projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
r300/compiler: implement DP2 opcode
[mesa.git] / src / gallium / drivers / nv50 / nv50_state_validate.c
1 /*
2 * Copyright 2008 Ben Skeggs
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
13 *
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
18 * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF
19 * OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
20 * SOFTWARE.
21 */
22
23 #include "util/u_format.h"
24
25 #include "nv50_context.h"
26 #include "nv50_resource.h"
27 #include "nouveau/nouveau_stateobj.h"
28
29 static struct nouveau_stateobj *
30 validate_fb(struct nv50_context *nv50)
31 {
32 struct nouveau_grobj *tesla = nv50->screen->tesla;
33 struct nouveau_stateobj *so = so_new(32, 79, 18);
34 struct pipe_framebuffer_state *fb = &nv50->framebuffer;
35 unsigned i, w = 0, h = 0, gw = 0;
36
37 /* Set nr of active RTs and select RT for each colour output.
38 * FP result 0 always goes to RT[0], bits 4 - 6 are ignored.
39 * Ambiguous assignment results in no rendering (no DATA_ERROR).
40 */
41 so_method(so, tesla, NV50TCL_RT_CONTROL, 1);
42 so_data (so, fb->nr_cbufs |
43 (0 << 4) | (1 << 7) | (2 << 10) | (3 << 13) |
44 (4 << 16) | (5 << 19) | (6 << 22) | (7 << 25));
45
46 for (i = 0; i < fb->nr_cbufs; i++) {
47 struct pipe_resource *pt = fb->cbufs[i]->texture;
48 struct nouveau_bo *bo = nv50_miptree(pt)->base.bo;
49
50 if (!gw) {
51 w = fb->cbufs[i]->width;
52 h = fb->cbufs[i]->height;
53 gw = 1;
54 } else {
55 assert(w == fb->cbufs[i]->width);
56 assert(h == fb->cbufs[i]->height);
57 }
58
59 so_method(so, tesla, NV50TCL_RT_HORIZ(i), 2);
60 so_data (so, fb->cbufs[i]->width);
61 so_data (so, fb->cbufs[i]->height);
62
63 so_method(so, tesla, NV50TCL_RT_ADDRESS_HIGH(i), 5);
64 so_reloc (so, bo, fb->cbufs[i]->offset, NOUVEAU_BO_VRAM |
65 NOUVEAU_BO_HIGH | NOUVEAU_BO_RDWR, 0, 0);
66 so_reloc (so, bo, fb->cbufs[i]->offset, NOUVEAU_BO_VRAM |
67 NOUVEAU_BO_LOW | NOUVEAU_BO_RDWR, 0, 0);
68 switch (fb->cbufs[i]->format) {
69 case PIPE_FORMAT_B8G8R8A8_UNORM:
70 so_data(so, NV50TCL_RT_FORMAT_A8R8G8B8_UNORM);
71 break;
72 case PIPE_FORMAT_B8G8R8X8_UNORM:
73 so_data(so, NV50TCL_RT_FORMAT_X8R8G8B8_UNORM);
74 break;
75 case PIPE_FORMAT_B5G6R5_UNORM:
76 so_data(so, NV50TCL_RT_FORMAT_R5G6B5_UNORM);
77 break;
78 case PIPE_FORMAT_R16G16B16A16_SNORM:
79 so_data(so, NV50TCL_RT_FORMAT_R16G16B16A16_SNORM);
80 break;
81 case PIPE_FORMAT_R16G16B16A16_UNORM:
82 so_data(so, NV50TCL_RT_FORMAT_R16G16B16A16_UNORM);
83 break;
84 case PIPE_FORMAT_R32G32B32A32_FLOAT:
85 so_data(so, NV50TCL_RT_FORMAT_R32G32B32A32_FLOAT);
86 break;
87 case PIPE_FORMAT_R16G16_SNORM:
88 so_data(so, NV50TCL_RT_FORMAT_R16G16_SNORM);
89 break;
90 case PIPE_FORMAT_R16G16_UNORM:
91 so_data(so, NV50TCL_RT_FORMAT_R16G16_UNORM);
92 break;
93 default:
94 NOUVEAU_ERR("AIIII unknown format %s\n",
95 util_format_name(fb->cbufs[i]->format));
96 so_data(so, NV50TCL_RT_FORMAT_X8R8G8B8_UNORM);
97 break;
98 }
99 so_data(so, nv50_miptree(pt)->
