projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
radeonsi: Implement POLYGON_OFFSET_UNITS_UNSCALED
[mesa.git] / src / gallium / drivers / radeonsi / si_state.c
1 /*
2 * Copyright 2012 Advanced Micro Devices, Inc.
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 * Authors:
24 * Christian König <christian.koenig@amd.com>
25 */
26
27 #include "si_pipe.h"
28 #include "si_shader.h"
29 #include "sid.h"
30 #include "radeon/r600_cs.h"
31
32 #include "util/u_dual_blend.h"
33 #include "util/u_format.h"
34 #include "util/u_format_s3tc.h"
35 #include "util/u_memory.h"
36 #include "util/u_pstipple.h"
37 #include "util/u_resource.h"
38
39 /* Initialize an external atom (owned by ../radeon). */
40 static void
41 si_init_external_atom(struct si_context *sctx, struct r600_atom *atom,
42 struct r600_atom **list_elem)
43 {
44 atom->id = list_elem - sctx->atoms.array + 1;
45 *list_elem = atom;
46 }
47
48 /* Initialize an atom owned by radeonsi. */
49 void si_init_atom(struct si_context *sctx, struct r600_atom *atom,
50 struct r600_atom **list_elem,
51 void (*emit_func)(struct si_context *ctx, struct r600_atom *state))
52 {
53 atom->emit = (void*)emit_func;
54 atom->id = list_elem - sctx->atoms.array + 1; /* index+1 in the atom array */
55 *list_elem = atom;
56 }
57
58 static unsigned si_map_swizzle(unsigned swizzle)
59 {
60 switch (swizzle) {
61 case PIPE_SWIZZLE_Y:
62 return V_008F0C_SQ_SEL_Y;
63 case PIPE_SWIZZLE_Z:
64 return V_008F0C_SQ_SEL_Z;
65 case PIPE_SWIZZLE_W:
66 return V_008F0C_SQ_SEL_W;
67 case PIPE_SWIZZLE_0:
68 return V_008F0C_SQ_SEL_0;
69 case PIPE_SWIZZLE_1:
70 return V_008F0C_SQ_SEL_1;
71 default: /* PIPE_SWIZZLE_X */
72 return V_008F0C_SQ_SEL_X;
73 }
74 }
75
76 static uint32_t S_FIXED(float value, uint32_t frac_bits)
77 {
78 return value * (1 << frac_bits);
79 }
80
81 /* 12.4 fixed-point */
82 static unsigned si_pack_float_12p4(float x)
83 {
84 return x <= 0 ? 0 :
85 x >= 4096 ? 0xffff : x * 16;
86 }
87
88 /*
89 * Inferred framebuffer and blender state.
90 *
91 * One of the reasons CB_TARGET_MASK must be derived from the framebuffer state
92 * is that:
93 * - The blend state mask is 0xf most of the time.
94 * - The COLOR1 format isn't INVALID because of possible dual-source blending,
95 * so COLOR1 is enabled pretty much all the time.
96 * So CB_TARGET_MASK is the only register that can disable COLOR1.
97 *
98 * Another reason is to avoid a hang with dual source blending.
99 */
100 static void si_emit_cb_render_state(struct si_context *sctx, struct r600_atom *atom)
101 {
102 struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
103 struct si_state_blend *blend = sctx->queued.named.blend;
104 uint32_t cb_target_mask = 0, i;
105
106 for (i = 0; i < sctx->framebuffer.state.nr_cbufs; i++)
107 if (sctx->framebuffer.state.cbufs[i])
108 cb_target_mask |= 0xf << (4*i);
109
110 if (blend)
111 cb_target_mask &= blend->cb_target_mask;
112
113 /* Avoid a hang that happens when dual source blending is enabled
114 * but there is not enough color outputs. This is undefined behavior,
115 * so disable color writes completely.
116 *
117 * Reproducible with Unigine Heaven 4.0 and drirc missing.
118 */
119 if (blend && blend->dual_src_blend &&
120 sctx->ps_shader.cso &&
121 (sctx->ps_shader.cso->info.colors_written & 0x3) != 0x3)
122 cb_target_mask = 0;
123
124 radeon_set_context_reg(cs, R_028238_CB_TARGET_MASK, cb_target_mask);
125
126 /* STONEY-specific register settings. */
127 if (sctx->b.family == CHIP_STONEY) {
128 unsigned spi_shader_col_format =
129 sctx->ps_shader.cso ?
130 sctx->ps_shader.current->key.ps.epilog.spi_shader_col_format : 0;
131 unsigned sx_ps_downconvert = 0;
132 unsigned sx_blend_opt_epsilon = 0;
133 unsigned sx_blend_opt_control = 0;
134
135 for (i = 0; i < sctx->framebuffer.state.nr_cbufs; i++) {
136 struct r600_surface *surf =
137 (struct r600_surface*)sctx->framebuffer.state.cbufs[i];
138 unsigned format, swap, spi_format, colormask;
139 bool has_alpha, has_rgb;
140
141 if (!surf)
142 continue;
143
144 format = G_028C70_FORMAT(surf->cb_color_info);
145 swap = G_028C70_COMP_SWAP(surf->cb_color_info);
146 spi_format = (spi_shader_col_format >> (i * 4)) & 0xf;
147 colormask = (cb_target_mask >> (i * 4)) & 0xf;
148
149 /* Set if RGB and A are present. */
150 has_alpha = !G_028C74_FORCE_DST_ALPHA_1(surf->cb_color_attrib);
151
152 if (format == V_028C70_COLOR_8 ||
153 format == V_028C70_COLOR_16 ||
154 format == V_028C70_COLOR_32)
155 has_rgb = !has_alpha;
156 else
157 has_rgb = true;
158
159 /* Check the colormask and export format. */
160 if (!(colormask & (PIPE_MASK_RGBA & ~PIPE_MASK_A)))
161 has_rgb = false;
162 if (!(colormask & PIPE_MASK_A))
163 has_alpha = false;
164
165 if (spi_format == V_028714_SPI_SHADER_ZERO) {
166 has_rgb = false;
167 has_alpha = false;
168 }
169
170 /* Disable value checking for disabled channels. */
171 if (!has_rgb)
172 sx_blend_opt_control |= S_02875C_MRT0_COLOR_OPT_DISABLE(1) << (i * 4);
173 if (!has_alpha)
174 sx_blend_opt_control |= S_02875C_MRT0_ALPHA_OPT_DISABLE(1) << (i * 4);
175
176 /* Enable down-conversion for 32bpp and smaller formats. */
177 switch (format) {
178 case V_028C70_COLOR_8:
179 case V_028C70_COLOR_8_8:
180 case V_028C70_COLOR_8_8_8_8:
181 /* For 1 and 2-channel formats, use the superset thereof. */
182 if (spi_format == V_028714_SPI_SHADER_FP16_ABGR ||
183 spi_format == V_028714_SPI_SHADER_UINT16_ABGR ||
184 spi_format == V_028714_SPI_SHADER_SINT16_ABGR) {
185 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_8_8_8_8 << (i * 4);
186 sx_blend_opt_epsilon |= V_028758_8BIT_FORMAT << (i * 4);
187 }
188 break;
189
190 case V_028C70_COLOR_5_6_5:
191 if (spi_format == V_028714_SPI_SHADER_FP16_ABGR) {
192 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_5_6_5 << (i * 4);
193 sx_blend_opt_epsilon |= V_028758_6BIT_FORMAT << (i * 4);
194 }
195 break;
196
197 case V_028C70_COLOR_1_5_5_5:
198 if (spi_format == V_028714_SPI_SHADER_FP16_ABGR) {
199 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_1_5_5_5 << (i * 4);
200 sx_blend_opt_epsilon |= V_028758_5BIT_FORMAT << (i * 4);
201 }
202 break;
203
204 case V_028C70_COLOR_4_4_4_4:
205 if (spi_format == V_028714_SPI_SHADER_FP16_ABGR) {
206 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_4_4_4_4 << (i * 4);
207 sx_blend_opt_epsilon |= V_028758_4BIT_FORMAT << (i * 4);
208 }
209 break;
210
211 case V_028C70_COLOR_32:
212 if (swap == V_0280A0_SWAP_STD &&
213 spi_format == V_028714_SPI_SHADER_32_R)
214 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_32_R << (i * 4);
215 else if (swap == V_0280A0_SWAP_ALT_REV &&
216 spi_format == V_028714_SPI_SHADER_32_AR)
217 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_32_A << (i * 4);
218 break;
219
220 case V_028C70_COLOR_16:
221 case V_028C70_COLOR_16_16:
222 /* For 1-channel formats, use the superset thereof. */
223 if (spi_format == V_028714_SPI_SHADER_UNORM16_ABGR ||
224 spi_format == V_028714_SPI_SHADER_SNORM16_ABGR ||
225 spi_format == V_028714_SPI_SHADER_UINT16_ABGR ||
226 spi_format == V_028714_SPI_SHADER_SINT16_ABGR) {
227 if (swap == V_0280A0_SWAP_STD ||
228 swap == V_0280A0_SWAP_STD_REV)
229 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_16_16_GR << (i * 4);
230 else
231 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_16_16_AR << (i * 4);
232 }
233 break;
234
235 case V_028C70_COLOR_10_11_11:
236 if (spi_format == V_028714_SPI_SHADER_FP16_ABGR) {
237 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_10_11_11 << (i * 4);
238 sx_blend_opt_epsilon |= V_028758_11BIT_FORMAT << (i * 4);
239 }
240 break;
241
242 case V_028C70_COLOR_2_10_10_10:
243 if (spi_format == V_028714_SPI_SHADER_FP16_ABGR) {
244 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_2_10_10_10 << (i * 4);
245 sx_blend_opt_epsilon |= V_028758_10BIT_FORMAT << (i * 4);
246 }
247 break;
248 }
249 }
250
251 if (sctx->screen->b.debug_flags & DBG_NO_RB_PLUS) {
252 sx_ps_downconvert = 0;
253 sx_blend_opt_epsilon = 0;
254 sx_blend_opt_control = 0;
255 }
256
257 radeon_set_context_reg_seq(cs, R_028754_SX_PS_DOWNCONVERT, 3);
258 radeon_emit(cs, sx_ps_downconvert); /* R_028754_SX_PS_DOWNCONVERT */
259 radeon_emit(cs, sx_blend_opt_epsilon); /* R_028758_SX_BLEND_OPT_EPSILON */
260 radeon_emit(cs, sx_blend_opt_control); /* R_02875C_SX_BLEND_OPT_CONTROL */
261 }
262 }
263
264 /*
265 * Blender functions
266 */
267
268 static uint32_t si_translate_blend_function(int blend_func)
269 {
270 switch (blend_func) {
271 case PIPE_BLEND_ADD:
272 return V_028780_COMB_DST_PLUS_SRC;
273 case PIPE_BLEND_SUBTRACT:
274 return V_028780_COMB_SRC_MINUS_DST;
275 case PIPE_BLEND_REVERSE_SUBTRACT:
276 return V_028780_COMB_DST_MINUS_SRC;
277 case PIPE_BLEND_MIN:
278 return V_028780_COMB_MIN_DST_SRC;
279 case PIPE_BLEND_MAX:
280 return V_028780_COMB_MAX_DST_SRC;
281 default:
282 R600_ERR("Unknown blend function %d\n", blend_func);
283 assert(0);
284 break;
285 }
286 return 0;
287 }
288
289 static uint32_t si_translate_blend_factor(int blend_fact)
290 {
291 switch (blend_fact) {
292 case PIPE_BLENDFACTOR_ONE:
293 return V_028780_BLEND_ONE;
294 case PIPE_BLENDFACTOR_SRC_COLOR:
295 return V_028780_BLEND_SRC_COLOR;
296 case PIPE_BLENDFACTOR_SRC_ALPHA:
297 return V_028780_BLEND_SRC_ALPHA;
298 case PIPE_BLENDFACTOR_DST_ALPHA:
299 return V_028780_BLEND_DST_ALPHA;
300 case PIPE_BLENDFACTOR_DST_COLOR:
301 return V_028780_BLEND_DST_COLOR;
302 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
303 return V_028780_BLEND_SRC_ALPHA_SATURATE;
304 case PIPE_BLENDFACTOR_CONST_COLOR:
305 return V_028780_BLEND_CONSTANT_COLOR;
306 case PIPE_BLENDFACTOR_CONST_ALPHA:
307 return V_028780_BLEND_CONSTANT_ALPHA;
308 case PIPE_BLENDFACTOR_ZERO:
309 return V_028780_BLEND_ZERO;
310 case PIPE_BLENDFACTOR_INV_SRC_COLOR:
311 return V_028780_BLEND_ONE_MINUS_SRC_COLOR;
312 case PIPE_BLENDFACTOR_INV_SRC_ALPHA:
313 return V_028780_BLEND_ONE_MINUS_SRC_ALPHA;
314 case PIPE_BLENDFACTOR_INV_DST_ALPHA:
315 return V_028780_BLEND_ONE_MINUS_DST_ALPHA;
316 case PIPE_BLENDFACTOR_INV_DST_COLOR:
317 return V_028780_BLEND_ONE_MINUS_DST_COLOR;
318 case PIPE_BLENDFACTOR_INV_CONST_COLOR:
319 return V_028780_BLEND_ONE_MINUS_CONSTANT_COLOR;
320 case PIPE_BLENDFACTOR_INV_CONST_ALPHA:
321 return V_028780_BLEND_ONE_MINUS_CONSTANT_ALPHA;
322 case PIPE_BLENDFACTOR_SRC1_COLOR:
323 return V_028780_BLEND_SRC1_COLOR;
324 case PIPE_BLENDFACTOR_SRC1_ALPHA:
325 return V_028780_BLEND_SRC1_ALPHA;
326 case PIPE_BLENDFACTOR_INV_SRC1_COLOR:
327 return V_028780_BLEND_INV_SRC1_COLOR;
328 case PIPE_BLENDFACTOR_INV_SRC1_ALPHA:
329 return V_028780_BLEND_INV_SRC1_ALPHA;
330 default:
331 R600_ERR("Bad blend factor %d not supported!\n", blend_fact);
332 assert(0);
333 break;
334 }
335 return 0;
336 }
337
338 static uint32_t si_translate_blend_opt_function(int blend_func)
339 {
340 switch (blend_func) {
341 case PIPE_BLEND_ADD:
342 return V_028760_OPT_COMB_ADD;
343 case PIPE_BLEND_SUBTRACT:
344 return V_028760_OPT_COMB_SUBTRACT;
345 case PIPE_BLEND_REVERSE_SUBTRACT:
346 return V_028760_OPT_COMB_REVSUBTRACT;
347 case PIPE_BLEND_MIN:
348 return V_028760_OPT_COMB_MIN;
349 case PIPE_BLEND_MAX:
350 return V_028760_OPT_COMB_MAX;
351 default:
352 return V_028760_OPT_COMB_BLEND_DISABLED;
353 }
354 }
355
356 static uint32_t si_translate_blend_opt_factor(int blend_fact, bool is_alpha)
357 {
358 switch (blend_fact) {
359 case PIPE_BLENDFACTOR_ZERO:
360 return V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_ALL;
361 case PIPE_BLENDFACTOR_ONE:
362 return V_028760_BLEND_OPT_PRESERVE_ALL_IGNORE_NONE;
363 case PIPE_BLENDFACTOR_SRC_COLOR:
364 return is_alpha ? V_028760_BLEND_OPT_PRESERVE_A1_IGNORE_A0
365 : V_028760_BLEND_OPT_PRESERVE_C1_IGNORE_C0;
366 case PIPE_BLENDFACTOR_INV_SRC_COLOR:
367 return is_alpha ? V_028760_BLEND_OPT_PRESERVE_A0_IGNORE_A1
368 : V_028760_BLEND_OPT_PRESERVE_C0_IGNORE_C1;
369 case PIPE_BLENDFACTOR_SRC_ALPHA:
370 return V_028760_BLEND_OPT_PRESERVE_A1_IGNORE_A0;
371 case PIPE_BLENDFACTOR_INV_SRC_ALPHA:
372 return V_028760_BLEND_OPT_PRESERVE_A0_IGNORE_A1;
373 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
374 return is_alpha ? V_028760_BLEND_OPT_PRESERVE_ALL_IGNORE_NONE
375 : V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_A0;
376 default:
377 return V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE;
378 }
379 }
380
381 /**
382 * Get rid of DST in the blend factors by commuting the operands:
383 * func(src * DST, dst * 0) ---> func(src * 0, dst * SRC)
384 */
385 static void si_blend_remove_dst(unsigned *func, unsigned *src_factor,
386 unsigned *dst_factor, unsigned expected_dst,
387 unsigned replacement_src)
388 {
389 if (*src_factor == expected_dst &&
390 *dst_factor == PIPE_BLENDFACTOR_ZERO) {
391 *src_factor = PIPE_BLENDFACTOR_ZERO;
392 *dst_factor = replacement_src;
393
394 /* Commuting the operands requires reversing subtractions. */
395 if (*func == PIPE_BLEND_SUBTRACT)
396 *func = PIPE_BLEND_REVERSE_SUBTRACT;
397 else if (*func == PIPE_BLEND_REVERSE_SUBTRACT)
398 *func = PIPE_BLEND_SUBTRACT;
399 }
400 }
401
402 static bool si_blend_factor_uses_dst(unsigned factor)
403 {
404 return factor == PIPE_BLENDFACTOR_DST_COLOR ||
405 factor == PIPE_BLENDFACTOR_DST_ALPHA ||
406 factor == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE ||
407 factor == PIPE_BLENDFACTOR_INV_DST_ALPHA ||
408 factor == PIPE_BLENDFACTOR_INV_DST_COLOR;
409 }
410
411 static void *si_create_blend_state_mode(struct pipe_context *ctx,
412 const struct pipe_blend_state *state,
413 unsigned mode)
414 {
415 struct si_context *sctx = (struct si_context*)ctx;
416 struct si_state_blend *blend = CALLOC_STRUCT(si_state_blend);
417 struct si_pm4_state *pm4 = &blend->pm4;
418 uint32_t sx_mrt_blend_opt[8] = {0};
419 uint32_t color_control = 0;
420
421 if (!blend)
422 return NULL;
423
424 blend->alpha_to_coverage = state->alpha_to_coverage;
425 blend->alpha_to_one = state->alpha_to_one;
426 blend->dual_src_blend = util_blend_state_is_dual(state, 0);
427
428 if (state->logicop_enable) {
429 color_control |= S_028808_ROP3(state->logicop_func | (state->logicop_func << 4));
430 } else {
431 color_control |= S_028808_ROP3(0xcc);
432 }
433
434 si_pm4_set_reg(pm4, R_028B70_DB_ALPHA_TO_MASK,
435 S_028B70_ALPHA_TO_MASK_ENABLE(state->alpha_to_coverage) |
436 S_028B70_ALPHA_TO_MASK_OFFSET0(2) |
437 S_028B70_ALPHA_TO_MASK_OFFSET1(2) |
438 S_028B70_ALPHA_TO_MASK_OFFSET2(2) |
439 S_028B70_ALPHA_TO_MASK_OFFSET3(2));
440
441 if (state->alpha_to_coverage)
442 blend->need_src_alpha_4bit |= 0xf;
443
444 blend->cb_target_mask = 0;
445 for (int i = 0; i < 8; i++) {
446 /* state->rt entries > 0 only written if independent blending */
447 const int j = state->independent_blend_enable ? i : 0;
448
449 unsigned eqRGB = state->rt[j].rgb_func;
450 unsigned srcRGB = state->rt[j].rgb_src_factor;
451 unsigned dstRGB = state->rt[j].rgb_dst_factor;
452 unsigned eqA = state->rt[j].alpha_func;
453 unsigned srcA = state->rt[j].alpha_src_factor;
454 unsigned dstA = state->rt[j].alpha_dst_factor;
455
456 unsigned srcRGB_opt, dstRGB_opt, srcA_opt, dstA_opt;
457 unsigned blend_cntl = 0;
458
459 sx_mrt_blend_opt[i] =
460 S_028760_COLOR_COMB_FCN(V_028760_OPT_COMB_BLEND_DISABLED) |
461 S_028760_ALPHA_COMB_FCN(V_028760_OPT_COMB_BLEND_DISABLED);
462
463 if (!state->rt[j].colormask)
464 continue;
465
466 /* cb_render_state will disable unused ones */
467 blend->cb_target_mask |= (unsigned)state->rt[j].colormask << (4 * i);
468
469 if (!state->rt[j].blend_enable) {
470 si_pm4_set_reg(pm4, R_028780_CB_BLEND0_CONTROL + i * 4, blend_cntl);
471 continue;
472 }
473
474 /* Blending optimizations for Stoney.
475 * These transformations don't change the behavior.
