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 __m128i vertx
, verty
;
395 __m128i shufx
, shufy
;
397 __m128i cdx02
, cdx13
, cdy02
, cdy13
, c02
, c13
;
398 __m128i c01
, c23
, unused
;
399 __m128i dcdx_neg_mask
;
400 __m128i dcdy_neg_mask
;
401 __m128i dcdx_zero_mask
;
402 __m128i top_left_flag
, c_dec
;
403 __m128i eo
, p0
, p1
, p2
;
404 __m128i zero
= _mm_setzero_si128();
406 vertx
= _mm_load_si128((__m128i
*)position
->x
); /* vertex x coords */
407 verty
= _mm_load_si128((__m128i
*)position
->y
); /* vertex y coords */
409 shufx
= _mm_shuffle_epi32(vertx
, _MM_SHUFFLE(3,0,2,1));
410 shufy
= _mm_shuffle_epi32(verty
, _MM_SHUFFLE(3,0,2,1));
412 dcdx
= _mm_sub_epi32(verty
, shufy
);
413 dcdy
= _mm_sub_epi32(vertx
, shufx
);
415 dcdx_neg_mask
= _mm_srai_epi32(dcdx
, 31);
416 dcdx_zero_mask
= _mm_cmpeq_epi32(dcdx
, zero
);
417 dcdy_neg_mask
= _mm_srai_epi32(dcdy
, 31);
419 top_left_flag
= _mm_set1_epi32((setup
->bottom_edge_rule
== 0) ? ~0 : 0);
421 c_dec
= _mm_or_si128(dcdx_neg_mask
,
422 _mm_and_si128(dcdx_zero_mask
,
423 _mm_xor_si128(dcdy_neg_mask
,
428 * Note we need _signed_ mul (_mm_mul_epi32) which we emulate.
430 cdx02
= mm_mullohi_epi32(dcdx
, vertx
, &cdx13
);
431 cdy02
= mm_mullohi_epi32(dcdy
, verty
, &cdy13
);
432 c02
= _mm_sub_epi64(cdx02
, cdy02
);
433 c13
= _mm_sub_epi64(cdx13
, cdy13
);
434 c02
= _mm_sub_epi64(c02
, _mm_shuffle_epi32(c_dec
,
435 _MM_SHUFFLE(2,2,0,0)));
436 c13
= _mm_sub_epi64(c13
, _mm_shuffle_epi32(c_dec
,
437 _MM_SHUFFLE(3,3,1,1)));
440 * Useful for very small fbs/tris (or fewer subpixel bits) only:
441 * c = _mm_sub_epi32(mm_mullo_epi32(dcdx, vertx),
442 * mm_mullo_epi32(dcdy, verty));
444 * c = _mm_sub_epi32(c, c_dec);
447 /* Scale up to match c:
449 dcdx
= _mm_slli_epi32(dcdx
, FIXED_ORDER
);
450 dcdy
= _mm_slli_epi32(dcdy
, FIXED_ORDER
);
453 * Calculate trivial reject values:
454 * Note eo cannot overflow even if dcdx/dcdy would already have
455 * 31 bits (which they shouldn't have). This is because eo
456 * is never negative (albeit if we rely on that need to be careful...)
458 eo
= _mm_sub_epi32(_mm_andnot_si128(dcdy_neg_mask
, dcdy
),
459 _mm_and_si128(dcdx_neg_mask
, dcdx
));
461 /* ei = _mm_sub_epi32(_mm_sub_epi32(dcdy, dcdx), eo); */
464 * Pointless transpose which gets undone immediately in
466 * It is actually difficult to do away with it - would essentially
467 * need GET_PLANES_DX, GET_PLANES_DY etc., but the calculations
468 * for this then would need to depend on the number of planes.
469 * The transpose is quite special here due to c being 64bit...
470 * The store has to be unaligned (unless we'd make the plane size
471 * a multiple of 128), and of course storing eo separately...
