nv50_state_validate.c revision ef715b866b6858b79530b44ec7be61f271591adc 
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
29static struct nouveau_stateobj *
30validate_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		assert(nv50_format_table[fb->cbufs[i]->format].rt);
60
61		so_method(so, tesla, NV50TCL_RT_HORIZ(i), 2);
62		so_data  (so, fb->cbufs[i]->width);
63		so_data  (so, fb->cbufs[i]->height);
64
65		so_method(so, tesla, NV50TCL_RT_ADDRESS_HIGH(i), 5);
66		so_reloc (so, bo, fb->cbufs[i]->offset, NOUVEAU_BO_VRAM |
67			      NOUVEAU_BO_HIGH | NOUVEAU_BO_RDWR, 0, 0);
68		so_reloc (so, bo, fb->cbufs[i]->offset, NOUVEAU_BO_VRAM |
69			      NOUVEAU_BO_LOW | NOUVEAU_BO_RDWR, 0, 0);
70		so_data  (so, nv50_format_table[fb->cbufs[i]->format].rt);
71		so_data  (so, nv50_miptree(pt)->
72			      level[fb->cbufs[i]->level].tile_mode << 4);
73		so_data(so, 0x00000000);
74
75		so_method(so, tesla, NV50TCL_RT_ARRAY_MODE, 1);
76		so_data  (so, 1);
77	}
78
79	if (fb->zsbuf) {
80		struct pipe_resource *pt = fb->zsbuf->texture;
81		struct nouveau_bo *bo = nv50_miptree(pt)->base.bo;
82
83		if (!gw) {
84			w = fb->zsbuf->width;
85			h = fb->zsbuf->height;
86			gw = 1;
87		} else {
88			assert(w == fb->zsbuf->width);
89			assert(h == fb->zsbuf->height);
90		}
91
92		assert(nv50_format_table[fb->zsbuf->format].rt);
93
94		so_method(so, tesla, NV50TCL_ZETA_ADDRESS_HIGH, 5);
95		so_reloc (so, bo, fb->zsbuf->offset, NOUVEAU_BO_VRAM |
96			      NOUVEAU_BO_HIGH | NOUVEAU_BO_RDWR, 0, 0);
97		so_reloc (so, bo, fb->zsbuf->offset, NOUVEAU_BO_VRAM |
98			      NOUVEAU_BO_LOW | NOUVEAU_BO_RDWR, 0, 0);
99		so_data  (so, nv50_format_table[fb->zsbuf->format].rt);
100		so_data  (so, nv50_miptree(pt)->
101			      level[fb->zsbuf->level].tile_mode << 4);
102		so_data  (so, 0x00000000);
103
104		so_method(so, tesla, NV50TCL_ZETA_ENABLE, 1);
105		so_data  (so, 1);
106		so_method(so, tesla, NV50TCL_ZETA_HORIZ, 3);
107		so_data  (so, fb->zsbuf->width);
108		so_data  (so, fb->zsbuf->height);
109		so_data  (so, 0x00010001);
110	} else {
111		so_method(so, tesla, NV50TCL_ZETA_ENABLE, 1);
112		so_data  (so, 0);
113	}
114
115	so_method(so, tesla, NV50TCL_VIEWPORT_HORIZ(0), 2);
116	so_data  (so, w << 16);
117	so_data  (so, h << 16);
118	/* set window lower left corner */
119	so_method(so, tesla, NV50TCL_WINDOW_OFFSET_X, 2);
120	so_data  (so, 0);
121	so_data  (so, 0);
122	/* set screen scissor rectangle */
123	so_method(so, tesla, NV50TCL_SCREEN_SCISSOR_HORIZ, 2);
124	so_data  (so, w << 16);
125	so_data  (so, h << 16);
126
127	return so;
128}
129
130static void
131nv50_validate_samplers(struct nv50_context *nv50, struct nouveau_stateobj *so,
132		       unsigned p)
133{
134	struct nouveau_grobj *eng2d = nv50->screen->eng2d;
135	unsigned i, j, dw = nv50->sampler_nr[p] * 8;
136
137	if (!dw)
138		return;
139	nv50_so_init_sifc(nv50, so, nv50->screen->tsc, NOUVEAU_BO_VRAM,
140			  p * (32 * 8 * 4), dw * 4);
141
142	so_method(so, eng2d, NV50_2D_SIFC_DATA | (2 << 29), dw);
143
144	for (i = 0; i < nv50->sampler_nr[p]; ++i) {
145		if (nv50->sampler[p][i])
146			so_datap(so, nv50->sampler[p][i]->tsc, 8);
147		else {
148			for (j = 0; j < 8; ++j) /* you get punished */
149				so_data(so, 0); /* ... for leaving holes */
150		}
151	}
152}
153
