projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
gallium: change set_constant_buffer to be UBO-friendly
[mesa.git] / src / gallium / drivers / r600 / evergreen_state.c
1 /*
2 * Copyright 2010 Jerome Glisse <glisse@freedesktop.org>
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 * on the rights to use, copy, modify, merge, publish, distribute, sub
8 * license, and/or sell copies of the Software, and to permit persons to whom
9 * the Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
19 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
20 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21 * USE OR OTHER DEALINGS IN THE SOFTWARE.
22 */
23 #include "r600_formats.h"
24 #include "evergreend.h"
25
26 #include "pipe/p_shader_tokens.h"
27 #include "util/u_pack_color.h"
28 #include "util/u_memory.h"
29 #include "util/u_framebuffer.h"
30 #include "util/u_dual_blend.h"
31
32 static uint32_t eg_num_banks(uint32_t nbanks)
33 {
34 switch (nbanks) {
35 case 2:
36 return 0;
37 case 4:
38 return 1;
39 case 8:
40 default:
41 return 2;
42 case 16:
43 return 3;
44 }
45 }
46
47
48 static unsigned eg_tile_split(unsigned tile_split)
49 {
50 switch (tile_split) {
51 case 64: tile_split = 0; break;
52 case 128: tile_split = 1; break;
53 case 256: tile_split = 2; break;
54 case 512: tile_split = 3; break;
55 default:
56 case 1024: tile_split = 4; break;
57 case 2048: tile_split = 5; break;
58 case 4096: tile_split = 6; break;
59 }
60 return tile_split;
61 }
62
63 static unsigned eg_macro_tile_aspect(unsigned macro_tile_aspect)
64 {
65 switch (macro_tile_aspect) {
66 default:
67 case 1: macro_tile_aspect = 0; break;
68 case 2: macro_tile_aspect = 1; break;
69 case 4: macro_tile_aspect = 2; break;
70 case 8: macro_tile_aspect = 3; break;
71 }
72 return macro_tile_aspect;
73 }
74
75 static unsigned eg_bank_wh(unsigned bankwh)
76 {
77 switch (bankwh) {
78 default:
79 case 1: bankwh = 0; break;
80 case 2: bankwh = 1; break;
81 case 4: bankwh = 2; break;
82 case 8: bankwh = 3; break;
83 }
84 return bankwh;
85 }
86
87 static uint32_t r600_translate_blend_function(int blend_func)
88 {
89 switch (blend_func) {
90 case PIPE_BLEND_ADD:
91 return V_028780_COMB_DST_PLUS_SRC;
92 case PIPE_BLEND_SUBTRACT:
93 return V_028780_COMB_SRC_MINUS_DST;
94 case PIPE_BLEND_REVERSE_SUBTRACT:
95 return V_028780_COMB_DST_MINUS_SRC;
96 case PIPE_BLEND_MIN:
97 return V_028780_COMB_MIN_DST_SRC;
98 case PIPE_BLEND_MAX:
99 return V_028780_COMB_MAX_DST_SRC;
100 default:
101 R600_ERR("Unknown blend function %d\n", blend_func);
102 assert(0);
103 break;
104 }
105 return 0;
106 }
107
108 static uint32_t r600_translate_blend_factor(int blend_fact)
109 {
110 switch (blend_fact) {
111 case PIPE_BLENDFACTOR_ONE:
112 return V_028780_BLEND_ONE;
113 case PIPE_BLENDFACTOR_SRC_COLOR:
114 return V_028780_BLEND_SRC_COLOR;
115 case PIPE_BLENDFACTOR_SRC_ALPHA:
116 return V_028780_BLEND_SRC_ALPHA;
117 case PIPE_BLENDFACTOR_DST_ALPHA:
118 return V_028780_BLEND_DST_ALPHA;
119 case PIPE_BLENDFACTOR_DST_COLOR:
120 return V_028780_BLEND_DST_COLOR;
121 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
122 return V_028780_BLEND_SRC_ALPHA_SATURATE;
123 case PIPE_BLENDFACTOR_CONST_COLOR:
124 return V_028780_BLEND_CONST_COLOR;
125 case PIPE_BLENDFACTOR_CONST_ALPHA:
126 return V_028780_BLEND_CONST_ALPHA;
127 case PIPE_BLENDFACTOR_ZERO:
128 return V_028780_BLEND_ZERO;
129 case PIPE_BLENDFACTOR_INV_SRC_COLOR:
130 return V_028780_BLEND_ONE_MINUS_SRC_COLOR;
131 case PIPE_BLENDFACTOR_INV_SRC_ALPHA:
132 return V_028780_BLEND_ONE_MINUS_SRC_ALPHA;
133 case PIPE_BLENDFACTOR_INV_DST_ALPHA:
134 return V_028780_BLEND_ONE_MINUS_DST_ALPHA;
135 case PIPE_BLENDFACTOR_INV_DST_COLOR:
136 return V_028780_BLEND_ONE_MINUS_DST_COLOR;
137 case PIPE_BLENDFACTOR_INV_CONST_COLOR:
138 return V_028780_BLEND_ONE_MINUS_CONST_COLOR;
139 case PIPE_BLENDFACTOR_INV_CONST_ALPHA:
140 return V_028780_BLEND_ONE_MINUS_CONST_ALPHA;
141 case PIPE_BLENDFACTOR_SRC1_COLOR:
142 return V_028780_BLEND_SRC1_COLOR;
143 case PIPE_BLENDFACTOR_SRC1_ALPHA:
144 return V_028780_BLEND_SRC1_ALPHA;
145 case PIPE_BLENDFACTOR_INV_SRC1_COLOR:
146 return V_028780_BLEND_INV_SRC1_COLOR;
147 case PIPE_BLENDFACTOR_INV_SRC1_ALPHA:
148 return V_028780_BLEND_INV_SRC1_ALPHA;
149 default:
150 R600_ERR("Bad blend factor %d not supported!\n", blend_fact);
151 assert(0);
152 break;
153 }
154 return 0;
155 }
156
157 static unsigned r600_tex_dim(unsigned dim)
158 {
159 switch (dim) {
160 default:
161 case PIPE_TEXTURE_1D:
162 return V_030000_SQ_TEX_DIM_1D;
163 case PIPE_TEXTURE_1D_ARRAY:
164 return V_030000_SQ_TEX_DIM_1D_ARRAY;
165 case PIPE_TEXTURE_2D:
166 case PIPE_TEXTURE_RECT:
167 return V_030000_SQ_TEX_DIM_2D;
168 case PIPE_TEXTURE_2D_ARRAY:
169 return V_030000_SQ_TEX_DIM_2D_ARRAY;
170 case PIPE_TEXTURE_3D:
171 return V_030000_SQ_TEX_DIM_3D;
172 case PIPE_TEXTURE_CUBE:
173 return V_030000_SQ_TEX_DIM_CUBEMAP;
174 }
175 }
176
177 static uint32_t r600_translate_dbformat(enum pipe_format format)
178 {
179 switch (format) {
180 case PIPE_FORMAT_Z16_UNORM:
181 return V_028040_Z_16;
182 case PIPE_FORMAT_Z24X8_UNORM:
183 case PIPE_FORMAT_Z24_UNORM_S8_UINT:
184 return V_028040_Z_24;
185 case PIPE_FORMAT_Z32_FLOAT:
186 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
187 return V_028040_Z_32_FLOAT;
188 default:
189 return ~0U;
190 }
191 }
192
193 static uint32_t r600_translate_colorswap(enum pipe_format format)
194 {
195 switch (format) {
196 /* 8-bit buffers. */
197 case PIPE_FORMAT_L4A4_UNORM:
198 case PIPE_FORMAT_A4R4_UNORM:
199 return V_028C70_SWAP_ALT;
200
201 case PIPE_FORMAT_A8_UNORM:
202 case PIPE_FORMAT_A8_SNORM:
203 case PIPE_FORMAT_A8_UINT:
204 case PIPE_FORMAT_A8_SINT:
205 case PIPE_FORMAT_A16_UNORM:
206 case PIPE_FORMAT_A16_SNORM:
207 case PIPE_FORMAT_A16_UINT:
208 case PIPE_FORMAT_A16_SINT:
209 case PIPE_FORMAT_A16_FLOAT:
210 case PIPE_FORMAT_A32_UINT:
211 case PIPE_FORMAT_A32_SINT:
212 case PIPE_FORMAT_A32_FLOAT:
213 case PIPE_FORMAT_R4A4_UNORM:
214 return V_028C70_SWAP_ALT_REV;
215 case PIPE_FORMAT_I8_UNORM:
216 case PIPE_FORMAT_I8_SNORM:
217 case PIPE_FORMAT_I8_UINT:
218 case PIPE_FORMAT_I8_SINT:
219 case PIPE_FORMAT_I16_UNORM:
220 case PIPE_FORMAT_I16_SNORM:
221 case PIPE_FORMAT_I16_UINT:
222 case PIPE_FORMAT_I16_SINT:
223 case PIPE_FORMAT_I16_FLOAT:
224 case PIPE_FORMAT_I32_UINT:
225 case PIPE_FORMAT_I32_SINT:
226 case PIPE_FORMAT_I32_FLOAT:
227 case PIPE_FORMAT_L8_UNORM:
228 case PIPE_FORMAT_L8_SNORM:
229 case PIPE_FORMAT_L8_UINT:
230 case PIPE_FORMAT_L8_SINT:
231 case PIPE_FORMAT_L8_SRGB:
232 case PIPE_FORMAT_L16_UNORM:
233 case PIPE_FORMAT_L16_SNORM:
234 case PIPE_FORMAT_L16_UINT:
235 case PIPE_FORMAT_L16_SINT:
236 case PIPE_FORMAT_L16_FLOAT:
237 case PIPE_FORMAT_L32_UINT:
238 case PIPE_FORMAT_L32_SINT:
239 case PIPE_FORMAT_L32_FLOAT:
240 case PIPE_FORMAT_R8_UNORM:
241 case PIPE_FORMAT_R8_SNORM:
242 case PIPE_FORMAT_R8_UINT:
243 case PIPE_FORMAT_R8_SINT:
244 return V_028C70_SWAP_STD;
245
246 /* 16-bit buffers. */
247 case PIPE_FORMAT_B5G6R5_UNORM:
248 return V_028C70_SWAP_STD_REV;
249
250 case PIPE_FORMAT_B5G5R5A1_UNORM:
251 case PIPE_FORMAT_B5G5R5X1_UNORM:
252 return V_028C70_SWAP_ALT;
253
254 case PIPE_FORMAT_B4G4R4A4_UNORM:
255 case PIPE_FORMAT_B4G4R4X4_UNORM:
256 return V_028C70_SWAP_ALT;
257
258 case PIPE_FORMAT_Z16_UNORM:
259 return V_028C70_SWAP_STD;
260
261 case PIPE_FORMAT_L8A8_UNORM:
262 case PIPE_FORMAT_L8A8_SNORM:
263 case PIPE_FORMAT_L8A8_UINT:
264 case PIPE_FORMAT_L8A8_SINT:
265 case PIPE_FORMAT_L8A8_SRGB:
266 case PIPE_FORMAT_L16A16_UNORM:
267 case PIPE_FORMAT_L16A16_SNORM:
268 case PIPE_FORMAT_L16A16_UINT:
269 case PIPE_FORMAT_L16A16_SINT:
270 case PIPE_FORMAT_L16A16_FLOAT:
271 case PIPE_FORMAT_L32A32_UINT:
272 case PIPE_FORMAT_L32A32_SINT:
273 case PIPE_FORMAT_L32A32_FLOAT:
274 return V_028C70_SWAP_ALT;
275 case PIPE_FORMAT_R8G8_UNORM:
276 case PIPE_FORMAT_R8G8_SNORM:
277 case PIPE_FORMAT_R8G8_UINT:
278 case PIPE_FORMAT_R8G8_SINT:
279 return V_028C70_SWAP_STD;
280
281 case PIPE_FORMAT_R16_UNORM:
282 case PIPE_FORMAT_R16_SNORM:
283 case PIPE_FORMAT_R16_UINT:
284 case PIPE_FORMAT_R16_SINT:
285 case PIPE_FORMAT_R16_FLOAT:
286 return V_028C70_SWAP_STD;
287
288 /* 32-bit buffers. */
289 case PIPE_FORMAT_A8B8G8R8_SRGB:
290 return V_028C70_SWAP_STD_REV;
291 case PIPE_FORMAT_B8G8R8A8_SRGB:
292 return V_028C70_SWAP_ALT;
293
294 case PIPE_FORMAT_B8G8R8A8_UNORM:
295 case PIPE_FORMAT_B8G8R8X8_UNORM:
296 return V_028C70_SWAP_ALT;
297
298 case PIPE_FORMAT_A8R8G8B8_UNORM:
299 case PIPE_FORMAT_X8R8G8B8_UNORM:
300 return V_028C70_SWAP_ALT_REV;
301 case PIPE_FORMAT_R8G8B8A8_SNORM:
302 case PIPE_FORMAT_R8G8B8A8_UNORM:
303 case PIPE_FORMAT_R8G8B8A8_SINT:
304 case PIPE_FORMAT_R8G8B8A8_UINT:
305 case PIPE_FORMAT_R8G8B8X8_UNORM:
306 return V_028C70_SWAP_STD;
307
308 case PIPE_FORMAT_A8B8G8R8_UNORM:
309 case PIPE_FORMAT_X8B8G8R8_UNORM:
310 /* case PIPE_FORMAT_R8SG8SB8UX8U_NORM: */
311 return V_028C70_SWAP_STD_REV;
312
313 case PIPE_FORMAT_Z24X8_UNORM:
314 case PIPE_FORMAT_Z24_UNORM_S8_UINT:
315 return V_028C70_SWAP_STD;
316
317 case PIPE_FORMAT_X8Z24_UNORM:
318 case PIPE_FORMAT_S8_UINT_Z24_UNORM:
319 return V_028C70_SWAP_STD;
320
321 case PIPE_FORMAT_R10G10B10A2_UNORM:
322 case PIPE_FORMAT_R10G10B10X2_SNORM:
323 case PIPE_FORMAT_R10SG10SB10SA2U_NORM:
324 return V_028C70_SWAP_STD;
325
326 case PIPE_FORMAT_B10G10R10A2_UNORM:
327 case PIPE_FORMAT_B10G10R10A2_UINT:
328 return V_028C70_SWAP_ALT;
329
330 case PIPE_FORMAT_R11G11B10_FLOAT:
331 case PIPE_FORMAT_R32_FLOAT:
332 case PIPE_FORMAT_R32_UINT:
333 case PIPE_FORMAT_R32_SINT:
334 case PIPE_FORMAT_Z32_FLOAT:
335 case PIPE_FORMAT_R16G16_FLOAT:
336 case PIPE_FORMAT_R16G16_UNORM:
337 case PIPE_FORMAT_R16G16_SNORM:
338 case PIPE_FORMAT_R16G16_UINT:
339 case PIPE_FORMAT_R16G16_SINT:
340 return V_028C70_SWAP_STD;
341
342 /* 64-bit buffers. */
343 case PIPE_FORMAT_R32G32_FLOAT:
344 case PIPE_FORMAT_R32G32_UINT:
345 case PIPE_FORMAT_R32G32_SINT:
346 case PIPE_FORMAT_R16G16B16A16_UNORM:
347 case PIPE_FORMAT_R16G16B16A16_SNORM:
348 case PIPE_FORMAT_R16G16B16A16_UINT:
349 case PIPE_FORMAT_R16G16B16A16_SINT:
350 case PIPE_FORMAT_R16G16B16A16_FLOAT:
351 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
352
353 /* 128-bit buffers. */
354 case PIPE_FORMAT_R32G32B32A32_FLOAT:
355 case PIPE_FORMAT_R32G32B32A32_SNORM:
356 case PIPE_FORMAT_R32G32B32A32_UNORM:
357 case PIPE_FORMAT_R32G32B32A32_SINT:
358 case PIPE_FORMAT_R32G32B32A32_UINT:
359 return V_028C70_SWAP_STD;
360 default:
361 R600_ERR("unsupported colorswap format %d\n", format);
362 return ~0U;
363 }
364 return ~0U;
365 }
366
367 static uint32_t r600_translate_colorformat(enum pipe_format format)
368 {
369 switch (format) {
370 /* 8-bit buffers. */
371 case PIPE_FORMAT_A8_UNORM:
372 case PIPE_FORMAT_A8_SNORM:
373 case PIPE_FORMAT_A8_UINT:
374 case PIPE_FORMAT_A8_SINT:
375 case PIPE_FORMAT_I8_UNORM:
376 case PIPE_FORMAT_I8_SNORM:
377 case PIPE_FORMAT_I8_UINT:
378 case PIPE_FORMAT_I8_SINT:
379 case PIPE_FORMAT_L8_UNORM:
380 case PIPE_FORMAT_L8_SNORM:
381 case PIPE_FORMAT_L8_UINT:
382 case PIPE_FORMAT_L8_SINT:
383 case PIPE_FORMAT_L8_SRGB:
384 case PIPE_FORMAT_R8_UNORM:
385 case PIPE_FORMAT_R8_SNORM:
386 case PIPE_FORMAT_R8_UINT:
387 case PIPE_FORMAT_R8_SINT:
388 return V_028C70_COLOR_8;
389
390 /* 16-bit buffers. */
391 case PIPE_FORMAT_B5G6R5_UNORM:
392 return V_028C70_COLOR_5_6_5;
393
394 case PIPE_FORMAT_B5G5R5A1_UNORM:
395 case PIPE_FORMAT_B5G5R5X1_UNORM:
396 return V_028C70_COLOR_1_5_5_5;
397
398 case PIPE_FORMAT_B4G4R4A4_UNORM:
399 case PIPE_FORMAT_B4G4R4X4_UNORM:
400 return V_028C70_COLOR_4_4_4_4;
401
402 case PIPE_FORMAT_Z16_UNORM:
403 return V_028C70_COLOR_16;
404
405 case PIPE_FORMAT_L8A8_UNORM:
406 case PIPE_FORMAT_L8A8_SNORM:
407 case PIPE_FORMAT_L8A8_UINT:
408 case PIPE_FORMAT_L8A8_SINT:
409 case PIPE_FORMAT_L8A8_SRGB:
410 case PIPE_FORMAT_R8G8_UNORM:
411 case PIPE_FORMAT_R8G8_SNORM:
412 case PIPE_FORMAT_R8G8_UINT:
413 case PIPE_FORMAT_R8G8_SINT:
414 return V_028C70_COLOR_8_8;
415
416 case PIPE_FORMAT_R16_UNORM:
417 case PIPE_FORMAT_R16_SNORM:
418 case PIPE_FORMAT_R16_UINT:
419 case PIPE_FORMAT_R16_SINT:
420 case PIPE_FORMAT_A16_UNORM:
421 case PIPE_FORMAT_A16_SNORM:
422 case PIPE_FORMAT_A16_UINT:
423 case PIPE_FORMAT_A16_SINT:
424 case PIPE_FORMAT_L16_UNORM:
425 case PIPE_FORMAT_L16_SNORM:
426 case PIPE_FORMAT_L16_UINT:
427 case PIPE_FORMAT_L16_SINT:
428 case PIPE_FORMAT_I16_UNORM:
429 case PIPE_FORMAT_I16_SNORM:
430 case PIPE_FORMAT_I16_UINT:
431 case PIPE_FORMAT_I16_SINT:
432 return V_028C70_COLOR_16;
433
434 case PIPE_FORMAT_R16_FLOAT:
435 case PIPE_FORMAT_A16_FLOAT:
436 case PIPE_FORMAT_L16_FLOAT:
437 case PIPE_FORMAT_I16_FLOAT:
438 return V_028C70_COLOR_16_FLOAT;
439
440 /* 32-bit buffers. */
441 case PIPE_FORMAT_A8B8G8R8_SRGB:
442 case PIPE_FORMAT_A8B8G8R8_UNORM:
