Merge branch 'mesa_7_7_branch'
[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 "nv50_context.h"
24 #include "nouveau/nouveau_stateobj.h"
25
26 #define NV50_CBUF_FORMAT_CASE(n) \
27 case PIPE_FORMAT_##n: so_data(so, NV50TCL_RT_FORMAT_##n); break
28
29 #define NV50_ZETA_FORMAT_CASE(n) \
30 case PIPE_FORMAT_##n: so_data(so, NV50TCL_ZETA_FORMAT_##n); break
31
32 static void
33 nv50_state_validate_fb(struct nv50_context *nv50)
34 {
35 struct nouveau_grobj *tesla = nv50->screen->tesla;
36 struct nouveau_stateobj *so = so_new(32, 79, 18);
37 struct pipe_framebuffer_state *fb = &nv50->framebuffer;
38 unsigned i, w, h, gw = 0;
39
40 /* Set nr of active RTs and select RT for each colour output.
41 * FP result 0 always goes to RT[0], bits 4 - 6 are ignored.
42 * Ambiguous assignment results in no rendering (no DATA_ERROR).
43 */
44 so_method(so, tesla, NV50TCL_RT_CONTROL, 1);
45 so_data (so, fb->nr_cbufs |
46 (0 << 4) | (1 << 7) | (2 << 10) | (3 << 13) |
47 (4 << 16) | (5 << 19) | (6 << 22) | (7 << 25));
48
49 for (i = 0; i < fb->nr_cbufs; i++) {
50 struct pipe_texture *pt = fb->cbufs[i]->texture;
51 struct nouveau_bo *bo = nv50_miptree(pt)->base.bo;
52
53 if (!gw) {
54 w = fb->cbufs[i]->width;
55 h = fb->cbufs[i]->height;
56 gw = 1;
57 } else {
58 assert(w == fb->cbufs[i]->width);
59 assert(h == fb->cbufs[i]->height);
60 }
61
62 so_method(so, tesla, NV50TCL_RT_HORIZ(i), 2);
63 so_data (so, fb->cbufs[i]->width);
64 so_data (so, fb->cbufs[i]->height);
65
66 so_method(so, tesla, NV50TCL_RT_ADDRESS_HIGH(i), 5);
67 so_reloc (so, bo, fb->cbufs[i]->offset, NOUVEAU_BO_VRAM |
68 NOUVEAU_BO_HIGH | NOUVEAU_BO_RDWR, 0, 0);
69 so_reloc (so, bo, fb->cbufs[i]->offset, NOUVEAU_BO_VRAM |
70 NOUVEAU_BO_LOW | NOUVEAU_BO_RDWR, 0, 0);
71 switch (fb->cbufs[i]->format) {
72 NV50_CBUF_FORMAT_CASE(A8R8G8B8_UNORM);
73 NV50_CBUF_FORMAT_CASE(X8R8G8B8_UNORM);
74 NV50_CBUF_FORMAT_CASE(R5G6B5_UNORM);
75 NV50_CBUF_FORMAT_CASE(R16G16B16A16_SNORM);
76 NV50_CBUF_FORMAT_CASE(R16G16B16A16_UNORM);
77 NV50_CBUF_FORMAT_CASE(R32G32B32A32_FLOAT);
78 NV50_CBUF_FORMAT_CASE(R16G16_SNORM);
79 NV50_CBUF_FORMAT_CASE(R16G16_UNORM);
80 default:
81 NOUVEAU_ERR("AIIII unknown format %s\n",
82 pf_name(fb->cbufs[i]->format));
83 so_data(so, NV50TCL_RT_FORMAT_X8R8G8B8_UNORM);
84 break;
85 }
86 so_data(so, nv50_miptree(pt)->
87 level[fb->cbufs[i]->level].tile_mode << 4);
88 so_data(so, 0x00000000);
89
90 so_method(so, tesla, NV50TCL_RT_ARRAY_MODE, 1);
91 so_data (so, 1);
92 }
93
94 if (fb->zsbuf) {
95 struct pipe_texture *pt = fb->zsbuf->texture;
96 struct nouveau_bo *bo = nv50_miptree(pt)->base.bo;
97
98 if (!gw) {
99 w = fb->zsbuf->width;
100 h = fb->zsbuf->height;
101 gw = 1;
102 } else {
103 assert(w == fb->zsbuf->width);
104 assert(h == fb->zsbuf->height);
105 }
106
107 so_method(so, tesla, NV50TCL_ZETA_ADDRESS_HIGH, 5);
108 so_reloc (so, bo, fb->zsbuf->offset, NOUVEAU_BO_VRAM |
109 NOUVEAU_BO_HIGH | NOUVEAU_BO_RDWR, 0, 0);
110 so_reloc (so, bo, fb->zsbuf->offset, NOUVEAU_BO_VRAM |