476 *
477 * First, get rid of DST in the blend factors:
478 * func(src * DST, dst * 0) ---> func(src * 0, dst * SRC)
479 */
480 si_blend_remove_dst(&eqRGB, &srcRGB, &dstRGB,
481 PIPE_BLENDFACTOR_DST_COLOR,
482 PIPE_BLENDFACTOR_SRC_COLOR);
483 si_blend_remove_dst(&eqA, &srcA, &dstA,
484 PIPE_BLENDFACTOR_DST_COLOR,
485 PIPE_BLENDFACTOR_SRC_COLOR);
486 si_blend_remove_dst(&eqA, &srcA, &dstA,
487 PIPE_BLENDFACTOR_DST_ALPHA,
488 PIPE_BLENDFACTOR_SRC_ALPHA);
489
490 /* Look up the ideal settings from tables. */
491 srcRGB_opt = si_translate_blend_opt_factor(srcRGB, false);
492 dstRGB_opt = si_translate_blend_opt_factor(dstRGB, false);
493 srcA_opt = si_translate_blend_opt_factor(srcA, true);
494 dstA_opt = si_translate_blend_opt_factor(dstA, true);
495
496 /* Handle interdependencies. */
497 if (si_blend_factor_uses_dst(srcRGB))
498 dstRGB_opt = V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE;
499 if (si_blend_factor_uses_dst(srcA))
500 dstA_opt = V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE;
501
502 if (srcRGB == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE &&
503 (dstRGB == PIPE_BLENDFACTOR_ZERO ||
504 dstRGB == PIPE_BLENDFACTOR_SRC_ALPHA ||
505 dstRGB == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE))
506 dstRGB_opt = V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_A0;
507
508 /* Set the final value. */
509 sx_mrt_blend_opt[i] =
510 S_028760_COLOR_SRC_OPT(srcRGB_opt) |
511 S_028760_COLOR_DST_OPT(dstRGB_opt) |
512 S_028760_COLOR_COMB_FCN(si_translate_blend_opt_function(eqRGB)) |
513 S_028760_ALPHA_SRC_OPT(srcA_opt) |
514 S_028760_ALPHA_DST_OPT(dstA_opt) |
515 S_028760_ALPHA_COMB_FCN(si_translate_blend_opt_function(eqA));
516
517 /* Set blend state. */
518 blend_cntl |= S_028780_ENABLE(1);
519 blend_cntl |= S_028780_COLOR_COMB_FCN(si_translate_blend_function(eqRGB));
520 blend_cntl |= S_028780_COLOR_SRCBLEND(si_translate_blend_factor(srcRGB));
521 blend_cntl |= S_028780_COLOR_DESTBLEND(si_translate_blend_factor(dstRGB));
522
523 if (srcA != srcRGB || dstA != dstRGB || eqA != eqRGB) {
524 blend_cntl |= S_028780_SEPARATE_ALPHA_BLEND(1);
525 blend_cntl |= S_028780_ALPHA_COMB_FCN(si_translate_blend_function(eqA));
526 blend_cntl |= S_028780_ALPHA_SRCBLEND(si_translate_blend_factor(srcA));
527 blend_cntl |= S_028780_ALPHA_DESTBLEND(si_translate_blend_factor(dstA));
528 }
529 si_pm4_set_reg(pm4, R_028780_CB_BLEND0_CONTROL + i * 4, blend_cntl);
530
531 blend->blend_enable_4bit |= 0xfu << (i * 4);
532
533 /* This is only important for formats without alpha. */
534 if (srcRGB == PIPE_BLENDFACTOR_SRC_ALPHA ||
535 dstRGB == PIPE_BLENDFACTOR_SRC_ALPHA ||
536 srcRGB == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE ||
537 dstRGB == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE ||
538 srcRGB == PIPE_BLENDFACTOR_INV_SRC_ALPHA ||
539 dstRGB == PIPE_BLENDFACTOR_INV_SRC_ALPHA)
540 blend->need_src_alpha_4bit |= 0xfu << (i * 4);
541 }
542
543 if (blend->cb_target_mask) {
544 color_control |= S_028808_MODE(mode);
545 } else {
546 color_control |= S_028808_MODE(V_028808_CB_DISABLE);
547 }
548
549 if (sctx->b.family == CHIP_STONEY) {
550 for (int i = 0; i < 8; i++)
551 si_pm4_set_reg(pm4, R_028760_SX_MRT0_BLEND_OPT + i * 4,
552 sx_mrt_blend_opt[i]);
553
554 /* RB+ doesn't work with dual source blending, logic op, and RESOLVE. */
555 if (blend->dual_src_blend || state->logicop_enable ||
556 mode == V_028808_CB_RESOLVE)
557 color_control |= S_028808_DISABLE_DUAL_QUAD(1);
558 }
559
560 si_pm4_set_reg(pm4, R_028808_CB_COLOR_CONTROL, color_control);
561 return blend;
562 }
563
564 static void *si_create_blend_state(struct pipe_context *ctx,
565 const struct pipe_blend_state *state)
566 {
567 return si_create_blend_state_mode(ctx, state, V_028808_CB_NORMAL);
568 }
569
570 static void si_bind_blend_state(struct pipe_context *ctx, void *state)
571 {
572 struct si_context *sctx = (struct si_context *)ctx;
573 si_pm4_bind_state(sctx, blend, (struct si_state_blend *)state);
574 si_mark_atom_dirty(sctx, &sctx->cb_render_state);
575 }
576
577 static void si_delete_blend_state(struct pipe_context *ctx, void *state)
578 {
579 struct si_context *sctx = (struct si_context *)ctx;
580 si_pm4_delete_state(sctx, blend, (struct si_state_blend *)state);
581 }
582
583 static void si_set_blend_color(struct pipe_context *ctx,
584 const struct pipe_blend_color *state)
585 {
586 struct si_context *sctx = (struct si_context *)ctx;
587
588 if (memcmp(&sctx->blend_color.state, state, sizeof(*state)) == 0)
589 return;
590
591 sctx->blend_color.state = *state;
592 si_mark_atom_dirty(sctx, &sctx->blend_color.atom);
593 }
594
595 static void si_emit_blend_color(struct si_context *sctx, struct r600_atom *atom)
596 {
597 struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
598
599 radeon_set_context_reg_seq(cs, R_028414_CB_BLEND_RED, 4);
600 radeon_emit_array(cs, (uint32_t*)sctx->blend_color.state.color, 4);
601 }
602
603 /*
604 * Clipping
605 */
606
607 static void si_set_clip_state(struct pipe_context *ctx,
608 const struct pipe_clip_state *state)
609 {
610 struct si_context *sctx = (struct si_context *)ctx;
611 struct pipe_constant_buffer cb;
612
613 if (memcmp(&sctx->clip_state.state, state, sizeof(*state)) == 0)
614 return;
615
616 sctx->clip_state.state = *state;
617 si_mark_atom_dirty(sctx, &sctx->clip_state.atom);
618
619 cb.buffer = NULL;
620 cb.user_buffer = state->ucp;
621 cb.buffer_offset = 0;
622 cb.buffer_size = 4*4*8;
623 si_set_rw_buffer(sctx, SI_VS_CONST_CLIP_PLANES, &cb);
624 pipe_resource_reference(&cb.buffer, NULL);
625 }
626
627 static void si_emit_clip_state(struct si_context *sctx, struct r600_atom *atom)
628 {
629 struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
630
631 radeon_set_context_reg_seq(cs, R_0285BC_PA_CL_UCP_0_X, 6*4);
632 radeon_emit_array(cs, (uint32_t*)sctx->clip_state.state.ucp, 6*4);
633 }
634
635 #define SIX_BITS 0x3F
636
637 static void si_emit_clip_regs(struct si_context *sctx, struct r600_atom *atom)
638 {
639 struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
640 struct tgsi_shader_info *info = si_get_vs_info(sctx);
641 unsigned window_space =
642 info->properties[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION];
643 unsigned clipdist_mask =
644 info->writes_clipvertex ? SIX_BITS : info->clipdist_writemask;
645
646 radeon_set_context_reg(cs, R_02881C_PA_CL_VS_OUT_CNTL,
647 S_02881C_USE_VTX_POINT_SIZE(info->writes_psize) |
648 S_02881C_USE_VTX_EDGE_FLAG(info->writes_edgeflag) |
649 S_02881C_USE_VTX_RENDER_TARGET_INDX(info->writes_layer) |
650 S_02881C_USE_VTX_VIEWPORT_INDX(info->writes_viewport_index) |
651 S_02881C_VS_OUT_CCDIST0_VEC_ENA((clipdist_mask & 0x0F) != 0) |
652 S_02881C_VS_OUT_CCDIST1_VEC_ENA((clipdist_mask & 0xF0) != 0) |
653 S_02881C_VS_OUT_MISC_VEC_ENA(info->writes_psize ||
654 info->writes_edgeflag ||
655 info->writes_layer ||
656 info->writes_viewport_index) |
657 S_02881C_VS_OUT_MISC_SIDE_BUS_ENA(1) |
658 (sctx->queued.named.rasterizer->clip_plane_enable &
659 clipdist_mask));
660 radeon_set_context_reg(cs, R_028810_PA_CL_CLIP_CNTL,
661 sctx->queued.named.rasterizer->pa_cl_clip_cntl |
662 (clipdist_mask ? 0 :
663 sctx->queued.named.rasterizer->clip_plane_enable & SIX_BITS) |
664 S_028810_CLIP_DISABLE(window_space));
665
666 /* reuse needs to be set off if we write oViewport */
667 radeon_set_context_reg(cs, R_028AB4_VGT_REUSE_OFF,
668 S_028AB4_REUSE_OFF(info->writes_viewport_index));
669 }
670
671 /*
672 * inferred state between framebuffer and rasterizer
673 */
674 static void si_update_poly_offset_state(struct si_context *sctx)
675 {
676 struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
677
678 if (!rs || !rs->uses_poly_offset || !sctx->framebuffer.state.zsbuf)
679 return;
680
681 switch (sctx->framebuffer.state.zsbuf->texture->format) {
682 case PIPE_FORMAT_Z16_UNORM:
683 si_pm4_bind_state(sctx, poly_offset, &rs->pm4_poly_offset[0]);
684 break;
685 default: /* 24-bit */
686 si_pm4_bind_state(sctx, poly_offset, &rs->pm4_poly_offset[1]);
687 break;
688 case PIPE_FORMAT_Z32_FLOAT:
689 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
690 si_pm4_bind_state(sctx, poly_offset, &rs->pm4_poly_offset[2]);
691 break;
692 }
693 }
694
695 /*
696 * Rasterizer
697 */
698
699 static uint32_t si_translate_fill(uint32_t func)
700 {
701 switch(func) {
702 case PIPE_POLYGON_MODE_FILL:
703 return V_028814_X_DRAW_TRIANGLES;
704 case PIPE_POLYGON_MODE_LINE:
705 return V_028814_X_DRAW_LINES;
706 case PIPE_POLYGON_MODE_POINT:
707 return V_028814_X_DRAW_POINTS;
708 default:
709 assert(0);
710 return V_028814_X_DRAW_POINTS;
711 }
712 }
713
714 static void *si_create_rs_state(struct pipe_context *ctx,
715 const struct pipe_rasterizer_state *state)
716 {
717 struct si_state_rasterizer *rs = CALLOC_STRUCT(si_state_rasterizer);
718 struct si_pm4_state *pm4 = &rs->pm4;
719 unsigned tmp, i;
720 float psize_min, psize_max;
721
722 if (!rs) {
723 return NULL;
724 }
725
726 rs->scissor_enable = state->scissor;
727 rs->two_side = state->light_twoside;
728 rs->multisample_enable = state->multisample;
729 rs->force_persample_interp = state->force_persample_interp;
730 rs->clip_plane_enable = state->clip_plane_enable;
731 rs->line_stipple_enable = state->line_stipple_enable;
732 rs->poly_stipple_enable = state->poly_stipple_enable;
733 rs->line_smooth = state->line_smooth;
734 rs->poly_smooth = state->poly_smooth;
735 rs->uses_poly_offset = state->offset_point || state->offset_line ||
736 state->offset_tri;
737 rs->clamp_fragment_color = state->clamp_fragment_color;
738 rs->flatshade = state->flatshade;
739 rs->sprite_coord_enable = state->sprite_coord_enable;
740 rs->rasterizer_discard = state->rasterizer_discard;
741 rs->pa_sc_line_stipple = state->line_stipple_enable ?
742 S_028A0C_LINE_PATTERN(state->line_stipple_pattern) |
743 S_028A0C_REPEAT_COUNT(state->line_stipple_factor) : 0;
744 rs->pa_cl_clip_cntl =
745 S_028810_DX_CLIP_SPACE_DEF(state->clip_halfz) |
746 S_028810_ZCLIP_NEAR_DISABLE(!state->depth_clip) |
747 S_028810_ZCLIP_FAR_DISABLE(!state->depth_clip) |
748 S_028810_DX_RASTERIZATION_KILL(state->rasterizer_discard) |
749 S_028810_DX_LINEAR_ATTR_CLIP_ENA(1);
750
751 si_pm4_set_reg(pm4, R_0286D4_SPI_INTERP_CONTROL_0,
752 S_0286D4_FLAT_SHADE_ENA(1) |
753 S_0286D4_PNT_SPRITE_ENA(1) |
754 S_0286D4_PNT_SPRITE_OVRD_X(V_0286D4_SPI_PNT_SPRITE_SEL_S) |
755 S_0286D4_PNT_SPRITE_OVRD_Y(V_0286D4_SPI_PNT_SPRITE_SEL_T) |
756 S_0286D4_PNT_SPRITE_OVRD_Z(V_0286D4_SPI_PNT_SPRITE_SEL_0) |
757 S_0286D4_PNT_SPRITE_OVRD_W(V_0286D4_SPI_PNT_SPRITE_SEL_1) |
758 S_0286D4_PNT_SPRITE_TOP_1(state->sprite_coord_mode != PIPE_SPRITE_COORD_UPPER_LEFT));
759
760 /* point size 12.4 fixed point */
761 tmp = (unsigned)(state->point_size * 8.0);
762 si_pm4_set_reg(pm4, R_028A00_PA_SU_POINT_SIZE, S_028A00_HEIGHT(tmp) | S_028A00_WIDTH(tmp));
763
764 if (state->point_size_per_vertex) {
765 psize_min = util_get_min_point_size(state);
766 psize_max = 8192;
767 } else {
768 /* Force the point size to be as if the vertex output was disabled. */
769 psize_min = state->point_size;
770 psize_max = state->point_size;
771 }
772 /* Divide by two, because 0.5 = 1 pixel. */
773 si_pm4_set_reg(pm4, R_028A04_PA_SU_POINT_MINMAX,
774 S_028A04_MIN_SIZE(si_pack_float_12p4(psize_min/2)) |
775 S_028A04_MAX_SIZE(si_pack_float_12p4(psize_max/2)));
776
777 tmp = (unsigned)state->line_width * 8;
778 si_pm4_set_reg(pm4, R_028A08_PA_SU_LINE_CNTL, S_028A08_WIDTH(tmp));
779 si_pm4_set_reg(pm4, R_028A48_PA_SC_MODE_CNTL_0,
780 S_028A48_LINE_STIPPLE_ENABLE(state->line_stipple_enable) |
781 S_028A48_MSAA_ENABLE(state->multisample ||
782 state->poly_smooth ||
783 state->line_smooth) |
784 S_028A48_VPORT_SCISSOR_ENABLE(1));
785
786 si_pm4_set_reg(pm4, R_028BE4_PA_SU_VTX_CNTL,
787 S_028BE4_PIX_CENTER(state->half_pixel_center) |
788 S_028BE4_QUANT_MODE(V_028BE4_X_16_8_FIXED_POINT_1_256TH));
789
790 si_pm4_set_reg(pm4, R_028B7C_PA_SU_POLY_OFFSET_CLAMP, fui(state->offset_clamp));
791 si_pm4_set_reg(pm4, R_028814_PA_SU_SC_MODE_CNTL,
792 S_028814_PROVOKING_VTX_LAST(!state->flatshade_first) |
793 S_028814_CULL_FRONT((state->cull_face & PIPE_FACE_FRONT) ? 1 : 0) |
794 S_028814_CULL_BACK((state->cull_face & PIPE_FACE_BACK) ? 1 : 0) |
795 S_028814_FACE(!state->front_ccw) |
796 S_028814_POLY_OFFSET_FRONT_ENABLE(util_get_offset(state, state->fill_front)) |
797 S_028814_POLY_OFFSET_BACK_ENABLE(util_get_offset(state, state->fill_back)) |
798 S_028814_POLY_OFFSET_PARA_ENABLE(state->offset_point || state->offset_line) |
799 S_028814_POLY_MODE(state->fill_front != PIPE_POLYGON_MODE_FILL ||
800 state->fill_back != PIPE_POLYGON_MODE_FILL) |
801 S_028814_POLYMODE_FRONT_PTYPE(si_translate_fill(state->fill_front)) |
802 S_028814_POLYMODE_BACK_PTYPE(si_translate_fill(state->fill_back)));
803 si_pm4_set_reg(pm4, R_00B130_SPI_SHADER_USER_DATA_VS_0 +
804 SI_SGPR_VS_STATE_BITS * 4, state->clamp_vertex_color);
805
806 /* Precalculate polygon offset states for 16-bit, 24-bit, and 32-bit zbuffers. */
807 for (i = 0; i < 3; i++) {
808 struct si_pm4_state *pm4 = &rs->pm4_poly_offset[i];
809 float offset_units = state->offset_units;
810 float offset_scale = state->offset_scale * 16.0f;
811 uint32_t pa_su_poly_offset_db_fmt_cntl = 0;
812
813 if (!state->offset_units_unscaled) {
814 switch (i) {
815 case 0: /* 16-bit zbuffer */
816 offset_units *= 4.0f;
817 pa_su_poly_offset_db_fmt_cntl =
818 S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-16);
819 break;
820 case 1: /* 24-bit zbuffer */
821 offset_units *= 2.0f;
822 pa_su_poly_offset_db_fmt_cntl =
823 S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-24);
824 break;
825 case 2: /* 32-bit zbuffer */
826 offset_units *= 1.0f;
827 pa_su_poly_offset_db_fmt_cntl = S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-23) |
828 S_028B78_POLY_OFFSET_DB_IS_FLOAT_FMT(1);
829 break;
830 }
831 }
832
833 si_pm4_set_reg(pm4, R_028B80_PA_SU_POLY_OFFSET_FRONT_SCALE,
834 fui(offset_scale));
835 si_pm4_set_reg(pm4, R_028B84_PA_SU_POLY_OFFSET_FRONT_OFFSET,
836 fui(offset_units));
837 si_pm4_set_reg(pm4, R_028B88_PA_SU_POLY_OFFSET_BACK_SCALE,
838 fui(offset_scale));
839 si_pm4_set_reg(pm4, R_028B8C_PA_SU_POLY_OFFSET_BACK_OFFSET,
840 fui(offset_units));
841 si_pm4_set_reg(pm4, R_028B78_PA_SU_POLY_OFFSET_DB_FMT_CNTL,
842 pa_su_poly_offset_db_fmt_cntl);
843 }
844
845 return rs;
846 }
847
848 static void si_bind_rs_state(struct pipe_context *ctx, void *state)
849 {
850 struct si_context *sctx = (struct si_context *)ctx;
851 struct si_state_rasterizer *old_rs =
852 (struct si_state_rasterizer*)sctx->queued.named.rasterizer;
853 struct si_state_rasterizer *rs = (struct si_state_rasterizer *)state;
854
855 if (!state)
856 return;
857
858 if (sctx->framebuffer.nr_samples > 1 &&
859 (!old_rs || old_rs->multisample_enable != rs->multisample_enable))
860 si_mark_atom_dirty(sctx, &sctx->db_render_state);
861
862 r600_set_scissor_enable(&sctx->b, rs->scissor_enable);
863
864 si_pm4_bind_state(sctx, rasterizer, rs);
865 si_update_poly_offset_state(sctx);
866
867 si_mark_atom_dirty(sctx, &sctx->clip_regs);
868 }
869
870 static void si_delete_rs_state(struct pipe_context *ctx, void *state)
871 {
872 struct si_context *sctx = (struct si_context *)ctx;
873
874 if (sctx->queued.named.rasterizer == state)
875 si_pm4_bind_state(sctx, poly_offset, NULL);
876 si_pm4_delete_state(sctx, rasterizer, (struct si_state_rasterizer *)state);
877 }
878
879 /*
880 * infeered state between dsa and stencil ref
881 */
882 static void si_emit_stencil_ref(struct si_context *sctx, struct r600_atom *atom)
883 {
884 struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
885 struct pipe_stencil_ref *ref = &sctx->stencil_ref.state;
886 struct si_dsa_stencil_ref_part *dsa = &sctx->stencil_ref.dsa_part;
887
888 radeon_set_context_reg_seq(cs, R_028430_DB_STENCILREFMASK, 2);
889 radeon_emit(cs, S_028430_STENCILTESTVAL(ref->ref_value[0]) |
890 S_028430_STENCILMASK(dsa->valuemask[0]) |
891 S_028430_STENCILWRITEMASK(dsa->writemask[0]) |
892 S_028430_STENCILOPVAL(1));
893 radeon_emit(cs, S_028434_STENCILTESTVAL_BF(ref->ref_value[1]) |
894 S_028434_STENCILMASK_BF(dsa->valuemask[1]) |
895 S_028434_STENCILWRITEMASK_BF(dsa->writemask[1]) |
896 S_028434_STENCILOPVAL_BF(1));
897 }
898
899 static void si_set_stencil_ref(struct pipe_context *ctx,
900 const struct pipe_stencil_ref *state)
901 {
902 struct si_context *sctx = (struct si_context *)ctx;
903
904 if (memcmp(&sctx->stencil_ref.state, state, sizeof(*state)) == 0)
905 return;
906
907 sctx->stencil_ref.state = *state;
908 si_mark_atom_dirty(sctx, &sctx->stencil_ref.atom);
909 }
910
911
912 /*
913 * DSA
914 */
915
916 static uint32_t si_translate_stencil_op(int s_op)
917 {
918 switch (s_op) {
919 case PIPE_STENCIL_OP_KEEP:
920 return V_02842C_STENCIL_KEEP;
921 case PIPE_STENCIL_OP_ZERO:
922 return V_02842C_STENCIL_ZERO;
923 case PIPE_STENCIL_OP_REPLACE:
924 return V_02842C_STENCIL_REPLACE_TEST;
925 case PIPE_STENCIL_OP_INCR:
926 return V_02842C_STENCIL_ADD_CLAMP;
927 case PIPE_STENCIL_OP_DECR:
928 return V_02842C_STENCIL_SUB_CLAMP;
929 case PIPE_STENCIL_OP_INCR_WRAP:
930 return V_02842C_STENCIL_ADD_WRAP;
931 case PIPE_STENCIL_OP_DECR_WRAP:
932 return V_02842C_STENCIL_SUB_WRAP;
933 case PIPE_STENCIL_OP_INVERT:
934 return V_02842C_STENCIL_INVERT;
935 default:
936 R600_ERR("Unknown stencil op %d", s_op);
937 assert(0);
938 break;
939 }
940 return 0;
941 }
942
943 static void *si_create_dsa_state(struct pipe_context *ctx,
944 const struct pipe_depth_stencil_alpha_state *state)
945 {
946 struct si_state_dsa *dsa = CALLOC_STRUCT(si_state_dsa);
947 struct si_pm4_state *pm4 = &dsa->pm4;
948 unsigned db_depth_control;
949 uint32_t db_stencil_control = 0;
950
951 if (!dsa) {
952 return NULL;
953 }
954
955 dsa->stencil_ref.valuemask[0] = state->stencil[0].valuemask;
956 dsa->stencil_ref.valuemask[1] = state->stencil[1].valuemask;
957 dsa->stencil_ref.writemask[0] = state->stencil[0].writemask;
958 dsa->stencil_ref.writemask[1] = state->stencil[1].writemask;
959
960 db_depth_control = S_028800_Z_ENABLE(state->depth.enabled) |
961 S_028800_Z_WRITE_ENABLE(state->depth.writemask) |
962 S_028800_ZFUNC(state->depth.func) |
963 S_028800_DEPTH_BOUNDS_ENABLE(state->depth.bounds_test);
964
965 /* stencil */
966 if (state->stencil[0].enabled) {
967 db_depth_control |= S_028800_STENCIL_ENABLE(1);
968 db_depth_control |= S_028800_STENCILFUNC(state->stencil[0].func);
969 db_stencil_control |= S_02842C_STENCILFAIL(si_translate_stencil_op(state->stencil[0].fail_op));
970 db_stencil_control |= S_02842C_STENCILZPASS(si_translate_stencil_op(state->stencil[0].zpass_op));
971 db_stencil_control |= S_02842C_STENCILZFAIL(si_translate_stencil_op(state->stencil[0].zfail_op));
972
973 if (state->stencil[1].enabled) {
974 db_depth_control |= S_028800_BACKFACE_ENABLE(1);
975 db_depth_control |= S_028800_STENCILFUNC_BF(state->stencil[1].func);
976 db_stencil_control |= S_02842C_STENCILFAIL_BF(si_translate_stencil_op(state->stencil[1].fail_op));
977 db_stencil_control |= S_02842C_STENCILZPASS_BF(si_translate_stencil_op(state->stencil[1].zpass_op));
978 db_stencil_control |= S_02842C_STENCILZFAIL_BF(si_translate_stencil_op(state->stencil[1].zfail_op));
979 }
980 }
981
982 /* alpha */
983 if (state->alpha.enabled) {
984 dsa->alpha_func = state->alpha.func;
985
986 si_pm4_set_reg(pm4, R_00B030_SPI_SHADER_USER_DATA_PS_0 +
987 SI_SGPR_ALPHA_REF * 4, fui(state->alpha.ref_value));
988 } else {
989 dsa->alpha_func = PIPE_FUNC_ALWAYS;
990 }
991
992 si_pm4_set_reg(pm4, R_028800_DB_DEPTH_CONTROL, db_depth_control);
993 si_pm4_set_reg(pm4, R_02842C_DB_STENCIL_CONTROL, db_stencil_control);
994 if (state->depth.bounds_test) {
995 si_pm4_set_reg(pm4, R_028020_DB_DEPTH_BOUNDS_MIN, fui(state->depth.bounds_min));
996 si_pm4_set_reg(pm4, R_028024_DB_DEPTH_BOUNDS_MAX, fui(state->depth.bounds_max));
997 }
998
999 return dsa;
1000 }
1001
1002 static void si_bind_dsa_state(struct pipe_context *ctx, void *state)
1003 {
1004 struct si_context *sctx = (struct si_context *)ctx;
1005 struct si_state_dsa *dsa = state;
1006
1007 if (!state)
1008 return;
1009
1010 si_pm4_bind_state(sctx, dsa, dsa);
1011
1012 if (memcmp(&dsa->stencil_ref, &sctx->stencil_ref.dsa_part,
1013 sizeof(struct si_dsa_stencil_ref_part)) != 0) {
1014 sctx->stencil_ref.dsa_part = dsa->stencil_ref;
1015 si_mark_atom_dirty(sctx, &sctx->stencil_ref.atom);
1016 }
1017 }
1018
1019 static void si_delete_dsa_state(struct pipe_context *ctx, void *state)
1020 {
1021 struct si_context *sctx = (struct si_context *)ctx;