443 case PIPE_FORMAT_A8R8G8B8_UNORM:
444 case PIPE_FORMAT_B8G8R8A8_SRGB:
445 case PIPE_FORMAT_B8G8R8A8_UNORM:
446 case PIPE_FORMAT_B8G8R8X8_UNORM:
447 case PIPE_FORMAT_R8G8B8A8_SNORM:
448 case PIPE_FORMAT_R8G8B8A8_UNORM:
449 case PIPE_FORMAT_R8G8B8X8_UNORM:
450 case PIPE_FORMAT_R8SG8SB8UX8U_NORM:
451 case PIPE_FORMAT_X8B8G8R8_UNORM:
452 case PIPE_FORMAT_X8R8G8B8_UNORM:
453 case PIPE_FORMAT_R8G8B8_UNORM:
454 case PIPE_FORMAT_R8G8B8A8_SINT:
455 case PIPE_FORMAT_R8G8B8A8_UINT:
456 return V_028C70_COLOR_8_8_8_8;
457
458 case PIPE_FORMAT_R10G10B10A2_UNORM:
459 case PIPE_FORMAT_R10G10B10X2_SNORM:
460 case PIPE_FORMAT_B10G10R10A2_UNORM:
461 case PIPE_FORMAT_B10G10R10A2_UINT:
462 case PIPE_FORMAT_R10SG10SB10SA2U_NORM:
463 return V_028C70_COLOR_2_10_10_10;
464
465 case PIPE_FORMAT_Z24X8_UNORM:
466 case PIPE_FORMAT_Z24_UNORM_S8_UINT:
467 return V_028C70_COLOR_8_24;
468
469 case PIPE_FORMAT_X8Z24_UNORM:
470 case PIPE_FORMAT_S8_UINT_Z24_UNORM:
471 return V_028C70_COLOR_24_8;
472
473 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
474 return V_028C70_COLOR_X24_8_32_FLOAT;
475
476 case PIPE_FORMAT_R32_UINT:
477 case PIPE_FORMAT_R32_SINT:
478 case PIPE_FORMAT_A32_UINT:
479 case PIPE_FORMAT_A32_SINT:
480 case PIPE_FORMAT_L32_UINT:
481 case PIPE_FORMAT_L32_SINT:
482 case PIPE_FORMAT_I32_UINT:
483 case PIPE_FORMAT_I32_SINT:
484 return V_028C70_COLOR_32;
485
486 case PIPE_FORMAT_R32_FLOAT:
487 case PIPE_FORMAT_A32_FLOAT:
488 case PIPE_FORMAT_L32_FLOAT:
489 case PIPE_FORMAT_I32_FLOAT:
490 case PIPE_FORMAT_Z32_FLOAT:
491 return V_028C70_COLOR_32_FLOAT;
492
493 case PIPE_FORMAT_R16G16_FLOAT:
494 case PIPE_FORMAT_L16A16_FLOAT:
495 return V_028C70_COLOR_16_16_FLOAT;
496
497 case PIPE_FORMAT_R16G16_UNORM:
498 case PIPE_FORMAT_R16G16_SNORM:
499 case PIPE_FORMAT_R16G16_UINT:
500 case PIPE_FORMAT_R16G16_SINT:
501 case PIPE_FORMAT_L16A16_UNORM:
502 case PIPE_FORMAT_L16A16_SNORM:
503 case PIPE_FORMAT_L16A16_UINT:
504 case PIPE_FORMAT_L16A16_SINT:
505 return V_028C70_COLOR_16_16;
506
507 case PIPE_FORMAT_R11G11B10_FLOAT:
508 return V_028C70_COLOR_10_11_11_FLOAT;
509
510 /* 64-bit buffers. */
511 case PIPE_FORMAT_R16G16B16A16_UINT:
512 case PIPE_FORMAT_R16G16B16A16_SINT:
513 case PIPE_FORMAT_R16G16B16A16_UNORM:
514 case PIPE_FORMAT_R16G16B16A16_SNORM:
515 return V_028C70_COLOR_16_16_16_16;
516
517 case PIPE_FORMAT_R16G16B16_FLOAT:
518 case PIPE_FORMAT_R16G16B16A16_FLOAT:
519 return V_028C70_COLOR_16_16_16_16_FLOAT;
520
521 case PIPE_FORMAT_R32G32_FLOAT:
522 case PIPE_FORMAT_L32A32_FLOAT:
523 return V_028C70_COLOR_32_32_FLOAT;
524
525 case PIPE_FORMAT_R32G32_SINT:
526 case PIPE_FORMAT_R32G32_UINT:
527 case PIPE_FORMAT_L32A32_UINT:
528 case PIPE_FORMAT_L32A32_SINT:
529 return V_028C70_COLOR_32_32;
530
531 /* 96-bit buffers. */
532 case PIPE_FORMAT_R32G32B32_FLOAT:
533 return V_028C70_COLOR_32_32_32_FLOAT;
534
535 /* 128-bit buffers. */
536 case PIPE_FORMAT_R32G32B32A32_SNORM:
537 case PIPE_FORMAT_R32G32B32A32_UNORM:
538 case PIPE_FORMAT_R32G32B32A32_SINT:
539 case PIPE_FORMAT_R32G32B32A32_UINT:
540 return V_028C70_COLOR_32_32_32_32;
541 case PIPE_FORMAT_R32G32B32A32_FLOAT:
542 return V_028C70_COLOR_32_32_32_32_FLOAT;
543
544 /* YUV buffers. */
545 case PIPE_FORMAT_UYVY:
546 case PIPE_FORMAT_YUYV:
547 default:
548 return ~0U; /* Unsupported. */
549 }
550 }
551
552 static uint32_t r600_colorformat_endian_swap(uint32_t colorformat)
553 {
554 if (R600_BIG_ENDIAN) {
555 switch(colorformat) {
556
557 /* 8-bit buffers. */
558 case V_028C70_COLOR_8:
559 return ENDIAN_NONE;
560
561 /* 16-bit buffers. */
562 case V_028C70_COLOR_5_6_5:
563 case V_028C70_COLOR_1_5_5_5:
564 case V_028C70_COLOR_4_4_4_4:
565 case V_028C70_COLOR_16:
566 case V_028C70_COLOR_8_8:
567 return ENDIAN_8IN16;
568
569 /* 32-bit buffers. */
570 case V_028C70_COLOR_8_8_8_8:
571 case V_028C70_COLOR_2_10_10_10:
572 case V_028C70_COLOR_8_24:
573 case V_028C70_COLOR_24_8:
574 case V_028C70_COLOR_32_FLOAT:
575 case V_028C70_COLOR_16_16_FLOAT:
576 case V_028C70_COLOR_16_16:
577 return ENDIAN_8IN32;
578
579 /* 64-bit buffers. */
580 case V_028C70_COLOR_16_16_16_16:
581 case V_028C70_COLOR_16_16_16_16_FLOAT:
582 return ENDIAN_8IN16;
583
584 case V_028C70_COLOR_32_32_FLOAT:
585 case V_028C70_COLOR_32_32:
586 case V_028C70_COLOR_X24_8_32_FLOAT:
587 return ENDIAN_8IN32;
588
589 /* 96-bit buffers. */
590 case V_028C70_COLOR_32_32_32_FLOAT:
591 /* 128-bit buffers. */
592 case V_028C70_COLOR_32_32_32_32_FLOAT:
593 case V_028C70_COLOR_32_32_32_32:
594 return ENDIAN_8IN32;
595 default:
596 return ENDIAN_NONE; /* Unsupported. */
597 }
598 } else {
599 return ENDIAN_NONE;
600 }
601 }
602
603 static bool r600_is_sampler_format_supported(struct pipe_screen *screen, enum pipe_format format)
604 {
605 return r600_translate_texformat(screen, format, NULL, NULL, NULL) != ~0U;
606 }
607
608 static bool r600_is_colorbuffer_format_supported(enum pipe_format format)
609 {
610 return r600_translate_colorformat(format) != ~0U &&
611 r600_translate_colorswap(format) != ~0U;
612 }
613
614 static bool r600_is_zs_format_supported(enum pipe_format format)
615 {
616 return r600_translate_dbformat(format) != ~0U;
617 }
618
619 boolean evergreen_is_format_supported(struct pipe_screen *screen,
620 enum pipe_format format,
621 enum pipe_texture_target target,
622 unsigned sample_count,
623 unsigned usage)
624 {
625 unsigned retval = 0;
626
627 if (target >= PIPE_MAX_TEXTURE_TYPES) {
628 R600_ERR("r600: unsupported texture type %d\n", target);
629 return FALSE;
630 }
631
632 if (!util_format_is_supported(format, usage))
633 return FALSE;
634
635 /* Multisample */
636 if (sample_count > 1)
637 return FALSE;
638
639 if ((usage & PIPE_BIND_SAMPLER_VIEW) &&
640 r600_is_sampler_format_supported(screen, format)) {
641 retval |= PIPE_BIND_SAMPLER_VIEW;
642 }
643
644 if ((usage & (PIPE_BIND_RENDER_TARGET |
645 PIPE_BIND_DISPLAY_TARGET |
646 PIPE_BIND_SCANOUT |
647 PIPE_BIND_SHARED)) &&
648 r600_is_colorbuffer_format_supported(format)) {
649 retval |= usage &
650 (PIPE_BIND_RENDER_TARGET |
651 PIPE_BIND_DISPLAY_TARGET |
652 PIPE_BIND_SCANOUT |
653 PIPE_BIND_SHARED);
654 }
655
656 if ((usage & PIPE_BIND_DEPTH_STENCIL) &&
657 r600_is_zs_format_supported(format)) {
658 retval |= PIPE_BIND_DEPTH_STENCIL;
659 }
660
661 if ((usage & PIPE_BIND_VERTEX_BUFFER) &&
662 r600_is_vertex_format_supported(format)) {
663 retval |= PIPE_BIND_VERTEX_BUFFER;
664 }
665
666 if (usage & PIPE_BIND_TRANSFER_READ)
667 retval |= PIPE_BIND_TRANSFER_READ;
668 if (usage & PIPE_BIND_TRANSFER_WRITE)
669 retval |= PIPE_BIND_TRANSFER_WRITE;
670
671 return retval == usage;
672 }
673
674 static void *evergreen_create_blend_state(struct pipe_context *ctx,
675 const struct pipe_blend_state *state)
676 {
677 struct r600_context *rctx = (struct r600_context *)ctx;
678 struct r600_pipe_blend *blend = CALLOC_STRUCT(r600_pipe_blend);
679 struct r600_pipe_state *rstate;
680 uint32_t color_control, target_mask;
681 /* XXX there is more then 8 framebuffer */
682 unsigned blend_cntl[8];
683
684 if (blend == NULL) {
685 return NULL;
686 }
687
688 rstate = &blend->rstate;
689
690 rstate->id = R600_PIPE_STATE_BLEND;
691
692 target_mask = 0;
693 color_control = S_028808_MODE(1);
694 if (state->logicop_enable) {
695 color_control |= (state->logicop_func << 16) | (state->logicop_func << 20);
696 } else {
697 color_control |= (0xcc << 16);
698 }
699 /* we pretend 8 buffer are used, CB_SHADER_MASK will disable unused one */
700 if (state->independent_blend_enable) {
701 for (int i = 0; i < 8; i++) {
702 target_mask |= (state->rt[i].colormask << (4 * i));
703 }
704 } else {
705 for (int i = 0; i < 8; i++) {
706 target_mask |= (state->rt[0].colormask << (4 * i));
707 }
708 }
709 blend->cb_target_mask = target_mask;
710
711 r600_pipe_state_add_reg(rstate, R_028808_CB_COLOR_CONTROL,
712 color_control);
713 /* only have dual source on MRT0 */
714 blend->dual_src_blend = util_blend_state_is_dual(state, 0);
715 for (int i = 0; i < 8; i++) {
716 /* state->rt entries > 0 only written if independent blending */
717 const int j = state->independent_blend_enable ? i : 0;
718
719 unsigned eqRGB = state->rt[j].rgb_func;
720 unsigned srcRGB = state->rt[j].rgb_src_factor;
721 unsigned dstRGB = state->rt[j].rgb_dst_factor;
722 unsigned eqA = state->rt[j].alpha_func;
723 unsigned srcA = state->rt[j].alpha_src_factor;
724 unsigned dstA = state->rt[j].alpha_dst_factor;
725
726 blend_cntl[i] = 0;
727 if (!state->rt[j].blend_enable)
728 continue;
729
730 blend_cntl[i] |= S_028780_BLEND_CONTROL_ENABLE(1);
731 blend_cntl[i] |= S_028780_COLOR_COMB_FCN(r600_translate_blend_function(eqRGB));
732 blend_cntl[i] |= S_028780_COLOR_SRCBLEND(r600_translate_blend_factor(srcRGB));
733 blend_cntl[i] |= S_028780_COLOR_DESTBLEND(r600_translate_blend_factor(dstRGB));
734
735 if (srcA != srcRGB || dstA != dstRGB || eqA != eqRGB) {
736 blend_cntl[i] |= S_028780_SEPARATE_ALPHA_BLEND(1);
737 blend_cntl[i] |= S_028780_ALPHA_COMB_FCN(r600_translate_blend_function(eqA));
738 blend_cntl[i] |= S_028780_ALPHA_SRCBLEND(r600_translate_blend_factor(srcA));
739 blend_cntl[i] |= S_028780_ALPHA_DESTBLEND(r600_translate_blend_factor(dstA));
740 }
741 }
742 for (int i = 0; i < 8; i++) {
743 r600_pipe_state_add_reg(rstate, R_028780_CB_BLEND0_CONTROL + i * 4, blend_cntl[i]);
744 }
745
746 return rstate;
747 }
748
749 static void *evergreen_create_dsa_state(struct pipe_context *ctx,
750 const struct pipe_depth_stencil_alpha_state *state)
751 {
752 struct r600_context *rctx = (struct r600_context *)ctx;
753 struct r600_pipe_dsa *dsa = CALLOC_STRUCT(r600_pipe_dsa);
754 unsigned db_depth_control, alpha_test_control, alpha_ref;
755 unsigned db_render_control;
756 struct r600_pipe_state *rstate;
757
758 if (dsa == NULL) {
759 return NULL;
760 }
761
762 dsa->valuemask[0] = state->stencil[0].valuemask;
763 dsa->valuemask[1] = state->stencil[1].valuemask;
764 dsa->writemask[0] = state->stencil[0].writemask;
765 dsa->writemask[1] = state->stencil[1].writemask;
766
767 rstate = &dsa->rstate;
768
769 rstate->id = R600_PIPE_STATE_DSA;
770 db_depth_control = S_028800_Z_ENABLE(state->depth.enabled) |
771 S_028800_Z_WRITE_ENABLE(state->depth.writemask) |
772 S_028800_ZFUNC(state->depth.func);
773
774 /* stencil */
775 if (state->stencil[0].enabled) {
776 db_depth_control |= S_028800_STENCIL_ENABLE(1);
777 db_depth_control |= S_028800_STENCILFUNC(state->stencil[0].func); /* translates straight */
778 db_depth_control |= S_028800_STENCILFAIL(r600_translate_stencil_op(state->stencil[0].fail_op));
779 db_depth_control |= S_028800_STENCILZPASS(r600_translate_stencil_op(state->stencil[0].zpass_op));
780 db_depth_control |= S_028800_STENCILZFAIL(r600_translate_stencil_op(state->stencil[0].zfail_op));
781
782 if (state->stencil[1].enabled) {
783 db_depth_control |= S_028800_BACKFACE_ENABLE(1);
784 db_depth_control |= S_028800_STENCILFUNC_BF(state->stencil[1].func); /* translates straight */
785 db_depth_control |= S_028800_STENCILFAIL_BF(r600_translate_stencil_op(state->stencil[1].fail_op));
786 db_depth_control |= S_028800_STENCILZPASS_BF(r600_translate_stencil_op(state->stencil[1].zpass_op));
787 db_depth_control |= S_028800_STENCILZFAIL_BF(r600_translate_stencil_op(state->stencil[1].zfail_op));
788 }
789 }
790
791 /* alpha */
792 alpha_test_control = 0;
793 alpha_ref = 0;
794 if (state->alpha.enabled) {
795 alpha_test_control = S_028410_ALPHA_FUNC(state->alpha.func);
796 alpha_test_control |= S_028410_ALPHA_TEST_ENABLE(1);
797 alpha_ref = fui(state->alpha.ref_value);
798 }
799 dsa->alpha_ref = alpha_ref;
800
801 /* misc */
802 db_render_control = 0;
803 r600_pipe_state_add_reg(rstate, R_028410_SX_ALPHA_TEST_CONTROL, alpha_test_control);
804 r600_pipe_state_add_reg(rstate, R_028800_DB_DEPTH_CONTROL, db_depth_control);
805 r600_pipe_state_add_reg(rstate, R_028000_DB_RENDER_CONTROL, db_render_control);
806 return rstate;
807 }
808
809 static void *evergreen_create_rs_state(struct pipe_context *ctx,
810 const struct pipe_rasterizer_state *state)
811 {
812 struct r600_context *rctx = (struct r600_context *)ctx;
813 struct r600_pipe_rasterizer *rs = CALLOC_STRUCT(r600_pipe_rasterizer);
814 struct r600_pipe_state *rstate;
815 unsigned tmp;
816 unsigned prov_vtx = 1, polygon_dual_mode;
817 float psize_min, psize_max;
818
819 if (rs == NULL) {
820 return NULL;
821 }
822
823 polygon_dual_mode = (state->fill_front != PIPE_POLYGON_MODE_FILL ||
824 state->fill_back != PIPE_POLYGON_MODE_FILL);
825
826 if (state->flatshade_first)
827 prov_vtx = 0;
828
829 rstate = &rs->rstate;
830 rs->flatshade = state->flatshade;
831 rs->sprite_coord_enable = state->sprite_coord_enable;
832 rs->two_side = state->light_twoside;
833 rs->clip_plane_enable = state->clip_plane_enable;
834 rs->pa_sc_line_stipple = state->line_stipple_enable ?
