nv50_state_validate.c revision 272bbbffb0138024564cbf8410a30ea9e25eb179 
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 "nouveau/nouveau_stateobj.h"
27
28static struct nouveau_stateobj *
29validate_fb(struct nv50_context *nv50)
30{
31	struct nouveau_grobj *tesla = nv50->screen->tesla;
32	struct nouveau_stateobj *so = so_new(32, 79, 18);
33	struct pipe_framebuffer_state *fb = &nv50->framebuffer;
34	unsigned i, w, h, gw = 0;
35
36	/* Set nr of active RTs and select RT for each colour output.
37	 * FP result 0 always goes to RT[0], bits 4 - 6 are ignored.
38	 * Ambiguous assignment results in no rendering (no DATA_ERROR).
39	 */
40	so_method(so, tesla, NV50TCL_RT_CONTROL, 1);
41	so_data  (so, fb->nr_cbufs |
42		  (0 <<  4) | (1 <<  7) | (2 << 10) | (3 << 13) |
43		  (4 << 16) | (5 << 19) | (6 << 22) | (7 << 25));
44
45	for (i = 0; i < fb->nr_cbufs; i++) {
46		struct pipe_texture *pt = fb->cbufs[i]->texture;
47		struct nouveau_bo *bo = nv50_miptree(pt)->base.bo;
48
49		if (!gw) {
50			w = fb->cbufs[i]->width;
51			h = fb->cbufs[i]->height;
52			gw = 1;
53		} else {
54			assert(w == fb->cbufs[i]->width);
55			assert(h == fb->cbufs[i]->height);
56		}
57
58		so_method(so, tesla, NV50TCL_RT_HORIZ(i), 2);
59		so_data  (so, fb->cbufs[i]->width);
60		so_data  (so, fb->cbufs[i]->height);
61
62		so_method(so, tesla, NV50TCL_RT_ADDRESS_HIGH(i), 5);
63		so_reloc (so, bo, fb->cbufs[i]->offset, NOUVEAU_BO_VRAM |
64			      NOUVEAU_BO_HIGH | NOUVEAU_BO_RDWR, 0, 0);
65		so_reloc (so, bo, fb->cbufs[i]->offset, NOUVEAU_BO_VRAM |
66			      NOUVEAU_BO_LOW | NOUVEAU_BO_RDWR, 0, 0);
67		switch (fb->cbufs[i]->format) {
68		case PIPE_FORMAT_B8G8R8A8_UNORM:
69			so_data(so, NV50TCL_RT_FORMAT_A8R8G8B8_UNORM);
70			break;
71		case PIPE_FORMAT_B8G8R8X8_UNORM:
72			so_data(so, NV50TCL_RT_FORMAT_X8R8G8B8_UNORM);
73			break;
74		case PIPE_FORMAT_B5G6R5_UNORM:
75			so_data(so, NV50TCL_RT_FORMAT_R5G6B5_UNORM);
76			break;
77		case PIPE_FORMAT_R16G16B16A16_SNORM:
78			so_data(so, NV50TCL_RT_FORMAT_R16G16B16A16_SNORM);
79			break;
80		case PIPE_FORMAT_R16G16B16A16_UNORM:
81			so_data(so, NV50TCL_RT_FORMAT_R16G16B16A16_UNORM);
82			break;
83		case PIPE_FORMAT_R32G32B32A32_FLOAT:
84			so_data(so, NV50TCL_RT_FORMAT_R32G32B32A32_FLOAT);
85			break;
86		case PIPE_FORMAT_R16G16_SNORM:
87			so_data(so, NV50TCL_RT_FORMAT_R16G16_SNORM);
88			break;
89		case PIPE_FORMAT_R16G16_UNORM:
90			so_data(so, NV50TCL_RT_FORMAT_R16G16_UNORM);
91			break;
92		default:
93			NOUVEAU_ERR("AIIII unknown format %s\n",
94			            util_format_name(fb->cbufs[i]->format));
95			so_data(so, NV50TCL_RT_FORMAT_X8R8G8B8_UNORM);
96			break;
97		}
98		so_data(so, nv50_miptree(pt)->
99				level[fb->cbufs[i]->level].tile_mode << 4);
100		so_data(so, 0x00000000);
101
102		so_method(so, tesla, NV50TCL_RT_ARRAY_MODE, 1);
103		so_data  (so, 1);
104	}
105
106	if (fb->zsbuf) {
107		struct pipe_texture *pt = fb->zsbuf->texture;
108		struct nouveau_bo *bo = nv50_miptree(pt)->base.bo;
109
110		if (!gw) {
111			w = fb->zsbuf->width;