1022 si_pm4_delete_state(sctx, dsa, (struct si_state_dsa *)state);
1023 }
1024
1025 static void *si_create_db_flush_dsa(struct si_context *sctx)
1026 {
1027 struct pipe_depth_stencil_alpha_state dsa = {};
1028
1029 return sctx->b.b.create_depth_stencil_alpha_state(&sctx->b.b, &dsa);
1030 }
1031
1032 /* DB RENDER STATE */
1033
1034 static void si_set_active_query_state(struct pipe_context *ctx, boolean enable)
1035 {
1036 struct si_context *sctx = (struct si_context*)ctx;
1037
1038 /* Pipeline stat & streamout queries. */
1039 if (enable) {
1040 sctx->b.flags &= ~R600_CONTEXT_STOP_PIPELINE_STATS;
1041 sctx->b.flags |= R600_CONTEXT_START_PIPELINE_STATS;
1042 } else {
1043 sctx->b.flags &= ~R600_CONTEXT_START_PIPELINE_STATS;
1044 sctx->b.flags |= R600_CONTEXT_STOP_PIPELINE_STATS;
1045 }
1046
1047 /* Occlusion queries. */
1048 if (sctx->occlusion_queries_disabled != !enable) {
1049 sctx->occlusion_queries_disabled = !enable;
1050 si_mark_atom_dirty(sctx, &sctx->db_render_state);
1051 }
1052 }
1053
1054 static void si_set_occlusion_query_state(struct pipe_context *ctx, bool enable)
1055 {
1056 struct si_context *sctx = (struct si_context*)ctx;
1057
1058 si_mark_atom_dirty(sctx, &sctx->db_render_state);
1059 }
1060
1061 static void si_emit_db_render_state(struct si_context *sctx, struct r600_atom *state)
1062 {
1063 struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
1064 struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
1065 unsigned db_shader_control;
1066
1067 radeon_set_context_reg_seq(cs, R_028000_DB_RENDER_CONTROL, 2);
1068
1069 /* DB_RENDER_CONTROL */
1070 if (sctx->dbcb_depth_copy_enabled ||
1071 sctx->dbcb_stencil_copy_enabled) {
1072 radeon_emit(cs,
1073 S_028000_DEPTH_COPY(sctx->dbcb_depth_copy_enabled) |
1074 S_028000_STENCIL_COPY(sctx->dbcb_stencil_copy_enabled) |
1075 S_028000_COPY_CENTROID(1) |
1076 S_028000_COPY_SAMPLE(sctx->dbcb_copy_sample));
1077 } else if (sctx->db_flush_depth_inplace || sctx->db_flush_stencil_inplace) {
1078 radeon_emit(cs,
1079 S_028000_DEPTH_COMPRESS_DISABLE(sctx->db_flush_depth_inplace) |
1080 S_028000_STENCIL_COMPRESS_DISABLE(sctx->db_flush_stencil_inplace));
1081 } else {
1082 radeon_emit(cs,
1083 S_028000_DEPTH_CLEAR_ENABLE(sctx->db_depth_clear) |
1084 S_028000_STENCIL_CLEAR_ENABLE(sctx->db_stencil_clear));
1085 }
1086
1087 /* DB_COUNT_CONTROL (occlusion queries) */
1088 if (sctx->b.num_occlusion_queries > 0 &&
1089 !sctx->occlusion_queries_disabled) {
1090 bool perfect = sctx->b.num_perfect_occlusion_queries > 0;
1091
1092 if (sctx->b.chip_class >= CIK) {
1093 radeon_emit(cs,
1094 S_028004_PERFECT_ZPASS_COUNTS(perfect) |
1095 S_028004_SAMPLE_RATE(sctx->framebuffer.log_samples) |
1096 S_028004_ZPASS_ENABLE(1) |
1097 S_028004_SLICE_EVEN_ENABLE(1) |
1098 S_028004_SLICE_ODD_ENABLE(1));
1099 } else {
1100 radeon_emit(cs,
1101 S_028004_PERFECT_ZPASS_COUNTS(perfect) |
1102 S_028004_SAMPLE_RATE(sctx->framebuffer.log_samples));
1103 }
1104 } else {
1105 /* Disable occlusion queries. */
1106 if (sctx->b.chip_class >= CIK) {
1107 radeon_emit(cs, 0);
1108 } else {
1109 radeon_emit(cs, S_028004_ZPASS_INCREMENT_DISABLE(1));
1110 }
1111 }
1112
1113 /* DB_RENDER_OVERRIDE2 */
1114 radeon_set_context_reg(cs, R_028010_DB_RENDER_OVERRIDE2,
1115 S_028010_DISABLE_ZMASK_EXPCLEAR_OPTIMIZATION(sctx->db_depth_disable_expclear) |
1116 S_028010_DISABLE_SMEM_EXPCLEAR_OPTIMIZATION(sctx->db_stencil_disable_expclear) |
1117 S_028010_DECOMPRESS_Z_ON_FLUSH(sctx->framebuffer.nr_samples >= 4));
1118
1119 db_shader_control = S_02880C_ALPHA_TO_MASK_DISABLE(sctx->framebuffer.cb0_is_integer) |
1120 sctx->ps_db_shader_control;
1121
1122 /* Bug workaround for smoothing (overrasterization) on SI. */
1123 if (sctx->b.chip_class == SI && sctx->smoothing_enabled) {
1124 db_shader_control &= C_02880C_Z_ORDER;
1125 db_shader_control |= S_02880C_Z_ORDER(V_02880C_LATE_Z);
1126 }
1127
1128 /* Disable the gl_SampleMask fragment shader output if MSAA is disabled. */
1129 if (sctx->framebuffer.nr_samples <= 1 || (rs && !rs->multisample_enable))
1130 db_shader_control &= C_02880C_MASK_EXPORT_ENABLE;
1131
1132 if (sctx->b.family == CHIP_STONEY &&
1133 sctx->screen->b.debug_flags & DBG_NO_RB_PLUS)
1134 db_shader_control |= S_02880C_DUAL_QUAD_DISABLE(1);
1135
1136 radeon_set_context_reg(cs, R_02880C_DB_SHADER_CONTROL,
1137 db_shader_control);
1138 }
1139
1140 /*
1141 * format translation
1142 */
1143 static uint32_t si_translate_colorformat(enum pipe_format format)
1144 {
1145 const struct util_format_description *desc = util_format_description(format);
1146
1147 #define HAS_SIZE(x,y,z,w) \
1148 (desc->channel[0].size == (x) && desc->channel[1].size == (y) && \
1149 desc->channel[2].size == (z) && desc->channel[3].size == (w))
1150
1151 if (format == PIPE_FORMAT_R11G11B10_FLOAT) /* isn't plain */
1152 return V_028C70_COLOR_10_11_11;
1153
1154 if (desc->layout != UTIL_FORMAT_LAYOUT_PLAIN)
1155 return V_028C70_COLOR_INVALID;
1156
1157 /* hw cannot support mixed formats (except depth/stencil, since
1158 * stencil is not written to). */
1159 if (desc->is_mixed && desc->colorspace != UTIL_FORMAT_COLORSPACE_ZS)
1160 return V_028C70_COLOR_INVALID;
1161
1162 switch (desc->nr_channels) {
1163 case 1:
1164 switch (desc->channel[0].size) {
1165 case 8:
1166 return V_028C70_COLOR_8;
1167 case 16:
1168 return V_028C70_COLOR_16;
1169 case 32:
1170 return V_028C70_COLOR_32;
1171 }
1172 break;
1173 case 2:
1174 if (desc->channel[0].size == desc->channel[1].size) {
1175 switch (desc->channel[0].size) {
1176 case 8:
1177 return V_028C70_COLOR_8_8;
1178 case 16:
1179 return V_028C70_COLOR_16_16;
1180 case 32:
1181 return V_028C70_COLOR_32_32;
1182 }
1183 } else if (HAS_SIZE(8,24,0,0)) {
1184 return V_028C70_COLOR_24_8;
1185 } else if (HAS_SIZE(24,8,0,0)) {
1186 return V_028C70_COLOR_8_24;
1187 }
1188 break;
1189 case 3:
1190 if (HAS_SIZE(5,6,5,0)) {
1191 return V_028C70_COLOR_5_6_5;
1192 } else if (HAS_SIZE(32,8,24,0)) {
1193 return V_028C70_COLOR_X24_8_32_FLOAT;
1194 }
1195 break;
1196 case 4:
1197 if (desc->channel[0].size == desc->channel[1].size &&
1198 desc->channel[0].size == desc->channel[2].size &&
1199 desc->channel[0].size == desc->channel[3].size) {
1200 switch (desc->channel[0].size) {
1201 case 4:
1202 return V_028C70_COLOR_4_4_4_4;
1203 case 8:
1204 return V_028C70_COLOR_8_8_8_8;
1205 case 16:
1206 return V_028C70_COLOR_16_16_16_16;
1207 case 32:
1208 return V_028C70_COLOR_32_32_32_32;
1209 }
1210 } else if (HAS_SIZE(5,5,5,1)) {
1211 return V_028C70_COLOR_1_5_5_5;
1212 } else if (HAS_SIZE(10,10,10,2)) {
1213 return V_028C70_COLOR_2_10_10_10;
1214 }
1215 break;
1216 }
1217 return V_028C70_COLOR_INVALID;
1218 }
1219
1220 static uint32_t si_colorformat_endian_swap(uint32_t colorformat)
1221 {
1222 if (SI_BIG_ENDIAN) {
1223 switch(colorformat) {
1224 /* 8-bit buffers. */
1225 case V_028C70_COLOR_8:
1226 return V_028C70_ENDIAN_NONE;
1227
1228 /* 16-bit buffers. */
1229 case V_028C70_COLOR_5_6_5:
1230 case V_028C70_COLOR_1_5_5_5:
1231 case V_028C70_COLOR_4_4_4_4:
1232 case V_028C70_COLOR_16:
1233 case V_028C70_COLOR_8_8:
1234 return V_028C70_ENDIAN_8IN16;
1235
1236 /* 32-bit buffers. */
1237 case V_028C70_COLOR_8_8_8_8:
1238 case V_028C70_COLOR_2_10_10_10:
1239 case V_028C70_COLOR_8_24:
1240 case V_028C70_COLOR_24_8:
1241 case V_028C70_COLOR_16_16:
1242 return V_028C70_ENDIAN_8IN32;
1243
1244 /* 64-bit buffers. */
1245 case V_028C70_COLOR_16_16_16_16:
1246 return V_028C70_ENDIAN_8IN16;
1247
1248 case V_028C70_COLOR_32_32:
1249 return V_028C70_ENDIAN_8IN32;
1250
1251 /* 128-bit buffers. */
1252 case V_028C70_COLOR_32_32_32_32:
1253 return V_028C70_ENDIAN_8IN32;
1254 default:
1255 return V_028C70_ENDIAN_NONE; /* Unsupported. */
1256 }
1257 } else {
1258 return V_028C70_ENDIAN_NONE;
1259 }
1260 }
1261
1262 static uint32_t si_translate_dbformat(enum pipe_format format)
1263 {
1264 switch (format) {
1265 case PIPE_FORMAT_Z16_UNORM:
1266 return V_028040_Z_16;
1267 case PIPE_FORMAT_S8_UINT_Z24_UNORM:
1268 case PIPE_FORMAT_X8Z24_UNORM:
1269 case PIPE_FORMAT_Z24X8_UNORM:
1270 case PIPE_FORMAT_Z24_UNORM_S8_UINT:
1271 return V_028040_Z_24; /* deprecated on SI */
1272 case PIPE_FORMAT_Z32_FLOAT:
1273 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
1274 return V_028040_Z_32_FLOAT;
1275 default:
1276 return V_028040_Z_INVALID;
1277 }
1278 }
1279
1280 /*
1281 * Texture translation
1282 */
1283
1284 static uint32_t si_translate_texformat(struct pipe_screen *screen,
1285 enum pipe_format format,
1286 const struct util_format_description *desc,
1287 int first_non_void)
1288 {
1289 struct si_screen *sscreen = (struct si_screen*)screen;
1290 bool enable_compressed_formats = (sscreen->b.info.drm_major == 2 &&
1291 sscreen->b.info.drm_minor >= 31) ||
1292 sscreen->b.info.drm_major == 3;
1293 boolean uniform = TRUE;
1294 int i;
1295
1296 /* Colorspace (return non-RGB formats directly). */
1297 switch (desc->colorspace) {
1298 /* Depth stencil formats */
1299 case UTIL_FORMAT_COLORSPACE_ZS:
1300 switch (format) {
1301 case PIPE_FORMAT_Z16_UNORM:
1302 return V_008F14_IMG_DATA_FORMAT_16;
1303 case PIPE_FORMAT_X24S8_UINT:
1304 case PIPE_FORMAT_Z24X8_UNORM:
1305 case PIPE_FORMAT_Z24_UNORM_S8_UINT:
1306 return V_008F14_IMG_DATA_FORMAT_8_24;
1307 case PIPE_FORMAT_X8Z24_UNORM:
1308 case PIPE_FORMAT_S8X24_UINT:
1309 case PIPE_FORMAT_S8_UINT_Z24_UNORM:
1310 return V_008F14_IMG_DATA_FORMAT_24_8;
1311 case PIPE_FORMAT_S8_UINT:
1312 return V_008F14_IMG_DATA_FORMAT_8;
1313 case PIPE_FORMAT_Z32_FLOAT:
1314 return V_008F14_IMG_DATA_FORMAT_32;
1315 case PIPE_FORMAT_X32_S8X24_UINT:
1316 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
1317 return V_008F14_IMG_DATA_FORMAT_X24_8_32;
1318 default:
1319 goto out_unknown;
1320 }
1321
1322 case UTIL_FORMAT_COLORSPACE_YUV:
1323 goto out_unknown; /* TODO */
1324
1325 case UTIL_FORMAT_COLORSPACE_SRGB:
1326 if (desc->nr_channels != 4 && desc->nr_channels != 1)
1327 goto out_unknown;
1328 break;
1329
1330 default:
1331 break;
1332 }
1333
1334 if (desc->layout == UTIL_FORMAT_LAYOUT_RGTC) {
1335 if (!enable_compressed_formats)
1336 goto out_unknown;
1337
1338 switch (format) {
1339 case PIPE_FORMAT_RGTC1_SNORM:
1340 case PIPE_FORMAT_LATC1_SNORM:
1341 case PIPE_FORMAT_RGTC1_UNORM:
1342 case PIPE_FORMAT_LATC1_UNORM:
1343 return V_008F14_IMG_DATA_FORMAT_BC4;
1344 case PIPE_FORMAT_RGTC2_SNORM:
1345 case PIPE_FORMAT_LATC2_SNORM:
1346 case PIPE_FORMAT_RGTC2_UNORM:
1347 case PIPE_FORMAT_LATC2_UNORM:
1348 return V_008F14_IMG_DATA_FORMAT_BC5;
1349 default:
1350 goto out_unknown;
1351 }
1352 }
1353
1354 if (desc->layout == UTIL_FORMAT_LAYOUT_ETC &&
1355 sscreen->b.family == CHIP_STONEY) {
1356 switch (format) {
1357 case PIPE_FORMAT_ETC1_RGB8:
1358 case PIPE_FORMAT_ETC2_RGB8:
1359 case PIPE_FORMAT_ETC2_SRGB8:
1360 return V_008F14_IMG_DATA_FORMAT_ETC2_RGB;
1361 case PIPE_FORMAT_ETC2_RGB8A1:
1362 case PIPE_FORMAT_ETC2_SRGB8A1:
1363 return V_008F14_IMG_DATA_FORMAT_ETC2_RGBA1;
1364 case PIPE_FORMAT_ETC2_RGBA8:
1365 case PIPE_FORMAT_ETC2_SRGBA8:
1366 return V_008F14_IMG_DATA_FORMAT_ETC2_RGBA;
1367 case PIPE_FORMAT_ETC2_R11_UNORM:
1368 case PIPE_FORMAT_ETC2_R11_SNORM:
1369 return V_008F14_IMG_DATA_FORMAT_ETC2_R;
1370 case PIPE_FORMAT_ETC2_RG11_UNORM:
1371 case PIPE_FORMAT_ETC2_RG11_SNORM:
1372 return V_008F14_IMG_DATA_FORMAT_ETC2_RG;
1373 default:
1374 goto out_unknown;
1375 }
1376 }
1377
1378 if (desc->layout == UTIL_FORMAT_LAYOUT_BPTC) {
1379 if (!enable_compressed_formats)
1380 goto out_unknown;
1381
1382 switch (format) {
1383 case PIPE_FORMAT_BPTC_RGBA_UNORM:
1384 case PIPE_FORMAT_BPTC_SRGBA:
1385 return V_008F14_IMG_DATA_FORMAT_BC7;
1386 case PIPE_FORMAT_BPTC_RGB_FLOAT:
1387 case PIPE_FORMAT_BPTC_RGB_UFLOAT:
1388 return V_008F14_IMG_DATA_FORMAT_BC6;
1389 default:
1390 goto out_unknown;
1391 }
1392 }
1393
1394 if (desc->layout == UTIL_FORMAT_LAYOUT_SUBSAMPLED) {
1395 switch (format) {
1396 case PIPE_FORMAT_R8G8_B8G8_UNORM:
1397 case PIPE_FORMAT_G8R8_B8R8_UNORM:
1398 return V_008F14_IMG_DATA_FORMAT_GB_GR;
1399 case PIPE_FORMAT_G8R8_G8B8_UNORM:
1400 case PIPE_FORMAT_R8G8_R8B8_UNORM:
1401 return V_008F14_IMG_DATA_FORMAT_BG_RG;
1402 default:
1403 goto out_unknown;
1404 }
1405 }
1406
1407 if (desc->layout == UTIL_FORMAT_LAYOUT_S3TC) {
1408 if (!enable_compressed_formats)
1409 goto out_unknown;
1410
1411 if (!util_format_s3tc_enabled) {
1412 goto out_unknown;
1413 }
1414
1415 switch (format) {
1416 case PIPE_FORMAT_DXT1_RGB:
1417 case PIPE_FORMAT_DXT1_RGBA:
1418 case PIPE_FORMAT_DXT1_SRGB:
1419 case PIPE_FORMAT_DXT1_SRGBA:
1420 return V_008F14_IMG_DATA_FORMAT_BC1;
1421 case PIPE_FORMAT_DXT3_RGBA:
1422 case PIPE_FORMAT_DXT3_SRGBA:
1423 return V_008F14_IMG_DATA_FORMAT_BC2;
1424 case PIPE_FORMAT_DXT5_RGBA:
1425 case PIPE_FORMAT_DXT5_SRGBA:
1426 return V_008F14_IMG_DATA_FORMAT_BC3;
1427 default:
1428 goto out_unknown;
1429 }
1430 }
1431
1432 if (format == PIPE_FORMAT_R9G9B9E5_FLOAT) {
1433 return V_008F14_IMG_DATA_FORMAT_5_9_9_9;
1434 } else if (format == PIPE_FORMAT_R11G11B10_FLOAT) {
1435 return V_008F14_IMG_DATA_FORMAT_10_11_11;
1436 }
1437
1438 /* R8G8Bx_SNORM - TODO CxV8U8 */
1439
1440 /* hw cannot support mixed formats (except depth/stencil, since only
1441 * depth is read).*/
1442 if (desc->is_mixed && desc->colorspace != UTIL_FORMAT_COLORSPACE_ZS)
1443 goto out_unknown;
1444
1445 /* See whether the components are of the same size. */
1446 for (i = 1; i < desc->nr_channels; i++) {
1447 uniform = uniform && desc->channel[0].size == desc->channel[i].size;
1448 }
1449
1450 /* Non-uniform formats. */
1451 if (!uniform) {
1452 switch(desc->nr_channels) {
1453 case 3:
1454 if (desc->channel[0].size == 5 &&
1455 desc->channel[1].size == 6 &&
1456 desc->channel[2].size == 5) {
1457 return V_008F14_IMG_DATA_FORMAT_5_6_5;
1458 }
1459 goto out_unknown;
1460 case 4:
1461 if (desc->channel[0].size == 5 &&
1462 desc->channel[1].size == 5 &&
1463 desc->channel[2].size == 5 &&
1464 desc->channel[3].size == 1) {
1465 return V_008F14_IMG_DATA_FORMAT_1_5_5_5;
1466 }
1467 if (desc->channel[0].size == 10 &&
1468 desc->channel[1].size == 10 &&
1469 desc->channel[2].size == 10 &&
1470 desc->channel[3].size == 2) {
1471 return V_008F14_IMG_DATA_FORMAT_2_10_10_10;
1472 }
1473 goto out_unknown;
1474 }
1475 goto out_unknown;
1476 }
1477
1478 if (first_non_void < 0 || first_non_void > 3)
1479 goto out_unknown;
1480
1481 /* uniform formats */
1482 switch (desc->channel[first_non_void].size) {
1483 case 4:
1484 switch (desc->nr_channels) {
1485 #if 0 /* Not supported for render targets */
1486 case 2:
1487 return V_008F14_IMG_DATA_FORMAT_4_4;
1488 #endif
1489 case 4:
1490 return V_008F14_IMG_DATA_FORMAT_4_4_4_4;
1491 }
1492 break;
1493 case 8:
1494 switch (desc->nr_channels) {
1495 case 1:
1496 return V_008F14_IMG_DATA_FORMAT_8;
1497 case 2:
1498 return V_008F14_IMG_DATA_FORMAT_8_8;
1499 case 4:
1500 return V_008F14_IMG_DATA_FORMAT_8_8_8_8;
1501 }
1502 break;
1503 case 16:
1504 switch (desc->nr_channels) {
1505 case 1:
1506 return V_008F14_IMG_DATA_FORMAT_16;
1507 case 2:
1508 return V_008F14_IMG_DATA_FORMAT_16_16;
1509 case 4:
1510 return V_008F14_IMG_DATA_FORMAT_16_16_16_16;
1511 }
1512 break;
1513 case 32:
1514 switch (desc->nr_channels) {
1515 case 1:
1516 return V_008F14_IMG_DATA_FORMAT_32;
1517 case 2:
1518 return V_008F14_IMG_DATA_FORMAT_32_32;
1519 #if 0 /* Not supported for render targets */
1520 case 3:
1521 return V_008F14_IMG_DATA_FORMAT_32_32_32;
1522 #endif
1523 case 4:
1524 return V_008F14_IMG_DATA_FORMAT_32_32_32_32;
1525 }
1526 }
1527
1528 out_unknown:
1529 /* R600_ERR("Unable to handle texformat %d %s\n", format, util_format_name(format)); */
1530 return ~0;
1531 }
1532
1533 static unsigned si_tex_wrap(unsigned wrap)
1534 {
1535 switch (wrap) {
1536 default:
1537 case PIPE_TEX_WRAP_REPEAT:
1538 return V_008F30_SQ_TEX_WRAP;
1539 case PIPE_TEX_WRAP_CLAMP:
1540 return V_008F30_SQ_TEX_CLAMP_HALF_BORDER;
1541 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
1542 return V_008F30_SQ_TEX_CLAMP_LAST_TEXEL;
1543 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
1544 return V_008F30_SQ_TEX_CLAMP_BORDER;
1545 case PIPE_TEX_WRAP_MIRROR_REPEAT:
1546 return V_008F30_SQ_TEX_MIRROR;
1547 case PIPE_TEX_WRAP_MIRROR_CLAMP:
1548 return V_008F30_SQ_TEX_MIRROR_ONCE_HALF_BORDER;
1549 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
1550 return V_008F30_SQ_TEX_MIRROR_ONCE_LAST_TEXEL;
1551 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
1552 return V_008F30_SQ_TEX_MIRROR_ONCE_BORDER;
1553 }
1554 }
1555
1556 static unsigned si_tex_mipfilter(unsigned filter)
1557 {
1558 switch (filter) {
1559 case PIPE_TEX_MIPFILTER_NEAREST:
1560 return V_008F38_SQ_TEX_Z_FILTER_POINT;
1561 case PIPE_TEX_MIPFILTER_LINEAR:
1562 return V_008F38_SQ_TEX_Z_FILTER_LINEAR;
1563 default:
1564 case PIPE_TEX_MIPFILTER_NONE:
1565 return V_008F38_SQ_TEX_Z_FILTER_NONE;
1566 }
1567 }
1568
1569 static unsigned si_tex_compare(unsigned compare)
1570 {
1571 switch (compare) {
1572 default:
1573 case PIPE_FUNC_NEVER:
1574 return V_008F30_SQ_TEX_DEPTH_COMPARE_NEVER;
1575 case PIPE_FUNC_LESS:
1576 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESS;
1577 case PIPE_FUNC_EQUAL:
1578 return V_008F30_SQ_TEX_DEPTH_COMPARE_EQUAL;
1579 case PIPE_FUNC_LEQUAL:
1580 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESSEQUAL;
1581 case PIPE_FUNC_GREATER:
1582 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATER;
1583 case PIPE_FUNC_NOTEQUAL:
1584 return V_008F30_SQ_TEX_DEPTH_COMPARE_NOTEQUAL;
1585 case PIPE_FUNC_GEQUAL:
1586 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATEREQUAL;
1587 case PIPE_FUNC_ALWAYS:
1588 return V_008F30_SQ_TEX_DEPTH_COMPARE_ALWAYS;
1589 }
1590 }
1591
1592 static unsigned si_tex_dim(unsigned res_target, unsigned view_target,
1593 unsigned nr_samples)
1594 {
1595 if (view_target == PIPE_TEXTURE_CUBE ||
1596 view_target == PIPE_TEXTURE_CUBE_ARRAY)
1597 res_target = view_target;
1598
1599 switch (res_target) {
1600 default:
1601 case PIPE_TEXTURE_1D:
1602 return V_008F1C_SQ_RSRC_IMG_1D;
1603 case PIPE_TEXTURE_1D_ARRAY:
1604 return V_008F1C_SQ_RSRC_IMG_1D_ARRAY;
1605 case PIPE_TEXTURE_2D:
1606 case PIPE_TEXTURE_RECT:
1607 return nr_samples > 1 ? V_008F1C_SQ_RSRC_IMG_2D_MSAA :
1608 V_008F1C_SQ_RSRC_IMG_2D;
1609 case PIPE_TEXTURE_2D_ARRAY:
1610 return nr_samples > 1 ? V_008F1C_SQ_RSRC_IMG_2D_MSAA_ARRAY :
1611 V_008F1C_SQ_RSRC_IMG_2D_ARRAY;
1612 case PIPE_TEXTURE_3D:
1613 return V_008F1C_SQ_RSRC_IMG_3D;
1614 case PIPE_TEXTURE_CUBE:
1615 case PIPE_TEXTURE_CUBE_ARRAY:
1616 return V_008F1C_SQ_RSRC_IMG_CUBE;
1617 }
1618 }
1619
1620 /*
1621 * Format support testing
1622 */
1623
1624 static bool si_is_sampler_format_supported(struct pipe_screen *screen, enum pipe_format format)
1625 {
1626 return si_translate_texformat(screen, format, util_format_description(format),
1627 util_format_get_first_non_void_channel(format)) != ~0U;
1628 }
1629
1630 static uint32_t si_translate_buffer_dataformat(struct pipe_screen *screen,
1631 const struct util_format_description *desc,
1632 int first_non_void)
1633 {
1634 unsigned type;
1635 int i;
1636
1637 if (desc->format == PIPE_FORMAT_R11G11B10_FLOAT)
1638 return V_008F0C_BUF_DATA_FORMAT_10_11_11;
1639
1640 assert(first_non_void >= 0);
1641 type = desc->channel[first_non_void].type;
1642
1643 if (type == UTIL_FORMAT_TYPE_FIXED)
1644 return V_008F0C_BUF_DATA_FORMAT_INVALID;
1645
1646 if (desc->nr_channels == 4 &&
1647 desc->channel[0].size == 10 &&
1648 desc->channel[1].size == 10 &&
1649 desc->channel[2].size == 10 &&
1650 desc->channel[3].size == 2)
1651 return V_008F0C_BUF_DATA_FORMAT_2_10_10_10;
1652
1653 /* See whether the components are of the same size. */
1654 for (i = 0; i < desc->nr_channels; i++) {
1655 if (desc->channel[first_non_void].size != desc->channel[i].size)
1656 return V_008F0C_BUF_DATA_FORMAT_INVALID;
1657 }
1658
1659 switch (desc->channel[first_non_void].size) {
1660 case 8:
1661 switch (desc->nr_channels) {
1662 case 1:
1663 return V_008F0C_BUF_DATA_FORMAT_8;
1664 case 2:
1665 return V_008F0C_BUF_DATA_FORMAT_8_8;
1666 case 3:
1667 case 4:
1668 return V_008F0C_BUF_DATA_FORMAT_8_8_8_8;
1669 }
1670 break;
1671 case 16:
1672 switch (desc->nr_channels) {
1673 case 1:
1674 return V_008F0C_BUF_DATA_FORMAT_16;
1675 case 2:
1676 return V_008F0C_BUF_DATA_FORMAT_16_16;
1677 case 3:
1678 case 4:
1679 return V_008F0C_BUF_DATA_FORMAT_16_16_16_16;
1680 }
1681 break;
1682 case 32:
1683 /* From the Southern Islands ISA documentation about MTBUF:
1684 * 'Memory reads of data in memory that is 32 or 64 bits do not
1685 * undergo any format conversion.'
