int64_t planeout
= cx
[i
] + eo
[i
];
874 int64_t planepartial
= cx
[i
] + ei
[i
] - 1;
875 out
|= (int) (planeout
>> 63);
876 partial
|= ((int) (planepartial
>> 63)) & (1<<i
);
882 break; /* exiting triangle, all done with this row */
883 LP_COUNT(nr_empty_64
);
886 /* Not trivially accepted by at least one plane -
887 * rasterize/shade partial tile
889 int count
= util_bitcount(partial
);
892 if (!lp_scene_bin_cmd_with_state( scene
, x
, y
,
895 lp_rast_32_tri_tab
[count
] :
896 lp_rast_tri_tab
[count
],
897 lp_rast_arg_triangle(tri
, partial
) ))
900 LP_COUNT(nr_partially_covered_64
);
903 /* triangle covers the whole tile- shade whole tile */
904 LP_COUNT(nr_fully_covered_64
);
906 if (!lp_setup_whole_tile(setup
, &tri
->inputs
, x
, y
))
910 /* Iterate cx values across the region: */
911 for (i
= 0; i
< nr_planes
; i
++)
915 /* Iterate c values down the region: */
916 for (i
= 0; i
< nr_planes
; i
++)
924 /* Need to disable any partially binned triangle. This is easier
925 * than trying to locate all the triangle, shade-tile, etc,
926 * commands which may have been binned.
928 tri
->inputs
.disable
= TRUE
;
934 * Try to draw the triangle, restart the scene on failure.
936 static void retry_triangle_ccw( struct lp_setup_context
*setup
,
937 struct fixed_position
* position
,
938 const float (*v0
)[4],
939 const float (*v1
)[4],
940 const float (*v2
)[4],
943 if (!do_triangle_ccw( setup
, position
, v0
, v1
, v2
, front
))
945 if (!lp_setup_flush_and_restart(setup
))
948 if (!do_triangle_ccw( setup
, position
, v0
, v1
, v2
, front
))
954 * Calculate fixed position data for a triangle
957 calc_fixed_position( struct lp_setup_context
*setup
,
958 struct fixed_position
* position
,
959 const float (*v0
)[4],
960 const float (*v1
)[4],
961 const float (*v2
)[4])
963 position
->x
[0] = subpixel_snap(v0
[0][0] - setup
->pixel_offset
);
964 position
->x
[1] = subpixel_snap(v1
[0][0] - setup
->pixel_offset
);
965 position
->x
[2] = subpixel_snap(v2
[0][0] - setup
->pixel_offset
);
968 position
->y
[0] = subpixel_snap(v0
[0][1] - setup
->pixel_offset
);
969 position
->y
[1] = subpixel_snap(v1
[0][1] - setup
->pixel_offset
);
970 position
->y
[2] = subpixel_snap(v2
[0][1] - setup
->pixel_offset
);
973 position
->dx01
= position
->x
[0] - position
->x
[1];
974 position
->dy01
= position
->y
[0] - position
->y
[1];
976 position
->dx20
= position
->x
[2] - position
->x
[0];
977 position
->dy20
= position
->y
[2] - position
->y
[0];
979 position
->area
= IMUL64(position
->dx01
, position
->dy20
) -
980 IMUL64(position
->dx20
, position
->dy01
);
985 * Rotate a triangle, flipping its clockwise direction,
986 * Swaps values for xy[0] and xy[1]
989 rotate_fixed_position_01( struct fixed_position
* position
)
995 position
->x
[1] = position
->x
[0];
996 position
->y
[1] = position
->y
[0];
1000 position
->dx01
= -position
->dx01
;
1001 position
->dy01
= -position
->dy01
;
1002 position
->dx20
= position
->x
[2] - position
->x
[0];
1003 position
->dy20
= position
->y
[2] - position
->y
[0];
1005 position
->area
= -position
->area
;
1010 * Rotate a triangle, flipping its clockwise direction,
1011 * Swaps values for xy[1] and xy[2]
1014 rotate_fixed_position_12( struct fixed_position
* position
)
1020 position
->x
[2] = position
->x
[1];
1021 position
->y
[2] = position
->y
[1];
1027 position
->dx01
= -position
->dx20
;
1028 position
->dy01
= -position
->dy20
;
1029 position
->dx20
= -x
;
1030 position
->dy20
= -y
;
1032 position
->area
= -position
->area
;
1037 * Draw triangle if it's CW, cull otherwise.
