projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
radeonsi: add an environment variable that forces EQAA for MSAA allocations
[mesa.git] / src / gallium / drivers / radeonsi / si_state.c
1 /*
2 * Copyright 2012 Advanced Micro Devices, Inc.
3 * All Rights Reserved.
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * on the rights to use, copy, modify, merge, publish, distribute, sub
9 * license, and/or sell copies of the Software, and to permit persons to whom
10 * the Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
14 * Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
20 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
21 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
22 * USE OR OTHER DEALINGS IN THE SOFTWARE.
23 */
24
25 #include "si_build_pm4.h"
26 #include "gfx9d.h"
27 #include "si_query.h"
28
29 #include "util/u_dual_blend.h"
30 #include "util/u_format.h"
31 #include "util/u_format_s3tc.h"
32 #include "util/u_memory.h"
33 #include "util/u_resource.h"
34 #include "util/u_upload_mgr.h"
35
36 static unsigned si_map_swizzle(unsigned swizzle)
37 {
38 switch (swizzle) {
39 case PIPE_SWIZZLE_Y:
40 return V_008F0C_SQ_SEL_Y;
41 case PIPE_SWIZZLE_Z:
42 return V_008F0C_SQ_SEL_Z;
43 case PIPE_SWIZZLE_W:
44 return V_008F0C_SQ_SEL_W;
45 case PIPE_SWIZZLE_0:
46 return V_008F0C_SQ_SEL_0;
47 case PIPE_SWIZZLE_1:
48 return V_008F0C_SQ_SEL_1;
49 default: /* PIPE_SWIZZLE_X */
50 return V_008F0C_SQ_SEL_X;
51 }
52 }
53
54 /* 12.4 fixed-point */
55 static unsigned si_pack_float_12p4(float x)
56 {
57 return x <= 0 ? 0 :
58 x >= 4096 ? 0xffff : x * 16;
59 }
60
61 /*
62 * Inferred framebuffer and blender state.
63 *
64 * CB_TARGET_MASK is emitted here to avoid a hang with dual source blending
65 * if there is not enough PS outputs.
66 */
67 static void si_emit_cb_render_state(struct si_context *sctx)
68 {
69 struct radeon_winsys_cs *cs = sctx->gfx_cs;
70 struct si_state_blend *blend = sctx->queued.named.blend;
71 /* CB_COLORn_INFO.FORMAT=INVALID should disable unbound colorbuffers,
72 * but you never know. */
73 uint32_t cb_target_mask = sctx->framebuffer.colorbuf_enabled_4bit;
74 unsigned i;
75
76 if (blend)
77 cb_target_mask &= blend->cb_target_mask;
78
79 /* Avoid a hang that happens when dual source blending is enabled
80 * but there is not enough color outputs. This is undefined behavior,
81 * so disable color writes completely.
82 *
83 * Reproducible with Unigine Heaven 4.0 and drirc missing.
84 */
85 if (blend && blend->dual_src_blend &&
86 sctx->ps_shader.cso &&
87 (sctx->ps_shader.cso->info.colors_written & 0x3) != 0x3)
88 cb_target_mask = 0;
89
90 radeon_set_context_reg(cs, R_028238_CB_TARGET_MASK, cb_target_mask);
91
92 /* GFX9: Flush DFSM when CB_TARGET_MASK changes.
93 * I think we don't have to do anything between IBs.
94 */
95 if (sctx->screen->dfsm_allowed &&
96 sctx->last_cb_target_mask != cb_target_mask) {
97 sctx->last_cb_target_mask = cb_target_mask;
98
99 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
100 radeon_emit(cs, EVENT_TYPE(V_028A90_FLUSH_DFSM) | EVENT_INDEX(0));
101 }
102
103 if (sctx->chip_class >= VI) {
104 /* DCC MSAA workaround for blending.
105 * Alternatively, we can set CB_COLORi_DCC_CONTROL.OVERWRITE_-
106 * COMBINER_DISABLE, but that would be more complicated.
107 */
108 bool oc_disable = (sctx->chip_class == VI ||
109 sctx->chip_class == GFX9) &&
110 blend &&
111 blend->blend_enable_4bit & cb_target_mask &&
112 sctx->framebuffer.nr_samples >= 2;
113
114 radeon_set_context_reg(cs, R_028424_CB_DCC_CONTROL,
115 S_028424_OVERWRITE_COMBINER_MRT_SHARING_DISABLE(1) |
116 S_028424_OVERWRITE_COMBINER_WATERMARK(4) |
117 S_028424_OVERWRITE_COMBINER_DISABLE(oc_disable));
118 }
119
120 /* RB+ register settings. */
121 if (sctx->screen->rbplus_allowed) {
122 unsigned spi_shader_col_format =
123 sctx->ps_shader.cso ?
124 sctx->ps_shader.current->key.part.ps.epilog.spi_shader_col_format : 0;
125 unsigned sx_ps_downconvert = 0;
126 unsigned sx_blend_opt_epsilon = 0;
127 unsigned sx_blend_opt_control = 0;
128
129 for (i = 0; i < sctx->framebuffer.state.nr_cbufs; i++) {
130 struct r600_surface *surf =
131 (struct r600_surface*)sctx->framebuffer.state.cbufs[i];
132 unsigned format, swap, spi_format, colormask;
133 bool has_alpha, has_rgb;
134
135 if (!surf)
136 continue;
137
138 format = G_028C70_FORMAT(surf->cb_color_info);
139 swap = G_028C70_COMP_SWAP(surf->cb_color_info);
140 spi_format = (spi_shader_col_format >> (i * 4)) & 0xf;
141 colormask = (cb_target_mask >> (i * 4)) & 0xf;
142
143 /* Set if RGB and A are present. */
144 has_alpha = !G_028C74_FORCE_DST_ALPHA_1(surf->cb_color_attrib);
145
146 if (format == V_028C70_COLOR_8 ||
147 format == V_028C70_COLOR_16 ||
148 format == V_028C70_COLOR_32)
149 has_rgb = !has_alpha;
150 else
151 has_rgb = true;
152
153 /* Check the colormask and export format. */
154 if (!(colormask & (PIPE_MASK_RGBA & ~PIPE_MASK_A)))
155 has_rgb = false;
156 if (!(colormask & PIPE_MASK_A))
157 has_alpha = false;
158
159 if (spi_format == V_028714_SPI_SHADER_ZERO) {
160 has_rgb = false;
161 has_alpha = false;
162 }
163
164 /* Disable value checking for disabled channels. */
165 if (!has_rgb)
166 sx_blend_opt_control |= S_02875C_MRT0_COLOR_OPT_DISABLE(1) << (i * 4);
167 if (!has_alpha)
168 sx_blend_opt_control |= S_02875C_MRT0_ALPHA_OPT_DISABLE(1) << (i * 4);
169
170 /* Enable down-conversion for 32bpp and smaller formats. */
171 switch (format) {
172 case V_028C70_COLOR_8:
173 case V_028C70_COLOR_8_8:
174 case V_028C70_COLOR_8_8_8_8:
175 /* For 1 and 2-channel formats, use the superset thereof. */
176 if (spi_format == V_028714_SPI_SHADER_FP16_ABGR ||
177 spi_format == V_028714_SPI_SHADER_UINT16_ABGR ||
178 spi_format == V_028714_SPI_SHADER_SINT16_ABGR) {
179 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_8_8_8_8 << (i * 4);
180 sx_blend_opt_epsilon |= V_028758_8BIT_FORMAT << (i * 4);
181 }
182 break;
183
184 case V_028C70_COLOR_5_6_5:
185 if (spi_format == V_028714_SPI_SHADER_FP16_ABGR) {
186 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_5_6_5 << (i * 4);
187 sx_blend_opt_epsilon |= V_028758_6BIT_FORMAT << (i * 4);
188 }
189 break;
190
191 case V_028C70_COLOR_1_5_5_5:
192 if (spi_format == V_028714_SPI_SHADER_FP16_ABGR) {
193 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_1_5_5_5 << (i * 4);
194 sx_blend_opt_epsilon |= V_028758_5BIT_FORMAT << (i * 4);
195 }
196 break;
197
198 case V_028C70_COLOR_4_4_4_4:
199 if (spi_format == V_028714_SPI_SHADER_FP16_ABGR) {
200 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_4_4_4_4 << (i * 4);
201 sx_blend_opt_epsilon |= V_028758_4BIT_FORMAT << (i * 4);
202 }
203 break;
204
205 case V_028C70_COLOR_32:
206 if (swap == V_028C70_SWAP_STD &&
207 spi_format == V_028714_SPI_SHADER_32_R)
208 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_32_R << (i * 4);
209 else if (swap == V_028C70_SWAP_ALT_REV &&
210 spi_format == V_028714_SPI_SHADER_32_AR)
211 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_32_A << (i * 4);
212 break;
213
214 case V_028C70_COLOR_16:
215 case V_028C70_COLOR_16_16:
216 /* For 1-channel formats, use the superset thereof. */
217 if (spi_format == V_028714_SPI_SHADER_UNORM16_ABGR ||
218 spi_format == V_028714_SPI_SHADER_SNORM16_ABGR ||
219 spi_format == V_028714_SPI_SHADER_UINT16_ABGR ||
220 spi_format == V_028714_SPI_SHADER_SINT16_ABGR) {
221 if (swap == V_028C70_SWAP_STD ||
222 swap == V_028C70_SWAP_STD_REV)
223 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_16_16_GR << (i * 4);
224 else
225 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_16_16_AR << (i * 4);
226 }
227 break;
228
229 case V_028C70_COLOR_10_11_11:
230 if (spi_format == V_028714_SPI_SHADER_FP16_ABGR) {
231 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_10_11_11 << (i * 4);
232 sx_blend_opt_epsilon |= V_028758_11BIT_FORMAT << (i * 4);
233 }
234 break;
235
236 case V_028C70_COLOR_2_10_10_10:
237 if (spi_format == V_028714_SPI_SHADER_FP16_ABGR) {
238 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_2_10_10_10 << (i * 4);
239 sx_blend_opt_epsilon |= V_028758_10BIT_FORMAT << (i * 4);
240 }
241 break;
242 }
243 }
244
245 radeon_set_context_reg_seq(cs, R_028754_SX_PS_DOWNCONVERT, 3);
246 radeon_emit(cs, sx_ps_downconvert); /* R_028754_SX_PS_DOWNCONVERT */
247 radeon_emit(cs, sx_blend_opt_epsilon); /* R_028758_SX_BLEND_OPT_EPSILON */
248 radeon_emit(cs, sx_blend_opt_control); /* R_02875C_SX_BLEND_OPT_CONTROL */
249 }
250 }
251
252 /*
253 * Blender functions
254 */
255
256 static uint32_t si_translate_blend_function(int blend_func)
257 {
258 switch (blend_func) {
259 case PIPE_BLEND_ADD:
260 return V_028780_COMB_DST_PLUS_SRC;
261 case PIPE_BLEND_SUBTRACT:
262 return V_028780_COMB_SRC_MINUS_DST;
263 case PIPE_BLEND_REVERSE_SUBTRACT:
264 return V_028780_COMB_DST_MINUS_SRC;
265 case PIPE_BLEND_MIN:
266 return V_028780_COMB_MIN_DST_SRC;
267 case PIPE_BLEND_MAX:
268 return V_028780_COMB_MAX_DST_SRC;
269 default:
270 PRINT_ERR("Unknown blend function %d\n", blend_func);
271 assert(0);
272 break;
273 }
274 return 0;
275 }
276
277 static uint32_t si_translate_blend_factor(int blend_fact)
278 {
279 switch (blend_fact) {
280 case PIPE_BLENDFACTOR_ONE:
281 return V_028780_BLEND_ONE;
282 case PIPE_BLENDFACTOR_SRC_COLOR:
283 return V_028780_BLEND_SRC_COLOR;
284 case PIPE_BLENDFACTOR_SRC_ALPHA:
285 return V_028780_BLEND_SRC_ALPHA;
286 case PIPE_BLENDFACTOR_DST_ALPHA:
287 return V_028780_BLEND_DST_ALPHA;
288 case PIPE_BLENDFACTOR_DST_COLOR:
289 return V_028780_BLEND_DST_COLOR;
290 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
291 return V_028780_BLEND_SRC_ALPHA_SATURATE;
292 case PIPE_BLENDFACTOR_CONST_COLOR:
293 return V_028780_BLEND_CONSTANT_COLOR;
294 case PIPE_BLENDFACTOR_CONST_ALPHA:
295 return V_028780_BLEND_CONSTANT_ALPHA;
296 case PIPE_BLENDFACTOR_ZERO:
297 return V_028780_BLEND_ZERO;
298 case PIPE_BLENDFACTOR_INV_SRC_COLOR:
299 return V_028780_BLEND_ONE_MINUS_SRC_COLOR;
300 case PIPE_BLENDFACTOR_INV_SRC_ALPHA:
301 return V_028780_BLEND_ONE_MINUS_SRC_ALPHA;
302 case PIPE_BLENDFACTOR_INV_DST_ALPHA:
303 return V_028780_BLEND_ONE_MINUS_DST_ALPHA;
304 case PIPE_BLENDFACTOR_INV_DST_COLOR:
305 return V_028780_BLEND_ONE_MINUS_DST_COLOR;
306 case PIPE_BLENDFACTOR_INV_CONST_COLOR:
307 return V_028780_BLEND_ONE_MINUS_CONSTANT_COLOR;
308 case PIPE_BLENDFACTOR_INV_CONST_ALPHA:
309 return V_028780_BLEND_ONE_MINUS_CONSTANT_ALPHA;
310 case PIPE_BLENDFACTOR_SRC1_COLOR:
311 return V_028780_BLEND_SRC1_COLOR;
312 case PIPE_BLENDFACTOR_SRC1_ALPHA:
313 return V_028780_BLEND_SRC1_ALPHA;
314 case PIPE_BLENDFACTOR_INV_SRC1_COLOR:
315 return V_028780_BLEND_INV_SRC1_COLOR;
316 case PIPE_BLENDFACTOR_INV_SRC1_ALPHA:
317 return V_028780_BLEND_INV_SRC1_ALPHA;
318 default:
319 PRINT_ERR("Bad blend factor %d not supported!\n", blend_fact);
320 assert(0);
321 break;
322 }
323 return 0;
324 }
325
326 static uint32_t si_translate_blend_opt_function(int blend_func)
327 {
328 switch (blend_func) {
329 case PIPE_BLEND_ADD:
330 return V_028760_OPT_COMB_ADD;
331 case PIPE_BLEND_SUBTRACT:
332 return V_028760_OPT_COMB_SUBTRACT;
333 case PIPE_BLEND_REVERSE_SUBTRACT:
334 return V_028760_OPT_COMB_REVSUBTRACT;
335 case PIPE_BLEND_MIN:
336 return V_028760_OPT_COMB_MIN;
337 case PIPE_BLEND_MAX:
338 return V_028760_OPT_COMB_MAX;
339 default:
340 return V_028760_OPT_COMB_BLEND_DISABLED;
341 }
342 }
343
344 static uint32_t si_translate_blend_opt_factor(int blend_fact, bool is_alpha)
345 {
346 switch (blend_fact) {
347 case PIPE_BLENDFACTOR_ZERO:
348 return V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_ALL;
349 case PIPE_BLENDFACTOR_ONE:
350 return V_028760_BLEND_OPT_PRESERVE_ALL_IGNORE_NONE;
351 case PIPE_BLENDFACTOR_SRC_COLOR:
352 return is_alpha ? V_028760_BLEND_OPT_PRESERVE_A1_IGNORE_A0
353 : V_028760_BLEND_OPT_PRESERVE_C1_IGNORE_C0;
354 case PIPE_BLENDFACTOR_INV_SRC_COLOR:
355 return is_alpha ? V_028760_BLEND_OPT_PRESERVE_A0_IGNORE_A1
356 : V_028760_BLEND_OPT_PRESERVE_C0_IGNORE_C1;
357 case PIPE_BLENDFACTOR_SRC_ALPHA:
358 return V_028760_BLEND_OPT_PRESERVE_A1_IGNORE_A0;
359 case PIPE_BLENDFACTOR_INV_SRC_ALPHA:
360 return V_028760_BLEND_OPT_PRESERVE_A0_IGNORE_A1;
361 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
362 return is_alpha ? V_028760_BLEND_OPT_PRESERVE_ALL_IGNORE_NONE
363 : V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_A0;
364 default:
365 return V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE;
366 }
367 }
368
369 static void si_blend_check_commutativity(struct si_screen *sscreen,
370 struct si_state_blend *blend,
371 enum pipe_blend_func func,
372 enum pipe_blendfactor src,
373 enum pipe_blendfactor dst,
374 unsigned chanmask)
375 {
376 /* Src factor is allowed when it does not depend on Dst */
377 static const uint32_t src_allowed =
378 (1u << PIPE_BLENDFACTOR_ONE) |
379 (1u << PIPE_BLENDFACTOR_SRC_COLOR) |
380 (1u << PIPE_BLENDFACTOR_SRC_ALPHA) |
381 (1u << PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE) |
382 (1u << PIPE_BLENDFACTOR_CONST_COLOR) |
383 (1u << PIPE_BLENDFACTOR_CONST_ALPHA) |
384 (1u << PIPE_BLENDFACTOR_SRC1_COLOR) |
385 (1u << PIPE_BLENDFACTOR_SRC1_ALPHA) |
386 (1u << PIPE_BLENDFACTOR_ZERO) |
387 (1u << PIPE_BLENDFACTOR_INV_SRC_COLOR) |
388 (1u << PIPE_BLENDFACTOR_INV_SRC_ALPHA) |
389 (1u << PIPE_BLENDFACTOR_INV_CONST_COLOR) |
390 (1u << PIPE_BLENDFACTOR_INV_CONST_ALPHA) |
391 (1u << PIPE_BLENDFACTOR_INV_SRC1_COLOR) |
392 (1u << PIPE_BLENDFACTOR_INV_SRC1_ALPHA);
393
394 if (dst == PIPE_BLENDFACTOR_ONE &&
395 (src_allowed & (1u << src))) {
396 /* Addition is commutative, but floating point addition isn't
397 * associative: subtle changes can be introduced via different
398 * rounding.
399 *
400 * Out-of-order is also non-deterministic, which means that
401 * this breaks OpenGL invariance requirements. So only enable
402 * out-of-order additive blending if explicitly allowed by a
403 * setting.
404 */
405 if (func == PIPE_BLEND_MAX || func == PIPE_BLEND_MIN ||
406 (func == PIPE_BLEND_ADD && sscreen->commutative_blend_add))
407 blend->commutative_4bit |= chanmask;
408 }
409 }
410
411 /**
412 * Get rid of DST in the blend factors by commuting the operands:
413 * func(src * DST, dst * 0) ---> func(src * 0, dst * SRC)
414 */
415 static void si_blend_remove_dst(unsigned *func, unsigned *src_factor,
416 unsigned *dst_factor, unsigned expected_dst,
417 unsigned replacement_src)
418 {
419 if (*src_factor == expected_dst &&
420 *dst_factor == PIPE_BLENDFACTOR_ZERO) {
421 *src_factor = PIPE_BLENDFACTOR_ZERO;
422 *dst_factor = replacement_src;
423
424 /* Commuting the operands requires reversing subtractions. */
425 if (*func == PIPE_BLEND_SUBTRACT)
426 *func = PIPE_BLEND_REVERSE_SUBTRACT;
427 else if (*func == PIPE_BLEND_REVERSE_SUBTRACT)
428 *func = PIPE_BLEND_SUBTRACT;
429 }
430 }
431
432 static bool si_blend_factor_uses_dst(unsigned factor)
433 {
434 return factor == PIPE_BLENDFACTOR_DST_COLOR ||
435 factor == PIPE_BLENDFACTOR_DST_ALPHA ||
436 factor == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE ||
437 factor == PIPE_BLENDFACTOR_INV_DST_ALPHA ||
438 factor == PIPE_BLENDFACTOR_INV_DST_COLOR;
439 }
440
441 static void *si_create_blend_state_mode(struct pipe_context *ctx,
442 const struct pipe_blend_state *state,
443 unsigned mode)
444 {
445 struct si_context *sctx = (struct si_context*)ctx;
446 struct si_state_blend *blend = CALLOC_STRUCT(si_state_blend);
447 struct si_pm4_state *pm4 = &blend->pm4;
448 uint32_t sx_mrt_blend_opt[8] = {0};
449 uint32_t color_control = 0;
450
451 if (!blend)
452 return NULL;
453
454 blend->alpha_to_coverage = state->alpha_to_coverage;
455 blend->alpha_to_one = state->alpha_to_one;
456 blend->dual_src_blend = util_blend_state_is_dual(state, 0);
457 blend->logicop_enable = state->logicop_enable;
458
459 if (state->logicop_enable) {
460 color_control |= S_028808_ROP3(state->logicop_func | (state->logicop_func << 4));
461 } else {
462 color_control |= S_028808_ROP3(0xcc);
463 }
464
465 si_pm4_set_reg(pm4, R_028B70_DB_ALPHA_TO_MASK,
466 S_028B70_ALPHA_TO_MASK_ENABLE(state->alpha_to_coverage) |
467 S_028B70_ALPHA_TO_MASK_OFFSET0(2) |
468 S_028B70_ALPHA_TO_MASK_OFFSET1(2) |
469 S_028B70_ALPHA_TO_MASK_OFFSET2(2) |
470 S_028B70_ALPHA_TO_MASK_OFFSET3(2));
471
472 if (state->alpha_to_coverage)
473 blend->need_src_alpha_4bit |= 0xf;
474
475 blend->cb_target_mask = 0;
476 blend->cb_target_enabled_4bit = 0;
477
478 for (int i = 0; i < 8; i++) {
479 /* state->rt entries > 0 only written if independent blending */
480 const int j = state->independent_blend_enable ? i : 0;
481
482 unsigned eqRGB = state->rt[j].rgb_func;
483 unsigned srcRGB = state->rt[j].rgb_src_factor;
484 unsigned dstRGB = state->rt[j].rgb_dst_factor;
485 unsigned eqA = state->rt[j].alpha_func;
486 unsigned srcA = state->rt[j].alpha_src_factor;
487 unsigned dstA = state->rt[j].alpha_dst_factor;
488
489 unsigned srcRGB_opt, dstRGB_opt, srcA_opt, dstA_opt;
490 unsigned blend_cntl = 0;
491
492 sx_mrt_blend_opt[i] =
493 S_028760_COLOR_COMB_FCN(V_028760_OPT_COMB_BLEND_DISABLED) |
494 S_028760_ALPHA_COMB_FCN(V_028760_OPT_COMB_BLEND_DISABLED);
495
496 /* Only set dual source blending for MRT0 to avoid a hang. */
497 if (i >= 1 && blend->dual_src_blend) {
498 /* Vulkan does this for dual source blending. */
499 if (i == 1)
500 blend_cntl |= S_028780_ENABLE(1);
501
502 si_pm4_set_reg(pm4, R_028780_CB_BLEND0_CONTROL + i * 4, blend_cntl);
503 continue;
504 }
505
506 /* Only addition and subtraction equations are supported with
507 * dual source blending.
508 */
509 if (blend->dual_src_blend &&
510 (eqRGB == PIPE_BLEND_MIN || eqRGB == PIPE_BLEND_MAX ||
511 eqA == PIPE_BLEND_MIN || eqA == PIPE_BLEND_MAX)) {
512 assert(!"Unsupported equation for dual source blending");
513 si_pm4_set_reg(pm4, R_028780_CB_BLEND0_CONTROL + i * 4, blend_cntl);
514 continue;
515 }
516
517 /* cb_render_state will disable unused ones */
518 blend->cb_target_mask |= (unsigned)state->rt[j].colormask << (4 * i);
519 if (state->rt[j].colormask)
520 blend->cb_target_enabled_4bit |= 0xf << (4 * i);
521
522 if (!state->rt[j].colormask || !state->rt[j].blend_enable) {
523 si_pm4_set_reg(pm4, R_028780_CB_BLEND0_CONTROL + i * 4, blend_cntl);
524 continue;
525 }
526
527 si_blend_check_commutativity(sctx->screen, blend,
528 eqRGB, srcRGB, dstRGB, 0x7 << (4 * i));
529 si_blend_check_commutativity(sctx->screen, blend,
530 eqA, srcA, dstA, 0x8 << (4 * i));
531
532 /* Blending optimizations for RB+.
533 * These transformations don't change the behavior.