111 NOUVEAU_BO_LOW | NOUVEAU_BO_RDWR, 0, 0);
112 switch (fb->zsbuf->format) {
113 NV50_ZETA_FORMAT_CASE(S8Z24_UNORM);
114 NV50_ZETA_FORMAT_CASE(X8Z24_UNORM);
115 NV50_ZETA_FORMAT_CASE(Z24S8_UNORM);
116 NV50_ZETA_FORMAT_CASE(Z32_FLOAT);
117 default:
118 NOUVEAU_ERR("AIIII unknown format %s\n",
119 pf_name(fb->zsbuf->format));
120 so_data(so, NV50TCL_ZETA_FORMAT_S8Z24_UNORM);
121 break;
122 }
123 so_data(so, nv50_miptree(pt)->
124 level[fb->zsbuf->level].tile_mode << 4);
125 so_data(so, 0x00000000);
126
127 so_method(so, tesla, NV50TCL_ZETA_ENABLE, 1);
128 so_data (so, 1);
129 so_method(so, tesla, NV50TCL_ZETA_HORIZ, 3);
130 so_data (so, fb->zsbuf->width);
131 so_data (so, fb->zsbuf->height);
132 so_data (so, 0x00010001);
133 } else {
134 so_method(so, tesla, NV50TCL_ZETA_ENABLE, 1);
135 so_data (so, 0);
136 }
137
138 so_method(so, tesla, NV50TCL_VIEWPORT_HORIZ(0), 2);
139 so_data (so, w << 16);
140 so_data (so, h << 16);
141 /* set window lower left corner */
142 so_method(so, tesla, NV50TCL_WINDOW_OFFSET_X, 2);
143 so_data (so, 0);
144 so_data (so, 0);
145 /* set screen scissor rectangle */
146 so_method(so, tesla, NV50TCL_SCREEN_SCISSOR_HORIZ, 2);
147 so_data (so, w << 16);
148 so_data (so, h << 16);
149
150 /* we set scissors to framebuffer size when they're 'turned off' */
151 nv50->dirty |= NV50_NEW_SCISSOR;
152 so_ref(NULL, &nv50->state.scissor);
153
154 so_ref(so, &nv50->state.fb);
155 so_ref(NULL, &so);
156 }
157
158 static void
159 nv50_validate_samplers(struct nv50_context *nv50, struct nouveau_stateobj *so,
160 unsigned p)
161 {
162 struct nouveau_grobj *eng2d = nv50->screen->eng2d;
163 unsigned i, j, dw = nv50->sampler_nr[p] * 8;
164
165 if (!dw)
166 return;
167 nv50_so_init_sifc(nv50, so, nv50->screen->tsc, NOUVEAU_BO_VRAM,
168 p * (32 * 8 * 4), dw * 4);
169
170 so_method(so, eng2d, NV50_2D_SIFC_DATA | (2 << 29), dw);
171
172 for (i = 0; i < nv50->sampler_nr[p]; ++i) {
173 if (nv50->sampler[p][i])
174 so_datap(so, nv50->sampler[p][i]->tsc, 8);
175 else {
176 for (j = 0; j < 8; ++j) /* you get punished */
177 so_data(so, 0); /* ... for leaving holes */
178 }
179 }
180 }
181
182 static void
183 nv50_state_emit(struct nv50_context *nv50)
184 {
185 struct nv50_screen *screen = nv50->screen;
186 struct nouveau_channel *chan = screen->base.channel;
187
188 /* I don't want to copy headers from the winsys. */
189 screen->cur_ctx = nv50;
190
191 if (nv50->pctx_id != screen->cur_pctx) {
192 if (nv50->state.fb)
193 nv50->state.dirty |= NV50_NEW_FRAMEBUFFER;
194 if (nv50->state.blend)
195 nv50->state.dirty |= NV50_NEW_BLEND;
196 if (nv50->state.zsa)
197 nv50->state.dirty |= NV50_NEW_ZSA;
198 if (nv50->state.vertprog)
199 nv50->state.dirty |= NV50_NEW_VERTPROG;
200 if (nv50->state.fragprog)
201 nv50->state.dirty |= NV50_NEW_FRAGPROG;
202 if (nv50->state.geomprog)
203 nv50->state.dirty |= NV50_NEW_GEOMPROG;
204 if (nv50->state.rast)
205 nv50->state.dirty |= NV50_NEW_RASTERIZER;
206 if (nv50->state.blend_colour)
207 nv50->state.dirty |= NV50_NEW_BLEND_COLOUR;