1686 */
1687 if (type != UTIL_FORMAT_TYPE_FLOAT &&
1688 !desc->channel[first_non_void].pure_integer)
1689 return V_008F0C_BUF_DATA_FORMAT_INVALID;
1690
1691 switch (desc->nr_channels) {
1692 case 1:
1693 return V_008F0C_BUF_DATA_FORMAT_32;
1694 case 2:
1695 return V_008F0C_BUF_DATA_FORMAT_32_32;
1696 case 3:
1697 return V_008F0C_BUF_DATA_FORMAT_32_32_32;
1698 case 4:
1699 return V_008F0C_BUF_DATA_FORMAT_32_32_32_32;
1700 }
1701 break;
1702 }
1703
1704 return V_008F0C_BUF_DATA_FORMAT_INVALID;
1705 }
1706
1707 static uint32_t si_translate_buffer_numformat(struct pipe_screen *screen,
1708 const struct util_format_description *desc,
1709 int first_non_void)
1710 {
1711 if (desc->format == PIPE_FORMAT_R11G11B10_FLOAT)
1712 return V_008F0C_BUF_NUM_FORMAT_FLOAT;
1713
1714 assert(first_non_void >= 0);
1715
1716 switch (desc->channel[first_non_void].type) {
1717 case UTIL_FORMAT_TYPE_SIGNED:
1718 if (desc->channel[first_non_void].normalized)
1719 return V_008F0C_BUF_NUM_FORMAT_SNORM;
1720 else if (desc->channel[first_non_void].pure_integer)
1721 return V_008F0C_BUF_NUM_FORMAT_SINT;
1722 else
1723 return V_008F0C_BUF_NUM_FORMAT_SSCALED;
1724 break;
1725 case UTIL_FORMAT_TYPE_UNSIGNED:
1726 if (desc->channel[first_non_void].normalized)
1727 return V_008F0C_BUF_NUM_FORMAT_UNORM;
1728 else if (desc->channel[first_non_void].pure_integer)
1729 return V_008F0C_BUF_NUM_FORMAT_UINT;
1730 else
1731 return V_008F0C_BUF_NUM_FORMAT_USCALED;
1732 break;
1733 case UTIL_FORMAT_TYPE_FLOAT:
1734 default:
1735 return V_008F0C_BUF_NUM_FORMAT_FLOAT;
1736 }
1737 }
1738
1739 static bool si_is_vertex_format_supported(struct pipe_screen *screen, enum pipe_format format)
1740 {
1741 const struct util_format_description *desc;
1742 int first_non_void;
1743 unsigned data_format;
1744
1745 desc = util_format_description(format);
1746 first_non_void = util_format_get_first_non_void_channel(format);
1747 data_format = si_translate_buffer_dataformat(screen, desc, first_non_void);
1748 return data_format != V_008F0C_BUF_DATA_FORMAT_INVALID;
1749 }
1750
1751 static bool si_is_colorbuffer_format_supported(enum pipe_format format)
1752 {
1753 return si_translate_colorformat(format) != V_028C70_COLOR_INVALID &&
1754 r600_translate_colorswap(format, FALSE) != ~0U;
1755 }
1756
1757 static bool si_is_zs_format_supported(enum pipe_format format)
1758 {
1759 return si_translate_dbformat(format) != V_028040_Z_INVALID;
1760 }
1761
1762 boolean si_is_format_supported(struct pipe_screen *screen,
1763 enum pipe_format format,
1764 enum pipe_texture_target target,
1765 unsigned sample_count,
1766 unsigned usage)
1767 {
1768 unsigned retval = 0;
1769
1770 if (target >= PIPE_MAX_TEXTURE_TYPES) {
1771 R600_ERR("r600: unsupported texture type %d\n", target);
1772 return FALSE;
1773 }
1774
1775 if (!util_format_is_supported(format, usage))
1776 return FALSE;
1777
1778 if (sample_count > 1) {
1779 if (!screen->get_param(screen, PIPE_CAP_TEXTURE_MULTISAMPLE))
1780 return FALSE;
1781
1782 switch (sample_count) {
1783 case 2:
1784 case 4:
1785 case 8:
1786 break;
1787 case 16:
1788 if (format == PIPE_FORMAT_NONE)
1789 return TRUE;
1790 else
1791 return FALSE;
1792 default:
1793 return FALSE;
1794 }
1795 }
1796
1797 if (usage & (PIPE_BIND_SAMPLER_VIEW |
1798 PIPE_BIND_SHADER_IMAGE)) {
1799 if (target == PIPE_BUFFER) {
1800 if (si_is_vertex_format_supported(screen, format))
1801 retval |= usage & (PIPE_BIND_SAMPLER_VIEW |
1802 PIPE_BIND_SHADER_IMAGE);
1803 } else {
1804 if (si_is_sampler_format_supported(screen, format))
1805 retval |= usage & (PIPE_BIND_SAMPLER_VIEW |
1806 PIPE_BIND_SHADER_IMAGE);
1807 }
1808 }
1809
1810 if ((usage & (PIPE_BIND_RENDER_TARGET |
1811 PIPE_BIND_DISPLAY_TARGET |
1812 PIPE_BIND_SCANOUT |
1813 PIPE_BIND_SHARED |
1814 PIPE_BIND_BLENDABLE)) &&
1815 si_is_colorbuffer_format_supported(format)) {
1816 retval |= usage &
1817 (PIPE_BIND_RENDER_TARGET |
1818 PIPE_BIND_DISPLAY_TARGET |
1819 PIPE_BIND_SCANOUT |
1820 PIPE_BIND_SHARED);
1821 if (!util_format_is_pure_integer(format) &&
1822 !util_format_is_depth_or_stencil(format))
1823 retval |= usage & PIPE_BIND_BLENDABLE;
1824 }
1825
1826 if ((usage & PIPE_BIND_DEPTH_STENCIL) &&
1827 si_is_zs_format_supported(format)) {
1828 retval |= PIPE_BIND_DEPTH_STENCIL;
1829 }
1830
1831 if ((usage & PIPE_BIND_VERTEX_BUFFER) &&
1832 si_is_vertex_format_supported(screen, format)) {
1833 retval |= PIPE_BIND_VERTEX_BUFFER;
1834 }
1835
1836 if (usage & PIPE_BIND_TRANSFER_READ)
1837 retval |= PIPE_BIND_TRANSFER_READ;
1838 if (usage & PIPE_BIND_TRANSFER_WRITE)
1839 retval |= PIPE_BIND_TRANSFER_WRITE;
1840
1841 if ((usage & PIPE_BIND_LINEAR) &&
1842 !util_format_is_compressed(format) &&
1843 !(usage & PIPE_BIND_DEPTH_STENCIL))
1844 retval |= PIPE_BIND_LINEAR;
1845
1846 return retval == usage;
1847 }
1848
1849 /*
1850 * framebuffer handling
1851 */
1852
1853 static void si_choose_spi_color_formats(struct r600_surface *surf,
1854 unsigned format, unsigned swap,
1855 unsigned ntype, bool is_depth)
1856 {
1857 /* Alpha is needed for alpha-to-coverage.
1858 * Blending may be with or without alpha.
1859 */
1860 unsigned normal = 0; /* most optimal, may not support blending or export alpha */
1861 unsigned alpha = 0; /* exports alpha, but may not support blending */
1862 unsigned blend = 0; /* supports blending, but may not export alpha */
1863 unsigned blend_alpha = 0; /* least optimal, supports blending and exports alpha */
1864
1865 /* Choose the SPI color formats. These are required values for Stoney/RB+.
1866 * Other chips have multiple choices, though they are not necessarily better.
1867 */
1868 switch (format) {
1869 case V_028C70_COLOR_5_6_5:
1870 case V_028C70_COLOR_1_5_5_5:
1871 case V_028C70_COLOR_5_5_5_1:
1872 case V_028C70_COLOR_4_4_4_4:
1873 case V_028C70_COLOR_10_11_11:
1874 case V_028C70_COLOR_11_11_10:
1875 case V_028C70_COLOR_8:
1876 case V_028C70_COLOR_8_8:
1877 case V_028C70_COLOR_8_8_8_8:
1878 case V_028C70_COLOR_10_10_10_2:
1879 case V_028C70_COLOR_2_10_10_10:
1880 if (ntype == V_028C70_NUMBER_UINT)
1881 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_UINT16_ABGR;
1882 else if (ntype == V_028C70_NUMBER_SINT)
1883 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_SINT16_ABGR;
1884 else
1885 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_FP16_ABGR;
1886 break;
1887
1888 case V_028C70_COLOR_16:
1889 case V_028C70_COLOR_16_16:
1890 case V_028C70_COLOR_16_16_16_16:
1891 if (ntype == V_028C70_NUMBER_UNORM ||
1892 ntype == V_028C70_NUMBER_SNORM) {
1893 /* UNORM16 and SNORM16 don't support blending */
1894 if (ntype == V_028C70_NUMBER_UNORM)
1895 normal = alpha = V_028714_SPI_SHADER_UNORM16_ABGR;
1896 else
1897 normal = alpha = V_028714_SPI_SHADER_SNORM16_ABGR;
1898
1899 /* Use 32 bits per channel for blending. */
1900 if (format == V_028C70_COLOR_16) {
1901 if (swap == V_028C70_SWAP_STD) { /* R */
1902 blend = V_028714_SPI_SHADER_32_R;
1903 blend_alpha = V_028714_SPI_SHADER_32_AR;
1904 } else if (swap == V_028C70_SWAP_ALT_REV) /* A */
1905 blend = blend_alpha = V_028714_SPI_SHADER_32_AR;
1906 else
1907 assert(0);
1908 } else if (format == V_028C70_COLOR_16_16) {
1909 if (swap == V_028C70_SWAP_STD) { /* RG */
1910 blend = V_028714_SPI_SHADER_32_GR;
1911 blend_alpha = V_028714_SPI_SHADER_32_ABGR;
1912 } else if (swap == V_028C70_SWAP_ALT) /* RA */
1913 blend = blend_alpha = V_028714_SPI_SHADER_32_AR;
1914 else
1915 assert(0);
1916 } else /* 16_16_16_16 */
1917 blend = blend_alpha = V_028714_SPI_SHADER_32_ABGR;
1918 } else if (ntype == V_028C70_NUMBER_UINT)
1919 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_UINT16_ABGR;
1920 else if (ntype == V_028C70_NUMBER_SINT)
1921 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_SINT16_ABGR;
1922 else if (ntype == V_028C70_NUMBER_FLOAT)
1923 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_FP16_ABGR;
1924 else
1925 assert(0);
1926 break;
1927
1928 case V_028C70_COLOR_32:
1929 if (swap == V_028C70_SWAP_STD) { /* R */
1930 blend = normal = V_028714_SPI_SHADER_32_R;
1931 alpha = blend_alpha = V_028714_SPI_SHADER_32_AR;
1932 } else if (swap == V_028C70_SWAP_ALT_REV) /* A */
1933 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_32_AR;
1934 else
1935 assert(0);
1936 break;
1937
1938 case V_028C70_COLOR_32_32:
1939 if (swap == V_028C70_SWAP_STD) { /* RG */
1940 blend = normal = V_028714_SPI_SHADER_32_GR;
1941 alpha = blend_alpha = V_028714_SPI_SHADER_32_ABGR;
1942 } else if (swap == V_028C70_SWAP_ALT) /* RA */
1943 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_32_AR;
1944 else
1945 assert(0);
1946 break;
1947
1948 case V_028C70_COLOR_32_32_32_32:
1949 case V_028C70_COLOR_8_24:
1950 case V_028C70_COLOR_24_8:
1951 case V_028C70_COLOR_X24_8_32_FLOAT:
1952 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_32_ABGR;
1953 break;
1954
1955 default:
1956 assert(0);
1957 return;
1958 }
1959
1960 /* The DB->CB copy needs 32_ABGR. */
1961 if (is_depth)
1962 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_32_ABGR;
1963
1964 surf->spi_shader_col_format = normal;
1965 surf->spi_shader_col_format_alpha = alpha;
1966 surf->spi_shader_col_format_blend = blend;
1967 surf->spi_shader_col_format_blend_alpha = blend_alpha;
1968 }
1969
1970 static void si_initialize_color_surface(struct si_context *sctx,
1971 struct r600_surface *surf)
1972 {
1973 struct r600_texture *rtex = (struct r600_texture*)surf->base.texture;
1974 unsigned color_info, color_attrib, color_view;
1975 unsigned format, swap, ntype, endian;
1976 const struct util_format_description *desc;
1977 int i;
1978 unsigned blend_clamp = 0, blend_bypass = 0;
1979
1980 color_view = S_028C6C_SLICE_START(surf->base.u.tex.first_layer) |
1981 S_028C6C_SLICE_MAX(surf->base.u.tex.last_layer);
1982
1983 desc = util_format_description(surf->base.format);
1984 for (i = 0; i < 4; i++) {
1985 if (desc->channel[i].type != UTIL_FORMAT_TYPE_VOID) {
1986 break;
1987 }
1988 }
1989 if (i == 4 || desc->channel[i].type == UTIL_FORMAT_TYPE_FLOAT) {
1990 ntype = V_028C70_NUMBER_FLOAT;
1991 } else {
1992 ntype = V_028C70_NUMBER_UNORM;
1993 if (desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB)
1994 ntype = V_028C70_NUMBER_SRGB;
1995 else if (desc->channel[i].type == UTIL_FORMAT_TYPE_SIGNED) {
1996 if (desc->channel[i].pure_integer) {
1997 ntype = V_028C70_NUMBER_SINT;
1998 } else {
1999 assert(desc->channel[i].normalized);
2000 ntype = V_028C70_NUMBER_SNORM;
2001 }
2002 } else if (desc->channel[i].type == UTIL_FORMAT_TYPE_UNSIGNED) {
2003 if (desc->channel[i].pure_integer) {
2004 ntype = V_028C70_NUMBER_UINT;
2005 } else {
2006 assert(desc->channel[i].normalized);
2007 ntype = V_028C70_NUMBER_UNORM;
2008 }
2009 }
2010 }
2011
2012 format = si_translate_colorformat(surf->base.format);
2013 if (format == V_028C70_COLOR_INVALID) {
2014 R600_ERR("Invalid CB format: %d, disabling CB.\n", surf->base.format);
2015 }
2016 assert(format != V_028C70_COLOR_INVALID);
2017 swap = r600_translate_colorswap(surf->base.format, FALSE);
2018 endian = si_colorformat_endian_swap(format);
2019
2020 /* blend clamp should be set for all NORM/SRGB types */
2021 if (ntype == V_028C70_NUMBER_UNORM ||
2022 ntype == V_028C70_NUMBER_SNORM ||
2023 ntype == V_028C70_NUMBER_SRGB)
2024 blend_clamp = 1;
2025
2026 /* set blend bypass according to docs if SINT/UINT or
2027 8/24 COLOR variants */
2028 if (ntype == V_028C70_NUMBER_UINT || ntype == V_028C70_NUMBER_SINT ||
2029 format == V_028C70_COLOR_8_24 || format == V_028C70_COLOR_24_8 ||
2030 format == V_028C70_COLOR_X24_8_32_FLOAT) {
2031 blend_clamp = 0;
2032 blend_bypass = 1;
2033 }
2034
2035 if ((ntype == V_028C70_NUMBER_UINT || ntype == V_028C70_NUMBER_SINT) &&
2036 (format == V_028C70_COLOR_8 ||
2037 format == V_028C70_COLOR_8_8 ||
2038 format == V_028C70_COLOR_8_8_8_8))
2039 surf->color_is_int8 = true;
2040
2041 color_info = S_028C70_FORMAT(format) |
2042 S_028C70_COMP_SWAP(swap) |
2043 S_028C70_BLEND_CLAMP(blend_clamp) |
2044 S_028C70_BLEND_BYPASS(blend_bypass) |
2045 S_028C70_NUMBER_TYPE(ntype) |
2046 S_028C70_ENDIAN(endian);
2047
2048 /* Intensity is implemented as Red, so treat it that way. */
2049 color_attrib = S_028C74_FORCE_DST_ALPHA_1(desc->swizzle[3] == PIPE_SWIZZLE_1 ||
2050 util_format_is_intensity(surf->base.format));
2051
2052 if (rtex->resource.b.b.nr_samples > 1) {
2053 unsigned log_samples = util_logbase2(rtex->resource.b.b.nr_samples);
2054
2055 color_attrib |= S_028C74_NUM_SAMPLES(log_samples) |
2056 S_028C74_NUM_FRAGMENTS(log_samples);
2057
2058 if (rtex->fmask.size) {
2059 color_info |= S_028C70_COMPRESSION(1);
2060 unsigned fmask_bankh = util_logbase2(rtex->fmask.bank_height);
2061
2062 if (sctx->b.chip_class == SI) {
2063 /* due to a hw bug, FMASK_BANK_HEIGHT must be set on SI too */
2064 color_attrib |= S_028C74_FMASK_BANK_HEIGHT(fmask_bankh);
2065 }
2066 }
2067 }
2068
2069 surf->cb_color_view = color_view;
2070 surf->cb_color_info = color_info;
2071 surf->cb_color_attrib = color_attrib;
2072
2073 if (sctx->b.chip_class >= VI) {
2074 unsigned max_uncompressed_block_size = 2;
2075
2076 if (rtex->surface.nsamples > 1) {
2077 if (rtex->surface.bpe == 1)
2078 max_uncompressed_block_size = 0;
2079 else if (rtex->surface.bpe == 2)
2080 max_uncompressed_block_size = 1;
2081 }
2082
2083 surf->cb_dcc_control = S_028C78_MAX_UNCOMPRESSED_BLOCK_SIZE(max_uncompressed_block_size) |
2084 S_028C78_INDEPENDENT_64B_BLOCKS(1);
2085 }
2086
2087 /* This must be set for fast clear to work without FMASK. */
2088 if (!rtex->fmask.size && sctx->b.chip_class == SI) {
2089 unsigned bankh = util_logbase2(rtex->surface.bankh);
2090 surf->cb_color_attrib |= S_028C74_FMASK_BANK_HEIGHT(bankh);
2091 }
2092
2093 /* Determine pixel shader export format */
2094 si_choose_spi_color_formats(surf, format, swap, ntype, rtex->is_depth);
2095
2096 surf->color_initialized = true;
2097 }
2098
2099 static void si_init_depth_surface(struct si_context *sctx,
2100 struct r600_surface *surf)
2101 {
2102 struct r600_texture *rtex = (struct r600_texture*)surf->base.texture;
2103 unsigned level = surf->base.u.tex.level;
2104 struct radeon_surf_level *levelinfo = &rtex->surface.level[level];
2105 unsigned format;
2106 uint32_t z_info, s_info, db_depth_info;
2107 uint64_t z_offs, s_offs;
2108 uint32_t db_htile_data_base, db_htile_surface;
2109
2110 format = si_translate_dbformat(rtex->resource.b.b.format);
2111
2112 if (format == V_028040_Z_INVALID) {
2113 R600_ERR("Invalid DB format: %d, disabling DB.\n", rtex->resource.b.b.format);
2114 }
2115 assert(format != V_028040_Z_INVALID);
2116
2117 s_offs = z_offs = rtex->resource.gpu_address;
2118 z_offs += rtex->surface.level[level].offset;
2119 s_offs += rtex->surface.stencil_level[level].offset;
2120
2121 db_depth_info = S_02803C_ADDR5_SWIZZLE_MASK(1);
2122
2123 z_info = S_028040_FORMAT(format);
2124 if (rtex->resource.b.b.nr_samples > 1) {
2125 z_info |= S_028040_NUM_SAMPLES(util_logbase2(rtex->resource.b.b.nr_samples));
2126 }
2127
2128 if (rtex->surface.flags & RADEON_SURF_SBUFFER)
2129 s_info = S_028044_FORMAT(V_028044_STENCIL_8);
2130 else
2131 s_info = S_028044_FORMAT(V_028044_STENCIL_INVALID);
2132
2133 if (sctx->b.chip_class >= CIK) {
2134 struct radeon_info *info = &sctx->screen->b.info;
2135 unsigned index = rtex->surface.tiling_index[level];
2136 unsigned stencil_index = rtex->surface.stencil_tiling_index[level];
2137 unsigned macro_index = rtex->surface.macro_tile_index;
2138 unsigned tile_mode = info->si_tile_mode_array[index];
2139 unsigned stencil_tile_mode = info->si_tile_mode_array[stencil_index];
2140 unsigned macro_mode = info->cik_macrotile_mode_array[macro_index];
2141
2142 db_depth_info |=
2143 S_02803C_ARRAY_MODE(G_009910_ARRAY_MODE(tile_mode)) |
2144 S_02803C_PIPE_CONFIG(G_009910_PIPE_CONFIG(tile_mode)) |
2145 S_02803C_BANK_WIDTH(G_009990_BANK_WIDTH(macro_mode)) |
2146 S_02803C_BANK_HEIGHT(G_009990_BANK_HEIGHT(macro_mode)) |
2147 S_02803C_MACRO_TILE_ASPECT(G_009990_MACRO_TILE_ASPECT(macro_mode)) |
2148 S_02803C_NUM_BANKS(G_009990_NUM_BANKS(macro_mode));
2149 z_info |= S_028040_TILE_SPLIT(G_009910_TILE_SPLIT(tile_mode));
2150 s_info |= S_028044_TILE_SPLIT(G_009910_TILE_SPLIT(stencil_tile_mode));
2151 } else {
2152 unsigned tile_mode_index = si_tile_mode_index(rtex, level, false);
2153 z_info |= S_028040_TILE_MODE_INDEX(tile_mode_index);
2154 tile_mode_index = si_tile_mode_index(rtex, level, true);
2155 s_info |= S_028044_TILE_MODE_INDEX(tile_mode_index);
2156 }
2157
2158 /* HiZ aka depth buffer htile */
2159 /* use htile only for first level */
2160 if (rtex->htile_buffer && !level) {
2161 z_info |= S_028040_TILE_SURFACE_ENABLE(1) |
2162 S_028040_ALLOW_EXPCLEAR(1);
2163
2164 if (rtex->surface.flags & RADEON_SURF_SBUFFER) {
2165 /* Workaround: For a not yet understood reason, the
2166 * combination of MSAA, fast stencil clear and stencil
2167 * decompress messes with subsequent stencil buffer
2168 * uses. Problem was reproduced on Verde, Bonaire,
2169 * Tonga, and Carrizo.