473 c01
= _mm_unpacklo_epi64(c02
, c13
);
474 c23
= _mm_unpackhi_epi64(c02
, c13
);
475 transpose2_64_2_32(&c01
, &c23
, &dcdx
, &dcdy
,
476 &p0
, &p1
, &p2
, &unused
);
477 _mm_storeu_si128((__m128i
*)&plane
[0], p0
);
478 plane
[0].eo
= (uint32_t)_mm_cvtsi128_si32(eo
);
479 _mm_storeu_si128((__m128i
*)&plane
[1], p1
);
480 eo
= _mm_shuffle_epi32(eo
, _MM_SHUFFLE(3,2,0,1));
481 plane
[1].eo
= (uint32_t)_mm_cvtsi128_si32(eo
);
482 _mm_storeu_si128((__m128i
*)&plane
[2], p2
);
483 eo
= _mm_shuffle_epi32(eo
, _MM_SHUFFLE(0,0,0,2));
484 plane
[2].eo
= (uint32_t)_mm_cvtsi128_si32(eo
);
486 #elif defined(_ARCH_PWR8) && defined(PIPE_ARCH_LITTLE_ENDIAN)
488 * XXX this code is effectively disabled for all practical purposes,
489 * as the allowed fb size is tiny if FIXED_ORDER is 8.
491 if (setup
->fb
.width
<= MAX_FIXED_LENGTH32
&&
492 setup
->fb
.height
<= MAX_FIXED_LENGTH32
&&
493 (bbox
.x1
- bbox
.x0
) <= MAX_FIXED_LENGTH32
&&
494 (bbox
.y1
- bbox
.y0
) <= MAX_FIXED_LENGTH32
) {
495 unsigned int bottom_edge
;
496 __m128i vertx
, verty
;
497 __m128i shufx
, shufy
;
498 __m128i dcdx
, dcdy
, c
;
500 __m128i dcdx_neg_mask
;
501 __m128i dcdy_neg_mask
;
502 __m128i dcdx_zero_mask
;
503 __m128i top_left_flag
;
504 __m128i c_inc_mask
, c_inc
;
505 __m128i eo
, p0
, p1
, p2
;
506 __m128i_union vshuf_mask
;
507 __m128i zero
= vec_splats((unsigned char) 0);
508 PIPE_ALIGN_VAR(16) int32_t temp_vec
[4];
510 #ifdef PIPE_ARCH_LITTLE_ENDIAN
511 vshuf_mask
.i
[0] = 0x07060504;
512 vshuf_mask
.i
[1] = 0x0B0A0908;
513 vshuf_mask
.i
[2] = 0x03020100;
514 vshuf_mask
.i
[3] = 0x0F0E0D0C;
516 vshuf_mask
.i
[0] = 0x00010203;
517 vshuf_mask
.i
[1] = 0x0C0D0E0F;
518 vshuf_mask
.i
[2] = 0x04050607;
519 vshuf_mask
.i
[3] = 0x08090A0B;
522 /* vertex x coords */
523 vertx
= vec_load_si128((const uint32_t *) position
->x
);
524 /* vertex y coords */
525 verty
= vec_load_si128((const uint32_t *) position
->y
);
527 shufx
= vec_perm (vertx
, vertx
, vshuf_mask
.m128i
);
528 shufy
= vec_perm (verty
, verty
, vshuf_mask
.m128i
);
530 dcdx
= vec_sub_epi32(verty
, shufy
);
531 dcdy
= vec_sub_epi32(vertx
, shufx
);
533 dcdx_neg_mask
= vec_srai_epi32(dcdx
, 31);
534 dcdx_zero_mask
= vec_cmpeq_epi32(dcdx
, zero
);
535 dcdy_neg_mask
= vec_srai_epi32(dcdy
, 31);
537 bottom_edge
= (setup
->bottom_edge_rule
== 0) ? ~0 : 0;
538 top_left_flag
= (__m128i
) vec_splats(bottom_edge
);
540 c_inc_mask
= vec_or(dcdx_neg_mask
,
541 vec_and(dcdx_zero_mask
,
542 vec_xor(dcdy_neg_mask
,
545 c_inc
= vec_srli_epi32(c_inc_mask
, 31);
547 c
= vec_sub_epi32(vec_mullo_epi32(dcdx
, vertx
),
548 vec_mullo_epi32(dcdy
, verty
));
550 c
= vec_add_epi32(c
, c_inc
);
552 /* Scale up to match c:
554 dcdx
= vec_slli_epi32(dcdx
, FIXED_ORDER
);
555 dcdy
= vec_slli_epi32(dcdy
, FIXED_ORDER
);
557 /* Calculate trivial reject values:
559 eo
= vec_sub_epi32(vec_andc(dcdy_neg_mask
, dcdy
),
560 vec_and(dcdx_neg_mask
, dcdx
));
562 /* ei = _mm_sub_epi32(_mm_sub_epi32(dcdy, dcdx), eo); */
564 /* Pointless transpose which gets undone immediately in
567 transpose4_epi32(&c
, &dcdx
, &dcdy
, &eo
,
568 &p0
, &p1
, &p2
, &unused
);
570 #define STORE_PLANE(plane, vec) do { \