534 *
535 * First, get rid of DST in the blend factors:
536 * func(src * DST, dst * 0) ---> func(src * 0, dst * SRC)
537 */
538 si_blend_remove_dst(&eqRGB, &srcRGB, &dstRGB,
539 PIPE_BLENDFACTOR_DST_COLOR,
540 PIPE_BLENDFACTOR_SRC_COLOR);
541 si_blend_remove_dst(&eqA, &srcA, &dstA,
542 PIPE_BLENDFACTOR_DST_COLOR,
543 PIPE_BLENDFACTOR_SRC_COLOR);
544 si_blend_remove_dst(&eqA, &srcA, &dstA,
545 PIPE_BLENDFACTOR_DST_ALPHA,
546 PIPE_BLENDFACTOR_SRC_ALPHA);
547
548 /* Look up the ideal settings from tables. */
549 srcRGB_opt = si_translate_blend_opt_factor(srcRGB, false);
550 dstRGB_opt = si_translate_blend_opt_factor(dstRGB, false);
551 srcA_opt = si_translate_blend_opt_factor(srcA, true);
552 dstA_opt = si_translate_blend_opt_factor(dstA, true);
553
554 /* Handle interdependencies. */
555 if (si_blend_factor_uses_dst(srcRGB))
556 dstRGB_opt = V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE;
557 if (si_blend_factor_uses_dst(srcA))
558 dstA_opt = V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE;
559
560 if (srcRGB == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE &&
561 (dstRGB == PIPE_BLENDFACTOR_ZERO ||
562 dstRGB == PIPE_BLENDFACTOR_SRC_ALPHA ||
563 dstRGB == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE))
564 dstRGB_opt = V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_A0;
565
566 /* Set the final value. */
567 sx_mrt_blend_opt[i] =
568 S_028760_COLOR_SRC_OPT(srcRGB_opt) |
569 S_028760_COLOR_DST_OPT(dstRGB_opt) |
570 S_028760_COLOR_COMB_FCN(si_translate_blend_opt_function(eqRGB)) |
571 S_028760_ALPHA_SRC_OPT(srcA_opt) |
572 S_028760_ALPHA_DST_OPT(dstA_opt) |
573 S_028760_ALPHA_COMB_FCN(si_translate_blend_opt_function(eqA));
574
575 /* Set blend state. */
576 blend_cntl |= S_028780_ENABLE(1);
577 blend_cntl |= S_028780_COLOR_COMB_FCN(si_translate_blend_function(eqRGB));
578 blend_cntl |= S_028780_COLOR_SRCBLEND(si_translate_blend_factor(srcRGB));
579 blend_cntl |= S_028780_COLOR_DESTBLEND(si_translate_blend_factor(dstRGB));
580
581 if (srcA != srcRGB || dstA != dstRGB || eqA != eqRGB) {
582 blend_cntl |= S_028780_SEPARATE_ALPHA_BLEND(1);
583 blend_cntl |= S_028780_ALPHA_COMB_FCN(si_translate_blend_function(eqA));
584 blend_cntl |= S_028780_ALPHA_SRCBLEND(si_translate_blend_factor(srcA));
585 blend_cntl |= S_028780_ALPHA_DESTBLEND(si_translate_blend_factor(dstA));
586 }
587 si_pm4_set_reg(pm4, R_028780_CB_BLEND0_CONTROL + i * 4, blend_cntl);
588
589 blend->blend_enable_4bit |= 0xfu << (i * 4);
590
591 /* This is only important for formats without alpha. */
592 if (srcRGB == PIPE_BLENDFACTOR_SRC_ALPHA ||
593 dstRGB == PIPE_BLENDFACTOR_SRC_ALPHA ||
594 srcRGB == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE ||
595 dstRGB == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE ||
596 srcRGB == PIPE_BLENDFACTOR_INV_SRC_ALPHA ||
597 dstRGB == PIPE_BLENDFACTOR_INV_SRC_ALPHA)
598 blend->need_src_alpha_4bit |= 0xfu << (i * 4);
599 }
600
601 if (blend->cb_target_mask) {
602 color_control |= S_028808_MODE(mode);
603 } else {
604 color_control |= S_028808_MODE(V_028808_CB_DISABLE);
605 }
606
607 if (sctx->screen->rbplus_allowed) {
608 /* Disable RB+ blend optimizations for dual source blending.
609 * Vulkan does this.
610 */
611 if (blend->dual_src_blend) {
612 for (int i = 0; i < 8; i++) {
613 sx_mrt_blend_opt[i] =
614 S_028760_COLOR_COMB_FCN(V_028760_OPT_COMB_NONE) |
615 S_028760_ALPHA_COMB_FCN(V_028760_OPT_COMB_NONE);
616 }
617 }
618
619 for (int i = 0; i < 8; i++)
620 si_pm4_set_reg(pm4, R_028760_SX_MRT0_BLEND_OPT + i * 4,
621 sx_mrt_blend_opt[i]);
622
623 /* RB+ doesn't work with dual source blending, logic op, and RESOLVE. */
624 if (blend->dual_src_blend || state->logicop_enable ||
625 mode == V_028808_CB_RESOLVE)
626 color_control |= S_028808_DISABLE_DUAL_QUAD(1);
627 }
628
629 si_pm4_set_reg(pm4, R_028808_CB_COLOR_CONTROL, color_control);
630 return blend;
631 }
632
633 static void *si_create_blend_state(struct pipe_context *ctx,
634 const struct pipe_blend_state *state)
635 {
636 return si_create_blend_state_mode(ctx, state, V_028808_CB_NORMAL);
637 }
638
639 static void si_bind_blend_state(struct pipe_context *ctx, void *state)
640 {
641 struct si_context *sctx = (struct si_context *)ctx;
642 struct si_state_blend *old_blend = sctx->queued.named.blend;
643 struct si_state_blend *blend = (struct si_state_blend *)state;
644
645 if (!state)
646 return;
647
648 si_pm4_bind_state(sctx, blend, state);
649
650 if (!old_blend ||
651 old_blend->cb_target_mask != blend->cb_target_mask ||
652 old_blend->dual_src_blend != blend->dual_src_blend ||
653 (old_blend->blend_enable_4bit != blend->blend_enable_4bit &&
654 sctx->framebuffer.nr_samples >= 2 &&
655 sctx->screen->dcc_msaa_allowed))
656 si_mark_atom_dirty(sctx, &sctx->atoms.s.cb_render_state);
657
658 if (!old_blend ||
659 old_blend->cb_target_mask != blend->cb_target_mask ||
660 old_blend->alpha_to_coverage != blend->alpha_to_coverage ||
661 old_blend->alpha_to_one != blend->alpha_to_one ||
662 old_blend->dual_src_blend != blend->dual_src_blend ||
663 old_blend->blend_enable_4bit != blend->blend_enable_4bit ||
664 old_blend->need_src_alpha_4bit != blend->need_src_alpha_4bit)
665 sctx->do_update_shaders = true;
666
667 if (sctx->screen->dpbb_allowed &&
668 (!old_blend ||
669 old_blend->alpha_to_coverage != blend->alpha_to_coverage ||
670 old_blend->blend_enable_4bit != blend->blend_enable_4bit ||
671 old_blend->cb_target_enabled_4bit != blend->cb_target_enabled_4bit))
672 si_mark_atom_dirty(sctx, &sctx->atoms.s.dpbb_state);
673
674 if (sctx->screen->has_out_of_order_rast &&
675 (!old_blend ||
676 (old_blend->blend_enable_4bit != blend->blend_enable_4bit ||
677 old_blend->cb_target_enabled_4bit != blend->cb_target_enabled_4bit ||
678 old_blend->commutative_4bit != blend->commutative_4bit ||
679 old_blend->logicop_enable != blend->logicop_enable)))
680 si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
681 }
682
683 static void si_delete_blend_state(struct pipe_context *ctx, void *state)
684 {
685 struct si_context *sctx = (struct si_context *)ctx;
686 si_pm4_delete_state(sctx, blend, (struct si_state_blend *)state);
687 }
688
689 static void si_set_blend_color(struct pipe_context *ctx,
690 const struct pipe_blend_color *state)
691 {
692 struct si_context *sctx = (struct si_context *)ctx;
693 static const struct pipe_blend_color zeros;
694
695 sctx->blend_color.state = *state;
696 sctx->blend_color.any_nonzeros = memcmp(state, &zeros, sizeof(*state)) != 0;
697 si_mark_atom_dirty(sctx, &sctx->atoms.s.blend_color);
698 }
699
700 static void si_emit_blend_color(struct si_context *sctx)
701 {
702 struct radeon_winsys_cs *cs = sctx->gfx_cs;
703
704 radeon_set_context_reg_seq(cs, R_028414_CB_BLEND_RED, 4);
705 radeon_emit_array(cs, (uint32_t*)sctx->blend_color.state.color, 4);
706 }
707
708 /*
709 * Clipping
710 */
711
712 static void si_set_clip_state(struct pipe_context *ctx,
713 const struct pipe_clip_state *state)
714 {
715 struct si_context *sctx = (struct si_context *)ctx;
716 struct pipe_constant_buffer cb;
717 static const struct pipe_clip_state zeros;
718
719 if (memcmp(&sctx->clip_state.state, state, sizeof(*state)) == 0)
720 return;
721
722 sctx->clip_state.state = *state;
723 sctx->clip_state.any_nonzeros = memcmp(state, &zeros, sizeof(*state)) != 0;
724 si_mark_atom_dirty(sctx, &sctx->atoms.s.clip_state);
725
726 cb.buffer = NULL;
727 cb.user_buffer = state->ucp;
728 cb.buffer_offset = 0;
729 cb.buffer_size = 4*4*8;
730 si_set_rw_buffer(sctx, SI_VS_CONST_CLIP_PLANES, &cb);
731 pipe_resource_reference(&cb.buffer, NULL);
732 }
733
734 static void si_emit_clip_state(struct si_context *sctx)
735 {
736 struct radeon_winsys_cs *cs = sctx->gfx_cs;
737
738 radeon_set_context_reg_seq(cs, R_0285BC_PA_CL_UCP_0_X, 6*4);
739 radeon_emit_array(cs, (uint32_t*)sctx->clip_state.state.ucp, 6*4);
740 }
741
742 static void si_emit_clip_regs(struct si_context *sctx)
743 {
744 struct radeon_winsys_cs *cs = sctx->gfx_cs;
745 struct si_shader *vs = si_get_vs_state(sctx);
746 struct si_shader_selector *vs_sel = vs->selector;
747 struct tgsi_shader_info *info = &vs_sel->info;
748 struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
749 unsigned window_space =
750 info->properties[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION];
751 unsigned clipdist_mask = vs_sel->clipdist_mask;
752 unsigned ucp_mask = clipdist_mask ? 0 : rs->clip_plane_enable & SIX_BITS;
753 unsigned culldist_mask = vs_sel->culldist_mask;
754 unsigned total_mask;
755
756 if (vs->key.opt.clip_disable) {
757 assert(!info->culldist_writemask);
758 clipdist_mask = 0;
759 culldist_mask = 0;
760 }
761 total_mask = clipdist_mask | culldist_mask;
762
763 /* Clip distances on points have no effect, so need to be implemented
764 * as cull distances. This applies for the clipvertex case as well.
765 *
766 * Setting this for primitives other than points should have no adverse
767 * effects.
768 */
769 clipdist_mask &= rs->clip_plane_enable;
770 culldist_mask |= clipdist_mask;
771
772 radeon_set_context_reg(cs, R_02881C_PA_CL_VS_OUT_CNTL,
773 vs_sel->pa_cl_vs_out_cntl |
774 S_02881C_VS_OUT_CCDIST0_VEC_ENA((total_mask & 0x0F) != 0) |
775 S_02881C_VS_OUT_CCDIST1_VEC_ENA((total_mask & 0xF0) != 0) |
776 clipdist_mask | (culldist_mask << 8));
777 radeon_set_context_reg(cs, R_028810_PA_CL_CLIP_CNTL,
778 rs->pa_cl_clip_cntl |
779 ucp_mask |
780 S_028810_CLIP_DISABLE(window_space));
781 }
782
783 /*
784 * inferred state between framebuffer and rasterizer
785 */
786 static void si_update_poly_offset_state(struct si_context *sctx)
787 {
788 struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
789
790 if (!rs || !rs->uses_poly_offset || !sctx->framebuffer.state.zsbuf) {
791 si_pm4_bind_state(sctx, poly_offset, NULL);
792 return;
793 }
794
795 /* Use the user format, not db_render_format, so that the polygon
796 * offset behaves as expected by applications.
797 */
798 switch (sctx->framebuffer.state.zsbuf->texture->format) {
799 case PIPE_FORMAT_Z16_UNORM:
800 si_pm4_bind_state(sctx, poly_offset, &rs->pm4_poly_offset[0]);
801 break;
802 default: /* 24-bit */
803 si_pm4_bind_state(sctx, poly_offset, &rs->pm4_poly_offset[1]);
804 break;
805 case PIPE_FORMAT_Z32_FLOAT:
806 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
807 si_pm4_bind_state(sctx, poly_offset, &rs->pm4_poly_offset[2]);
808 break;
809 }
810 }
811
812 /*
813 * Rasterizer
814 */
815
816 static uint32_t si_translate_fill(uint32_t func)
817 {
818 switch(func) {
819 case PIPE_POLYGON_MODE_FILL:
820 return V_028814_X_DRAW_TRIANGLES;
821 case PIPE_POLYGON_MODE_LINE:
822 return V_028814_X_DRAW_LINES;
823 case PIPE_POLYGON_MODE_POINT:
824 return V_028814_X_DRAW_POINTS;
825 default:
826 assert(0);
827 return V_028814_X_DRAW_POINTS;
828 }
829 }
830
831 static void *si_create_rs_state(struct pipe_context *ctx,
832 const struct pipe_rasterizer_state *state)
833 {
834 struct si_screen *sscreen = ((struct si_context *)ctx)->screen;
835 struct si_state_rasterizer *rs = CALLOC_STRUCT(si_state_rasterizer);
836 struct si_pm4_state *pm4 = &rs->pm4;
837 unsigned tmp, i;
838 float psize_min, psize_max;
839
840 if (!rs) {
841 return NULL;
842 }
843
844 rs->scissor_enable = state->scissor;
845 rs->clip_halfz = state->clip_halfz;
846 rs->two_side = state->light_twoside;
847 rs->multisample_enable = state->multisample;
848 rs->force_persample_interp = state->force_persample_interp;
849 rs->clip_plane_enable = state->clip_plane_enable;
850 rs->line_stipple_enable = state->line_stipple_enable;
851 rs->poly_stipple_enable = state->poly_stipple_enable;
852 rs->line_smooth = state->line_smooth;
853 rs->line_width = state->line_width;
854 rs->poly_smooth = state->poly_smooth;
855 rs->uses_poly_offset = state->offset_point || state->offset_line ||
856 state->offset_tri;
857 rs->clamp_fragment_color = state->clamp_fragment_color;
858 rs->clamp_vertex_color = state->clamp_vertex_color;
859 rs->flatshade = state->flatshade;
860 rs->sprite_coord_enable = state->sprite_coord_enable;
861 rs->rasterizer_discard = state->rasterizer_discard;
862 rs->pa_sc_line_stipple = state->line_stipple_enable ?
863 S_028A0C_LINE_PATTERN(state->line_stipple_pattern) |
864 S_028A0C_REPEAT_COUNT(state->line_stipple_factor) : 0;
865 rs->pa_cl_clip_cntl =
866 S_028810_DX_CLIP_SPACE_DEF(state->clip_halfz) |
867 S_028810_ZCLIP_NEAR_DISABLE(!state->depth_clip) |
868 S_028810_ZCLIP_FAR_DISABLE(!state->depth_clip) |
869 S_028810_DX_RASTERIZATION_KILL(state->rasterizer_discard) |
870 S_028810_DX_LINEAR_ATTR_CLIP_ENA(1);
871
872 si_pm4_set_reg(pm4, R_0286D4_SPI_INTERP_CONTROL_0,
873 S_0286D4_FLAT_SHADE_ENA(1) |
874 S_0286D4_PNT_SPRITE_ENA(state->point_quad_rasterization) |
875 S_0286D4_PNT_SPRITE_OVRD_X(V_0286D4_SPI_PNT_SPRITE_SEL_S) |
876 S_0286D4_PNT_SPRITE_OVRD_Y(V_0286D4_SPI_PNT_SPRITE_SEL_T) |
877 S_0286D4_PNT_SPRITE_OVRD_Z(V_0286D4_SPI_PNT_SPRITE_SEL_0) |
878 S_0286D4_PNT_SPRITE_OVRD_W(V_0286D4_SPI_PNT_SPRITE_SEL_1) |
879 S_0286D4_PNT_SPRITE_TOP_1(state->sprite_coord_mode != PIPE_SPRITE_COORD_UPPER_LEFT));
880
881 /* point size 12.4 fixed point */
882 tmp = (unsigned)(state->point_size * 8.0);
883 si_pm4_set_reg(pm4, R_028A00_PA_SU_POINT_SIZE, S_028A00_HEIGHT(tmp) | S_028A00_WIDTH(tmp));
884
885 if (state->point_size_per_vertex) {
886 psize_min = util_get_min_point_size(state);
887 psize_max = 8192;
888 } else {
889 /* Force the point size to be as if the vertex output was disabled. */
890 psize_min = state->point_size;
891 psize_max = state->point_size;
892 }
893 rs->max_point_size = psize_max;
894
895 /* Divide by two, because 0.5 = 1 pixel. */
896 si_pm4_set_reg(pm4, R_028A04_PA_SU_POINT_MINMAX,
897 S_028A04_MIN_SIZE(si_pack_float_12p4(psize_min/2)) |
898 S_028A04_MAX_SIZE(si_pack_float_12p4(psize_max/2)));
899
900 si_pm4_set_reg(pm4, R_028A08_PA_SU_LINE_CNTL,
901 S_028A08_WIDTH(si_pack_float_12p4(state->line_width/2)));
902 si_pm4_set_reg(pm4, R_028A48_PA_SC_MODE_CNTL_0,
903 S_028A48_LINE_STIPPLE_ENABLE(state->line_stipple_enable) |
904 S_028A48_MSAA_ENABLE(state->multisample ||
905 state->poly_smooth ||
906 state->line_smooth) |
907 S_028A48_VPORT_SCISSOR_ENABLE(1) |
908 S_028A48_ALTERNATE_RBS_PER_TILE(sscreen->info.chip_class >= GFX9));
909
910 si_pm4_set_reg(pm4, R_028BE4_PA_SU_VTX_CNTL,
911 S_028BE4_PIX_CENTER(state->half_pixel_center) |
912 S_028BE4_QUANT_MODE(V_028BE4_X_16_8_FIXED_POINT_1_256TH));
913
914 si_pm4_set_reg(pm4, R_028B7C_PA_SU_POLY_OFFSET_CLAMP, fui(state->offset_clamp));
915 si_pm4_set_reg(pm4, R_028814_PA_SU_SC_MODE_CNTL,
916 S_028814_PROVOKING_VTX_LAST(!state->flatshade_first) |
917 S_028814_CULL_FRONT((state->cull_face & PIPE_FACE_FRONT) ? 1 : 0) |
918 S_028814_CULL_BACK((state->cull_face & PIPE_FACE_BACK) ? 1 : 0) |
919 S_028814_FACE(!state->front_ccw) |
920 S_028814_POLY_OFFSET_FRONT_ENABLE(util_get_offset(state, state->fill_front)) |
921 S_028814_POLY_OFFSET_BACK_ENABLE(util_get_offset(state, state->fill_back)) |
922 S_028814_POLY_OFFSET_PARA_ENABLE(state->offset_point || state->offset_line) |
923 S_028814_POLY_MODE(state->fill_front != PIPE_POLYGON_MODE_FILL ||
924 state->fill_back != PIPE_POLYGON_MODE_FILL) |
925 S_028814_POLYMODE_FRONT_PTYPE(si_translate_fill(state->fill_front)) |
926 S_028814_POLYMODE_BACK_PTYPE(si_translate_fill(state->fill_back)));
927
928 if (!rs->uses_poly_offset)
929 return rs;
930
931 rs->pm4_poly_offset = CALLOC(3, sizeof(struct si_pm4_state));
932 if (!rs->pm4_poly_offset) {
933 FREE(rs);
934 return NULL;
935 }
936
937 /* Precalculate polygon offset states for 16-bit, 24-bit, and 32-bit zbuffers. */
938 for (i = 0; i < 3; i++) {
939 struct si_pm4_state *pm4 = &rs->pm4_poly_offset[i];
940 float offset_units = state->offset_units;
941 float offset_scale = state->offset_scale * 16.0f;
942 uint32_t pa_su_poly_offset_db_fmt_cntl = 0;
943
944 if (!state->offset_units_unscaled) {
945 switch (i) {
946 case 0: /* 16-bit zbuffer */
947 offset_units *= 4.0f;
948 pa_su_poly_offset_db_fmt_cntl =
949 S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-16);
950 break;
951 case 1: /* 24-bit zbuffer */
952 offset_units *= 2.0f;
953 pa_su_poly_offset_db_fmt_cntl =
954 S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-24);
955 break;
956 case 2: /* 32-bit zbuffer */
957 offset_units *= 1.0f;
958 pa_su_poly_offset_db_fmt_cntl = S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-23) |
959 S_028B78_POLY_OFFSET_DB_IS_FLOAT_FMT(1);
960 break;
961 }
962 }
963
964 si_pm4_set_reg(pm4, R_028B80_PA_SU_POLY_OFFSET_FRONT_SCALE,
965 fui(offset_scale));
966 si_pm4_set_reg(pm4, R_028B84_PA_SU_POLY_OFFSET_FRONT_OFFSET,
967 fui(offset_units));
968 si_pm4_set_reg(pm4, R_028B88_PA_SU_POLY_OFFSET_BACK_SCALE,
969 fui(offset_scale));
970 si_pm4_set_reg(pm4, R_028B8C_PA_SU_POLY_OFFSET_BACK_OFFSET,
971 fui(offset_units));
972 si_pm4_set_reg(pm4, R_028B78_PA_SU_POLY_OFFSET_DB_FMT_CNTL,
973 pa_su_poly_offset_db_fmt_cntl);
974 }
975
976 return rs;
977 }
978
979 static void si_bind_rs_state(struct pipe_context *ctx, void *state)
980 {
981 struct si_context *sctx = (struct si_context *)ctx;
982 struct si_state_rasterizer *old_rs =
983 (struct si_state_rasterizer*)sctx->queued.named.rasterizer;
984 struct si_state_rasterizer *rs = (struct si_state_rasterizer *)state;
985
986 if (!state)
987 return;
988
989 if (!old_rs || old_rs->multisample_enable != rs->multisample_enable) {
990 si_mark_atom_dirty(sctx, &sctx->atoms.s.db_render_state);
991
992 /* Update the small primitive filter workaround if necessary. */
993 if (sctx->screen->has_msaa_sample_loc_bug &&
994 sctx->framebuffer.nr_samples > 1)
995 si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_sample_locs);
996 }
997
998 sctx->current_vs_state &= C_VS_STATE_CLAMP_VERTEX_COLOR;
999 sctx->current_vs_state |= S_VS_STATE_CLAMP_VERTEX_COLOR(rs->clamp_vertex_color);
1000
1001 si_pm4_bind_state(sctx, rasterizer, rs);
1002 si_update_poly_offset_state(sctx);
1003
1004 if (!old_rs ||
1005 (old_rs->scissor_enable != rs->scissor_enable ||
1006 old_rs->line_width != rs->line_width ||
1007 old_rs->max_point_size != rs->max_point_size)) {
1008 sctx->scissors.dirty_mask = (1 << SI_MAX_VIEWPORTS) - 1;
1009 si_mark_atom_dirty(sctx, &sctx->atoms.s.scissors);
1010 }
1011
1012 if (!old_rs ||
1013 old_rs->clip_halfz != rs->clip_halfz) {
1014 sctx->viewports.depth_range_dirty_mask = (1 << SI_MAX_VIEWPORTS) - 1;
1015 si_mark_atom_dirty(sctx, &sctx->atoms.s.viewports);
1016 }
1017
1018 if (!old_rs ||
1019 old_rs->clip_plane_enable != rs->clip_plane_enable ||
1020 old_rs->pa_cl_clip_cntl != rs->pa_cl_clip_cntl)
1021 si_mark_atom_dirty(sctx, &sctx->atoms.s.clip_regs);
1022
1023 sctx->ia_multi_vgt_param_key.u.line_stipple_enabled =
1024 rs->line_stipple_enable;
1025
1026 if (!old_rs ||
1027 old_rs->clip_plane_enable != rs->clip_plane_enable ||
1028 old_rs->rasterizer_discard != rs->rasterizer_discard ||
1029 old_rs->sprite_coord_enable != rs->sprite_coord_enable ||
1030 old_rs->flatshade != rs->flatshade ||
1031 old_rs->two_side != rs->two_side ||
1032 old_rs->multisample_enable != rs->multisample_enable ||
1033 old_rs->poly_stipple_enable != rs->poly_stipple_enable ||
1034 old_rs->poly_smooth != rs->poly_smooth ||
1035 old_rs->line_smooth != rs->line_smooth ||
1036 old_rs->clamp_fragment_color != rs->clamp_fragment_color ||
1037 old_rs->force_persample_interp != rs->force_persample_interp)
1038 sctx->do_update_shaders = true;
