nv50_state_validate.c revision 63cb6f59eac91ba34cf80ff3736568e40b094fe1
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
28#define NV50_CBUF_FORMAT_CASE(n) \
29	case PIPE_FORMAT_##n: so_data(so, NV50TCL_RT_FORMAT_##n); break
30
31#define NV50_ZETA_FORMAT_CASE(n) \
32	case PIPE_FORMAT_##n: so_data(so, NV50TCL_ZETA_FORMAT_##n); break
33
34static void
35nv50_state_validate_fb(struct nv50_context *nv50)
36{
37	struct nouveau_grobj *tesla = nv50->screen->tesla;
38	struct nouveau_stateobj *so = so_new(32, 79, 18);
39	struct pipe_framebuffer_state *fb = &nv50->framebuffer;
40	unsigned i, w, h, gw = 0;
41
42	/* Set nr of active RTs and select RT for each colour output.
43	 * FP result 0 always goes to RT[0], bits 4 - 6 are ignored.
44	 * Ambiguous assignment results in no rendering (no DATA_ERROR).
45	 */
46	so_method(so, tesla, NV50TCL_RT_CONTROL, 1);
47	so_data  (so, fb->nr_cbufs |
48		  (0 <<  4) | (1 <<  7) | (2 << 10) | (3 << 13) |
49		  (4 << 16) | (5 << 19) | (6 << 22) | (7 << 25));
50
51	for (i = 0; i < fb->nr_cbufs; i++) {
52		struct pipe_texture *pt = fb->cbufs[i]->texture;
53		struct nouveau_bo *bo = nv50_miptree(pt)->base.bo;
54
55		if (!gw) {
56			w = fb->cbufs[i]->width;
57			h = fb->cbufs[i]->height;
58			gw = 1;
59		} else {
60			assert(w == fb->cbufs[i]->width);
61			assert(h == fb->cbufs[i]->height);
62		}
63
64		so_method(so, tesla, NV50TCL_RT_HORIZ(i), 2);
65		so_data  (so, fb->cbufs[i]->width);
66		so_data  (so, fb->cbufs[i]->height);
67
68		so_method(so, tesla, NV50TCL_RT_ADDRESS_HIGH(i), 5);
69		so_reloc (so, bo, fb->cbufs[i]->offset, NOUVEAU_BO_VRAM |
70			      NOUVEAU_BO_HIGH | NOUVEAU_BO_RDWR, 0, 0);
71		so_reloc (so, bo, fb->cbufs[i]->offset, NOUVEAU_BO_VRAM |
72			      NOUVEAU_BO_LOW | NOUVEAU_BO_RDWR, 0, 0);
73		switch (fb->cbufs[i]->format) {
74		NV50_CBUF_FORMAT_CASE(A8R8G8B8_UNORM);
75		NV50_CBUF_FORMAT_CASE(X8R8G8B8_UNORM);
76		NV50_CBUF_FORMAT_CASE(R5G6B5_UNORM);
77		NV50_CBUF_FORMAT_CASE(R16G16B16A16_SNORM);
78		NV50_CBUF_FORMAT_CASE(R16G16B16A16_UNORM);
79		NV50_CBUF_FORMAT_CASE(R32G32B32A32_FLOAT);
80		NV50_CBUF_FORMAT_CASE(R16G16_SNORM);
81		NV50_CBUF_FORMAT_CASE(R16G16_UNORM);
82		default:
83			NOUVEAU_ERR("AIIII unknown format %s\n",
84			            util_format_name(fb->cbufs[i]->format));
85			so_data(so, NV50TCL_RT_FORMAT_X8R8G8B8_UNORM);
86			break;
87		}
88		so_data(so, nv50_miptree(pt)->
89				level[fb->cbufs[i]->level].tile_mode << 4);
90		so_data(so, 0x00000000);
91
92		so_method(so, tesla, NV50TCL_RT_ARRAY_MODE, 1);
93		so_data  (so, 1);
94	}
95
96	if (fb->zsbuf) {
97		struct pipe_texture *pt = fb->zsbuf->texture;
98		struct nouveau_bo *bo = nv50_miptree(pt)->base.bo;
99
100		if (!gw) {
101			w = fb->zsbuf->width;
102			h = fb->zsbuf->height;
103			gw = 1;
104		} else {