571 vec_store_si128((uint32_t *)&temp_vec, vec); \
572 plane.c = (int64_t)temp_vec[0]; \
573 plane.dcdx = temp_vec[1]; \
574 plane.dcdy = temp_vec[2]; \
575 plane.eo = temp_vec[3]; \
578 STORE_PLANE(plane
[0], p0
);
579 STORE_PLANE(plane
[1], p1
);
580 STORE_PLANE(plane
[2], p2
);
586 plane
[0].dcdy
= position
->dx01
;
587 plane
[1].dcdy
= position
->x
[1] - position
->x
[2];
588 plane
[2].dcdy
= position
->dx20
;
589 plane
[0].dcdx
= position
->dy01
;
590 plane
[1].dcdx
= position
->y
[1] - position
->y
[2];
591 plane
[2].dcdx
= position
->dy20
;
593 for (i
= 0; i
< 3; i
++) {
594 /* half-edge constants, will be iterated over the whole render
597 plane
[i
].c
= IMUL64(plane
[i
].dcdx
, position
->x
[i
]) -
598 IMUL64(plane
[i
].dcdy
, position
->y
[i
]);
600 /* correct for top-left vs. bottom-left fill convention.
602 if (plane
[i
].dcdx
< 0) {
603 /* both fill conventions want this - adjust for left edges */
606 else if (plane
[i
].dcdx
== 0) {
607 if (setup
->bottom_edge_rule
== 0){
608 /* correct for top-left fill convention:
610 if (plane
[i
].dcdy
> 0) plane
[i
].c
++;
613 /* correct for bottom-left fill convention:
615 if (plane
[i
].dcdy
< 0) plane
[i
].c
++;
619 /* Scale up to match c:
621 assert((plane
[i
].dcdx
<< FIXED_ORDER
) >> FIXED_ORDER
== plane
[i
].dcdx
);
622 assert((plane
[i
].dcdy
<< FIXED_ORDER
) >> FIXED_ORDER
== plane
[i
].dcdy
);
623 plane
[i
].dcdx
<<= FIXED_ORDER
;
624 plane
[i
].dcdy
<<= FIXED_ORDER
;
626 /* find trivial reject offsets for each edge for a single-pixel
627 * sized block. These will be scaled up at each recursive level to
628 * match the active blocksize. Scaling in this way works best if
629 * the blocks are square.
632 if (plane
[i
].dcdx
< 0) plane
[i
].eo
-= plane
[i
].dcdx
;
633 if (plane
[i
].dcdy
> 0) plane
[i
].eo
+= plane
[i
].dcdy
;
638 debug_printf("p0: %"PRIx64
"/%08x/%08x/%08x\n",
644 debug_printf("p1: %"PRIx64
"/%08x/%08x/%08x\n",
650 debug_printf("p2: %"PRIx64
"/%08x/%08x/%08x\n",
659 * When rasterizing scissored tris, use the intersection of the
660 * triangle bounding box and the scissor rect to generate the
663 * This permits us to cut off the triangle "tails" that are present
664 * in the intermediate recursive levels caused when two of the
665 * triangles edges don't diverge quickly enough to trivially reject
666 * exterior blocks from the triangle.
668 * It's not really clear if it's worth worrying about these tails,
669 * but since we generate the planes for each scissored tri, it's
670 * free to trim them in this case.
672 * Note that otherwise, the scissor planes only vary in 'C' value,
673 * and even then only on state-changes. Could alternatively store
674 * these planes elsewhere.
676 if (nr_planes
== 7) {
677 const struct u_rect
*scissor
= &setup
->scissors
[viewport_index
];
681 plane
[3].c
= 1-scissor
->x0
;
686 plane
[4].c
= scissor
->x1
+1;
691 plane
[5].c
= 1-scissor
->y0
;
696 plane
[6].c
= scissor
->y1
+1;
700 return lp_setup_bin_triangle(setup
, tri
, &bbox
, nr_planes
, viewport_index
);
704 * Round to nearest less or equal power of two of the input.
706 * Undefined if no bit set exists, so code should check against 0 first.