2170 *
2171 * Disabling EXPCLEAR works around the problem.
2172 *
2173 * Check piglit's arb_texture_multisample-stencil-clear
2174 * test if you want to try changing this.
2175 */
2176 if (rtex->resource.b.b.nr_samples <= 1)
2177 s_info |= S_028044_ALLOW_EXPCLEAR(1);
2178 } else
2179 /* Use all of the htile_buffer for depth if there's no stencil. */
2180 s_info |= S_028044_TILE_STENCIL_DISABLE(1);
2181
2182 uint64_t va = rtex->htile_buffer->gpu_address;
2183 db_htile_data_base = va >> 8;
2184 db_htile_surface = S_028ABC_FULL_CACHE(1);
2185 } else {
2186 db_htile_data_base = 0;
2187 db_htile_surface = 0;
2188 }
2189
2190 assert(levelinfo->nblk_x % 8 == 0 && levelinfo->nblk_y % 8 == 0);
2191
2192 surf->db_depth_view = S_028008_SLICE_START(surf->base.u.tex.first_layer) |
2193 S_028008_SLICE_MAX(surf->base.u.tex.last_layer);
2194 surf->db_htile_data_base = db_htile_data_base;
2195 surf->db_depth_info = db_depth_info;
2196 surf->db_z_info = z_info;
2197 surf->db_stencil_info = s_info;
2198 surf->db_depth_base = z_offs >> 8;
2199 surf->db_stencil_base = s_offs >> 8;
2200 surf->db_depth_size = S_028058_PITCH_TILE_MAX((levelinfo->nblk_x / 8) - 1) |
2201 S_028058_HEIGHT_TILE_MAX((levelinfo->nblk_y / 8) - 1);
2202 surf->db_depth_slice = S_02805C_SLICE_TILE_MAX((levelinfo->nblk_x *
2203 levelinfo->nblk_y) / 64 - 1);
2204 surf->db_htile_surface = db_htile_surface;
2205
2206 surf->depth_initialized = true;
2207 }
2208
2209 void si_dec_framebuffer_counters(const struct pipe_framebuffer_state *state)
2210 {
2211 for (int i = 0; i < state->nr_cbufs; ++i) {
2212 struct r600_surface *surf = NULL;
2213 struct r600_texture *rtex;
2214
2215 if (!state->cbufs[i])
2216 continue;
2217 surf = (struct r600_surface*)state->cbufs[i];
2218 rtex = (struct r600_texture*)surf->base.texture;
2219
2220 p_atomic_dec(&rtex->framebuffers_bound);
2221 }
2222 }
2223
2224 static void si_set_framebuffer_state(struct pipe_context *ctx,
2225 const struct pipe_framebuffer_state *state)
2226 {
2227 struct si_context *sctx = (struct si_context *)ctx;
2228 struct pipe_constant_buffer constbuf = {0};
2229 struct r600_surface *surf = NULL;
2230 struct r600_texture *rtex;
2231 bool old_cb0_is_integer = sctx->framebuffer.cb0_is_integer;
2232 bool old_any_dst_linear = sctx->framebuffer.any_dst_linear;
2233 unsigned old_nr_samples = sctx->framebuffer.nr_samples;
2234 int i;
2235
2236 /* Only flush TC when changing the framebuffer state, because
2237 * the only client not using TC that can change textures is
2238 * the framebuffer.
2239 *
2240 * Flush all CB and DB caches here because all buffers can be used
2241 * for write by both TC (with shader image stores) and CB/DB.
2242 */
2243 sctx->b.flags |= SI_CONTEXT_INV_VMEM_L1 |
2244 SI_CONTEXT_INV_GLOBAL_L2 |
2245 SI_CONTEXT_FLUSH_AND_INV_FRAMEBUFFER |
2246 SI_CONTEXT_CS_PARTIAL_FLUSH;
2247
2248 /* Take the maximum of the old and new count. If the new count is lower,
2249 * dirtying is needed to disable the unbound colorbuffers.
2250 */
2251 sctx->framebuffer.dirty_cbufs |=
2252 (1 << MAX2(sctx->framebuffer.state.nr_cbufs, state->nr_cbufs)) - 1;
2253 sctx->framebuffer.dirty_zsbuf |= sctx->framebuffer.state.zsbuf != state->zsbuf;
2254
2255 si_dec_framebuffer_counters(&sctx->framebuffer.state);
2256 util_copy_framebuffer_state(&sctx->framebuffer.state, state);
2257
2258 sctx->framebuffer.spi_shader_col_format = 0;
2259 sctx->framebuffer.spi_shader_col_format_alpha = 0;
2260 sctx->framebuffer.spi_shader_col_format_blend = 0;
2261 sctx->framebuffer.spi_shader_col_format_blend_alpha = 0;
2262 sctx->framebuffer.color_is_int8 = 0;
2263
2264 sctx->framebuffer.compressed_cb_mask = 0;
2265 sctx->framebuffer.nr_samples = util_framebuffer_get_num_samples(state);
2266 sctx->framebuffer.log_samples = util_logbase2(sctx->framebuffer.nr_samples);
2267 sctx->framebuffer.cb0_is_integer = state->nr_cbufs && state->cbufs[0] &&
2268 util_format_is_pure_integer(state->cbufs[0]->format);
2269 sctx->framebuffer.any_dst_linear = false;
2270
2271 if (sctx->framebuffer.cb0_is_integer != old_cb0_is_integer)
2272 si_mark_atom_dirty(sctx, &sctx->db_render_state);
2273
2274 for (i = 0; i < state->nr_cbufs; i++) {
2275 if (!state->cbufs[i])
2276 continue;
2277
2278 surf = (struct r600_surface*)state->cbufs[i];
2279 rtex = (struct r600_texture*)surf->base.texture;
2280
2281 if (!surf->color_initialized) {
2282 si_initialize_color_surface(sctx, surf);
2283 }
2284
2285 sctx->framebuffer.spi_shader_col_format |=
2286 surf->spi_shader_col_format << (i * 4);
2287 sctx->framebuffer.spi_shader_col_format_alpha |=
2288 surf->spi_shader_col_format_alpha << (i * 4);
2289 sctx->framebuffer.spi_shader_col_format_blend |=
2290 surf->spi_shader_col_format_blend << (i * 4);
2291 sctx->framebuffer.spi_shader_col_format_blend_alpha |=
2292 surf->spi_shader_col_format_blend_alpha << (i * 4);
2293
2294 if (surf->color_is_int8)
2295 sctx->framebuffer.color_is_int8 |= 1 << i;
2296
2297 if (rtex->fmask.size && rtex->cmask.size) {
2298 sctx->framebuffer.compressed_cb_mask |= 1 << i;
2299 }
2300
2301 if (surf->level_info->mode == RADEON_SURF_MODE_LINEAR_ALIGNED)
2302 sctx->framebuffer.any_dst_linear = true;
2303
2304 r600_context_add_resource_size(ctx, surf->base.texture);
2305
2306 p_atomic_inc(&rtex->framebuffers_bound);
2307 }
2308 /* Set the second SPI format for possible dual-src blending. */
2309 if (i == 1 && surf) {
2310 sctx->framebuffer.spi_shader_col_format |=
2311 surf->spi_shader_col_format << (i * 4);
2312 sctx->framebuffer.spi_shader_col_format_alpha |=
2313 surf->spi_shader_col_format_alpha << (i * 4);
2314 sctx->framebuffer.spi_shader_col_format_blend |=
2315 surf->spi_shader_col_format_blend << (i * 4);
2316 sctx->framebuffer.spi_shader_col_format_blend_alpha |=
2317 surf->spi_shader_col_format_blend_alpha << (i * 4);
2318 }
2319
2320 if (state->zsbuf) {
2321 surf = (struct r600_surface*)state->zsbuf;
2322
2323 if (!surf->depth_initialized) {
2324 si_init_depth_surface(sctx, surf);
2325 }
2326 r600_context_add_resource_size(ctx, surf->base.texture);
2327 }
2328
2329 si_update_poly_offset_state(sctx);
2330 si_mark_atom_dirty(sctx, &sctx->cb_render_state);
2331 si_mark_atom_dirty(sctx, &sctx->framebuffer.atom);
2332
2333 if (sctx->framebuffer.any_dst_linear != old_any_dst_linear)
2334 si_mark_atom_dirty(sctx, &sctx->msaa_config);
2335
2336 if (sctx->framebuffer.nr_samples != old_nr_samples) {
2337 si_mark_atom_dirty(sctx, &sctx->msaa_config);
2338 si_mark_atom_dirty(sctx, &sctx->db_render_state);
2339
2340 /* Set sample locations as fragment shader constants. */
2341 switch (sctx->framebuffer.nr_samples) {
2342 case 1:
2343 constbuf.user_buffer = sctx->b.sample_locations_1x;
2344 break;
2345 case 2:
2346 constbuf.user_buffer = sctx->b.sample_locations_2x;
2347 break;
2348 case 4:
2349 constbuf.user_buffer = sctx->b.sample_locations_4x;
2350 break;
2351 case 8:
2352 constbuf.user_buffer = sctx->b.sample_locations_8x;
2353 break;
2354 case 16:
2355 constbuf.user_buffer = sctx->b.sample_locations_16x;
2356 break;
2357 default:
2358 R600_ERR("Requested an invalid number of samples %i.\n",
2359 sctx->framebuffer.nr_samples);
2360 assert(0);
2361 }
2362 constbuf.buffer_size = sctx->framebuffer.nr_samples * 2 * 4;
2363 si_set_rw_buffer(sctx, SI_PS_CONST_SAMPLE_POSITIONS, &constbuf);
2364
2365 /* Smoothing (only possible with nr_samples == 1) uses the same
2366 * sample locations as the MSAA it simulates.
2367 *
2368 * Therefore, don't update the sample locations when
2369 * transitioning from no AA to smoothing-equivalent AA, and
2370 * vice versa.
2371 */
2372 if ((sctx->framebuffer.nr_samples != 1 ||
2373 old_nr_samples != SI_NUM_SMOOTH_AA_SAMPLES) &&
2374 (sctx->framebuffer.nr_samples != SI_NUM_SMOOTH_AA_SAMPLES ||
2375 old_nr_samples != 1))
2376 si_mark_atom_dirty(sctx, &sctx->msaa_sample_locs);
2377 }
2378
2379 sctx->need_check_render_feedback = true;
2380 }
2381
2382 static void si_emit_framebuffer_state(struct si_context *sctx, struct r600_atom *atom)
2383 {
2384 struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
2385 struct pipe_framebuffer_state *state = &sctx->framebuffer.state;
2386 unsigned i, nr_cbufs = state->nr_cbufs;
2387 struct r600_texture *tex = NULL;
2388 struct r600_surface *cb = NULL;
2389 unsigned cb_color_info = 0;
2390
2391 /* Colorbuffers. */
2392 for (i = 0; i < nr_cbufs; i++) {
2393 unsigned pitch_tile_max, slice_tile_max, tile_mode_index;
2394 unsigned cb_color_base, cb_color_fmask, cb_color_attrib;
2395 unsigned cb_color_pitch, cb_color_slice, cb_color_fmask_slice;
2396
2397 if (!(sctx->framebuffer.dirty_cbufs & (1 << i)))
2398 continue;
2399
2400 cb = (struct r600_surface*)state->cbufs[i];
2401 if (!cb) {
2402 radeon_set_context_reg(cs, R_028C70_CB_COLOR0_INFO + i * 0x3C,
2403 S_028C70_FORMAT(V_028C70_COLOR_INVALID));
2404 continue;
2405 }
2406
2407 tex = (struct r600_texture *)cb->base.texture;
2408 radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx,
2409 &tex->resource, RADEON_USAGE_READWRITE,
2410 tex->surface.nsamples > 1 ?
2411 RADEON_PRIO_COLOR_BUFFER_MSAA :
2412 RADEON_PRIO_COLOR_BUFFER);
2413
2414 if (tex->cmask_buffer && tex->cmask_buffer != &tex->resource) {
2415 radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx,
2416 tex->cmask_buffer, RADEON_USAGE_READWRITE,
2417 RADEON_PRIO_CMASK);
2418 }
2419
2420 /* Compute mutable surface parameters. */
2421 pitch_tile_max = cb->level_info->nblk_x / 8 - 1;
2422 slice_tile_max = cb->level_info->nblk_x *
2423 cb->level_info->nblk_y / 64 - 1;
2424 tile_mode_index = si_tile_mode_index(tex, cb->base.u.tex.level, false);
2425
2426 cb_color_base = (tex->resource.gpu_address + cb->level_info->offset) >> 8;
2427 cb_color_pitch = S_028C64_TILE_MAX(pitch_tile_max);
2428 cb_color_slice = S_028C68_TILE_MAX(slice_tile_max);
2429 cb_color_attrib = cb->cb_color_attrib |
2430 S_028C74_TILE_MODE_INDEX(tile_mode_index);
2431
2432 if (tex->fmask.size) {
2433 if (sctx->b.chip_class >= CIK)
2434 cb_color_pitch |= S_028C64_FMASK_TILE_MAX(tex->fmask.pitch_in_pixels / 8 - 1);
2435 cb_color_attrib |= S_028C74_FMASK_TILE_MODE_INDEX(tex->fmask.tile_mode_index);
2436 cb_color_fmask = (tex->resource.gpu_address + tex->fmask.offset) >> 8;
2437 cb_color_fmask_slice = S_028C88_TILE_MAX(tex->fmask.slice_tile_max);
2438 } else {
2439 /* This must be set for fast clear to work without FMASK. */
2440 if (sctx->b.chip_class >= CIK)
2441 cb_color_pitch |= S_028C64_FMASK_TILE_MAX(pitch_tile_max);
2442 cb_color_attrib |= S_028C74_FMASK_TILE_MODE_INDEX(tile_mode_index);
2443 cb_color_fmask = cb_color_base;
2444 cb_color_fmask_slice = S_028C88_TILE_MAX(slice_tile_max);
2445 }
2446
2447 cb_color_info = cb->cb_color_info | tex->cb_color_info;
2448
2449 if (tex->dcc_offset && cb->level_info->dcc_enabled) {
2450 bool is_msaa_resolve_dst = state->cbufs[0] &&
2451 state->cbufs[0]->texture->nr_samples > 1 &&
2452 state->cbufs[1] == &cb->base &&
2453 state->cbufs[1]->texture->nr_samples <= 1;
2454
2455 if (!is_msaa_resolve_dst)
2456 cb_color_info |= S_028C70_DCC_ENABLE(1);
2457 }
2458
2459 radeon_set_context_reg_seq(cs, R_028C60_CB_COLOR0_BASE + i * 0x3C,
2460 sctx->b.chip_class >= VI ? 14 : 13);
2461 radeon_emit(cs, cb_color_base); /* R_028C60_CB_COLOR0_BASE */
2462 radeon_emit(cs, cb_color_pitch); /* R_028C64_CB_COLOR0_PITCH */
2463 radeon_emit(cs, cb_color_slice); /* R_028C68_CB_COLOR0_SLICE */
2464 radeon_emit(cs, cb->cb_color_view); /* R_028C6C_CB_COLOR0_VIEW */
2465 radeon_emit(cs, cb_color_info); /* R_028C70_CB_COLOR0_INFO */
2466 radeon_emit(cs, cb_color_attrib); /* R_028C74_CB_COLOR0_ATTRIB */
2467 radeon_emit(cs, cb->cb_dcc_control); /* R_028C78_CB_COLOR0_DCC_CONTROL */
2468 radeon_emit(cs, tex->cmask.base_address_reg); /* R_028C7C_CB_COLOR0_CMASK */
2469 radeon_emit(cs, tex->cmask.slice_tile_max); /* R_028C80_CB_COLOR0_CMASK_SLICE */
2470 radeon_emit(cs, cb_color_fmask); /* R_028C84_CB_COLOR0_FMASK */
2471 radeon_emit(cs, cb_color_fmask_slice); /* R_028C88_CB_COLOR0_FMASK_SLICE */
2472 radeon_emit(cs, tex->color_clear_value[0]); /* R_028C8C_CB_COLOR0_CLEAR_WORD0 */
2473 radeon_emit(cs, tex->color_clear_value[1]); /* R_028C90_CB_COLOR0_CLEAR_WORD1 */
2474
2475 if (sctx->b.chip_class >= VI) /* R_028C94_CB_COLOR0_DCC_BASE */
2476 radeon_emit(cs, (tex->resource.gpu_address +
2477 tex->dcc_offset +
2478 tex->surface.level[cb->base.u.tex.level].dcc_offset) >> 8);
2479 }
2480 /* set CB_COLOR1_INFO for possible dual-src blending */
2481 if (i == 1 && state->cbufs[0] &&
2482 sctx->framebuffer.dirty_cbufs & (1 << 0)) {
2483 radeon_set_context_reg(cs, R_028C70_CB_COLOR0_INFO + 1 * 0x3C,
2484 cb_color_info);
2485 i++;
2486 }
2487 for (; i < 8 ; i++)
2488 if (sctx->framebuffer.dirty_cbufs & (1 << i))
2489 radeon_set_context_reg(cs, R_028C70_CB_COLOR0_INFO + i * 0x3C, 0);
2490
2491 /* ZS buffer. */
2492 if (state->zsbuf && sctx->framebuffer.dirty_zsbuf) {
2493 struct r600_surface *zb = (struct r600_surface*)state->zsbuf;
2494 struct r600_texture *rtex = (struct r600_texture*)zb->base.texture;
2495
2496 radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx,
2497 &rtex->resource, RADEON_USAGE_READWRITE,
2498 zb->base.texture->nr_samples > 1 ?