208 if (nv50->state.stipple)
209 nv50->state.dirty |= NV50_NEW_STIPPLE;
210 if (nv50->state.scissor)
211 nv50->state.dirty |= NV50_NEW_SCISSOR;
212 if (nv50->state.viewport)
213 nv50->state.dirty |= NV50_NEW_VIEWPORT;
214 if (nv50->state.tsc_upload)
215 nv50->state.dirty |= NV50_NEW_SAMPLER;
216 if (nv50->state.tic_upload)
217 nv50->state.dirty |= NV50_NEW_TEXTURE;
218 if (nv50->state.vtxfmt && nv50->state.vtxbuf)
219 nv50->state.dirty |= NV50_NEW_ARRAYS;
220 screen->cur_pctx = nv50->pctx_id;
221 }
222
223 if (nv50->state.dirty & NV50_NEW_FRAMEBUFFER)
224 so_emit(chan, nv50->state.fb);
225 if (nv50->state.dirty & NV50_NEW_BLEND)
226 so_emit(chan, nv50->state.blend);
227 if (nv50->state.dirty & NV50_NEW_ZSA)
228 so_emit(chan, nv50->state.zsa);
229 if (nv50->state.dirty & NV50_NEW_VERTPROG)
230 so_emit(chan, nv50->state.vertprog);
231 if (nv50->state.dirty & NV50_NEW_FRAGPROG)
232 so_emit(chan, nv50->state.fragprog);
233 if (nv50->state.dirty & NV50_NEW_GEOMPROG && nv50->state.geomprog)
234 so_emit(chan, nv50->state.geomprog);
235 if (nv50->state.dirty & (NV50_NEW_FRAGPROG | NV50_NEW_VERTPROG |
236 NV50_NEW_GEOMPROG | NV50_NEW_RASTERIZER))
237 so_emit(chan, nv50->state.fp_linkage);
238 if ((nv50->state.dirty & (NV50_NEW_VERTPROG | NV50_NEW_GEOMPROG))
239 && nv50->state.gp_linkage)
240 so_emit(chan, nv50->state.gp_linkage);
241 if (nv50->state.dirty & NV50_NEW_RASTERIZER)
242 so_emit(chan, nv50->state.rast);
243 if (nv50->state.dirty & NV50_NEW_BLEND_COLOUR)
244 so_emit(chan, nv50->state.blend_colour);
245 if (nv50->state.dirty & NV50_NEW_STIPPLE)
246 so_emit(chan, nv50->state.stipple);
247 if (nv50->state.dirty & NV50_NEW_SCISSOR)
248 so_emit(chan, nv50->state.scissor);
249 if (nv50->state.dirty & NV50_NEW_VIEWPORT)
250 so_emit(chan, nv50->state.viewport);
251 if (nv50->state.dirty & NV50_NEW_SAMPLER)
252 so_emit(chan, nv50->state.tsc_upload);
253 if (nv50->state.dirty & NV50_NEW_TEXTURE)
254 so_emit(chan, nv50->state.tic_upload);
255 if (nv50->state.dirty & NV50_NEW_ARRAYS) {
256 so_emit(chan, nv50->state.vtxfmt);
257 so_emit(chan, nv50->state.vtxbuf);
258 if (nv50->state.vtxattr)
259 so_emit(chan, nv50->state.vtxattr);
260 }
261 nv50->state.dirty = 0;
262 }
263
264 void
265 nv50_state_flush_notify(struct nouveau_channel *chan)
266 {
267 struct nv50_context *nv50 = chan->user_private;
268
269 if (nv50->state.tic_upload && !(nv50->dirty & NV50_NEW_TEXTURE))
270 so_emit(chan, nv50->state.tic_upload);
271
272 so_emit_reloc_markers(chan, nv50->state.fb);
273 so_emit_reloc_markers(chan, nv50->state.vertprog);
274 so_emit_reloc_markers(chan, nv50->state.fragprog);
275 so_emit_reloc_markers(chan, nv50->state.vtxbuf);
276 so_emit_reloc_markers(chan, nv50->screen->static_init);
277
278 if (nv50->state.instbuf)
279 so_emit_reloc_markers(chan, nv50->state.instbuf);
280 }
281
282 boolean
283 nv50_state_validate(struct nv50_context *nv50)
284 {
285 struct nouveau_grobj *tesla = nv50->screen->tesla;
286 struct nouveau_stateobj *so;
287 unsigned i;
288