708 static inline uint32_t
709 floor_pot(uint32_t n
)
711 #if defined(PIPE_CC_GCC) && (defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64))
731 lp_setup_bin_triangle( struct lp_setup_context
*setup
,
732 struct lp_rast_triangle
*tri
,
733 const struct u_rect
*bbox
,
735 unsigned viewport_index
)
737 struct lp_scene
*scene
= setup
->scene
;
738 struct u_rect trimmed_box
= *bbox
;
740 /* What is the largest power-of-two boundary this triangle crosses:
742 int dx
= floor_pot((bbox
->x0
^ bbox
->x1
) |
743 (bbox
->y0
^ bbox
->y1
));
745 /* The largest dimension of the rasterized area of the triangle
746 * (aligned to a 4x4 grid), rounded down to the nearest power of two:
748 int max_sz
= ((bbox
->x1
- (bbox
->x0
& ~3)) |
749 (bbox
->y1
- (bbox
->y0
& ~3)));
750 int sz
= floor_pot(max_sz
);
751 boolean use_32bits
= max_sz
<= MAX_FIXED_LENGTH32
;
753 /* Now apply scissor, etc to the bounding box. Could do this
754 * earlier, but it confuses the logic for tri-16 and would force
755 * the rasterizer to also respect scissor, etc, just for the rare
756 * cases where a small triangle extends beyond the scissor.
758 u_rect_find_intersection(&setup
->draw_regions
[viewport_index
],
761 /* Determine which tile(s) intersect the triangle's bounding box
765 int ix0
= bbox
->x0
/ TILE_SIZE
;
766 int iy0
= bbox
->y0
/ TILE_SIZE
;
767 unsigned px
= bbox
->x0
& 63 & ~3;
768 unsigned py
= bbox
->y0
& 63 & ~3;
770 assert(iy0
== bbox
->y1
/ TILE_SIZE
&&
771 ix0
== bbox
->x1
/ TILE_SIZE
);
773 if (nr_planes
== 3) {
776 /* Triangle is contained in a single 4x4 stamp:
778 assert(px
+ 4 <= TILE_SIZE
);
779 assert(py
+ 4 <= TILE_SIZE
);
780 return lp_scene_bin_cmd_with_state( scene
, ix0
, iy0
,
783 LP_RAST_OP_TRIANGLE_32_3_4
:
784 LP_RAST_OP_TRIANGLE_3_4
,
785 lp_rast_arg_triangle_contained(tri
, px
, py
) );
790 /* Triangle is contained in a single 16x16 block:
794 * The 16x16 block is only 4x4 aligned, and can exceed the tile
795 * dimensions if the triangle is 16 pixels in one dimension but 4
796 * in the other. So budge the 16x16 back inside the tile.
798 px
= MIN2(px
, TILE_SIZE
- 16);
799 py
= MIN2(py
, TILE_SIZE
- 16);
801 assert(px
+ 16 <= TILE_SIZE
);
802 assert(py
+ 16 <= TILE_SIZE
);
804 return lp_scene_bin_cmd_with_state( scene
, ix0
, iy0
,
807 LP_RAST_OP_TRIANGLE_32_3_16
:
808 LP_RAST_OP_TRIANGLE_3_16
,
809 lp_rast_arg_triangle_contained(tri
, px
, py
) );
812 else if (nr_planes
== 4 && sz
< 16)
814 px
= MIN2(px
, TILE_SIZE
- 16);
815 py
= MIN2(py
, TILE_SIZE
- 16);
817 assert(px
+ 16 <= TILE_SIZE
);
818 assert(py
+ 16 <= TILE_SIZE
);
820 return lp_scene_bin_cmd_with_state(scene
, ix0
, iy0
,
823 LP_RAST_OP_TRIANGLE_32_4_16
:
824 LP_RAST_OP_TRIANGLE_4_16
,
825 lp_rast_arg_triangle_contained(tri
, px
, py
));
829 /* Triangle is contained in a single tile:
831 return lp_scene_bin_cmd_with_state(
832 scene
, ix0
, iy0
, setup
->fs
.stored
,
833 use_32bits
? lp_rast_32_tri_tab
[nr_planes
] : lp_rast_tri_tab
[nr_planes
],
834 lp_rast_arg_triangle(tri
, (1<<nr_planes
)-1));
838 struct lp_rast_plane
*plane
= GET_PLANES(tri
);
839 int64_t c
[MAX_PLANES
];
840 int64_t ei
[MAX_PLANES
];
842 int64_t eo
[MAX_PLANES
];
843 int64_t xstep
[MAX_PLANES
];
844 int64_t ystep
[MAX_PLANES
];
847 int ix0
= trimmed_box
.x0
/ TILE_SIZE
;
848 int iy0
= trimmed_box
.y0
/ TILE_SIZE
;
849 int ix1
= trimmed_box
.x1
/ TILE_SIZE
;
850 int iy1
= trimmed_box
.y1
/ TILE_SIZE
;
852 for (i
= 0; i
< nr_planes
; i
++) {
854 IMUL64(plane
[i
].dcdy
, iy0
) * TILE_SIZE
-
855 IMUL64(plane
[i
].dcdx
, ix0
) * TILE_SIZE
);
857 ei
[i
] = (plane
[i
].dcdy
-
859 (int64_t)plane
[i
].eo
) << TILE_ORDER
;
861 eo
[i
] = (int64_t)plane
[i
].eo
<< TILE_ORDER
;
862 xstep
[i
] = -(((int64_t)plane
[i
].dcdx
) << TILE_ORDER
);
863 ystep
[i
] = ((int64_t)plane
[i
].dcdy
) << TILE_ORDER
;
868 /* Test tile-sized blocks against the triangle.
