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"
46
#if defined(PIPE_ARCH_SSE)
47
#include <emmintrin.h>
48
#elif defined(_ARCH_PWR8) && UTIL_ARCH_LITTLE_ENDIAN
50
#include "util/u_pwr8.h"
53
#if !defined(PIPE_ARCH_SSE)
56
subpixel_snap(float a)
58
return util_iround(FIXED_ONE * a);
63
/* Position and area in fixed point coordinates */
64
struct fixed_position {
75
* Alloc space for a new triangle plus the input.a0/dadx/dady arrays
76
* immediately after it.
77
* The memory is allocated from the per-scene pool, not per-tile.
78
* \param tri_size returns number of bytes allocated
79
* \param num_inputs number of fragment shader inputs
80
* \return pointer to triangle space
82
struct lp_rast_triangle *
83
lp_setup_alloc_triangle(struct lp_scene *scene,
88
// add 1 for XYZW position
89
unsigned input_array_sz = (nr_inputs + 1) * sizeof(float[4]);
90
unsigned plane_sz = nr_planes * sizeof(struct lp_rast_plane);
92
STATIC_ASSERT(sizeof(struct lp_rast_plane) % 8 == 0);
94
*tri_size = (sizeof(struct lp_rast_triangle) +
95
3 * input_array_sz + // 3 = da + dadx + dady
98
struct lp_rast_triangle *tri = lp_scene_alloc_aligned(scene, *tri_size, 16);
102
tri->inputs.stride = input_array_sz;
105
ASSERTED char *a = (char *)tri;
106
ASSERTED char *b = (char *)&GET_PLANES(tri)[nr_planes];
108
assert(b - a == *tri_size);
115
lp_setup_print_vertex(struct lp_setup_context *setup,
119
const struct lp_setup_variant_key *key = &setup->setup.variant->key;
121
debug_printf(" wpos (%s[0]) xyzw %f %f %f %f\n",
123
v[0][0], v[0][1], v[0][2], v[0][3]);
125
for (int i = 0; i < key->num_inputs; i++) {
126
const float *in = v[key->inputs[i].src_index];
128
debug_printf(" in[%d] (%s[%d]) %s%s%s%s ",
130
name, key->inputs[i].src_index,
131
(key->inputs[i].usage_mask & 0x1) ? "x" : " ",
132
(key->inputs[i].usage_mask & 0x2) ? "y" : " ",
133
(key->inputs[i].usage_mask & 0x4) ? "z" : " ",
134
(key->inputs[i].usage_mask & 0x8) ? "w" : " ");
136
for (int j = 0; j < 4; j++)
137
if (key->inputs[i].usage_mask & (1<<j))
138
debug_printf("%.5f ", in[j]);
146
* Print triangle vertex attribs (for debug).
149
lp_setup_print_triangle(struct lp_setup_context *setup,
150
const float (*v0)[4],
151
const float (*v1)[4],
152
const float (*v2)[4])
154
debug_printf("triangle\n");
157
const float ex = v0[0][0] - v2[0][0];
158
const float ey = v0[0][1] - v2[0][1];
159
const float fx = v1[0][0] - v2[0][0];
160
const float fy = v1[0][1] - v2[0][1];
162
/* det = cross(e,f).z */
163
const float det = ex * fy - ey * fx;
165
debug_printf(" - ccw\n");
167
debug_printf(" - cw\n");
169
debug_printf(" - zero area\n");
172
lp_setup_print_vertex(setup, "v0", v0);
173
lp_setup_print_vertex(setup, "v1", v1);
174
lp_setup_print_vertex(setup, "v2", v2);
180
lp_rast_tri_tab[MAX_PLANES+1] = {
181
0, /* should be impossible */
182
LP_RAST_OP_TRIANGLE_1,
183
LP_RAST_OP_TRIANGLE_2,
184
LP_RAST_OP_TRIANGLE_3,
185
LP_RAST_OP_TRIANGLE_4,
186
LP_RAST_OP_TRIANGLE_5,
187
LP_RAST_OP_TRIANGLE_6,
188
LP_RAST_OP_TRIANGLE_7,
189
LP_RAST_OP_TRIANGLE_8
193
lp_rast_32_tri_tab[MAX_PLANES+1] = {
194
0, /* should be impossible */
195
LP_RAST_OP_TRIANGLE_32_1,
196
LP_RAST_OP_TRIANGLE_32_2,
197
LP_RAST_OP_TRIANGLE_32_3,
198
LP_RAST_OP_TRIANGLE_32_4,
199
LP_RAST_OP_TRIANGLE_32_5,
200
LP_RAST_OP_TRIANGLE_32_6,
201
LP_RAST_OP_TRIANGLE_32_7,
202
LP_RAST_OP_TRIANGLE_32_8
207
lp_rast_ms_tri_tab[MAX_PLANES+1] = {
208
0, /* should be impossible */
209
LP_RAST_OP_MS_TRIANGLE_1,
210
LP_RAST_OP_MS_TRIANGLE_2,
211
LP_RAST_OP_MS_TRIANGLE_3,
212
LP_RAST_OP_MS_TRIANGLE_4,
213
LP_RAST_OP_MS_TRIANGLE_5,
214
LP_RAST_OP_MS_TRIANGLE_6,
215
LP_RAST_OP_MS_TRIANGLE_7,
216
LP_RAST_OP_MS_TRIANGLE_8
221
* Detect big primitives drawn with an alpha == 1.0.