1039 }
1040
1041 static void si_delete_rs_state(struct pipe_context *ctx, void *state)
1042 {
1043 struct si_context *sctx = (struct si_context *)ctx;
1044 struct si_state_rasterizer *rs = (struct si_state_rasterizer *)state;
1045
1046 if (sctx->queued.named.rasterizer == state)
1047 si_pm4_bind_state(sctx, poly_offset, NULL);
1048
1049 FREE(rs->pm4_poly_offset);
1050 si_pm4_delete_state(sctx, rasterizer, rs);
1051 }
1052
1053 /*
1054 * infeered state between dsa and stencil ref
1055 */
1056 static void si_emit_stencil_ref(struct si_context *sctx)
1057 {
1058 struct radeon_winsys_cs *cs = sctx->gfx_cs;
1059 struct pipe_stencil_ref *ref = &sctx->stencil_ref.state;
1060 struct si_dsa_stencil_ref_part *dsa = &sctx->stencil_ref.dsa_part;
1061
1062 radeon_set_context_reg_seq(cs, R_028430_DB_STENCILREFMASK, 2);
1063 radeon_emit(cs, S_028430_STENCILTESTVAL(ref->ref_value[0]) |
1064 S_028430_STENCILMASK(dsa->valuemask[0]) |
1065 S_028430_STENCILWRITEMASK(dsa->writemask[0]) |
1066 S_028430_STENCILOPVAL(1));
1067 radeon_emit(cs, S_028434_STENCILTESTVAL_BF(ref->ref_value[1]) |
1068 S_028434_STENCILMASK_BF(dsa->valuemask[1]) |
1069 S_028434_STENCILWRITEMASK_BF(dsa->writemask[1]) |
1070 S_028434_STENCILOPVAL_BF(1));
1071 }
1072
1073 static void si_set_stencil_ref(struct pipe_context *ctx,
1074 const struct pipe_stencil_ref *state)
1075 {
1076 struct si_context *sctx = (struct si_context *)ctx;
1077
1078 if (memcmp(&sctx->stencil_ref.state, state, sizeof(*state)) == 0)
1079 return;
1080
1081 sctx->stencil_ref.state = *state;
1082 si_mark_atom_dirty(sctx, &sctx->atoms.s.stencil_ref);
1083 }
1084
1085
1086 /*
1087 * DSA
1088 */
1089
1090 static uint32_t si_translate_stencil_op(int s_op)
1091 {
1092 switch (s_op) {
1093 case PIPE_STENCIL_OP_KEEP:
1094 return V_02842C_STENCIL_KEEP;
1095 case PIPE_STENCIL_OP_ZERO:
1096 return V_02842C_STENCIL_ZERO;
1097 case PIPE_STENCIL_OP_REPLACE:
1098 return V_02842C_STENCIL_REPLACE_TEST;
1099 case PIPE_STENCIL_OP_INCR:
1100 return V_02842C_STENCIL_ADD_CLAMP;
1101 case PIPE_STENCIL_OP_DECR:
1102 return V_02842C_STENCIL_SUB_CLAMP;
1103 case PIPE_STENCIL_OP_INCR_WRAP:
1104 return V_02842C_STENCIL_ADD_WRAP;
1105 case PIPE_STENCIL_OP_DECR_WRAP:
1106 return V_02842C_STENCIL_SUB_WRAP;
1107 case PIPE_STENCIL_OP_INVERT:
1108 return V_02842C_STENCIL_INVERT;
1109 default:
1110 PRINT_ERR("Unknown stencil op %d", s_op);
1111 assert(0);
1112 break;
1113 }
1114 return 0;
1115 }
1116
1117 static bool si_dsa_writes_stencil(const struct pipe_stencil_state *s)
1118 {
1119 return s->enabled && s->writemask &&
1120 (s->fail_op != PIPE_STENCIL_OP_KEEP ||
1121 s->zfail_op != PIPE_STENCIL_OP_KEEP ||
1122 s->zpass_op != PIPE_STENCIL_OP_KEEP);
1123 }
1124
1125 static bool si_order_invariant_stencil_op(enum pipe_stencil_op op)
1126 {
1127 /* REPLACE is normally order invariant, except when the stencil
1128 * reference value is written by the fragment shader. Tracking this
1129 * interaction does not seem worth the effort, so be conservative. */
1130 return op != PIPE_STENCIL_OP_INCR &&
1131 op != PIPE_STENCIL_OP_DECR &&
1132 op != PIPE_STENCIL_OP_REPLACE;
1133 }
1134
1135 /* Compute whether, assuming Z writes are disabled, this stencil state is order
1136 * invariant in the sense that the set of passing fragments as well as the
1137 * final stencil buffer result does not depend on the order of fragments. */
1138 static bool si_order_invariant_stencil_state(const struct pipe_stencil_state *state)
1139 {
1140 return !state->enabled || !state->writemask ||
1141 /* The following assumes that Z writes are disabled. */
1142 (state->func == PIPE_FUNC_ALWAYS &&
1143 si_order_invariant_stencil_op(state->zpass_op) &&
1144 si_order_invariant_stencil_op(state->zfail_op)) ||
1145 (state->func == PIPE_FUNC_NEVER &&
1146 si_order_invariant_stencil_op(state->fail_op));
1147 }
1148
1149 static void *si_create_dsa_state(struct pipe_context *ctx,
1150 const struct pipe_depth_stencil_alpha_state *state)
1151 {
1152 struct si_context *sctx = (struct si_context *)ctx;
1153 struct si_state_dsa *dsa = CALLOC_STRUCT(si_state_dsa);
1154 struct si_pm4_state *pm4 = &dsa->pm4;
1155 unsigned db_depth_control;
1156 uint32_t db_stencil_control = 0;
1157
1158 if (!dsa) {
1159 return NULL;
1160 }
1161
1162 dsa->stencil_ref.valuemask[0] = state->stencil[0].valuemask;
1163 dsa->stencil_ref.valuemask[1] = state->stencil[1].valuemask;
1164 dsa->stencil_ref.writemask[0] = state->stencil[0].writemask;
1165 dsa->stencil_ref.writemask[1] = state->stencil[1].writemask;
1166
1167 db_depth_control = S_028800_Z_ENABLE(state->depth.enabled) |
1168 S_028800_Z_WRITE_ENABLE(state->depth.writemask) |
1169 S_028800_ZFUNC(state->depth.func) |
1170 S_028800_DEPTH_BOUNDS_ENABLE(state->depth.bounds_test);
1171
1172 /* stencil */
1173 if (state->stencil[0].enabled) {
1174 db_depth_control |= S_028800_STENCIL_ENABLE(1);
1175 db_depth_control |= S_028800_STENCILFUNC(state->stencil[0].func);
1176 db_stencil_control |= S_02842C_STENCILFAIL(si_translate_stencil_op(state->stencil[0].fail_op));
1177 db_stencil_control |= S_02842C_STENCILZPASS(si_translate_stencil_op(state->stencil[0].zpass_op));
1178 db_stencil_control |= S_02842C_STENCILZFAIL(si_translate_stencil_op(state->stencil[0].zfail_op));
1179
1180 if (state->stencil[1].enabled) {
1181 db_depth_control |= S_028800_BACKFACE_ENABLE(1);
1182 db_depth_control |= S_028800_STENCILFUNC_BF(state->stencil[1].func);
1183 db_stencil_control |= S_02842C_STENCILFAIL_BF(si_translate_stencil_op(state->stencil[1].fail_op));
1184 db_stencil_control |= S_02842C_STENCILZPASS_BF(si_translate_stencil_op(state->stencil[1].zpass_op));
1185 db_stencil_control |= S_02842C_STENCILZFAIL_BF(si_translate_stencil_op(state->stencil[1].zfail_op));
1186 }
1187 }
1188
1189 /* alpha */
1190 if (state->alpha.enabled) {
1191 dsa->alpha_func = state->alpha.func;
1192
1193 si_pm4_set_reg(pm4, R_00B030_SPI_SHADER_USER_DATA_PS_0 +
1194 SI_SGPR_ALPHA_REF * 4, fui(state->alpha.ref_value));
1195 } else {
1196 dsa->alpha_func = PIPE_FUNC_ALWAYS;
1197 }
1198
1199 si_pm4_set_reg(pm4, R_028800_DB_DEPTH_CONTROL, db_depth_control);
1200 if (state->stencil[0].enabled)
1201 si_pm4_set_reg(pm4, R_02842C_DB_STENCIL_CONTROL, db_stencil_control);
1202 if (state->depth.bounds_test) {
1203 si_pm4_set_reg(pm4, R_028020_DB_DEPTH_BOUNDS_MIN, fui(state->depth.bounds_min));
1204 si_pm4_set_reg(pm4, R_028024_DB_DEPTH_BOUNDS_MAX, fui(state->depth.bounds_max));
1205 }
1206
1207 dsa->depth_enabled = state->depth.enabled;
1208 dsa->depth_write_enabled = state->depth.enabled &&
1209 state->depth.writemask;
1210 dsa->stencil_enabled = state->stencil[0].enabled;
1211 dsa->stencil_write_enabled = state->stencil[0].enabled &&
1212 (si_dsa_writes_stencil(&state->stencil[0]) ||
1213 si_dsa_writes_stencil(&state->stencil[1]));
1214 dsa->db_can_write = dsa->depth_write_enabled ||
1215 dsa->stencil_write_enabled;
1216
1217 bool zfunc_is_ordered =
1218 state->depth.func == PIPE_FUNC_NEVER ||
1219 state->depth.func == PIPE_FUNC_LESS ||
1220 state->depth.func == PIPE_FUNC_LEQUAL ||
1221 state->depth.func == PIPE_FUNC_GREATER ||
1222 state->depth.func == PIPE_FUNC_GEQUAL;
1223
1224 bool nozwrite_and_order_invariant_stencil =
1225 !dsa->db_can_write ||
1226 (!dsa->depth_write_enabled &&
1227 si_order_invariant_stencil_state(&state->stencil[0]) &&
1228 si_order_invariant_stencil_state(&state->stencil[1]));
1229
1230 dsa->order_invariance[1].zs =
1231 nozwrite_and_order_invariant_stencil ||
1232 (!dsa->stencil_write_enabled && zfunc_is_ordered);
1233 dsa->order_invariance[0].zs = !dsa->depth_write_enabled || zfunc_is_ordered;
1234
1235 dsa->order_invariance[1].pass_set =
1236 nozwrite_and_order_invariant_stencil ||
1237 (!dsa->stencil_write_enabled &&
1238 (state->depth.func == PIPE_FUNC_ALWAYS ||
1239 state->depth.func == PIPE_FUNC_NEVER));
1240 dsa->order_invariance[0].pass_set =
1241 !dsa->depth_write_enabled ||
1242 (state->depth.func == PIPE_FUNC_ALWAYS ||
1243 state->depth.func == PIPE_FUNC_NEVER);
1244
1245 dsa->order_invariance[1].pass_last =
1246 sctx->screen->assume_no_z_fights &&
1247 !dsa->stencil_write_enabled &&
1248 dsa->depth_write_enabled && zfunc_is_ordered;
1249 dsa->order_invariance[0].pass_last =
1250 sctx->screen->assume_no_z_fights &&
1251 dsa->depth_write_enabled && zfunc_is_ordered;
1252
1253 return dsa;
1254 }
1255
1256 static void si_bind_dsa_state(struct pipe_context *ctx, void *state)
1257 {
1258 struct si_context *sctx = (struct si_context *)ctx;
1259 struct si_state_dsa *old_dsa = sctx->queued.named.dsa;
1260 struct si_state_dsa *dsa = state;
1261
1262 if (!state)
1263 return;
1264
1265 si_pm4_bind_state(sctx, dsa, dsa);
1266
1267 if (memcmp(&dsa->stencil_ref, &sctx->stencil_ref.dsa_part,
1268 sizeof(struct si_dsa_stencil_ref_part)) != 0) {
1269 sctx->stencil_ref.dsa_part = dsa->stencil_ref;
1270 si_mark_atom_dirty(sctx, &sctx->atoms.s.stencil_ref);
1271 }
1272
1273 if (!old_dsa || old_dsa->alpha_func != dsa->alpha_func)
1274 sctx->do_update_shaders = true;
1275
1276 if (sctx->screen->dpbb_allowed &&
1277 (!old_dsa ||
1278 (old_dsa->depth_enabled != dsa->depth_enabled ||
1279 old_dsa->stencil_enabled != dsa->stencil_enabled ||
1280 old_dsa->db_can_write != dsa->db_can_write)))
1281 si_mark_atom_dirty(sctx, &sctx->atoms.s.dpbb_state);
1282
1283 if (sctx->screen->has_out_of_order_rast &&
1284 (!old_dsa ||
1285 memcmp(old_dsa->order_invariance, dsa->order_invariance,
1286 sizeof(old_dsa->order_invariance))))
1287 si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
1288 }
1289
1290 static void si_delete_dsa_state(struct pipe_context *ctx, void *state)
1291 {
1292 struct si_context *sctx = (struct si_context *)ctx;
1293 si_pm4_delete_state(sctx, dsa, (struct si_state_dsa *)state);
1294 }
1295
1296 static void *si_create_db_flush_dsa(struct si_context *sctx)
1297 {
1298 struct pipe_depth_stencil_alpha_state dsa = {};
1299
1300 return sctx->b.create_depth_stencil_alpha_state(&sctx->b, &dsa);
1301 }
1302
1303 /* DB RENDER STATE */
1304
1305 static void si_set_active_query_state(struct pipe_context *ctx, boolean enable)
1306 {
1307 struct si_context *sctx = (struct si_context*)ctx;
1308
1309 /* Pipeline stat & streamout queries. */
1310 if (enable) {
1311 sctx->flags &= ~SI_CONTEXT_STOP_PIPELINE_STATS;
1312 sctx->flags |= SI_CONTEXT_START_PIPELINE_STATS;
1313 } else {
1314 sctx->flags &= ~SI_CONTEXT_START_PIPELINE_STATS;
1315 sctx->flags |= SI_CONTEXT_STOP_PIPELINE_STATS;
1316 }
1317
1318 /* Occlusion queries. */
1319 if (sctx->occlusion_queries_disabled != !enable) {
1320 sctx->occlusion_queries_disabled = !enable;
1321 si_mark_atom_dirty(sctx, &sctx->atoms.s.db_render_state);
1322 }
1323 }
1324
1325 void si_set_occlusion_query_state(struct si_context *sctx,
1326 bool old_perfect_enable)
1327 {
1328 si_mark_atom_dirty(sctx, &sctx->atoms.s.db_render_state);
1329
1330 bool perfect_enable = sctx->num_perfect_occlusion_queries != 0;
1331
1332 if (perfect_enable != old_perfect_enable)
1333 si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
1334 }
1335
1336 void si_save_qbo_state(struct si_context *sctx, struct si_qbo_state *st)
1337 {
1338 st->saved_compute = sctx->cs_shader_state.program;
1339
1340 si_get_pipe_constant_buffer(sctx, PIPE_SHADER_COMPUTE, 0, &st->saved_const0);
1341 si_get_shader_buffers(sctx, PIPE_SHADER_COMPUTE, 0, 3, st->saved_ssbo);
1342 }
1343
1344 static void si_emit_db_render_state(struct si_context *sctx)
1345 {
1346 struct radeon_winsys_cs *cs = sctx->gfx_cs;
1347 struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
1348 unsigned db_shader_control;
1349
1350 radeon_set_context_reg_seq(cs, R_028000_DB_RENDER_CONTROL, 2);
1351
1352 /* DB_RENDER_CONTROL */
1353 if (sctx->dbcb_depth_copy_enabled ||
1354 sctx->dbcb_stencil_copy_enabled) {
1355 radeon_emit(cs,
1356 S_028000_DEPTH_COPY(sctx->dbcb_depth_copy_enabled) |
1357 S_028000_STENCIL_COPY(sctx->dbcb_stencil_copy_enabled) |
1358 S_028000_COPY_CENTROID(1) |
1359 S_028000_COPY_SAMPLE(sctx->dbcb_copy_sample));
1360 } else if (sctx->db_flush_depth_inplace || sctx->db_flush_stencil_inplace) {
1361 radeon_emit(cs,
1362 S_028000_DEPTH_COMPRESS_DISABLE(sctx->db_flush_depth_inplace) |
1363 S_028000_STENCIL_COMPRESS_DISABLE(sctx->db_flush_stencil_inplace));
1364 } else {
1365 radeon_emit(cs,
1366 S_028000_DEPTH_CLEAR_ENABLE(sctx->db_depth_clear) |
1367 S_028000_STENCIL_CLEAR_ENABLE(sctx->db_stencil_clear));
1368 }
1369
1370 /* DB_COUNT_CONTROL (occlusion queries) */
1371 if (sctx->num_occlusion_queries > 0 &&
1372 !sctx->occlusion_queries_disabled) {
1373 bool perfect = sctx->num_perfect_occlusion_queries > 0;
1374
1375 if (sctx->chip_class >= CIK) {
1376 radeon_emit(cs,
1377 S_028004_PERFECT_ZPASS_COUNTS(perfect) |
1378 S_028004_SAMPLE_RATE(sctx->framebuffer.log_samples) |
1379 S_028004_ZPASS_ENABLE(1) |
1380 S_028004_SLICE_EVEN_ENABLE(1) |
1381 S_028004_SLICE_ODD_ENABLE(1));
1382 } else {
1383 radeon_emit(cs,
1384 S_028004_PERFECT_ZPASS_COUNTS(perfect) |
1385 S_028004_SAMPLE_RATE(sctx->framebuffer.log_samples));
1386 }
1387 } else {
1388 /* Disable occlusion queries. */
1389 if (sctx->chip_class >= CIK) {
1390 radeon_emit(cs, 0);
1391 } else {
1392 radeon_emit(cs, S_028004_ZPASS_INCREMENT_DISABLE(1));
1393 }
1394 }
1395
1396 /* DB_RENDER_OVERRIDE2 */
1397 radeon_set_context_reg(cs, R_028010_DB_RENDER_OVERRIDE2,
1398 S_028010_DISABLE_ZMASK_EXPCLEAR_OPTIMIZATION(sctx->db_depth_disable_expclear) |
1399 S_028010_DISABLE_SMEM_EXPCLEAR_OPTIMIZATION(sctx->db_stencil_disable_expclear) |
1400 S_028010_DECOMPRESS_Z_ON_FLUSH(sctx->framebuffer.nr_samples >= 4));
1401
1402 db_shader_control = sctx->ps_db_shader_control;
1403
1404 /* Bug workaround for smoothing (overrasterization) on SI. */
1405 if (sctx->chip_class == SI && sctx->smoothing_enabled) {
1406 db_shader_control &= C_02880C_Z_ORDER;
1407 db_shader_control |= S_02880C_Z_ORDER(V_02880C_LATE_Z);
1408 }
1409
1410 /* Disable the gl_SampleMask fragment shader output if MSAA is disabled. */
1411 if (!rs || !rs->multisample_enable)
1412 db_shader_control &= C_02880C_MASK_EXPORT_ENABLE;
1413
1414 if (sctx->screen->has_rbplus &&
1415 !sctx->screen->rbplus_allowed)
1416 db_shader_control |= S_02880C_DUAL_QUAD_DISABLE(1);
1417
1418 radeon_set_context_reg(cs, R_02880C_DB_SHADER_CONTROL,
1419 db_shader_control);
1420 }
1421
1422 /*
1423 * format translation
1424 */
1425 static uint32_t si_translate_colorformat(enum pipe_format format)
1426 {
1427 const struct util_format_description *desc = util_format_description(format);
1428 if (!desc)
1429 return V_028C70_COLOR_INVALID;
1430
1431 #define HAS_SIZE(x,y,z,w) \
1432 (desc->channel[0].size == (x) && desc->channel[1].size == (y) && \
1433 desc->channel[2].size == (z) && desc->channel[3].size == (w))
1434
1435 if (format == PIPE_FORMAT_R11G11B10_FLOAT) /* isn't plain */
1436 return V_028C70_COLOR_10_11_11;
1437
1438 if (desc->layout != UTIL_FORMAT_LAYOUT_PLAIN)
1439 return V_028C70_COLOR_INVALID;
1440
1441 /* hw cannot support mixed formats (except depth/stencil, since
1442 * stencil is not written to). */
1443 if (desc->is_mixed && desc->colorspace != UTIL_FORMAT_COLORSPACE_ZS)
1444 return V_028C70_COLOR_INVALID;
1445
1446 switch (desc->nr_channels) {
1447 case 1:
1448 switch (desc->channel[0].size) {
1449 case 8:
1450 return V_028C70_COLOR_8;
1451 case 16:
1452 return V_028C70_COLOR_16;
1453 case 32:
1454 return V_028C70_COLOR_32;
1455 }
1456 break;
1457 case 2:
1458 if (desc->channel[0].size == desc->channel[1].size) {
1459 switch (desc->channel[0].size) {
1460 case 8:
1461 return V_028C70_COLOR_8_8;
1462 case 16:
1463 return V_028C70_COLOR_16_16;
1464 case 32:
1465 return V_028C70_COLOR_32_32;
1466 }
1467 } else if (HAS_SIZE(8,24,0,0)) {
1468 return V_028C70_COLOR_24_8;
1469 } else if (HAS_SIZE(24,8,0,0)) {
1470 return V_028C70_COLOR_8_24;
1471 }
1472 break;
1473 case 3:
1474 if (HAS_SIZE(5,6,5,0)) {
1475 return V_028C70_COLOR_5_6_5;
1476 } else if (HAS_SIZE(32,8,24,0)) {
1477 return V_028C70_COLOR_X24_8_32_FLOAT;
1478 }
1479 break;
1480 case 4:
1481 if (desc->channel[0].size == desc->channel[1].size &&
1482 desc->channel[0].size == desc->channel[2].size &&
1483 desc->channel[0].size == desc->channel[3].size) {
1484 switch (desc->channel[0].size) {
1485 case 4:
1486 return V_028C70_COLOR_4_4_4_4;
1487 case 8:
1488 return V_028C70_COLOR_8_8_8_8;
1489 case 16:
1490 return V_028C70_COLOR_16_16_16_16;
1491 case 32:
1492 return V_028C70_COLOR_32_32_32_32;
1493 }
1494 } else if (HAS_SIZE(5,5,5,1)) {
1495 return V_028C70_COLOR_1_5_5_5;
1496 } else if (HAS_SIZE(1,5,5,5)) {
1497 return V_028C70_COLOR_5_5_5_1;
1498 } else if (HAS_SIZE(10,10,10,2)) {
1499 return V_028C70_COLOR_2_10_10_10;
1500 }
1501 break;
1502 }
1503 return V_028C70_COLOR_INVALID;
1504 }
1505
1506 static uint32_t si_colorformat_endian_swap(uint32_t colorformat)
1507 {
1508 if (SI_BIG_ENDIAN) {
1509 switch(colorformat) {
1510 /* 8-bit buffers. */
1511 case V_028C70_COLOR_8:
1512 return V_028C70_ENDIAN_NONE;
1513
1514 /* 16-bit buffers. */
1515 case V_028C70_COLOR_5_6_5:
1516 case V_028C70_COLOR_1_5_5_5:
1517 case V_028C70_COLOR_4_4_4_4:
1518 case V_028C70_COLOR_16:
1519 case V_028C70_COLOR_8_8:
1520 return V_028C70_ENDIAN_8IN16;
1521
1522 /* 32-bit buffers. */
1523 case V_028C70_COLOR_8_8_8_8:
1524 case V_028C70_COLOR_2_10_10_10:
1525 case V_028C70_COLOR_8_24:
1526 case V_028C70_COLOR_24_8:
1527 case V_028C70_COLOR_16_16:
1528 return V_028C70_ENDIAN_8IN32;
1529
1530 /* 64-bit buffers. */
1531 case V_028C70_COLOR_16_16_16_16:
1532 return V_028C70_ENDIAN_8IN16;
1533
1534 case V_028C70_COLOR_32_32:
1535 return V_028C70_ENDIAN_8IN32;
1536
1537 /* 128-bit buffers. */
1538 case V_028C70_COLOR_32_32_32_32:
1539 return V_028C70_ENDIAN_8IN32;
1540 default:
1541 return V_028C70_ENDIAN_NONE; /* Unsupported. */
1542 }
1543 } else {
1544 return V_028C70_ENDIAN_NONE;
1545 }
1546 }
1547
1548 static uint32_t si_translate_dbformat(enum pipe_format format)
1549 {
1550 switch (format) {
1551 case PIPE_FORMAT_Z16_UNORM:
1552 return V_028040_Z_16;
1553 case PIPE_FORMAT_S8_UINT_Z24_UNORM:
1554 case PIPE_FORMAT_X8Z24_UNORM:
1555 case PIPE_FORMAT_Z24X8_UNORM:
1556 case PIPE_FORMAT_Z24_UNORM_S8_UINT:
1557 return V_028040_Z_24; /* deprecated on SI */
1558 case PIPE_FORMAT_Z32_FLOAT:
1559 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
1560 return V_028040_Z_32_FLOAT;
1561 default:
1562 return V_028040_Z_INVALID;
1563 }
1564 }
1565
1566 /*
1567 * Texture translation
1568 */
1569
1570 static uint32_t si_translate_texformat(struct pipe_screen *screen,
1571 enum pipe_format format,
1572 const struct util_format_description *desc,
1573 int first_non_void)
1574 {
1575 struct si_screen *sscreen = (struct si_screen*)screen;
1576 bool enable_compressed_formats = (sscreen->info.drm_major == 2 &&
1577 sscreen->info.drm_minor >= 31) ||
1578 sscreen->info.drm_major == 3;
1579 bool uniform = true;
1580 int i;
1581
1582 /* Colorspace (return non-RGB formats directly). */
1583 switch (desc->colorspace) {
1584 /* Depth stencil formats */
1585 case UTIL_FORMAT_COLORSPACE_ZS:
1586 switch (format) {
1587 case PIPE_FORMAT_Z16_UNORM:
1588 return V_008F14_IMG_DATA_FORMAT_16;
1589 case PIPE_FORMAT_X24S8_UINT:
1590 case PIPE_FORMAT_S8X24_UINT:
1591 /*
1592 * Implemented as an 8_8_8_8 data format to fix texture
1593 * gathers in stencil sampling. This affects at least
1594 * GL45-CTS.texture_cube_map_array.sampling on VI.