154static struct nouveau_stateobj *
155validate_blend(struct nv50_context *nv50)
156{
157	struct nouveau_stateobj *so = NULL;
158	so_ref(nv50->blend->so, &so);
159	return so;
160}
161
162static struct nouveau_stateobj *
163validate_zsa(struct nv50_context *nv50)
164{
165	struct nouveau_stateobj *so = NULL;
166	so_ref(nv50->zsa->so, &so);
167	return so;
168}
169
170static struct nouveau_stateobj *
171validate_rast(struct nv50_context *nv50)
172{
173	struct nouveau_stateobj *so = NULL;
174	so_ref(nv50->rasterizer->so, &so);
175	return so;
176}
177
178static struct nouveau_stateobj *
179validate_blend_colour(struct nv50_context *nv50)
180{
181	struct nouveau_grobj *tesla = nv50->screen->tesla;
182	struct nouveau_stateobj *so = so_new(1, 4, 0);
183
184	so_method(so, tesla, NV50TCL_BLEND_COLOR(0), 4);
185	so_data  (so, fui(nv50->blend_colour.color[0]));
186	so_data  (so, fui(nv50->blend_colour.color[1]));
187	so_data  (so, fui(nv50->blend_colour.color[2]));
188	so_data  (so, fui(nv50->blend_colour.color[3]));
189	return so;
190}
191
192static struct nouveau_stateobj *
193validate_stencil_ref(struct nv50_context *nv50)
194{
195	struct nouveau_grobj *tesla = nv50->screen->tesla;
196	struct nouveau_stateobj *so = so_new(2, 2, 0);
197
198	so_method(so, tesla, NV50TCL_STENCIL_FRONT_FUNC_REF, 1);
199	so_data  (so, nv50->stencil_ref.ref_value[0]);
200	so_method(so, tesla, NV50TCL_STENCIL_BACK_FUNC_REF, 1);
201	so_data  (so, nv50->stencil_ref.ref_value[1]);
202	return so;
203}
204
205static struct nouveau_stateobj *
206validate_stipple(struct nv50_context *nv50)
207{
208	struct nouveau_grobj *tesla = nv50->screen->tesla;
209	struct nouveau_stateobj *so = so_new(1, 32, 0);
210	int i;
211
212	so_method(so, tesla, NV50TCL_POLYGON_STIPPLE_PATTERN(0), 32);
213	for (i = 0; i < 32; i++)
214		so_data(so, util_bswap32(nv50->stipple.stipple[i]));
215	return so;
216}
217
218static struct nouveau_stateobj *
219validate_scissor(struct nv50_context *nv50)
220{
221	struct nouveau_grobj *tesla = nv50->screen->tesla;
222        struct pipe_scissor_state *s = &nv50->scissor;
223	struct nouveau_stateobj *so;
224
225	so = so_new(1, 2, 0);
226	so_method(so, tesla, NV50TCL_SCISSOR_HORIZ(0), 2);
227	so_data  (so, (s->maxx << 16) | s->minx);
228	so_data  (so, (s->maxy << 16) | s->miny);
229	return so;
230}
231
232static struct nouveau_stateobj *
233validate_viewport(struct nv50_context *nv50)
234{
235	struct nouveau_grobj *tesla = nv50->screen->tesla;
236	struct nouveau_stateobj *so = so_new(3, 7, 0);
237
238	so_method(so, tesla, NV50TCL_VIEWPORT_TRANSLATE_X(0), 3);
239	so_data  (so, fui(nv50->viewport.translate[0]));
240	so_data  (so, fui(nv50->viewport.translate[1]));
241	so_data  (so, fui(nv50->viewport.translate[2]));
242	so_method(so, tesla, NV50TCL_VIEWPORT_SCALE_X(0), 3);
243	so_data  (so, fui(nv50->viewport.scale[0]));
244	so_data  (so, fui(nv50->viewport.scale[1]));
245	so_data  (so, fui(nv50->viewport.scale[2]));
246
247	/* no idea what 0f90 does */
248	so_method(so, tesla, 0x0f90, 1);
249	so_data  (so, 0);
250
251	return so;
252}
253
254static struct nouveau_stateobj *
255validate_sampler(struct nv50_context *nv50)
256{
257	struct nouveau_grobj *tesla = nv50->screen->tesla;
258	struct nouveau_stateobj *so;
259	unsigned nr = 0, i;
260