2499 RADEON_PRIO_DEPTH_BUFFER_MSAA :
2500 RADEON_PRIO_DEPTH_BUFFER);
2501
2502 if (zb->db_htile_data_base) {
2503 radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx,
2504 rtex->htile_buffer, RADEON_USAGE_READWRITE,
2505 RADEON_PRIO_HTILE);
2506 }
2507
2508 radeon_set_context_reg(cs, R_028008_DB_DEPTH_VIEW, zb->db_depth_view);
2509 radeon_set_context_reg(cs, R_028014_DB_HTILE_DATA_BASE, zb->db_htile_data_base);
2510
2511 radeon_set_context_reg_seq(cs, R_02803C_DB_DEPTH_INFO, 9);
2512 radeon_emit(cs, zb->db_depth_info); /* R_02803C_DB_DEPTH_INFO */
2513 radeon_emit(cs, zb->db_z_info | /* R_028040_DB_Z_INFO */
2514 S_028040_ZRANGE_PRECISION(rtex->depth_clear_value != 0));
2515 radeon_emit(cs, zb->db_stencil_info); /* R_028044_DB_STENCIL_INFO */
2516 radeon_emit(cs, zb->db_depth_base); /* R_028048_DB_Z_READ_BASE */
2517 radeon_emit(cs, zb->db_stencil_base); /* R_02804C_DB_STENCIL_READ_BASE */
2518 radeon_emit(cs, zb->db_depth_base); /* R_028050_DB_Z_WRITE_BASE */
2519 radeon_emit(cs, zb->db_stencil_base); /* R_028054_DB_STENCIL_WRITE_BASE */
2520 radeon_emit(cs, zb->db_depth_size); /* R_028058_DB_DEPTH_SIZE */
2521 radeon_emit(cs, zb->db_depth_slice); /* R_02805C_DB_DEPTH_SLICE */
2522
2523 radeon_set_context_reg_seq(cs, R_028028_DB_STENCIL_CLEAR, 2);
2524 radeon_emit(cs, rtex->stencil_clear_value); /* R_028028_DB_STENCIL_CLEAR */
2525 radeon_emit(cs, fui(rtex->depth_clear_value)); /* R_02802C_DB_DEPTH_CLEAR */
2526
2527 radeon_set_context_reg(cs, R_028ABC_DB_HTILE_SURFACE, zb->db_htile_surface);
2528 } else if (sctx->framebuffer.dirty_zsbuf) {
2529 radeon_set_context_reg_seq(cs, R_028040_DB_Z_INFO, 2);
2530 radeon_emit(cs, S_028040_FORMAT(V_028040_Z_INVALID)); /* R_028040_DB_Z_INFO */
2531 radeon_emit(cs, S_028044_FORMAT(V_028044_STENCIL_INVALID)); /* R_028044_DB_STENCIL_INFO */
2532 }
2533
2534 /* Framebuffer dimensions. */
2535 /* PA_SC_WINDOW_SCISSOR_TL is set in si_init_config() */
2536 radeon_set_context_reg(cs, R_028208_PA_SC_WINDOW_SCISSOR_BR,
2537 S_028208_BR_X(state->width) | S_028208_BR_Y(state->height));
2538
2539 sctx->framebuffer.dirty_cbufs = 0;
2540 sctx->framebuffer.dirty_zsbuf = false;
2541 }
2542
2543 static void si_emit_msaa_sample_locs(struct si_context *sctx,
2544 struct r600_atom *atom)
2545 {
2546 struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
2547 unsigned nr_samples = sctx->framebuffer.nr_samples;
2548
2549 cayman_emit_msaa_sample_locs(cs, nr_samples > 1 ? nr_samples :
2550 SI_NUM_SMOOTH_AA_SAMPLES);
2551 }
2552
2553 static void si_emit_msaa_config(struct si_context *sctx, struct r600_atom *atom)
2554 {
2555 struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
2556 unsigned num_tile_pipes = sctx->screen->b.info.num_tile_pipes;
2557 /* 33% faster rendering to linear color buffers */
2558 bool dst_is_linear = sctx->framebuffer.any_dst_linear;
2559 unsigned sc_mode_cntl_1 =
2560 S_028A4C_WALK_SIZE(dst_is_linear) |
2561 S_028A4C_WALK_FENCE_ENABLE(!dst_is_linear) |
2562 S_028A4C_WALK_FENCE_SIZE(num_tile_pipes == 2 ? 2 : 3) |
2563 /* always 1: */
2564 S_028A4C_WALK_ALIGN8_PRIM_FITS_ST(1) |
2565 S_028A4C_SUPERTILE_WALK_ORDER_ENABLE(1) |
2566 S_028A4C_TILE_WALK_ORDER_ENABLE(1) |
2567 S_028A4C_MULTI_SHADER_ENGINE_PRIM_DISCARD_ENABLE(1) |
2568 S_028A4C_FORCE_EOV_CNTDWN_ENABLE(1) |
2569 S_028A4C_FORCE_EOV_REZ_ENABLE(1);
2570
2571 cayman_emit_msaa_config(cs, sctx->framebuffer.nr_samples,
2572 sctx->ps_iter_samples,
2573 sctx->smoothing_enabled ? SI_NUM_SMOOTH_AA_SAMPLES : 0,
2574 sc_mode_cntl_1);
2575 }
2576
2577 static void si_set_min_samples(struct pipe_context *ctx, unsigned min_samples)
2578 {
2579 struct si_context *sctx = (struct si_context *)ctx;
2580
2581 if (sctx->ps_iter_samples == min_samples)
2582 return;
2583
2584 sctx->ps_iter_samples = min_samples;
2585
2586 if (sctx->framebuffer.nr_samples > 1)
2587 si_mark_atom_dirty(sctx, &sctx->msaa_config);
2588 }
2589
2590 /*
2591 * Samplers
2592 */
2593
2594 /**
2595 * Build the sampler view descriptor for a buffer texture.
2596 * @param state 256-bit descriptor; only the high 128 bits are filled in
2597 */
2598 void
2599 si_make_buffer_descriptor(struct si_screen *screen, struct r600_resource *buf,
2600 enum pipe_format format,
2601 unsigned first_element, unsigned last_element,
2602 uint32_t *state)
2603 {
2604 const struct util_format_description *desc;
2605 int first_non_void;
2606 uint64_t va;
2607 unsigned stride;
2608 unsigned num_records;
2609 unsigned num_format, data_format;
2610
2611 desc = util_format_description(format);
2612 first_non_void = util_format_get_first_non_void_channel(format);
2613 stride = desc->block.bits / 8;
2614 va = buf->gpu_address + first_element * stride;
2615 num_format = si_translate_buffer_numformat(&screen->b.b, desc, first_non_void);
2616 data_format = si_translate_buffer_dataformat(&screen->b.b, desc, first_non_void);
2617
2618 num_records = last_element + 1 - first_element;
2619 num_records = MIN2(num_records, buf->b.b.width0 / stride);
2620
2621 if (screen->b.chip_class >= VI)
2622 num_records *= stride;
2623
2624 state[4] = va;
2625 state[5] = S_008F04_BASE_ADDRESS_HI(va >> 32) |
2626 S_008F04_STRIDE(stride);
2627 state[6] = num_records;
2628 state[7] = S_008F0C_DST_SEL_X(si_map_swizzle(desc->swizzle[0])) |
2629 S_008F0C_DST_SEL_Y(si_map_swizzle(desc->swizzle[1])) |
2630 S_008F0C_DST_SEL_Z(si_map_swizzle(desc->swizzle[2])) |
2631 S_008F0C_DST_SEL_W(si_map_swizzle(desc->swizzle[3])) |
2632 S_008F0C_NUM_FORMAT(num_format) |
2633 S_008F0C_DATA_FORMAT(data_format);
2634 }
2635
2636 /**
2637 * Build the sampler view descriptor for a texture.
2638 */
2639 void
2640 si_make_texture_descriptor(struct si_screen *screen,
2641 struct r600_texture *tex,
2642 bool sampler,
2643 enum pipe_texture_target target,
2644 enum pipe_format pipe_format,
2645 const unsigned char state_swizzle[4],
2646 unsigned first_level, unsigned last_level,
2647 unsigned first_layer, unsigned last_layer,
2648 unsigned width, unsigned height, unsigned depth,
2649 uint32_t *state,
2650 uint32_t *fmask_state)
2651 {
2652 struct pipe_resource *res = &tex->resource.b.b;
2653 const struct util_format_description *desc;
2654 unsigned char swizzle[4];
2655 int first_non_void;
2656 unsigned num_format, data_format, type;
2657 uint64_t va;
2658
2659 desc = util_format_description(pipe_format);
2660
2661 if (desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS) {
2662 const unsigned char swizzle_xxxx[4] = {0, 0, 0, 0};
2663 const unsigned char swizzle_yyyy[4] = {1, 1, 1, 1};
2664
2665 switch (pipe_format) {
2666 case PIPE_FORMAT_S8_UINT_Z24_UNORM:
2667 case PIPE_FORMAT_X24S8_UINT:
2668 case PIPE_FORMAT_X32_S8X24_UINT:
2669 case PIPE_FORMAT_X8Z24_UNORM:
2670 util_format_compose_swizzles(swizzle_yyyy, state_swizzle, swizzle);
2671 break;
2672 default:
2673 util_format_compose_swizzles(swizzle_xxxx, state_swizzle, swizzle);
2674 }
2675 } else {
2676 util_format_compose_swizzles(desc->swizzle, state_swizzle, swizzle);
2677 }
2678
2679 first_non_void = util_format_get_first_non_void_channel(pipe_format);
2680
2681 switch (pipe_format) {
2682 case PIPE_FORMAT_S8_UINT_Z24_UNORM:
2683 num_format = V_008F14_IMG_NUM_FORMAT_UNORM;
2684 break;
2685 default:
2686 if (first_non_void < 0) {
2687 if (util_format_is_compressed(pipe_format)) {
2688 switch (pipe_format) {
2689 case PIPE_FORMAT_DXT1_SRGB:
2690 case PIPE_FORMAT_DXT1_SRGBA:
2691 case PIPE_FORMAT_DXT3_SRGBA:
2692 case PIPE_FORMAT_DXT5_SRGBA:
2693 case PIPE_FORMAT_BPTC_SRGBA:
2694 case PIPE_FORMAT_ETC2_SRGB8:
2695 case PIPE_FORMAT_ETC2_SRGB8A1:
2696 case PIPE_FORMAT_ETC2_SRGBA8:
2697 num_format = V_008F14_IMG_NUM_FORMAT_SRGB;
2698 break;
2699 case PIPE_FORMAT_RGTC1_SNORM:
2700 case PIPE_FORMAT_LATC1_SNORM:
2701 case PIPE_FORMAT_RGTC2_SNORM:
2702 case PIPE_FORMAT_LATC2_SNORM:
2703 case PIPE_FORMAT_ETC2_R11_SNORM:
2704 case PIPE_FORMAT_ETC2_RG11_SNORM:
2705 /* implies float, so use SNORM/UNORM to determine
2706 whether data is signed or not */
2707 case PIPE_FORMAT_BPTC_RGB_FLOAT:
2708 num_format = V_008F14_IMG_NUM_FORMAT_SNORM;
2709 break;
2710 default:
2711 num_format = V_008F14_IMG_NUM_FORMAT_UNORM;
2712 break;
2713 }
2714 } else if (desc->layout == UTIL_FORMAT_LAYOUT_SUBSAMPLED) {
2715 num_format = V_008F14_IMG_NUM_FORMAT_UNORM;
2716 } else {
2717 num_format = V_008F14_IMG_NUM_FORMAT_FLOAT;
2718 }
2719 } else if (desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
2720 num_format = V_008F14_IMG_NUM_FORMAT_SRGB;
2721 } else {
2722 num_format = V_008F14_IMG_NUM_FORMAT_UNORM;
2723
2724 switch (desc->channel[first_non_void].type) {
2725 case UTIL_FORMAT_TYPE_FLOAT:
2726 num_format = V_008F14_IMG_NUM_FORMAT_FLOAT;
2727 break;
2728 case UTIL_FORMAT_TYPE_SIGNED:
2729 if (desc->channel[first_non_void].normalized)
2730 num_format = V_008F14_IMG_NUM_FORMAT_SNORM;
2731 else if (desc->channel[first_non_void].pure_integer)
2732 num_format = V_008F14_IMG_NUM_FORMAT_SINT;
2733 else
2734 num_format = V_008F14_IMG_NUM_FORMAT_SSCALED;
2735 break;
2736 case UTIL_FORMAT_TYPE_UNSIGNED:
2737 if (desc->channel[first_non_void].normalized)
2738 num_format = V_008F14_IMG_NUM_FORMAT_UNORM;
2739 else if (desc->channel[first_non_void].pure_integer)
2740 num_format = V_008F14_IMG_NUM_FORMAT_UINT;
2741 else
2742 num_format = V_008F14_IMG_NUM_FORMAT_USCALED;
2743 }
2744 }
2745 }
2746
2747 data_format = si_translate_texformat(&screen->b.b, pipe_format, desc, first_non_void);
2748 if (data_format == ~0) {
2749 data_format = 0;
2750 }
2751
2752 if (!sampler &&
2753 (res->target == PIPE_TEXTURE_CUBE ||
2754 res->target == PIPE_TEXTURE_CUBE_ARRAY ||
2755 res->target == PIPE_TEXTURE_3D)) {
2756 /* For the purpose of shader images, treat cube maps and 3D
2757 * textures as 2D arrays. For 3D textures, the address
2758 * calculations for mipmaps are different, so we rely on the
2759 * caller to effectively disable mipmaps.
2760 */
2761 type = V_008F1C_SQ_RSRC_IMG_2D_ARRAY;
2762
2763 assert(res->target != PIPE_TEXTURE_3D || (first_level == 0 && last_level == 0));
2764 } else {
2765 type = si_tex_dim(res->target, target, res->nr_samples);
2766 }
2767
2768 if (type == V_008F1C_SQ_RSRC_IMG_1D_ARRAY) {
2769 height = 1;
2770 depth = res->array_size;
2771 } else if (type == V_008F1C_SQ_RSRC_IMG_2D_ARRAY ||
2772 type == V_008F1C_SQ_RSRC_IMG_2D_MSAA_ARRAY) {
2773 if (sampler || res->target != PIPE_TEXTURE_3D)
2774 depth = res->array_size;
2775 } else if (type == V_008F1C_SQ_RSRC_IMG_CUBE)
2776 depth = res->array_size / 6;
2777
2778 state[0] = 0;
2779 state[1] = (S_008F14_DATA_FORMAT(data_format) |
2780 S_008F14_NUM_FORMAT(num_format));
2781 state[2] = (S_008F18_WIDTH(width - 1) |
2782 S_008F18_HEIGHT(height - 1));
2783 state[3] = (S_008F1C_DST_SEL_X(si_map_swizzle(swizzle[0])) |
2784 S_008F1C_DST_SEL_Y(si_map_swizzle(swizzle[1])) |
2785 S_008F1C_DST_SEL_Z(si_map_swizzle(swizzle[2])) |
2786 S_008F1C_DST_SEL_W(si_map_swizzle(swizzle[3])) |
2787 S_008F1C_BASE_LEVEL(res->nr_samples > 1 ?
2788 0 : first_level) |
2789 S_008F1C_LAST_LEVEL(res->nr_samples > 1 ?
2790 util_logbase2(res->nr_samples) :
2791 last_level) |
2792 S_008F1C_POW2_PAD(res->last_level > 0) |
2793 S_008F1C_TYPE(type));
2794 state[4] = S_008F20_DEPTH(depth - 1);
2795 state[5] = (S_008F24_BASE_ARRAY(first_layer) |
2796 S_008F24_LAST_ARRAY(last_layer));
2797 state[6] = 0;
2798 state[7] = 0;
2799
2800 if (tex->dcc_offset) {
2801 unsigned swap = r600_translate_colorswap(pipe_format, FALSE);
2802
2803 state[6] = S_008F28_ALPHA_IS_ON_MSB(swap <= 1);
2804 } else {
2805 /* The last dword is unused by hw. The shader uses it to clear
2806 * bits in the first dword of sampler state.
2807 */
2808 if (screen->b.chip_class <= CIK && res->nr_samples <= 1) {
2809 if (first_level == last_level)
2810 state[7] = C_008F30_MAX_ANISO_RATIO;
2811 else
2812 state[7] = 0xffffffff;
2813 }
2814 }
2815
2816 /* Initialize the sampler view for FMASK. */
2817 if (tex->fmask.size) {
2818 uint32_t fmask_format;
2819
2820 va = tex->resource.gpu_address + tex->fmask.offset;
2821
2822 switch (res->nr_samples) {
2823 case 2:
2824 fmask_format = V_008F14_IMG_DATA_FORMAT_FMASK8_S2_F2;
2825 break;
2826 case 4:
2827 fmask_format = V_008F14_IMG_DATA_FORMAT_FMASK8_S4_F4;
2828 break;
2829 case 8:
2830 fmask_format = V_008F14_IMG_DATA_FORMAT_FMASK32_S8_F8;
2831 break;
2832 default:
2833 assert(0);
2834 fmask_format = V_008F14_IMG_DATA_FORMAT_INVALID;
2835 }
2836
2837 fmask_state[0] = va >> 8;
2838 fmask_state[1] = S_008F14_BASE_ADDRESS_HI(va >> 40) |
2839 S_008F14_DATA_FORMAT(fmask_format) |
2840 S_008F14_NUM_FORMAT(V_008F14_IMG_NUM_FORMAT_UINT);
2841 fmask_state[2] = S_008F18_WIDTH(width - 1) |
2842 S_008F18_HEIGHT(height - 1);
2843 fmask_state[3] = S_008F1C_DST_SEL_X(V_008F1C_SQ_SEL_X) |
2844 S_008F1C_DST_SEL_Y(V_008F1C_SQ_SEL_X) |
2845 S_008F1C_DST_SEL_Z(V_008F1C_SQ_SEL_X) |
2846 S_008F1C_DST_SEL_W(V_008F1C_SQ_SEL_X) |
2847 S_008F1C_TILING_INDEX(tex->fmask.tile_mode_index) |
2848 S_008F1C_TYPE(si_tex_dim(res->target, target, 0));
2849 fmask_state[4] = S_008F20_DEPTH(depth - 1) |
2850 S_008F20_PITCH(tex->fmask.pitch_in_pixels - 1);
2851 fmask_state[5] = S_008F24_BASE_ARRAY(first_layer) |
2852 S_008F24_LAST_ARRAY(last_layer);
2853 fmask_state[6] = 0;
2854 fmask_state[7] = 0;
2855 }
2856 }
2857
2858 /**
2859 * Create a sampler view.