105			assert(w == fb->zsbuf->width);
106			assert(h == fb->zsbuf->height);
107		}
108
109		so_method(so, tesla, NV50TCL_ZETA_ADDRESS_HIGH, 5);
110		so_reloc (so, bo, fb->zsbuf->offset, NOUVEAU_BO_VRAM |
111			      NOUVEAU_BO_HIGH | NOUVEAU_BO_RDWR, 0, 0);
112		so_reloc (so, bo, fb->zsbuf->offset, NOUVEAU_BO_VRAM |
113			      NOUVEAU_BO_LOW | NOUVEAU_BO_RDWR, 0, 0);
114		switch (fb->zsbuf->format) {
115		NV50_ZETA_FORMAT_CASE(S8Z24_UNORM);
116		NV50_ZETA_FORMAT_CASE(X8Z24_UNORM);
117		NV50_ZETA_FORMAT_CASE(Z24S8_UNORM);
118		NV50_ZETA_FORMAT_CASE(Z32_FLOAT);
119		default:
120			NOUVEAU_ERR("AIIII unknown format %s\n",
121			            util_format_name(fb->zsbuf->format));
122			so_data(so, NV50TCL_ZETA_FORMAT_S8Z24_UNORM);
123			break;
124		}
125		so_data(so, nv50_miptree(pt)->
126				level[fb->zsbuf->level].tile_mode << 4);
127		so_data(so, 0x00000000);
128
129		so_method(so, tesla, NV50TCL_ZETA_ENABLE, 1);
130		so_data  (so, 1);
131		so_method(so, tesla, NV50TCL_ZETA_HORIZ, 3);
132		so_data  (so, fb->zsbuf->width);
133		so_data  (so, fb->zsbuf->height);
134		so_data  (so, 0x00010001);
135	} else {
136		so_method(so, tesla, NV50TCL_ZETA_ENABLE, 1);
137		so_data  (so, 0);
138	}
139
140	so_method(so, tesla, NV50TCL_VIEWPORT_HORIZ(0), 2);
141	so_data  (so, w << 16);
142	so_data  (so, h << 16);
143	/* set window lower left corner */
144	so_method(so, tesla, NV50TCL_WINDOW_OFFSET_X, 2);
145	so_data  (so, 0);
146	so_data  (so, 0);
147	/* set screen scissor rectangle */
148	so_method(so, tesla, NV50TCL_SCREEN_SCISSOR_HORIZ, 2);
149	so_data  (so, w << 16);
150	so_data  (so, h << 16);
151
152	/* we set scissors to framebuffer size when they're 'turned off' */
153	nv50->dirty |= NV50_NEW_SCISSOR;
154	so_ref(NULL, &nv50->state.scissor);
155
156	so_ref(so, &nv50->state.fb);
157	so_ref(NULL, &so);
158}
159
160static void
161nv50_validate_samplers(struct nv50_context *nv50, struct nouveau_stateobj *so,
162		       unsigned p)
163{
164	struct nouveau_grobj *eng2d = nv50->screen->eng2d;
165	unsigned i, j, dw = nv50->sampler_nr[p] * 8;
166
167	if (!dw)
168		return;
169	nv50_so_init_sifc(nv50, so, nv50->screen->tsc, NOUVEAU_BO_VRAM,
170			  p * (32 * 8 * 4), dw * 4);
171
172	so_method(so, eng2d, NV50_2D_SIFC_DATA | (2 << 29), dw);
173
174	for (i = 0; i < nv50->sampler_nr[p]; ++i) {
175		if (nv50->sampler[p][i])
176			so_datap(so, nv50->sampler[p][i]->tsc, 8);
177		else {
178			for (j = 0; j < 8; ++j) /* you get punished */
179				so_data(so, 0); /* ... for leaving holes */
180		}
181	}
182}
183
184static void
185nv50_state_emit(struct nv50_context *nv50)
186{
187	struct nv50_screen *screen = nv50->screen;
188	struct nouveau_channel *chan = screen->base.channel;
189
190	/* XXX: this is racy for multiple contexts active on separate
191	 * threads.