835 S_028A0C_LINE_PATTERN(state->line_stipple_pattern) |
836 S_028A0C_REPEAT_COUNT(state->line_stipple_factor) : 0;
837 rs->pa_cl_clip_cntl =
838 S_028810_PS_UCP_MODE(3) |
839 S_028810_ZCLIP_NEAR_DISABLE(!state->depth_clip) |
840 S_028810_ZCLIP_FAR_DISABLE(!state->depth_clip) |
841 S_028810_DX_LINEAR_ATTR_CLIP_ENA(1);
842
843 /* offset */
844 rs->offset_units = state->offset_units;
845 rs->offset_scale = state->offset_scale * 12.0f;
846
847 rstate->id = R600_PIPE_STATE_RASTERIZER;
848 tmp = S_0286D4_FLAT_SHADE_ENA(1);
849 if (state->sprite_coord_enable) {
850 tmp |= S_0286D4_PNT_SPRITE_ENA(1) |
851 S_0286D4_PNT_SPRITE_OVRD_X(2) |
852 S_0286D4_PNT_SPRITE_OVRD_Y(3) |
853 S_0286D4_PNT_SPRITE_OVRD_Z(0) |
854 S_0286D4_PNT_SPRITE_OVRD_W(1);
855 if (state->sprite_coord_mode != PIPE_SPRITE_COORD_UPPER_LEFT) {
856 tmp |= S_0286D4_PNT_SPRITE_TOP_1(1);
857 }
858 }
859 r600_pipe_state_add_reg(rstate, R_0286D4_SPI_INTERP_CONTROL_0, tmp);
860
861 /* point size 12.4 fixed point */
862 tmp = (unsigned)(state->point_size * 8.0);
863 r600_pipe_state_add_reg(rstate, R_028A00_PA_SU_POINT_SIZE, S_028A00_HEIGHT(tmp) | S_028A00_WIDTH(tmp));
864
865 if (state->point_size_per_vertex) {
866 psize_min = util_get_min_point_size(state);
867 psize_max = 8192;
868 } else {
869 /* Force the point size to be as if the vertex output was disabled. */
870 psize_min = state->point_size;
871 psize_max = state->point_size;
872 }
873 /* Divide by two, because 0.5 = 1 pixel. */
874 r600_pipe_state_add_reg(rstate, R_028A04_PA_SU_POINT_MINMAX,
875 S_028A04_MIN_SIZE(r600_pack_float_12p4(psize_min/2)) |
876 S_028A04_MAX_SIZE(r600_pack_float_12p4(psize_max/2)));
877
878 tmp = (unsigned)state->line_width * 8;
879 r600_pipe_state_add_reg(rstate, R_028A08_PA_SU_LINE_CNTL, S_028A08_WIDTH(tmp));
880 r600_pipe_state_add_reg(rstate, R_028A48_PA_SC_MODE_CNTL_0,
881 S_028A48_VPORT_SCISSOR_ENABLE(state->scissor) |
882 S_028A48_LINE_STIPPLE_ENABLE(state->line_stipple_enable));
883
884 if (rctx->chip_class == CAYMAN) {
885 r600_pipe_state_add_reg(rstate, CM_R_028BE4_PA_SU_VTX_CNTL,
886 S_028C08_PIX_CENTER_HALF(state->gl_rasterization_rules));
887 } else {
888 r600_pipe_state_add_reg(rstate, R_028C08_PA_SU_VTX_CNTL,
889 S_028C08_PIX_CENTER_HALF(state->gl_rasterization_rules));
890 }
891 r600_pipe_state_add_reg(rstate, R_028B7C_PA_SU_POLY_OFFSET_CLAMP, fui(state->offset_clamp));
892 r600_pipe_state_add_reg(rstate, R_028814_PA_SU_SC_MODE_CNTL,
893 S_028814_PROVOKING_VTX_LAST(prov_vtx) |
894 S_028814_CULL_FRONT(state->rasterizer_discard || (state->cull_face & PIPE_FACE_FRONT) ? 1 : 0) |
895 S_028814_CULL_BACK(state->rasterizer_discard || (state->cull_face & PIPE_FACE_BACK) ? 1 : 0) |
896 S_028814_FACE(!state->front_ccw) |
897 S_028814_POLY_OFFSET_FRONT_ENABLE(state->offset_tri) |
898 S_028814_POLY_OFFSET_BACK_ENABLE(state->offset_tri) |
899 S_028814_POLY_OFFSET_PARA_ENABLE(state->offset_tri) |
900 S_028814_POLY_MODE(polygon_dual_mode) |
901 S_028814_POLYMODE_FRONT_PTYPE(r600_translate_fill(state->fill_front)) |
902 S_028814_POLYMODE_BACK_PTYPE(r600_translate_fill(state->fill_back)));
903 r600_pipe_state_add_reg(rstate, R_028350_SX_MISC, S_028350_MULTIPASS(state->rasterizer_discard));
904 return rstate;
905 }
906
907 static void *evergreen_create_sampler_state(struct pipe_context *ctx,
908 const struct pipe_sampler_state *state)
909 {
910 struct r600_pipe_state *rstate = CALLOC_STRUCT(r600_pipe_state);
911 union util_color uc;
912 unsigned aniso_flag_offset = state->max_anisotropy > 1 ? 2 : 0;
913
914 if (rstate == NULL) {
915 return NULL;
916 }
917
918 rstate->id = R600_PIPE_STATE_SAMPLER;
919 util_pack_color(state->border_color.f, PIPE_FORMAT_B8G8R8A8_UNORM, &uc);
920 r600_pipe_state_add_reg_noblock(rstate, R_03C000_SQ_TEX_SAMPLER_WORD0_0,
921 S_03C000_CLAMP_X(r600_tex_wrap(state->wrap_s)) |
922 S_03C000_CLAMP_Y(r600_tex_wrap(state->wrap_t)) |
923 S_03C000_CLAMP_Z(r600_tex_wrap(state->wrap_r)) |
924 S_03C000_XY_MAG_FILTER(r600_tex_filter(state->mag_img_filter) | aniso_flag_offset) |
925 S_03C000_XY_MIN_FILTER(r600_tex_filter(state->min_img_filter) | aniso_flag_offset) |
926 S_03C000_MIP_FILTER(r600_tex_mipfilter(state->min_mip_filter)) |
927 S_03C000_MAX_ANISO(r600_tex_aniso_filter(state->max_anisotropy)) |
928 S_03C000_DEPTH_COMPARE_FUNCTION(r600_tex_compare(state->compare_func)) |
929 S_03C000_BORDER_COLOR_TYPE(uc.ui ? V_03C000_SQ_TEX_BORDER_COLOR_REGISTER : 0), NULL, 0);
930 r600_pipe_state_add_reg_noblock(rstate, R_03C004_SQ_TEX_SAMPLER_WORD1_0,
931 S_03C004_MIN_LOD(S_FIXED(CLAMP(state->min_lod, 0, 15), 8)) |
932 S_03C004_MAX_LOD(S_FIXED(CLAMP(state->max_lod, 0, 15), 8)),
933 NULL, 0);
934 r600_pipe_state_add_reg_noblock(rstate, R_03C008_SQ_TEX_SAMPLER_WORD2_0,
935 S_03C008_LOD_BIAS(S_FIXED(CLAMP(state->lod_bias, -16, 16), 8)) |
936 (state->seamless_cube_map ? 0 : S_03C008_DISABLE_CUBE_WRAP(1)) |
937 S_03C008_TYPE(1),
938 NULL, 0);
939
940 if (uc.ui) {
941 r600_pipe_state_add_reg_noblock(rstate, R_00A404_TD_PS_SAMPLER0_BORDER_RED, fui(state->border_color.f[0]), NULL, 0);
942 r600_pipe_state_add_reg_noblock(rstate, R_00A408_TD_PS_SAMPLER0_BORDER_GREEN, fui(state->border_color.f[1]), NULL, 0);
943 r600_pipe_state_add_reg_noblock(rstate, R_00A40C_TD_PS_SAMPLER0_BORDER_BLUE, fui(state->border_color.f[2]), NULL, 0);
944 r600_pipe_state_add_reg_noblock(rstate, R_00A410_TD_PS_SAMPLER0_BORDER_ALPHA, fui(state->border_color.f[3]), NULL, 0);
945 }
946 return rstate;
947 }
948
949 static struct pipe_sampler_view *evergreen_create_sampler_view(struct pipe_context *ctx,
950 struct pipe_resource *texture,
951 const struct pipe_sampler_view *state)
952 {
953 struct r600_screen *rscreen = (struct r600_screen*)ctx->screen;
954 struct r600_pipe_sampler_view *view = CALLOC_STRUCT(r600_pipe_sampler_view);
955 struct r600_pipe_resource_state *rstate;
956 struct r600_resource_texture *tmp = (struct r600_resource_texture*)texture;
957 unsigned format, endian;
958 uint32_t word4 = 0, yuv_format = 0, pitch = 0;
959 unsigned char swizzle[4], array_mode = 0, tile_type = 0;
960 unsigned height, depth, width;
961 unsigned macro_aspect, tile_split, bankh, bankw, nbanks;
962
963 if (view == NULL)
964 return NULL;
965 rstate = &view->state;
966
967 /* initialize base object */
968 view->base = *state;
969 view->base.texture = NULL;
970 pipe_reference(NULL, &texture->reference);
971 view->base.texture = texture;
972 view->base.reference.count = 1;
973 view->base.context = ctx;
974
975 swizzle[0] = state->swizzle_r;
976 swizzle[1] = state->swizzle_g;
977 swizzle[2] = state->swizzle_b;
978 swizzle[3] = state->swizzle_a;
979
980 format = r600_translate_texformat(ctx->screen, state->format,
981 swizzle,
982 &word4, &yuv_format);
983 if (format == ~0) {
984 format = 0;
985 }
986
987 if (tmp->is_depth && !tmp->is_flushing_texture) {
988 r600_texture_depth_flush(ctx, texture, TRUE);
989 tmp = tmp->flushed_depth_texture;
990 }
991
992 endian = r600_colorformat_endian_swap(format);
993
994 if (!rscreen->use_surface_alloc) {
995 height = texture->height0;
996 depth = texture->depth0;
997 width = texture->width0;
998 pitch = align(tmp->pitch_in_blocks[0] *
999 util_format_get_blockwidth(state->format), 8);
1000 array_mode = tmp->array_mode[0];
1001 tile_type = tmp->tile_type;
1002 tile_split = 0;
1003 macro_aspect = 0;
1004 bankw = 0;
1005 bankh = 0;
1006 } else {
1007 width = tmp->surface.level[0].npix_x;
1008 height = tmp->surface.level[0].npix_y;
1009 depth = tmp->surface.level[0].npix_z;
1010 pitch = tmp->surface.level[0].nblk_x * util_format_get_blockwidth(state->format);
1011 tile_type = tmp->tile_type;
1012
1013 switch (tmp->surface.level[0].mode) {
1014 case RADEON_SURF_MODE_LINEAR_ALIGNED:
1015 array_mode = V_028C70_ARRAY_LINEAR_ALIGNED;
1016 break;
1017 case RADEON_SURF_MODE_2D:
1018 array_mode = V_028C70_ARRAY_2D_TILED_THIN1;
1019 break;
1020 case RADEON_SURF_MODE_1D:
1021 array_mode = V_028C70_ARRAY_1D_TILED_THIN1;
1022 break;
1023 case RADEON_SURF_MODE_LINEAR:
1024 default:
1025 array_mode = V_028C70_ARRAY_LINEAR_GENERAL;
1026 break;
1027 }
1028 tile_split = tmp->surface.tile_split;
1029 macro_aspect = tmp->surface.mtilea;
1030 bankw = tmp->surface.bankw;
1031 bankh = tmp->surface.bankh;
1032 tile_split = eg_tile_split(tile_split);
1033 macro_aspect = eg_macro_tile_aspect(macro_aspect);
1034 bankw = eg_bank_wh(bankw);
1035 bankh = eg_bank_wh(bankh);
1036 }
1037 /* 128 bit formats require tile type = 1 */
1038 if (rscreen->chip_class == CAYMAN) {
1039 if (util_format_get_blocksize(state->format) >= 16)
1040 tile_type = 1;
1041 }
1042 nbanks = eg_num_banks(rscreen->tiling_info.num_banks);
1043
1044 if (texture->target == PIPE_TEXTURE_1D_ARRAY) {
1045 height = 1;
1046 depth = texture->array_size;
1047 } else if (texture->target == PIPE_TEXTURE_2D_ARRAY) {
1048 depth = texture->array_size;
1049 }
1050
1051 rstate->bo[0] = &tmp->resource;
1052 rstate->bo[1] = &tmp->resource;
1053 rstate->bo_usage[0] = RADEON_USAGE_READ;
1054 rstate->bo_usage[1] = RADEON_USAGE_READ;
1055
1056 rstate->val[0] = (S_030000_DIM(r600_tex_dim(texture->target)) |
1057 S_030000_PITCH((pitch / 8) - 1) |
1058 S_030000_TEX_WIDTH(width - 1));
1059 if (rscreen->chip_class == CAYMAN)
1060 rstate->val[0] |= CM_S_030000_NON_DISP_TILING_ORDER(tile_type);
1061 else
1062 rstate->val[0] |= S_030000_NON_DISP_TILING_ORDER(tile_type);
1063 rstate->val[1] = (S_030004_TEX_HEIGHT(height - 1) |
1064 S_030004_TEX_DEPTH(depth - 1) |
1065 S_030004_ARRAY_MODE(array_mode));
1066 rstate->val[2] = (tmp->offset[0] + r600_resource_va(ctx->screen, texture)) >> 8;
1067 if (state->u.tex.last_level) {
1068 rstate->val[3] = (tmp->offset[1] + r600_resource_va(ctx->screen, texture)) >> 8;
1069 } else {
1070 rstate->val[3] = (tmp->offset[0] + r600_resource_va(ctx->screen, texture)) >> 8;
1071 }
1072 rstate->val[4] = (word4 |
1073 S_030010_SRF_MODE_ALL(V_030010_SRF_MODE_ZERO_CLAMP_MINUS_ONE) |
1074 S_030010_ENDIAN_SWAP(endian) |
1075 S_030010_BASE_LEVEL(state->u.tex.first_level));
1076 rstate->val[5] = (S_030014_LAST_LEVEL(state->u.tex.last_level) |
1077 S_030014_BASE_ARRAY(state->u.tex.first_layer) |
1078 S_030014_LAST_ARRAY(state->u.tex.last_layer));
1079 /* aniso max 16 samples */
1080 rstate->val[6] = (S_030018_MAX_ANISO(4)) |
1081 (S_030018_TILE_SPLIT(tile_split));
1082 rstate->val[7] = S_03001C_DATA_FORMAT(format) |
1083 S_03001C_TYPE(V_03001C_SQ_TEX_VTX_VALID_TEXTURE) |
1084 S_03001C_BANK_WIDTH(bankw) |
1085 S_03001C_BANK_HEIGHT(bankh) |
1086 S_03001C_MACRO_TILE_ASPECT(macro_aspect) |
1087 S_03001C_NUM_BANKS(nbanks);
1088
1089 return &view->base;
1090 }
1091
1092 static void evergreen_set_vs_sampler_view(struct pipe_context *ctx, unsigned count,
1093 struct pipe_sampler_view **views)
1094 {
1095 struct r600_context *rctx = (struct r600_context *)ctx;
1096 struct r600_pipe_sampler_view **resource = (struct r600_pipe_sampler_view **)views;
1097
1098 for (int i = 0; i < count; i++) {
1099 if (resource[i]) {
1100 r600_context_pipe_state_set_vs_resource(rctx, &resource[i]->state,
1101 i + R600_MAX_CONST_BUFFERS);
1102 }
1103 }
1104 }
1105
1106 static void evergreen_set_ps_sampler_view(struct pipe_context *ctx, unsigned count,
1107 struct pipe_sampler_view **views)
1108 {
1109 struct r600_context *rctx = (struct r600_context *)ctx;
1110 struct r600_pipe_sampler_view **resource = (struct r600_pipe_sampler_view **)views;
1111 int i;
1112 int has_depth = 0;
1113
1114 for (i = 0; i < count; i++) {
1115 if (&rctx->ps_samplers.views[i]->base != views[i]) {
1116 if (resource[i]) {
1117 if (((struct r600_resource_texture *)resource[i]->base.texture)->is_depth)