289 if (nv50->dirty & NV50_NEW_FRAMEBUFFER)
290 nv50_state_validate_fb(nv50);
291
292 if (nv50->dirty & NV50_NEW_BLEND)
293 so_ref(nv50->blend->so, &nv50->state.blend);
294
295 if (nv50->dirty & NV50_NEW_ZSA)
296 so_ref(nv50->zsa->so, &nv50->state.zsa);
297
298 if (nv50->dirty & (NV50_NEW_VERTPROG | NV50_NEW_VERTPROG_CB))
299 nv50_vertprog_validate(nv50);
300
301 if (nv50->dirty & (NV50_NEW_FRAGPROG | NV50_NEW_FRAGPROG_CB))
302 nv50_fragprog_validate(nv50);
303
304 if (nv50->dirty & (NV50_NEW_GEOMPROG | NV50_NEW_GEOMPROG_CB))
305 nv50_geomprog_validate(nv50);
306
307 if (nv50->dirty & (NV50_NEW_FRAGPROG | NV50_NEW_VERTPROG |
308 NV50_NEW_GEOMPROG | NV50_NEW_RASTERIZER))
309 nv50_fp_linkage_validate(nv50);
310
311 if (nv50->dirty & (NV50_NEW_GEOMPROG | NV50_NEW_VERTPROG))
312 nv50_gp_linkage_validate(nv50);
313
314 if (nv50->dirty & NV50_NEW_RASTERIZER)
315 so_ref(nv50->rasterizer->so, &nv50->state.rast);
316
317 if (nv50->dirty & NV50_NEW_BLEND_COLOUR) {
318 so = so_new(1, 4, 0);
319 so_method(so, tesla, NV50TCL_BLEND_COLOR(0), 4);
320 so_data (so, fui(nv50->blend_colour.color[0]));
321 so_data (so, fui(nv50->blend_colour.color[1]));
322 so_data (so, fui(nv50->blend_colour.color[2]));
323 so_data (so, fui(nv50->blend_colour.color[3]));
324 so_ref(so, &nv50->state.blend_colour);
325 so_ref(NULL, &so);
326 }
327
328 if (nv50->dirty & NV50_NEW_STIPPLE) {
329 so = so_new(1, 32, 0);
330 so_method(so, tesla, NV50TCL_POLYGON_STIPPLE_PATTERN(0), 32);
331 for (i = 0; i < 32; i++)
332 so_data(so, util_bswap32(nv50->stipple.stipple[i]));
333 so_ref(so, &nv50->state.stipple);
334 so_ref(NULL, &so);
335 }
336
337 if (nv50->dirty & (NV50_NEW_SCISSOR | NV50_NEW_RASTERIZER)) {
338 struct pipe_rasterizer_state *rast = &nv50->rasterizer->pipe;
339 struct pipe_scissor_state *s = &nv50->scissor;
340
341 if (nv50->state.scissor &&
342 (rast->scissor == 0 && nv50->state.scissor_enabled == 0))
343 goto scissor_uptodate;
344 nv50->state.scissor_enabled = rast->scissor;
345
346 so = so_new(1, 2, 0);
347 so_method(so, tesla, NV50TCL_SCISSOR_HORIZ(0), 2);
348 if (nv50->state.scissor_enabled) {
349 so_data(so, (s->maxx << 16) | s->minx);
350 so_data(so, (s->maxy << 16) | s->miny);
351 } else {
352 so_data(so, (nv50->framebuffer.width << 16));
353 so_data(so, (nv50->framebuffer.height << 16));
354 }
355 so_ref(so, &nv50->state.scissor);
356 so_ref(NULL, &so);
357 nv50->state.dirty |= NV50_NEW_SCISSOR;
358 }
359 scissor_uptodate:
360
361 if (nv50->dirty & (NV50_NEW_VIEWPORT | NV50_NEW_RASTERIZER)) {
362 unsigned bypass;
363
364 if (!nv50->rasterizer->pipe.bypass_vs_clip_and_viewport)
365 bypass = 0;
366 else
367 bypass = 1;
368
369 if (nv50->state.viewport &&
370 (bypass || !(nv50->dirty & NV50_NEW_VIEWPORT)) &&
371 nv50->state.viewport_bypass == bypass)
372 goto viewport_uptodate;
373 nv50->state.viewport_bypass = bypass;
374
375 so = so_new(5, 9, 0);
376 if (!bypass) {
377 so_method(so, tesla, NV50TCL_VIEWPORT_TRANSLATE_X(0), 3);
378 so_data (so, fui(nv50->viewport.translate[0]));