100 level[fb->cbufs[i]->level].tile_mode << 4);
101 so_data(so, 0x00000000);
102
103 so_method(so, tesla, NV50TCL_RT_ARRAY_MODE, 1);
104 so_data (so, 1);
105 }
106
107 if (fb->zsbuf) {
108 struct pipe_resource *pt = fb->zsbuf->texture;
109 struct nouveau_bo *bo = nv50_miptree(pt)->base.bo;
110
111 if (!gw) {
112 w = fb->zsbuf->width;
113 h = fb->zsbuf->height;
114 gw = 1;
115 } else {
116 assert(w == fb->zsbuf->width);
117 assert(h == fb->zsbuf->height);
118 }
119
120 so_method(so, tesla, NV50TCL_ZETA_ADDRESS_HIGH, 5);
121 so_reloc (so, bo, fb->zsbuf->offset, NOUVEAU_BO_VRAM |
122 NOUVEAU_BO_HIGH | NOUVEAU_BO_RDWR, 0, 0);
123 so_reloc (so, bo, fb->zsbuf->offset, NOUVEAU_BO_VRAM |
124 NOUVEAU_BO_LOW | NOUVEAU_BO_RDWR, 0, 0);
125 switch (fb->zsbuf->format) {
126 case PIPE_FORMAT_Z24_UNORM_S8_USCALED:
127 so_data(so, NV50TCL_ZETA_FORMAT_S8Z24_UNORM);
128 break;
129 case PIPE_FORMAT_Z24X8_UNORM:
130 so_data(so, NV50TCL_ZETA_FORMAT_X8Z24_UNORM);
131 break;
132 case PIPE_FORMAT_S8_USCALED_Z24_UNORM:
133 so_data(so, NV50TCL_ZETA_FORMAT_Z24S8_UNORM);
134 break;
135 case PIPE_FORMAT_Z32_FLOAT:
136 so_data(so, NV50TCL_ZETA_FORMAT_Z32_FLOAT);
137 break;
138 default:
139 NOUVEAU_ERR("AIIII unknown format %s\n",
140 util_format_name(fb->zsbuf->format));
141 so_data(so, NV50TCL_ZETA_FORMAT_S8Z24_UNORM);
142 break;
143 }
144 so_data(so, nv50_miptree(pt)->
145 level[fb->zsbuf->level].tile_mode << 4);
146 so_data(so, 0x00000000);
147
148 so_method(so, tesla, NV50TCL_ZETA_ENABLE, 1);
149 so_data (so, 1);
150 so_method(so, tesla, NV50TCL_ZETA_HORIZ, 3);
151 so_data (so, fb->zsbuf->width);
152 so_data (so, fb->zsbuf->height);
153 so_data (so, 0x00010001);
154 } else {
155 so_method(so, tesla, NV50TCL_ZETA_ENABLE, 1);
156 so_data (so, 0);
157 }
158
159 so_method(so, tesla, NV50TCL_VIEWPORT_HORIZ(0), 2);
160 so_data (so, w << 16);
161 so_data (so, h << 16);
162 /* set window lower left corner */
163 so_method(so, tesla, NV50TCL_WINDOW_OFFSET_X, 2);
164 so_data (so, 0);
165 so_data (so, 0);
166 /* set screen scissor rectangle */
167 so_method(so, tesla, NV50TCL_SCREEN_SCISSOR_HORIZ, 2);
168 so_data (so, w << 16);
169 so_data (so, h << 16);
170
171 return so;
172 }
173
174 static void
175 nv50_validate_samplers(struct nv50_context *nv50, struct nouveau_stateobj *so,
176 unsigned p)
177 {
178 struct nouveau_grobj *eng2d = nv50->screen->eng2d;
179 unsigned i, j, dw = nv50->sampler_nr[p] * 8;
180
181 if (!dw)
182 return;
183 nv50_so_init_sifc(nv50, so, nv50->screen->tsc, NOUVEAU_BO_VRAM,
184 p * (32 * 8 * 4), dw * 4);
185
186 so_method(so, eng2d, NV50_2D_SIFC_DATA | (2 << 29), dw);
187
188 for (i = 0; i < nv50->sampler_nr[p]; ++i) {
189 if (nv50->sampler[p][i])
190 so_datap(so, nv50->sampler[p][i]->tsc, 8);
191 else {
192 for (j = 0; j < 8; ++j) /* you get punished */
193 so_data(so, 0); /* ... for leaving holes */
194 }
195 }
196 }
197
198 static struct nouveau_stateobj *
199 validate_blend(struct nv50_context *nv50)
200 {
201 struct nouveau_stateobj *so = NULL;
202 so_ref(nv50->blend->so, &so);