1118 has_depth = 1;
1119 r600_context_pipe_state_set_ps_resource(rctx, &resource[i]->state,
1120 i + R600_MAX_CONST_BUFFERS);
1121 } else
1122 r600_context_pipe_state_set_ps_resource(rctx, NULL,
1123 i + R600_MAX_CONST_BUFFERS);
1124
1125 pipe_sampler_view_reference(
1126 (struct pipe_sampler_view **)&rctx->ps_samplers.views[i],
1127 views[i]);
1128 } else {
1129 if (resource[i]) {
1130 if (((struct r600_resource_texture *)resource[i]->base.texture)->is_depth)
1131 has_depth = 1;
1132 }
1133 }
1134 }
1135 for (i = count; i < NUM_TEX_UNITS; i++) {
1136 if (rctx->ps_samplers.views[i]) {
1137 r600_context_pipe_state_set_ps_resource(rctx, NULL,
1138 i + R600_MAX_CONST_BUFFERS);
1139 pipe_sampler_view_reference((struct pipe_sampler_view **)&rctx->ps_samplers.views[i], NULL);
1140 }
1141 }
1142 rctx->have_depth_texture = has_depth;
1143 rctx->ps_samplers.n_views = count;
1144 }
1145
1146 static void evergreen_bind_ps_sampler(struct pipe_context *ctx, unsigned count, void **states)
1147 {
1148 struct r600_context *rctx = (struct r600_context *)ctx;
1149 struct r600_pipe_state **rstates = (struct r600_pipe_state **)states;
1150
1151 if (count)
1152 r600_inval_texture_cache(rctx);
1153
1154 memcpy(rctx->ps_samplers.samplers, states, sizeof(void*) * count);
1155 rctx->ps_samplers.n_samplers = count;
1156
1157 for (int i = 0; i < count; i++) {
1158 evergreen_context_pipe_state_set_ps_sampler(rctx, rstates[i], i);
1159 }
1160 }
1161
1162 static void evergreen_bind_vs_sampler(struct pipe_context *ctx, unsigned count, void **states)
1163 {
1164 struct r600_context *rctx = (struct r600_context *)ctx;
1165 struct r600_pipe_state **rstates = (struct r600_pipe_state **)states;
1166
1167 if (count)
1168 r600_inval_texture_cache(rctx);
1169
1170 for (int i = 0; i < count; i++) {
1171 evergreen_context_pipe_state_set_vs_sampler(rctx, rstates[i], i);
1172 }
1173 }
1174
1175 static void evergreen_set_clip_state(struct pipe_context *ctx,
1176 const struct pipe_clip_state *state)
1177 {
1178 struct r600_context *rctx = (struct r600_context *)ctx;
1179 struct r600_pipe_state *rstate = CALLOC_STRUCT(r600_pipe_state);
1180 struct pipe_constant_buffer cb;
1181
1182 if (rstate == NULL)
1183 return;
1184
1185 rctx->clip = *state;
1186 rstate->id = R600_PIPE_STATE_CLIP;
1187 for (int i = 0; i < 6; i++) {
1188 r600_pipe_state_add_reg(rstate,
1189 R_0285BC_PA_CL_UCP0_X + i * 16,
1190 fui(state->ucp[i][0]));
1191 r600_pipe_state_add_reg(rstate,
1192 R_0285C0_PA_CL_UCP0_Y + i * 16,
1193 fui(state->ucp[i][1]) );
1194 r600_pipe_state_add_reg(rstate,
1195 R_0285C4_PA_CL_UCP0_Z + i * 16,
1196 fui(state->ucp[i][2]));
1197 r600_pipe_state_add_reg(rstate,
1198 R_0285C8_PA_CL_UCP0_W + i * 16,
1199 fui(state->ucp[i][3]));
1200 }
1201
1202 free(rctx->states[R600_PIPE_STATE_CLIP]);
1203 rctx->states[R600_PIPE_STATE_CLIP] = rstate;
1204 r600_context_pipe_state_set(rctx, rstate);
1205
1206 cb.buffer = pipe_user_buffer_create(ctx->screen,
1207 state->ucp,
1208 4*4*8, /* 8*4 floats */
1209 PIPE_BIND_CONSTANT_BUFFER);
1210 cb.buffer_offset = 0;
1211 cb.buffer_size = 4*4*8;
1212 r600_set_constant_buffer(ctx, PIPE_SHADER_VERTEX, 1, &cb);
1213 pipe_resource_reference(&cb.buffer, NULL);
1214 }
1215
1216 static void evergreen_set_polygon_stipple(struct pipe_context *ctx,
1217 const struct pipe_poly_stipple *state)
1218 {
1219 }
1220
1221 static void evergreen_set_sample_mask(struct pipe_context *pipe, unsigned sample_mask)
1222 {
1223 }
1224
1225 static void evergreen_get_scissor_rect(struct r600_context *rctx,
1226 unsigned tl_x, unsigned tl_y, unsigned br_x, unsigned br_y,
1227 uint32_t *tl, uint32_t *br)
1228 {
1229 /* EG hw workaround */
1230 if (br_x == 0)
1231 tl_x = 1;
1232 if (br_y == 0)
1233 tl_y = 1;
1234
1235 /* cayman hw workaround */
1236 if (rctx->chip_class == CAYMAN) {
1237 if (br_x == 1 && br_y == 1)
1238 br_x = 2;
1239 }
1240
1241 *tl = S_028240_TL_X(tl_x) | S_028240_TL_Y(tl_y);
1242 *br = S_028244_BR_X(br_x) | S_028244_BR_Y(br_y);
1243 }
1244
1245 static void evergreen_set_scissor_state(struct pipe_context *ctx,
1246 const struct pipe_scissor_state *state)
1247 {
1248 struct r600_context *rctx = (struct r600_context *)ctx;
1249 struct r600_pipe_state *rstate = CALLOC_STRUCT(r600_pipe_state);
1250 uint32_t tl, br;
1251
1252 if (rstate == NULL)
1253 return;
1254
1255 evergreen_get_scissor_rect(rctx, state->minx, state->miny, state->maxx, state->maxy, &tl, &br);
1256
1257 rstate->id = R600_PIPE_STATE_SCISSOR;
1258 r600_pipe_state_add_reg(rstate, R_028250_PA_SC_VPORT_SCISSOR_0_TL, tl);
1259 r600_pipe_state_add_reg(rstate, R_028254_PA_SC_VPORT_SCISSOR_0_BR, br);
1260
1261 free(rctx->states[R600_PIPE_STATE_SCISSOR]);
1262 rctx->states[R600_PIPE_STATE_SCISSOR] = rstate;
1263 r600_context_pipe_state_set(rctx, rstate);
1264 }
1265
1266 static void evergreen_set_viewport_state(struct pipe_context *ctx,
1267 const struct pipe_viewport_state *state)
1268 {
1269 struct r600_context *rctx = (struct r600_context *)ctx;
1270 struct r600_pipe_state *rstate = CALLOC_STRUCT(r600_pipe_state);
1271
1272 if (rstate == NULL)
1273 return;
1274
1275 rctx->viewport = *state;
1276 rstate->id = R600_PIPE_STATE_VIEWPORT;
1277 r600_pipe_state_add_reg(rstate, R_02843C_PA_CL_VPORT_XSCALE_0, fui(state->scale[0]));
1278 r600_pipe_state_add_reg(rstate, R_028444_PA_CL_VPORT_YSCALE_0, fui(state->scale[1]));
1279 r600_pipe_state_add_reg(rstate, R_02844C_PA_CL_VPORT_ZSCALE_0, fui(state->scale[2]));
1280 r600_pipe_state_add_reg(rstate, R_028440_PA_CL_VPORT_XOFFSET_0, fui(state->translate[0]));
1281 r600_pipe_state_add_reg(rstate, R_028448_PA_CL_VPORT_YOFFSET_0, fui(state->translate[1]));
1282 r600_pipe_state_add_reg(rstate, R_028450_PA_CL_VPORT_ZOFFSET_0, fui(state->translate[2]));
1283
1284 free(rctx->states[R600_PIPE_STATE_VIEWPORT]);
1285 rctx->states[R600_PIPE_STATE_VIEWPORT] = rstate;
1286 r600_context_pipe_state_set(rctx, rstate);
1287 }
1288
1289 static void evergreen_cb(struct r600_context *rctx, struct r600_pipe_state *rstate,
1290 const struct pipe_framebuffer_state *state, int cb)
1291 {
1292 struct r600_screen *rscreen = rctx->screen;
1293 struct r600_resource_texture *rtex;
1294 struct r600_surface *surf;
1295 unsigned level = state->cbufs[cb]->u.tex.level;
1296 unsigned pitch, slice;
1297 unsigned color_info, color_attrib;
1298 unsigned format, swap, ntype, endian;
1299 uint64_t offset;
1300 unsigned tile_type, macro_aspect, tile_split, bankh, bankw, nbanks;
1301 const struct util_format_description *desc;
1302 int i;
1303 unsigned blend_clamp = 0, blend_bypass = 0;
1304
1305 surf = (struct r600_surface *)state->cbufs[cb];
1306 rtex = (struct r600_resource_texture*)state->cbufs[cb]->texture;
1307
1308 if (rtex->is_depth)
1309 rctx->have_depth_fb = TRUE;
1310
1311 if (rtex->is_depth && !rtex->is_flushing_texture) {
1312 r600_texture_depth_flush(&rctx->context, state->cbufs[cb]->texture, TRUE);
1313 rtex = rtex->flushed_depth_texture;
1314 }
1315
1316 /* XXX quite sure for dx10+ hw don't need any offset hacks */
1317 if (!rscreen->use_surface_alloc) {
1318 offset = r600_texture_get_offset(rtex,
1319 level, state->cbufs[cb]->u.tex.first_layer);
1320 pitch = rtex->pitch_in_blocks[level] / 8 - 1;
1321 slice = rtex->pitch_in_blocks[level] * surf->aligned_height / 64;
1322 if (slice) {
1323 slice = slice - 1;
1324 }
1325 color_info = S_028C70_ARRAY_MODE(rtex->array_mode[level]);
1326 tile_split = 0;
1327 macro_aspect = 0;
1328 bankw = 0;
1329 bankh = 0;
1330 if (rtex->array_mode[level] > V_028C70_ARRAY_LINEAR_ALIGNED) {
1331 tile_type = rtex->tile_type;
1332 } else {
1333 /* workaround for linear buffers */
1334 tile_type = 1;
1335 }
1336 } else {
1337 offset = rtex->surface.level[level].offset;
1338 if (rtex->surface.level[level].mode < RADEON_SURF_MODE_1D) {
1339 offset += rtex->surface.level[level].slice_size *
1340 state->cbufs[cb]->u.tex.first_layer;
1341 }
1342 pitch = (rtex->surface.level[level].nblk_x) / 8 - 1;
1343 slice = (rtex->surface.level[level].nblk_x * rtex->surface.level[level].nblk_y) / 64;
1344 if (slice) {
1345 slice = slice - 1;
1346 }
1347 color_info = 0;
1348 switch (rtex->surface.level[level].mode) {
1349 case RADEON_SURF_MODE_LINEAR_ALIGNED:
1350 color_info = S_028C70_ARRAY_MODE(V_028C70_ARRAY_LINEAR_ALIGNED);
1351 tile_type = 1;
1352 break;
1353 case RADEON_SURF_MODE_1D:
1354 color_info = S_028C70_ARRAY_MODE(V_028C70_ARRAY_1D_TILED_THIN1);
1355 tile_type = rtex->tile_type;
1356 break;
1357 case RADEON_SURF_MODE_2D:
1358 color_info = S_028C70_ARRAY_MODE(V_028C70_ARRAY_2D_TILED_THIN1);
1359 tile_type = rtex->tile_type;
1360 break;
1361 case RADEON_SURF_MODE_LINEAR:
1362 default:
1363 color_info = S_028C70_ARRAY_MODE(V_028C70_ARRAY_LINEAR_GENERAL);
1364 tile_type = 1;
1365 break;
1366 }
1367 tile_split = rtex->surface.tile_split;
1368 macro_aspect = rtex->surface.mtilea;
1369 bankw = rtex->surface.bankw;
1370 bankh = rtex->surface.bankh;
1371 tile_split = eg_tile_split(tile_split);
1372 macro_aspect = eg_macro_tile_aspect(macro_aspect);
1373 bankw = eg_bank_wh(bankw);
1374 bankh = eg_bank_wh(bankh);
1375 }
1376 /* 128 bit formats require tile type = 1 */
1377 if (rscreen->chip_class == CAYMAN) {
1378 if (util_format_get_blocksize(surf->base.format) >= 16)
1379 tile_type = 1;
1380 }
1381 nbanks = eg_num_banks(rscreen->tiling_info.num_banks);
1382 desc = util_format_description(surf->base.format);
1383 for (i = 0; i < 4; i++) {
1384 if (desc->channel[i].type != UTIL_FORMAT_TYPE_VOID) {
1385 break;
1386 }
1387 }
1388
1389 color_attrib = S_028C74_TILE_SPLIT(tile_split)|
1390 S_028C74_NUM_BANKS(nbanks) |
1391 S_028C74_BANK_WIDTH(bankw) |
1392 S_028C74_BANK_HEIGHT(bankh) |
1393 S_028C74_MACRO_TILE_ASPECT(macro_aspect) |
1394 S_028C74_NON_DISP_TILING_ORDER(tile_type);
1395
1396 ntype = V_028C70_NUMBER_UNORM;
1397 if (desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB)
1398 ntype = V_028C70_NUMBER_SRGB;
1399 else if (desc->channel[i].type == UTIL_FORMAT_TYPE_SIGNED) {
1400 if (desc->channel[i].normalized)
1401 ntype = V_028C70_NUMBER_SNORM;
1402 else if (desc->channel[i].pure_integer)
1403 ntype = V_028C70_NUMBER_SINT;
1404 } else if (desc->channel[i].type == UTIL_FORMAT_TYPE_UNSIGNED) {
1405 if (desc->channel[i].normalized)
1406 ntype = V_028C70_NUMBER_UNORM;
1407 else if (desc->channel[i].pure_integer)
1408 ntype = V_028C70_NUMBER_UINT;
1409 }
1410
1411 format = r600_translate_colorformat(surf->base.format);
1412 swap = r600_translate_colorswap(surf->base.format);
1413 if (rtex->resource.b.b.usage == PIPE_USAGE_STAGING) {
1414 endian = ENDIAN_NONE;
1415 } else {
1416 endian = r600_colorformat_endian_swap(format);
1417 }
1418
1419 /* blend clamp should be set for all NORM/SRGB types */
1420 if (ntype == V_028C70_NUMBER_UNORM || ntype == V_028C70_NUMBER_SNORM ||
1421 ntype == V_028C70_NUMBER_SRGB)
1422 blend_clamp = 1;
1423
1424 /* set blend bypass according to docs if SINT/UINT or
1425 8/24 COLOR variants */
1426 if (ntype == V_028C70_NUMBER_UINT || ntype == V_028C70_NUMBER_SINT ||
1427 format == V_028C70_COLOR_8_24 || format == V_028C70_COLOR_24_8 ||
1428 format == V_028C70_COLOR_X24_8_32_FLOAT) {
1429 blend_clamp = 0;
1430 blend_bypass = 1;
1431 }
1432
1433 color_info |= S_028C70_FORMAT(format) |
1434 S_028C70_COMP_SWAP(swap) |
1435 S_028C70_BLEND_CLAMP(blend_clamp) |
1436 S_028C70_BLEND_BYPASS(blend_bypass) |
1437 S_028C70_NUMBER_TYPE(ntype) |
1438 S_028C70_ENDIAN(endian);
1439
1440 /* EXPORT_NORM is an optimzation that can be enabled for better
1441 * performance in certain cases.