869 * Discard blocks fully outside the tri. If the block is fully
870 * contained inside the tri, bin an lp_rast_shade_tile command.
871 * Else, bin a lp_rast_triangle command.
873 for (y
= iy0
; y
<= iy1
; y
++)
875 boolean in
= FALSE
; /* are we inside the triangle? */
876 int64_t cx
[MAX_PLANES
];
878 for (i
= 0; i
< nr_planes
; i
++)
881 for (x
= ix0
; x
<= ix1
; x
++)
886 for (i
= 0; i
< nr_planes
; i
++) {
887 int64_t planeout
= cx
[i
] + eo
[i
];
888 int64_t planepartial
= cx
[i
] + ei
[i
] - 1;
889 out
|= (int) (planeout
>> 63);
890 partial
|= ((int) (planepartial
>> 63)) & (1<<i
);
896 break; /* exiting triangle, all done with this row */
897 LP_COUNT(nr_empty_64
);
900 /* Not trivially accepted by at least one plane -
901 * rasterize/shade partial tile
903 int count
= util_bitcount(partial
);
906 if (!lp_scene_bin_cmd_with_state( scene
, x
, y
,
909 lp_rast_32_tri_tab
[count
] :
910 lp_rast_tri_tab
[count
],
911 lp_rast_arg_triangle(tri
, partial
) ))
914 LP_COUNT(nr_partially_covered_64
);
917 /* triangle covers the whole tile- shade whole tile */
918 LP_COUNT(nr_fully_covered_64
);
920 if (!lp_setup_whole_tile(setup
, &tri
->inputs
, x
, y
))
924 /* Iterate cx values across the region: */
925 for (i
= 0; i
< nr_planes
; i
++)
929 /* Iterate c values down the region: */
930 for (i
= 0; i
< nr_planes
; i
++)
938 /* Need to disable any partially binned triangle. This is easier
939 * than trying to locate all the triangle, shade-tile, etc,
940 * commands which may have been binned.
942 tri
->inputs
.disable
= TRUE
;
948 * Try to draw the triangle, restart the scene on failure.
950 static void retry_triangle_ccw( struct lp_setup_context
*setup
,
951 struct fixed_position
* position
,
952 const float (*v0
)[4],
953 const float (*v1
)[4],
954 const float (*v2
)[4],
957 if (!do_triangle_ccw( setup
, position
, v0
, v1
, v2
, front
))
959 if (!lp_setup_flush_and_restart(setup
))
962 if (!do_triangle_ccw( setup
, position
, v0
, v1
, v2
, front
))
968 * Calculate fixed position data for a triangle
969 * It is unfortunate we need to do that here (as we need area
970 * calculated in fixed point), as there's quite some code duplication
971 * to what is done in the jit setup prog.
974 calc_fixed_position(struct lp_setup_context
*setup
,
975 struct fixed_position
* position
,
976 const float (*v0
)[4],
977 const float (*v1
)[4],
978 const float (*v2
)[4])
981 * The rounding may not be quite the same with PIPE_ARCH_SSE
982 * (util_iround right now only does nearest/even on x87,
983 * otherwise nearest/away-from-zero).