2860 *
2861 * @param ctx context
2862 * @param texture texture
2863 * @param state sampler view template
2864 * @param width0 width0 override (for compressed textures as int)
2865 * @param height0 height0 override (for compressed textures as int)
2866 * @param force_level set the base address to the level (for compressed textures)
2867 */
2868 struct pipe_sampler_view *
2869 si_create_sampler_view_custom(struct pipe_context *ctx,
2870 struct pipe_resource *texture,
2871 const struct pipe_sampler_view *state,
2872 unsigned width0, unsigned height0,
2873 unsigned force_level)
2874 {
2875 struct si_context *sctx = (struct si_context*)ctx;
2876 struct si_sampler_view *view = CALLOC_STRUCT(si_sampler_view);
2877 struct r600_texture *tmp = (struct r600_texture*)texture;
2878 unsigned base_level, first_level, last_level;
2879 unsigned char state_swizzle[4];
2880 unsigned height, depth, width;
2881 unsigned last_layer = state->u.tex.last_layer;
2882 enum pipe_format pipe_format;
2883 const struct radeon_surf_level *surflevel;
2884
2885 if (!view)
2886 return NULL;
2887
2888 /* initialize base object */
2889 view->base = *state;
2890 view->base.texture = NULL;
2891 view->base.reference.count = 1;
2892 view->base.context = ctx;
2893
2894 /* NULL resource, obey swizzle (only ZERO and ONE make sense). */
2895 if (!texture) {
2896 view->state[3] = S_008F1C_DST_SEL_X(si_map_swizzle(state->swizzle_r)) |
2897 S_008F1C_DST_SEL_Y(si_map_swizzle(state->swizzle_g)) |
2898 S_008F1C_DST_SEL_Z(si_map_swizzle(state->swizzle_b)) |
2899 S_008F1C_DST_SEL_W(si_map_swizzle(state->swizzle_a)) |
2900 S_008F1C_TYPE(V_008F1C_SQ_RSRC_IMG_1D);
2901 return &view->base;
2902 }
2903
2904 pipe_resource_reference(&view->base.texture, texture);
2905
2906 if (state->format == PIPE_FORMAT_X24S8_UINT ||
2907 state->format == PIPE_FORMAT_S8X24_UINT ||
2908 state->format == PIPE_FORMAT_X32_S8X24_UINT ||
2909 state->format == PIPE_FORMAT_S8_UINT)
2910 view->is_stencil_sampler = true;
2911
2912 /* Buffer resource. */
2913 if (texture->target == PIPE_BUFFER) {
2914 si_make_buffer_descriptor(sctx->screen,
2915 (struct r600_resource *)texture,
2916 state->format,
2917 state->u.buf.first_element,
2918 state->u.buf.last_element,
2919 view->state);
2920
2921 LIST_ADDTAIL(&view->list, &sctx->b.texture_buffers);
2922 return &view->base;
2923 }
2924
2925 state_swizzle[0] = state->swizzle_r;
2926 state_swizzle[1] = state->swizzle_g;
2927 state_swizzle[2] = state->swizzle_b;
2928 state_swizzle[3] = state->swizzle_a;
2929
2930 base_level = 0;
2931 first_level = state->u.tex.first_level;
2932 last_level = state->u.tex.last_level;
2933 width = width0;
2934 height = height0;
2935 depth = texture->depth0;
2936
2937 if (force_level) {
2938 assert(force_level == first_level &&
2939 force_level == last_level);
2940 base_level = force_level;
2941 first_level = 0;
2942 last_level = 0;
2943 width = u_minify(width, force_level);
2944 height = u_minify(height, force_level);
2945 depth = u_minify(depth, force_level);
2946 }
2947
2948 /* This is not needed if state trackers set last_layer correctly. */
2949 if (state->target == PIPE_TEXTURE_1D ||
2950 state->target == PIPE_TEXTURE_2D ||
2951 state->target == PIPE_TEXTURE_RECT ||
2952 state->target == PIPE_TEXTURE_CUBE)
2953 last_layer = state->u.tex.first_layer;
2954
2955 /* Texturing with separate depth and stencil. */
2956 pipe_format = state->format;
2957 surflevel = tmp->surface.level;
2958
2959 if (tmp->is_depth && !tmp->is_flushing_texture) {
2960 switch (pipe_format) {
2961 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
2962 pipe_format = PIPE_FORMAT_Z32_FLOAT;
2963 break;
2964 case PIPE_FORMAT_X8Z24_UNORM:
2965 case PIPE_FORMAT_S8_UINT_Z24_UNORM:
2966 /* Z24 is always stored like this. */
2967 pipe_format = PIPE_FORMAT_Z24X8_UNORM;
2968 break;
2969 case PIPE_FORMAT_X24S8_UINT:
2970 case PIPE_FORMAT_S8X24_UINT:
2971 case PIPE_FORMAT_X32_S8X24_UINT:
2972 pipe_format = PIPE_FORMAT_S8_UINT;
2973 surflevel = tmp->surface.stencil_level;
2974 break;
2975 default:;
2976 }
2977 }
2978
2979 si_make_texture_descriptor(sctx->screen, tmp, true,
2980 state->target, pipe_format, state_swizzle,
2981 first_level, last_level,
2982 state->u.tex.first_layer, last_layer,
2983 width, height, depth,
2984 view->state, view->fmask_state);
2985
2986 view->base_level_info = &surflevel[base_level];
2987 view->base_level = base_level;
2988 view->block_width = util_format_get_blockwidth(pipe_format);
2989 return &view->base;
2990 }
2991
2992 static struct pipe_sampler_view *
2993 si_create_sampler_view(struct pipe_context *ctx,
2994 struct pipe_resource *texture,
2995 const struct pipe_sampler_view *state)
2996 {
2997 return si_create_sampler_view_custom(ctx, texture, state,
2998 texture ? texture->width0 : 0,
2999 texture ? texture->height0 : 0, 0);
3000 }
3001
3002 static void si_sampler_view_destroy(struct pipe_context *ctx,
3003 struct pipe_sampler_view *state)
3004 {
3005 struct si_sampler_view *view = (struct si_sampler_view *)state;
3006
3007 if (state->texture && state->texture->target == PIPE_BUFFER)
3008 LIST_DELINIT(&view->list);
3009
3010 pipe_resource_reference(&state->texture, NULL);
3011 FREE(view);
3012 }
3013
3014 static bool wrap_mode_uses_border_color(unsigned wrap, bool linear_filter)
3015 {
3016 return wrap == PIPE_TEX_WRAP_CLAMP_TO_BORDER ||
3017 wrap == PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER ||
3018 (linear_filter &&
3019 (wrap == PIPE_TEX_WRAP_CLAMP ||
3020 wrap == PIPE_TEX_WRAP_MIRROR_CLAMP));
3021 }
3022
3023 static bool sampler_state_needs_border_color(const struct pipe_sampler_state *state)
3024 {
3025 bool linear_filter = state->min_img_filter != PIPE_TEX_FILTER_NEAREST ||
3026 state->mag_img_filter != PIPE_TEX_FILTER_NEAREST;
3027
3028 return (state->border_color.ui[0] || state->border_color.ui[1] ||
3029 state->border_color.ui[2] || state->border_color.ui[3]) &&
3030 (wrap_mode_uses_border_color(state->wrap_s, linear_filter) ||
3031 wrap_mode_uses_border_color(state->wrap_t, linear_filter) ||
3032 wrap_mode_uses_border_color(state->wrap_r, linear_filter));
3033 }
3034
3035 static void *si_create_sampler_state(struct pipe_context *ctx,
3036 const struct pipe_sampler_state *state)
3037 {
3038 struct si_context *sctx = (struct si_context *)ctx;
3039 struct r600_common_screen *rscreen = sctx->b.screen;
3040 struct si_sampler_state *rstate = CALLOC_STRUCT(si_sampler_state);
3041 unsigned border_color_type, border_color_index = 0;
3042 unsigned max_aniso = rscreen->force_aniso >= 0 ? rscreen->force_aniso
3043 : state->max_anisotropy;
3044 unsigned max_aniso_ratio = r600_tex_aniso_filter(max_aniso);
3045
3046 if (!rstate) {
3047 return NULL;
3048 }
3049
3050 if (!sampler_state_needs_border_color(state))
3051 border_color_type = V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK;
3052 else if (state->border_color.f[0] == 0 &&
3053 state->border_color.f[1] == 0 &&
3054 state->border_color.f[2] == 0 &&
3055 state->border_color.f[3] == 0)
3056 border_color_type = V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK;
3057 else if (state->border_color.f[0] == 0 &&
3058 state->border_color.f[1] == 0 &&
3059 state->border_color.f[2] == 0 &&
3060 state->border_color.f[3] == 1)
3061 border_color_type = V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_BLACK;
3062 else if (state->border_color.f[0] == 1 &&
3063 state->border_color.f[1] == 1 &&
3064 state->border_color.f[2] == 1 &&
3065 state->border_color.f[3] == 1)
3066 border_color_type = V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_WHITE;
3067 else {
3068 int i;
3069
3070 border_color_type = V_008F3C_SQ_TEX_BORDER_COLOR_REGISTER;
3071
3072 /* Check if the border has been uploaded already. */
3073 for (i = 0; i < sctx->border_color_count; i++)
3074 if (memcmp(&sctx->border_color_table[i], &state->border_color,
3075 sizeof(state->border_color)) == 0)
3076 break;
3077
3078 if (i >= SI_MAX_BORDER_COLORS) {
3079 /* Getting 4096 unique border colors is very unlikely. */
3080 fprintf(stderr, "radeonsi: The border color table is full. "
3081 "Any new border colors will be just black. "
3082 "Please file a bug.\n");
3083 border_color_type = V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK;
3084 } else {
3085 if (i == sctx->border_color_count) {
3086 /* Upload a new border color. */
3087 memcpy(&sctx->border_color_table[i], &state->border_color,
3088 sizeof(state->border_color));
3089 util_memcpy_cpu_to_le32(&sctx->border_color_map[i],
3090 &state->border_color,
3091 sizeof(state->border_color));
3092 sctx->border_color_count++;
3093 }
3094
3095 border_color_index = i;
3096 }
3097 }
3098
3099 rstate->val[0] = (S_008F30_CLAMP_X(si_tex_wrap(state->wrap_s)) |
3100 S_008F30_CLAMP_Y(si_tex_wrap(state->wrap_t)) |
3101 S_008F30_CLAMP_Z(si_tex_wrap(state->wrap_r)) |
3102 S_008F30_MAX_ANISO_RATIO(max_aniso_ratio) |
3103 S_008F30_DEPTH_COMPARE_FUNC(si_tex_compare(state->compare_func)) |
3104 S_008F30_FORCE_UNNORMALIZED(!state->normalized_coords) |
3105 S_008F30_DISABLE_CUBE_WRAP(!state->seamless_cube_map) |
3106 S_008F30_COMPAT_MODE(sctx->b.chip_class >= VI));
3107 rstate->val[1] = (S_008F34_MIN_LOD(S_FIXED(CLAMP(state->min_lod, 0, 15), 8)) |
3108 S_008F34_MAX_LOD(S_FIXED(CLAMP(state->max_lod, 0, 15), 8)));
3109 rstate->val[2] = (S_008F38_LOD_BIAS(S_FIXED(CLAMP(state->lod_bias, -16, 16), 8)) |
3110 S_008F38_XY_MAG_FILTER(eg_tex_filter(state->mag_img_filter, max_aniso)) |
3111 S_008F38_XY_MIN_FILTER(eg_tex_filter(state->min_img_filter, max_aniso)) |
3112 S_008F38_MIP_FILTER(si_tex_mipfilter(state->min_mip_filter)) |
3113 S_008F38_MIP_POINT_PRECLAMP(1) |
3114 S_008F38_DISABLE_LSB_CEIL(1) |
3115 S_008F38_FILTER_PREC_FIX(1) |
3116 S_008F38_ANISO_OVERRIDE(sctx->b.chip_class >= VI));
3117 rstate->val[3] = S_008F3C_BORDER_COLOR_PTR(border_color_index) |
3118 S_008F3C_BORDER_COLOR_TYPE(border_color_type);
3119 return rstate;
3120 }
3121
3122 static void si_set_sample_mask(struct pipe_context *ctx, unsigned sample_mask)
3123 {
3124 struct si_context *sctx = (struct si_context *)ctx;
3125
3126 if (sctx->sample_mask.sample_mask == (uint16_t)sample_mask)
3127 return;
3128
3129 sctx->sample_mask.sample_mask = sample_mask;
3130 si_mark_atom_dirty(sctx, &sctx->sample_mask.atom);
3131 }
3132
3133 static void si_emit_sample_mask(struct si_context *sctx, struct r600_atom *atom)
3134 {
3135 struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
3136 unsigned mask = sctx->sample_mask.sample_mask;
3137
3138 radeon_set_context_reg_seq(cs, R_028C38_PA_SC_AA_MASK_X0Y0_X1Y0, 2);
3139 radeon_emit(cs, mask | (mask << 16));
3140 radeon_emit(cs, mask | (mask << 16));
3141 }
3142
3143 static void si_delete_sampler_state(struct pipe_context *ctx, void *state)
3144 {
3145 free(state);
3146 }
3147
3148 /*
3149 * Vertex elements & buffers
3150 */
3151
3152 static void *si_create_vertex_elements(struct pipe_context *ctx,
3153 unsigned count,
3154 const struct pipe_vertex_element *elements)
3155 {
3156 struct si_vertex_element *v = CALLOC_STRUCT(si_vertex_element);
3157 int i;
3158
3159 assert(count <= SI_MAX_ATTRIBS);
3160 if (!v)
3161 return NULL;
3162
3163 v->count = count;
3164 for (i = 0; i < count; ++i) {
3165 const struct util_format_description *desc;
3166 unsigned data_format, num_format;
3167 int first_non_void;
3168
3169 desc = util_format_description(elements[i].src_format);
3170 first_non_void = util_format_get_first_non_void_channel(elements[i].src_format);
3171 data_format = si_translate_buffer_dataformat(ctx->screen, desc, first_non_void);
3172 num_format = si_translate_buffer_numformat(ctx->screen, desc, first_non_void);
3173
3174 v->rsrc_word3[i] = S_008F0C_DST_SEL_X(si_map_swizzle(desc->swizzle[0])) |
3175 S_008F0C_DST_SEL_Y(si_map_swizzle(desc->swizzle[1])) |
3176 S_008F0C_DST_SEL_Z(si_map_swizzle(desc->swizzle[2])) |
3177 S_008F0C_DST_SEL_W(si_map_swizzle(desc->swizzle[3])) |
3178 S_008F0C_NUM_FORMAT(num_format) |
3179 S_008F0C_DATA_FORMAT(data_format);
3180 v->format_size[i] = desc->block.bits / 8;
3181 }
3182 memcpy(v->elements, elements, sizeof(struct pipe_vertex_element) * count);
3183
3184 return v;
3185 }
3186
3187 static void si_bind_vertex_elements(struct pipe_context *ctx, void *state)
3188 {
3189 struct si_context *sctx = (struct si_context *)ctx;
3190 struct si_vertex_element *v = (struct si_vertex_element*)state;
3191
3192 sctx->vertex_elements = v;
3193 sctx->vertex_buffers_dirty = true;
3194 }
3195
3196 static void si_delete_vertex_element(struct pipe_context *ctx, void *state)
3197 {
3198 struct si_context *sctx = (struct si_context *)ctx;
3199
3200 if (sctx->vertex_elements == state)
3201 sctx->vertex_elements = NULL;
3202 FREE(state);
3203 }
3204
3205 static void si_set_vertex_buffers(struct pipe_context *ctx,
3206 unsigned start_slot, unsigned count,
3207 const struct pipe_vertex_buffer *buffers)
3208 {
3209 struct si_context *sctx = (struct si_context *)ctx;
3210 struct pipe_vertex_buffer *dst = sctx->vertex_buffer + start_slot;
3211 int i;
3212
3213 assert(start_slot + count <= ARRAY_SIZE(sctx->vertex_buffer));
3214
3215 if (buffers) {
3216 for (i = 0; i < count; i++) {
3217 const struct pipe_vertex_buffer *src = buffers + i;
3218 struct pipe_vertex_buffer *dsti = dst + i;
3219
3220 pipe_resource_reference(&dsti->buffer, src->buffer);
3221 dsti->buffer_offset = src->buffer_offset;
3222 dsti->stride = src->stride;
3223 r600_context_add_resource_size(ctx, src->buffer);
3224 }
3225 } else {
3226 for (i = 0; i < count; i++) {
3227 pipe_resource_reference(&dst[i].buffer, NULL);
3228 }
3229 }
3230 sctx->vertex_buffers_dirty = true;
3231 }
3232
3233 static void si_set_index_buffer(struct pipe_context *ctx,
3234 const struct pipe_index_buffer *ib)
3235 {
3236 struct si_context *sctx = (struct si_context *)ctx;
3237
3238 if (ib) {
3239 pipe_resource_reference(&sctx->index_buffer.buffer, ib->buffer);
3240 memcpy(&sctx->index_buffer, ib, sizeof(*ib));
3241 r600_context_add_resource_size(ctx, ib->buffer);
3242 } else {
3243 pipe_resource_reference(&sctx->index_buffer.buffer, NULL);
3244 }
3245 }
3246
3247 /*
3248 * Misc
3249 */
3250
3251 static void si_set_tess_state(struct pipe_context *ctx,
3252 const float default_outer_level[4],
3253 const float default_inner_level[2])
3254 {
3255 struct si_context *sctx = (struct si_context *)ctx;
3256 struct pipe_constant_buffer cb;
3257 float array[8];
3258
3259 memcpy(array, default_outer_level, sizeof(float) * 4);
3260 memcpy(array+4, default_inner_level, sizeof(float) * 2);
3261
3262 cb.buffer = NULL;
3263 cb.user_buffer = NULL;
3264 cb.buffer_size = sizeof(array);
3265
3266 si_upload_const_buffer(sctx, (struct r600_resource**)&cb.buffer,
3267 (void*)array, sizeof(array),
3268 &cb.buffer_offset);
3269
3270 si_set_rw_buffer(sctx, SI_HS_CONST_DEFAULT_TESS_LEVELS, &cb);
3271 pipe_resource_reference(&cb.buffer, NULL);
3272 }
3273
3274 static void si_texture_barrier(struct pipe_context *ctx)
3275 {
3276 struct si_context *sctx = (struct si_context *)ctx;
3277
3278 sctx->b.flags |= SI_CONTEXT_INV_VMEM_L1 |
3279 SI_CONTEXT_INV_GLOBAL_L2 |
3280 SI_CONTEXT_FLUSH_AND_INV_CB |
3281 SI_CONTEXT_CS_PARTIAL_FLUSH;
3282 }
3283
3284 static void si_memory_barrier(struct pipe_context *ctx, unsigned flags)
3285 {
3286 struct si_context *sctx = (struct si_context *)ctx;
3287
3288 /* Subsequent commands must wait for all shader invocations to
3289 * complete. */
3290 sctx->b.flags |= SI_CONTEXT_PS_PARTIAL_FLUSH |
3291 SI_CONTEXT_CS_PARTIAL_FLUSH;
3292
3293 if (flags & PIPE_BARRIER_CONSTANT_BUFFER)
3294 sctx->b.flags |= SI_CONTEXT_INV_SMEM_L1 |
3295 SI_CONTEXT_INV_VMEM_L1;
3296
3297 if (flags & (PIPE_BARRIER_VERTEX_BUFFER |
3298 PIPE_BARRIER_SHADER_BUFFER |
3299 PIPE_BARRIER_TEXTURE |
3300 PIPE_BARRIER_IMAGE |
3301 PIPE_BARRIER_STREAMOUT_BUFFER |
3302 PIPE_BARRIER_GLOBAL_BUFFER)) {
3303 /* As far as I can tell, L1 contents are written back to L2
3304 * automatically at end of shader, but the contents of other
3305 * L1 caches might still be stale. */
3306 sctx->b.flags |= SI_CONTEXT_INV_VMEM_L1;
3307 }
3308
3309 if (flags & PIPE_BARRIER_INDEX_BUFFER) {
3310 sctx->b.flags |= SI_CONTEXT_INV_VMEM_L1;
3311
3312 /* Indices are read through TC L2 since VI. */
3313 if (sctx->screen->b.chip_class <= CIK)
3314 sctx->b.flags |= SI_CONTEXT_INV_GLOBAL_L2;
3315 }
3316
3317 if (flags & PIPE_BARRIER_FRAMEBUFFER)
3318 sctx->b.flags |= SI_CONTEXT_FLUSH_AND_INV_FRAMEBUFFER;
3319
3320 if (flags & (PIPE_BARRIER_MAPPED_BUFFER |
3321 PIPE_BARRIER_FRAMEBUFFER |
3322 PIPE_BARRIER_INDIRECT_BUFFER)) {
3323 /* Not sure if INV_GLOBAL_L2 is the best thing here.
3324 *
3325 * We need to make sure that TC L1 & L2 are written back to
3326 * memory, because neither CPU accesses nor CB fetches consider
3327 * TC, but there's no need to invalidate any TC cache lines. */
3328 sctx->b.flags |= SI_CONTEXT_INV_GLOBAL_L2;
3329 }
3330 }
3331
3332 static void *si_create_blend_custom(struct si_context *sctx, unsigned mode)
3333 {
3334 struct pipe_blend_state blend;
3335
3336 memset(&blend, 0, sizeof(blend));
3337 blend.independent_blend_enable = true;
3338 blend.rt[0].colormask = 0xf;
3339 return si_create_blend_state_mode(&sctx->b.b, &blend, mode);
3340 }
3341
3342 static void si_need_gfx_cs_space(struct pipe_context *ctx, unsigned num_dw,
3343 bool include_draw_vbo)
3344 {
3345 si_need_cs_space((struct si_context*)ctx);
3346 }
3347
3348 static void si_init_config(struct si_context *sctx);
3349
3350 void si_init_state_functions(struct si_context *sctx)
3351 {
3352 si_init_external_atom(sctx, &sctx->b.render_cond_atom, &sctx->atoms.s.render_cond);
3353 si_init_external_atom(sctx, &sctx->b.streamout.begin_atom, &sctx->atoms.s.streamout_begin);
3354 si_init_external_atom(sctx, &sctx->b.streamout.enable_atom, &sctx->atoms.s.streamout_enable);
3355 si_init_external_atom(sctx, &sctx->b.scissors.atom, &sctx->atoms.s.scissors);
3356 si_init_external_atom(sctx, &sctx->b.viewports.atom, &sctx->atoms.s.viewports);
3357
3358 si_init_atom(sctx, &sctx->cache_flush, &sctx->atoms.s.cache_flush, si_emit_cache_flush);
3359 si_init_atom(sctx, &sctx->framebuffer.atom, &sctx->atoms.s.framebuffer, si_emit_framebuffer_state);
3360 si_init_atom(sctx, &sctx->msaa_sample_locs, &sctx->atoms.s.msaa_sample_locs, si_emit_msaa_sample_locs);
3361 si_init_atom(sctx, &sctx->db_render_state, &sctx->atoms.s.db_render_state, si_emit_db_render_state);
3362 si_init_atom(sctx, &sctx->msaa_config, &sctx->atoms.s.msaa_config, si_emit_msaa_config);
3363 si_init_atom(sctx, &sctx->sample_mask.atom, &sctx->atoms.s.sample_mask, si_emit_sample_mask);
3364 si_init_atom(sctx, &sctx->cb_render_state, &sctx->atoms.s.cb_render_state, si_emit_cb_render_state);
3365 si_init_atom(sctx, &sctx->blend_color.atom, &sctx->atoms.s.blend_color, si_emit_blend_color);
3366 si_init_atom(sctx, &sctx->clip_regs, &sctx->atoms.s.clip_regs, si_emit_clip_regs);
3367 si_init_atom(sctx, &sctx->clip_state.atom, &sctx->atoms.s.clip_state, si_emit_clip_state);
3368 si_init_atom(sctx, &sctx->stencil_ref.atom, &sctx->atoms.s.stencil_ref, si_emit_stencil_ref);
3369
3370 sctx->b.b.create_blend_state = si_create_blend_state;
3371 sctx->b.b.bind_blend_state = si_bind_blend_state;
3372 sctx->b.b.delete_blend_state = si_delete_blend_state;
3373 sctx->b.b.set_blend_color = si_set_blend_color;
3374
3375 sctx->b.b.create_rasterizer_state = si_create_rs_state;
3376 sctx->b.b.bind_rasterizer_state = si_bind_rs_state;
3377 sctx->b.b.delete_rasterizer_state = si_delete_rs_state;
3378
3379 sctx->b.b.create_depth_stencil_alpha_state = si_create_dsa_state;
3380 sctx->b.b.bind_depth_stencil_alpha_state = si_bind_dsa_state;
3381 sctx->b.b.delete_depth_stencil_alpha_state = si_delete_dsa_state;
3382
3383 sctx->custom_dsa_flush = si_create_db_flush_dsa(sctx);
3384 sctx->custom_blend_resolve = si_create_blend_custom(sctx, V_028808_CB_RESOLVE);
3385 sctx->custom_blend_decompress = si_create_blend_custom(sctx, V_028808_CB_FMASK_DECOMPRESS);
3386 sctx->custom_blend_fastclear = si_create_blend_custom(sctx, V_028808_CB_ELIMINATE_FAST_CLEAR);
3387 sctx->custom_blend_dcc_decompress = si_create_blend_custom(sctx, V_028808_CB_DCC_DECOMPRESS);
3388
3389 sctx->b.b.set_clip_state = si_set_clip_state;
3390 sctx->b.b.set_stencil_ref = si_set_stencil_ref;
3391
3392 sctx->b.b.set_framebuffer_state = si_set_framebuffer_state;
3393 sctx->b.b.get_sample_position = cayman_get_sample_position;
3394
3395 sctx->b.b.create_sampler_state = si_create_sampler_state;
3396 sctx->b.b.delete_sampler_state = si_delete_sampler_state;
3397
3398 sctx->b.b.create_sampler_view = si_create_sampler_view;
3399 sctx->b.b.sampler_view_destroy = si_sampler_view_destroy;
3400
3401 sctx->b.b.set_sample_mask = si_set_sample_mask;
3402
3403 sctx->b.b.create_vertex_elements_state = si_create_vertex_elements;
3404 sctx->b.b.bind_vertex_elements_state = si_bind_vertex_elements;
3405 sctx->b.b.delete_vertex_elements_state = si_delete_vertex_element;
3406 sctx->b.b.set_vertex_buffers = si_set_vertex_buffers;
3407 sctx->b.b.set_index_buffer = si_set_index_buffer;
3408
3409 sctx->b.b.texture_barrier = si_texture_barrier;
3410 sctx->b.b.memory_barrier = si_memory_barrier;
3411 sctx->b.b.set_min_samples = si_set_min_samples;
3412 sctx->b.b.set_tess_state = si_set_tess_state;
3413
3414 sctx->b.b.set_active_query_state = si_set_active_query_state;
3415 sctx->b.set_occlusion_query_state = si_set_occlusion_query_state;
3416 sctx->b.need_gfx_cs_space = si_need_gfx_cs_space;
3417
3418 sctx->b.b.draw_vbo = si_draw_vbo;
3419
3420 si_init_config(sctx);
3421 }
3422
3423 static uint32_t si_get_bo_metadata_word1(struct r600_common_screen *rscreen)
3424 {
3425 return (ATI_VENDOR_ID << 16) | rscreen->info.pci_id;
3426 }
3427
3428 static void si_query_opaque_metadata(struct r600_common_screen *rscreen,
3429 struct r600_texture *rtex,
3430 struct radeon_bo_metadata *md)
3431 {
3432 struct si_screen *sscreen = (struct si_screen*)rscreen;
3433 struct pipe_resource *res = &rtex->resource.b.b;
3434 static const unsigned char swizzle[] = {
3435 PIPE_SWIZZLE_X,
3436 PIPE_SWIZZLE_Y,
3437 PIPE_SWIZZLE_Z,
3438 PIPE_SWIZZLE_W
3439 };
3440 uint32_t desc[8], i;
3441 bool is_array = util_resource_is_array_texture(res);
3442
3443 /* DRM 2.x.x doesn't support this. */
3444 if (rscreen->info.drm_major != 3)
3445 return;
3446
3447 assert(rtex->fmask.size == 0);
3448
3449 /* Metadata image format format version 1:
3450 * [0] = 1 (metadata format identifier)
3451 * [1] = (VENDOR_ID << 16) | PCI_ID
3452 * [2:9] = image descriptor for the whole resource
3453 * [2] is always 0, because the base address is cleared
3454 * [9] is the DCC offset bits [39:8] from the beginning of
3455 * the buffer
3456 * [10:10+LAST_LEVEL] = mipmap level offset bits [39:8] for each level
3457 */
3458
3459 md->metadata[0] = 1; /* metadata image format version 1 */
3460
3461 /* TILE_MODE_INDEX is ambiguous without a PCI ID. */
3462 md->metadata[1] = si_get_bo_metadata_word1(rscreen);
3463
3464 si_make_texture_descriptor(sscreen, rtex, true,
3465 res->target, res->format,
3466 swizzle, 0, res->last_level, 0,
3467 is_array ? res->array_size - 1 : 0,
3468 res->width0, res->height0, res->depth0,
3469 desc, NULL);
3470
3471 si_set_mutable_tex_desc_fields(rtex, &rtex->surface.level[0], 0, 0,
3472 rtex->surface.blk_w, false, desc);
3473
3474 /* Clear the base address and set the relative DCC offset. */
3475 desc[0] = 0;
3476 desc[1] &= C_008F14_BASE_ADDRESS_HI;
3477 desc[7] = rtex->dcc_offset >> 8;
3478
3479 /* Dwords [2:9] contain the image descriptor. */
3480 memcpy(&md->metadata[2], desc, sizeof(desc));
3481
3482 /* Dwords [10:..] contain the mipmap level offsets. */
3483 for (i = 0; i <= res->last_level; i++)
3484 md->metadata[10+i] = rtex->surface.level[i].offset >> 8;
3485
3486 md->size_metadata = (11 + res->last_level) * 4;
3487 }
3488
3489 static void si_apply_opaque_metadata(struct r600_common_screen *rscreen,
3490 struct r600_texture *rtex,
3491 struct radeon_bo_metadata *md)
3492 {
3493 uint32_t *desc = &md->metadata[2];
3494
3495 if (rscreen->chip_class < VI)
3496 return;
3497
3498 /* Return if DCC is enabled. The texture should be set up with it
3499 * already.
