1 /**************************************************************************
2 *
3 * Copyright 2007-2010 VMware, Inc.
4 * All Rights Reserved.
5 *
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:
13 *
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
16 * of the Software.
17 *
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.
25 *
26 **************************************************************************/
27
28 /*
29 * Rasterization for binned triangles within a tile
30 */
31
32
33
34 /**
35 * Prototype for a 8 plane rasterizer function. Will codegenerate
36 * several of these.
37 *
38 * XXX: Varients for more/fewer planes.
39 * XXX: Need ways of dropping planes as we descend.
40 * XXX: SIMD
41 */
42 static void
TAG(do_block_4)43 TAG(do_block_4)(struct lp_rasterizer_task *task,
44 const struct lp_rast_triangle *tri,
45 const struct lp_rast_plane *plane,
46 int x, int y,
47 const int64_t *c)
48 {
49 #ifndef MULTISAMPLE
50 unsigned mask = 0xffff;
51 #else
52 uint64_t mask = UINT64_MAX;
53 #endif
54
55 for (unsigned j = 0; j < NR_PLANES; j++) {
56 #ifndef MULTISAMPLE
57 #ifdef RASTER_64
58 mask &= ~BUILD_MASK_LINEAR(((c[j] - 1) >> (int64_t)FIXED_ORDER),
59 -plane[j].dcdx >> FIXED_ORDER,
60 plane[j].dcdy >> FIXED_ORDER);
61 #else
62 mask &= ~BUILD_MASK_LINEAR((c[j] - 1),
63 -plane[j].dcdx,
64 plane[j].dcdy);
65 #endif
66 #else
67 for (unsigned s = 0; s < 4; s++) {
68 int64_t new_c = (c[j]) + ((IMUL64(task->scene->fixed_sample_pos[s][1], plane[j].dcdy) + IMUL64(task->scene->fixed_sample_pos[s][0], -plane[j].dcdx)) >> FIXED_ORDER);
69 uint32_t build_mask;
70 #ifdef RASTER_64
71 build_mask = BUILD_MASK_LINEAR((int32_t)((new_c - 1) >> (int64_t)FIXED_ORDER),
72 -plane[j].dcdx >> FIXED_ORDER,
73 plane[j].dcdy >> FIXED_ORDER);
74 #else
75 build_mask = BUILD_MASK_LINEAR((new_c - 1),
76 -plane[j].dcdx,
77 plane[j].dcdy);
78 #endif
79 mask &= ~((uint64_t)build_mask << (s * 16));
80 }
81 #endif
82 }
83
84 /* Now pass to the shader:
85 */
86 if (mask)
87 lp_rast_shade_quads_mask_sample(task, &tri->inputs, x, y, mask);
88 }
89
90
91 /**
92 * Evaluate a 16x16 block of pixels to determine which 4x4 subblocks are in/out
93 * of the triangle's bounds.
94 */
95 static void
TAG(do_block_16)96 TAG(do_block_16)(struct lp_rasterizer_task *task,
97 const struct lp_rast_triangle *tri,
98 const struct lp_rast_plane *plane,
99 int x, int y,
100 const int64_t *c)
101 {
102 unsigned outmask = 0; /* outside one or more trivial reject planes */
103 unsigned partmask = 0; /* outside one or more trivial accept planes */
104
105 for (unsigned j = 0; j < NR_PLANES; j++) {
106 #ifdef RASTER_64
107 int32_t dcdx = -plane[j].dcdx >> FIXED_ORDER;
108 int32_t dcdy = plane[j].dcdy >> FIXED_ORDER;
109 const int32_t cox = plane[j].eo >> FIXED_ORDER;
110 const int32_t ei = (dcdy + dcdx - cox) << 2;
111 const int32_t cox_s = cox << 2;
112 const int32_t co = (int32_t)(c[j] >> (int64_t)FIXED_ORDER) + cox_s;
113 int32_t cdiff;
114 cdiff = ei - cox_s + ((int32_t)((c[j] - 1) >> (int64_t)FIXED_ORDER) -
115 (int32_t)(c[j] >> (int64_t)FIXED_ORDER));
116 dcdx <<= 2;
117 dcdy <<= 2;
118 #else
119 const int64_t dcdx = -IMUL64(plane[j].dcdx, 4);
120 const int64_t dcdy = IMUL64(plane[j].dcdy, 4);
121 const int64_t cox = IMUL64(plane[j].eo, 4);
122 const int32_t ei = plane[j].dcdy - plane[j].dcdx - (int64_t)plane[j].eo;
123 const int64_t cio = IMUL64(ei, 4) - 1;
124 int32_t co, cdiff;
125 co = c[j] + cox;
126 cdiff = cio - cox;
127 #endif
128
129 BUILD_MASKS(co, cdiff,
130 dcdx, dcdy,
131 &outmask, /* sign bits from c[i][0..15] + cox */