1595 */
1596 if (sscreen->info.chip_class <= VI)
1597 return V_008F14_IMG_DATA_FORMAT_8_8_8_8;
1598
1599 if (format == PIPE_FORMAT_X24S8_UINT)
1600 return V_008F14_IMG_DATA_FORMAT_8_24;
1601 else
1602 return V_008F14_IMG_DATA_FORMAT_24_8;
1603 case PIPE_FORMAT_Z24X8_UNORM:
1604 case PIPE_FORMAT_Z24_UNORM_S8_UINT:
1605 return V_008F14_IMG_DATA_FORMAT_8_24;
1606 case PIPE_FORMAT_X8Z24_UNORM:
1607 case PIPE_FORMAT_S8_UINT_Z24_UNORM:
1608 return V_008F14_IMG_DATA_FORMAT_24_8;
1609 case PIPE_FORMAT_S8_UINT:
1610 return V_008F14_IMG_DATA_FORMAT_8;
1611 case PIPE_FORMAT_Z32_FLOAT:
1612 return V_008F14_IMG_DATA_FORMAT_32;
1613 case PIPE_FORMAT_X32_S8X24_UINT:
1614 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
1615 return V_008F14_IMG_DATA_FORMAT_X24_8_32;
1616 default:
1617 goto out_unknown;
1618 }
1619
1620 case UTIL_FORMAT_COLORSPACE_YUV:
1621 goto out_unknown; /* TODO */
1622
1623 case UTIL_FORMAT_COLORSPACE_SRGB:
1624 if (desc->nr_channels != 4 && desc->nr_channels != 1)
1625 goto out_unknown;
1626 break;
1627
1628 default:
1629 break;
1630 }
1631
1632 if (desc->layout == UTIL_FORMAT_LAYOUT_RGTC) {
1633 if (!enable_compressed_formats)
1634 goto out_unknown;
1635
1636 switch (format) {
1637 case PIPE_FORMAT_RGTC1_SNORM:
1638 case PIPE_FORMAT_LATC1_SNORM:
1639 case PIPE_FORMAT_RGTC1_UNORM:
1640 case PIPE_FORMAT_LATC1_UNORM:
1641 return V_008F14_IMG_DATA_FORMAT_BC4;
1642 case PIPE_FORMAT_RGTC2_SNORM:
1643 case PIPE_FORMAT_LATC2_SNORM:
1644 case PIPE_FORMAT_RGTC2_UNORM:
1645 case PIPE_FORMAT_LATC2_UNORM:
1646 return V_008F14_IMG_DATA_FORMAT_BC5;
1647 default:
1648 goto out_unknown;
1649 }
1650 }
1651
1652 if (desc->layout == UTIL_FORMAT_LAYOUT_ETC &&
1653 (sscreen->info.family == CHIP_STONEY ||
1654 sscreen->info.family == CHIP_VEGA10 ||
1655 sscreen->info.family == CHIP_RAVEN)) {
1656 switch (format) {
1657 case PIPE_FORMAT_ETC1_RGB8:
1658 case PIPE_FORMAT_ETC2_RGB8:
1659 case PIPE_FORMAT_ETC2_SRGB8:
1660 return V_008F14_IMG_DATA_FORMAT_ETC2_RGB;
1661 case PIPE_FORMAT_ETC2_RGB8A1:
1662 case PIPE_FORMAT_ETC2_SRGB8A1:
1663 return V_008F14_IMG_DATA_FORMAT_ETC2_RGBA1;
1664 case PIPE_FORMAT_ETC2_RGBA8:
1665 case PIPE_FORMAT_ETC2_SRGBA8:
1666 return V_008F14_IMG_DATA_FORMAT_ETC2_RGBA;
1667 case PIPE_FORMAT_ETC2_R11_UNORM:
1668 case PIPE_FORMAT_ETC2_R11_SNORM:
1669 return V_008F14_IMG_DATA_FORMAT_ETC2_R;
1670 case PIPE_FORMAT_ETC2_RG11_UNORM:
1671 case PIPE_FORMAT_ETC2_RG11_SNORM:
1672 return V_008F14_IMG_DATA_FORMAT_ETC2_RG;
1673 default:
1674 goto out_unknown;
1675 }
1676 }
1677
1678 if (desc->layout == UTIL_FORMAT_LAYOUT_BPTC) {
1679 if (!enable_compressed_formats)
1680 goto out_unknown;
1681
1682 switch (format) {
1683 case PIPE_FORMAT_BPTC_RGBA_UNORM:
1684 case PIPE_FORMAT_BPTC_SRGBA:
1685 return V_008F14_IMG_DATA_FORMAT_BC7;
1686 case PIPE_FORMAT_BPTC_RGB_FLOAT:
1687 case PIPE_FORMAT_BPTC_RGB_UFLOAT:
1688 return V_008F14_IMG_DATA_FORMAT_BC6;
1689 default:
1690 goto out_unknown;
1691 }
1692 }
1693
1694 if (desc->layout == UTIL_FORMAT_LAYOUT_SUBSAMPLED) {
1695 switch (format) {
1696 case PIPE_FORMAT_R8G8_B8G8_UNORM:
1697 case PIPE_FORMAT_G8R8_B8R8_UNORM:
1698 return V_008F14_IMG_DATA_FORMAT_GB_GR;
1699 case PIPE_FORMAT_G8R8_G8B8_UNORM:
1700 case PIPE_FORMAT_R8G8_R8B8_UNORM:
1701 return V_008F14_IMG_DATA_FORMAT_BG_RG;
1702 default:
1703 goto out_unknown;
1704 }
1705 }
1706
1707 if (desc->layout == UTIL_FORMAT_LAYOUT_S3TC) {
1708 if (!enable_compressed_formats)
1709 goto out_unknown;
1710
1711 switch (format) {
1712 case PIPE_FORMAT_DXT1_RGB:
1713 case PIPE_FORMAT_DXT1_RGBA:
1714 case PIPE_FORMAT_DXT1_SRGB:
1715 case PIPE_FORMAT_DXT1_SRGBA:
1716 return V_008F14_IMG_DATA_FORMAT_BC1;
1717 case PIPE_FORMAT_DXT3_RGBA:
1718 case PIPE_FORMAT_DXT3_SRGBA:
1719 return V_008F14_IMG_DATA_FORMAT_BC2;
1720 case PIPE_FORMAT_DXT5_RGBA:
1721 case PIPE_FORMAT_DXT5_SRGBA:
1722 return V_008F14_IMG_DATA_FORMAT_BC3;
1723 default:
1724 goto out_unknown;
1725 }
1726 }
1727
1728 if (format == PIPE_FORMAT_R9G9B9E5_FLOAT) {
1729 return V_008F14_IMG_DATA_FORMAT_5_9_9_9;
1730 } else if (format == PIPE_FORMAT_R11G11B10_FLOAT) {
1731 return V_008F14_IMG_DATA_FORMAT_10_11_11;
1732 }
1733
1734 /* R8G8Bx_SNORM - TODO CxV8U8 */
1735
1736 /* hw cannot support mixed formats (except depth/stencil, since only
1737 * depth is read).*/
1738 if (desc->is_mixed && desc->colorspace != UTIL_FORMAT_COLORSPACE_ZS)
1739 goto out_unknown;
1740
1741 /* See whether the components are of the same size. */
1742 for (i = 1; i < desc->nr_channels; i++) {
1743 uniform = uniform && desc->channel[0].size == desc->channel[i].size;
1744 }
1745
1746 /* Non-uniform formats. */
1747 if (!uniform) {
1748 switch(desc->nr_channels) {
1749 case 3:
1750 if (desc->channel[0].size == 5 &&
1751 desc->channel[1].size == 6 &&
1752 desc->channel[2].size == 5) {
1753 return V_008F14_IMG_DATA_FORMAT_5_6_5;
1754 }
1755 goto out_unknown;
1756 case 4:
1757 if (desc->channel[0].size == 5 &&
1758 desc->channel[1].size == 5 &&
1759 desc->channel[2].size == 5 &&
1760 desc->channel[3].size == 1) {
1761 return V_008F14_IMG_DATA_FORMAT_1_5_5_5;
1762 }
1763 if (desc->channel[0].size == 1 &&
1764 desc->channel[1].size == 5 &&
1765 desc->channel[2].size == 5 &&
1766 desc->channel[3].size == 5) {
1767 return V_008F14_IMG_DATA_FORMAT_5_5_5_1;
1768 }
1769 if (desc->channel[0].size == 10 &&
1770 desc->channel[1].size == 10 &&
1771 desc->channel[2].size == 10 &&
1772 desc->channel[3].size == 2) {
1773 return V_008F14_IMG_DATA_FORMAT_2_10_10_10;
1774 }
1775 goto out_unknown;
1776 }
1777 goto out_unknown;
1778 }
1779
1780 if (first_non_void < 0 || first_non_void > 3)
1781 goto out_unknown;
1782
1783 /* uniform formats */
1784 switch (desc->channel[first_non_void].size) {
1785 case 4:
1786 switch (desc->nr_channels) {
1787 #if 0 /* Not supported for render targets */
1788 case 2:
1789 return V_008F14_IMG_DATA_FORMAT_4_4;
1790 #endif
1791 case 4:
1792 return V_008F14_IMG_DATA_FORMAT_4_4_4_4;
1793 }
1794 break;
1795 case 8:
1796 switch (desc->nr_channels) {
1797 case 1:
1798 return V_008F14_IMG_DATA_FORMAT_8;
1799 case 2:
1800 return V_008F14_IMG_DATA_FORMAT_8_8;
1801 case 4:
1802 return V_008F14_IMG_DATA_FORMAT_8_8_8_8;
1803 }
1804 break;
1805 case 16:
1806 switch (desc->nr_channels) {
1807 case 1:
1808 return V_008F14_IMG_DATA_FORMAT_16;
1809 case 2:
1810 return V_008F14_IMG_DATA_FORMAT_16_16;
1811 case 4:
1812 return V_008F14_IMG_DATA_FORMAT_16_16_16_16;
1813 }
1814 break;
1815 case 32:
1816 switch (desc->nr_channels) {
1817 case 1:
1818 return V_008F14_IMG_DATA_FORMAT_32;
1819 case 2:
1820 return V_008F14_IMG_DATA_FORMAT_32_32;
1821 #if 0 /* Not supported for render targets */
1822 case 3:
1823 return V_008F14_IMG_DATA_FORMAT_32_32_32;
1824 #endif
1825 case 4:
1826 return V_008F14_IMG_DATA_FORMAT_32_32_32_32;
1827 }
1828 }
1829
1830 out_unknown:
1831 return ~0;
1832 }
1833
1834 static unsigned si_tex_wrap(unsigned wrap)
1835 {
1836 switch (wrap) {
1837 default:
1838 case PIPE_TEX_WRAP_REPEAT:
1839 return V_008F30_SQ_TEX_WRAP;
1840 case PIPE_TEX_WRAP_CLAMP:
1841 return V_008F30_SQ_TEX_CLAMP_HALF_BORDER;
1842 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
1843 return V_008F30_SQ_TEX_CLAMP_LAST_TEXEL;
1844 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
1845 return V_008F30_SQ_TEX_CLAMP_BORDER;
1846 case PIPE_TEX_WRAP_MIRROR_REPEAT:
1847 return V_008F30_SQ_TEX_MIRROR;
1848 case PIPE_TEX_WRAP_MIRROR_CLAMP:
1849 return V_008F30_SQ_TEX_MIRROR_ONCE_HALF_BORDER;
1850 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
1851 return V_008F30_SQ_TEX_MIRROR_ONCE_LAST_TEXEL;
1852 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
1853 return V_008F30_SQ_TEX_MIRROR_ONCE_BORDER;
1854 }
1855 }
1856
1857 static unsigned si_tex_mipfilter(unsigned filter)
1858 {
1859 switch (filter) {
1860 case PIPE_TEX_MIPFILTER_NEAREST:
1861 return V_008F38_SQ_TEX_Z_FILTER_POINT;
1862 case PIPE_TEX_MIPFILTER_LINEAR:
1863 return V_008F38_SQ_TEX_Z_FILTER_LINEAR;
1864 default:
1865 case PIPE_TEX_MIPFILTER_NONE:
1866 return V_008F38_SQ_TEX_Z_FILTER_NONE;
1867 }
1868 }
1869
1870 static unsigned si_tex_compare(unsigned compare)
1871 {
1872 switch (compare) {
1873 default:
1874 case PIPE_FUNC_NEVER:
1875 return V_008F30_SQ_TEX_DEPTH_COMPARE_NEVER;
1876 case PIPE_FUNC_LESS:
1877 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESS;
1878 case PIPE_FUNC_EQUAL:
1879 return V_008F30_SQ_TEX_DEPTH_COMPARE_EQUAL;
1880 case PIPE_FUNC_LEQUAL:
1881 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESSEQUAL;
1882 case PIPE_FUNC_GREATER:
1883 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATER;
1884 case PIPE_FUNC_NOTEQUAL:
1885 return V_008F30_SQ_TEX_DEPTH_COMPARE_NOTEQUAL;
1886 case PIPE_FUNC_GEQUAL:
1887 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATEREQUAL;
1888 case PIPE_FUNC_ALWAYS:
1889 return V_008F30_SQ_TEX_DEPTH_COMPARE_ALWAYS;
1890 }
1891 }
1892
1893 static unsigned si_tex_dim(struct si_screen *sscreen, struct r600_texture *rtex,
1894 unsigned view_target, unsigned nr_samples)
1895 {
1896 unsigned res_target = rtex->buffer.b.b.target;
1897
1898 if (view_target == PIPE_TEXTURE_CUBE ||
1899 view_target == PIPE_TEXTURE_CUBE_ARRAY)
1900 res_target = view_target;
1901 /* If interpreting cubemaps as something else, set 2D_ARRAY. */
1902 else if (res_target == PIPE_TEXTURE_CUBE ||
1903 res_target == PIPE_TEXTURE_CUBE_ARRAY)
1904 res_target = PIPE_TEXTURE_2D_ARRAY;
1905
1906 /* GFX9 allocates 1D textures as 2D. */
1907 if ((res_target == PIPE_TEXTURE_1D ||
1908 res_target == PIPE_TEXTURE_1D_ARRAY) &&
1909 sscreen->info.chip_class >= GFX9 &&
1910 rtex->surface.u.gfx9.resource_type == RADEON_RESOURCE_2D) {
1911 if (res_target == PIPE_TEXTURE_1D)
1912 res_target = PIPE_TEXTURE_2D;
1913 else
1914 res_target = PIPE_TEXTURE_2D_ARRAY;
1915 }
1916
1917 switch (res_target) {
1918 default:
1919 case PIPE_TEXTURE_1D:
1920 return V_008F1C_SQ_RSRC_IMG_1D;
1921 case PIPE_TEXTURE_1D_ARRAY:
1922 return V_008F1C_SQ_RSRC_IMG_1D_ARRAY;
1923 case PIPE_TEXTURE_2D:
1924 case PIPE_TEXTURE_RECT:
1925 return nr_samples > 1 ? V_008F1C_SQ_RSRC_IMG_2D_MSAA :
1926 V_008F1C_SQ_RSRC_IMG_2D;
1927 case PIPE_TEXTURE_2D_ARRAY:
1928 return nr_samples > 1 ? V_008F1C_SQ_RSRC_IMG_2D_MSAA_ARRAY :
1929 V_008F1C_SQ_RSRC_IMG_2D_ARRAY;
1930 case PIPE_TEXTURE_3D:
1931 return V_008F1C_SQ_RSRC_IMG_3D;
1932 case PIPE_TEXTURE_CUBE:
1933 case PIPE_TEXTURE_CUBE_ARRAY:
1934 return V_008F1C_SQ_RSRC_IMG_CUBE;
1935 }
1936 }
1937
1938 /*
1939 * Format support testing
1940 */
1941
1942 static bool si_is_sampler_format_supported(struct pipe_screen *screen, enum pipe_format format)
1943 {
1944 const struct util_format_description *desc = util_format_description(format);
1945 if (!desc)
1946 return false;
1947
1948 return si_translate_texformat(screen, format, desc,
1949 util_format_get_first_non_void_channel(format)) != ~0U;
1950 }
1951
1952 static uint32_t si_translate_buffer_dataformat(struct pipe_screen *screen,
1953 const struct util_format_description *desc,
1954 int first_non_void)
1955 {
1956 int i;
1957
1958 if (desc->format == PIPE_FORMAT_R11G11B10_FLOAT)
1959 return V_008F0C_BUF_DATA_FORMAT_10_11_11;
1960
1961 assert(first_non_void >= 0);
1962
1963 if (desc->nr_channels == 4 &&
1964 desc->channel[0].size == 10 &&
1965 desc->channel[1].size == 10 &&
1966 desc->channel[2].size == 10 &&
1967 desc->channel[3].size == 2)
1968 return V_008F0C_BUF_DATA_FORMAT_2_10_10_10;
1969
1970 /* See whether the components are of the same size. */
1971 for (i = 0; i < desc->nr_channels; i++) {
1972 if (desc->channel[first_non_void].size != desc->channel[i].size)
1973 return V_008F0C_BUF_DATA_FORMAT_INVALID;
1974 }
1975
1976 switch (desc->channel[first_non_void].size) {
1977 case 8:
1978 switch (desc->nr_channels) {
1979 case 1:
1980 case 3: /* 3 loads */
1981 return V_008F0C_BUF_DATA_FORMAT_8;
1982 case 2:
1983 return V_008F0C_BUF_DATA_FORMAT_8_8;
1984 case 4:
1985 return V_008F0C_BUF_DATA_FORMAT_8_8_8_8;
1986 }
1987 break;
1988 case 16:
1989 switch (desc->nr_channels) {
1990 case 1:
1991 case 3: /* 3 loads */
1992 return V_008F0C_BUF_DATA_FORMAT_16;
1993 case 2:
1994 return V_008F0C_BUF_DATA_FORMAT_16_16;
1995 case 4:
1996 return V_008F0C_BUF_DATA_FORMAT_16_16_16_16;
1997 }
1998 break;
1999 case 32:
2000 switch (desc->nr_channels) {
2001 case 1:
2002 return V_008F0C_BUF_DATA_FORMAT_32;
2003 case 2:
2004 return V_008F0C_BUF_DATA_FORMAT_32_32;
2005 case 3:
2006 return V_008F0C_BUF_DATA_FORMAT_32_32_32;
2007 case 4:
2008 return V_008F0C_BUF_DATA_FORMAT_32_32_32_32;
2009 }
2010 break;
2011 case 64:
2012 /* Legacy double formats. */
2013 switch (desc->nr_channels) {
2014 case 1: /* 1 load */
2015 return V_008F0C_BUF_DATA_FORMAT_32_32;
2016 case 2: /* 1 load */
2017 return V_008F0C_BUF_DATA_FORMAT_32_32_32_32;
2018 case 3: /* 3 loads */
2019 return V_008F0C_BUF_DATA_FORMAT_32_32;
2020 case 4: /* 2 loads */
2021 return V_008F0C_BUF_DATA_FORMAT_32_32_32_32;
2022 }
2023 break;
2024 }
2025
2026 return V_008F0C_BUF_DATA_FORMAT_INVALID;
2027 }
2028
2029 static uint32_t si_translate_buffer_numformat(struct pipe_screen *screen,
2030 const struct util_format_description *desc,
2031 int first_non_void)
2032 {
2033 if (desc->format == PIPE_FORMAT_R11G11B10_FLOAT)
2034 return V_008F0C_BUF_NUM_FORMAT_FLOAT;
2035
2036 assert(first_non_void >= 0);
2037
2038 switch (desc->channel[first_non_void].type) {
2039 case UTIL_FORMAT_TYPE_SIGNED:
2040 case UTIL_FORMAT_TYPE_FIXED:
2041 if (desc->channel[first_non_void].size >= 32 ||
2042 desc->channel[first_non_void].pure_integer)
2043 return V_008F0C_BUF_NUM_FORMAT_SINT;
2044 else if (desc->channel[first_non_void].normalized)
2045 return V_008F0C_BUF_NUM_FORMAT_SNORM;
2046 else
2047 return V_008F0C_BUF_NUM_FORMAT_SSCALED;
2048 break;
2049 case UTIL_FORMAT_TYPE_UNSIGNED:
2050 if (desc->channel[first_non_void].size >= 32 ||
2051 desc->channel[first_non_void].pure_integer)
2052 return V_008F0C_BUF_NUM_FORMAT_UINT;
2053 else if (desc->channel[first_non_void].normalized)
2054 return V_008F0C_BUF_NUM_FORMAT_UNORM;
2055 else
2056 return V_008F0C_BUF_NUM_FORMAT_USCALED;
2057 break;
2058 case UTIL_FORMAT_TYPE_FLOAT:
2059 default:
2060 return V_008F0C_BUF_NUM_FORMAT_FLOAT;
2061 }
2062 }
2063
2064 static unsigned si_is_vertex_format_supported(struct pipe_screen *screen,
2065 enum pipe_format format,
2066 unsigned usage)
2067 {
2068 const struct util_format_description *desc;
2069 int first_non_void;
2070 unsigned data_format;
2071
2072 assert((usage & ~(PIPE_BIND_SHADER_IMAGE |
2073 PIPE_BIND_SAMPLER_VIEW |
2074 PIPE_BIND_VERTEX_BUFFER)) == 0);
2075
2076 desc = util_format_description(format);
2077 if (!desc)
2078 return 0;
2079
2080 /* There are no native 8_8_8 or 16_16_16 data formats, and we currently
2081 * select 8_8_8_8 and 16_16_16_16 instead. This works reasonably well
2082 * for read-only access (with caveats surrounding bounds checks), but
2083 * obviously fails for write access which we have to implement for
2084 * shader images. Luckily, OpenGL doesn't expect this to be supported
2085 * anyway, and so the only impact is on PBO uploads / downloads, which
2086 * shouldn't be expected to be fast for GL_RGB anyway.
2087 */
2088 if (desc->block.bits == 3 * 8 ||
2089 desc->block.bits == 3 * 16) {
2090 if (usage & (PIPE_BIND_SHADER_IMAGE | PIPE_BIND_SAMPLER_VIEW)) {
2091 usage &= ~(PIPE_BIND_SHADER_IMAGE | PIPE_BIND_SAMPLER_VIEW);
2092 if (!usage)
2093 return 0;
2094 }
2095 }
2096
2097 first_non_void = util_format_get_first_non_void_channel(format);
2098 data_format = si_translate_buffer_dataformat(screen, desc, first_non_void);
2099 if (data_format == V_008F0C_BUF_DATA_FORMAT_INVALID)
2100 return 0;
2101
2102 return usage;
2103 }
2104
2105 static bool si_is_colorbuffer_format_supported(enum pipe_format format)
2106 {
2107 return si_translate_colorformat(format) != V_028C70_COLOR_INVALID &&
2108 si_translate_colorswap(format, false) != ~0U;
2109 }
2110
2111 static bool si_is_zs_format_supported(enum pipe_format format)
2112 {
2113 return si_translate_dbformat(format) != V_028040_Z_INVALID;
2114 }
2115
2116 static boolean si_is_format_supported(struct pipe_screen *screen,
2117 enum pipe_format format,
2118 enum pipe_texture_target target,
2119 unsigned sample_count,
2120 unsigned usage)
2121 {
2122 struct si_screen *sscreen = (struct si_screen *)screen;
2123 unsigned retval = 0;
2124
2125 if (target >= PIPE_MAX_TEXTURE_TYPES) {
2126 PRINT_ERR("r600: unsupported texture type %d\n", target);
2127 return false;
2128 }
2129
2130 if (!util_format_is_supported(format, usage))
2131 return false;
2132
2133 if (sample_count > 1) {
2134 if (!screen->get_param(screen, PIPE_CAP_TEXTURE_MULTISAMPLE))
2135 return false;
2136
2137 if (usage & PIPE_BIND_SHADER_IMAGE)
2138 return false;
2139
2140 switch (sample_count) {
2141 case 2:
2142 case 4:
2143 case 8:
2144 break;
2145 case 16:
2146 /* Allow resource_copy_region with nr_samples == 16. */
2147 if (sscreen->eqaa_force_coverage_samples == 16 &&
2148 !util_format_is_depth_or_stencil(format))
2149 return true;
2150 if (format == PIPE_FORMAT_NONE)
2151 return true;
2152 else
2153 return false;
2154 default:
2155 return false;
2156 }
2157 }
2158
2159 if (usage & (PIPE_BIND_SAMPLER_VIEW |
2160 PIPE_BIND_SHADER_IMAGE)) {
2161 if (target == PIPE_BUFFER) {
2162 retval |= si_is_vertex_format_supported(
2163 screen, format, usage & (PIPE_BIND_SAMPLER_VIEW |
2164 PIPE_BIND_SHADER_IMAGE));
2165 } else {
2166 if (si_is_sampler_format_supported(screen, format))
2167 retval |= usage & (PIPE_BIND_SAMPLER_VIEW |
2168 PIPE_BIND_SHADER_IMAGE);
2169 }
2170 }
2171
2172 if ((usage & (PIPE_BIND_RENDER_TARGET |
2173 PIPE_BIND_DISPLAY_TARGET |
2174 PIPE_BIND_SCANOUT |
2175 PIPE_BIND_SHARED |
2176 PIPE_BIND_BLENDABLE)) &&
2177 si_is_colorbuffer_format_supported(format)) {
2178 retval |= usage &
2179 (PIPE_BIND_RENDER_TARGET |
2180 PIPE_BIND_DISPLAY_TARGET |
2181 PIPE_BIND_SCANOUT |
2182 PIPE_BIND_SHARED);
2183 if (!util_format_is_pure_integer(format) &&
2184 !util_format_is_depth_or_stencil(format))
2185 retval |= usage & PIPE_BIND_BLENDABLE;
2186 }
2187
2188 if ((usage & PIPE_BIND_DEPTH_STENCIL) &&
2189 si_is_zs_format_supported(format)) {
2190 retval |= PIPE_BIND_DEPTH_STENCIL;
2191 }
2192
2193 if (usage & PIPE_BIND_VERTEX_BUFFER) {
2194 retval |= si_is_vertex_format_supported(screen, format,
2195 PIPE_BIND_VERTEX_BUFFER);
2196 }
2197
2198 if ((usage & PIPE_BIND_LINEAR) &&
2199 !util_format_is_compressed(format) &&
2200 !(usage & PIPE_BIND_DEPTH_STENCIL))
2201 retval |= PIPE_BIND_LINEAR;
2202
2203 return retval == usage;
2204 }
2205
2206 /*
2207 * framebuffer handling
2208 */
2209
2210 static void si_choose_spi_color_formats(struct r600_surface *surf,
2211 unsigned format, unsigned swap,
2212 unsigned ntype, bool is_depth)
2213 {
2214 /* Alpha is needed for alpha-to-coverage.