112			h = fb->zsbuf->height;
113			gw = 1;
114		} else {
115			assert(w == fb->zsbuf->width);
116			assert(h == fb->zsbuf->height);
117		}
118
119		so_method(so, tesla, NV50TCL_ZETA_ADDRESS_HIGH, 5);
120		so_reloc (so, bo, fb->zsbuf->offset, NOUVEAU_BO_VRAM |
121			      NOUVEAU_BO_HIGH | NOUVEAU_BO_RDWR, 0, 0);
122		so_reloc (so, bo, fb->zsbuf->offset, NOUVEAU_BO_VRAM |
123			      NOUVEAU_BO_LOW | NOUVEAU_BO_RDWR, 0, 0);
124		switch (fb->zsbuf->format) {
125		case PIPE_FORMAT_Z24S8_UNORM:
126			so_data(so, NV50TCL_ZETA_FORMAT_S8Z24_UNORM);
127			break;
128		case PIPE_FORMAT_Z24X8_UNORM:
129			so_data(so, NV50TCL_ZETA_FORMAT_X8Z24_UNORM);
130			break;
131		case PIPE_FORMAT_S8Z24_UNORM:
132			so_data(so, NV50TCL_ZETA_FORMAT_Z24S8_UNORM);
133			break;
134		case PIPE_FORMAT_Z32_FLOAT:
135			so_data(so, NV50TCL_ZETA_FORMAT_Z32_FLOAT);
136			break;
137		default:
138			NOUVEAU_ERR("AIIII unknown format %s\n",
139			            util_format_name(fb->zsbuf->format));
140			so_data(so, NV50TCL_ZETA_FORMAT_S8Z24_UNORM);
141			break;
142		}
143		so_data(so, nv50_miptree(pt)->
144				level[fb->zsbuf->level].tile_mode << 4);
145		so_data(so, 0x00000000);
146
147		so_method(so, tesla, NV50TCL_ZETA_ENABLE, 1);
148		so_data  (so, 1);
149		so_method(so, tesla, NV50TCL_ZETA_HORIZ, 3);
150		so_data  (so, fb->zsbuf->width);
151		so_data  (so, fb->zsbuf->height);
152		so_data  (so, 0x00010001);
153	} else {
154		so_method(so, tesla, NV50TCL_ZETA_ENABLE, 1);
155		so_data  (so, 0);
156	}
157
158	so_method(so, tesla, NV50TCL_VIEWPORT_HORIZ(0), 2);
159	so_data  (so, w << 16);
160	so_data  (so, h << 16);
161	/* set window lower left corner */
162	so_method(so, tesla, NV50TCL_WINDOW_OFFSET_X, 2);
163	so_data  (so, 0);
164	so_data  (so, 0);
165	/* set screen scissor rectangle */
166	so_method(so, tesla, NV50TCL_SCREEN_SCISSOR_HORIZ, 2);
167	so_data  (so, w << 16);
168	so_data  (so, h << 16);
169
170	return so;
171}
172
173static void
174nv50_validate_samplers(struct nv50_context *nv50, struct nouveau_stateobj *so,
175		       unsigned p)
176{
177	struct nouveau_grobj *eng2d = nv50->screen->eng2d;
178	unsigned i, j, dw = nv50->sampler_nr[p] * 8;
179
180	if (!dw)
181		return;
182	nv50_so_init_sifc(nv50, so, nv50->screen->tsc, NOUVEAU_BO_VRAM,
183			  p * (32 * 8 * 4), dw * 4);
184
185	so_method(so, eng2d, NV50_2D_SIFC_DATA | (2 << 29), dw);
186
187	for (i = 0; i < nv50->sampler_nr[p]; ++i) {
188		if (nv50->sampler[p][i])
189			so_datap(so, nv50->sampler[p][i]->tsc, 8);
190		else {
191			for (j = 0; j < 8; ++j) /* you get punished */
192				so_data(so, 0); /* ... for leaving holes */
193		}
194	}
195}
196
197static struct nouveau_stateobj *
198validate_blend(struct nv50_context *nv50)
199{
200	struct nouveau_stateobj *so = NULL;
201	so_ref(nv50->blend->so, &so);
202	return so;
203}
204
205static struct nouveau_stateobj *
206validate_zsa(struct nv50_context *nv50)
207{
208	struct nouveau_stateobj *so = NULL;
209	so_ref(nv50->zsa->so, &so);
210	return so;
211}
212
213static struct nouveau_stateobj *