132 &partmask); /* sign bits from c[i][0..15] + cio */
133 }
134
135 if (outmask == 0xffff)
136 return;
137
138 /* Mask of sub-blocks which are inside all trivial accept planes:
139 */
140 unsigned inmask = ~partmask & 0xffff;
141
142 /* Mask of sub-blocks which are inside all trivial reject planes,
143 * but outside at least one trivial accept plane:
144 */
145 unsigned partial_mask = partmask & ~outmask;
146
147 assert((partial_mask & inmask) == 0);
148
149 LP_COUNT_ADD(nr_empty_4, util_bitcount(0xffff & ~(partial_mask | inmask)));
150
151 /* Iterate over partials:
152 */
153 while (partial_mask) {
154 int i = ffs(partial_mask) - 1;
155 int ix = (i & 3) * 4;
156 int iy = (i >> 2) * 4;
157 int px = x + ix;
158 int py = y + iy;
159 int64_t cx[NR_PLANES];
160
161 partial_mask &= ~(1 << i);
162
163 LP_COUNT(nr_partially_covered_4);
164
165 for (unsigned j = 0; j < NR_PLANES; j++) {
166 cx[j] = (c[j]
167 - IMUL64(plane[j].dcdx, ix)
168 + IMUL64(plane[j].dcdy, iy));
169 }
170
171 TAG(do_block_4)(task, tri, plane, px, py, cx);
172 }
173
174 /* Iterate over fulls:
175 */
176 while (inmask) {
177 int i = ffs(inmask) - 1;
178 int ix = (i & 3) * 4;
179 int iy = (i >> 2) * 4;
180 int px = x + ix;
181 int py = y + iy;
182
183 inmask &= ~(1 << i);
184
185 LP_COUNT(nr_fully_covered_4);
186 block_full_4(task, tri, px, py);
187 }
188 }
189
190
191 /**
192 * Scan the tile in chunks and figure out which pixels to rasterize
193 * for this triangle.
194 */
195 void
TAG(lp_rast_triangle)196 TAG(lp_rast_triangle)(struct lp_rasterizer_task *task,
197 const union lp_rast_cmd_arg arg)
198 {
199 const struct lp_rast_triangle *tri = arg.triangle.tri;
200 unsigned plane_mask = arg.triangle.plane_mask;
201 const struct lp_rast_plane *tri_plane = GET_PLANES(tri);
202 const int x = task->x, y = task->y;
203 struct lp_rast_plane plane[NR_PLANES];
204 int64_t c[NR_PLANES];
205 unsigned outmask, inmask, partmask, partial_mask;
206 unsigned j = 0;
207
208 if (tri->inputs.disable) {
209 /* This triangle was partially binned and has been disabled */
210 return;
211 }
212
213 outmask = 0; /* outside one or more trivial reject planes */
214 partmask = 0; /* outside one or more trivial accept planes */
215
216 while (plane_mask) {
217 int i = ffs(plane_mask) - 1;
218 plane[j] = tri_plane[i];
219 plane_mask &= ~(1 << i);
220 c[j] = plane[j].c + IMUL64(plane[j].dcdy, y) - IMUL64(plane[j].dcdx, x);
221
222 {
223 #ifdef RASTER_64
224 /*
225 * Strip off lower FIXED_ORDER bits. Note that those bits from
226 * dcdx, dcdy, eo are always 0 (by definition).
227 * c values, however, are not. This means that for every
228 * addition of the form c + n*dcdx the lower FIXED_ORDER bits will
229 * NOT change. And those bits are not relevant to the sign bit (which
230 * is only what we need!) that is,
231 * sign(c + n*dcdx) == sign((c >> FIXED_ORDER) + n*(dcdx >> FIXED_ORDER))
232 * This means we can get away with using 32bit math for the most part.
233 * Only tricky part is the -1 adjustment for cdiff.
234 */
235 int32_t dcdx = -plane[j].dcdx >> FIXED_ORDER;
236 int32_t dcdy = plane[j].dcdy >> FIXED_ORDER;
237 const int32_t cox = plane[j].eo >> FIXED_ORDER;
238 const int32_t ei = (dcdy + dcdx - cox) << 4;
239 const int32_t cox_s = cox << 4;
240 const int32_t co = (int32_t)(c[j] >> (int64_t)FIXED_ORDER) + cox_s;
241 int32_t cdiff;
242 /*
243 * Plausibility check to ensure the 32bit math works.
244 * Note that within a tile, the max we can move the edge function
245 * is essentially dcdx * TILE_SIZE + dcdy * TILE_SIZE.