1442 * EXPORT_NORM can be enabled if:
1443 * - 11-bit or smaller UNORM/SNORM/SRGB
1444 * - 16-bit or smaller FLOAT
1445 */
1446 /* XXX: This should probably be the same for all CBs if we want
1447 * useful alpha tests. */
1448 if (desc->colorspace != UTIL_FORMAT_COLORSPACE_ZS &&
1449 ((desc->channel[i].size < 12 &&
1450 desc->channel[i].type != UTIL_FORMAT_TYPE_FLOAT &&
1451 ntype != V_028C70_NUMBER_UINT && ntype != V_028C70_NUMBER_SINT) ||
1452 (desc->channel[i].size < 17 &&
1453 desc->channel[i].type == UTIL_FORMAT_TYPE_FLOAT))) {
1454 color_info |= S_028C70_SOURCE_FORMAT(V_028C70_EXPORT_4C_16BPC);
1455 rctx->export_16bpc = true;
1456 } else {
1457 rctx->export_16bpc = false;
1458 }
1459 rctx->alpha_ref_dirty = true;
1460
1461 if (cb == 0)
1462 rctx->color0_format = color_info;
1463
1464 offset += r600_resource_va(rctx->context.screen, state->cbufs[cb]->texture);
1465 offset >>= 8;
1466
1467 /* XXX handle enabling of CB beyond BASE8 which has different offset */
1468 r600_pipe_state_add_reg_bo(rstate,
1469 R_028C60_CB_COLOR0_BASE + cb * 0x3C,
1470 offset, &rtex->resource, RADEON_USAGE_READWRITE);
1471 r600_pipe_state_add_reg(rstate,
1472 R_028C78_CB_COLOR0_DIM + cb * 0x3C,
1473 0x0);
1474 r600_pipe_state_add_reg_bo(rstate,
1475 R_028C70_CB_COLOR0_INFO + cb * 0x3C,
1476 color_info, &rtex->resource, RADEON_USAGE_READWRITE);
1477 r600_pipe_state_add_reg(rstate,
1478 R_028C64_CB_COLOR0_PITCH + cb * 0x3C,
1479 S_028C64_PITCH_TILE_MAX(pitch));
1480 r600_pipe_state_add_reg(rstate,
1481 R_028C68_CB_COLOR0_SLICE + cb * 0x3C,
1482 S_028C68_SLICE_TILE_MAX(slice));
1483 if (!rscreen->use_surface_alloc) {
1484 r600_pipe_state_add_reg(rstate,
1485 R_028C6C_CB_COLOR0_VIEW + cb * 0x3C,
1486 0x00000000);
1487 } else {
1488 if (rtex->surface.level[level].mode < RADEON_SURF_MODE_1D) {
1489 r600_pipe_state_add_reg(rstate,
1490 R_028C6C_CB_COLOR0_VIEW + cb * 0x3C,
1491 0x00000000);
1492 } else {
1493 r600_pipe_state_add_reg(rstate,
1494 R_028C6C_CB_COLOR0_VIEW + cb * 0x3C,
1495 S_028C6C_SLICE_START(state->cbufs[cb]->u.tex.first_layer) |
1496 S_028C6C_SLICE_MAX(state->cbufs[cb]->u.tex.last_layer));
1497 }
1498 }
1499 r600_pipe_state_add_reg_bo(rstate,
1500 R_028C74_CB_COLOR0_ATTRIB + cb * 0x3C,
1501 color_attrib,
1502 &rtex->resource, RADEON_USAGE_READWRITE);
1503 }
1504
1505 static void evergreen_db(struct r600_context *rctx, struct r600_pipe_state *rstate,
1506 const struct pipe_framebuffer_state *state)
1507 {
1508 struct r600_screen *rscreen = rctx->screen;
1509 struct r600_resource_texture *rtex;
1510 struct r600_surface *surf;
1511 uint64_t offset;
1512 unsigned level, first_layer, pitch, slice, format, array_mode;
1513 unsigned macro_aspect, tile_split, bankh, bankw, z_info, nbanks;
1514
1515 if (state->zsbuf == NULL)
1516 return;
1517
1518 surf = (struct r600_surface *)state->zsbuf;
1519 level = surf->base.u.tex.level;
1520 rtex = (struct r600_resource_texture*)surf->base.texture;
1521 first_layer = surf->base.u.tex.first_layer;
1522 format = r600_translate_dbformat(rtex->real_format);
1523
1524 offset = r600_resource_va(rctx->context.screen, surf->base.texture);
1525 /* XXX remove this once tiling is properly supported */
1526 if (!rscreen->use_surface_alloc) {
1527 /* XXX remove this once tiling is properly supported */
1528 array_mode = rtex->array_mode[level] ? rtex->array_mode[level] :
1529 V_028C70_ARRAY_1D_TILED_THIN1;
1530
1531 offset += r600_texture_get_offset(rtex, level, first_layer);
1532 pitch = (rtex->pitch_in_blocks[level] / 8) - 1;
1533 slice = ((rtex->pitch_in_blocks[level] * surf->aligned_height) / 64);
1534 if (slice) {
1535 slice = slice - 1;
1536 }
1537 tile_split = 0;
1538 macro_aspect = 0;
1539 bankw = 0;
1540 bankh = 0;
1541 } else {
1542 offset += rtex->surface.level[level].offset;
1543 pitch = (rtex->surface.level[level].nblk_x / 8) - 1;
1544 slice = (rtex->surface.level[level].nblk_x * rtex->surface.level[level].nblk_y) / 64;
1545 if (slice) {
1546 slice = slice - 1;
1547 }
1548 switch (rtex->surface.level[level].mode) {
1549 case RADEON_SURF_MODE_2D:
1550 array_mode = V_028C70_ARRAY_2D_TILED_THIN1;
1551 break;
1552 case RADEON_SURF_MODE_1D:
1553 case RADEON_SURF_MODE_LINEAR_ALIGNED:
1554 case RADEON_SURF_MODE_LINEAR:
1555 default:
1556 array_mode = V_028C70_ARRAY_1D_TILED_THIN1;
1557 break;
1558 }
1559 tile_split = rtex->surface.tile_split;
1560 macro_aspect = rtex->surface.mtilea;
1561 bankw = rtex->surface.bankw;
1562 bankh = rtex->surface.bankh;
1563 tile_split = eg_tile_split(tile_split);
1564 macro_aspect = eg_macro_tile_aspect(macro_aspect);
1565 bankw = eg_bank_wh(bankw);
1566 bankh = eg_bank_wh(bankh);
1567 }
1568 nbanks = eg_num_banks(rscreen->tiling_info.num_banks);
1569 offset >>= 8;
1570
1571 z_info = S_028040_ARRAY_MODE(array_mode) |
1572 S_028040_FORMAT(format) |
1573 S_028040_TILE_SPLIT(tile_split)|
1574 S_028040_NUM_BANKS(nbanks) |
1575 S_028040_BANK_WIDTH(bankw) |
1576 S_028040_BANK_HEIGHT(bankh) |
1577 S_028040_MACRO_TILE_ASPECT(macro_aspect);
1578
1579 r600_pipe_state_add_reg_bo(rstate, R_028048_DB_Z_READ_BASE,
1580 offset, &rtex->resource, RADEON_USAGE_READWRITE);
1581 r600_pipe_state_add_reg_bo(rstate, R_028050_DB_Z_WRITE_BASE,
1582 offset, &rtex->resource, RADEON_USAGE_READWRITE);
1583 if (!rscreen->use_surface_alloc) {
1584 r600_pipe_state_add_reg(rstate, R_028008_DB_DEPTH_VIEW,
1585 0x00000000);
1586 } else {
1587 r600_pipe_state_add_reg(rstate, R_028008_DB_DEPTH_VIEW,
1588 S_028008_SLICE_START(state->zsbuf->u.tex.first_layer) |
1589 S_028008_SLICE_MAX(state->zsbuf->u.tex.last_layer));
1590 }
1591
1592 if (rtex->stencil) {
1593 uint64_t stencil_offset =
1594 r600_texture_get_offset(rtex->stencil, level, first_layer);
1595 unsigned stile_split;
1596
1597 stile_split = eg_tile_split(rtex->stencil->surface.tile_split);
1598 stencil_offset += r600_resource_va(rctx->context.screen, (void*)rtex->stencil);
1599 stencil_offset >>= 8;
1600
1601 r600_pipe_state_add_reg_bo(rstate, R_02804C_DB_STENCIL_READ_BASE,
1602 stencil_offset, &rtex->stencil->resource, RADEON_USAGE_READWRITE);
1603 r600_pipe_state_add_reg_bo(rstate, R_028054_DB_STENCIL_WRITE_BASE,
1604 stencil_offset, &rtex->stencil->resource, RADEON_USAGE_READWRITE);
1605 r600_pipe_state_add_reg_bo(rstate, R_028044_DB_STENCIL_INFO,
1606 1 | S_028044_TILE_SPLIT(stile_split),
1607 &rtex->stencil->resource, RADEON_USAGE_READWRITE);
1608 } else {
1609 if (rscreen->use_surface_alloc && rtex->surface.flags & RADEON_SURF_SBUFFER) {
1610 uint64_t stencil_offset = rtex->surface.stencil_offset;
1611 unsigned stile_split = rtex->surface.stencil_tile_split;
1612
1613 stile_split = eg_tile_split(stile_split);
1614 stencil_offset += r600_resource_va(rctx->context.screen, surf->base.texture);
1615 stencil_offset += rtex->surface.level[level].offset / 4;
1616 stencil_offset >>= 8;
1617
1618 r600_pipe_state_add_reg_bo(rstate, R_02804C_DB_STENCIL_READ_BASE,
1619 stencil_offset, &rtex->resource,
1620 RADEON_USAGE_READWRITE);
1621 r600_pipe_state_add_reg_bo(rstate, R_028054_DB_STENCIL_WRITE_BASE,
1622 stencil_offset, &rtex->resource,
1623 RADEON_USAGE_READWRITE);
1624 r600_pipe_state_add_reg_bo(rstate, R_028044_DB_STENCIL_INFO,
1625 1 | S_028044_TILE_SPLIT(stile_split),
1626 &rtex->resource,
1627 RADEON_USAGE_READWRITE);
1628 } else {
1629 r600_pipe_state_add_reg_bo(rstate, R_02804C_DB_STENCIL_READ_BASE,
1630 offset, &rtex->resource,
1631 RADEON_USAGE_READWRITE);
1632 r600_pipe_state_add_reg_bo(rstate, R_028054_DB_STENCIL_WRITE_BASE,
1633 offset, &rtex->resource,
1634 RADEON_USAGE_READWRITE);
1635 r600_pipe_state_add_reg_bo(rstate, R_028044_DB_STENCIL_INFO,
1636 0, NULL, RADEON_USAGE_READWRITE);
1637 }
1638 }
1639
1640 r600_pipe_state_add_reg_bo(rstate, R_028040_DB_Z_INFO, z_info,
1641 &rtex->resource, RADEON_USAGE_READWRITE);
1642 r600_pipe_state_add_reg(rstate, R_028058_DB_DEPTH_SIZE,
1643 S_028058_PITCH_TILE_MAX(pitch));
1644 r600_pipe_state_add_reg(rstate, R_02805C_DB_DEPTH_SLICE,
1645 S_02805C_SLICE_TILE_MAX(slice));
1646 }
1647
1648 static void evergreen_set_framebuffer_state(struct pipe_context *ctx,
1649 const struct pipe_framebuffer_state *state)
1650 {
1651 struct r600_context *rctx = (struct r600_context *)ctx;
1652 struct r600_pipe_state *rstate = CALLOC_STRUCT(r600_pipe_state);
1653 uint32_t tl, br;
1654
1655 if (rstate == NULL)
1656 return;
1657
1658 r600_flush_framebuffer(rctx, false);
1659
1660 /* unreference old buffer and reference new one */
1661 rstate->id = R600_PIPE_STATE_FRAMEBUFFER;
1662
1663 util_copy_framebuffer_state(&rctx->framebuffer, state);
1664
1665 /* build states */
1666 rctx->have_depth_fb = 0;
1667 rctx->nr_cbufs = state->nr_cbufs;
1668 for (int i = 0; i < state->nr_cbufs; i++) {
1669 evergreen_cb(rctx, rstate, state, i);
1670 }
1671 if (state->zsbuf) {
1672 evergreen_db(rctx, rstate, state);
1673 }
1674
1675 rctx->fb_cb_shader_mask = 0;
1676 for (int i = 0; i < state->nr_cbufs; i++) {
1677 rctx->fb_cb_shader_mask |= 0xf << (i * 4);
1678 }
1679
1680 evergreen_get_scissor_rect(rctx, 0, 0, state->width, state->height, &tl, &br);
1681
1682 r600_pipe_state_add_reg(rstate,
1683 R_028204_PA_SC_WINDOW_SCISSOR_TL, tl);
1684 r600_pipe_state_add_reg(rstate,
1685 R_028208_PA_SC_WINDOW_SCISSOR_BR, br);
1686
1687 free(rctx->states[R600_PIPE_STATE_FRAMEBUFFER]);
1688 rctx->states[R600_PIPE_STATE_FRAMEBUFFER] = rstate;
1689 r600_context_pipe_state_set(rctx, rstate);
1690
1691 if (state->zsbuf) {
1692 evergreen_polygon_offset_update(rctx);
1693 }
1694 }
1695
1696 static void evergreen_emit_db_misc_state(struct r600_context *rctx, struct r600_atom *atom)
1697 {
1698 struct radeon_winsys_cs *cs = rctx->cs;
1699 struct r600_db_misc_state *a = (struct r600_db_misc_state*)atom;
1700 unsigned db_count_control = 0;
1701 unsigned db_render_override =
1702 S_02800C_FORCE_HIZ_ENABLE(V_02800C_FORCE_DISABLE) |
1703 S_02800C_FORCE_HIS_ENABLE0(V_02800C_FORCE_DISABLE) |
1704 S_02800C_FORCE_HIS_ENABLE1(V_02800C_FORCE_DISABLE);
1705
1706 if (a->occlusion_query_enabled) {
1707 db_count_control |= S_028004_PERFECT_ZPASS_COUNTS(1);
1708 db_render_override |= S_02800C_NOOP_CULL_DISABLE(1);
1709 }
1710
1711 r600_write_context_reg(cs, R_028004_DB_COUNT_CONTROL, db_count_control);
1712 r600_write_context_reg(cs, R_02800C_DB_RENDER_OVERRIDE, db_render_override);
1713 }
1714
1715 static void evergreen_emit_vertex_buffers(struct r600_context *rctx, struct r600_atom *atom)
1716 {
1717 struct radeon_winsys_cs *cs = rctx->cs;
1718 struct pipe_vertex_buffer *vb = rctx->vertex_buffer;
1719 unsigned count = rctx->nr_vertex_buffers;
1720 unsigned i;
1721 uint64_t va;
1722
1723 for (i = 0; i < count; i++) {
1724 struct r600_resource *rbuffer = (struct r600_resource*)vb[i].buffer;
1725
1726 if (!rbuffer) {
1727 continue;
1728 }
1729
1730 va = r600_resource_va(&rctx->screen->screen, &rbuffer->b.b);
1731 va += vb[i].buffer_offset;
1732
1733 /* fetch resources start at index 992 */
1734 r600_write_value(cs, PKT3(PKT3_SET_RESOURCE, 8, 0));
1735 r600_write_value(cs, (992 + i) * 8);
1736 r600_write_value(cs, va); /* RESOURCEi_WORD0 */
1737 r600_write_value(cs, rbuffer->buf->size - vb[i].buffer_offset - 1); /* RESOURCEi_WORD1 */
1738 r600_write_value(cs, /* RESOURCEi_WORD2 */
1739 S_030008_ENDIAN_SWAP(r600_endian_swap(32)) |
1740 S_030008_STRIDE(vb[i].stride) |
1741 S_030008_BASE_ADDRESS_HI(va >> 32UL));
1742 r600_write_value(cs, /* RESOURCEi_WORD3 */
1743 S_03000C_DST_SEL_X(V_03000C_SQ_SEL_X) |
1744 S_03000C_DST_SEL_Y(V_03000C_SQ_SEL_Y) |
1745 S_03000C_DST_SEL_Z(V_03000C_SQ_SEL_Z) |
1746 S_03000C_DST_SEL_W(V_03000C_SQ_SEL_W));
1747 r600_write_value(cs, 0); /* RESOURCEi_WORD4 */
1748 r600_write_value(cs, 0); /* RESOURCEi_WORD5 */
1749 r600_write_value(cs, 0); /* RESOURCEi_WORD6 */
1750 r600_write_value(cs, 0xc0000000); /* RESOURCEi_WORD7 */
1751
1752 r600_write_value(cs, PKT3(PKT3_NOP, 0, 0));
1753 r600_write_value(cs, r600_context_bo_reloc(rctx, rbuffer, RADEON_USAGE_READ));
1754 }
1755 }
1756
1757 static void evergreen_emit_constant_buffer(struct r600_context *rctx,
1758 struct r600_constbuf_state *state,
1759 unsigned buffer_id_base,
1760 unsigned reg_alu_constbuf_size,
1761 unsigned reg_alu_const_cache)
1762 {
1763 struct radeon_winsys_cs *cs = rctx->cs;
1764 uint32_t dirty_mask = state->dirty_mask;
1765
1766 while (dirty_mask) {
1767 struct pipe_constant_buffer *cb;
1768 struct r600_resource *rbuffer;
1769 uint64_t va;
1770 unsigned buffer_index = ffs(dirty_mask) - 1;
1771
1772 cb = &state->cb[buffer_index];
1773 rbuffer = (struct r600_resource*)cb->buffer;
1774 assert(rbuffer);
1775
1776 va = r600_resource_va(&rctx->screen->screen, &rbuffer->b.b);
1777 va += cb->buffer_offset;
1778
1779 r600_write_context_reg(cs, reg_alu_constbuf_size + buffer_index * 4,
1780 ALIGN_DIVUP(cb->buffer_size >> 4, 16));
1781 r600_write_context_reg(cs, reg_alu_const_cache + buffer_index * 4, va >> 8);
1782
1783 r600_write_value(cs, PKT3(PKT3_NOP, 0, 0));
1784 r600_write_value(cs, r600_context_bo_reloc(rctx, rbuffer, RADEON_USAGE_READ));
1785
1786 r600_write_value(cs, PKT3(PKT3_SET_RESOURCE, 8, 0));
1787 r600_write_value(cs, (buffer_id_base + buffer_index) * 8);
1788 r600_write_value(cs, va); /* RESOURCEi_WORD0 */