214validate_rast(struct nv50_context *nv50)
215{
216	struct nouveau_stateobj *so = NULL;
217	so_ref(nv50->rasterizer->so, &so);
218	return so;
219}
220
221static struct nouveau_stateobj *
222validate_blend_colour(struct nv50_context *nv50)
223{
224	struct nouveau_grobj *tesla = nv50->screen->tesla;
225	struct nouveau_stateobj *so = so_new(1, 4, 0);
226
227	so_method(so, tesla, NV50TCL_BLEND_COLOR(0), 4);
228	so_data  (so, fui(nv50->blend_colour.color[0]));
229	so_data  (so, fui(nv50->blend_colour.color[1]));
230	so_data  (so, fui(nv50->blend_colour.color[2]));
231	so_data  (so, fui(nv50->blend_colour.color[3]));
232	return so;
233}
234
235static struct nouveau_stateobj *
236validate_stencil_ref(struct nv50_context *nv50)
237{
238	struct nouveau_grobj *tesla = nv50->screen->tesla;
239	struct nouveau_stateobj *so = so = so_new(2, 2, 0);
240
241	so_method(so, tesla, NV50TCL_STENCIL_FRONT_FUNC_REF, 1);
242	so_data  (so, nv50->stencil_ref.ref_value[0]);
243	so_method(so, tesla, NV50TCL_STENCIL_BACK_FUNC_REF, 1);
244	so_data  (so, nv50->stencil_ref.ref_value[1]);
245	return so;
246}
247
248static struct nouveau_stateobj *
249validate_stipple(struct nv50_context *nv50)
250{
251	struct nouveau_grobj *tesla = nv50->screen->tesla;
252	struct nouveau_stateobj *so = so_new(1, 32, 0);
253	int i;
254
255	so_method(so, tesla, NV50TCL_POLYGON_STIPPLE_PATTERN(0), 32);
256	for (i = 0; i < 32; i++)
257		so_data(so, util_bswap32(nv50->stipple.stipple[i]));
258	return so;
259}
260
261static struct nouveau_stateobj *
262validate_scissor(struct nv50_context *nv50)
263{
264	struct nouveau_grobj *tesla = nv50->screen->tesla;
265        struct pipe_scissor_state *s = &nv50->scissor;
266	struct nouveau_stateobj *so;
267
268	so = so_new(1, 2, 0);
269	so_method(so, tesla, NV50TCL_SCISSOR_HORIZ(0), 2);
270	so_data  (so, (s->maxx << 16) | s->minx);
271	so_data  (so, (s->maxy << 16) | s->miny);
272	return so;
273}
274
275static struct nouveau_stateobj *
276validate_viewport(struct nv50_context *nv50)
277{
278	struct nouveau_grobj *tesla = nv50->screen->tesla;
279	struct nouveau_stateobj *so = so_new(5, 9, 0);
280
281	so_method(so, tesla, NV50TCL_VIEWPORT_TRANSLATE_X(0), 3);
282	so_data  (so, fui(nv50->viewport.translate[0]));
283	so_data  (so, fui(nv50->viewport.translate[1]));
284	so_data  (so, fui(nv50->viewport.translate[2]));
285	so_method(so, tesla, NV50TCL_VIEWPORT_SCALE_X(0), 3);
286	so_data  (so, fui(nv50->viewport.scale[0]));
287	so_data  (so, fui(nv50->viewport.scale[1]));
288	so_data  (so, fui(nv50->viewport.scale[2]));
289
290	so_method(so, tesla, NV50TCL_VIEWPORT_TRANSFORM_EN, 1);
291	so_data  (so, 1);
292	/* 0x0000 = remove whole primitive only (xyz)
293	 * 0x1018 = remove whole primitive only (xy), clamp z
294	 * 0x1080 = clip primitive (xyz)
295	 * 0x1098 = clip primitive (xy), clamp z
296	 */
297	so_method(so, tesla, NV50TCL_VIEW_VOLUME_CLIP_CTRL, 1);
298	so_data  (so, 0x1080);
299	/* no idea what 0f90 does */
300	so_method(so, tesla, 0x0f90, 1);
301	so_data  (so, 0);
302
303	return so;