261	for (i = 0; i < 3; ++i)
262		nr += nv50->sampler_nr[i];
263
264	so = so_new(1 + 5 * 3, 1 + 19 * 3 + nr * 8, 3 * 2);
265
266	nv50_validate_samplers(nv50, so, 0); /* VP */
267	nv50_validate_samplers(nv50, so, 2); /* FP */
268
269	so_method(so, tesla, 0x1334, 1); /* flush TSC */
270	so_data  (so, 0);
271
272	return so;
273}
274
275static struct nouveau_stateobj *
276validate_vtxbuf(struct nv50_context *nv50)
277{
278	struct nouveau_stateobj *so = NULL;
279	so_ref(nv50->state.vtxbuf, &so);
280	return so;
281}
282
283static struct nouveau_stateobj *
284validate_vtxattr(struct nv50_context *nv50)
285{
286	struct nouveau_stateobj *so = NULL;
287	so_ref(nv50->state.vtxattr, &so);
288	return so;
289}
290
291static struct nouveau_stateobj *
292validate_clip(struct nv50_context *nv50)
293{
294	struct nouveau_grobj *tesla = nv50->screen->tesla;
295	struct nouveau_stateobj *so = so_new(1, 1, 0);
296	uint32_t vvcc;
297
298	/* 0x0000 = remove whole primitive only (xyz)
299	 * 0x1018 = remove whole primitive only (xy), clamp z
300	 * 0x1080 = clip primitive (xyz)
301	 * 0x1098 = clip primitive (xy), clamp z
302	 */
303	vvcc = nv50->clip.depth_clamp ? 0x1098 : 0x1080;
304
305	so_method(so, tesla, NV50TCL_VIEW_VOLUME_CLIP_CTRL, 1);
306	so_data  (so, vvcc);
307
308	return so;
309}
310
311struct state_validate {
312	struct nouveau_stateobj *(*func)(struct nv50_context *nv50);
313	unsigned states;
314} validate_list[] = {
315	{ validate_fb             , NV50_NEW_FRAMEBUFFER                      },
316	{ validate_blend          , NV50_NEW_BLEND                            },
317	{ validate_zsa            , NV50_NEW_ZSA                              },
318	{ nv50_vertprog_validate  , NV50_NEW_VERTPROG | NV50_NEW_VERTPROG_CB  },
319	{ nv50_fragprog_validate  , NV50_NEW_FRAGPROG | NV50_NEW_FRAGPROG_CB  },
320	{ nv50_geomprog_validate  , NV50_NEW_GEOMPROG | NV50_NEW_GEOMPROG_CB  },
321	{ nv50_fp_linkage_validate, NV50_NEW_VERTPROG | NV50_NEW_GEOMPROG |
322				    NV50_NEW_FRAGPROG | NV50_NEW_RASTERIZER   },
323	{ nv50_gp_linkage_validate, NV50_NEW_VERTPROG | NV50_NEW_GEOMPROG     },
324	{ validate_rast           , NV50_NEW_RASTERIZER                       },
325	{ validate_blend_colour   , NV50_NEW_BLEND_COLOUR                     },
326	{ validate_stencil_ref    , NV50_NEW_STENCIL_REF                      },
327	{ validate_stipple        , NV50_NEW_STIPPLE                          },
328	{ validate_scissor        , NV50_NEW_SCISSOR                          },
329	{ validate_viewport       , NV50_NEW_VIEWPORT                         },
330	{ validate_sampler        , NV50_NEW_SAMPLER                          },
331	{ nv50_tex_validate       , NV50_NEW_TEXTURE | NV50_NEW_SAMPLER       },
332	{ nv50_vbo_validate       , NV50_NEW_ARRAYS                           },
333	{ validate_vtxbuf         , NV50_NEW_ARRAYS                           },
334	{ validate_vtxattr        , NV50_NEW_ARRAYS                           },
335	{ validate_clip           , NV50_NEW_CLIP                             },
336	{ NULL                    , 0                                         }
337};
338#define validate_list_len (sizeof(validate_list) / sizeof(validate_list[0]))
339
340boolean
341nv50_state_validate(struct nv50_context *nv50, unsigned wait_dwords)
342{