1789 r600_write_value(cs, rbuffer->buf->size - cb->buffer_offset - 1); /* RESOURCEi_WORD1 */
1790 r600_write_value(cs, /* RESOURCEi_WORD2 */
1791 S_030008_ENDIAN_SWAP(r600_endian_swap(32)) |
1792 S_030008_STRIDE(16) |
1793 S_030008_BASE_ADDRESS_HI(va >> 32UL));
1794 r600_write_value(cs, /* RESOURCEi_WORD3 */
1795 S_03000C_DST_SEL_X(V_03000C_SQ_SEL_X) |
1796 S_03000C_DST_SEL_Y(V_03000C_SQ_SEL_Y) |
1797 S_03000C_DST_SEL_Z(V_03000C_SQ_SEL_Z) |
1798 S_03000C_DST_SEL_W(V_03000C_SQ_SEL_W));
1799 r600_write_value(cs, 0); /* RESOURCEi_WORD4 */
1800 r600_write_value(cs, 0); /* RESOURCEi_WORD5 */
1801 r600_write_value(cs, 0); /* RESOURCEi_WORD6 */
1802 r600_write_value(cs, 0xc0000000); /* RESOURCEi_WORD7 */
1803
1804 r600_write_value(cs, PKT3(PKT3_NOP, 0, 0));
1805 r600_write_value(cs, r600_context_bo_reloc(rctx, rbuffer, RADEON_USAGE_READ));
1806
1807 dirty_mask &= ~(1 << buffer_index);
1808 }
1809 state->dirty_mask = 0;
1810 }
1811
1812 static void evergreen_emit_vs_constant_buffer(struct r600_context *rctx, struct r600_atom *atom)
1813 {
1814 evergreen_emit_constant_buffer(rctx, &rctx->vs_constbuf_state, 176,
1815 R_028180_ALU_CONST_BUFFER_SIZE_VS_0,
1816 R_028980_ALU_CONST_CACHE_VS_0);
1817 }
1818
1819 static void evergreen_emit_ps_constant_buffer(struct r600_context *rctx, struct r600_atom *atom)
1820 {
1821 evergreen_emit_constant_buffer(rctx, &rctx->ps_constbuf_state, 0,
1822 R_028140_ALU_CONST_BUFFER_SIZE_PS_0,
1823 R_028940_ALU_CONST_CACHE_PS_0);
1824 }
1825
1826 void evergreen_init_state_functions(struct r600_context *rctx)
1827 {
1828 r600_init_atom(&rctx->db_misc_state.atom, evergreen_emit_db_misc_state, 6, 0);
1829 r600_atom_dirty(rctx, &rctx->db_misc_state.atom);
1830 r600_init_atom(&rctx->vertex_buffer_state, evergreen_emit_vertex_buffers, 0, 0);
1831 r600_init_atom(&rctx->vs_constbuf_state.atom, evergreen_emit_vs_constant_buffer, 0, 0);
1832 r600_init_atom(&rctx->ps_constbuf_state.atom, evergreen_emit_ps_constant_buffer, 0, 0);
1833
1834 rctx->context.create_blend_state = evergreen_create_blend_state;
1835 rctx->context.create_depth_stencil_alpha_state = evergreen_create_dsa_state;
1836 rctx->context.create_fs_state = r600_create_shader_state;
1837 rctx->context.create_rasterizer_state = evergreen_create_rs_state;
1838 rctx->context.create_sampler_state = evergreen_create_sampler_state;
1839 rctx->context.create_sampler_view = evergreen_create_sampler_view;
1840 rctx->context.create_vertex_elements_state = r600_create_vertex_elements;
1841 rctx->context.create_vs_state = r600_create_shader_state;
1842 rctx->context.bind_blend_state = r600_bind_blend_state;
1843 rctx->context.bind_depth_stencil_alpha_state = r600_bind_dsa_state;
1844 rctx->context.bind_fragment_sampler_states = evergreen_bind_ps_sampler;
1845 rctx->context.bind_fs_state = r600_bind_ps_shader;
1846 rctx->context.bind_rasterizer_state = r600_bind_rs_state;
1847 rctx->context.bind_vertex_elements_state = r600_bind_vertex_elements;
1848 rctx->context.bind_vertex_sampler_states = evergreen_bind_vs_sampler;
1849 rctx->context.bind_vs_state = r600_bind_vs_shader;
1850 rctx->context.delete_blend_state = r600_delete_state;
1851 rctx->context.delete_depth_stencil_alpha_state = r600_delete_state;
1852 rctx->context.delete_fs_state = r600_delete_ps_shader;
1853 rctx->context.delete_rasterizer_state = r600_delete_rs_state;
1854 rctx->context.delete_sampler_state = r600_delete_state;
1855 rctx->context.delete_vertex_elements_state = r600_delete_vertex_element;
1856 rctx->context.delete_vs_state = r600_delete_vs_shader;
1857 rctx->context.set_blend_color = r600_set_blend_color;
1858 rctx->context.set_clip_state = evergreen_set_clip_state;
1859 rctx->context.set_constant_buffer = r600_set_constant_buffer;
1860 rctx->context.set_fragment_sampler_views = evergreen_set_ps_sampler_view;
1861 rctx->context.set_framebuffer_state = evergreen_set_framebuffer_state;
1862 rctx->context.set_polygon_stipple = evergreen_set_polygon_stipple;
1863 rctx->context.set_sample_mask = evergreen_set_sample_mask;
1864 rctx->context.set_scissor_state = evergreen_set_scissor_state;
1865 rctx->context.set_stencil_ref = r600_set_pipe_stencil_ref;
1866 rctx->context.set_vertex_buffers = r600_set_vertex_buffers;
1867 rctx->context.set_index_buffer = r600_set_index_buffer;
1868 rctx->context.set_vertex_sampler_views = evergreen_set_vs_sampler_view;
1869 rctx->context.set_viewport_state = evergreen_set_viewport_state;
1870 rctx->context.sampler_view_destroy = r600_sampler_view_destroy;
1871 rctx->context.redefine_user_buffer = u_default_redefine_user_buffer;
1872 rctx->context.texture_barrier = r600_texture_barrier;
1873 rctx->context.create_stream_output_target = r600_create_so_target;
1874 rctx->context.stream_output_target_destroy = r600_so_target_destroy;
1875 rctx->context.set_stream_output_targets = r600_set_so_targets;
1876 }
1877
1878 static void cayman_init_atom_start_cs(struct r600_context *rctx)
1879 {
1880 struct r600_command_buffer *cb = &rctx->start_cs_cmd;
1881
1882 r600_init_command_buffer(cb, 256, EMIT_EARLY);
1883
1884 /* This must be first. */
1885 r600_store_value(cb, PKT3(PKT3_CONTEXT_CONTROL, 1, 0));
1886 r600_store_value(cb, 0x80000000);
1887 r600_store_value(cb, 0x80000000);
1888
1889 r600_store_config_reg_seq(cb, R_008C00_SQ_CONFIG, 2);
1890 r600_store_value(cb, S_008C00_EXPORT_SRC_C(1)); /* R_008C00_SQ_CONFIG */
1891 /* always set the temp clauses */
1892 r600_store_value(cb, S_008C04_NUM_CLAUSE_TEMP_GPRS(4)); /* R_008C04_SQ_GPR_RESOURCE_MGMT_1 */
1893
1894 r600_store_config_reg_seq(cb, R_008C10_SQ_GLOBAL_GPR_RESOURCE_MGMT_1, 2);
1895 r600_store_value(cb, 0); /* R_008C10_SQ_GLOBAL_GPR_RESOURCE_MGMT_1 */
1896 r600_store_value(cb, 0); /* R_008C14_SQ_GLOBAL_GPR_RESOURCE_MGMT_2 */
1897
1898 r600_store_config_reg(cb, R_008D8C_SQ_DYN_GPR_CNTL_PS_FLUSH_REQ, (1 << 8));
1899
1900 r600_store_context_reg(cb, R_028A4C_PA_SC_MODE_CNTL_1, 0);
1901
1902 r600_store_context_reg_seq(cb, R_028A10_VGT_OUTPUT_PATH_CNTL, 13);
1903 r600_store_value(cb, 0); /* R_028A10_VGT_OUTPUT_PATH_CNTL */
1904 r600_store_value(cb, 0); /* R_028A14_VGT_HOS_CNTL */
1905 r600_store_value(cb, 0); /* R_028A18_VGT_HOS_MAX_TESS_LEVEL */
1906 r600_store_value(cb, 0); /* R_028A1C_VGT_HOS_MIN_TESS_LEVEL */
1907 r600_store_value(cb, 0); /* R_028A20_VGT_HOS_REUSE_DEPTH */
1908 r600_store_value(cb, 0); /* R_028A24_VGT_GROUP_PRIM_TYPE */
1909 r600_store_value(cb, 0); /* R_028A28_VGT_GROUP_FIRST_DECR */
1910 r600_store_value(cb, 0); /* R_028A2C_VGT_GROUP_DECR */
1911 r600_store_value(cb, 0); /* R_028A30_VGT_GROUP_VECT_0_CNTL */
1912 r600_store_value(cb, 0); /* R_028A34_VGT_GROUP_VECT_1_CNTL */
1913 r600_store_value(cb, 0); /* R_028A38_VGT_GROUP_VECT_0_FMT_CNTL */
1914 r600_store_value(cb, 0); /* R_028A3C_VGT_GROUP_VECT_1_FMT_CNTL */
1915 r600_store_value(cb, 0); /* R_028A40_VGT_GS_MODE */
1916
1917 r600_store_context_reg_seq(cb, R_028B94_VGT_STRMOUT_CONFIG, 2);
1918 r600_store_value(cb, 0); /* R_028B94_VGT_STRMOUT_CONFIG */
1919 r600_store_value(cb, 0); /* R_028B98_VGT_STRMOUT_BUFFER_CONFIG */
1920
1921 r600_store_context_reg_seq(cb, R_028AB4_VGT_REUSE_OFF, 2);
1922 r600_store_value(cb, 0); /* R_028AB4_VGT_REUSE_OFF */
1923 r600_store_value(cb, 0); /* R_028AB8_VGT_VTX_CNT_EN */
1924
1925 r600_store_config_reg(cb, R_008A14_PA_CL_ENHANCE, (3 << 1) | 1);
1926
1927 r600_store_context_reg(cb, CM_R_028AA8_IA_MULTI_VGT_PARAM, S_028AA8_SWITCH_ON_EOP(1) | S_028AA8_PARTIAL_VS_WAVE_ON(1) | S_028AA8_PRIMGROUP_SIZE(63));
1928
1929 r600_store_context_reg_seq(cb, CM_R_028BD4_PA_SC_CENTROID_PRIORITY_0, 2);
1930 r600_store_value(cb, 0x76543210); /* CM_R_028BD4_PA_SC_CENTROID_PRIORITY_0 */
1931 r600_store_value(cb, 0xfedcba98); /* CM_R_028BD8_PA_SC_CENTROID_PRIORITY_1 */
1932
1933 r600_store_context_reg_seq(cb, CM_R_0288E8_SQ_LDS_ALLOC, 2);
1934 r600_store_value(cb, 0); /* CM_R_0288E8_SQ_LDS_ALLOC */
1935 r600_store_value(cb, 0); /* R_0288EC_SQ_LDS_ALLOC_PS */
1936
1937 r600_store_context_reg(cb, CM_R_028804_DB_EQAA, 0x110000);
1938
1939 r600_store_context_reg_seq(cb, R_028380_SQ_VTX_SEMANTIC_0, 34);
1940 r600_store_value(cb, 0); /* R_028380_SQ_VTX_SEMANTIC_0 */
1941 r600_store_value(cb, 0);
1942 r600_store_value(cb, 0);
1943 r600_store_value(cb, 0);
1944 r600_store_value(cb, 0);
1945 r600_store_value(cb, 0);
1946 r600_store_value(cb, 0);
1947 r600_store_value(cb, 0);
1948 r600_store_value(cb, 0);
1949 r600_store_value(cb, 0);
1950 r600_store_value(cb, 0);
1951 r600_store_value(cb, 0);
1952 r600_store_value(cb, 0);
1953 r600_store_value(cb, 0);
1954 r600_store_value(cb, 0);
1955 r600_store_value(cb, 0);
1956 r600_store_value(cb, 0);
1957 r600_store_value(cb, 0);
1958 r600_store_value(cb, 0);
1959 r600_store_value(cb, 0);
1960 r600_store_value(cb, 0);
1961 r600_store_value(cb, 0);
1962 r600_store_value(cb, 0);
1963 r600_store_value(cb, 0);
1964 r600_store_value(cb, 0);
1965 r600_store_value(cb, 0);
1966 r600_store_value(cb, 0);
1967 r600_store_value(cb, 0);
1968 r600_store_value(cb, 0);
1969 r600_store_value(cb, 0);
1970 r600_store_value(cb, 0);
1971 r600_store_value(cb, 0); /* R_0283FC_SQ_VTX_SEMANTIC_31 */
1972 r600_store_value(cb, ~0); /* R_028400_VGT_MAX_VTX_INDX */
1973 r600_store_value(cb, 0); /* R_028404_VGT_MIN_VTX_INDX */
1974
1975 r600_store_ctl_const(cb, R_03CFF0_SQ_VTX_BASE_VTX_LOC, 0);
1976
1977 r600_store_context_reg_seq(cb, CM_R_028C38_PA_SC_AA_MASK_X0Y0_X1Y0, 2);
1978 r600_store_value(cb, ~0); /* CM_R_028C38_PA_SC_AA_MASK_X0Y0_X1Y0 */
1979 r600_store_value(cb, ~0); /* CM_R_028C3C_PA_SC_AA_MASK_X0Y1_X1Y1 */
1980
1981 r600_store_context_reg_seq(cb, R_028028_DB_STENCIL_CLEAR, 2);
1982 r600_store_value(cb, 0); /* R_028028_DB_STENCIL_CLEAR */
1983 r600_store_value(cb, 0x3F800000); /* R_02802C_DB_DEPTH_CLEAR */
1984
1985 r600_store_context_reg(cb, R_0286DC_SPI_FOG_CNTL, 0);
1986
1987 r600_store_context_reg_seq(cb, R_028AC0_DB_SRESULTS_COMPARE_STATE0, 3);
1988 r600_store_value(cb, 0); /* R_028AC0_DB_SRESULTS_COMPARE_STATE0 */
1989 r600_store_value(cb, 0); /* R_028AC4_DB_SRESULTS_COMPARE_STATE1 */
1990 r600_store_value(cb, 0); /* R_028AC8_DB_PRELOAD_CONTROL */
1991
1992 r600_store_context_reg(cb, R_028200_PA_SC_WINDOW_OFFSET, 0);
1993 r600_store_context_reg(cb, R_02820C_PA_SC_CLIPRECT_RULE, 0xFFFF);
1994
1995 r600_store_context_reg_seq(cb, R_0282D0_PA_SC_VPORT_ZMIN_0, 2);
1996 r600_store_value(cb, 0); /* R_0282D0_PA_SC_VPORT_ZMIN_0 */
1997 r600_store_value(cb, 0x3F800000); /* R_0282D4_PA_SC_VPORT_ZMAX_0 */
1998
1999 r600_store_context_reg(cb, R_028230_PA_SC_EDGERULE, 0xAAAAAAAA);
2000 r600_store_context_reg(cb, R_028818_PA_CL_VTE_CNTL, 0x0000043F);
2001 r600_store_context_reg(cb, R_028820_PA_CL_NANINF_CNTL, 0);
2002 r600_store_context_reg(cb, R_028B70_DB_ALPHA_TO_MASK, 0x0000AA00);
2003
2004 r600_store_context_reg_seq(cb, CM_R_028BDC_PA_SC_LINE_CNTL, 2);
2005 r600_store_value(cb, 0x00000400); /* CM_R_028BDC_PA_SC_LINE_CNTL */
2006 r600_store_value(cb, 0); /* CM_R_028BE0_PA_SC_AA_CONFIG */
2007
2008 r600_store_context_reg_seq(cb, CM_R_028BE8_PA_CL_GB_VERT_CLIP_ADJ, 4);
2009 r600_store_value(cb, 0x3F800000); /* CM_R_028BE8_PA_CL_GB_VERT_CLIP_ADJ */
2010 r600_store_value(cb, 0x3F800000); /* CM_R_028BEC_PA_CL_GB_VERT_DISC_ADJ */
2011 r600_store_value(cb, 0x3F800000); /* CM_R_028BF0_PA_CL_GB_HORZ_CLIP_ADJ */
2012 r600_store_value(cb, 0x3F800000); /* CM_R_028BF4_PA_CL_GB_HORZ_DISC_ADJ */
2013
2014 r600_store_context_reg_seq(cb, R_028240_PA_SC_GENERIC_SCISSOR_TL, 2);
2015 r600_store_value(cb, 0); /* R_028240_PA_SC_GENERIC_SCISSOR_TL */
2016 r600_store_value(cb, S_028244_BR_X(16384) | S_028244_BR_Y(16384)); /* R_028244_PA_SC_GENERIC_SCISSOR_BR */
2017
2018 r600_store_context_reg_seq(cb, R_028030_PA_SC_SCREEN_SCISSOR_TL, 2);
2019 r600_store_value(cb, 0); /* R_028030_PA_SC_SCREEN_SCISSOR_TL */
2020 r600_store_value(cb, S_028034_BR_X(16384) | S_028034_BR_Y(16384)); /* R_028034_PA_SC_SCREEN_SCISSOR_BR */
2021
2022 r600_store_context_reg(cb, R_028848_SQ_PGM_RESOURCES_2_PS, S_028848_SINGLE_ROUND(V_SQ_ROUND_TO_ZERO));
2023 r600_store_context_reg(cb, R_028864_SQ_PGM_RESOURCES_2_VS, S_028864_SINGLE_ROUND(V_SQ_ROUND_TO_ZERO));
2024 r600_store_context_reg(cb, R_0288A8_SQ_PGM_RESOURCES_FS, 0);
2025
2026 r600_store_context_reg(cb, R_028354_SX_SURFACE_SYNC, S_028354_SURFACE_SYNC_MASK(0xf));
2027 r600_store_context_reg(cb, R_028800_DB_DEPTH_CONTROL, 0);
2028
2029 eg_store_loop_const(cb, R_03A200_SQ_LOOP_CONST_0, 0x01000FFF);
2030 eg_store_loop_const(cb, R_03A200_SQ_LOOP_CONST_0 + (32 * 4), 0x01000FFF);
2031 }
2032
2033 void evergreen_init_atom_start_cs(struct r600_context *rctx)
2034 {
2035 struct r600_command_buffer *cb = &rctx->start_cs_cmd;
2036 int ps_prio;
2037 int vs_prio;
2038 int gs_prio;
2039 int es_prio;
2040 int hs_prio, cs_prio, ls_prio;
2041 int num_ps_gprs;
2042 int num_vs_gprs;
2043 int num_gs_gprs;
2044 int num_es_gprs;
2045 int num_hs_gprs;
2046 int num_ls_gprs;
2047 int num_temp_gprs;
2048 int num_ps_threads;
2049 int num_vs_threads;
2050 int num_gs_threads;
2051 int num_es_threads;
2052 int num_hs_threads;
2053 int num_ls_threads;
2054 int num_ps_stack_entries;
2055 int num_vs_stack_entries;
2056 int num_gs_stack_entries;
2057 int num_es_stack_entries;
2058 int num_hs_stack_entries;