223
* This is used when simulating anti-aliasing primitives in shaders, e.g.,
224
* when drawing the windows client area in Aero's flip-3d effect.
227
check_opaque(const struct lp_setup_context *setup,
228
const float (*v1)[4],
229
const float (*v2)[4],
230
const float (*v3)[4])
232
const struct lp_fragment_shader_variant *variant =
233
setup->fs.current.variant;
238
if (!variant->potentially_opaque)
241
const struct lp_tgsi_channel_info *alpha_info = &variant->shader->info.cbuf[0][3];
242
if (alpha_info->file == TGSI_FILE_CONSTANT) {
243
const float *constants = setup->fs.current.jit_context.constants[0];
244
float alpha = constants[alpha_info->u.index*4 +
245
alpha_info->swizzle];
246
return alpha == 1.0f;
249
if (alpha_info->file == TGSI_FILE_INPUT) {
250
return (v1[1 + alpha_info->u.index][alpha_info->swizzle] == 1.0f &&
251
v2[1 + alpha_info->u.index][alpha_info->swizzle] == 1.0f &&
252
v3[1 + alpha_info->u.index][alpha_info->swizzle] == 1.0f);
260
* Do basic setup for triangle rasterization and determine which
261
* framebuffer tiles are touched. Put the triangle in the scene's
262
* bins for the tiles which we overlap.
265
do_triangle_ccw(struct lp_setup_context *setup,
266
struct fixed_position *position,
267
const float (*v0)[4],
268
const float (*v1)[4],
269
const float (*v2)[4],
272
struct lp_scene *scene = setup->scene;
275
lp_setup_print_triangle(setup, v0, v1, v2);
277
const float (*pv)[4];
278
if (setup->flatshade_first) {
284
unsigned viewport_index = 0;
285
if (setup->viewport_index_slot > 0) {
286
unsigned *udata = (unsigned*)pv[setup->viewport_index_slot];
287
viewport_index = lp_clamp_viewport_idx(*udata);
291
if (setup->layer_slot > 0) {
292
layer = *(unsigned*)pv[setup->layer_slot];
293
layer = MIN2(layer, scene->fb_max_layer);
296
/* Bounding rectangle (in pixels) */
299
/* Yes this is necessary to accurately calculate bounding boxes
300
* with the two fill-conventions we support. GL (normally) ends
301
* up needing a bottom-left fill convention, which requires
302
* slightly different rounding.
304
int adj = (setup->bottom_edge_rule != 0) ? 1 : 0;
306
/* Inclusive x0, exclusive x1 */
307
bbox.x0 = MIN3(position->x[0], position->x[1], position->x[2]) >> FIXED_ORDER;
308
bbox.x1 = (MAX3(position->x[0], position->x[1], position->x[2]) - 1) >> FIXED_ORDER;
310
/* Inclusive / exclusive depending upon adj (bottom-left or top-right) */
311
bbox.y0 = (MIN3(position->y[0], position->y[1], position->y[2]) + adj) >> FIXED_ORDER;
312
bbox.y1 = (MAX3(position->y[0], position->y[1], position->y[2]) - 1 + adj) >> FIXED_ORDER;
315
if (!u_rect_test_intersection(&setup->draw_regions[viewport_index], &bbox)) {
316
if (0) debug_printf("no intersection\n");
317
LP_COUNT(nr_culled_tris);
321
int max_szorig = ((bbox.x1 - (bbox.x0 & ~3)) |
322
(bbox.y1 - (bbox.y0 & ~3)));
323
boolean use_32bits = max_szorig <= MAX_FIXED_LENGTH32;
324
#if defined(_ARCH_PWR8) && UTIL_ARCH_LITTLE_ENDIAN
325
boolean pwr8_limit_check = (bbox.x1 - bbox.x0) <= MAX_FIXED_LENGTH32 &&
326
(bbox.y1 - bbox.y0) <= MAX_FIXED_LENGTH32;
329
/* Can safely discard negative regions, but need to keep hold of
330
* information about when the triangle extends past screen
331
* boundaries. See trimmed_box in lp_setup_bin_triangle().
333
bbox.x0 = MAX2(bbox.x0, 0);
334
bbox.y0 = MAX2(bbox.y0, 0);
339
* Determine how many scissor planes we need, that is drop scissor
340
* edges if the bounding box of the tri is fully inside that edge.
342
const struct u_rect *scissor = &setup->draw_regions[viewport_index];
344
scissor_planes_needed(s_planes, &bbox, scissor);
345
nr_planes += s_planes[0] + s_planes[1] + s_planes[2] + s_planes[3];
348
const struct lp_setup_variant_key *key = &setup->setup.variant->key;
349
struct lp_rast_triangle *tri =
350
lp_setup_alloc_triangle(scene, key->num_inputs, nr_planes, &tri_bytes);
355
tri->v[0][0] = v0[0][0];
356
tri->v[1][0] = v1[0][0];
357
tri->v[2][0] = v2[0][0];
358
tri->v[0][1] = v0[0][1];
359
tri->v[1][1] = v1[0][1];
360
tri->v[2][1] = v2[0][1];
366
* Rotate the tri such that v0 is closest to the fb origin.
367
* This can give more accurate a0 value (which is at fb origin)
368
* when calculating the interpolants.
369
* It can't work when there's flat shading for instance in one
370
* of the attributes, hence restrict this to just a single attribute
371
* which is what causes some test failures.