192	 */
193	if (screen->cur_ctx != nv50) {
194		if (nv50->state.fb)
195			nv50->state.dirty |= NV50_NEW_FRAMEBUFFER;
196		if (nv50->state.blend)
197			nv50->state.dirty |= NV50_NEW_BLEND;
198		if (nv50->state.zsa)
199			nv50->state.dirty |= NV50_NEW_ZSA;
200		if (nv50->state.vertprog)
201			nv50->state.dirty |= NV50_NEW_VERTPROG;
202		if (nv50->state.fragprog)
203			nv50->state.dirty |= NV50_NEW_FRAGPROG;
204		if (nv50->state.geomprog)
205			nv50->state.dirty |= NV50_NEW_GEOMPROG;
206		if (nv50->state.rast)
207			nv50->state.dirty |= NV50_NEW_RASTERIZER;
208		if (nv50->state.blend_colour)
209			nv50->state.dirty |= NV50_NEW_BLEND_COLOUR;
210		if (nv50->state.stencil_ref)
211			nv50->state.dirty |= NV50_NEW_STENCIL_REF;
212		if (nv50->state.stipple)
213			nv50->state.dirty |= NV50_NEW_STIPPLE;
214		if (nv50->state.scissor)
215			nv50->state.dirty |= NV50_NEW_SCISSOR;
216		if (nv50->state.viewport)
217			nv50->state.dirty |= NV50_NEW_VIEWPORT;
218		if (nv50->state.tsc_upload)
219			nv50->state.dirty |= NV50_NEW_SAMPLER;
220		if (nv50->state.tic_upload)
221			nv50->state.dirty |= NV50_NEW_TEXTURE;
222		if (nv50->state.vtxfmt && nv50->state.vtxbuf)
223			nv50->state.dirty |= NV50_NEW_ARRAYS;
224		screen->cur_ctx = nv50;
225	}
226
227	if (nv50->state.dirty & NV50_NEW_FRAMEBUFFER)
228		so_emit(chan, nv50->state.fb);
229	if (nv50->state.dirty & NV50_NEW_BLEND)
230		so_emit(chan, nv50->state.blend);
231	if (nv50->state.dirty & NV50_NEW_ZSA)
232		so_emit(chan, nv50->state.zsa);
233	if (nv50->state.dirty & NV50_NEW_VERTPROG)
234		so_emit(chan, nv50->state.vertprog);
235	if (nv50->state.dirty & NV50_NEW_FRAGPROG)
236		so_emit(chan, nv50->state.fragprog);
237	if (nv50->state.dirty & NV50_NEW_GEOMPROG && nv50->state.geomprog)
238		so_emit(chan, nv50->state.geomprog);
239	if (nv50->state.dirty & (NV50_NEW_FRAGPROG | NV50_NEW_VERTPROG |
240				 NV50_NEW_GEOMPROG | NV50_NEW_RASTERIZER))
241		so_emit(chan, nv50->state.fp_linkage);
242	if ((nv50->state.dirty & (NV50_NEW_VERTPROG | NV50_NEW_GEOMPROG))
243	    && nv50->state.gp_linkage)
244		so_emit(chan, nv50->state.gp_linkage);
245	if (nv50->state.dirty & NV50_NEW_RASTERIZER)
246		so_emit(chan, nv50->state.rast);
247	if (nv50->state.dirty & NV50_NEW_BLEND_COLOUR)
248		so_emit(chan, nv50->state.blend_colour);
249	if (nv50->state.dirty & NV50_NEW_STENCIL_REF)
250		so_emit(chan, nv50->state.stencil_ref);
251	if (nv50->state.dirty & NV50_NEW_STIPPLE)
252		so_emit(chan, nv50->state.stipple);
253	if (nv50->state.dirty & NV50_NEW_SCISSOR)
254		so_emit(chan, nv50->state.scissor);
255	if (nv50->state.dirty & NV50_NEW_VIEWPORT)
256		so_emit(chan, nv50->state.viewport);
257	if (nv50->state.dirty & NV50_NEW_SAMPLER)
258		so_emit(chan, nv50->state.tsc_upload);
259	if (nv50->state.dirty & NV50_NEW_TEXTURE)
260		so_emit(chan, nv50->state.tic_upload);
261	if (nv50->state.dirty & NV50_NEW_ARRAYS) {
262		so_emit(chan, nv50->state.vtxfmt);
263		so_emit(chan, nv50->state.vtxbuf);
264		if (nv50->state.vtxattr)
265			so_emit(chan, nv50->state.vtxattr);
266	}
267	nv50->state.dirty = 0;
268}
269
270void
271nv50_state_flush_notify(struct nouveau_channel *chan)
272{
273	struct nv50_context *nv50 = chan->user_private;
274
275	if (nv50->state.tic_upload && !(nv50->dirty & NV50_NEW_TEXTURE))
276		so_emit(chan, nv50->state.tic_upload);