984 * Both should be acceptable, I think.
986 #if defined(PIPE_ARCH_SSE)
987 __m128d v0r
, v1r
, v2r
;
988 __m128 vxy0xy2
, vxy1xy0
;
989 __m128i vxy0xy2i
, vxy1xy0i
;
990 __m128i dxdy0120
, x0x2y0y2
, x1x0y1y0
, x0120
, y0120
;
991 __m128 pix_offset
= _mm_set1_ps(setup
->pixel_offset
);
992 __m128 fixed_one
= _mm_set1_ps((float)FIXED_ONE
);
993 v0r
= _mm_load_sd((const double *)v0
[0]);
994 v1r
= _mm_load_sd((const double *)v1
[0]);
995 v2r
= _mm_load_sd((const double *)v2
[0]);
996 vxy0xy2
= (__m128
)_mm_unpacklo_pd(v0r
, v2r
);
997 vxy1xy0
= (__m128
)_mm_unpacklo_pd(v1r
, v0r
);
998 vxy0xy2
= _mm_sub_ps(vxy0xy2
, pix_offset
);
999 vxy1xy0
= _mm_sub_ps(vxy1xy0
, pix_offset
);
1000 vxy0xy2
= _mm_mul_ps(vxy0xy2
, fixed_one
);
1001 vxy1xy0
= _mm_mul_ps(vxy1xy0
, fixed_one
);
1002 vxy0xy2i
= _mm_cvtps_epi32(vxy0xy2
);
1003 vxy1xy0i
= _mm_cvtps_epi32(vxy1xy0
);
1004 dxdy0120
= _mm_sub_epi32(vxy0xy2i
, vxy1xy0i
);
1005 _mm_store_si128((__m128i
*)&position
->dx01
, dxdy0120
);
1007 * For the mul, would need some more shuffles, plus emulation
1008 * for the signed mul (without sse41), so don't bother.
1010 x0x2y0y2
= _mm_shuffle_epi32(vxy0xy2i
, _MM_SHUFFLE(3,1,2,0));
1011 x1x0y1y0
= _mm_shuffle_epi32(vxy1xy0i
, _MM_SHUFFLE(3,1,2,0));
1012 x0120
= _mm_unpacklo_epi32(x0x2y0y2
, x1x0y1y0
);
1013 y0120
= _mm_unpackhi_epi32(x0x2y0y2
, x1x0y1y0
);
1014 _mm_store_si128((__m128i
*)&position
->x
[0], x0120
);
1015 _mm_store_si128((__m128i
*)&position
->y
[0], y0120
);
1018 position
->x
[0] = subpixel_snap(v0
[0][0] - setup
->pixel_offset
);
1019 position
->x
[1] = subpixel_snap(v1
[0][0] - setup
->pixel_offset
);
1020 position
->x
[2] = subpixel_snap(v2
[0][0] - setup
->pixel_offset
);
1021 position
->x
[3] = 0; // should be unused
1023 position
->y
[0] = subpixel_snap(v0
[0][1] - setup
->pixel_offset
);
1024 position
->y
[1] = subpixel_snap(v1
[0][1] - setup
->pixel_offset
);
1025 position
->y
[2] = subpixel_snap(v2
[0][1] - setup
->pixel_offset
);
1026 position
->y
[3] = 0; // should be unused
1028 position
->dx01
= position
->x
[0] - position
->x
[1];
1029 position
->dy01
= position
->y
[0] - position
->y
[1];
1031 position
->dx20
= position
->x
[2] - position
->x
[0];
1032 position
->dy20
= position
->y
[2] - position
->y
[0];
1035 position
->area
= IMUL64(position
->dx01
, position
->dy20
) -
1036 IMUL64(position
->dx20
, position
->dy01
);
1041 * Rotate a triangle, flipping its clockwise direction,
1042 * Swaps values for xy[0] and xy[1]
1045 rotate_fixed_position_01( struct fixed_position
* position
)
1051 position
->x
[1] = position
->x
[0];
1052 position
->y
[1] = position
->y
[0];
1056 position
->dx01
= -position
->dx01
;
1057 position
->dy01
= -position
->dy01
;
1058 position
->dx20
= position
->x
[2] - position
->x
[0];
1059 position
->dy20
= position
->y
[2] - position
->y
[0];
1061 position
->area
= -position
->area
;
1066 * Rotate a triangle, flipping its clockwise direction,
1067 * Swaps values for xy[1] and xy[2]
1070 rotate_fixed_position_12( struct fixed_position
* position
)
1076 position
->x
[2] = position
->x
[1];
1077 position
->y
[2] = position
->y
[1];
1083 position
->dx01
= -position
->dx20
;
1084 position
->dy01
= -position
->dy20
;
1085 position
->dx20
= -x
;
1086 position
->dy20
= -y
;
1088 position
->area
= -position
->area
;
1093 * Draw triangle if it's CW, cull otherwise.