2215 * Blending may be with or without alpha.
2216 */
2217 unsigned normal = 0; /* most optimal, may not support blending or export alpha */
2218 unsigned alpha = 0; /* exports alpha, but may not support blending */
2219 unsigned blend = 0; /* supports blending, but may not export alpha */
2220 unsigned blend_alpha = 0; /* least optimal, supports blending and exports alpha */
2221
2222 /* Choose the SPI color formats. These are required values for RB+.
2223 * Other chips have multiple choices, though they are not necessarily better.
2224 */
2225 switch (format) {
2226 case V_028C70_COLOR_5_6_5:
2227 case V_028C70_COLOR_1_5_5_5:
2228 case V_028C70_COLOR_5_5_5_1:
2229 case V_028C70_COLOR_4_4_4_4:
2230 case V_028C70_COLOR_10_11_11:
2231 case V_028C70_COLOR_11_11_10:
2232 case V_028C70_COLOR_8:
2233 case V_028C70_COLOR_8_8:
2234 case V_028C70_COLOR_8_8_8_8:
2235 case V_028C70_COLOR_10_10_10_2:
2236 case V_028C70_COLOR_2_10_10_10:
2237 if (ntype == V_028C70_NUMBER_UINT)
2238 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_UINT16_ABGR;
2239 else if (ntype == V_028C70_NUMBER_SINT)
2240 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_SINT16_ABGR;
2241 else
2242 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_FP16_ABGR;
2243 break;
2244
2245 case V_028C70_COLOR_16:
2246 case V_028C70_COLOR_16_16:
2247 case V_028C70_COLOR_16_16_16_16:
2248 if (ntype == V_028C70_NUMBER_UNORM ||
2249 ntype == V_028C70_NUMBER_SNORM) {
2250 /* UNORM16 and SNORM16 don't support blending */
2251 if (ntype == V_028C70_NUMBER_UNORM)
2252 normal = alpha = V_028714_SPI_SHADER_UNORM16_ABGR;
2253 else
2254 normal = alpha = V_028714_SPI_SHADER_SNORM16_ABGR;
2255
2256 /* Use 32 bits per channel for blending. */
2257 if (format == V_028C70_COLOR_16) {
2258 if (swap == V_028C70_SWAP_STD) { /* R */
2259 blend = V_028714_SPI_SHADER_32_R;
2260 blend_alpha = V_028714_SPI_SHADER_32_AR;
2261 } else if (swap == V_028C70_SWAP_ALT_REV) /* A */
2262 blend = blend_alpha = V_028714_SPI_SHADER_32_AR;
2263 else
2264 assert(0);
2265 } else if (format == V_028C70_COLOR_16_16) {
2266 if (swap == V_028C70_SWAP_STD) { /* RG */
2267 blend = V_028714_SPI_SHADER_32_GR;
2268 blend_alpha = V_028714_SPI_SHADER_32_ABGR;
2269 } else if (swap == V_028C70_SWAP_ALT) /* RA */
2270 blend = blend_alpha = V_028714_SPI_SHADER_32_AR;
2271 else
2272 assert(0);
2273 } else /* 16_16_16_16 */
2274 blend = blend_alpha = V_028714_SPI_SHADER_32_ABGR;
2275 } else if (ntype == V_028C70_NUMBER_UINT)
2276 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_UINT16_ABGR;
2277 else if (ntype == V_028C70_NUMBER_SINT)
2278 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_SINT16_ABGR;
2279 else if (ntype == V_028C70_NUMBER_FLOAT)
2280 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_FP16_ABGR;
2281 else
2282 assert(0);
2283 break;
2284
2285 case V_028C70_COLOR_32:
2286 if (swap == V_028C70_SWAP_STD) { /* R */
2287 blend = normal = V_028714_SPI_SHADER_32_R;
2288 alpha = blend_alpha = V_028714_SPI_SHADER_32_AR;
2289 } else if (swap == V_028C70_SWAP_ALT_REV) /* A */
2290 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_32_AR;
2291 else
2292 assert(0);
2293 break;
2294
2295 case V_028C70_COLOR_32_32:
2296 if (swap == V_028C70_SWAP_STD) { /* RG */
2297 blend = normal = V_028714_SPI_SHADER_32_GR;
2298 alpha = blend_alpha = V_028714_SPI_SHADER_32_ABGR;
2299 } else if (swap == V_028C70_SWAP_ALT) /* RA */
2300 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_32_AR;
2301 else
2302 assert(0);
2303 break;
2304
2305 case V_028C70_COLOR_32_32_32_32:
2306 case V_028C70_COLOR_8_24:
2307 case V_028C70_COLOR_24_8:
2308 case V_028C70_COLOR_X24_8_32_FLOAT:
2309 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_32_ABGR;
2310 break;
2311
2312 default:
2313 assert(0);
2314 return;
2315 }
2316
2317 /* The DB->CB copy needs 32_ABGR. */
2318 if (is_depth)
2319 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_32_ABGR;
2320
2321 surf->spi_shader_col_format = normal;
2322 surf->spi_shader_col_format_alpha = alpha;
2323 surf->spi_shader_col_format_blend = blend;
2324 surf->spi_shader_col_format_blend_alpha = blend_alpha;
2325 }
2326
2327 static void si_initialize_color_surface(struct si_context *sctx,
2328 struct r600_surface *surf)
2329 {
2330 struct r600_texture *rtex = (struct r600_texture*)surf->base.texture;
2331 unsigned color_info, color_attrib;
2332 unsigned format, swap, ntype, endian;
2333 const struct util_format_description *desc;
2334 int firstchan;
2335 unsigned blend_clamp = 0, blend_bypass = 0;
2336
2337 desc = util_format_description(surf->base.format);
2338 for (firstchan = 0; firstchan < 4; firstchan++) {
2339 if (desc->channel[firstchan].type != UTIL_FORMAT_TYPE_VOID) {
2340 break;
2341 }
2342 }
2343 if (firstchan == 4 || desc->channel[firstchan].type == UTIL_FORMAT_TYPE_FLOAT) {
2344 ntype = V_028C70_NUMBER_FLOAT;
2345 } else {
2346 ntype = V_028C70_NUMBER_UNORM;
2347 if (desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB)
2348 ntype = V_028C70_NUMBER_SRGB;
2349 else if (desc->channel[firstchan].type == UTIL_FORMAT_TYPE_SIGNED) {
2350 if (desc->channel[firstchan].pure_integer) {
2351 ntype = V_028C70_NUMBER_SINT;
2352 } else {
2353 assert(desc->channel[firstchan].normalized);
2354 ntype = V_028C70_NUMBER_SNORM;
2355 }
2356 } else if (desc->channel[firstchan].type == UTIL_FORMAT_TYPE_UNSIGNED) {
2357 if (desc->channel[firstchan].pure_integer) {
2358 ntype = V_028C70_NUMBER_UINT;
2359 } else {
2360 assert(desc->channel[firstchan].normalized);
2361 ntype = V_028C70_NUMBER_UNORM;
2362 }
2363 }
2364 }
2365
2366 format = si_translate_colorformat(surf->base.format);
2367 if (format == V_028C70_COLOR_INVALID) {
2368 PRINT_ERR("Invalid CB format: %d, disabling CB.\n", surf->base.format);
2369 }
2370 assert(format != V_028C70_COLOR_INVALID);
2371 swap = si_translate_colorswap(surf->base.format, false);
2372 endian = si_colorformat_endian_swap(format);
2373
2374 /* blend clamp should be set for all NORM/SRGB types */
2375 if (ntype == V_028C70_NUMBER_UNORM ||
2376 ntype == V_028C70_NUMBER_SNORM ||
2377 ntype == V_028C70_NUMBER_SRGB)
2378 blend_clamp = 1;
2379
2380 /* set blend bypass according to docs if SINT/UINT or
2381 8/24 COLOR variants */
2382 if (ntype == V_028C70_NUMBER_UINT || ntype == V_028C70_NUMBER_SINT ||
2383 format == V_028C70_COLOR_8_24 || format == V_028C70_COLOR_24_8 ||
2384 format == V_028C70_COLOR_X24_8_32_FLOAT) {
2385 blend_clamp = 0;
2386 blend_bypass = 1;
2387 }
2388
2389 if (ntype == V_028C70_NUMBER_UINT || ntype == V_028C70_NUMBER_SINT) {
2390 if (format == V_028C70_COLOR_8 ||
2391 format == V_028C70_COLOR_8_8 ||
2392 format == V_028C70_COLOR_8_8_8_8)
2393 surf->color_is_int8 = true;
2394 else if (format == V_028C70_COLOR_10_10_10_2 ||
2395 format == V_028C70_COLOR_2_10_10_10)
2396 surf->color_is_int10 = true;
2397 }
2398
2399 color_info = S_028C70_FORMAT(format) |
2400 S_028C70_COMP_SWAP(swap) |
2401 S_028C70_BLEND_CLAMP(blend_clamp) |
2402 S_028C70_BLEND_BYPASS(blend_bypass) |
2403 S_028C70_SIMPLE_FLOAT(1) |
2404 S_028C70_ROUND_MODE(ntype != V_028C70_NUMBER_UNORM &&
2405 ntype != V_028C70_NUMBER_SNORM &&
2406 ntype != V_028C70_NUMBER_SRGB &&
2407 format != V_028C70_COLOR_8_24 &&
2408 format != V_028C70_COLOR_24_8) |
2409 S_028C70_NUMBER_TYPE(ntype) |
2410 S_028C70_ENDIAN(endian);
2411
2412 /* Intensity is implemented as Red, so treat it that way. */
2413 color_attrib = S_028C74_FORCE_DST_ALPHA_1(desc->swizzle[3] == PIPE_SWIZZLE_1 ||
2414 util_format_is_intensity(surf->base.format));
2415
2416 if (rtex->buffer.b.b.nr_samples > 1) {
2417 unsigned log_samples = util_logbase2(rtex->buffer.b.b.nr_samples);
2418 unsigned log_fragments = util_logbase2(rtex->num_color_samples);
2419
2420 color_attrib |= S_028C74_NUM_SAMPLES(log_samples) |
2421 S_028C74_NUM_FRAGMENTS(log_fragments);
2422
2423 if (rtex->surface.fmask_size) {
2424 color_info |= S_028C70_COMPRESSION(1);
2425 unsigned fmask_bankh = util_logbase2(rtex->surface.u.legacy.fmask.bankh);
2426
2427 if (sctx->chip_class == SI) {
2428 /* due to a hw bug, FMASK_BANK_HEIGHT must be set on SI too */
2429 color_attrib |= S_028C74_FMASK_BANK_HEIGHT(fmask_bankh);
2430 }
2431 }
2432 }
2433
2434 if (sctx->chip_class >= VI) {
2435 unsigned max_uncompressed_block_size = V_028C78_MAX_BLOCK_SIZE_256B;
2436 unsigned min_compressed_block_size = V_028C78_MIN_BLOCK_SIZE_32B;
2437
2438 /* amdvlk: [min-compressed-block-size] should be set to 32 for dGPU and
2439 64 for APU because all of our APUs to date use DIMMs which have
2440 a request granularity size of 64B while all other chips have a
2441 32B request size */
2442 if (!sctx->screen->info.has_dedicated_vram)
2443 min_compressed_block_size = V_028C78_MIN_BLOCK_SIZE_64B;
2444
2445 if (rtex->num_color_samples > 1) {
2446 if (rtex->surface.bpe == 1)
2447 max_uncompressed_block_size = V_028C78_MAX_BLOCK_SIZE_64B;
2448 else if (rtex->surface.bpe == 2)
2449 max_uncompressed_block_size = V_028C78_MAX_BLOCK_SIZE_128B;
2450 }
2451
2452 surf->cb_dcc_control = S_028C78_MAX_UNCOMPRESSED_BLOCK_SIZE(max_uncompressed_block_size) |
2453 S_028C78_MIN_COMPRESSED_BLOCK_SIZE(min_compressed_block_size) |
2454 S_028C78_INDEPENDENT_64B_BLOCKS(1);
2455 }
2456
2457 /* This must be set for fast clear to work without FMASK. */
2458 if (!rtex->surface.fmask_size && sctx->chip_class == SI) {
2459 unsigned bankh = util_logbase2(rtex->surface.u.legacy.bankh);
2460 color_attrib |= S_028C74_FMASK_BANK_HEIGHT(bankh);
2461 }
2462
2463 unsigned color_view = S_028C6C_SLICE_START(surf->base.u.tex.first_layer) |
2464 S_028C6C_SLICE_MAX(surf->base.u.tex.last_layer);
2465
2466 if (sctx->chip_class >= GFX9) {
2467 unsigned mip0_depth = util_max_layer(&rtex->buffer.b.b, 0);
2468
2469 color_view |= S_028C6C_MIP_LEVEL(surf->base.u.tex.level);
2470 color_attrib |= S_028C74_MIP0_DEPTH(mip0_depth) |
2471 S_028C74_RESOURCE_TYPE(rtex->surface.u.gfx9.resource_type);
2472 surf->cb_color_attrib2 = S_028C68_MIP0_WIDTH(surf->width0 - 1) |
2473 S_028C68_MIP0_HEIGHT(surf->height0 - 1) |
2474 S_028C68_MAX_MIP(rtex->buffer.b.b.last_level);
2475 }
2476
2477 surf->cb_color_view = color_view;
2478 surf->cb_color_info = color_info;
2479 surf->cb_color_attrib = color_attrib;
2480
2481 /* Determine pixel shader export format */
2482 si_choose_spi_color_formats(surf, format, swap, ntype, rtex->is_depth);
2483
2484 surf->color_initialized = true;
2485 }
2486
2487 static void si_init_depth_surface(struct si_context *sctx,
2488 struct r600_surface *surf)
2489 {
2490 struct r600_texture *rtex = (struct r600_texture*)surf->base.texture;
2491 unsigned level = surf->base.u.tex.level;
2492 unsigned format, stencil_format;
2493 uint32_t z_info, s_info;
2494
2495 format = si_translate_dbformat(rtex->db_render_format);
2496 stencil_format = rtex->surface.has_stencil ?
2497 V_028044_STENCIL_8 : V_028044_STENCIL_INVALID;
2498
2499 assert(format != V_028040_Z_INVALID);
2500 if (format == V_028040_Z_INVALID)
2501 PRINT_ERR("Invalid DB format: %d, disabling DB.\n", rtex->buffer.b.b.format);
2502
2503 surf->db_depth_view = S_028008_SLICE_START(surf->base.u.tex.first_layer) |
2504 S_028008_SLICE_MAX(surf->base.u.tex.last_layer);
2505 surf->db_htile_data_base = 0;
2506 surf->db_htile_surface = 0;
2507
2508 if (sctx->chip_class >= GFX9) {
2509 assert(rtex->surface.u.gfx9.surf_offset == 0);
2510 surf->db_depth_base = rtex->buffer.gpu_address >> 8;
2511 surf->db_stencil_base = (rtex->buffer.gpu_address +
2512 rtex->surface.u.gfx9.stencil_offset) >> 8;
2513 z_info = S_028038_FORMAT(format) |
2514 S_028038_NUM_SAMPLES(util_logbase2(rtex->buffer.b.b.nr_samples)) |
2515 S_028038_SW_MODE(rtex->surface.u.gfx9.surf.swizzle_mode) |
2516 S_028038_MAXMIP(rtex->buffer.b.b.last_level);
2517 s_info = S_02803C_FORMAT(stencil_format) |
2518 S_02803C_SW_MODE(rtex->surface.u.gfx9.stencil.swizzle_mode);
2519 surf->db_z_info2 = S_028068_EPITCH(rtex->surface.u.gfx9.surf.epitch);
2520 surf->db_stencil_info2 = S_02806C_EPITCH(rtex->surface.u.gfx9.stencil.epitch);
2521 surf->db_depth_view |= S_028008_MIPID(level);
2522 surf->db_depth_size = S_02801C_X_MAX(rtex->buffer.b.b.width0 - 1) |
2523 S_02801C_Y_MAX(rtex->buffer.b.b.height0 - 1);
2524
2525 if (si_htile_enabled(rtex, level)) {
2526 z_info |= S_028038_TILE_SURFACE_ENABLE(1) |
2527 S_028038_ALLOW_EXPCLEAR(1);
2528
2529 if (rtex->tc_compatible_htile) {
2530 unsigned max_zplanes = 4;
2531
2532 if (rtex->db_render_format == PIPE_FORMAT_Z16_UNORM &&
2533 rtex->buffer.b.b.nr_samples > 1)
2534 max_zplanes = 2;
2535
2536 z_info |= S_028038_DECOMPRESS_ON_N_ZPLANES(max_zplanes + 1) |
2537 S_028038_ITERATE_FLUSH(1);
2538 s_info |= S_02803C_ITERATE_FLUSH(1);
2539 }
2540
2541 if (rtex->surface.has_stencil) {
2542 /* Stencil buffer workaround ported from the SI-CI-VI code.
2543 * See that for explanation.
2544 */
2545 s_info |= S_02803C_ALLOW_EXPCLEAR(rtex->buffer.b.b.nr_samples <= 1);
2546 } else {
2547 /* Use all HTILE for depth if there's no stencil. */
2548 s_info |= S_02803C_TILE_STENCIL_DISABLE(1);
2549 }
2550
2551 surf->db_htile_data_base = (rtex->buffer.gpu_address +
2552 rtex->htile_offset) >> 8;
2553 surf->db_htile_surface = S_028ABC_FULL_CACHE(1) |
2554 S_028ABC_PIPE_ALIGNED(rtex->surface.u.gfx9.htile.pipe_aligned) |
2555 S_028ABC_RB_ALIGNED(rtex->surface.u.gfx9.htile.rb_aligned);
2556 }
2557 } else {
2558 /* SI-CI-VI */
2559 struct legacy_surf_level *levelinfo = &rtex->surface.u.legacy.level[level];
2560
2561 assert(levelinfo->nblk_x % 8 == 0 && levelinfo->nblk_y % 8 == 0);
2562
2563 surf->db_depth_base = (rtex->buffer.gpu_address +
2564 rtex->surface.u.legacy.level[level].offset) >> 8;
2565 surf->db_stencil_base = (rtex->buffer.gpu_address +
2566 rtex->surface.u.legacy.stencil_level[level].offset) >> 8;
2567
2568 z_info = S_028040_FORMAT(format) |
2569 S_028040_NUM_SAMPLES(util_logbase2(rtex->buffer.b.b.nr_samples));
2570 s_info = S_028044_FORMAT(stencil_format);
2571 surf->db_depth_info = S_02803C_ADDR5_SWIZZLE_MASK(!rtex->tc_compatible_htile);
2572
2573 if (sctx->chip_class >= CIK) {
2574 struct radeon_info *info = &sctx->screen->info;
2575 unsigned index = rtex->surface.u.legacy.tiling_index[level];
2576 unsigned stencil_index = rtex->surface.u.legacy.stencil_tiling_index[level];
2577 unsigned macro_index = rtex->surface.u.legacy.macro_tile_index;
2578 unsigned tile_mode = info->si_tile_mode_array[index];
2579 unsigned stencil_tile_mode = info->si_tile_mode_array[stencil_index];
2580 unsigned macro_mode = info->cik_macrotile_mode_array[macro_index];
2581
2582 surf->db_depth_info |=
2583 S_02803C_ARRAY_MODE(G_009910_ARRAY_MODE(tile_mode)) |
2584 S_02803C_PIPE_CONFIG(G_009910_PIPE_CONFIG(tile_mode)) |
2585 S_02803C_BANK_WIDTH(G_009990_BANK_WIDTH(macro_mode)) |
2586 S_02803C_BANK_HEIGHT(G_009990_BANK_HEIGHT(macro_mode)) |
2587 S_02803C_MACRO_TILE_ASPECT(G_009990_MACRO_TILE_ASPECT(macro_mode)) |
2588 S_02803C_NUM_BANKS(G_009990_NUM_BANKS(macro_mode));
2589 z_info |= S_028040_TILE_SPLIT(G_009910_TILE_SPLIT(tile_mode));
2590 s_info |= S_028044_TILE_SPLIT(G_009910_TILE_SPLIT(stencil_tile_mode));
2591 } else {
2592 unsigned tile_mode_index = si_tile_mode_index(rtex, level, false);
2593 z_info |= S_028040_TILE_MODE_INDEX(tile_mode_index);
2594 tile_mode_index = si_tile_mode_index(rtex, level, true);
2595 s_info |= S_028044_TILE_MODE_INDEX(tile_mode_index);
2596 }
2597
2598 surf->db_depth_size = S_028058_PITCH_TILE_MAX((levelinfo->nblk_x / 8) - 1) |
2599 S_028058_HEIGHT_TILE_MAX((levelinfo->nblk_y / 8) - 1);
2600 surf->db_depth_slice = S_02805C_SLICE_TILE_MAX((levelinfo->nblk_x *
2601 levelinfo->nblk_y) / 64 - 1);
2602
2603 if (si_htile_enabled(rtex, level)) {
2604 z_info |= S_028040_TILE_SURFACE_ENABLE(1) |
2605 S_028040_ALLOW_EXPCLEAR(1);
2606
2607 if (rtex->surface.has_stencil) {
2608 /* Workaround: For a not yet understood reason, the
2609 * combination of MSAA, fast stencil clear and stencil
2610 * decompress messes with subsequent stencil buffer
2611 * uses. Problem was reproduced on Verde, Bonaire,
2612 * Tonga, and Carrizo.
2613 *
2614 * Disabling EXPCLEAR works around the problem.
2615 *
2616 * Check piglit's arb_texture_multisample-stencil-clear
2617 * test if you want to try changing this.
2618 */
2619 if (rtex->buffer.b.b.nr_samples <= 1)
2620 s_info |= S_028044_ALLOW_EXPCLEAR(1);
2621 } else if (!rtex->tc_compatible_htile) {
2622 /* Use all of the htile_buffer for depth if there's no stencil.
2623 * This must not be set when TC-compatible HTILE is enabled
2624 * due to a hw bug.
2625 */
2626 s_info |= S_028044_TILE_STENCIL_DISABLE(1);
2627 }
2628
2629 surf->db_htile_data_base = (rtex->buffer.gpu_address +
2630 rtex->htile_offset) >> 8;
2631 surf->db_htile_surface = S_028ABC_FULL_CACHE(1);
2632
2633 if (rtex->tc_compatible_htile) {
2634 surf->db_htile_surface |= S_028ABC_TC_COMPATIBLE(1);
2635
2636 /* 0 = full compression. N = only compress up to N-1 Z planes. */
2637 if (rtex->buffer.b.b.nr_samples <= 1)
2638 z_info |= S_028040_DECOMPRESS_ON_N_ZPLANES(5);
2639 else if (rtex->buffer.b.b.nr_samples <= 4)
2640 z_info |= S_028040_DECOMPRESS_ON_N_ZPLANES(3);
2641 else
2642 z_info |= S_028040_DECOMPRESS_ON_N_ZPLANES(2);
2643 }
2644 }
2645 }
2646
2647 surf->db_z_info = z_info;
2648 surf->db_stencil_info = s_info;
2649
2650 surf->depth_initialized = true;
2651 }
2652
2653 void si_update_fb_dirtiness_after_rendering(struct si_context *sctx)
2654 {
2655 if (sctx->decompression_enabled)
2656 return;
2657
2658 if (sctx->framebuffer.state.zsbuf) {
2659 struct pipe_surface *surf = sctx->framebuffer.state.zsbuf;
2660 struct r600_texture *rtex = (struct r600_texture *)surf->texture;
2661
2662 rtex->dirty_level_mask |= 1 << surf->u.tex.level;
2663
2664 if (rtex->surface.has_stencil)
2665 rtex->stencil_dirty_level_mask |= 1 << surf->u.tex.level;
2666 }
2667
2668 unsigned compressed_cb_mask = sctx->framebuffer.compressed_cb_mask;
2669 while (compressed_cb_mask) {
2670 unsigned i = u_bit_scan(&compressed_cb_mask);
2671 struct pipe_surface *surf = sctx->framebuffer.state.cbufs[i];
2672 struct r600_texture *rtex = (struct r600_texture*)surf->texture;
2673
2674 if (rtex->surface.fmask_size)
2675 rtex->dirty_level_mask |= 1 << surf->u.tex.level;
2676 if (rtex->dcc_gather_statistics)
2677 rtex->separate_dcc_dirty = true;
2678 }
2679 }
2680
2681 static void si_dec_framebuffer_counters(const struct pipe_framebuffer_state *state)
2682 {
2683 for (int i = 0; i < state->nr_cbufs; ++i) {
2684 struct r600_surface *surf = NULL;
2685 struct r600_texture *rtex;
2686
2687 if (!state->cbufs[i])
2688 continue;
2689 surf = (struct r600_surface*)state->cbufs[i];
2690 rtex = (struct r600_texture*)surf->base.texture;
2691
2692 p_atomic_dec(&rtex->framebuffers_bound);
2693 }
2694 }
2695
2696 static void si_set_framebuffer_state(struct pipe_context *ctx,
2697 const struct pipe_framebuffer_state *state)
2698 {
2699 struct si_context *sctx = (struct si_context *)ctx;
2700 struct pipe_constant_buffer constbuf = {0};
2701 struct r600_surface *surf = NULL;
2702 struct r600_texture *rtex;
2703 bool old_any_dst_linear = sctx->framebuffer.any_dst_linear;
2704 unsigned old_nr_samples = sctx->framebuffer.nr_samples;
2705 unsigned old_colorbuf_enabled_4bit = sctx->framebuffer.colorbuf_enabled_4bit;
2706 bool old_has_zsbuf = !!sctx->framebuffer.state.zsbuf;
2707 bool old_has_stencil =
2708 old_has_zsbuf &&
2709 ((struct r600_texture*)sctx->framebuffer.state.zsbuf->texture)->surface.has_stencil;
2710 bool unbound = false;
2711 int i;
2712
2713 si_update_fb_dirtiness_after_rendering(sctx);
2714
2715 for (i = 0; i < sctx->framebuffer.state.nr_cbufs; i++) {
2716 if (!sctx->framebuffer.state.cbufs[i])
2717 continue;
2718
2719 rtex = (struct r600_texture*)sctx->framebuffer.state.cbufs[i]->texture;
2720 if (rtex->dcc_gather_statistics)
2721 vi_separate_dcc_stop_query(sctx, rtex);
2722 }
2723
2724 /* Disable DCC if the formats are incompatible. */
2725 for (i = 0; i < state->nr_cbufs; i++) {
2726 if (!state->cbufs[i])
2727 continue;
2728
2729 surf = (struct r600_surface*)state->cbufs[i];
2730 rtex = (struct r600_texture*)surf->base.texture;
2731
2732 if (!surf->dcc_incompatible)
2733 continue;
2734
2735 /* Since the DCC decompression calls back into set_framebuffer-
2736 * _state, we need to unbind the framebuffer, so that
2737 * vi_separate_dcc_stop_query isn't called twice with the same
2738 * color buffer.