379 so_data (so, fui(nv50->viewport.translate[1]));
380 so_data (so, fui(nv50->viewport.translate[2]));
381 so_method(so, tesla, NV50TCL_VIEWPORT_SCALE_X(0), 3);
382 so_data (so, fui(nv50->viewport.scale[0]));
383 so_data (so, fui(nv50->viewport.scale[1]));
384 so_data (so, fui(nv50->viewport.scale[2]));
385
386 so_method(so, tesla, NV50TCL_VIEWPORT_TRANSFORM_EN, 1);
387 so_data (so, 1);
388 /* 0x0000 = remove whole primitive only (xyz)
389 * 0x1018 = remove whole primitive only (xy), clamp z
390 * 0x1080 = clip primitive (xyz)
391 * 0x1098 = clip primitive (xy), clamp z
392 */
393 so_method(so, tesla, NV50TCL_VIEW_VOLUME_CLIP_CTRL, 1);
394 so_data (so, 0x1080);
395 /* no idea what 0f90 does */
396 so_method(so, tesla, 0x0f90, 1);
397 so_data (so, 0);
398 } else {
399 so_method(so, tesla, NV50TCL_VIEWPORT_TRANSFORM_EN, 1);
400 so_data (so, 0);
401 so_method(so, tesla, NV50TCL_VIEW_VOLUME_CLIP_CTRL, 1);
402 so_data (so, 0x0000);
403 so_method(so, tesla, 0x0f90, 1);
404 so_data (so, 1);
405 }
406
407 so_ref(so, &nv50->state.viewport);
408 so_ref(NULL, &so);
409 nv50->state.dirty |= NV50_NEW_VIEWPORT;
410 }
411 viewport_uptodate:
412
413 if (nv50->dirty & NV50_NEW_SAMPLER) {
414 unsigned nr = 0;
415
416 for (i = 0; i < PIPE_SHADER_TYPES; ++i)
417 nr += nv50->sampler_nr[i];
418
419 so = so_new(1 + 5 * PIPE_SHADER_TYPES,
420 1 + 19 * PIPE_SHADER_TYPES + nr * 8,
421 PIPE_SHADER_TYPES * 2);
422
423 nv50_validate_samplers(nv50, so, PIPE_SHADER_VERTEX);
424 nv50_validate_samplers(nv50, so, PIPE_SHADER_FRAGMENT);
425
426 so_method(so, tesla, 0x1334, 1); /* flush TSC */
427 so_data (so, 0);
428
429 so_ref(so, &nv50->state.tsc_upload);
430 so_ref(NULL, &so);
431 }
432
433 if (nv50->dirty & (NV50_NEW_TEXTURE | NV50_NEW_SAMPLER))
434 nv50_tex_validate(nv50);
435
436 if (nv50->dirty & NV50_NEW_ARRAYS)
437 nv50_vbo_validate(nv50);
438
439 nv50->state.dirty |= nv50->dirty;
440 nv50->dirty = 0;
441 nv50_state_emit(nv50);
442
443 return TRUE;
444 }
445
446 void nv50_so_init_sifc(struct nv50_context *nv50,
447 struct nouveau_stateobj *so,
448 struct nouveau_bo *bo, unsigned reloc,
449 unsigned offset, unsigned size)
450 {
451 struct nouveau_grobj *eng2d = nv50->screen->eng2d;
452
453 reloc |= NOUVEAU_BO_WR;
454
455 so_method(so, eng2d, NV50_2D_DST_FORMAT, 2);
456 so_data (so, NV50_2D_DST_FORMAT_R8_UNORM);
457 so_data (so, 1);
458 so_method(so, eng2d, NV50_2D_DST_PITCH, 5);
459 so_data (so, 262144);
460 so_data (so, 65536);
461 so_data (so, 1);
462 so_reloc (so, bo, offset, reloc | NOUVEAU_BO_HIGH, 0, 0);
463 so_reloc (so, bo, offset, reloc | NOUVEAU_BO_LOW, 0, 0);
464 so_method(so, eng2d, NV50_2D_SIFC_BITMAP_ENABLE, 2);
465 so_data (so, 0);
466 so_data (so, NV50_2D_SIFC_FORMAT_R8_UNORM);
467 so_method(so, eng2d, NV50_2D_SIFC_WIDTH, 10);
468 so_data (so, size);
469 so_data (so, 1);
470 so_data (so, 0);
471 so_data (so, 1);
472 so_data (so, 0);
473 so_data (so, 1);
474 so_data (so, 0);
475 so_data (so, 0);
476 so_data (so, 0);
477 so_data (so, 0);
478 }