372
* (This does not address the problem that interpolation may be
373
* inaccurate if gradients are relatively steep in small tris far
374
* away from the origin. It does however fix the (silly) wgf11rasterizer
375
* Interpolator test.)
376
* XXX This causes problems with mipgen -EmuTexture for not yet really
377
* understood reasons (if the vertices would be submitted in a different
378
* order, we'd also generate the same "wrong" results here without
379
* rotation). In any case, that we generate different values if a prim
380
* has the vertices rotated but is otherwise the same (which is due to
381
* numerical issues) is not a nice property. An additional problem by
382
* swapping the vertices here (which is possibly worse) is that
383
* the same primitive coming in twice might generate different values
384
* (in particular for z) due to the swapping potentially not happening
385
* both times, if the attributes to be interpolated are different. For now,
386
* just restrict this to not get used with dx9 (by checking pixel offset),
387
* could also restrict it further to only trigger with wgf11Interpolator
388
* Rasterizer test (the only place which needs it, with always the same
391
if ((LP_DEBUG & DEBUG_ACCURATE_A0) &&
392
setup->pixel_offset == 0.5f &&
393
key->num_inputs == 1 &&
394
(key->inputs[0].interp == LP_INTERP_LINEAR ||
395
key->inputs[0].interp == LP_INTERP_PERSPECTIVE)) {
396
float dist0 = v0[0][0] * v0[0][0] + v0[0][1] * v0[0][1];
397
float dist1 = v1[0][0] * v1[0][0] + v1[0][1] * v1[0][1];
398
float dist2 = v2[0][0] * v2[0][0] + v2[0][1] * v2[0][1];
399
if (dist0 > dist1 && dist1 < dist2) {
400
const float (*vt)[4];
408
position->x[0] = position->x[1];
409
position->y[0] = position->y[1];
410
position->x[1] = position->x[2];
411
position->y[1] = position->y[2];
415
position->dx20 = position->dx01;
416
position->dy20 = position->dy01;
417
position->dx01 = position->x[0] - position->x[1];
418
position->dy01 = position->y[0] - position->y[1];
419
} else if (dist0 > dist2) {
420
const float (*vt)[4];
428
position->x[0] = position->x[2];
429
position->y[0] = position->y[2];
430
position->x[2] = position->x[1];
431
position->y[2] = position->y[1];
435
position->dx01 = position->dx20;
436
position->dy01 = position->dy20;
437
position->dx20 = position->x[2] - position->x[0];
438
position->dy20 = position->y[2] - position->y[0];
442
/* Setup parameter interpolants:
444
setup->setup.variant->jit_function(v0, v1, v2,
446
GET_A0(&tri->inputs),
447
GET_DADX(&tri->inputs),
448
GET_DADY(&tri->inputs),
449
&setup->setup.variant->key);
451
tri->inputs.frontfacing = frontfacing;
452
tri->inputs.disable = FALSE;
453
tri->inputs.is_blit = FALSE;
454
tri->inputs.layer = layer;
455
tri->inputs.viewport_index = viewport_index;
456
tri->inputs.view_index = setup->view_index;
459
lp_dump_setup_coef(&setup->setup.variant->key,
460
GET_A0(&tri->inputs),
461
GET_DADX(&tri->inputs),
462
GET_DADY(&tri->inputs));
464
struct lp_rast_plane *plane = GET_PLANES(tri);
466
#if defined(PIPE_ARCH_SSE)
468
__m128i vertx, verty;
469
__m128i shufx, shufy;
471
__m128i cdx02, cdx13, cdy02, cdy13, c02, c13;
472
__m128i c01, c23, unused;
473
__m128i dcdx_neg_mask;
474
__m128i dcdy_neg_mask;
475
__m128i dcdx_zero_mask;
476
__m128i top_left_flag, c_dec;
477
__m128i eo, p0, p1, p2;
478
__m128i zero = _mm_setzero_si128();
480
vertx = _mm_load_si128((__m128i *)position->x); /* vertex x coords */
481
verty = _mm_load_si128((__m128i *)position->y); /* vertex y coords */
483
shufx = _mm_shuffle_epi32(vertx, _MM_SHUFFLE(3,0,2,1));
484
shufy = _mm_shuffle_epi32(verty, _MM_SHUFFLE(3,0,2,1));
486
dcdx = _mm_sub_epi32(verty, shufy);
487
dcdy = _mm_sub_epi32(vertx, shufx);
489
dcdx_neg_mask = _mm_srai_epi32(dcdx, 31);
490
dcdx_zero_mask = _mm_cmpeq_epi32(dcdx, zero);
491
dcdy_neg_mask = _mm_srai_epi32(dcdy, 31);
493
top_left_flag = _mm_set1_epi32((setup->bottom_edge_rule == 0) ? ~0 : 0);
495
c_dec = _mm_or_si128(dcdx_neg_mask,
496
_mm_and_si128(dcdx_zero_mask,
497
_mm_xor_si128(dcdy_neg_mask,
502
* Note we need _signed_ mul (_mm_mul_epi32) which we emulate.