3500 */
3501 if (md->size_metadata >= 11 * 4 &&
3502 md->metadata[0] != 0 &&
3503 md->metadata[1] == si_get_bo_metadata_word1(rscreen) &&
3504 G_008F28_COMPRESSION_EN(desc[6])) {
3505 assert(rtex->dcc_offset == ((uint64_t)desc[7] << 8));
3506 return;
3507 }
3508
3509 /* Disable DCC. These are always set by texture_from_handle and must
3510 * be cleared here.
3511 */
3512 rtex->dcc_offset = 0;
3513 }
3514
3515 void si_init_screen_state_functions(struct si_screen *sscreen)
3516 {
3517 sscreen->b.query_opaque_metadata = si_query_opaque_metadata;
3518 sscreen->b.apply_opaque_metadata = si_apply_opaque_metadata;
3519 }
3520
3521 static void
3522 si_write_harvested_raster_configs(struct si_context *sctx,
3523 struct si_pm4_state *pm4,
3524 unsigned raster_config,
3525 unsigned raster_config_1)
3526 {
3527 unsigned sh_per_se = MAX2(sctx->screen->b.info.max_sh_per_se, 1);
3528 unsigned num_se = MAX2(sctx->screen->b.info.max_se, 1);
3529 unsigned rb_mask = sctx->screen->b.info.enabled_rb_mask;
3530 unsigned num_rb = MIN2(sctx->screen->b.info.num_render_backends, 16);
3531 unsigned rb_per_pkr = MIN2(num_rb / num_se / sh_per_se, 2);
3532 unsigned rb_per_se = num_rb / num_se;
3533 unsigned se_mask[4];
3534 unsigned se;
3535
3536 se_mask[0] = ((1 << rb_per_se) - 1);
3537 se_mask[1] = (se_mask[0] << rb_per_se);
3538 se_mask[2] = (se_mask[1] << rb_per_se);
3539 se_mask[3] = (se_mask[2] << rb_per_se);
3540
3541 se_mask[0] &= rb_mask;
3542 se_mask[1] &= rb_mask;
3543 se_mask[2] &= rb_mask;
3544 se_mask[3] &= rb_mask;
3545
3546 assert(num_se == 1 || num_se == 2 || num_se == 4);
3547 assert(sh_per_se == 1 || sh_per_se == 2);
3548 assert(rb_per_pkr == 1 || rb_per_pkr == 2);
3549
3550 /* XXX: I can't figure out what the *_XSEL and *_YSEL
3551 * fields are for, so I'm leaving them as their default
3552 * values. */
3553
3554 for (se = 0; se < num_se; se++) {
3555 unsigned raster_config_se = raster_config;
3556 unsigned pkr0_mask = ((1 << rb_per_pkr) - 1) << (se * rb_per_se);
3557 unsigned pkr1_mask = pkr0_mask << rb_per_pkr;
3558 int idx = (se / 2) * 2;
3559
3560 if ((num_se > 1) && (!se_mask[idx] || !se_mask[idx + 1])) {
3561 raster_config_se &= C_028350_SE_MAP;
3562
3563 if (!se_mask[idx]) {
3564 raster_config_se |=
3565 S_028350_SE_MAP(V_028350_RASTER_CONFIG_SE_MAP_3);
3566 } else {
3567 raster_config_se |=
3568 S_028350_SE_MAP(V_028350_RASTER_CONFIG_SE_MAP_0);
3569 }
3570 }
3571
3572 pkr0_mask &= rb_mask;
3573 pkr1_mask &= rb_mask;
3574 if (rb_per_se > 2 && (!pkr0_mask || !pkr1_mask)) {
3575 raster_config_se &= C_028350_PKR_MAP;
3576
3577 if (!pkr0_mask) {
3578 raster_config_se |=
3579 S_028350_PKR_MAP(V_028350_RASTER_CONFIG_PKR_MAP_3);
3580 } else {
3581 raster_config_se |=
3582 S_028350_PKR_MAP(V_028350_RASTER_CONFIG_PKR_MAP_0);
3583 }
3584 }
3585
3586 if (rb_per_se >= 2) {
3587 unsigned rb0_mask = 1 << (se * rb_per_se);
3588 unsigned rb1_mask = rb0_mask << 1;
3589
3590 rb0_mask &= rb_mask;
3591 rb1_mask &= rb_mask;
3592 if (!rb0_mask || !rb1_mask) {
3593 raster_config_se &= C_028350_RB_MAP_PKR0;
3594
3595 if (!rb0_mask) {
3596 raster_config_se |=
3597 S_028350_RB_MAP_PKR0(V_028350_RASTER_CONFIG_RB_MAP_3);
3598 } else {
3599 raster_config_se |=
3600 S_028350_RB_MAP_PKR0(V_028350_RASTER_CONFIG_RB_MAP_0);
3601 }
3602 }
3603
3604 if (rb_per_se > 2) {
3605 rb0_mask = 1 << (se * rb_per_se + rb_per_pkr);
3606 rb1_mask = rb0_mask << 1;
3607 rb0_mask &= rb_mask;
3608 rb1_mask &= rb_mask;
3609 if (!rb0_mask || !rb1_mask) {
3610 raster_config_se &= C_028350_RB_MAP_PKR1;
3611
3612 if (!rb0_mask) {
3613 raster_config_se |=
3614 S_028350_RB_MAP_PKR1(V_028350_RASTER_CONFIG_RB_MAP_3);
3615 } else {
3616 raster_config_se |=
3617 S_028350_RB_MAP_PKR1(V_028350_RASTER_CONFIG_RB_MAP_0);
3618 }
3619 }
3620 }
3621 }
3622
3623 /* GRBM_GFX_INDEX has a different offset on SI and CI+ */
3624 if (sctx->b.chip_class < CIK)
3625 si_pm4_set_reg(pm4, GRBM_GFX_INDEX,
3626 SE_INDEX(se) | SH_BROADCAST_WRITES |
3627 INSTANCE_BROADCAST_WRITES);
3628 else
3629 si_pm4_set_reg(pm4, R_030800_GRBM_GFX_INDEX,
3630 S_030800_SE_INDEX(se) | S_030800_SH_BROADCAST_WRITES(1) |
3631 S_030800_INSTANCE_BROADCAST_WRITES(1));
3632 si_pm4_set_reg(pm4, R_028350_PA_SC_RASTER_CONFIG, raster_config_se);
3633 }
3634
3635 /* GRBM_GFX_INDEX has a different offset on SI and CI+ */
3636 if (sctx->b.chip_class < CIK)
3637 si_pm4_set_reg(pm4, GRBM_GFX_INDEX,
3638 SE_BROADCAST_WRITES | SH_BROADCAST_WRITES |
3639 INSTANCE_BROADCAST_WRITES);
3640 else {
3641 si_pm4_set_reg(pm4, R_030800_GRBM_GFX_INDEX,
3642 S_030800_SE_BROADCAST_WRITES(1) | S_030800_SH_BROADCAST_WRITES(1) |
3643 S_030800_INSTANCE_BROADCAST_WRITES(1));
3644
3645 if ((num_se > 2) && ((!se_mask[0] && !se_mask[1]) ||
3646 (!se_mask[2] && !se_mask[3]))) {
3647 raster_config_1 &= C_028354_SE_PAIR_MAP;
3648
3649 if (!se_mask[0] && !se_mask[1]) {
3650 raster_config_1 |=
3651 S_028354_SE_PAIR_MAP(V_028354_RASTER_CONFIG_SE_PAIR_MAP_3);
3652 } else {
3653 raster_config_1 |=
3654 S_028354_SE_PAIR_MAP(V_028354_RASTER_CONFIG_SE_PAIR_MAP_0);
3655 }
3656 }
3657
3658 si_pm4_set_reg(pm4, R_028354_PA_SC_RASTER_CONFIG_1, raster_config_1);
3659 }
3660 }
3661
3662 static void si_init_config(struct si_context *sctx)
3663 {
3664 struct si_screen *sscreen = sctx->screen;
3665 unsigned num_rb = MIN2(sctx->screen->b.info.num_render_backends, 16);
3666 unsigned rb_mask = sctx->screen->b.info.enabled_rb_mask;
3667 unsigned raster_config, raster_config_1;
3668 uint64_t border_color_va = sctx->border_color_buffer->gpu_address;
3669 struct si_pm4_state *pm4 = CALLOC_STRUCT(si_pm4_state);
3670 int i;
3671
3672 if (!pm4)
3673 return;
3674
3675 si_pm4_cmd_begin(pm4, PKT3_CONTEXT_CONTROL);
3676 si_pm4_cmd_add(pm4, CONTEXT_CONTROL_LOAD_ENABLE(1));
3677 si_pm4_cmd_add(pm4, CONTEXT_CONTROL_SHADOW_ENABLE(1));
3678 si_pm4_cmd_end(pm4, false);
3679
3680 si_pm4_set_reg(pm4, R_028A18_VGT_HOS_MAX_TESS_LEVEL, fui(64));
3681 si_pm4_set_reg(pm4, R_028A1C_VGT_HOS_MIN_TESS_LEVEL, fui(0));
3682
3683 /* FIXME calculate these values somehow ??? */
3684 si_pm4_set_reg(pm4, R_028A54_VGT_GS_PER_ES, SI_GS_PER_ES);
3685 si_pm4_set_reg(pm4, R_028A58_VGT_ES_PER_GS, 0x40);
3686 si_pm4_set_reg(pm4, R_028A5C_VGT_GS_PER_VS, 0x2);
3687
3688 si_pm4_set_reg(pm4, R_028A8C_VGT_PRIMITIVEID_RESET, 0x0);
3689 si_pm4_set_reg(pm4, R_028B28_VGT_STRMOUT_DRAW_OPAQUE_OFFSET, 0);
3690
3691 si_pm4_set_reg(pm4, R_028B98_VGT_STRMOUT_BUFFER_CONFIG, 0x0);
3692 si_pm4_set_reg(pm4, R_028AB8_VGT_VTX_CNT_EN, 0x0);
3693 if (sctx->b.chip_class < CIK)
3694 si_pm4_set_reg(pm4, R_008A14_PA_CL_ENHANCE, S_008A14_NUM_CLIP_SEQ(3) |
3695 S_008A14_CLIP_VTX_REORDER_ENA(1));
3696
3697 si_pm4_set_reg(pm4, R_028BD4_PA_SC_CENTROID_PRIORITY_0, 0x76543210);
3698 si_pm4_set_reg(pm4, R_028BD8_PA_SC_CENTROID_PRIORITY_1, 0xfedcba98);
3699
3700 si_pm4_set_reg(pm4, R_02882C_PA_SU_PRIM_FILTER_CNTL, 0);
3701
3702 for (i = 0; i < 16; i++) {
3703 si_pm4_set_reg(pm4, R_0282D0_PA_SC_VPORT_ZMIN_0 + i*8, 0);
3704 si_pm4_set_reg(pm4, R_0282D4_PA_SC_VPORT_ZMAX_0 + i*8, fui(1.0));
3705 }
3706
3707 switch (sctx->screen->b.family) {
3708 case CHIP_TAHITI:
3709 case CHIP_PITCAIRN:
3710 raster_config = 0x2a00126a;
3711 raster_config_1 = 0x00000000;
3712 break;
3713 case CHIP_VERDE:
3714 raster_config = 0x0000124a;
3715 raster_config_1 = 0x00000000;
3716 break;
3717 case CHIP_OLAND:
3718 raster_config = 0x00000082;
3719 raster_config_1 = 0x00000000;
3720 break;
3721 case CHIP_HAINAN:
3722 raster_config = 0x00000000;
3723 raster_config_1 = 0x00000000;
3724 break;
3725 case CHIP_BONAIRE:
3726 raster_config = 0x16000012;
3727 raster_config_1 = 0x00000000;
3728 break;
3729 case CHIP_HAWAII:
3730 raster_config = 0x3a00161a;
3731 raster_config_1 = 0x0000002e;
3732 break;
3733 case CHIP_FIJI:
3734 if (sscreen->b.info.cik_macrotile_mode_array[0] == 0x000000e8) {
3735 /* old kernels with old tiling config */
3736 raster_config = 0x16000012;
3737 raster_config_1 = 0x0000002a;
3738 } else {
3739 raster_config = 0x3a00161a;
3740 raster_config_1 = 0x0000002e;
3741 }
3742 break;
3743 case CHIP_POLARIS10:
3744 raster_config = 0x16000012;
3745 raster_config_1 = 0x0000002a;
3746 break;
3747 case CHIP_POLARIS11:
3748 raster_config = 0x16000012;
3749 raster_config_1 = 0x00000000;
3750 break;
3751 case CHIP_TONGA:
3752 raster_config = 0x16000012;
3753 raster_config_1 = 0x0000002a;
3754 break;
3755 case CHIP_ICELAND:
3756 if (num_rb == 1)
3757 raster_config = 0x00000000;
3758 else
3759 raster_config = 0x00000002;
3760 raster_config_1 = 0x00000000;
3761 break;
3762 case CHIP_CARRIZO:
3763 raster_config = 0x00000002;
3764 raster_config_1 = 0x00000000;
3765 break;
3766 case CHIP_KAVERI:
3767 /* KV should be 0x00000002, but that causes problems with radeon */
3768 raster_config = 0x00000000; /* 0x00000002 */
3769 raster_config_1 = 0x00000000;
3770 break;
3771 case CHIP_KABINI:
3772 case CHIP_MULLINS:
3773 case CHIP_STONEY:
3774 raster_config = 0x00000000;
3775 raster_config_1 = 0x00000000;
3776 break;
3777 default:
3778 fprintf(stderr,
3779 "radeonsi: Unknown GPU, using 0 for raster_config\n");
3780 raster_config = 0x00000000;
3781 raster_config_1 = 0x00000000;
3782 break;
3783 }
3784
3785 /* Always use the default config when all backends are enabled
3786 * (or when we failed to determine the enabled backends).
3787 */
3788 if (!rb_mask || util_bitcount(rb_mask) >= num_rb) {
3789 si_pm4_set_reg(pm4, R_028350_PA_SC_RASTER_CONFIG,
3790 raster_config);
3791 if (sctx->b.chip_class >= CIK)
3792 si_pm4_set_reg(pm4, R_028354_PA_SC_RASTER_CONFIG_1,
3793 raster_config_1);
3794 } else {
3795 si_write_harvested_raster_configs(sctx, pm4, raster_config, raster_config_1);
3796 }
3797
3798 si_pm4_set_reg(pm4, R_028204_PA_SC_WINDOW_SCISSOR_TL, S_028204_WINDOW_OFFSET_DISABLE(1));
3799 si_pm4_set_reg(pm4, R_028240_PA_SC_GENERIC_SCISSOR_TL, S_028240_WINDOW_OFFSET_DISABLE(1));
3800 si_pm4_set_reg(pm4, R_028244_PA_SC_GENERIC_SCISSOR_BR,
3801 S_028244_BR_X(16384) | S_028244_BR_Y(16384));
3802 si_pm4_set_reg(pm4, R_028030_PA_SC_SCREEN_SCISSOR_TL, 0);
3803 si_pm4_set_reg(pm4, R_028034_PA_SC_SCREEN_SCISSOR_BR,
3804 S_028034_BR_X(16384) | S_028034_BR_Y(16384));
3805
3806 si_pm4_set_reg(pm4, R_02820C_PA_SC_CLIPRECT_RULE, 0xFFFF);
3807 si_pm4_set_reg(pm4, R_028230_PA_SC_EDGERULE, 0xAAAAAAAA);
3808 /* PA_SU_HARDWARE_SCREEN_OFFSET must be 0 due to hw bug on SI */
3809 si_pm4_set_reg(pm4, R_028234_PA_SU_HARDWARE_SCREEN_OFFSET, 0);
3810 si_pm4_set_reg(pm4, R_028820_PA_CL_NANINF_CNTL, 0);
3811 si_pm4_set_reg(pm4, R_028AC0_DB_SRESULTS_COMPARE_STATE0, 0x0);
3812 si_pm4_set_reg(pm4, R_028AC4_DB_SRESULTS_COMPARE_STATE1, 0x0);
3813 si_pm4_set_reg(pm4, R_028AC8_DB_PRELOAD_CONTROL, 0x0);
3814 si_pm4_set_reg(pm4, R_02800C_DB_RENDER_OVERRIDE,
3815 S_02800C_FORCE_HIS_ENABLE0(V_02800C_FORCE_DISABLE) |
3816 S_02800C_FORCE_HIS_ENABLE1(V_02800C_FORCE_DISABLE));
3817
3818 si_pm4_set_reg(pm4, R_028400_VGT_MAX_VTX_INDX, ~0);
3819 si_pm4_set_reg(pm4, R_028404_VGT_MIN_VTX_INDX, 0);
3820 si_pm4_set_reg(pm4, R_028408_VGT_INDX_OFFSET, 0);
3821
3822 if (sctx->b.chip_class >= CIK) {
3823 si_pm4_set_reg(pm4, R_00B41C_SPI_SHADER_PGM_RSRC3_HS, 0);
3824 si_pm4_set_reg(pm4, R_00B31C_SPI_SHADER_PGM_RSRC3_ES, S_00B31C_CU_EN(0xffff));
3825 si_pm4_set_reg(pm4, R_00B21C_SPI_SHADER_PGM_RSRC3_GS, S_00B21C_CU_EN(0xffff));
3826
3827 if (sscreen->b.info.num_good_compute_units /
3828 (sscreen->b.info.max_se * sscreen->b.info.max_sh_per_se) <= 4) {
3829 /* Too few available compute units per SH. Disallowing
3830 * VS to run on CU0 could hurt us more than late VS
3831 * allocation would help.
3832 *
3833 * LATE_ALLOC_VS = 2 is the highest safe number.
3834 */
3835 si_pm4_set_reg(pm4, R_00B51C_SPI_SHADER_PGM_RSRC3_LS, S_00B51C_CU_EN(0xffff));
3836 si_pm4_set_reg(pm4, R_00B118_SPI_SHADER_PGM_RSRC3_VS, S_00B118_CU_EN(0xffff));
3837 si_pm4_set_reg(pm4, R_00B11C_SPI_SHADER_LATE_ALLOC_VS, S_00B11C_LIMIT(2));
3838 } else {
3839 /* Set LATE_ALLOC_VS == 31. It should be less than
3840 * the number of scratch waves. Limitations:
3841 * - VS can't execute on CU0.
3842 * - If HS writes outputs to LDS, LS can't execute on CU0.
3843 */
3844 si_pm4_set_reg(pm4, R_00B51C_SPI_SHADER_PGM_RSRC3_LS, S_00B51C_CU_EN(0xfffe));
3845 si_pm4_set_reg(pm4, R_00B118_SPI_SHADER_PGM_RSRC3_VS, S_00B118_CU_EN(0xfffe));
3846 si_pm4_set_reg(pm4, R_00B11C_SPI_SHADER_LATE_ALLOC_VS, S_00B11C_LIMIT(31));
3847 }
3848
3849 si_pm4_set_reg(pm4, R_00B01C_SPI_SHADER_PGM_RSRC3_PS, S_00B01C_CU_EN(0xffff));
3850 }
3851
3852 if (sctx->b.chip_class >= VI) {
3853 unsigned vgt_tess_distribution;
3854
3855 si_pm4_set_reg(pm4, R_028424_CB_DCC_CONTROL,
3856 S_028424_OVERWRITE_COMBINER_MRT_SHARING_DISABLE(1) |
3857 S_028424_OVERWRITE_COMBINER_WATERMARK(4));
3858 if (sctx->b.family < CHIP_POLARIS10)
3859 si_pm4_set_reg(pm4, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL, 30);
3860 si_pm4_set_reg(pm4, R_028C5C_VGT_OUT_DEALLOC_CNTL, 32);
3861
3862 vgt_tess_distribution =
3863 S_028B50_ACCUM_ISOLINE(32) |
3864 S_028B50_ACCUM_TRI(11) |
3865 S_028B50_ACCUM_QUAD(11) |
3866 S_028B50_DONUT_SPLIT(16);
3867
3868 /* Testing with Unigine Heaven extreme tesselation yielded best results
3869 * with TRAP_SPLIT = 3.
3870 */
3871 if (sctx->b.family == CHIP_FIJI ||
3872 sctx->b.family >= CHIP_POLARIS10)
3873 vgt_tess_distribution |= S_028B50_TRAP_SPLIT(3);
3874
3875 si_pm4_set_reg(pm4, R_028B50_VGT_TESS_DISTRIBUTION, vgt_tess_distribution);
3876 } else {
3877 si_pm4_set_reg(pm4, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL, 14);
3878 si_pm4_set_reg(pm4, R_028C5C_VGT_OUT_DEALLOC_CNTL, 16);
3879 }
3880
3881 if (sctx->b.family == CHIP_STONEY)
3882 si_pm4_set_reg(pm4, R_028C40_PA_SC_SHADER_CONTROL, 0);
3883
3884 si_pm4_set_reg(pm4, R_028080_TA_BC_BASE_ADDR, border_color_va >> 8);
3885 if (sctx->b.chip_class >= CIK)
3886 si_pm4_set_reg(pm4, R_028084_TA_BC_BASE_ADDR_HI, border_color_va >> 40);
3887 si_pm4_add_bo(pm4, sctx->border_color_buffer, RADEON_USAGE_READ,
3888 RADEON_PRIO_BORDER_COLORS);
3889
3890 si_pm4_upload_indirect_buffer(sctx, pm4);
3891 sctx->init_config = pm4;
3892 }