203 return so;
204 }
205
206 static struct nouveau_stateobj *
207 validate_zsa(struct nv50_context *nv50)
208 {
209 struct nouveau_stateobj *so = NULL;
210 so_ref(nv50->zsa->so, &so);
211 return so;
212 }
213
214 static struct nouveau_stateobj *
215 validate_rast(struct nv50_context *nv50)
216 {
217 struct nouveau_stateobj *so = NULL;
218 so_ref(nv50->rasterizer->so, &so);
219 return so;
220 }
221
222 static struct nouveau_stateobj *
223 validate_blend_colour(struct nv50_context *nv50)
224 {
225 struct nouveau_grobj *tesla = nv50->screen->tesla;
226 struct nouveau_stateobj *so = so_new(1, 4, 0);
227
228 so_method(so, tesla, NV50TCL_BLEND_COLOR(0), 4);
229 so_data (so, fui(nv50->blend_colour.color[0]));
230 so_data (so, fui(nv50->blend_colour.color[1]));
231 so_data (so, fui(nv50->blend_colour.color[2]));
232 so_data (so, fui(nv50->blend_colour.color[3]));
233 return so;
234 }
235
236 static struct nouveau_stateobj *
237 validate_stencil_ref(struct nv50_context *nv50)
238 {
239 struct nouveau_grobj *tesla = nv50->screen->tesla;
240 struct nouveau_stateobj *so = so_new(2, 2, 0);
241
242 so_method(so, tesla, NV50TCL_STENCIL_FRONT_FUNC_REF, 1);
243 so_data (so, nv50->stencil_ref.ref_value[0]);
244 so_method(so, tesla, NV50TCL_STENCIL_BACK_FUNC_REF, 1);
245 so_data (so, nv50->stencil_ref.ref_value[1]);
246 return so;
247 }
248
249 static struct nouveau_stateobj *
250 validate_stipple(struct nv50_context *nv50)
251 {
252 struct nouveau_grobj *tesla = nv50->screen->tesla;
253 struct nouveau_stateobj *so = so_new(1, 32, 0);
254 int i;
255
256 so_method(so, tesla, NV50TCL_POLYGON_STIPPLE_PATTERN(0), 32);
257 for (i = 0; i < 32; i++)
258 so_data(so, util_bswap32(nv50->stipple.stipple[i]));
259 return so;
260 }
261
262 static struct nouveau_stateobj *
263 validate_scissor(struct nv50_context *nv50)
264 {
265 struct nouveau_grobj *tesla = nv50->screen->tesla;
266 struct pipe_scissor_state *s = &nv50->scissor;
267 struct nouveau_stateobj *so;
268
269 so = so_new(1, 2, 0);
270 so_method(so, tesla, NV50TCL_SCISSOR_HORIZ(0), 2);
271 so_data (so, (s->maxx << 16) | s->minx);
272 so_data (so, (s->maxy << 16) | s->miny);
273 return so;
274 }
275
276 static struct nouveau_stateobj *
277 validate_viewport(struct nv50_context *nv50)
278 {
279 struct nouveau_grobj *tesla = nv50->screen->tesla;
280 struct nouveau_stateobj *so = so_new(3, 7, 0);
281
282 so_method(so, tesla, NV50TCL_VIEWPORT_TRANSLATE_X(0), 3);
283 so_data (so, fui(nv50->viewport.translate[0]));
284 so_data (so, fui(nv50->viewport.translate[1]));
285 so_data (so, fui(nv50->viewport.translate[2]));
286 so_method(so, tesla, NV50TCL_VIEWPORT_SCALE_X(0), 3);
287 so_data (so, fui(nv50->viewport.scale[0]));
288 so_data (so, fui(nv50->viewport.scale[1]));
289 so_data (so, fui(nv50->viewport.scale[2]));
290
291 /* no idea what 0f90 does */
292 so_method(so, tesla, 0x0f90, 1);
293 so_data (so, 0);
294
295 return so;
296 }
297
298 static struct nouveau_stateobj *
299 validate_sampler(struct nv50_context *nv50)
300 {