504
cdx02 = mm_mullohi_epi32(dcdx, vertx, &cdx13);
505
cdy02 = mm_mullohi_epi32(dcdy, verty, &cdy13);
506
c02 = _mm_sub_epi64(cdx02, cdy02);
507
c13 = _mm_sub_epi64(cdx13, cdy13);
508
c02 = _mm_sub_epi64(c02, _mm_shuffle_epi32(c_dec,
509
_MM_SHUFFLE(2,2,0,0)));
510
c13 = _mm_sub_epi64(c13, _mm_shuffle_epi32(c_dec,
511
_MM_SHUFFLE(3,3,1,1)));
514
* Useful for very small fbs/tris (or fewer subpixel bits) only:
515
* c = _mm_sub_epi32(mm_mullo_epi32(dcdx, vertx),
516
* mm_mullo_epi32(dcdy, verty));
518
* c = _mm_sub_epi32(c, c_dec);
521
/* Scale up to match c:
523
dcdx = _mm_slli_epi32(dcdx, FIXED_ORDER);
524
dcdy = _mm_slli_epi32(dcdy, FIXED_ORDER);
527
* Calculate trivial reject values:
528
* Note eo cannot overflow even if dcdx/dcdy would already have
529
* 31 bits (which they shouldn't have). This is because eo
530
* is never negative (albeit if we rely on that need to be careful...)
532
eo = _mm_sub_epi32(_mm_andnot_si128(dcdy_neg_mask, dcdy),
533
_mm_and_si128(dcdx_neg_mask, dcdx));
535
/* ei = _mm_sub_epi32(_mm_sub_epi32(dcdy, dcdx), eo); */
538
* Pointless transpose which gets undone immediately in
540
* It is actually difficult to do away with it - would essentially
541
* need GET_PLANES_DX, GET_PLANES_DY etc., but the calculations
542
* for this then would need to depend on the number of planes.
543
* The transpose is quite special here due to c being 64bit...
544
* The store has to be unaligned (unless we'd make the plane size
545
* a multiple of 128), and of course storing eo separately...
547
c01 = _mm_unpacklo_epi64(c02, c13);
548
c23 = _mm_unpackhi_epi64(c02, c13);
549
transpose2_64_2_32(&c01, &c23, &dcdx, &dcdy,
550
&p0, &p1, &p2, &unused);
551
_mm_storeu_si128((__m128i *)&plane[0], p0);
552
plane[0].eo = (uint32_t)_mm_cvtsi128_si32(eo);
553
_mm_storeu_si128((__m128i *)&plane[1], p1);
554
eo = _mm_shuffle_epi32(eo, _MM_SHUFFLE(3,2,0,1));
555
plane[1].eo = (uint32_t)_mm_cvtsi128_si32(eo);
556
_mm_storeu_si128((__m128i *)&plane[2], p2);
557
eo = _mm_shuffle_epi32(eo, _MM_SHUFFLE(0,0,0,2));
558
plane[2].eo = (uint32_t)_mm_cvtsi128_si32(eo);
560
#elif defined(_ARCH_PWR8) && UTIL_ARCH_LITTLE_ENDIAN
562
* XXX this code is effectively disabled for all practical purposes,
563
* as the allowed fb size is tiny if FIXED_ORDER is 8.
565
if (setup->fb.width <= MAX_FIXED_LENGTH32 &&
566
setup->fb.height <= MAX_FIXED_LENGTH32 &&
568
unsigned int bottom_edge;
569
__m128i vertx, verty;
570
__m128i shufx, shufy;
571
__m128i dcdx, dcdy, c;
573
__m128i dcdx_neg_mask;
574
__m128i dcdy_neg_mask;
575
__m128i dcdx_zero_mask;
576
__m128i top_left_flag;
577
__m128i c_inc_mask, c_inc;
578
__m128i eo, p0, p1, p2;
579
__m128i_union vshuf_mask;
580
__m128i zero = vec_splats((unsigned char) 0);
581
alignas(16) int32_t temp_vec[4];
583
#if UTIL_ARCH_LITTLE_ENDIAN
584
vshuf_mask.i[0] = 0x07060504;
585
vshuf_mask.i[1] = 0x0B0A0908;
586
vshuf_mask.i[2] = 0x03020100;
587
vshuf_mask.i[3] = 0x0F0E0D0C;
589
vshuf_mask.i[0] = 0x00010203;
590
vshuf_mask.i[1] = 0x0C0D0E0F;
591
vshuf_mask.i[2] = 0x04050607;
592
vshuf_mask.i[3] = 0x08090A0B;
595
/* vertex x coords */
596
vertx = vec_load_si128((const uint32_t *) position->x);
597
/* vertex y coords */
598
verty = vec_load_si128((const uint32_t *) position->y);
600
shufx = vec_perm (vertx, vertx, vshuf_mask.m128i);
601
shufy = vec_perm (verty, verty, vshuf_mask.m128i);
603
dcdx = vec_sub_epi32(verty, shufy);
604
dcdy = vec_sub_epi32(vertx, shufx);
606
dcdx_neg_mask = vec_srai_epi32(dcdx, 31);
607
dcdx_zero_mask = vec_cmpeq_epi32(dcdx, zero);
608
dcdy_neg_mask = vec_srai_epi32(dcdy, 31);
610
bottom_edge = (setup->bottom_edge_rule == 0) ? ~0 : 0;
611
top_left_flag = (__m128i) vec_splats(bottom_edge);
613
c_inc_mask = vec_or(dcdx_neg_mask,
614
vec_and(dcdx_zero_mask,
615
vec_xor(dcdy_neg_mask,
618
c_inc = vec_srli_epi32(c_inc_mask, 31);
620
c = vec_sub_epi32(vec_mullo_epi32(dcdx, vertx),
621
vec_mullo_epi32(dcdy, verty));
623
c = vec_add_epi32(c, c_inc);
625
/* Scale up to match c:
627
dcdx = vec_slli_epi32(dcdx, FIXED_ORDER);
628
dcdy = vec_slli_epi32(dcdy, FIXED_ORDER);
630
/* Calculate trivial reject values:
632
eo = vec_sub_epi32(vec_andnot_si128(dcdy_neg_mask, dcdy),
633
vec_and(dcdx_neg_mask, dcdx));
635
/* ei = _mm_sub_epi32(_mm_sub_epi32(dcdy, dcdx), eo); */
637
/* Pointless transpose which gets undone immediately in
640
transpose4_epi32(&c, &dcdx, &dcdy, &eo,
641
&p0, &p1, &p2, &unused);
643
#define STORE_PLANE(plane, vec) do { \
644
vec_store_si128((uint32_t *)&temp_vec, vec); \
645
plane.c = (int64_t)temp_vec[0]; \
646
plane.dcdx = temp_vec[1]; \
647
plane.dcdy = temp_vec[2]; \
648
plane.eo = temp_vec[3]; \
651
STORE_PLANE(plane[0], p0);
652
STORE_PLANE(plane[1], p1);
653
STORE_PLANE(plane[2], p2);
658
plane[0].dcdy = position->dx01;
659
plane[1].dcdy = position->x[1] - position->x[2];
660
plane[2].dcdy = position->dx20;
661
plane[0].dcdx = position->dy01;
662
plane[1].dcdx = position->y[1] - position->y[2];
663
plane[2].dcdx = position->dy20;
665
for (int i = 0; i < 3; i++) {
666
/* half-edge constants, will be iterated over the whole render
669
plane[i].c = IMUL64(plane[i].dcdx, position->x[i]) -
670
IMUL64(plane[i].dcdy, position->y[i]);
672
/* correct for top-left vs. bottom-left fill convention.
674
if (plane[i].dcdx < 0) {
675
/* both fill conventions want this - adjust for left edges */
678
else if (plane[i].dcdx == 0) {
679
if (setup->bottom_edge_rule == 0) {
680
/* correct for top-left fill convention:
682
if (plane[i].dcdy > 0)
685
/* correct for bottom-left fill convention:
687
if (plane[i].dcdy < 0)
692
/* Scale up to match c:
694
assert((plane[i].dcdx << FIXED_ORDER) >> FIXED_ORDER == plane[i].dcdx);
695
assert((plane[i].dcdy << FIXED_ORDER) >> FIXED_ORDER == plane[i].dcdy);
696
plane[i].dcdx <<= FIXED_ORDER;
697
plane[i].dcdy <<= FIXED_ORDER;
699
/* find trivial reject offsets for each edge for a single-pixel
700
* sized block. These will be scaled up at each recursive level to
701
* match the active blocksize. Scaling in this way works best if
702
* the blocks are square.
705
if (plane[i].dcdx < 0) plane[i].eo -= plane[i].dcdx;
706
if (plane[i].dcdy > 0) plane[i].eo += plane[i].dcdy;
711
debug_printf("p0: %"PRIx64"/%08x/%08x/%08x\n",
717
debug_printf("p1: %"PRIx64"/%08x/%08x/%08x\n",
723
debug_printf("p2: %"PRIx64"/%08x/%08x/%08x\n",
731
lp_setup_add_scissor_planes(scissor, &plane[3], s_planes, setup->multisample);
734
return lp_setup_bin_triangle(setup, tri, use_32bits,
735
check_opaque(setup, v0, v1, v2),
736
&bbox, nr_planes, viewport_index);
740
* Round to nearest less or equal power of two of the input.
742
* Undefined if no bit set exists, so code should check against 0 first.
744
static inline uint32_t
745
floor_pot(uint32_t n)
747
#if defined(PIPE_CC_GCC) && (defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64))
768
lp_setup_bin_triangle(struct lp_setup_context *setup,
769
struct lp_rast_triangle *tri,
772
const struct u_rect *bbox,
774
unsigned viewport_index)
776
struct lp_scene *scene = setup->scene;
779
/* What is the largest power-of-two boundary this triangle crosses:
781
const int dx = floor_pot((bbox->x0 ^ bbox->x1) |
782
(bbox->y0 ^ bbox->y1));
784
/* The largest dimension of the rasterized area of the triangle
785
* (aligned to a 4x4 grid), rounded down to the nearest power of two:
787
const int max_sz = ((bbox->x1 - (bbox->x0 & ~3)) |
788
(bbox->y1 - (bbox->y0 & ~3)));
789
const int sz = floor_pot(max_sz);
792
* NOTE: It is important to use the original bounding box
793
* which might contain negative values here, because if the
794
* plane math may overflow or not with the 32bit rasterization
795
* functions depends on the original extent of the triangle.
798
/* Now apply scissor, etc to the bounding box. Could do this
799
* earlier, but it confuses the logic for tri-16 and would force
800
* the rasterizer to also respect scissor, etc, just for the rare
801
* cases where a small triangle extends beyond the scissor.