2059 int num_ls_stack_entries;
2060 enum radeon_family family;
2061 unsigned tmp;
2062
2063 if (rctx->chip_class == CAYMAN) {
2064 cayman_init_atom_start_cs(rctx);
2065 return;
2066 }
2067
2068 r600_init_command_buffer(cb, 256, EMIT_EARLY);
2069
2070 /* This must be first. */
2071 r600_store_value(cb, PKT3(PKT3_CONTEXT_CONTROL, 1, 0));
2072 r600_store_value(cb, 0x80000000);
2073 r600_store_value(cb, 0x80000000);
2074
2075 family = rctx->family;
2076 ps_prio = 0;
2077 vs_prio = 1;
2078 gs_prio = 2;
2079 es_prio = 3;
2080 hs_prio = 0;
2081 ls_prio = 0;
2082 cs_prio = 0;
2083
2084 switch (family) {
2085 case CHIP_CEDAR:
2086 default:
2087 num_ps_gprs = 93;
2088 num_vs_gprs = 46;
2089 num_temp_gprs = 4;
2090 num_gs_gprs = 31;
2091 num_es_gprs = 31;
2092 num_hs_gprs = 23;
2093 num_ls_gprs = 23;
2094 num_ps_threads = 96;
2095 num_vs_threads = 16;
2096 num_gs_threads = 16;
2097 num_es_threads = 16;
2098 num_hs_threads = 16;
2099 num_ls_threads = 16;
2100 num_ps_stack_entries = 42;
2101 num_vs_stack_entries = 42;
2102 num_gs_stack_entries = 42;
2103 num_es_stack_entries = 42;
2104 num_hs_stack_entries = 42;
2105 num_ls_stack_entries = 42;
2106 break;
2107 case CHIP_REDWOOD:
2108 num_ps_gprs = 93;
2109 num_vs_gprs = 46;
2110 num_temp_gprs = 4;
2111 num_gs_gprs = 31;
2112 num_es_gprs = 31;
2113 num_hs_gprs = 23;
2114 num_ls_gprs = 23;
2115 num_ps_threads = 128;
2116 num_vs_threads = 20;
2117 num_gs_threads = 20;
2118 num_es_threads = 20;
2119 num_hs_threads = 20;
2120 num_ls_threads = 20;
2121 num_ps_stack_entries = 42;
2122 num_vs_stack_entries = 42;
2123 num_gs_stack_entries = 42;
2124 num_es_stack_entries = 42;
2125 num_hs_stack_entries = 42;
2126 num_ls_stack_entries = 42;
2127 break;
2128 case CHIP_JUNIPER:
2129 num_ps_gprs = 93;
2130 num_vs_gprs = 46;
2131 num_temp_gprs = 4;
2132 num_gs_gprs = 31;
2133 num_es_gprs = 31;
2134 num_hs_gprs = 23;
2135 num_ls_gprs = 23;
2136 num_ps_threads = 128;
2137 num_vs_threads = 20;
2138 num_gs_threads = 20;
2139 num_es_threads = 20;
2140 num_hs_threads = 20;
2141 num_ls_threads = 20;
2142 num_ps_stack_entries = 85;
2143 num_vs_stack_entries = 85;
2144 num_gs_stack_entries = 85;
2145 num_es_stack_entries = 85;
2146 num_hs_stack_entries = 85;
2147 num_ls_stack_entries = 85;
2148 break;
2149 case CHIP_CYPRESS:
2150 case CHIP_HEMLOCK:
2151 num_ps_gprs = 93;
2152 num_vs_gprs = 46;
2153 num_temp_gprs = 4;
2154 num_gs_gprs = 31;
2155 num_es_gprs = 31;
2156 num_hs_gprs = 23;
2157 num_ls_gprs = 23;
2158 num_ps_threads = 128;
2159 num_vs_threads = 20;
2160 num_gs_threads = 20;
2161 num_es_threads = 20;
2162 num_hs_threads = 20;
2163 num_ls_threads = 20;
2164 num_ps_stack_entries = 85;
2165 num_vs_stack_entries = 85;
2166 num_gs_stack_entries = 85;
2167 num_es_stack_entries = 85;
2168 num_hs_stack_entries = 85;
2169 num_ls_stack_entries = 85;
2170 break;
2171 case CHIP_PALM:
2172 num_ps_gprs = 93;
2173 num_vs_gprs = 46;
2174 num_temp_gprs = 4;
2175 num_gs_gprs = 31;
2176 num_es_gprs = 31;
2177 num_hs_gprs = 23;
2178 num_ls_gprs = 23;
2179 num_ps_threads = 96;
2180 num_vs_threads = 16;
2181 num_gs_threads = 16;
2182 num_es_threads = 16;
2183 num_hs_threads = 16;
2184 num_ls_threads = 16;
2185 num_ps_stack_entries = 42;
2186 num_vs_stack_entries = 42;
2187 num_gs_stack_entries = 42;
2188 num_es_stack_entries = 42;
2189 num_hs_stack_entries = 42;
2190 num_ls_stack_entries = 42;
2191 break;
2192 case CHIP_SUMO:
2193 num_ps_gprs = 93;
2194 num_vs_gprs = 46;
2195 num_temp_gprs = 4;
2196 num_gs_gprs = 31;
2197 num_es_gprs = 31;
2198 num_hs_gprs = 23;
2199 num_ls_gprs = 23;
2200 num_ps_threads = 96;
2201 num_vs_threads = 25;
2202 num_gs_threads = 25;
2203 num_es_threads = 25;
2204 num_hs_threads = 25;
2205 num_ls_threads = 25;
2206 num_ps_stack_entries = 42;
2207 num_vs_stack_entries = 42;
2208 num_gs_stack_entries = 42;
2209 num_es_stack_entries = 42;
2210 num_hs_stack_entries = 42;
2211 num_ls_stack_entries = 42;
2212 break;
2213 case CHIP_SUMO2:
2214 num_ps_gprs = 93;
2215 num_vs_gprs = 46;
2216 num_temp_gprs = 4;
2217 num_gs_gprs = 31;
2218 num_es_gprs = 31;
2219 num_hs_gprs = 23;
2220 num_ls_gprs = 23;
2221 num_ps_threads = 96;
2222 num_vs_threads = 25;
2223 num_gs_threads = 25;
2224 num_es_threads = 25;
2225 num_hs_threads = 25;
2226 num_ls_threads = 25;
2227 num_ps_stack_entries = 85;
2228 num_vs_stack_entries = 85;
2229 num_gs_stack_entries = 85;
2230 num_es_stack_entries = 85;
2231 num_hs_stack_entries = 85;
2232 num_ls_stack_entries = 85;
2233 break;
2234 case CHIP_BARTS:
2235 num_ps_gprs = 93;
2236 num_vs_gprs = 46;
2237 num_temp_gprs = 4;
2238 num_gs_gprs = 31;
2239 num_es_gprs = 31;
2240 num_hs_gprs = 23;
2241 num_ls_gprs = 23;
2242 num_ps_threads = 128;
2243 num_vs_threads = 20;
2244 num_gs_threads = 20;
2245 num_es_threads = 20;
2246 num_hs_threads = 20;
2247 num_ls_threads = 20;
2248 num_ps_stack_entries = 85;
2249 num_vs_stack_entries = 85;
2250 num_gs_stack_entries = 85;
2251 num_es_stack_entries = 85;
2252 num_hs_stack_entries = 85;
2253 num_ls_stack_entries = 85;
2254 break;
2255 case CHIP_TURKS:
2256 num_ps_gprs = 93;
2257 num_vs_gprs = 46;
2258 num_temp_gprs = 4;
2259 num_gs_gprs = 31;
2260 num_es_gprs = 31;
2261 num_hs_gprs = 23;
2262 num_ls_gprs = 23;
2263 num_ps_threads = 128;
2264 num_vs_threads = 20;
2265 num_gs_threads = 20;
2266 num_es_threads = 20;
2267 num_hs_threads = 20;
2268 num_ls_threads = 20;
2269 num_ps_stack_entries = 42;
2270 num_vs_stack_entries = 42;
2271 num_gs_stack_entries = 42;
2272 num_es_stack_entries = 42;
2273 num_hs_stack_entries = 42;
2274 num_ls_stack_entries = 42;
2275 break;
2276 case CHIP_CAICOS:
2277 num_ps_gprs = 93;
2278 num_vs_gprs = 46;
2279 num_temp_gprs = 4;
2280 num_gs_gprs = 31;
2281 num_es_gprs = 31;
2282 num_hs_gprs = 23;
2283 num_ls_gprs = 23;
2284 num_ps_threads = 128;
2285 num_vs_threads = 10;
2286 num_gs_threads = 10;
2287 num_es_threads = 10;
2288 num_hs_threads = 10;
2289 num_ls_threads = 10;
2290 num_ps_stack_entries = 42;
2291 num_vs_stack_entries = 42;
2292 num_gs_stack_entries = 42;
2293 num_es_stack_entries = 42;
2294 num_hs_stack_entries = 42;
2295 num_ls_stack_entries = 42;
2296 break;
2297 }
2298
2299 tmp = 0;
2300 switch (family) {
2301 case CHIP_CEDAR:
2302 case CHIP_PALM:
2303 case CHIP_SUMO:
2304 case CHIP_SUMO2:
2305 case CHIP_CAICOS:
2306 break;
2307 default:
2308 tmp |= S_008C00_VC_ENABLE(1);
2309 break;
2310 }
2311 tmp |= S_008C00_EXPORT_SRC_C(1);
2312 tmp |= S_008C00_CS_PRIO(cs_prio);
2313 tmp |= S_008C00_LS_PRIO(ls_prio);
2314 tmp |= S_008C00_HS_PRIO(hs_prio);
2315 tmp |= S_008C00_PS_PRIO(ps_prio);
2316 tmp |= S_008C00_VS_PRIO(vs_prio);
2317 tmp |= S_008C00_GS_PRIO(gs_prio);
2318 tmp |= S_008C00_ES_PRIO(es_prio);
2319
2320 /* enable dynamic GPR resource management */
2321 if (rctx->screen->info.drm_minor >= 7) {
2322 r600_store_config_reg_seq(cb, R_008C00_SQ_CONFIG, 2);
2323 r600_store_value(cb, tmp); /* R_008C00_SQ_CONFIG */
2324 /* always set temp clauses */
2325 r600_store_value(cb, S_008C04_NUM_CLAUSE_TEMP_GPRS(num_temp_gprs)); /* R_008C04_SQ_GPR_RESOURCE_MGMT_1 */
2326 r600_store_config_reg_seq(cb, R_008C10_SQ_GLOBAL_GPR_RESOURCE_MGMT_1, 2);
2327 r600_store_value(cb, 0); /* R_008C10_SQ_GLOBAL_GPR_RESOURCE_MGMT_1 */
2328 r600_store_value(cb, 0); /* R_008C14_SQ_GLOBAL_GPR_RESOURCE_MGMT_2 */
2329 r600_store_config_reg(cb, R_008D8C_SQ_DYN_GPR_CNTL_PS_FLUSH_REQ, (1 << 8));
2330 r600_store_context_reg(cb, R_028838_SQ_DYN_GPR_RESOURCE_LIMIT_1,
2331 S_028838_PS_GPRS(0x1e) |
2332 S_028838_VS_GPRS(0x1e) |
2333 S_028838_GS_GPRS(0x1e) |
2334 S_028838_ES_GPRS(0x1e) |
2335 S_028838_HS_GPRS(0x1e) |
2336 S_028838_LS_GPRS(0x1e)); /* workaround for hw issues with dyn gpr - must set all limits to 240 instead of 0, 0x1e == 240 / 8*/
2337 } else {
2338 r600_store_config_reg_seq(cb, R_008C00_SQ_CONFIG, 4);
2339 r600_store_value(cb, tmp); /* R_008C00_SQ_CONFIG */
2340
2341 tmp = S_008C04_NUM_PS_GPRS(num_ps_gprs);
2342 tmp |= S_008C04_NUM_VS_GPRS(num_vs_gprs);
2343 tmp |= S_008C04_NUM_CLAUSE_TEMP_GPRS(num_temp_gprs);
2344 r600_store_value(cb, tmp); /* R_008C04_SQ_GPR_RESOURCE_MGMT_1 */
2345
2346 tmp = S_008C08_NUM_GS_GPRS(num_gs_gprs);
2347 tmp |= S_008C08_NUM_ES_GPRS(num_es_gprs);
2348 r600_store_value(cb, tmp); /* R_008C08_SQ_GPR_RESOURCE_MGMT_2 */
2349
2350 tmp = S_008C0C_NUM_HS_GPRS(num_hs_gprs);
2351 tmp |= S_008C0C_NUM_HS_GPRS(num_ls_gprs);
2352 r600_store_value(cb, tmp); /* R_008C0C_SQ_GPR_RESOURCE_MGMT_3 */
2353 }
2354
2355 tmp = S_008C18_NUM_PS_THREADS(num_ps_threads);
2356 tmp |= S_008C18_NUM_VS_THREADS(num_vs_threads);
2357 tmp |= S_008C18_NUM_GS_THREADS(num_gs_threads);
2358 tmp |= S_008C18_NUM_ES_THREADS(num_es_threads);
2359 r600_store_config_reg_seq(cb, R_008C18_SQ_THREAD_RESOURCE_MGMT_1, 5);
2360 r600_store_value(cb, tmp); /* R_008C18_SQ_THREAD_RESOURCE_MGMT_1 */
2361
2362 tmp = S_008C1C_NUM_HS_THREADS(num_hs_threads);
2363 tmp |= S_008C1C_NUM_LS_THREADS(num_ls_threads);
2364 r600_store_value(cb, tmp); /* R_008C1C_SQ_THREAD_RESOURCE_MGMT_2 */
2365
2366 tmp = S_008C20_NUM_PS_STACK_ENTRIES(num_ps_stack_entries);
2367 tmp |= S_008C20_NUM_VS_STACK_ENTRIES(num_vs_stack_entries);
2368 r600_store_value(cb, tmp); /* R_008C20_SQ_STACK_RESOURCE_MGMT_1 */
2369
2370 tmp = S_008C24_NUM_GS_STACK_ENTRIES(num_gs_stack_entries);
2371 tmp |= S_008C24_NUM_ES_STACK_ENTRIES(num_es_stack_entries);
2372 r600_store_value(cb, tmp); /* R_008C24_SQ_STACK_RESOURCE_MGMT_2 */
2373
2374 tmp = S_008C28_NUM_HS_STACK_ENTRIES(num_hs_stack_entries);
2375 tmp |= S_008C28_NUM_LS_STACK_ENTRIES(num_ls_stack_entries);
2376 r600_store_value(cb, tmp); /* R_008C28_SQ_STACK_RESOURCE_MGMT_3 */
2377
2378 r600_store_config_reg(cb, R_008E2C_SQ_LDS_RESOURCE_MGMT,
2379 S_008E2C_NUM_PS_LDS(0x1000) | S_008E2C_NUM_LS_LDS(0x1000));
2380
2381 r600_store_config_reg(cb, R_009100_SPI_CONFIG_CNTL, 0);
2382 r600_store_config_reg(cb, R_00913C_SPI_CONFIG_CNTL_1, S_00913C_VTX_DONE_DELAY(4));
2383
2384 r600_store_context_reg(cb, R_028A4C_PA_SC_MODE_CNTL_1, 0);
2385
2386 r600_store_context_reg_seq(cb, R_028900_SQ_ESGS_RING_ITEMSIZE, 6);
2387 r600_store_value(cb, 0); /* R_028900_SQ_ESGS_RING_ITEMSIZE */
2388 r600_store_value(cb, 0); /* R_028904_SQ_GSVS_RING_ITEMSIZE */
2389 r600_store_value(cb, 0); /* R_028908_SQ_ESTMP_RING_ITEMSIZE */
2390 r600_store_value(cb, 0); /* R_02890C_SQ_GSTMP_RING_ITEMSIZE */
2391 r600_store_value(cb, 0); /* R_028910_SQ_VSTMP_RING_ITEMSIZE */
2392 r600_store_value(cb, 0); /* R_028914_SQ_PSTMP_RING_ITEMSIZE */
2393
2394 r600_store_context_reg_seq(cb, R_02891C_SQ_GS_VERT_ITEMSIZE, 4);
2395 r600_store_value(cb, 0); /* R_02891C_SQ_GS_VERT_ITEMSIZE */
2396 r600_store_value(cb, 0); /* R_028920_SQ_GS_VERT_ITEMSIZE_1 */
2397 r600_store_value(cb, 0); /* R_028924_SQ_GS_VERT_ITEMSIZE_2 */
2398 r600_store_value(cb, 0); /* R_028928_SQ_GS_VERT_ITEMSIZE_3 */
2399
2400 r600_store_context_reg_seq(cb, R_028A10_VGT_OUTPUT_PATH_CNTL, 13);
2401 r600_store_value(cb, 0); /* R_028A10_VGT_OUTPUT_PATH_CNTL */
2402 r600_store_value(cb, 0); /* R_028A14_VGT_HOS_CNTL */
2403 r600_store_value(cb, 0); /* R_028A18_VGT_HOS_MAX_TESS_LEVEL */
2404 r600_store_value(cb, 0); /* R_028A1C_VGT_HOS_MIN_TESS_LEVEL */
2405 r600_store_value(cb, 0); /* R_028A20_VGT_HOS_REUSE_DEPTH */
2406 r600_store_value(cb, 0); /* R_028A24_VGT_GROUP_PRIM_TYPE */
2407 r600_store_value(cb, 0); /* R_028A28_VGT_GROUP_FIRST_DECR */
2408 r600_store_value(cb, 0); /* R_028A2C_VGT_GROUP_DECR */
2409 r600_store_value(cb, 0); /* R_028A30_VGT_GROUP_VECT_0_CNTL */
2410 r600_store_value(cb, 0); /* R_028A34_VGT_GROUP_VECT_1_CNTL */
2411 r600_store_value(cb, 0); /* R_028A38_VGT_GROUP_VECT_0_FMT_CNTL */
2412 r600_store_value(cb, 0); /* R_028A3C_VGT_GROUP_VECT_1_FMT_CNTL */
2413 r600_store_value(cb, 0); /* R_028A40_VGT_GS_MODE */
2414
2415 r600_store_context_reg_seq(cb, R_028B94_VGT_STRMOUT_CONFIG, 2);
2416 r600_store_value(cb, 0); /* R_028B94_VGT_STRMOUT_CONFIG */
2417 r600_store_value(cb, 0); /* R_028B98_VGT_STRMOUT_BUFFER_CONFIG */
2418
2419 r600_store_context_reg_seq(cb, R_028AB4_VGT_REUSE_OFF, 2);
2420 r600_store_value(cb, 0); /* R_028AB4_VGT_REUSE_OFF */
2421 r600_store_value(cb, 0); /* R_028AB8_VGT_VTX_CNT_EN */
2422
2423 r600_store_config_reg(cb, R_008A14_PA_CL_ENHANCE, (3 << 1) | 1);
2424
2425 r600_store_context_reg_seq(cb, R_028380_SQ_VTX_SEMANTIC_0, 34);
2426 r600_store_value(cb, 0); /* R_028380_SQ_VTX_SEMANTIC_0 */
2427 r600_store_value(cb, 0);
2428 r600_store_value(cb, 0);
2429 r600_store_value(cb, 0);
2430 r600_store_value(cb, 0);
2431 r600_store_value(cb, 0);
2432 r600_store_value(cb, 0);
2433 r600_store_value(cb, 0);
2434 r600_store_value(cb, 0);
2435 r600_store_value(cb, 0);
2436 r600_store_value(cb, 0);
2437 r600_store_value(cb, 0);
2438 r600_store_value(cb, 0);
2439 r600_store_value(cb, 0);
2440 r600_store_value(cb, 0);
2441 r600_store_value(cb, 0);
2442 r600_store_value(cb, 0);
2443 r600_store_value(cb, 0);
2444 r600_store_value(cb, 0);
2445 r600_store_value(cb, 0);
2446 r600_store_value(cb, 0);
2447 r600_store_value(cb, 0);
2448 r600_store_value(cb, 0);
2449 r600_store_value(cb, 0);
2450 r600_store_value(cb, 0);
2451 r600_store_value(cb, 0);
2452 r600_store_value(cb, 0);
2453 r600_store_value(cb, 0);
2454 r600_store_value(cb, 0);
2455 r600_store_value(cb, 0);