246 * TILE_SIZE is 64, dcdx/dcdy are nominally 21 bit (for 8192 max size
247 * and 8 subpixel bits), I'd be happy with 2 bits more too (1 for
248 * increasing fb size to 16384, the required d3d11 value, another one
249 * because I'm not quite sure we can't be _just_ above the max value
250 * here). This gives us 30 bits max - hence if c would exceed that here
251 * that means the plane is either trivial reject for the whole tile
252 * (in which case the tri will not get binned), or trivial accept for
253 * the whole tile (in which case plane_mask will not include it).
254 */
255 #if 0
256 assert((c[j] >> (int64_t)FIXED_ORDER) > (int32_t)0xb0000000 &&
257 (c[j] >> (int64_t)FIXED_ORDER) < (int32_t)0x3fffffff);
258 #endif
259 /*
260 * Note the fixup part is constant throughout the tile - thus could
261 * just calculate this and avoid _all_ 64bit math in rasterization
262 * (except exactly this fixup calc).
263 * In fact theoretically could move that even to setup, albeit that
264 * seems tricky (pre-bin certainly can have values larger than 32bit,
265 * and would need to communicate that fixup value through).
266 * And if we want to support msaa, we'd probably don't want to do the
267 * downscaling in setup in any case...
268 */
269 cdiff = ei - cox_s + ((int32_t)((c[j] - 1) >> (int64_t)FIXED_ORDER) -
270 (int32_t)(c[j] >> (int64_t)FIXED_ORDER));
271 dcdx <<= 4;
272 dcdy <<= 4;
273 #else
274 const int32_t dcdx = -plane[j].dcdx << 4;
275 const int32_t dcdy = plane[j].dcdy << 4;
276 const int32_t cox = plane[j].eo << 4;
277 const int32_t ei = plane[j].dcdy - plane[j].dcdx - (int32_t)plane[j].eo;
278 const int32_t cio = (ei << 4) - 1;
279 int32_t co, cdiff;
280 co = c[j] + cox;
281 cdiff = cio - cox;
282 #endif
283 BUILD_MASKS(co, cdiff,
284 dcdx, dcdy,
285 &outmask, /* sign bits from c[i][0..15] + cox */
286 &partmask); /* sign bits from c[i][0..15] + cio */
287 }
288
289 j++;
290 }
291
292 if (outmask == 0xffff)
293 return;
294
295 /* Mask of sub-blocks which are inside all trivial accept planes:
296 */
297 inmask = ~partmask & 0xffff;
298
299 /* Mask of sub-blocks which are inside all trivial reject planes,
300 * but outside at least one trivial accept plane:
301 */
302 partial_mask = partmask & ~outmask;
303
304 assert((partial_mask & inmask) == 0);
305
306 LP_COUNT_ADD(nr_empty_16, util_bitcount(0xffff & ~(partial_mask | inmask)));
307
308 /* Iterate over partials:
309 */
310 while (partial_mask) {
311 int i = ffs(partial_mask) - 1;
312 int ix = (i & 3) * 16;
313 int iy = (i >> 2) * 16;
314 int px = x + ix;
315 int py = y + iy;
316 int64_t cx[NR_PLANES];
317
318 for (j = 0; j < NR_PLANES; j++)
319 cx[j] = (c[j]
320 - IMUL64(plane[j].dcdx, ix)
321 + IMUL64(plane[j].dcdy, iy));
322
323 partial_mask &= ~(1 << i);
324
325 LP_COUNT(nr_partially_covered_16);
326 TAG(do_block_16)(task, tri, plane, px, py, cx);
327 }
328
329 /* Iterate over fulls:
330 */
331 while (inmask) {
332 int i = ffs(inmask) - 1;
333 int ix = (i & 3) * 16;
334 int iy = (i >> 2) * 16;
335 int px = x + ix;
336 int py = y + iy;
337
338 inmask &= ~(1 << i);
339
340 LP_COUNT(nr_fully_covered_16);
341 block_full_16(task, tri, px, py);
342 }
343 }
344
345
346 #if defined(PIPE_ARCH_SSE) && defined(TRI_16)
347 /* XXX: special case this when intersection is not required.
348 * - tile completely within bbox,
349 * - bbox completely within tile.