277
278	so_emit_reloc_markers(chan, nv50->state.fb);
279	so_emit_reloc_markers(chan, nv50->state.vertprog);
280	so_emit_reloc_markers(chan, nv50->state.fragprog);
281	so_emit_reloc_markers(chan, nv50->state.vtxbuf);
282	so_emit_reloc_markers(chan, nv50->screen->static_init);
283
284	if (nv50->state.instbuf)
285		so_emit_reloc_markers(chan, nv50->state.instbuf);
286}
287
288boolean
289nv50_state_validate(struct nv50_context *nv50)
290{
291	struct nouveau_grobj *tesla = nv50->screen->tesla;
292	struct nouveau_stateobj *so;
293	unsigned i;
294
295	if (nv50->dirty & NV50_NEW_FRAMEBUFFER)
296		nv50_state_validate_fb(nv50);
297
298	if (nv50->dirty & NV50_NEW_BLEND)
299		so_ref(nv50->blend->so, &nv50->state.blend);
300
301	if (nv50->dirty & NV50_NEW_ZSA)
302		so_ref(nv50->zsa->so, &nv50->state.zsa);
303
304	if (nv50->dirty & (NV50_NEW_VERTPROG | NV50_NEW_VERTPROG_CB))
305		nv50_vertprog_validate(nv50);
306
307	if (nv50->dirty & (NV50_NEW_FRAGPROG | NV50_NEW_FRAGPROG_CB))
308		nv50_fragprog_validate(nv50);
309
310	if (nv50->dirty & (NV50_NEW_GEOMPROG | NV50_NEW_GEOMPROG_CB))
311		nv50_geomprog_validate(nv50);
312
313	if (nv50->dirty & (NV50_NEW_FRAGPROG | NV50_NEW_VERTPROG |
314			   NV50_NEW_GEOMPROG | NV50_NEW_RASTERIZER))
315		nv50_fp_linkage_validate(nv50);
316
317	if (nv50->dirty & (NV50_NEW_GEOMPROG | NV50_NEW_VERTPROG))
318		nv50_gp_linkage_validate(nv50);
319
320	if (nv50->dirty & NV50_NEW_RASTERIZER)
321		so_ref(nv50->rasterizer->so, &nv50->state.rast);
322
323	if (nv50->dirty & NV50_NEW_BLEND_COLOUR) {
324		so = so_new(1, 4, 0);
325		so_method(so, tesla, NV50TCL_BLEND_COLOR(0), 4);
326		so_data  (so, fui(nv50->blend_colour.color[0]));
327		so_data  (so, fui(nv50->blend_colour.color[1]));
328		so_data  (so, fui(nv50->blend_colour.color[2]));
329		so_data  (so, fui(nv50->blend_colour.color[3]));
330		so_ref(so, &nv50->state.blend_colour);
331		so_ref(NULL, &so);
332	}
333
334	if (nv50->dirty & NV50_NEW_STENCIL_REF) {
335		so = so_new(2, 2, 0);
336		so_method(so, tesla, NV50TCL_STENCIL_FRONT_FUNC_REF, 1);
337		so_data  (so, nv50->stencil_ref.ref_value[0]);
338		so_method(so, tesla, NV50TCL_STENCIL_BACK_FUNC_REF, 1);
339		so_data  (so, nv50->stencil_ref.ref_value[1]);
340		so_ref(so, &nv50->state.stencil_ref);
341		so_ref(NULL, &so);
342	}
343
344	if (nv50->dirty & NV50_NEW_STIPPLE) {
345		so = so_new(1, 32, 0);
346		so_method(so, tesla, NV50TCL_POLYGON_STIPPLE_PATTERN(0), 32);
347		for (i = 0; i < 32; i++)
348			so_data(so, util_bswap32(nv50->stipple.stipple[i]));
349		so_ref(so, &nv50->state.stipple);
350		so_ref(NULL, &so);
351	}
352
353	if (nv50->dirty & (NV50_NEW_SCISSOR | NV50_NEW_RASTERIZER)) {
354		struct pipe_rasterizer_state *rast = &nv50->rasterizer->pipe;
355		struct pipe_scissor_state *s = &nv50->scissor;
356
357		if (nv50->state.scissor &&
358		    (rast->scissor == 0 && nv50->state.scissor_enabled == 0))
359			goto scissor_uptodate;
360		nv50->state.scissor_enabled = rast->scissor;
361
362		so = so_new(1, 2, 0);
363		so_method(so, tesla, NV50TCL_SCISSOR_HORIZ(0), 2);
364		if (nv50->state.scissor_enabled) {
365			so_data(so, (s->maxx << 16) | s->minx);
366			so_data(so, (s->maxy << 16) | s->miny);
367		} else {
368			so_data(so, (nv50->framebuffer.width << 16));
369			so_data(so, (nv50->framebuffer.height << 16));
370		}