2456 r600_store_value(cb, 0);
2457 r600_store_value(cb, 0); /* R_0283FC_SQ_VTX_SEMANTIC_31 */
2458 r600_store_value(cb, ~0); /* R_028400_VGT_MAX_VTX_INDX */
2459 r600_store_value(cb, 0); /* R_028404_VGT_MIN_VTX_INDX */
2460
2461 r600_store_ctl_const(cb, R_03CFF0_SQ_VTX_BASE_VTX_LOC, 0);
2462
2463 r600_store_context_reg_seq(cb, R_028028_DB_STENCIL_CLEAR, 2);
2464 r600_store_value(cb, 0); /* R_028028_DB_STENCIL_CLEAR */
2465 r600_store_value(cb, 0x3F800000); /* R_02802C_DB_DEPTH_CLEAR */
2466
2467 r600_store_context_reg(cb, R_028200_PA_SC_WINDOW_OFFSET, 0);
2468 r600_store_context_reg(cb, R_02820C_PA_SC_CLIPRECT_RULE, 0xFFFF);
2469 r600_store_context_reg(cb, R_028230_PA_SC_EDGERULE, 0xAAAAAAAA);
2470
2471 r600_store_context_reg_seq(cb, R_0282D0_PA_SC_VPORT_ZMIN_0, 2);
2472 r600_store_value(cb, 0); /* R_0282D0_PA_SC_VPORT_ZMIN_0 */
2473 r600_store_value(cb, 0x3F800000); /* R_0282D4_PA_SC_VPORT_ZMAX_0 */
2474
2475 r600_store_context_reg(cb, R_0286DC_SPI_FOG_CNTL, 0);
2476 r600_store_context_reg(cb, R_028818_PA_CL_VTE_CNTL, 0x0000043F);
2477 r600_store_context_reg(cb, R_028820_PA_CL_NANINF_CNTL, 0);
2478
2479 r600_store_context_reg_seq(cb, R_028AC0_DB_SRESULTS_COMPARE_STATE0, 3);
2480 r600_store_value(cb, 0); /* R_028AC0_DB_SRESULTS_COMPARE_STATE0 */
2481 r600_store_value(cb, 0); /* R_028AC4_DB_SRESULTS_COMPARE_STATE1 */
2482 r600_store_value(cb, 0); /* R_028AC8_DB_PRELOAD_CONTROL */
2483
2484 r600_store_context_reg(cb, R_028B70_DB_ALPHA_TO_MASK, 0x0000AA00);
2485
2486 r600_store_context_reg_seq(cb, R_028C00_PA_SC_LINE_CNTL, 2);
2487 r600_store_value(cb, 0x00000400); /* R_028C00_PA_SC_LINE_CNTL */
2488 r600_store_value(cb, 0); /* R_028C04_PA_SC_AA_CONFIG */
2489
2490 r600_store_context_reg_seq(cb, R_028C0C_PA_CL_GB_VERT_CLIP_ADJ, 5);
2491 r600_store_value(cb, 0x3F800000); /* R_028C0C_PA_CL_GB_VERT_CLIP_ADJ */
2492 r600_store_value(cb, 0x3F800000); /* R_028C10_PA_CL_GB_VERT_DISC_ADJ */
2493 r600_store_value(cb, 0x3F800000); /* R_028C14_PA_CL_GB_HORZ_CLIP_ADJ */
2494 r600_store_value(cb, 0x3F800000); /* R_028C18_PA_CL_GB_HORZ_DISC_ADJ */
2495 r600_store_value(cb, 0); /* R_028C1C_PA_SC_AA_SAMPLE_LOCS_0 */
2496
2497 r600_store_context_reg(cb, R_028C3C_PA_SC_AA_MASK, ~0);
2498
2499 r600_store_context_reg_seq(cb, R_028240_PA_SC_GENERIC_SCISSOR_TL, 2);
2500 r600_store_value(cb, 0); /* R_028240_PA_SC_GENERIC_SCISSOR_TL */
2501 r600_store_value(cb, S_028244_BR_X(16384) | S_028244_BR_Y(16384)); /* R_028244_PA_SC_GENERIC_SCISSOR_BR */
2502
2503 r600_store_context_reg_seq(cb, R_028030_PA_SC_SCREEN_SCISSOR_TL, 2);
2504 r600_store_value(cb, 0); /* R_028030_PA_SC_SCREEN_SCISSOR_TL */
2505 r600_store_value(cb, S_028034_BR_X(16384) | S_028034_BR_Y(16384)); /* R_028034_PA_SC_SCREEN_SCISSOR_BR */
2506
2507 r600_store_context_reg(cb, R_028848_SQ_PGM_RESOURCES_2_PS, S_028848_SINGLE_ROUND(V_SQ_ROUND_TO_ZERO));
2508 r600_store_context_reg(cb, R_028864_SQ_PGM_RESOURCES_2_VS, S_028864_SINGLE_ROUND(V_SQ_ROUND_TO_ZERO));
2509 r600_store_context_reg(cb, R_0288A8_SQ_PGM_RESOURCES_FS, 0);
2510
2511 r600_store_context_reg(cb, R_028354_SX_SURFACE_SYNC, S_028354_SURFACE_SYNC_MASK(0xf));
2512 r600_store_context_reg(cb, R_028800_DB_DEPTH_CONTROL, 0);
2513
2514 eg_store_loop_const(cb, R_03A200_SQ_LOOP_CONST_0, 0x01000FFF);
2515 eg_store_loop_const(cb, R_03A200_SQ_LOOP_CONST_0 + (32 * 4), 0x01000FFF);
2516 }
2517
2518 void evergreen_polygon_offset_update(struct r600_context *rctx)
2519 {
2520 struct r600_pipe_state state;
2521
2522 state.id = R600_PIPE_STATE_POLYGON_OFFSET;
2523 state.nregs = 0;
2524 if (rctx->rasterizer && rctx->framebuffer.zsbuf) {
2525 float offset_units = rctx->rasterizer->offset_units;
2526 unsigned offset_db_fmt_cntl = 0, depth;
2527
2528 switch (rctx->framebuffer.zsbuf->texture->format) {
2529 case PIPE_FORMAT_Z24X8_UNORM:
2530 case PIPE_FORMAT_Z24_UNORM_S8_UINT:
2531 depth = -24;
2532 offset_units *= 2.0f;
2533 break;
2534 case PIPE_FORMAT_Z32_FLOAT:
2535 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
2536 depth = -23;
2537 offset_units *= 1.0f;
2538 offset_db_fmt_cntl |= S_028B78_POLY_OFFSET_DB_IS_FLOAT_FMT(1);
2539 break;
2540 case PIPE_FORMAT_Z16_UNORM:
2541 depth = -16;
2542 offset_units *= 4.0f;
2543 break;
2544 default:
2545 return;
2546 }
2547 /* XXX some of those reg can be computed with cso */
2548 offset_db_fmt_cntl |= S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(depth);
2549 r600_pipe_state_add_reg(&state,
2550 R_028B80_PA_SU_POLY_OFFSET_FRONT_SCALE,
2551 fui(rctx->rasterizer->offset_scale));
2552 r600_pipe_state_add_reg(&state,
2553 R_028B84_PA_SU_POLY_OFFSET_FRONT_OFFSET,
2554 fui(offset_units));
2555 r600_pipe_state_add_reg(&state,
2556 R_028B88_PA_SU_POLY_OFFSET_BACK_SCALE,
2557 fui(rctx->rasterizer->offset_scale));
2558 r600_pipe_state_add_reg(&state,
2559 R_028B8C_PA_SU_POLY_OFFSET_BACK_OFFSET,
2560 fui(offset_units));
2561 r600_pipe_state_add_reg(&state,
2562 R_028B78_PA_SU_POLY_OFFSET_DB_FMT_CNTL,
2563 offset_db_fmt_cntl);
2564 r600_context_pipe_state_set(rctx, &state);
2565 }
2566 }
2567
2568 void evergreen_pipe_shader_ps(struct pipe_context *ctx, struct r600_pipe_shader *shader)
2569 {
2570 struct r600_context *rctx = (struct r600_context *)ctx;
2571 struct r600_pipe_state *rstate = &shader->rstate;
2572 struct r600_shader *rshader = &shader->shader;
2573 unsigned i, exports_ps, num_cout, spi_ps_in_control_0, spi_input_z, spi_ps_in_control_1, db_shader_control;
2574 int pos_index = -1, face_index = -1;
2575 int ninterp = 0;
2576 boolean have_linear = FALSE, have_centroid = FALSE, have_perspective = FALSE;
2577 unsigned spi_baryc_cntl, sid, tmp, idx = 0;
2578
2579 rstate->nregs = 0;
2580
2581 db_shader_control = S_02880C_Z_ORDER(V_02880C_EARLY_Z_THEN_LATE_Z);
2582 for (i = 0; i < rshader->ninput; i++) {
2583 /* evergreen NUM_INTERP only contains values interpolated into the LDS,
2584 POSITION goes via GPRs from the SC so isn't counted */
2585 if (rshader->input[i].name == TGSI_SEMANTIC_POSITION)
2586 pos_index = i;
2587 else if (rshader->input[i].name == TGSI_SEMANTIC_FACE)
2588 face_index = i;
2589 else {
2590 ninterp++;
2591 if (rshader->input[i].interpolate == TGSI_INTERPOLATE_LINEAR)
2592 have_linear = TRUE;
2593 if (rshader->input[i].interpolate == TGSI_INTERPOLATE_PERSPECTIVE)
2594 have_perspective = TRUE;
2595 if (rshader->input[i].centroid)
2596 have_centroid = TRUE;
2597 }
2598
2599 sid = rshader->input[i].spi_sid;
2600
2601 if (sid) {
2602
2603 tmp = S_028644_SEMANTIC(sid);
2604
2605 if (rshader->input[i].name == TGSI_SEMANTIC_POSITION ||
2606 rshader->input[i].interpolate == TGSI_INTERPOLATE_CONSTANT ||
2607 (rshader->input[i].interpolate == TGSI_INTERPOLATE_COLOR &&
2608 rctx->rasterizer && rctx->rasterizer->flatshade)) {
2609 tmp |= S_028644_FLAT_SHADE(1);
2610 }
2611
2612 if (rshader->input[i].name == TGSI_SEMANTIC_GENERIC &&
2613 (rctx->sprite_coord_enable & (1 << rshader->input[i].sid))) {
2614 tmp |= S_028644_PT_SPRITE_TEX(1);
2615 }
2616
2617 r600_pipe_state_add_reg(rstate, R_028644_SPI_PS_INPUT_CNTL_0 + idx * 4,
2618 tmp);
2619
2620 idx++;
2621 }
2622 }
2623
2624 for (i = 0; i < rshader->noutput; i++) {
2625 if (rshader->output[i].name == TGSI_SEMANTIC_POSITION)
2626 db_shader_control |= S_02880C_Z_EXPORT_ENABLE(1);
2627 if (rshader->output[i].name == TGSI_SEMANTIC_STENCIL)
2628 db_shader_control |= S_02880C_STENCIL_EXPORT_ENABLE(1);
2629 }
2630 if (rshader->uses_kill)
2631 db_shader_control |= S_02880C_KILL_ENABLE(1);
2632
2633 exports_ps = 0;
2634 num_cout = 0;
2635 for (i = 0; i < rshader->noutput; i++) {
2636 if (rshader->output[i].name == TGSI_SEMANTIC_POSITION ||
2637 rshader->output[i].name == TGSI_SEMANTIC_STENCIL)
2638 exports_ps |= 1;
2639 else if (rshader->output[i].name == TGSI_SEMANTIC_COLOR) {
2640 if (rshader->fs_write_all)
2641 num_cout = rshader->nr_cbufs;
2642 else
2643 num_cout++;
2644 }
2645 }
2646 exports_ps |= S_02884C_EXPORT_COLORS(num_cout);
2647 if (!exports_ps) {
2648 /* always at least export 1 component per pixel */
2649 exports_ps = 2;
2650 }
2651 shader->ps_cb_shader_mask = (1ULL << ((unsigned)num_cout * 4)) - 1;
2652 if (ninterp == 0) {
2653 ninterp = 1;
2654 have_perspective = TRUE;
2655 }
2656
2657 if (!have_perspective && !have_linear)
2658 have_perspective = TRUE;
2659
2660 spi_ps_in_control_0 = S_0286CC_NUM_INTERP(ninterp) |
2661 S_0286CC_PERSP_GRADIENT_ENA(have_perspective) |
2662 S_0286CC_LINEAR_GRADIENT_ENA(have_linear);
2663 spi_input_z = 0;
2664 if (pos_index != -1) {
2665 spi_ps_in_control_0 |= S_0286CC_POSITION_ENA(1) |
2666 S_0286CC_POSITION_CENTROID(rshader->input[pos_index].centroid) |
2667 S_0286CC_POSITION_ADDR(rshader->input[pos_index].gpr);
2668 spi_input_z |= 1;
2669 }
2670
2671 spi_ps_in_control_1 = 0;
2672 if (face_index != -1) {
2673 spi_ps_in_control_1 |= S_0286D0_FRONT_FACE_ENA(1) |
2674 S_0286D0_FRONT_FACE_ADDR(rshader->input[face_index].gpr);
2675 }
2676
2677 spi_baryc_cntl = 0;
2678 if (have_perspective)
2679 spi_baryc_cntl |= S_0286E0_PERSP_CENTER_ENA(1) |
2680 S_0286E0_PERSP_CENTROID_ENA(have_centroid);
2681 if (have_linear)
2682 spi_baryc_cntl |= S_0286E0_LINEAR_CENTER_ENA(1) |
2683 S_0286E0_LINEAR_CENTROID_ENA(have_centroid);
2684
2685 r600_pipe_state_add_reg(rstate, R_0286CC_SPI_PS_IN_CONTROL_0,
2686 spi_ps_in_control_0);
2687 r600_pipe_state_add_reg(rstate, R_0286D0_SPI_PS_IN_CONTROL_1,
2688 spi_ps_in_control_1);
2689 r600_pipe_state_add_reg(rstate, R_0286E4_SPI_PS_IN_CONTROL_2,
2690 0);
2691 r600_pipe_state_add_reg(rstate, R_0286D8_SPI_INPUT_Z, spi_input_z);
2692 r600_pipe_state_add_reg(rstate,
2693 R_0286E0_SPI_BARYC_CNTL,
2694 spi_baryc_cntl);
2695
2696 r600_pipe_state_add_reg_bo(rstate,
2697 R_028840_SQ_PGM_START_PS,
2698 r600_resource_va(ctx->screen, (void *)shader->bo) >> 8,
2699 shader->bo, RADEON_USAGE_READ);
2700 r600_pipe_state_add_reg(rstate,
2701 R_028844_SQ_PGM_RESOURCES_PS,
2702 S_028844_NUM_GPRS(rshader->bc.ngpr) |
2703 S_028844_PRIME_CACHE_ON_DRAW(1) |
2704 S_028844_STACK_SIZE(rshader->bc.nstack));
2705 r600_pipe_state_add_reg(rstate,
2706 R_02884C_SQ_PGM_EXPORTS_PS,
2707 exports_ps);
2708 r600_pipe_state_add_reg(rstate, R_02880C_DB_SHADER_CONTROL,
2709 db_shader_control);
2710
2711 shader->sprite_coord_enable = rctx->sprite_coord_enable;
2712 if (rctx->rasterizer)
2713 shader->flatshade = rctx->rasterizer->flatshade;
2714 }
2715
2716 void evergreen_pipe_shader_vs(struct pipe_context *ctx, struct r600_pipe_shader *shader)
2717 {
2718 struct r600_context *rctx = (struct r600_context *)ctx;
2719 struct r600_pipe_state *rstate = &shader->rstate;
2720 struct r600_shader *rshader = &shader->shader;
2721 unsigned spi_vs_out_id[10] = {};
2722 unsigned i, tmp, nparams = 0;
2723
2724 /* clear previous register */
2725 rstate->nregs = 0;
2726
2727 for (i = 0; i < rshader->noutput; i++) {
2728 if (rshader->output[i].spi_sid) {
2729 tmp = rshader->output[i].spi_sid << ((nparams & 3) * 8);
2730 spi_vs_out_id[nparams / 4] |= tmp;
2731 nparams++;
2732 }
2733 }
2734
2735 for (i = 0; i < 10; i++) {
2736 r600_pipe_state_add_reg(rstate,
2737 R_02861C_SPI_VS_OUT_ID_0 + i * 4,
2738 spi_vs_out_id[i]);
2739 }
2740
2741 /* Certain attributes (position, psize, etc.) don't count as params.
2742 * VS is required to export at least one param and r600_shader_from_tgsi()
2743 * takes care of adding a dummy export.
2744 */
2745 if (nparams < 1)
2746 nparams = 1;
2747
2748 r600_pipe_state_add_reg(rstate,
2749 R_0286C4_SPI_VS_OUT_CONFIG,
2750 S_0286C4_VS_EXPORT_COUNT(nparams - 1));
2751 r600_pipe_state_add_reg(rstate,
2752 R_028860_SQ_PGM_RESOURCES_VS,
2753 S_028860_NUM_GPRS(rshader->bc.ngpr) |
2754 S_028860_STACK_SIZE(rshader->bc.nstack));
2755 r600_pipe_state_add_reg_bo(rstate,
2756 R_02885C_SQ_PGM_START_VS,
2757 r600_resource_va(ctx->screen, (void *)shader->bo) >> 8,
2758 shader->bo, RADEON_USAGE_READ);
2759
2760 shader->pa_cl_vs_out_cntl =
2761 S_02881C_VS_OUT_CCDIST0_VEC_ENA((rshader->clip_dist_write & 0x0F) != 0) |
2762 S_02881C_VS_OUT_CCDIST1_VEC_ENA((rshader->clip_dist_write & 0xF0) != 0) |
2763 S_02881C_VS_OUT_MISC_VEC_ENA(rshader->vs_out_misc_write) |
2764 S_02881C_USE_VTX_POINT_SIZE(rshader->vs_out_point_size);
2765 }
2766
2767 void evergreen_fetch_shader(struct pipe_context *ctx,
2768 struct r600_vertex_element *ve)
2769 {
2770 struct r600_context *rctx = (struct r600_context *)ctx;
2771 struct r600_pipe_state *rstate = &ve->rstate;
2772 rstate->id = R600_PIPE_STATE_FETCH_SHADER;
2773 rstate->nregs = 0;
2774 r600_pipe_state_add_reg_bo(rstate, R_0288A4_SQ_PGM_START_FS,
2775 r600_resource_va(ctx->screen, (void *)ve->fetch_shader) >> 8,
2776 ve->fetch_shader, RADEON_USAGE_READ);
2777 }
2778
2779 void *evergreen_create_db_flush_dsa(struct r600_context *rctx)
2780 {
2781 struct pipe_depth_stencil_alpha_state dsa;
2782 struct r600_pipe_state *rstate;
2783
2784 memset(&dsa, 0, sizeof(dsa));
2785
2786 rstate = rctx->context.create_depth_stencil_alpha_state(&rctx->context, &dsa);
2787 r600_pipe_state_add_reg(rstate,
2788 R_028000_DB_RENDER_CONTROL,
2789 S_028000_DEPTH_COPY_ENABLE(1) |
2790 S_028000_STENCIL_COPY_ENABLE(1) |
2791 S_028000_COPY_CENTROID(1));
2792 /* Don't set the 'is_flush' flag in r600_pipe_dsa, evergreen doesn't need it. */
2793 return rstate;
2794 }