304}
305
306static struct nouveau_stateobj *
307validate_sampler(struct nv50_context *nv50)
308{
309	struct nouveau_grobj *tesla = nv50->screen->tesla;
310	struct nouveau_stateobj *so;
311	unsigned nr = 0, i;
312
313	for (i = 0; i < 3; ++i)
314		nr += nv50->sampler_nr[i];
315
316	so = so_new(1 + 5 * 3, 1 + 19 * 3 + nr * 8, 3 * 2);
317
318	nv50_validate_samplers(nv50, so, 0); /* VP */
319	nv50_validate_samplers(nv50, so, 2); /* FP */
320
321	so_method(so, tesla, 0x1334, 1); /* flush TSC */
322	so_data  (so, 0);
323
324	return so;
325}
326
327static struct nouveau_stateobj *
328validate_vtxbuf(struct nv50_context *nv50)
329{
330	struct nouveau_stateobj *so = NULL;
331	so_ref(nv50->state.vtxbuf, &so);
332	return so;
333}
334
335static struct nouveau_stateobj *
336validate_vtxattr(struct nv50_context *nv50)
337{
338	struct nouveau_stateobj *so = NULL;
339	so_ref(nv50->state.vtxattr, &so);
340	return so;
341}
342
343struct state_validate {
344	struct nouveau_stateobj *(*func)(struct nv50_context *nv50);
345	unsigned states;
346} validate_list[] = {
347	{ validate_fb             , NV50_NEW_FRAMEBUFFER                      },
348	{ validate_blend          , NV50_NEW_BLEND                            },
349	{ validate_zsa            , NV50_NEW_ZSA                              },
350	{ nv50_vertprog_validate  , NV50_NEW_VERTPROG | NV50_NEW_VERTPROG_CB  },
351	{ nv50_fragprog_validate  , NV50_NEW_FRAGPROG | NV50_NEW_FRAGPROG_CB  },
352	{ nv50_geomprog_validate  , NV50_NEW_GEOMPROG | NV50_NEW_GEOMPROG_CB  },
353	{ nv50_fp_linkage_validate, NV50_NEW_VERTPROG | NV50_NEW_GEOMPROG |
354				    NV50_NEW_FRAGPROG | NV50_NEW_RASTERIZER   },
355	{ nv50_gp_linkage_validate, NV50_NEW_VERTPROG | NV50_NEW_GEOMPROG     },
356	{ validate_rast           , NV50_NEW_RASTERIZER                       },
357	{ validate_blend_colour   , NV50_NEW_BLEND_COLOUR                     },
358	{ validate_stencil_ref    , NV50_NEW_STENCIL_REF                      },
359	{ validate_stipple        , NV50_NEW_STIPPLE                          },
360	{ validate_scissor        , NV50_NEW_SCISSOR                          },
361	{ validate_viewport       , NV50_NEW_VIEWPORT                         },
362	{ validate_sampler        , NV50_NEW_SAMPLER                          },
363	{ nv50_tex_validate       , NV50_NEW_TEXTURE | NV50_NEW_SAMPLER       },
364	{ nv50_vbo_validate       , NV50_NEW_ARRAYS                           },
365	{ validate_vtxbuf         , NV50_NEW_ARRAYS                           },
366	{ validate_vtxattr        , NV50_NEW_ARRAYS                           },
367	{}
368};
369#define validate_list_len (sizeof(validate_list) / sizeof(validate_list[0]))
370
371boolean
372nv50_state_validate(struct nv50_context *nv50, unsigned wait_dwords)
373{
374	struct nouveau_channel *chan = nv50->screen->base.channel;
375	struct nouveau_grobj *tesla = nv50->screen->tesla;
376	unsigned nr_relocs = 128, nr_dwords = wait_dwords + 128 + 4;
377	int ret, i;
378
379	for (i = 0; i < validate_list_len; i++) {