350 */
351 void
TRI_16(struct lp_rasterizer_task * task,const union lp_rast_cmd_arg arg)352 TRI_16(struct lp_rasterizer_task *task,
353 const union lp_rast_cmd_arg arg)
354 {
355 const struct lp_rast_triangle *tri = arg.triangle.tri;
356 const struct lp_rast_plane *plane = GET_PLANES(tri);
357 unsigned mask = arg.triangle.plane_mask;
358 __m128i cstep4[NR_PLANES][4];
359 int x = (mask & 0xff);
360 int y = (mask >> 8);
361 unsigned outmask = 0; /* outside one or more trivial reject planes */
362
363 if (x + 12 >= 64) {
364 int i = ((x + 12) - 64) / 4;
365 outmask |= right_mask_tab[i];
366 }
367
368 if (y + 12 >= 64) {
369 int i = ((y + 12) - 64) / 4;
370 outmask |= bottom_mask_tab[i];
371 }
372
373 x += task->x;
374 y += task->y;
375
376 for (unsigned j = 0; j < NR_PLANES; j++) {
377 const int dcdx = -plane[j].dcdx * 4;
378 const int dcdy = plane[j].dcdy * 4;
379 __m128i xdcdy = _mm_set1_epi32(dcdy);
380
381 cstep4[j][0] = _mm_setr_epi32(0, dcdx, dcdx*2, dcdx*3);
382 cstep4[j][1] = _mm_add_epi32(cstep4[j][0], xdcdy);
383 cstep4[j][2] = _mm_add_epi32(cstep4[j][1], xdcdy);
384 cstep4[j][3] = _mm_add_epi32(cstep4[j][2], xdcdy);
385
386 {
387 const int c = plane[j].c + plane[j].dcdy * y - plane[j].dcdx * x;
388 const int cox = plane[j].eo * 4;
389
390 outmask |= sign_bits4(cstep4[j], c + cox);
391 }
392 }
393
394 if (outmask == 0xffff)
395 return;
396
397
398 /* Mask of sub-blocks which are inside all trivial reject planes,
399 * but outside at least one trivial accept plane:
400 */
401 unsigned partial_mask = 0xffff & ~outmask;
402
403 /* Iterate over partials:
404 */
405 while (partial_mask) {
406 int i = ffs(partial_mask) - 1;
407 int ix = (i & 3) * 4;
408 int iy = (i >> 2) * 4;
409 int px = x + ix;
410 int py = y + iy;
411 unsigned mask = 0xffff;
412
413 partial_mask &= ~(1 << i);
414
415 for (unsigned j = 0; j < NR_PLANES; j++) {
416 const int cx = (plane[j].c - 1
417 - plane[j].dcdx * px
418 + plane[j].dcdy * py) * 4;
419
420 mask &= ~sign_bits4(cstep4[j], cx);
421 }
422
423 if (mask)
424 lp_rast_shade_quads_mask(task, &tri->inputs, px, py, mask);
425 }
426 }
427 #endif
428
429
430 #if defined(PIPE_ARCH_SSE) && defined(TRI_4)
431 void
TRI_4(struct lp_rasterizer_task * task,const union lp_rast_cmd_arg arg)432 TRI_4(struct lp_rasterizer_task *task,
433 const union lp_rast_cmd_arg arg)
434 {
435 const struct lp_rast_triangle *tri = arg.triangle.tri;
436 const struct lp_rast_plane *plane = GET_PLANES(tri);
437 unsigned mask = arg.triangle.plane_mask;
438 const int x = task->x + (mask & 0xff);
439 const int y = task->y + (mask >> 8);
440
441 /* Iterate over partials:
442 */
443 unsigned mask = 0xffff;
444
445 for (unsigned j = 0; j < NR_PLANES; j++) {
446 const int cx = (plane[j].c
447 - plane[j].dcdx * x
448 + plane[j].dcdy * y);
449
450 const int dcdx = -plane[j].dcdx;
451 const int dcdy = plane[j].dcdy;
452 __m128i xdcdy = _mm_set1_epi32(dcdy);
453
454 __m128i cstep0 = _mm_setr_epi32(cx, cx + dcdx, cx + dcdx*2, cx + dcdx*3);
455 __m128i cstep1 = _mm_add_epi32(cstep0, xdcdy);
456 __m128i cstep2 = _mm_add_epi32(cstep1, xdcdy);
457 __m128i cstep3 = _mm_add_epi32(cstep2, xdcdy);
458
459 __m128i cstep01 = _mm_packs_epi32(cstep0, cstep1);
460 __m128i cstep23 = _mm_packs_epi32(cstep2, cstep3);
461 __m128i result = _mm_packs_epi16(cstep01, cstep23);
462
463 /* Extract the sign bits
464 */
465 mask &= ~_mm_movemask_epi8(result);
466 }
467
468 if (mask)
469 lp_rast_shade_quads_mask(task, &tri->inputs, x, y, mask);
470 }
471 #endif
472
473
474 #undef TAG
475 #undef TRI_4
476 #undef TRI_16
477 #undef NR_PLANES
478