1095 static void triangle_cw(struct lp_setup_context
*setup
,
1096 const float (*v0
)[4],
1097 const float (*v1
)[4],
1098 const float (*v2
)[4])
1100 PIPE_ALIGN_VAR(16) struct fixed_position position
;
1102 calc_fixed_position(setup
, &position
, v0
, v1
, v2
);
1104 if (position
.area
< 0) {
1105 if (setup
->flatshade_first
) {
1106 rotate_fixed_position_12(&position
);
1107 retry_triangle_ccw(setup
, &position
, v0
, v2
, v1
, !setup
->ccw_is_frontface
);
1109 rotate_fixed_position_01(&position
);
1110 retry_triangle_ccw(setup
, &position
, v1
, v0
, v2
, !setup
->ccw_is_frontface
);
1116 static void triangle_ccw(struct lp_setup_context
*setup
,
1117 const float (*v0
)[4],
1118 const float (*v1
)[4],
1119 const float (*v2
)[4])
1121 PIPE_ALIGN_VAR(16) struct fixed_position position
;
1123 calc_fixed_position(setup
, &position
, v0
, v1
, v2
);
1125 if (position
.area
> 0)
1126 retry_triangle_ccw(setup
, &position
, v0
, v1
, v2
, setup
->ccw_is_frontface
);
1130 * Draw triangle whether it's CW or CCW.
1132 static void triangle_both(struct lp_setup_context
*setup
,
1133 const float (*v0
)[4],
1134 const float (*v1
)[4],
1135 const float (*v2
)[4])
1137 PIPE_ALIGN_VAR(16) struct fixed_position position
;
1138 struct llvmpipe_context
*lp_context
= (struct llvmpipe_context
*)setup
->pipe
;
1140 if (lp_context
->active_statistics_queries
&&
1141 !llvmpipe_rasterization_disabled(lp_context
)) {
1142 lp_context
->pipeline_statistics
.c_primitives
++;
1145 calc_fixed_position(setup
, &position
, v0
, v1
, v2
);
1148 assert(!util_is_inf_or_nan(v0
[0][0]));
1149 assert(!util_is_inf_or_nan(v0
[0][1]));
1150 assert(!util_is_inf_or_nan(v1
[0][0]));
1151 assert(!util_is_inf_or_nan(v1
[0][1]));
1152 assert(!util_is_inf_or_nan(v2
[0][0]));
1153 assert(!util_is_inf_or_nan(v2
[0][1]));
1156 if (position
.area
> 0)
1157 retry_triangle_ccw( setup
, &position
, v0
, v1
, v2
, setup
->ccw_is_frontface
);
1158 else if (position
.area
< 0) {
1159 if (setup
->flatshade_first
) {
1160 rotate_fixed_position_12( &position
);
1161 retry_triangle_ccw( setup
, &position
, v0
, v2
, v1
, !setup
->ccw_is_frontface
);
1163 rotate_fixed_position_01( &position
);
1164 retry_triangle_ccw( setup
, &position
, v1
, v0
, v2
, !setup
->ccw_is_frontface
);
1170 static void triangle_nop( struct lp_setup_context
*setup
,
1171 const float (*v0
)[4],
1172 const float (*v1
)[4],
1173 const float (*v2
)[4] )
1179 lp_setup_choose_triangle( struct lp_setup_context
*setup
)
1181 switch (setup
->cullmode
) {
1182 case PIPE_FACE_NONE
:
1183 setup
->triangle
= triangle_both
;
1185 case PIPE_FACE_BACK
:
1186 setup
->triangle
= setup
->ccw_is_frontface
? triangle_ccw
: triangle_cw
;
1188 case PIPE_FACE_FRONT
:
1189 setup
->triangle
= setup
->ccw_is_frontface
? triangle_cw
: triangle_ccw
;
1192 setup
->triangle
= triangle_nop
;