803
struct u_rect trimmed_box = *bbox;
804
u_rect_find_intersection(&setup->draw_regions[viewport_index],
807
/* Determine which tile(s) intersect the triangle's bounding box
809
if (dx < TILE_SIZE) {
810
const int ix0 = bbox->x0 / TILE_SIZE;
811
const int iy0 = bbox->y0 / TILE_SIZE;
812
unsigned px = bbox->x0 & 63 & ~3;
813
unsigned py = bbox->y0 & 63 & ~3;
815
assert(iy0 == bbox->y1 / TILE_SIZE &&
816
ix0 == bbox->x1 / TILE_SIZE);
818
if (nr_planes == 3) {
820
/* Triangle is contained in a single 4x4 stamp:
822
assert(px + 4 <= TILE_SIZE);
823
assert(py + 4 <= TILE_SIZE);
824
if (setup->multisample)
825
cmd = LP_RAST_OP_MS_TRIANGLE_3_4;
827
cmd = use_32bits ? LP_RAST_OP_TRIANGLE_32_3_4 : LP_RAST_OP_TRIANGLE_3_4;
828
return lp_scene_bin_cmd_with_state(scene, ix0, iy0,
829
setup->fs.stored, cmd,
830
lp_rast_arg_triangle_contained(tri, px, py));
834
/* Triangle is contained in a single 16x16 block:
838
* The 16x16 block is only 4x4 aligned, and can exceed the tile
839
* dimensions if the triangle is 16 pixels in one dimension but 4
840
* in the other. So budge the 16x16 back inside the tile.
842
px = MIN2(px, TILE_SIZE - 16);
843
py = MIN2(py, TILE_SIZE - 16);
845
assert(px + 16 <= TILE_SIZE);
846
assert(py + 16 <= TILE_SIZE);
848
if (setup->multisample)
849
cmd = LP_RAST_OP_MS_TRIANGLE_3_16;
851
cmd = use_32bits ? LP_RAST_OP_TRIANGLE_32_3_16 : LP_RAST_OP_TRIANGLE_3_16;
852
return lp_scene_bin_cmd_with_state(scene, ix0, iy0,
853
setup->fs.stored, cmd,
854
lp_rast_arg_triangle_contained(tri, px, py));
856
} else if (nr_planes == 4 && sz < 16) {
857
px = MIN2(px, TILE_SIZE - 16);
858
py = MIN2(py, TILE_SIZE - 16);
860
assert(px + 16 <= TILE_SIZE);
861
assert(py + 16 <= TILE_SIZE);
863
if (setup->multisample)
864
cmd = LP_RAST_OP_MS_TRIANGLE_4_16;
866
cmd = use_32bits ? LP_RAST_OP_TRIANGLE_32_4_16 : LP_RAST_OP_TRIANGLE_4_16;
867
return lp_scene_bin_cmd_with_state(scene, ix0, iy0,
868
setup->fs.stored, cmd,
869
lp_rast_arg_triangle_contained(tri, px, py));
872
/* Triangle is contained in a single tile:
874
if (setup->multisample)
875
cmd = lp_rast_ms_tri_tab[nr_planes];
877
cmd = use_32bits ? lp_rast_32_tri_tab[nr_planes] : lp_rast_tri_tab[nr_planes];
878
return lp_scene_bin_cmd_with_state(scene, ix0, iy0, setup->fs.stored, cmd,
879
lp_rast_arg_triangle(tri, (1<<nr_planes)-1));
881
struct lp_rast_plane *plane = GET_PLANES(tri);
882
int64_t c[MAX_PLANES];
883
int64_t ei[MAX_PLANES];
885
int64_t eo[MAX_PLANES];
886
int64_t xstep[MAX_PLANES];
887
int64_t ystep[MAX_PLANES];
890
const int ix0 = trimmed_box.x0 / TILE_SIZE;
891
const int iy0 = trimmed_box.y0 / TILE_SIZE;
892
const int ix1 = trimmed_box.x1 / TILE_SIZE;
893
const int iy1 = trimmed_box.y1 / TILE_SIZE;
895
for (int i = 0; i < nr_planes; i++) {
897
IMUL64(plane[i].dcdy, iy0) * TILE_SIZE -
898
IMUL64(plane[i].dcdx, ix0) * TILE_SIZE);
900
ei[i] = (plane[i].dcdy -
902
(int64_t)plane[i].eo) << TILE_ORDER;
904
eo[i] = (int64_t)plane[i].eo << TILE_ORDER;
905
xstep[i] = -(((int64_t)plane[i].dcdx) << TILE_ORDER);
906
ystep[i] = ((int64_t)plane[i].dcdy) << TILE_ORDER;
909
tri->inputs.is_blit = lp_setup_is_blit(setup, &tri->inputs);
911
/* Test tile-sized blocks against the triangle.
912
* Discard blocks fully outside the tri. If the block is fully
913
* contained inside the tri, bin an lp_rast_shade_tile command.
914
* Else, bin a lp_rast_triangle command.