380		struct state_validate *validate = &validate_list[i];
381		struct nouveau_stateobj *so;
382
383		if (!(nv50->dirty & validate->states))
384			continue;
385
386		so = validate->func(nv50);
387		if (!so)
388			continue;
389
390		nr_dwords += (so->total + so->cur);
391		nr_relocs += so->cur_reloc;
392
393		so_ref(so, &nv50->state.hw[i]);
394		so_ref(NULL, &so);
395		nv50->state.hw_dirty |= (1 << i);
396	}
397	nv50->dirty = 0;
398
399	if (nv50->screen->cur_ctx != nv50) {
400		for (i = 0; i < validate_list_len; i++) {
401			if (!nv50->state.hw[i] ||
402			    (nv50->state.hw_dirty & (1 << i)))
403				continue;
404
405			nr_dwords += (nv50->state.hw[i]->total +
406				      nv50->state.hw[i]->cur);
407			nr_relocs += nv50->state.hw[i]->cur_reloc;
408			nv50->state.hw_dirty |= (1 << i);
409		}
410
411		nv50->screen->cur_ctx = nv50;
412	}
413
414	ret = MARK_RING(chan, nr_dwords, nr_relocs);
415	if (ret) {
416		debug_printf("MARK_RING(%d, %d) failed: %d\n",
417			     nr_dwords, nr_relocs, ret);
418		return FALSE;
419	}
420
421	while (nv50->state.hw_dirty) {
422		i = ffs(nv50->state.hw_dirty) - 1;
423		nv50->state.hw_dirty &= ~(1 << i);
424
425		so_emit(chan, nv50->state.hw[i]);
426	}
427
428	/* Yes, really, we need to do this.  If a buffer that is referenced
429	 * on the hardware isn't part of changed state above, without doing
430	 * this the kernel is given no clue that the buffer is being used
431	 * still.  This can cause all sorts of fun issues.
432	 */
433	nv50_tex_relocs(nv50);
434	so_emit_reloc_markers(chan, nv50->state.hw[0]); /* fb */
435	so_emit_reloc_markers(chan, nv50->state.hw[3]); /* vp */
436	so_emit_reloc_markers(chan, nv50->state.hw[4]); /* fp */
437	so_emit_reloc_markers(chan, nv50->state.hw[17]); /* vb */
438	nv50_screen_relocs(nv50->screen);
439
440	/* No idea.. */
441	BEGIN_RING(chan, tesla, 0x142c, 1);
442	OUT_RING  (chan, 0);
443	BEGIN_RING(chan, tesla, 0x142c, 1);
444	OUT_RING  (chan, 0);
445	return TRUE;
446}
447
448void nv50_so_init_sifc(struct nv50_context *nv50,
449		       struct nouveau_stateobj *so,
450		       struct nouveau_bo *bo, unsigned reloc,
451		       unsigned offset, unsigned size)
452{
453	struct nouveau_grobj *eng2d = nv50->screen->eng2d;
454
455	reloc |= NOUVEAU_BO_WR;
456
457	so_method(so, eng2d, NV50_2D_DST_FORMAT, 2);
458	so_data  (so, NV50_2D_DST_FORMAT_R8_UNORM);
459	so_data  (so, 1);
460	so_method(so, eng2d, NV50_2D_DST_PITCH, 5);
461	so_data  (so, 262144);
462	so_data  (so, 65536);
463	so_data  (so, 1);
464	so_reloc (so, bo, offset, reloc | NOUVEAU_BO_HIGH, 0, 0);
465	so_reloc (so, bo, offset, reloc | NOUVEAU_BO_LOW, 0, 0);
466	so_method(so, eng2d, NV50_2D_SIFC_BITMAP_ENABLE, 2);
467	so_data  (so, 0);
468	so_data  (so, NV50_2D_SIFC_FORMAT_R8_UNORM);
469	so_method(so, eng2d, NV50_2D_SIFC_WIDTH, 10);
470	so_data  (so, size);
471	so_data  (so, 1);
472	so_data  (so, 0);
473	so_data  (so, 1);
474	so_data  (so, 0);
475	so_data  (so, 1);
476	so_data  (so, 0);
477	so_data  (so, 0);
478	so_data  (so, 0);
479	so_data  (so, 0);
480}
481