371		so_ref(so, &nv50->state.scissor);
372		so_ref(NULL, &so);
373		nv50->state.dirty |= NV50_NEW_SCISSOR;
374	}
375scissor_uptodate:
376
377	if (nv50->dirty & (NV50_NEW_VIEWPORT | NV50_NEW_RASTERIZER)) {
378		if (nv50->state.viewport &&
379		    !(nv50->dirty & NV50_NEW_VIEWPORT))
380			goto viewport_uptodate;
381
382		so = so_new(5, 9, 0);
383		so_method(so, tesla, NV50TCL_VIEWPORT_TRANSLATE_X(0), 3);
384		so_data  (so, fui(nv50->viewport.translate[0]));
385		so_data  (so, fui(nv50->viewport.translate[1]));
386		so_data  (so, fui(nv50->viewport.translate[2]));
387		so_method(so, tesla, NV50TCL_VIEWPORT_SCALE_X(0), 3);
388		so_data  (so, fui(nv50->viewport.scale[0]));
389		so_data  (so, fui(nv50->viewport.scale[1]));
390		so_data  (so, fui(nv50->viewport.scale[2]));
391
392		so_method(so, tesla, NV50TCL_VIEWPORT_TRANSFORM_EN, 1);
393		so_data  (so, 1);
394		/* 0x0000 = remove whole primitive only (xyz)
395		 * 0x1018 = remove whole primitive only (xy), clamp z
396		 * 0x1080 = clip primitive (xyz)
397		 * 0x1098 = clip primitive (xy), clamp z
398		 */
399		so_method(so, tesla, NV50TCL_VIEW_VOLUME_CLIP_CTRL, 1);
400		so_data  (so, 0x1080);
401		/* no idea what 0f90 does */
402		so_method(so, tesla, 0x0f90, 1);
403		so_data  (so, 0);
404
405		so_ref(so, &nv50->state.viewport);
406		so_ref(NULL, &so);
407		nv50->state.dirty |= NV50_NEW_VIEWPORT;
408	}
409viewport_uptodate:
410
411	if (nv50->dirty & NV50_NEW_SAMPLER) {
412		unsigned nr = 0;
413
414		for (i = 0; i < PIPE_SHADER_TYPES; ++i)
415			nr += nv50->sampler_nr[i];
416
417		so = so_new(1 + 5 * PIPE_SHADER_TYPES,
418			    1 + 19 * PIPE_SHADER_TYPES + nr * 8,
419			    PIPE_SHADER_TYPES * 2);
420
421		nv50_validate_samplers(nv50, so, PIPE_SHADER_VERTEX);
422		nv50_validate_samplers(nv50, so, PIPE_SHADER_FRAGMENT);
423
424		so_method(so, tesla, 0x1334, 1); /* flush TSC */
425		so_data  (so, 0);
426
427		so_ref(so, &nv50->state.tsc_upload);
428		so_ref(NULL, &so);
429	}
430
431	if (nv50->dirty & (NV50_NEW_TEXTURE | NV50_NEW_SAMPLER))
432		nv50_tex_validate(nv50);
433
434	if (nv50->dirty & NV50_NEW_ARRAYS)
435		nv50_vbo_validate(nv50);
436
437	nv50->state.dirty |= nv50->dirty;
438	nv50->dirty = 0;
439	nv50_state_emit(nv50);
440
441	return TRUE;
442}
443
444void nv50_so_init_sifc(struct nv50_context *nv50,
445		       struct nouveau_stateobj *so,
446		       struct nouveau_bo *bo, unsigned reloc,
447		       unsigned offset, unsigned size)
448{
449	struct nouveau_grobj *eng2d = nv50->screen->eng2d;
450
451	reloc |= NOUVEAU_BO_WR;
452
453	so_method(so, eng2d, NV50_2D_DST_FORMAT, 2);
454	so_data  (so, NV50_2D_DST_FORMAT_R8_UNORM);
455	so_data  (so, 1);
456	so_method(so, eng2d, NV50_2D_DST_PITCH, 5);
457	so_data  (so, 262144);
458	so_data  (so, 65536);
459	so_data  (so, 1);
460	so_reloc (so, bo, offset, reloc | NOUVEAU_BO_HIGH, 0, 0);
461	so_reloc (so, bo, offset, reloc | NOUVEAU_BO_LOW, 0, 0);
462	so_method(so, eng2d, NV50_2D_SIFC_BITMAP_ENABLE, 2);
463	so_data  (so, 0);
464	so_data  (so, NV50_2D_SIFC_FORMAT_R8_UNORM);
465	so_method(so, eng2d, NV50_2D_SIFC_WIDTH, 10);
466	so_data  (so, size);
467	so_data  (so, 1);
468	so_data  (so, 0);
469	so_data  (so, 1);
470	so_data  (so, 0);
471	so_data  (so, 1);
472	so_data  (so, 0);
473	so_data  (so, 0);
474	so_data  (so, 0);
475	so_data  (so, 0);
476}
477