301 struct nouveau_grobj *tesla = nv50->screen->tesla;
302 struct nouveau_stateobj *so;
303 unsigned nr = 0, i;
304
305 for (i = 0; i < 3; ++i)
306 nr += nv50->sampler_nr[i];
307
308 so = so_new(1 + 5 * 3, 1 + 19 * 3 + nr * 8, 3 * 2);
309
310 nv50_validate_samplers(nv50, so, 0); /* VP */
311 nv50_validate_samplers(nv50, so, 2); /* FP */
312
313 so_method(so, tesla, 0x1334, 1); /* flush TSC */
314 so_data (so, 0);
315
316 return so;
317 }
318
319 static struct nouveau_stateobj *
320 validate_vtxbuf(struct nv50_context *nv50)
321 {
322 struct nouveau_stateobj *so = NULL;
323 so_ref(nv50->state.vtxbuf, &so);
324 return so;
325 }
326
327 static struct nouveau_stateobj *
328 validate_vtxattr(struct nv50_context *nv50)
329 {
330 struct nouveau_stateobj *so = NULL;
331 so_ref(nv50->state.vtxattr, &so);
332 return so;
333 }
334
335 static struct nouveau_stateobj *
336 validate_clip(struct nv50_context *nv50)
337 {
338 struct nouveau_grobj *tesla = nv50->screen->tesla;
339 struct nouveau_stateobj *so = so_new(1, 1, 0);
340 uint32_t vvcc;
341
342 /* 0x0000 = remove whole primitive only (xyz)
343 * 0x1018 = remove whole primitive only (xy), clamp z
344 * 0x1080 = clip primitive (xyz)
345 * 0x1098 = clip primitive (xy), clamp z
346 */
347 vvcc = nv50->clip.depth_clamp ? 0x1098 : 0x1080;
348
349 so_method(so, tesla, NV50TCL_VIEW_VOLUME_CLIP_CTRL, 1);
350 so_data (so, vvcc);
351
352 return so;
353 }
354
355 struct state_validate {
356 struct nouveau_stateobj *(*func)(struct nv50_context *nv50);
357 unsigned states;
358 } validate_list[] = {
359 { validate_fb , NV50_NEW_FRAMEBUFFER },
360 { validate_blend , NV50_NEW_BLEND },
361 { validate_zsa , NV50_NEW_ZSA },
362 { nv50_vertprog_validate , NV50_NEW_VERTPROG | NV50_NEW_VERTPROG_CB },
363 { nv50_fragprog_validate , NV50_NEW_FRAGPROG | NV50_NEW_FRAGPROG_CB },
364 { nv50_geomprog_validate , NV50_NEW_GEOMPROG | NV50_NEW_GEOMPROG_CB },
365 { nv50_fp_linkage_validate, NV50_NEW_VERTPROG | NV50_NEW_GEOMPROG |
366 NV50_NEW_FRAGPROG | NV50_NEW_RASTERIZER },
367 { nv50_gp_linkage_validate, NV50_NEW_VERTPROG | NV50_NEW_GEOMPROG },
368 { validate_rast , NV50_NEW_RASTERIZER },
369 { validate_blend_colour , NV50_NEW_BLEND_COLOUR },
370 { validate_stencil_ref , NV50_NEW_STENCIL_REF },
371 { validate_stipple , NV50_NEW_STIPPLE },
372 { validate_scissor , NV50_NEW_SCISSOR },
373 { validate_viewport , NV50_NEW_VIEWPORT },
374 { validate_sampler , NV50_NEW_SAMPLER },
375 { nv50_tex_validate , NV50_NEW_TEXTURE | NV50_NEW_SAMPLER },
376 { nv50_vbo_validate , NV50_NEW_ARRAYS },
377 { validate_vtxbuf , NV50_NEW_ARRAYS },
378 { validate_vtxattr , NV50_NEW_ARRAYS },
379 { validate_clip , NV50_NEW_CLIP },
380 {}
381 };
382 #define validate_list_len (sizeof(validate_list) / sizeof(validate_list[0]))
383
384 boolean
385 nv50_state_validate(struct nv50_context *nv50, unsigned wait_dwords)
386 {
387 struct nouveau_channel *chan = nv50->screen->base.channel;
388 struct nouveau_grobj *tesla = nv50->screen->tesla;
389 unsigned nr_relocs = 128, nr_dwords = wait_dwords + 128 + 4;