2739 */
2740 if (!unbound) {
2741 util_copy_framebuffer_state(&sctx->framebuffer.state, NULL);
2742 unbound = true;
2743 }
2744
2745 if (vi_dcc_enabled(rtex, surf->base.u.tex.level))
2746 if (!si_texture_disable_dcc(sctx, rtex))
2747 si_decompress_dcc(sctx, rtex);
2748
2749 surf->dcc_incompatible = false;
2750 }
2751
2752 /* Only flush TC when changing the framebuffer state, because
2753 * the only client not using TC that can change textures is
2754 * the framebuffer.
2755 *
2756 * Wait for compute shaders because of possible transitions:
2757 * - FB write -> shader read
2758 * - shader write -> FB read
2759 *
2760 * DB caches are flushed on demand (using si_decompress_textures).
2761 *
2762 * When MSAA is enabled, CB and TC caches are flushed on demand
2763 * (after FMASK decompression). Shader write -> FB read transitions
2764 * cannot happen for MSAA textures, because MSAA shader images are
2765 * not supported.
2766 *
2767 * Only flush and wait for CB if there is actually a bound color buffer.
2768 */
2769 if (sctx->framebuffer.uncompressed_cb_mask)
2770 si_make_CB_shader_coherent(sctx, sctx->framebuffer.nr_samples,
2771 sctx->framebuffer.CB_has_shader_readable_metadata);
2772
2773 sctx->flags |= SI_CONTEXT_CS_PARTIAL_FLUSH;
2774
2775 /* u_blitter doesn't invoke depth decompression when it does multiple
2776 * blits in a row, but the only case when it matters for DB is when
2777 * doing generate_mipmap. So here we flush DB manually between
2778 * individual generate_mipmap blits.
2779 * Note that lower mipmap levels aren't compressed.
2780 */
2781 if (sctx->generate_mipmap_for_depth) {
2782 si_make_DB_shader_coherent(sctx, 1, false,
2783 sctx->framebuffer.DB_has_shader_readable_metadata);
2784 } else if (sctx->chip_class == GFX9) {
2785 /* It appears that DB metadata "leaks" in a sequence of:
2786 * - depth clear
2787 * - DCC decompress for shader image writes (with DB disabled)
2788 * - render with DEPTH_BEFORE_SHADER=1
2789 * Flushing DB metadata works around the problem.
2790 */
2791 sctx->flags |= SI_CONTEXT_FLUSH_AND_INV_DB_META;
2792 }
2793
2794 /* Take the maximum of the old and new count. If the new count is lower,
2795 * dirtying is needed to disable the unbound colorbuffers.
2796 */
2797 sctx->framebuffer.dirty_cbufs |=
2798 (1 << MAX2(sctx->framebuffer.state.nr_cbufs, state->nr_cbufs)) - 1;
2799 sctx->framebuffer.dirty_zsbuf |= sctx->framebuffer.state.zsbuf != state->zsbuf;
2800
2801 si_dec_framebuffer_counters(&sctx->framebuffer.state);
2802 util_copy_framebuffer_state(&sctx->framebuffer.state, state);
2803
2804 sctx->framebuffer.colorbuf_enabled_4bit = 0;
2805 sctx->framebuffer.spi_shader_col_format = 0;
2806 sctx->framebuffer.spi_shader_col_format_alpha = 0;
2807 sctx->framebuffer.spi_shader_col_format_blend = 0;
2808 sctx->framebuffer.spi_shader_col_format_blend_alpha = 0;
2809 sctx->framebuffer.color_is_int8 = 0;
2810 sctx->framebuffer.color_is_int10 = 0;
2811
2812 sctx->framebuffer.compressed_cb_mask = 0;
2813 sctx->framebuffer.uncompressed_cb_mask = 0;
2814 sctx->framebuffer.nr_samples = util_framebuffer_get_num_samples(state);
2815 sctx->framebuffer.nr_color_samples = sctx->framebuffer.nr_samples;
2816 sctx->framebuffer.log_samples = util_logbase2(sctx->framebuffer.nr_samples);
2817 sctx->framebuffer.any_dst_linear = false;
2818 sctx->framebuffer.CB_has_shader_readable_metadata = false;
2819 sctx->framebuffer.DB_has_shader_readable_metadata = false;
2820
2821 for (i = 0; i < state->nr_cbufs; i++) {
2822 if (!state->cbufs[i])
2823 continue;
2824
2825 surf = (struct r600_surface*)state->cbufs[i];
2826 rtex = (struct r600_texture*)surf->base.texture;
2827
2828 if (!surf->color_initialized) {
2829 si_initialize_color_surface(sctx, surf);
2830 }
2831
2832 sctx->framebuffer.colorbuf_enabled_4bit |= 0xf << (i * 4);
2833 sctx->framebuffer.spi_shader_col_format |=
2834 surf->spi_shader_col_format << (i * 4);
2835 sctx->framebuffer.spi_shader_col_format_alpha |=
2836 surf->spi_shader_col_format_alpha << (i * 4);
2837 sctx->framebuffer.spi_shader_col_format_blend |=
2838 surf->spi_shader_col_format_blend << (i * 4);
2839 sctx->framebuffer.spi_shader_col_format_blend_alpha |=
2840 surf->spi_shader_col_format_blend_alpha << (i * 4);
2841
2842 if (surf->color_is_int8)
2843 sctx->framebuffer.color_is_int8 |= 1 << i;
2844 if (surf->color_is_int10)
2845 sctx->framebuffer.color_is_int10 |= 1 << i;
2846
2847 if (rtex->surface.fmask_size)
2848 sctx->framebuffer.compressed_cb_mask |= 1 << i;
2849 else
2850 sctx->framebuffer.uncompressed_cb_mask |= 1 << i;
2851
2852 /* Don't update nr_color_samples for non-AA buffers.
2853 * (e.g. destination of MSAA resolve)
2854 */
2855 if (rtex->buffer.b.b.nr_samples >= 2 &&
2856 rtex->num_color_samples < rtex->buffer.b.b.nr_samples) {
2857 sctx->framebuffer.nr_color_samples =
2858 MIN2(sctx->framebuffer.nr_color_samples,
2859 rtex->num_color_samples);
2860 }
2861
2862 if (rtex->surface.is_linear)
2863 sctx->framebuffer.any_dst_linear = true;
2864
2865 if (vi_dcc_enabled(rtex, surf->base.u.tex.level))
2866 sctx->framebuffer.CB_has_shader_readable_metadata = true;
2867
2868 si_context_add_resource_size(sctx, surf->base.texture);
2869
2870 p_atomic_inc(&rtex->framebuffers_bound);
2871
2872 if (rtex->dcc_gather_statistics) {
2873 /* Dirty tracking must be enabled for DCC usage analysis. */
2874 sctx->framebuffer.compressed_cb_mask |= 1 << i;
2875 vi_separate_dcc_start_query(sctx, rtex);
2876 }
2877 }
2878
2879 struct r600_texture *zstex = NULL;
2880
2881 if (state->zsbuf) {
2882 surf = (struct r600_surface*)state->zsbuf;
2883 zstex = (struct r600_texture*)surf->base.texture;
2884
2885 if (!surf->depth_initialized) {
2886 si_init_depth_surface(sctx, surf);
2887 }
2888
2889 if (vi_tc_compat_htile_enabled(zstex, surf->base.u.tex.level))
2890 sctx->framebuffer.DB_has_shader_readable_metadata = true;
2891
2892 si_context_add_resource_size(sctx, surf->base.texture);
2893 }
2894
2895 si_update_ps_colorbuf0_slot(sctx);
2896 si_update_poly_offset_state(sctx);
2897 si_mark_atom_dirty(sctx, &sctx->atoms.s.cb_render_state);
2898 si_mark_atom_dirty(sctx, &sctx->atoms.s.framebuffer);
2899
2900 if (sctx->screen->dpbb_allowed)
2901 si_mark_atom_dirty(sctx, &sctx->atoms.s.dpbb_state);
2902
2903 if (sctx->framebuffer.any_dst_linear != old_any_dst_linear)
2904 si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
2905
2906 if (sctx->screen->has_out_of_order_rast &&
2907 (sctx->framebuffer.colorbuf_enabled_4bit != old_colorbuf_enabled_4bit ||
2908 !!sctx->framebuffer.state.zsbuf != old_has_zsbuf ||
2909 (zstex && zstex->surface.has_stencil != old_has_stencil)))
2910 si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
2911
2912 if (sctx->framebuffer.nr_samples != old_nr_samples) {
2913 si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
2914 si_mark_atom_dirty(sctx, &sctx->atoms.s.db_render_state);
2915
2916 /* Set sample locations as fragment shader constants. */
2917 switch (sctx->framebuffer.nr_samples) {
2918 case 1:
2919 constbuf.user_buffer = sctx->sample_locations_1x;
2920 break;
2921 case 2:
2922 constbuf.user_buffer = sctx->sample_locations_2x;
2923 break;
2924 case 4:
2925 constbuf.user_buffer = sctx->sample_locations_4x;
2926 break;
2927 case 8:
2928 constbuf.user_buffer = sctx->sample_locations_8x;
2929 break;
2930 case 16:
2931 constbuf.user_buffer = sctx->sample_locations_16x;
2932 break;
2933 default:
2934 PRINT_ERR("Requested an invalid number of samples %i.\n",
2935 sctx->framebuffer.nr_samples);
2936 assert(0);
2937 }
2938 constbuf.buffer_size = sctx->framebuffer.nr_samples * 2 * 4;
2939 si_set_rw_buffer(sctx, SI_PS_CONST_SAMPLE_POSITIONS, &constbuf);
2940
2941 si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_sample_locs);
2942 }
2943
2944 sctx->do_update_shaders = true;
2945
2946 if (!sctx->decompression_enabled) {
2947 /* Prevent textures decompression when the framebuffer state
2948 * changes come from the decompression passes themselves.
2949 */
2950 sctx->need_check_render_feedback = true;
2951 }
2952 }
2953
2954 static void si_emit_framebuffer_state(struct si_context *sctx)
2955 {
2956 struct radeon_winsys_cs *cs = sctx->gfx_cs;
2957 struct pipe_framebuffer_state *state = &sctx->framebuffer.state;
2958 unsigned i, nr_cbufs = state->nr_cbufs;
2959 struct r600_texture *tex = NULL;
2960 struct r600_surface *cb = NULL;
2961 unsigned cb_color_info = 0;
2962
2963 /* Colorbuffers. */
2964 for (i = 0; i < nr_cbufs; i++) {
2965 uint64_t cb_color_base, cb_color_fmask, cb_color_cmask, cb_dcc_base;
2966 unsigned cb_color_attrib;
2967
2968 if (!(sctx->framebuffer.dirty_cbufs & (1 << i)))
2969 continue;
2970
2971 cb = (struct r600_surface*)state->cbufs[i];
2972 if (!cb) {
2973 radeon_set_context_reg(cs, R_028C70_CB_COLOR0_INFO + i * 0x3C,
2974 S_028C70_FORMAT(V_028C70_COLOR_INVALID));
2975 continue;
2976 }
2977
2978 tex = (struct r600_texture *)cb->base.texture;
2979 radeon_add_to_buffer_list(sctx, sctx->gfx_cs,
2980 &tex->buffer, RADEON_USAGE_READWRITE,
2981 tex->buffer.b.b.nr_samples > 1 ?
2982 RADEON_PRIO_COLOR_BUFFER_MSAA :
2983 RADEON_PRIO_COLOR_BUFFER);
2984
2985 if (tex->cmask_buffer && tex->cmask_buffer != &tex->buffer) {
2986 radeon_add_to_buffer_list(sctx, sctx->gfx_cs,
2987 tex->cmask_buffer, RADEON_USAGE_READWRITE,
2988 RADEON_PRIO_CMASK);
2989 }
2990
2991 if (tex->dcc_separate_buffer)
2992 radeon_add_to_buffer_list(sctx, sctx->gfx_cs,
2993 tex->dcc_separate_buffer,
2994 RADEON_USAGE_READWRITE,
2995 RADEON_PRIO_DCC);
2996
2997 /* Compute mutable surface parameters. */
2998 cb_color_base = tex->buffer.gpu_address >> 8;
2999 cb_color_fmask = 0;
3000 cb_color_cmask = tex->cmask.base_address_reg;
3001 cb_dcc_base = 0;
3002 cb_color_info = cb->cb_color_info | tex->cb_color_info;
3003 cb_color_attrib = cb->cb_color_attrib;
3004
3005 if (cb->base.u.tex.level > 0)
3006 cb_color_info &= C_028C70_FAST_CLEAR;
3007
3008 if (tex->surface.fmask_size) {
3009 cb_color_fmask = (tex->buffer.gpu_address + tex->fmask_offset) >> 8;
3010 cb_color_fmask |= tex->surface.fmask_tile_swizzle;
3011 }
3012
3013 /* Set up DCC. */
3014 if (vi_dcc_enabled(tex, cb->base.u.tex.level)) {
3015 bool is_msaa_resolve_dst = state->cbufs[0] &&
3016 state->cbufs[0]->texture->nr_samples > 1 &&
3017 state->cbufs[1] == &cb->base &&
3018 state->cbufs[1]->texture->nr_samples <= 1;
3019
3020 if (!is_msaa_resolve_dst)
3021 cb_color_info |= S_028C70_DCC_ENABLE(1);
3022
3023 cb_dcc_base = ((!tex->dcc_separate_buffer ? tex->buffer.gpu_address : 0) +
3024 tex->dcc_offset) >> 8;
3025 cb_dcc_base |= tex->surface.tile_swizzle;
3026 }
3027
3028 if (sctx->chip_class >= GFX9) {
3029 struct gfx9_surf_meta_flags meta;
3030
3031 if (tex->dcc_offset)
3032 meta = tex->surface.u.gfx9.dcc;
3033 else
3034 meta = tex->surface.u.gfx9.cmask;
3035
3036 /* Set mutable surface parameters. */
3037 cb_color_base += tex->surface.u.gfx9.surf_offset >> 8;
3038 cb_color_base |= tex->surface.tile_swizzle;
3039 if (!tex->surface.fmask_size)
3040 cb_color_fmask = cb_color_base;
3041 if (cb->base.u.tex.level > 0)
3042 cb_color_cmask = cb_color_base;
3043 cb_color_attrib |= S_028C74_COLOR_SW_MODE(tex->surface.u.gfx9.surf.swizzle_mode) |
3044 S_028C74_FMASK_SW_MODE(tex->surface.u.gfx9.fmask.swizzle_mode) |
3045 S_028C74_RB_ALIGNED(meta.rb_aligned) |
3046 S_028C74_PIPE_ALIGNED(meta.pipe_aligned);
3047
3048 radeon_set_context_reg_seq(cs, R_028C60_CB_COLOR0_BASE + i * 0x3C, 15);
3049 radeon_emit(cs, cb_color_base); /* CB_COLOR0_BASE */
3050 radeon_emit(cs, S_028C64_BASE_256B(cb_color_base >> 32)); /* CB_COLOR0_BASE_EXT */
3051 radeon_emit(cs, cb->cb_color_attrib2); /* CB_COLOR0_ATTRIB2 */
3052 radeon_emit(cs, cb->cb_color_view); /* CB_COLOR0_VIEW */
3053 radeon_emit(cs, cb_color_info); /* CB_COLOR0_INFO */
3054 radeon_emit(cs, cb_color_attrib); /* CB_COLOR0_ATTRIB */
3055 radeon_emit(cs, cb->cb_dcc_control); /* CB_COLOR0_DCC_CONTROL */
3056 radeon_emit(cs, cb_color_cmask); /* CB_COLOR0_CMASK */
3057 radeon_emit(cs, S_028C80_BASE_256B(cb_color_cmask >> 32)); /* CB_COLOR0_CMASK_BASE_EXT */
3058 radeon_emit(cs, cb_color_fmask); /* CB_COLOR0_FMASK */
3059 radeon_emit(cs, S_028C88_BASE_256B(cb_color_fmask >> 32)); /* CB_COLOR0_FMASK_BASE_EXT */
3060 radeon_emit(cs, tex->color_clear_value[0]); /* CB_COLOR0_CLEAR_WORD0 */
3061 radeon_emit(cs, tex->color_clear_value[1]); /* CB_COLOR0_CLEAR_WORD1 */
3062 radeon_emit(cs, cb_dcc_base); /* CB_COLOR0_DCC_BASE */
3063 radeon_emit(cs, S_028C98_BASE_256B(cb_dcc_base >> 32)); /* CB_COLOR0_DCC_BASE_EXT */
3064
3065 radeon_set_context_reg(cs, R_0287A0_CB_MRT0_EPITCH + i * 4,
3066 S_0287A0_EPITCH(tex->surface.u.gfx9.surf.epitch));
3067 } else {
3068 /* Compute mutable surface parameters (SI-CI-VI). */
3069 const struct legacy_surf_level *level_info =
3070 &tex->surface.u.legacy.level[cb->base.u.tex.level];
3071 unsigned pitch_tile_max, slice_tile_max, tile_mode_index;
3072 unsigned cb_color_pitch, cb_color_slice, cb_color_fmask_slice;
3073
3074 cb_color_base += level_info->offset >> 8;
3075 /* Only macrotiled modes can set tile swizzle. */
3076 if (level_info->mode == RADEON_SURF_MODE_2D)
3077 cb_color_base |= tex->surface.tile_swizzle;
3078
3079 if (!tex->surface.fmask_size)
3080 cb_color_fmask = cb_color_base;
3081 if (cb->base.u.tex.level > 0)
3082 cb_color_cmask = cb_color_base;
3083 if (cb_dcc_base)
3084 cb_dcc_base += level_info->dcc_offset >> 8;
3085
3086 pitch_tile_max = level_info->nblk_x / 8 - 1;
3087 slice_tile_max = level_info->nblk_x *
3088 level_info->nblk_y / 64 - 1;
3089 tile_mode_index = si_tile_mode_index(tex, cb->base.u.tex.level, false);
3090
3091 cb_color_attrib |= S_028C74_TILE_MODE_INDEX(tile_mode_index);
3092 cb_color_pitch = S_028C64_TILE_MAX(pitch_tile_max);
3093 cb_color_slice = S_028C68_TILE_MAX(slice_tile_max);
3094
3095 if (tex->surface.fmask_size) {
3096 if (sctx->chip_class >= CIK)
3097 cb_color_pitch |= S_028C64_FMASK_TILE_MAX(tex->surface.u.legacy.fmask.pitch_in_pixels / 8 - 1);
3098 cb_color_attrib |= S_028C74_FMASK_TILE_MODE_INDEX(tex->surface.u.legacy.fmask.tiling_index);
3099 cb_color_fmask_slice = S_028C88_TILE_MAX(tex->surface.u.legacy.fmask.slice_tile_max);
3100 } else {
3101 /* This must be set for fast clear to work without FMASK. */
3102 if (sctx->chip_class >= CIK)
3103 cb_color_pitch |= S_028C64_FMASK_TILE_MAX(pitch_tile_max);
3104 cb_color_attrib |= S_028C74_FMASK_TILE_MODE_INDEX(tile_mode_index);
3105 cb_color_fmask_slice = S_028C88_TILE_MAX(slice_tile_max);
3106 }
3107
3108 radeon_set_context_reg_seq(cs, R_028C60_CB_COLOR0_BASE + i * 0x3C,
3109 sctx->chip_class >= VI ? 14 : 13);
3110 radeon_emit(cs, cb_color_base); /* CB_COLOR0_BASE */
3111 radeon_emit(cs, cb_color_pitch); /* CB_COLOR0_PITCH */
3112 radeon_emit(cs, cb_color_slice); /* CB_COLOR0_SLICE */
3113 radeon_emit(cs, cb->cb_color_view); /* CB_COLOR0_VIEW */
3114 radeon_emit(cs, cb_color_info); /* CB_COLOR0_INFO */
3115 radeon_emit(cs, cb_color_attrib); /* CB_COLOR0_ATTRIB */
3116 radeon_emit(cs, cb->cb_dcc_control); /* CB_COLOR0_DCC_CONTROL */
3117 radeon_emit(cs, cb_color_cmask); /* CB_COLOR0_CMASK */
3118 radeon_emit(cs, tex->cmask.slice_tile_max); /* CB_COLOR0_CMASK_SLICE */
3119 radeon_emit(cs, cb_color_fmask); /* CB_COLOR0_FMASK */
3120 radeon_emit(cs, cb_color_fmask_slice); /* CB_COLOR0_FMASK_SLICE */
3121 radeon_emit(cs, tex->color_clear_value[0]); /* CB_COLOR0_CLEAR_WORD0 */
3122 radeon_emit(cs, tex->color_clear_value[1]); /* CB_COLOR0_CLEAR_WORD1 */
3123
3124 if (sctx->chip_class >= VI) /* R_028C94_CB_COLOR0_DCC_BASE */
3125 radeon_emit(cs, cb_dcc_base);
3126 }
3127 }
3128 for (; i < 8 ; i++)
3129 if (sctx->framebuffer.dirty_cbufs & (1 << i))
3130 radeon_set_context_reg(cs, R_028C70_CB_COLOR0_INFO + i * 0x3C, 0);
3131
3132 /* ZS buffer. */
3133 if (state->zsbuf && sctx->framebuffer.dirty_zsbuf) {
3134 struct r600_surface *zb = (struct r600_surface*)state->zsbuf;
3135 struct r600_texture *rtex = (struct r600_texture*)zb->base.texture;
3136
3137 radeon_add_to_buffer_list(sctx, sctx->gfx_cs,
3138 &rtex->buffer, RADEON_USAGE_READWRITE,
3139 zb->base.texture->nr_samples > 1 ?