343	struct nouveau_channel *chan = nv50->screen->base.channel;
344	struct nouveau_grobj *tesla = nv50->screen->tesla;
345	unsigned nr_relocs = 128, nr_dwords = wait_dwords + 128 + 4;
346	int ret, i;
347
348	for (i = 0; i < validate_list_len; i++) {
349		struct state_validate *validate = &validate_list[i];
350		struct nouveau_stateobj *so;
351
352		if (!(nv50->dirty & validate->states))
353			continue;
354
355		so = validate->func(nv50);
356		if (!so)
357			continue;
358
359		nr_dwords += (so->total + so->cur);
360		nr_relocs += so->cur_reloc;
361
362		so_ref(so, &nv50->state.hw[i]);
363		so_ref(NULL, &so);
364		nv50->state.hw_dirty |= (1 << i);
365	}
366	nv50->dirty = 0;
367
368	if (nv50->screen->cur_ctx != nv50) {
369		for (i = 0; i < validate_list_len; i++) {
370			if (!nv50->state.hw[i] ||
371			    (nv50->state.hw_dirty & (1 << i)))
372				continue;
373
374			nr_dwords += (nv50->state.hw[i]->total +
375				      nv50->state.hw[i]->cur);
376			nr_relocs += nv50->state.hw[i]->cur_reloc;
377			nv50->state.hw_dirty |= (1 << i);
378		}
379
380		nv50->screen->cur_ctx = nv50;
381	}
382
383	ret = MARK_RING(chan, nr_dwords, nr_relocs);
384	if (ret) {
385		debug_printf("MARK_RING(%d, %d) failed: %d\n",
386			     nr_dwords, nr_relocs, ret);
387		return FALSE;
388	}
389
390	while (nv50->state.hw_dirty) {
391		i = ffs(nv50->state.hw_dirty) - 1;
392		nv50->state.hw_dirty &= ~(1 << i);
393
394		so_emit(chan, nv50->state.hw[i]);
395	}
396
397	/* Yes, really, we need to do this.  If a buffer that is referenced
398	 * on the hardware isn't part of changed state above, without doing
399	 * this the kernel is given no clue that the buffer is being used
400	 * still.  This can cause all sorts of fun issues.
401	 */
402	nv50_tex_relocs(nv50);
403	so_emit_reloc_markers(chan, nv50->state.hw[0]); /* fb */
404	so_emit_reloc_markers(chan, nv50->state.hw[3]); /* vp */
405	so_emit_reloc_markers(chan, nv50->state.hw[4]); /* fp */
406	so_emit_reloc_markers(chan, nv50->state.hw[17]); /* vb */
407	nv50_screen_relocs(nv50->screen);
408
409	/* No idea.. */
410	BEGIN_RING(chan, tesla, 0x142c, 1);
411	OUT_RING  (chan, 0);
412	BEGIN_RING(chan, tesla, 0x142c, 1);
413	OUT_RING  (chan, 0);
414	return TRUE;
415}
416
417void nv50_so_init_sifc(struct nv50_context *nv50,
418		       struct nouveau_stateobj *so,
419		       struct nouveau_bo *bo, unsigned reloc,
420		       unsigned offset, unsigned size)
421{
422	struct nouveau_grobj *eng2d = nv50->screen->eng2d;
423
424	reloc |= NOUVEAU_BO_WR;
425
426	so_method(so, eng2d, NV50_2D_DST_FORMAT, 2);
427	so_data  (so, NV50_2D_DST_FORMAT_R8_UNORM);
428	so_data  (so, 1);
429	so_method(so, eng2d, NV50_2D_DST_PITCH, 5);
430	so_data  (so, 262144);
431	so_data  (so, 65536);
432	so_data  (so, 1);
433	so_reloc (so, bo, offset, reloc | NOUVEAU_BO_HIGH, 0, 0);
434	so_reloc (so, bo, offset, reloc | NOUVEAU_BO_LOW, 0, 0);
435	so_method(so, eng2d, NV50_2D_SIFC_BITMAP_ENABLE, 2);
436	so_data  (so, 0);
437	so_data  (so, NV50_2D_SIFC_FORMAT_R8_UNORM);
438	so_method(so, eng2d, NV50_2D_SIFC_WIDTH, 10);
439	so_data  (so, size);
440	so_data  (so, 1);
441	so_data  (so, 0);
442	so_data  (so, 1);
443	so_data  (so, 0);
444	so_data  (so, 1);
445	so_data  (so, 0);
446	so_data  (so, 0);
447	so_data  (so, 0);
448	so_data  (so, 0);
449}
450