916
for (y = iy0; y <= iy1; y++) {
917
boolean in = FALSE; /* are we inside the triangle? */
918
int64_t cx[MAX_PLANES];
920
for (int i = 0; i < nr_planes; i++)
923
for (x = ix0; x <= ix1; x++) {
927
for (int i = 0; i < nr_planes; i++) {
928
int64_t planeout = cx[i] + eo[i];
929
int64_t planepartial = cx[i] + ei[i] - 1;
930
out |= (int) (planeout >> 63);
931
partial |= ((int) (planepartial >> 63)) & (1<<i);
937
break; /* exiting triangle, all done with this row */
938
LP_COUNT(nr_empty_64);
939
} else if (partial) {
940
/* Not trivially accepted by at least one plane -
941
* rasterize/shade partial tile
943
int count = util_bitcount(partial);
946
if (setup->multisample)
947
cmd = lp_rast_ms_tri_tab[count];
949
cmd = use_32bits ? lp_rast_32_tri_tab[count] : lp_rast_tri_tab[count];
950
if (!lp_scene_bin_cmd_with_state(scene, x, y,
951
setup->fs.stored, cmd,
952
lp_rast_arg_triangle(tri, partial)))
955
LP_COUNT(nr_partially_covered_64);
957
/* triangle covers the whole tile- shade whole tile */
958
LP_COUNT(nr_fully_covered_64);
960
if (!lp_setup_whole_tile(setup, &tri->inputs, x, y, opaque))
964
/* Iterate cx values across the region: */
965
for (int i = 0; i < nr_planes; i++)
969
/* Iterate c values down the region: */
970
for (int i = 0; i < nr_planes; i++)
978
/* Need to disable any partially binned triangle. This is easier
979
* than trying to locate all the triangle, shade-tile, etc,
980
* commands which may have been binned.
982
tri->inputs.disable = TRUE;
988
* Try to draw the triangle, restart the scene on failure.
991
retry_triangle_ccw(struct lp_setup_context *setup,
992
struct fixed_position *position,
993
const float (*v0)[4],
994
const float (*v1)[4],
995
const float (*v2)[4],
998
if (!do_triangle_ccw(setup, position, v0, v1, v2, front)) {
999
if (!lp_setup_flush_and_restart(setup))
1002
if (!do_triangle_ccw(setup, position, v0, v1, v2, front))
1009
* Calculate fixed position data for a triangle
1010
* It is unfortunate we need to do that here (as we need area
1011
* calculated in fixed point), as there's quite some code duplication
1012
* to what is done in the jit setup prog.
1014
static inline int8_t
1015
calc_fixed_position(struct lp_setup_context *setup,
1016
struct fixed_position* position,
1017
const float (*v0)[4],
1018
const float (*v1)[4],
1019
const float (*v2)[4])
1021
float pixel_offset = setup->multisample ? 0.0 : setup->pixel_offset;
1023
* The rounding may not be quite the same with PIPE_ARCH_SSE
1024
* (util_iround right now only does nearest/even on x87,
1025
* otherwise nearest/away-from-zero).
1026
* Both should be acceptable, I think.
1028
#if defined(PIPE_ARCH_SSE)
1030
__m128 vxy0xy2, vxy1xy0;
1031
__m128i vxy0xy2i, vxy1xy0i;
1032
__m128i dxdy0120, x0x2y0y2, x1x0y1y0, x0120, y0120;
1033
__m128 pix_offset = _mm_set1_ps(pixel_offset);
1034
__m128 fixed_one = _mm_set1_ps((float)FIXED_ONE);
1035
v0r = _mm_castpd_ps(_mm_load_sd((double *)v0[0]));
1036
vxy0xy2 = _mm_loadh_pi(v0r, (__m64 *)v2[0]);
1037
v1r = _mm_castpd_ps(_mm_load_sd((double *)v1[0]));
1038
vxy1xy0 = _mm_movelh_ps(v1r, vxy0xy2);
1039
vxy0xy2 = _mm_sub_ps(vxy0xy2, pix_offset);
1040
vxy1xy0 = _mm_sub_ps(vxy1xy0, pix_offset);
1041
vxy0xy2 = _mm_mul_ps(vxy0xy2, fixed_one);
1042
vxy1xy0 = _mm_mul_ps(vxy1xy0, fixed_one);
1043
vxy0xy2i = _mm_cvtps_epi32(vxy0xy2);
1044
vxy1xy0i = _mm_cvtps_epi32(vxy1xy0);
1045
dxdy0120 = _mm_sub_epi32(vxy0xy2i, vxy1xy0i);
1046
_mm_store_si128((__m128i *)&position->dx01, dxdy0120);
1048
* For the mul, would need some more shuffles, plus emulation
1049
* for the signed mul (without sse41), so don't bother.