390 int ret, i;
391
392 for (i = 0; i < validate_list_len; i++) {
393 struct state_validate *validate = &validate_list[i];
394 struct nouveau_stateobj *so;
395
396 if (!(nv50->dirty & validate->states))
397 continue;
398
399 so = validate->func(nv50);
400 if (!so)
401 continue;
402
403 nr_dwords += (so->total + so->cur);
404 nr_relocs += so->cur_reloc;
405
406 so_ref(so, &nv50->state.hw[i]);
407 so_ref(NULL, &so);
408 nv50->state.hw_dirty |= (1 << i);
409 }
410 nv50->dirty = 0;
411
412 if (nv50->screen->cur_ctx != nv50) {
413 for (i = 0; i < validate_list_len; i++) {
414 if (!nv50->state.hw[i] ||
415 (nv50->state.hw_dirty & (1 << i)))
416 continue;
417
418 nr_dwords += (nv50->state.hw[i]->total +
419 nv50->state.hw[i]->cur);
420 nr_relocs += nv50->state.hw[i]->cur_reloc;
421 nv50->state.hw_dirty |= (1 << i);
422 }
423
424 nv50->screen->cur_ctx = nv50;
425 }
426
427 ret = MARK_RING(chan, nr_dwords, nr_relocs);
428 if (ret) {
429 debug_printf("MARK_RING(%d, %d) failed: %d\n",
430 nr_dwords, nr_relocs, ret);
431 return FALSE;
432 }
433
434 while (nv50->state.hw_dirty) {
435 i = ffs(nv50->state.hw_dirty) - 1;
436 nv50->state.hw_dirty &= ~(1 << i);
437
438 so_emit(chan, nv50->state.hw[i]);
439 }
440
441 /* Yes, really, we need to do this. If a buffer that is referenced
442 * on the hardware isn't part of changed state above, without doing
443 * this the kernel is given no clue that the buffer is being used
444 * still. This can cause all sorts of fun issues.
445 */
446 nv50_tex_relocs(nv50);
447 so_emit_reloc_markers(chan, nv50->state.hw[0]); /* fb */
448 so_emit_reloc_markers(chan, nv50->state.hw[3]); /* vp */
449 so_emit_reloc_markers(chan, nv50->state.hw[4]); /* fp */
450 so_emit_reloc_markers(chan, nv50->state.hw[17]); /* vb */
451 nv50_screen_relocs(nv50->screen);
452
453 /* No idea.. */
454 BEGIN_RING(chan, tesla, 0x142c, 1);
455 OUT_RING (chan, 0);
456 BEGIN_RING(chan, tesla, 0x142c, 1);
457 OUT_RING (chan, 0);
458 return TRUE;
459 }
460
461 void nv50_so_init_sifc(struct nv50_context *nv50,
462 struct nouveau_stateobj *so,
463 struct nouveau_bo *bo, unsigned reloc,
464 unsigned offset, unsigned size)
465 {
466 struct nouveau_grobj *eng2d = nv50->screen->eng2d;
467
468 reloc |= NOUVEAU_BO_WR;
469
470 so_method(so, eng2d, NV50_2D_DST_FORMAT, 2);
471 so_data (so, NV50_2D_DST_FORMAT_R8_UNORM);
472 so_data (so, 1);
473 so_method(so, eng2d, NV50_2D_DST_PITCH, 5);
474 so_data (so, 262144);
475 so_data (so, 65536);
476 so_data (so, 1);
477 so_reloc (so, bo, offset, reloc | NOUVEAU_BO_HIGH, 0, 0);
478 so_reloc (so, bo, offset, reloc | NOUVEAU_BO_LOW, 0, 0);
479 so_method(so, eng2d, NV50_2D_SIFC_BITMAP_ENABLE, 2);
480 so_data (so, 0);
481 so_data (so, NV50_2D_SIFC_FORMAT_R8_UNORM);
482 so_method(so, eng2d, NV50_2D_SIFC_WIDTH, 10);
483 so_data (so, size);
484 so_data (so, 1);
485 so_data (so, 0);
486 so_data (so, 1);
487 so_data (so, 0);
488 so_data (so, 1);
489 so_data (so, 0);
490 so_data (so, 0);
491 so_data (so, 0);
492 so_data (so, 0);
493 }