1039 static void triangle_cw(struct lp_setup_context
*setup
,
1040 const float (*v0
)[4],
1041 const float (*v1
)[4],
1042 const float (*v2
)[4])
1044 PIPE_ALIGN_VAR(16) struct fixed_position position
;
1046 calc_fixed_position(setup
, &position
, v0
, v1
, v2
);
1048 if (position
.area
< 0) {
1049 if (setup
->flatshade_first
) {
1050 rotate_fixed_position_12(&position
);
1051 retry_triangle_ccw(setup
, &position
, v0
, v2
, v1
, !setup
->ccw_is_frontface
);
1053 rotate_fixed_position_01(&position
);
1054 retry_triangle_ccw(setup
, &position
, v1
, v0
, v2
, !setup
->ccw_is_frontface
);
1060 static void triangle_ccw(struct lp_setup_context
*setup
,
1061 const float (*v0
)[4],
1062 const float (*v1
)[4],
1063 const float (*v2
)[4])
1065 PIPE_ALIGN_VAR(16) struct fixed_position position
;
1067 calc_fixed_position(setup
, &position
, v0
, v1
, v2
);
1069 if (position
.area
> 0)
1070 retry_triangle_ccw(setup
, &position
, v0
, v1
, v2
, setup
->ccw_is_frontface
);
1074 * Draw triangle whether it's CW or CCW.
1076 static void triangle_both(struct lp_setup_context
*setup
,
1077 const float (*v0
)[4],
1078 const float (*v1
)[4],
1079 const float (*v2
)[4])
1081 PIPE_ALIGN_VAR(16) struct fixed_position position
;
1082 struct llvmpipe_context
*lp_context
= (struct llvmpipe_context
*)setup
->pipe
;
1084 if (lp_context
->active_statistics_queries
&&
1085 !llvmpipe_rasterization_disabled(lp_context
)) {
1086 lp_context
->pipeline_statistics
.c_primitives
++;
1089 calc_fixed_position(setup
, &position
, v0
, v1
, v2
);
1092 assert(!util_is_inf_or_nan(v0
[0][0]));
1093 assert(!util_is_inf_or_nan(v0
[0][1]));
1094 assert(!util_is_inf_or_nan(v1
[0][0]));
1095 assert(!util_is_inf_or_nan(v1
[0][1]));
1096 assert(!util_is_inf_or_nan(v2
[0][0]));
1097 assert(!util_is_inf_or_nan(v2
[0][1]));
1100 if (position
.area
> 0)
1101 retry_triangle_ccw( setup
, &position
, v0
, v1
, v2
, setup
->ccw_is_frontface
);
1102 else if (position
.area
< 0) {
1103 if (setup
->flatshade_first
) {
1104 rotate_fixed_position_12( &position
);
1105 retry_triangle_ccw( setup
, &position
, v0
, v2
, v1
, !setup
->ccw_is_frontface
);
1107 rotate_fixed_position_01( &position
);
1108 retry_triangle_ccw( setup
, &position
, v1
, v0
, v2
, !setup
->ccw_is_frontface
);
1114 static void triangle_nop( struct lp_setup_context
*setup
,
1115 const float (*v0
)[4],
1116 const float (*v1
)[4],
1117 const float (*v2
)[4] )
1123 lp_setup_choose_triangle( struct lp_setup_context
*setup
)
1125 switch (setup
->cullmode
) {
1126 case PIPE_FACE_NONE
:
1127 setup
->triangle
= triangle_both
;
1129 case PIPE_FACE_BACK
:
1130 setup
->triangle
= setup
->ccw_is_frontface
? triangle_ccw
: triangle_cw
;
1132 case PIPE_FACE_FRONT
:
1133 setup
->triangle
= setup
->ccw_is_frontface
? triangle_cw
: triangle_ccw
;
1136 setup
->triangle
= triangle_nop
;