3140 RADEON_PRIO_DEPTH_BUFFER_MSAA :
3141 RADEON_PRIO_DEPTH_BUFFER);
3142
3143 if (sctx->chip_class >= GFX9) {
3144 radeon_set_context_reg_seq(cs, R_028014_DB_HTILE_DATA_BASE, 3);
3145 radeon_emit(cs, zb->db_htile_data_base); /* DB_HTILE_DATA_BASE */
3146 radeon_emit(cs, S_028018_BASE_HI(zb->db_htile_data_base >> 32)); /* DB_HTILE_DATA_BASE_HI */
3147 radeon_emit(cs, zb->db_depth_size); /* DB_DEPTH_SIZE */
3148
3149 radeon_set_context_reg_seq(cs, R_028038_DB_Z_INFO, 10);
3150 radeon_emit(cs, zb->db_z_info | /* DB_Z_INFO */
3151 S_028038_ZRANGE_PRECISION(rtex->depth_clear_value != 0));
3152 radeon_emit(cs, zb->db_stencil_info); /* DB_STENCIL_INFO */
3153 radeon_emit(cs, zb->db_depth_base); /* DB_Z_READ_BASE */
3154 radeon_emit(cs, S_028044_BASE_HI(zb->db_depth_base >> 32)); /* DB_Z_READ_BASE_HI */
3155 radeon_emit(cs, zb->db_stencil_base); /* DB_STENCIL_READ_BASE */
3156 radeon_emit(cs, S_02804C_BASE_HI(zb->db_stencil_base >> 32)); /* DB_STENCIL_READ_BASE_HI */
3157 radeon_emit(cs, zb->db_depth_base); /* DB_Z_WRITE_BASE */
3158 radeon_emit(cs, S_028054_BASE_HI(zb->db_depth_base >> 32)); /* DB_Z_WRITE_BASE_HI */
3159 radeon_emit(cs, zb->db_stencil_base); /* DB_STENCIL_WRITE_BASE */
3160 radeon_emit(cs, S_02805C_BASE_HI(zb->db_stencil_base >> 32)); /* DB_STENCIL_WRITE_BASE_HI */
3161
3162 radeon_set_context_reg_seq(cs, R_028068_DB_Z_INFO2, 2);
3163 radeon_emit(cs, zb->db_z_info2); /* DB_Z_INFO2 */
3164 radeon_emit(cs, zb->db_stencil_info2); /* DB_STENCIL_INFO2 */
3165 } else {
3166 radeon_set_context_reg(cs, R_028014_DB_HTILE_DATA_BASE, zb->db_htile_data_base);
3167
3168 radeon_set_context_reg_seq(cs, R_02803C_DB_DEPTH_INFO, 9);
3169 radeon_emit(cs, zb->db_depth_info); /* DB_DEPTH_INFO */
3170 radeon_emit(cs, zb->db_z_info | /* DB_Z_INFO */
3171 S_028040_ZRANGE_PRECISION(rtex->depth_clear_value != 0));
3172 radeon_emit(cs, zb->db_stencil_info); /* DB_STENCIL_INFO */
3173 radeon_emit(cs, zb->db_depth_base); /* DB_Z_READ_BASE */
3174 radeon_emit(cs, zb->db_stencil_base); /* DB_STENCIL_READ_BASE */
3175 radeon_emit(cs, zb->db_depth_base); /* DB_Z_WRITE_BASE */
3176 radeon_emit(cs, zb->db_stencil_base); /* DB_STENCIL_WRITE_BASE */
3177 radeon_emit(cs, zb->db_depth_size); /* DB_DEPTH_SIZE */
3178 radeon_emit(cs, zb->db_depth_slice); /* DB_DEPTH_SLICE */
3179 }
3180
3181 radeon_set_context_reg_seq(cs, R_028028_DB_STENCIL_CLEAR, 2);
3182 radeon_emit(cs, rtex->stencil_clear_value); /* R_028028_DB_STENCIL_CLEAR */
3183 radeon_emit(cs, fui(rtex->depth_clear_value)); /* R_02802C_DB_DEPTH_CLEAR */
3184
3185 radeon_set_context_reg(cs, R_028008_DB_DEPTH_VIEW, zb->db_depth_view);
3186 radeon_set_context_reg(cs, R_028ABC_DB_HTILE_SURFACE, zb->db_htile_surface);
3187 } else if (sctx->framebuffer.dirty_zsbuf) {
3188 if (sctx->chip_class >= GFX9)
3189 radeon_set_context_reg_seq(cs, R_028038_DB_Z_INFO, 2);
3190 else
3191 radeon_set_context_reg_seq(cs, R_028040_DB_Z_INFO, 2);
3192
3193 radeon_emit(cs, S_028040_FORMAT(V_028040_Z_INVALID)); /* DB_Z_INFO */
3194 radeon_emit(cs, S_028044_FORMAT(V_028044_STENCIL_INVALID)); /* DB_STENCIL_INFO */
3195 }
3196
3197 /* Framebuffer dimensions. */
3198 /* PA_SC_WINDOW_SCISSOR_TL is set in si_init_config() */
3199 radeon_set_context_reg(cs, R_028208_PA_SC_WINDOW_SCISSOR_BR,
3200 S_028208_BR_X(state->width) | S_028208_BR_Y(state->height));
3201
3202 if (sctx->screen->dfsm_allowed) {
3203 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
3204 radeon_emit(cs, EVENT_TYPE(V_028A90_BREAK_BATCH) | EVENT_INDEX(0));
3205 }
3206
3207 sctx->framebuffer.dirty_cbufs = 0;
3208 sctx->framebuffer.dirty_zsbuf = false;
3209 }
3210
3211 static void si_emit_msaa_sample_locs(struct si_context *sctx)
3212 {
3213 struct radeon_winsys_cs *cs = sctx->gfx_cs;
3214 unsigned nr_samples = sctx->framebuffer.nr_samples;
3215 bool has_msaa_sample_loc_bug = sctx->screen->has_msaa_sample_loc_bug;
3216
3217 /* Smoothing (only possible with nr_samples == 1) uses the same
3218 * sample locations as the MSAA it simulates.
3219 */
3220 if (nr_samples <= 1 && sctx->smoothing_enabled)
3221 nr_samples = SI_NUM_SMOOTH_AA_SAMPLES;
3222
3223 /* On Polaris, the small primitive filter uses the sample locations
3224 * even when MSAA is off, so we need to make sure they're set to 0.
3225 */
3226 if (has_msaa_sample_loc_bug)
3227 nr_samples = MAX2(nr_samples, 1);
3228
3229 if (nr_samples != sctx->sample_locs_num_samples) {
3230 sctx->sample_locs_num_samples = nr_samples;
3231 si_emit_sample_locations(cs, nr_samples);
3232 }
3233
3234 if (sctx->family >= CHIP_POLARIS10) {
3235 struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
3236 unsigned small_prim_filter_cntl =
3237 S_028830_SMALL_PRIM_FILTER_ENABLE(1) |
3238 /* line bug */
3239 S_028830_LINE_FILTER_DISABLE(sctx->family <= CHIP_POLARIS12);
3240
3241 /* The alternative of setting sample locations to 0 would
3242 * require a DB flush to avoid Z errors, see
3243 * https://bugs.freedesktop.org/show_bug.cgi?id=96908
3244 */
3245 if (has_msaa_sample_loc_bug &&
3246 sctx->framebuffer.nr_samples > 1 &&
3247 rs && !rs->multisample_enable)
3248 small_prim_filter_cntl &= C_028830_SMALL_PRIM_FILTER_ENABLE;
3249
3250 radeon_set_context_reg(cs, R_028830_PA_SU_SMALL_PRIM_FILTER_CNTL,
3251 small_prim_filter_cntl);
3252 }
3253 }
3254
3255 static bool si_out_of_order_rasterization(struct si_context *sctx)
3256 {
3257 struct si_state_blend *blend = sctx->queued.named.blend;
3258 struct si_state_dsa *dsa = sctx->queued.named.dsa;
3259
3260 if (!sctx->screen->has_out_of_order_rast)
3261 return false;
3262
3263 unsigned colormask = sctx->framebuffer.colorbuf_enabled_4bit;
3264
3265 if (blend) {
3266 colormask &= blend->cb_target_enabled_4bit;
3267 } else {
3268 colormask = 0;
3269 }
3270
3271 /* Conservative: No logic op. */
3272 if (colormask && blend->logicop_enable)
3273 return false;
3274
3275 struct si_dsa_order_invariance dsa_order_invariant = {
3276 .zs = true, .pass_set = true, .pass_last = false
3277 };
3278
3279 if (sctx->framebuffer.state.zsbuf) {
3280 struct r600_texture *zstex =
3281 (struct r600_texture*)sctx->framebuffer.state.zsbuf->texture;
3282 bool has_stencil = zstex->surface.has_stencil;
3283 dsa_order_invariant = dsa->order_invariance[has_stencil];
3284 if (!dsa_order_invariant.zs)
3285 return false;
3286
3287 /* The set of PS invocations is always order invariant,
3288 * except when early Z/S tests are requested. */
3289 if (sctx->ps_shader.cso &&
3290 sctx->ps_shader.cso->info.writes_memory &&
3291 sctx->ps_shader.cso->info.properties[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL] &&
3292 !dsa_order_invariant.pass_set)
3293 return false;
3294
3295 if (sctx->num_perfect_occlusion_queries != 0 &&
3296 !dsa_order_invariant.pass_set)
3297 return false;
3298 }
3299
3300 if (!colormask)
3301 return true;
3302
3303 unsigned blendmask = colormask & blend->blend_enable_4bit;
3304
3305 if (blendmask) {
3306 /* Only commutative blending. */
3307 if (blendmask & ~blend->commutative_4bit)
3308 return false;
3309
3310 if (!dsa_order_invariant.pass_set)
3311 return false;
3312 }
3313
3314 if (colormask & ~blendmask) {
3315 if (!dsa_order_invariant.pass_last)
3316 return false;
3317 }
3318
3319 return true;
3320 }
3321
3322 static void si_emit_msaa_config(struct si_context *sctx)
3323 {
3324 struct radeon_winsys_cs *cs = sctx->gfx_cs;
3325 unsigned num_tile_pipes = sctx->screen->info.num_tile_pipes;
3326 /* 33% faster rendering to linear color buffers */
3327 bool dst_is_linear = sctx->framebuffer.any_dst_linear;
3328 bool out_of_order_rast = si_out_of_order_rasterization(sctx);
3329 unsigned sc_mode_cntl_1 =
3330 S_028A4C_WALK_SIZE(dst_is_linear) |
3331 S_028A4C_WALK_FENCE_ENABLE(!dst_is_linear) |
3332 S_028A4C_WALK_FENCE_SIZE(num_tile_pipes == 2 ? 2 : 3) |
3333 S_028A4C_OUT_OF_ORDER_PRIMITIVE_ENABLE(out_of_order_rast) |
3334 S_028A4C_OUT_OF_ORDER_WATER_MARK(0x7) |
3335 /* always 1: */
3336 S_028A4C_WALK_ALIGN8_PRIM_FITS_ST(1) |
3337 S_028A4C_SUPERTILE_WALK_ORDER_ENABLE(1) |
3338 S_028A4C_TILE_WALK_ORDER_ENABLE(1) |
3339 S_028A4C_MULTI_SHADER_ENGINE_PRIM_DISCARD_ENABLE(1) |
3340 S_028A4C_FORCE_EOV_CNTDWN_ENABLE(1) |
3341 S_028A4C_FORCE_EOV_REZ_ENABLE(1);
3342 unsigned db_eqaa = S_028804_HIGH_QUALITY_INTERSECTIONS(1) |
3343 S_028804_INCOHERENT_EQAA_READS(1) |
3344 S_028804_INTERPOLATE_COMP_Z(1) |
3345 S_028804_STATIC_ANCHOR_ASSOCIATIONS(1);
3346 unsigned coverage_samples, color_samples;
3347
3348 /* S: Coverage samples (up to 16x):
3349 * - Scan conversion samples (PA_SC_AA_CONFIG.MSAA_NUM_SAMPLES)
3350 * - CB FMASK samples (CB_COLORi_ATTRIB.NUM_SAMPLES)
3351 *
3352 * Z: Z/S samples (up to 8x, must be <= coverage samples and >= color samples):
3353 * - Value seen by DB (DB_Z_INFO.NUM_SAMPLES)
3354 * - Value seen by CB, must be correct even if Z/S is unbound (DB_EQAA.MAX_ANCHOR_SAMPLES)
3355 * # Missing samples are derived from Z planes if Z is compressed (up to 16x quality), or
3356 * # from the closest defined sample if Z is uncompressed (same quality as the number of
3357 * # Z samples).
3358 *
3359 * F: Color samples (up to 8x, must be <= coverage samples):
3360 * - CB color samples (CB_COLORi_ATTRIB.NUM_FRAGMENTS)
3361 * - PS iter samples (DB_EQAA.PS_ITER_SAMPLES)
3362 *
3363 * Can be anything between coverage and color samples:
3364 * - SampleMaskIn samples (PA_SC_AA_CONFIG.MSAA_EXPOSED_SAMPLES)
3365 * - SampleMaskOut samples (DB_EQAA.MASK_EXPORT_NUM_SAMPLES)
3366 * - Alpha-to-coverage samples (DB_EQAA.ALPHA_TO_MASK_NUM_SAMPLES)
3367 * - Occlusion query samples (DB_COUNT_CONTROL.SAMPLE_RATE)
3368 * # All are currently set the same as coverage samples.
3369 *
3370 * If color samples < coverage samples, FMASK has a higher bpp to store an "unknown"
3371 * flag for undefined color samples. A shader-based resolve must handle unknowns
3372 * or mask them out with AND. Unknowns can also be guessed from neighbors via
3373 * an edge-detect shader-based resolve, which is required to make "color samples = 1"
3374 * useful. The CB resolve always drops unknowns.
3375 *
3376 * Sensible AA configurations:
3377 * EQAA 16s 8z 8f - might look the same as 16x MSAA if Z is compressed
3378 * EQAA 16s 8z 4f - might look the same as 16x MSAA if Z is compressed
3379 * EQAA 16s 4z 4f - might look the same as 16x MSAA if Z is compressed
3380 * EQAA 8s 8z 8f = 8x MSAA
3381 * EQAA 8s 8z 4f - might look the same as 8x MSAA
3382 * EQAA 8s 8z 2f - might look the same as 8x MSAA with low-density geometry
3383 * EQAA 8s 4z 4f - might look the same as 8x MSAA if Z is compressed
3384 * EQAA 8s 4z 2f - might look the same as 8x MSAA with low-density geometry if Z is compressed
3385 * EQAA 4s 4z 4f = 4x MSAA
3386 * EQAA 4s 4z 2f - might look the same as 4x MSAA with low-density geometry
3387 * EQAA 2s 2z 2f = 2x MSAA
3388 */
3389 if (sctx->framebuffer.nr_samples > 1) {
3390 coverage_samples = sctx->framebuffer.nr_samples;
3391 color_samples = sctx->framebuffer.nr_color_samples;
3392 } else if (sctx->smoothing_enabled) {
3393 coverage_samples = color_samples = SI_NUM_SMOOTH_AA_SAMPLES;
3394 } else {
3395 coverage_samples = color_samples = 1;
3396 }
3397
3398 /* Required by OpenGL line rasterization.
3399 *
3400 * TODO: We should also enable perpendicular endcaps for AA lines,
3401 * but that requires implementing line stippling in the pixel
3402 * shader. SC can only do line stippling with axis-aligned
3403 * endcaps.
3404 */
3405 unsigned sc_line_cntl = S_028BDC_DX10_DIAMOND_TEST_ENA(1);
3406
3407 if (coverage_samples > 1) {
3408 /* distance from the pixel center, indexed by log2(nr_samples) */
3409 static unsigned max_dist[] = {
3410 0, /* unused */
3411 4, /* 2x MSAA */
3412 6, /* 4x MSAA */
3413 7, /* 8x MSAA */
3414 8, /* 16x MSAA */
3415 };
3416 unsigned log_samples = util_logbase2(coverage_samples);
3417 unsigned ps_iter_samples = si_get_ps_iter_samples(sctx);
3418 ps_iter_samples = MIN2(ps_iter_samples, color_samples);
3419 unsigned log_ps_iter_samples =
3420 util_logbase2(util_next_power_of_two(ps_iter_samples));
3421
3422 radeon_set_context_reg_seq(cs, R_028BDC_PA_SC_LINE_CNTL, 2);
3423 radeon_emit(cs, sc_line_cntl |
3424 S_028BDC_EXPAND_LINE_WIDTH(1)); /* R_028BDC_PA_SC_LINE_CNTL */
3425 radeon_emit(cs, S_028BE0_MSAA_NUM_SAMPLES(log_samples) |
3426 S_028BE0_MAX_SAMPLE_DIST(max_dist[log_samples]) |
3427 S_028BE0_MSAA_EXPOSED_SAMPLES(log_samples)); /* R_028BE0_PA_SC_AA_CONFIG */
3428
3429 if (sctx->framebuffer.nr_samples > 1) {
3430 radeon_set_context_reg(cs, R_028804_DB_EQAA,
3431 db_eqaa |
3432 S_028804_MAX_ANCHOR_SAMPLES(log_samples) |
3433 S_028804_PS_ITER_SAMPLES(log_ps_iter_samples) |
3434 S_028804_MASK_EXPORT_NUM_SAMPLES(log_samples) |
3435 S_028804_ALPHA_TO_MASK_NUM_SAMPLES(log_samples));
3436 radeon_set_context_reg(cs, R_028A4C_PA_SC_MODE_CNTL_1,
3437 S_028A4C_PS_ITER_SAMPLE(ps_iter_samples > 1) |
3438 sc_mode_cntl_1);
3439 } else if (sctx->smoothing_enabled) {
3440 radeon_set_context_reg(cs, R_028804_DB_EQAA,
3441 db_eqaa |
3442 S_028804_OVERRASTERIZATION_AMOUNT(log_samples));
3443 radeon_set_context_reg(cs, R_028A4C_PA_SC_MODE_CNTL_1,
3444 sc_mode_cntl_1);
3445 }
3446 } else {
3447 radeon_set_context_reg_seq(cs, R_028BDC_PA_SC_LINE_CNTL, 2);
3448 radeon_emit(cs, sc_line_cntl); /* R_028BDC_PA_SC_LINE_CNTL */
3449 radeon_emit(cs, 0); /* R_028BE0_PA_SC_AA_CONFIG */
3450
3451 radeon_set_context_reg(cs, R_028804_DB_EQAA, db_eqaa);
3452 radeon_set_context_reg(cs, R_028A4C_PA_SC_MODE_CNTL_1,
3453 sc_mode_cntl_1);
3454 }
3455
3456 /* GFX9: Flush DFSM when the AA mode changes. */
3457 if (sctx->screen->dfsm_allowed) {
3458 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
3459 radeon_emit(cs, EVENT_TYPE(V_028A90_FLUSH_DFSM) | EVENT_INDEX(0));
3460 }
3461 }
3462
3463 void si_update_ps_iter_samples(struct si_context *sctx)
3464 {
3465 if (sctx->framebuffer.nr_samples > 1)
3466 si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
3467 if (sctx->screen->dpbb_allowed)
3468 si_mark_atom_dirty(sctx, &sctx->atoms.s.dpbb_state);
3469 }
3470
3471 static void si_set_min_samples(struct pipe_context *ctx, unsigned min_samples)
3472 {
3473 struct si_context *sctx = (struct si_context *)ctx;
3474
3475 if (sctx->ps_iter_samples == min_samples)
3476 return;
3477
3478 sctx->ps_iter_samples = min_samples;
3479 sctx->do_update_shaders = true;
3480
3481 si_update_ps_iter_samples(sctx);
3482 }
3483
3484 /*
3485 * Samplers
3486 */
3487
3488 /**
3489 * Build the sampler view descriptor for a buffer texture.
3490 * @param state 256-bit descriptor; only the high 128 bits are filled in
3491 */
3492 void
3493 si_make_buffer_descriptor(struct si_screen *screen, struct r600_resource *buf,
3494 enum pipe_format format,
3495 unsigned offset, unsigned size,
3496 uint32_t *state)
3497 {
3498 const struct util_format_description *desc;
3499 int first_non_void;
3500 unsigned stride;
3501 unsigned num_records;
3502 unsigned num_format, data_format;
3503
3504 desc = util_format_description(format);
3505 first_non_void = util_format_get_first_non_void_channel(format);
3506 stride = desc->block.bits / 8;
3507 num_format = si_translate_buffer_numformat(&screen->b, desc, first_non_void);
3508 data_format = si_translate_buffer_dataformat(&screen->b, desc, first_non_void);
3509
3510 num_records = size / stride;
3511 num_records = MIN2(num_records, (buf->b.b.width0 - offset) / stride);
3512
3513 /* The NUM_RECORDS field has a different meaning depending on the chip,
3514 * instruction type, STRIDE, and SWIZZLE_ENABLE.
3515 *
3516 * SI-CIK:
3517 * - If STRIDE == 0, it's in byte units.
3518 * - If STRIDE != 0, it's in units of STRIDE, used with inst.IDXEN.
3519 *
3520 * VI:
3521 * - For SMEM and STRIDE == 0, it's in byte units.
3522 * - For SMEM and STRIDE != 0, it's in units of STRIDE.
3523 * - For VMEM and STRIDE == 0 or SWIZZLE_ENABLE == 0, it's in byte units.
3524 * - For VMEM and STRIDE != 0 and SWIZZLE_ENABLE == 1, it's in units of STRIDE.
3525 * NOTE: There is incompatibility between VMEM and SMEM opcodes due to SWIZZLE_-
3526 * ENABLE. The workaround is to set STRIDE = 0 if SWIZZLE_ENABLE == 0 when
3527 * using SMEM. This can be done in the shader by clearing STRIDE with s_and.
3528 * That way the same descriptor can be used by both SMEM and VMEM.
3529 *
3530 * GFX9:
3531 * - For SMEM and STRIDE == 0, it's in byte units.
3532 * - For SMEM and STRIDE != 0, it's in units of STRIDE.
3533 * - For VMEM and inst.IDXEN == 0 or STRIDE == 0, it's in byte units.
3534 * - For VMEM and inst.IDXEN == 1 and STRIDE != 0, it's in units of STRIDE.
3535 */
3536 if (screen->info.chip_class >= GFX9)
3537 /* When vindex == 0, LLVM sets IDXEN = 0, thus changing units
3538 * from STRIDE to bytes. This works around it by setting
3539 * NUM_RECORDS to at least the size of one element, so that
3540 * the first element is readable when IDXEN == 0.
3541 *
3542 * TODO: Fix this in LLVM, but do we need a new intrinsic where
3543 * IDXEN is enforced?
3544 */
3545 num_records = num_records ? MAX2(num_records, stride) : 0;
3546 else if (screen->info.chip_class == VI)
3547 num_records *= stride;
3548
3549 state[4] = 0;
3550 state[5] = S_008F04_STRIDE(stride);
3551 state[6] = num_records;
3552 state[7] = S_008F0C_DST_SEL_X(si_map_swizzle(desc->swizzle[0])) |
3553 S_008F0C_DST_SEL_Y(si_map_swizzle(desc->swizzle[1])) |
3554 S_008F0C_DST_SEL_Z(si_map_swizzle(desc->swizzle[2])) |
3555 S_008F0C_DST_SEL_W(si_map_swizzle(desc->swizzle[3])) |
3556 S_008F0C_NUM_FORMAT(num_format) |
3557 S_008F0C_DATA_FORMAT(data_format);
3558 }
3559
3560 static unsigned gfx9_border_color_swizzle(const unsigned char swizzle[4])
3561 {
3562 unsigned bc_swizzle = V_008F20_BC_SWIZZLE_XYZW;
3563
3564 if (swizzle[3] == PIPE_SWIZZLE_X) {
3565 /* For the pre-defined border color values (white, opaque
3566 * black, transparent black), the only thing that matters is
3567 * that the alpha channel winds up in the correct place
3568 * (because the RGB channels are all the same) so either of
3569 * these enumerations will work.
3570 */
3571 if (swizzle[2] == PIPE_SWIZZLE_Y)
3572 bc_swizzle = V_008F20_BC_SWIZZLE_WZYX;
3573 else
3574 bc_swizzle = V_008F20_BC_SWIZZLE_WXYZ;
3575 } else if (swizzle[0] == PIPE_SWIZZLE_X) {
3576 if (swizzle[1] == PIPE_SWIZZLE_Y)
3577 bc_swizzle = V_008F20_BC_SWIZZLE_XYZW;
3578 else
3579 bc_swizzle = V_008F20_BC_SWIZZLE_XWYZ;
3580 } else if (swizzle[1] == PIPE_SWIZZLE_X) {
3581 bc_swizzle = V_008F20_BC_SWIZZLE_YXWZ;
3582 } else if (swizzle[2] == PIPE_SWIZZLE_X) {
3583 bc_swizzle = V_008F20_BC_SWIZZLE_ZYXW;
3584 }
3585
3586 return bc_swizzle;
3587 }
3588
3589 /**
3590 * Build the sampler view descriptor for a texture.
3591 */
3592 void
3593 si_make_texture_descriptor(struct si_screen *screen,
3594 struct r600_texture *tex,
3595 bool sampler,
3596 enum pipe_texture_target target,
3597 enum pipe_format pipe_format,
3598 const unsigned char state_swizzle[4],
3599 unsigned first_level, unsigned last_level,
3600 unsigned first_layer, unsigned last_layer,
3601 unsigned width, unsigned height, unsigned depth,
3602 uint32_t *state,
3603 uint32_t *fmask_state)
3604 {
3605 struct pipe_resource *res = &tex->buffer.b.b;
3606 const struct util_format_description *desc;
3607 unsigned char swizzle[4];
3608 int first_non_void;
3609 unsigned num_format, data_format, type, num_samples;
3610 uint64_t va;
3611
3612 desc = util_format_description(pipe_format);
3613
3614 num_samples = desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS ?
3615 MAX2(1, res->nr_samples) : tex->num_color_samples;
3616
3617 if (desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS) {
3618 const unsigned char swizzle_xxxx[4] = {0, 0, 0, 0};
3619 const unsigned char swizzle_yyyy[4] = {1, 1, 1, 1};
3620 const unsigned char swizzle_wwww[4] = {3, 3, 3, 3};
3621
3622 switch (pipe_format) {
3623 case PIPE_FORMAT_S8_UINT_Z24_UNORM:
3624 case PIPE_FORMAT_X32_S8X24_UINT:
3625 case PIPE_FORMAT_X8Z24_UNORM:
3626 util_format_compose_swizzles(swizzle_yyyy, state_swizzle, swizzle);
3627 break;
3628 case PIPE_FORMAT_X24S8_UINT:
3629 /*
3630 * X24S8 is implemented as an 8_8_8_8 data format, to
3631 * fix texture gathers. This affects at least
3632 * GL45-CTS.texture_cube_map_array.sampling on VI.