1051
x0x2y0y2 = _mm_shuffle_epi32(vxy0xy2i, _MM_SHUFFLE(3,1,2,0));
1052
x1x0y1y0 = _mm_shuffle_epi32(vxy1xy0i, _MM_SHUFFLE(3,1,2,0));
1053
x0120 = _mm_unpacklo_epi32(x0x2y0y2, x1x0y1y0);
1054
y0120 = _mm_unpackhi_epi32(x0x2y0y2, x1x0y1y0);
1055
_mm_store_si128((__m128i *)&position->x[0], x0120);
1056
_mm_store_si128((__m128i *)&position->y[0], y0120);
1059
position->x[0] = subpixel_snap(v0[0][0] - pixel_offset);
1060
position->x[1] = subpixel_snap(v1[0][0] - pixel_offset);
1061
position->x[2] = subpixel_snap(v2[0][0] - pixel_offset);
1062
position->x[3] = 0; // should be unused
1064
position->y[0] = subpixel_snap(v0[0][1] - pixel_offset);
1065
position->y[1] = subpixel_snap(v1[0][1] - pixel_offset);
1066
position->y[2] = subpixel_snap(v2[0][1] - pixel_offset);
1067
position->y[3] = 0; // should be unused
1069
position->dx01 = position->x[0] - position->x[1];
1070
position->dy01 = position->y[0] - position->y[1];
1072
position->dx20 = position->x[2] - position->x[0];
1073
position->dy20 = position->y[2] - position->y[0];
1076
uint64_t area = IMUL64(position->dx01, position->dy20) -
1077
IMUL64(position->dx20, position->dy01);
1078
return area == 0 ? 0 : (area & (1ULL << 63)) ? -1 : 1;
1083
* Rotate a triangle, flipping its clockwise direction,
1084
* Swaps values for xy[0] and xy[1]
1087
rotate_fixed_position_01(struct fixed_position* position)
1089
int x = position->x[1];
1090
int y = position->y[1];
1092
position->x[1] = position->x[0];
1093
position->y[1] = position->y[0];
1097
position->dx01 = -position->dx01;
1098
position->dy01 = -position->dy01;
1099
position->dx20 = position->x[2] - position->x[0];
1100
position->dy20 = position->y[2] - position->y[0];
1105
* Rotate a triangle, flipping its clockwise direction,
1106
* Swaps values for xy[1] and xy[2]
1109
rotate_fixed_position_12(struct fixed_position* position)
1111
int x = position->x[2];
1112
int y = position->y[2];
1114
position->x[2] = position->x[1];
1115
position->y[2] = position->y[1];
1121
position->dx01 = -position->dx20;
1122
position->dy01 = -position->dy20;
1123
position->dx20 = -x;
1124
position->dy20 = -y;
1129
* Draw triangle if it's CW, cull otherwise.
1132
triangle_cw(struct lp_setup_context *setup,
1133
const float (*v0)[4],
1134
const float (*v1)[4],
1135
const float (*v2)[4])
1137
alignas(16) struct fixed_position position;
1138
struct llvmpipe_context *lp_context = (struct llvmpipe_context *)setup->pipe;
1140
if (lp_context->active_statistics_queries) {
1141
lp_context->pipeline_statistics.c_primitives++;
1144
int8_t area_sign = calc_fixed_position(setup, &position, v0, v1, v2);
1146
if (area_sign < 0) {
1147
if (setup->flatshade_first) {
1148
rotate_fixed_position_12(&position);
1149
retry_triangle_ccw(setup, &position, v0, v2, v1, !setup->ccw_is_frontface);
1151
rotate_fixed_position_01(&position);
1152
retry_triangle_ccw(setup, &position, v1, v0, v2, !setup->ccw_is_frontface);
1159
triangle_ccw(struct lp_setup_context *setup,
1160
const float (*v0)[4],
1161
const float (*v1)[4],
1162
const float (*v2)[4])
1164
alignas(16) struct fixed_position position;
1165
struct llvmpipe_context *lp_context = (struct llvmpipe_context *)setup->pipe;
1167
if (lp_context->active_statistics_queries) {
1168
lp_context->pipeline_statistics.c_primitives++;
1171
int8_t area_sign = calc_fixed_position(setup, &position, v0, v1, v2);
1174
retry_triangle_ccw(setup, &position, v0, v1, v2, setup->ccw_is_frontface);
1179
* Draw triangle whether it's CW or CCW.
1182
triangle_both(struct lp_setup_context *setup,
1183
const float (*v0)[4],
1184
const float (*v1)[4],
1185
const float (*v2)[4])
1187
alignas(16) struct fixed_position position;
1188
struct llvmpipe_context *lp_context = (struct llvmpipe_context *)setup->pipe;
1190
if (lp_context->active_statistics_queries) {
1191
lp_context->pipeline_statistics.c_primitives++;
1194
int8_t area_sign = calc_fixed_position(setup, &position, v0, v1, v2);
1197
assert(!util_is_inf_or_nan(v0[0][0]));
1198
assert(!util_is_inf_or_nan(v0[0][1]));
1199
assert(!util_is_inf_or_nan(v1[0][0]));
1200
assert(!util_is_inf_or_nan(v1[0][1]));
1201
assert(!util_is_inf_or_nan(v2[0][0]));
1202
assert(!util_is_inf_or_nan(v2[0][1]));
1205
if (area_sign > 0) {
1206
retry_triangle_ccw(setup, &position, v0, v1, v2, setup->ccw_is_frontface);
1207
} else if (area_sign < 0) {
1208
if (setup->flatshade_first) {
1209
rotate_fixed_position_12(&position);
1210
retry_triangle_ccw(setup, &position, v0, v2, v1, !setup->ccw_is_frontface);
1212
rotate_fixed_position_01(&position);
1213
retry_triangle_ccw(setup, &position, v1, v0, v2, !setup->ccw_is_frontface);
1220
triangle_noop(struct lp_setup_context *setup,
1221
const float (*v0)[4],
1222
const float (*v1)[4],
1223
const float (*v2)[4])
1229
lp_setup_choose_triangle(struct lp_setup_context *setup)
1231
if (setup->rasterizer_discard) {
1232
setup->triangle = triangle_noop;
1235
switch (setup->cullmode) {
1236
case PIPE_FACE_NONE:
1237
setup->triangle = triangle_both;
1239
case PIPE_FACE_BACK:
1240
setup->triangle = setup->ccw_is_frontface ? triangle_ccw : triangle_cw;
1242
case PIPE_FACE_FRONT:
1243
setup->triangle = setup->ccw_is_frontface ? triangle_cw : triangle_ccw;
1246
setup->triangle = triangle_noop;
added
removed
src/gallium/drivers/llvmpipe/lp_setup_tri.c