3633 */
3634 if (screen->info.chip_class <= VI)
3635 util_format_compose_swizzles(swizzle_wwww, state_swizzle, swizzle);
3636 else
3637 util_format_compose_swizzles(swizzle_yyyy, state_swizzle, swizzle);
3638 break;
3639 default:
3640 util_format_compose_swizzles(swizzle_xxxx, state_swizzle, swizzle);
3641 }
3642 } else {
3643 util_format_compose_swizzles(desc->swizzle, state_swizzle, swizzle);
3644 }
3645
3646 first_non_void = util_format_get_first_non_void_channel(pipe_format);
3647
3648 switch (pipe_format) {
3649 case PIPE_FORMAT_S8_UINT_Z24_UNORM:
3650 num_format = V_008F14_IMG_NUM_FORMAT_UNORM;
3651 break;
3652 default:
3653 if (first_non_void < 0) {
3654 if (util_format_is_compressed(pipe_format)) {
3655 switch (pipe_format) {
3656 case PIPE_FORMAT_DXT1_SRGB:
3657 case PIPE_FORMAT_DXT1_SRGBA:
3658 case PIPE_FORMAT_DXT3_SRGBA:
3659 case PIPE_FORMAT_DXT5_SRGBA:
3660 case PIPE_FORMAT_BPTC_SRGBA:
3661 case PIPE_FORMAT_ETC2_SRGB8:
3662 case PIPE_FORMAT_ETC2_SRGB8A1:
3663 case PIPE_FORMAT_ETC2_SRGBA8:
3664 num_format = V_008F14_IMG_NUM_FORMAT_SRGB;
3665 break;
3666 case PIPE_FORMAT_RGTC1_SNORM:
3667 case PIPE_FORMAT_LATC1_SNORM:
3668 case PIPE_FORMAT_RGTC2_SNORM:
3669 case PIPE_FORMAT_LATC2_SNORM:
3670 case PIPE_FORMAT_ETC2_R11_SNORM:
3671 case PIPE_FORMAT_ETC2_RG11_SNORM:
3672 /* implies float, so use SNORM/UNORM to determine
3673 whether data is signed or not */
3674 case PIPE_FORMAT_BPTC_RGB_FLOAT:
3675 num_format = V_008F14_IMG_NUM_FORMAT_SNORM;
3676 break;
3677 default:
3678 num_format = V_008F14_IMG_NUM_FORMAT_UNORM;
3679 break;
3680 }
3681 } else if (desc->layout == UTIL_FORMAT_LAYOUT_SUBSAMPLED) {
3682 num_format = V_008F14_IMG_NUM_FORMAT_UNORM;
3683 } else {
3684 num_format = V_008F14_IMG_NUM_FORMAT_FLOAT;
3685 }
3686 } else if (desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
3687 num_format = V_008F14_IMG_NUM_FORMAT_SRGB;
3688 } else {
3689 num_format = V_008F14_IMG_NUM_FORMAT_UNORM;
3690
3691 switch (desc->channel[first_non_void].type) {
3692 case UTIL_FORMAT_TYPE_FLOAT:
3693 num_format = V_008F14_IMG_NUM_FORMAT_FLOAT;
3694 break;
3695 case UTIL_FORMAT_TYPE_SIGNED:
3696 if (desc->channel[first_non_void].normalized)
3697 num_format = V_008F14_IMG_NUM_FORMAT_SNORM;
3698 else if (desc->channel[first_non_void].pure_integer)
3699 num_format = V_008F14_IMG_NUM_FORMAT_SINT;
3700 else
3701 num_format = V_008F14_IMG_NUM_FORMAT_SSCALED;
3702 break;
3703 case UTIL_FORMAT_TYPE_UNSIGNED:
3704 if (desc->channel[first_non_void].normalized)
3705 num_format = V_008F14_IMG_NUM_FORMAT_UNORM;
3706 else if (desc->channel[first_non_void].pure_integer)
3707 num_format = V_008F14_IMG_NUM_FORMAT_UINT;
3708 else
3709 num_format = V_008F14_IMG_NUM_FORMAT_USCALED;
3710 }
3711 }
3712 }
3713
3714 data_format = si_translate_texformat(&screen->b, pipe_format, desc, first_non_void);
3715 if (data_format == ~0) {
3716 data_format = 0;
3717 }
3718
3719 /* S8 with Z32 HTILE needs a special format. */
3720 if (screen->info.chip_class >= GFX9 &&
3721 pipe_format == PIPE_FORMAT_S8_UINT &&
3722 tex->tc_compatible_htile)
3723 data_format = V_008F14_IMG_DATA_FORMAT_S8_32;
3724
3725 if (!sampler &&
3726 (res->target == PIPE_TEXTURE_CUBE ||
3727 res->target == PIPE_TEXTURE_CUBE_ARRAY ||
3728 (screen->info.chip_class <= VI &&
3729 res->target == PIPE_TEXTURE_3D))) {
3730 /* For the purpose of shader images, treat cube maps and 3D
3731 * textures as 2D arrays. For 3D textures, the address
3732 * calculations for mipmaps are different, so we rely on the
3733 * caller to effectively disable mipmaps.
3734 */
3735 type = V_008F1C_SQ_RSRC_IMG_2D_ARRAY;
3736
3737 assert(res->target != PIPE_TEXTURE_3D || (first_level == 0 && last_level == 0));
3738 } else {
3739 type = si_tex_dim(screen, tex, target, num_samples);
3740 }
3741
3742 if (type == V_008F1C_SQ_RSRC_IMG_1D_ARRAY) {
3743 height = 1;
3744 depth = res->array_size;
3745 } else if (type == V_008F1C_SQ_RSRC_IMG_2D_ARRAY ||
3746 type == V_008F1C_SQ_RSRC_IMG_2D_MSAA_ARRAY) {
3747 if (sampler || res->target != PIPE_TEXTURE_3D)
3748 depth = res->array_size;
3749 } else if (type == V_008F1C_SQ_RSRC_IMG_CUBE)
3750 depth = res->array_size / 6;
3751
3752 state[0] = 0;
3753 state[1] = (S_008F14_DATA_FORMAT_GFX6(data_format) |
3754 S_008F14_NUM_FORMAT_GFX6(num_format));
3755 state[2] = (S_008F18_WIDTH(width - 1) |
3756 S_008F18_HEIGHT(height - 1) |
3757 S_008F18_PERF_MOD(4));
3758 state[3] = (S_008F1C_DST_SEL_X(si_map_swizzle(swizzle[0])) |
3759 S_008F1C_DST_SEL_Y(si_map_swizzle(swizzle[1])) |
3760 S_008F1C_DST_SEL_Z(si_map_swizzle(swizzle[2])) |
3761 S_008F1C_DST_SEL_W(si_map_swizzle(swizzle[3])) |
3762 S_008F1C_BASE_LEVEL(num_samples > 1 ? 0 : first_level) |
3763 S_008F1C_LAST_LEVEL(num_samples > 1 ?
3764 util_logbase2(num_samples) :
3765 last_level) |
3766 S_008F1C_TYPE(type));
3767 state[4] = 0;
3768 state[5] = S_008F24_BASE_ARRAY(first_layer);
3769 state[6] = 0;
3770 state[7] = 0;
3771
3772 if (screen->info.chip_class >= GFX9) {
3773 unsigned bc_swizzle = gfx9_border_color_swizzle(desc->swizzle);
3774
3775 /* Depth is the the last accessible layer on Gfx9.
3776 * The hw doesn't need to know the total number of layers.
3777 */
3778 if (type == V_008F1C_SQ_RSRC_IMG_3D)
3779 state[4] |= S_008F20_DEPTH(depth - 1);
3780 else
3781 state[4] |= S_008F20_DEPTH(last_layer);
3782
3783 state[4] |= S_008F20_BC_SWIZZLE(bc_swizzle);
3784 state[5] |= S_008F24_MAX_MIP(num_samples > 1 ?
3785 util_logbase2(num_samples) :
3786 tex->buffer.b.b.last_level);
3787 } else {
3788 state[3] |= S_008F1C_POW2_PAD(res->last_level > 0);
3789 state[4] |= S_008F20_DEPTH(depth - 1);
3790 state[5] |= S_008F24_LAST_ARRAY(last_layer);
3791 }
3792
3793 if (tex->dcc_offset) {
3794 state[6] = S_008F28_ALPHA_IS_ON_MSB(vi_alpha_is_on_msb(pipe_format));
3795 } else {
3796 /* The last dword is unused by hw. The shader uses it to clear
3797 * bits in the first dword of sampler state.
3798 */
3799 if (screen->info.chip_class <= CIK && res->nr_samples <= 1) {
3800 if (first_level == last_level)
3801 state[7] = C_008F30_MAX_ANISO_RATIO;
3802 else
3803 state[7] = 0xffffffff;
3804 }
3805 }
3806
3807 /* Initialize the sampler view for FMASK. */
3808 if (tex->surface.fmask_size) {
3809 uint32_t data_format, num_format;
3810
3811 va = tex->buffer.gpu_address + tex->fmask_offset;
3812
3813 #define FMASK(s,f) (((unsigned)(s) * 16) + (f))
3814 if (screen->info.chip_class >= GFX9) {
3815 data_format = V_008F14_IMG_DATA_FORMAT_FMASK;
3816 switch (FMASK(res->nr_samples, tex->num_color_samples)) {
3817 case FMASK(2,1):
3818 num_format = V_008F14_IMG_FMASK_8_2_1;
3819 break;
3820 case FMASK(2,2):
3821 num_format = V_008F14_IMG_FMASK_8_2_2;
3822 break;
3823 case FMASK(4,1):
3824 num_format = V_008F14_IMG_FMASK_8_4_1;
3825 break;
3826 case FMASK(4,2):
3827 num_format = V_008F14_IMG_FMASK_8_4_2;
3828 break;
3829 case FMASK(4,4):
3830 num_format = V_008F14_IMG_FMASK_8_4_4;
3831 break;
3832 case FMASK(8,1):
3833 num_format = V_008F14_IMG_FMASK_8_8_1;
3834 break;
3835 case FMASK(8,2):
3836 num_format = V_008F14_IMG_FMASK_16_8_2;
3837 break;
3838 case FMASK(8,4):
3839 num_format = V_008F14_IMG_FMASK_32_8_4;
3840 break;
3841 case FMASK(8,8):
3842 num_format = V_008F14_IMG_FMASK_32_8_8;
3843 break;
3844 case FMASK(16,1):
3845 num_format = V_008F14_IMG_FMASK_16_16_1;
3846 break;
3847 case FMASK(16,2):
3848 num_format = V_008F14_IMG_FMASK_32_16_2;
3849 break;
3850 case FMASK(16,4):
3851 num_format = V_008F14_IMG_FMASK_64_16_4;
3852 break;
3853 case FMASK(16,8):
3854 num_format = V_008F14_IMG_FMASK_64_16_8;
3855 break;
3856 default:
3857 unreachable("invalid nr_samples");
3858 }
3859 } else {
3860 switch (FMASK(res->nr_samples, tex->num_color_samples)) {
3861 case FMASK(2,1):
3862 data_format = V_008F14_IMG_DATA_FORMAT_FMASK8_S2_F1;
3863 break;
3864 case FMASK(2,2):
3865 data_format = V_008F14_IMG_DATA_FORMAT_FMASK8_S2_F2;
3866 break;
3867 case FMASK(4,1):
3868 data_format = V_008F14_IMG_DATA_FORMAT_FMASK8_S4_F1;
3869 break;
3870 case FMASK(4,2):
3871 data_format = V_008F14_IMG_DATA_FORMAT_FMASK8_S4_F2;
3872 break;
3873 case FMASK(4,4):
3874 data_format = V_008F14_IMG_DATA_FORMAT_FMASK8_S4_F4;
3875 break;
3876 case FMASK(8,1):
3877 data_format = V_008F14_IMG_DATA_FORMAT_FMASK8_S8_F1;
3878 break;
3879 case FMASK(8,2):
3880 data_format = V_008F14_IMG_DATA_FORMAT_FMASK16_S8_F2;
3881 break;
3882 case FMASK(8,4):
3883 data_format = V_008F14_IMG_DATA_FORMAT_FMASK32_S8_F4;
3884 break;
3885 case FMASK(8,8):
3886 data_format = V_008F14_IMG_DATA_FORMAT_FMASK32_S8_F8;
3887 break;
3888 case FMASK(16,1):
3889 data_format = V_008F14_IMG_DATA_FORMAT_FMASK16_S16_F1;
3890 break;
3891 case FMASK(16,2):
3892 data_format = V_008F14_IMG_DATA_FORMAT_FMASK32_S16_F2;
3893 break;
3894 case FMASK(16,4):
3895 data_format = V_008F14_IMG_DATA_FORMAT_FMASK64_S16_F4;
3896 break;
3897 case FMASK(16,8):
3898 data_format = V_008F14_IMG_DATA_FORMAT_FMASK64_S16_F8;
3899 break;
3900 default:
3901 unreachable("invalid nr_samples");
3902 }
3903 num_format = V_008F14_IMG_NUM_FORMAT_UINT;
3904 }
3905 #undef FMASK
3906
3907 fmask_state[0] = (va >> 8) | tex->surface.fmask_tile_swizzle;
3908 fmask_state[1] = S_008F14_BASE_ADDRESS_HI(va >> 40) |
3909 S_008F14_DATA_FORMAT_GFX6(data_format) |
3910 S_008F14_NUM_FORMAT_GFX6(num_format);
3911 fmask_state[2] = S_008F18_WIDTH(width - 1) |
3912 S_008F18_HEIGHT(height - 1);
3913 fmask_state[3] = S_008F1C_DST_SEL_X(V_008F1C_SQ_SEL_X) |
3914 S_008F1C_DST_SEL_Y(V_008F1C_SQ_SEL_X) |
3915 S_008F1C_DST_SEL_Z(V_008F1C_SQ_SEL_X) |
3916 S_008F1C_DST_SEL_W(V_008F1C_SQ_SEL_X) |
3917 S_008F1C_TYPE(si_tex_dim(screen, tex, target, 0));
3918 fmask_state[4] = 0;
3919 fmask_state[5] = S_008F24_BASE_ARRAY(first_layer);
3920 fmask_state[6] = 0;
3921 fmask_state[7] = 0;
3922
3923 if (screen->info.chip_class >= GFX9) {
3924 fmask_state[3] |= S_008F1C_SW_MODE(tex->surface.u.gfx9.fmask.swizzle_mode);
3925 fmask_state[4] |= S_008F20_DEPTH(last_layer) |
3926 S_008F20_PITCH_GFX9(tex->surface.u.gfx9.fmask.epitch);
3927 fmask_state[5] |= S_008F24_META_PIPE_ALIGNED(tex->surface.u.gfx9.cmask.pipe_aligned) |
3928 S_008F24_META_RB_ALIGNED(tex->surface.u.gfx9.cmask.rb_aligned);
3929 } else {
3930 fmask_state[3] |= S_008F1C_TILING_INDEX(tex->surface.u.legacy.fmask.tiling_index);
3931 fmask_state[4] |= S_008F20_DEPTH(depth - 1) |
3932 S_008F20_PITCH_GFX6(tex->surface.u.legacy.fmask.pitch_in_pixels - 1);
3933 fmask_state[5] |= S_008F24_LAST_ARRAY(last_layer);
3934 }
3935 }
3936 }
3937
3938 /**
3939 * Create a sampler view.
3940 *
3941 * @param ctx context
3942 * @param texture texture
3943 * @param state sampler view template
3944 * @param width0 width0 override (for compressed textures as int)
3945 * @param height0 height0 override (for compressed textures as int)
3946 * @param force_level set the base address to the level (for compressed textures)
3947 */
3948 struct pipe_sampler_view *
3949 si_create_sampler_view_custom(struct pipe_context *ctx,
3950 struct pipe_resource *texture,
3951 const struct pipe_sampler_view *state,
3952 unsigned width0, unsigned height0,
3953 unsigned force_level)
3954 {
3955 struct si_context *sctx = (struct si_context*)ctx;
3956 struct si_sampler_view *view = CALLOC_STRUCT(si_sampler_view);
3957 struct r600_texture *tmp = (struct r600_texture*)texture;
3958 unsigned base_level, first_level, last_level;
3959 unsigned char state_swizzle[4];
3960 unsigned height, depth, width;
3961 unsigned last_layer = state->u.tex.last_layer;
3962 enum pipe_format pipe_format;
3963 const struct legacy_surf_level *surflevel;
3964
3965 if (!view)
3966 return NULL;
3967
3968 /* initialize base object */
3969 view->base = *state;
3970 view->base.texture = NULL;
3971 view->base.reference.count = 1;
3972 view->base.context = ctx;
3973
3974 assert(texture);
3975 pipe_resource_reference(&view->base.texture, texture);
3976
3977 if (state->format == PIPE_FORMAT_X24S8_UINT ||
3978 state->format == PIPE_FORMAT_S8X24_UINT ||
3979 state->format == PIPE_FORMAT_X32_S8X24_UINT ||
3980 state->format == PIPE_FORMAT_S8_UINT)
3981 view->is_stencil_sampler = true;
3982
3983 /* Buffer resource. */
3984 if (texture->target == PIPE_BUFFER) {
3985 si_make_buffer_descriptor(sctx->screen,
3986 r600_resource(texture),
3987 state->format,
3988 state->u.buf.offset,
3989 state->u.buf.size,
3990 view->state);
3991 return &view->base;
3992 }
3993
3994 state_swizzle[0] = state->swizzle_r;
3995 state_swizzle[1] = state->swizzle_g;
3996 state_swizzle[2] = state->swizzle_b;
3997 state_swizzle[3] = state->swizzle_a;
3998
3999 base_level = 0;
4000 first_level = state->u.tex.first_level;
4001 last_level = state->u.tex.last_level;
4002 width = width0;
4003 height = height0;
4004 depth = texture->depth0;
4005
4006 if (sctx->chip_class <= VI && force_level) {
4007 assert(force_level == first_level &&
4008 force_level == last_level);
4009 base_level = force_level;
4010 first_level = 0;
4011 last_level = 0;
4012 width = u_minify(width, force_level);
4013 height = u_minify(height, force_level);
4014 depth = u_minify(depth, force_level);
4015 }
4016
4017 /* This is not needed if state trackers set last_layer correctly. */
4018 if (state->target == PIPE_TEXTURE_1D ||
4019 state->target == PIPE_TEXTURE_2D ||
4020 state->target == PIPE_TEXTURE_RECT ||
4021 state->target == PIPE_TEXTURE_CUBE)
4022 last_layer = state->u.tex.first_layer;
4023
4024 /* Texturing with separate depth and stencil. */
4025 pipe_format = state->format;
4026
4027 /* Depth/stencil texturing sometimes needs separate texture. */
4028 if (tmp->is_depth && !si_can_sample_zs(tmp, view->is_stencil_sampler)) {
4029 if (!tmp->flushed_depth_texture &&
4030 !si_init_flushed_depth_texture(ctx, texture, NULL)) {
4031 pipe_resource_reference(&view->base.texture, NULL);
4032 FREE(view);
4033 return NULL;
4034 }
4035
4036 assert(tmp->flushed_depth_texture);
4037
4038 /* Override format for the case where the flushed texture
4039 * contains only Z or only S.
4040 */
4041 if (tmp->flushed_depth_texture->buffer.b.b.format != tmp->buffer.b.b.format)
4042 pipe_format = tmp->flushed_depth_texture->buffer.b.b.format;
4043
4044 tmp = tmp->flushed_depth_texture;
4045 }
4046
4047 surflevel = tmp->surface.u.legacy.level;
4048
4049 if (tmp->db_compatible) {
4050 if (!view->is_stencil_sampler)
4051 pipe_format = tmp->db_render_format;
4052
4053 switch (pipe_format) {
4054 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
4055 pipe_format = PIPE_FORMAT_Z32_FLOAT;
4056 break;
4057 case PIPE_FORMAT_X8Z24_UNORM:
4058 case PIPE_FORMAT_S8_UINT_Z24_UNORM:
4059 /* Z24 is always stored like this for DB
4060 * compatibility.
4061 */
4062 pipe_format = PIPE_FORMAT_Z24X8_UNORM;
4063 break;
4064 case PIPE_FORMAT_X24S8_UINT:
4065 case PIPE_FORMAT_S8X24_UINT:
4066 case PIPE_FORMAT_X32_S8X24_UINT:
4067 pipe_format = PIPE_FORMAT_S8_UINT;
4068 surflevel = tmp->surface.u.legacy.stencil_level;
4069 break;
4070 default:;
4071 }
4072 }
4073
4074 view->dcc_incompatible =
4075 vi_dcc_formats_are_incompatible(texture,
4076 state->u.tex.first_level,
4077 state->format);
4078
4079 si_make_texture_descriptor(sctx->screen, tmp, true,
4080 state->target, pipe_format, state_swizzle,
4081 first_level, last_level,
4082 state->u.tex.first_layer, last_layer,
4083 width, height, depth,
4084 view->state, view->fmask_state);
4085
4086 unsigned num_format = G_008F14_NUM_FORMAT_GFX6(view->state[1]);
4087 view->is_integer =
4088 num_format == V_008F14_IMG_NUM_FORMAT_USCALED ||
4089 num_format == V_008F14_IMG_NUM_FORMAT_SSCALED ||
4090 num_format == V_008F14_IMG_NUM_FORMAT_UINT ||
4091 num_format == V_008F14_IMG_NUM_FORMAT_SINT;
4092 view->base_level_info = &surflevel[base_level];
4093 view->base_level = base_level;
4094 view->block_width = util_format_get_blockwidth(pipe_format);
4095 return &view->base;
4096 }
4097
4098 static struct pipe_sampler_view *
4099 si_create_sampler_view(struct pipe_context *ctx,
4100 struct pipe_resource *texture,
4101 const struct pipe_sampler_view *state)
4102 {
4103 return si_create_sampler_view_custom(ctx, texture, state,
4104 texture ? texture->width0 : 0,
4105 texture ? texture->height0 : 0, 0);
4106 }
4107
4108 static void si_sampler_view_destroy(struct pipe_context *ctx,
4109 struct pipe_sampler_view *state)
4110 {
4111 struct si_sampler_view *view = (struct si_sampler_view *)state;
4112
4113 pipe_resource_reference(&state->texture, NULL);
4114 FREE(view);
4115 }
4116
4117 static bool wrap_mode_uses_border_color(unsigned wrap, bool linear_filter)
4118 {
4119 return wrap == PIPE_TEX_WRAP_CLAMP_TO_BORDER ||
4120 wrap == PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER ||
4121 (linear_filter &&
4122 (wrap == PIPE_TEX_WRAP_CLAMP ||
4123 wrap == PIPE_TEX_WRAP_MIRROR_CLAMP));
4124 }
4125
4126 static uint32_t si_translate_border_color(struct si_context *sctx,
4127 const struct pipe_sampler_state *state,
4128 const union pipe_color_union *color,
4129 bool is_integer)
4130 {
4131 bool linear_filter = state->min_img_filter != PIPE_TEX_FILTER_NEAREST ||
4132 state->mag_img_filter != PIPE_TEX_FILTER_NEAREST;
4133
4134 if (!wrap_mode_uses_border_color(state->wrap_s, linear_filter) &&
4135 !wrap_mode_uses_border_color(state->wrap_t, linear_filter) &&
4136 !wrap_mode_uses_border_color(state->wrap_r, linear_filter))
4137 return S_008F3C_BORDER_COLOR_TYPE(V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK);
4138
4139 #define simple_border_types(elt) \
4140 do { \
4141 if (color->elt[0] == 0 && color->elt[1] == 0 && \
4142 color->elt[2] == 0 && color->elt[3] == 0) \
4143 return S_008F3C_BORDER_COLOR_TYPE(V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK); \
4144 if (color->elt[0] == 0 && color->elt[1] == 0 && \
4145 color->elt[2] == 0 && color->elt[3] == 1) \
4146 return S_008F3C_BORDER_COLOR_TYPE(V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_BLACK); \
4147 if (color->elt[0] == 1 && color->elt[1] == 1 && \
4148 color->elt[2] == 1 && color->elt[3] == 1) \
4149 return S_008F3C_BORDER_COLOR_TYPE(V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_WHITE); \
4150 } while (false)
4151
4152 if (is_integer)
4153 simple_border_types(ui);
4154 else
4155 simple_border_types(f);
4156
4157 #undef simple_border_types
4158
4159 int i;
4160
4161 /* Check if the border has been uploaded already. */
4162 for (i = 0; i < sctx->border_color_count; i++)
4163 if (memcmp(&sctx->border_color_table[i], color,
4164 sizeof(*color)) == 0)
4165 break;
4166
4167 if (i >= SI_MAX_BORDER_COLORS) {
4168 /* Getting 4096 unique border colors is very unlikely. */
4169 fprintf(stderr, "radeonsi: The border color table is full. "
4170 "Any new border colors will be just black. "
4171 "Please file a bug.\n");
4172 return S_008F3C_BORDER_COLOR_TYPE(V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK);
4173 }
4174
4175 if (i == sctx->border_color_count) {
4176 /* Upload a new border color. */
4177 memcpy(&sctx->border_color_table[i], color,
4178 sizeof(*color));
4179 util_memcpy_cpu_to_le32(&sctx->border_color_map[i],
4180 color, sizeof(*color));
4181 sctx->border_color_count++;
4182 }
4183
4184 return S_008F3C_BORDER_COLOR_PTR(i) |
4185 S_008F3C_BORDER_COLOR_TYPE(V_008F3C_SQ_TEX_BORDER_COLOR_REGISTER);
4186 }
4187
4188 static inline int S_FIXED(float value, unsigned frac_bits)
4189 {
4190 return value * (1 << frac_bits);
4191 }
4192
4193 static inline unsigned si_tex_filter(unsigned filter, unsigned max_aniso)
4194 {
4195 if (filter == PIPE_TEX_FILTER_LINEAR)
4196 return max_aniso > 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_BILINEAR
4197 : V_008F38_SQ_TEX_XY_FILTER_BILINEAR;
4198 else
4199 return max_aniso > 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_POINT
4200 : V_008F38_SQ_TEX_XY_FILTER_POINT;
4201 }
4202
4203 static inline unsigned si_tex_aniso_filter(unsigned filter)
4204 {
4205 if (filter < 2)
4206 return 0;
4207 if (filter < 4)
4208 return 1;
4209 if (filter < 8)
4210 return 2;
4211 if (filter < 16)
4212 return 3;
4213 return 4;
4214 }
4215
4216 static void *si_create_sampler_state(struct pipe_context *ctx,
4217 const struct pipe_sampler_state *state)
4218 {
4219 struct si_context *sctx = (struct si_context *)ctx;
4220 struct si_screen *sscreen = sctx->screen;
4221 struct si_sampler_state *rstate = CALLOC_STRUCT(si_sampler_state);
4222 unsigned max_aniso = sscreen->force_aniso >= 0 ? sscreen->force_aniso
4223 : state->max_anisotropy;
4224 unsigned max_aniso_ratio = si_tex_aniso_filter(max_aniso);
4225 union pipe_color_union clamped_border_color;
4226
4227 if (!rstate) {
4228 return NULL;
4229 }
4230
4231 #ifdef DEBUG
4232 rstate->magic = SI_SAMPLER_STATE_MAGIC;
4233 #endif
4234 rstate->val[0] = (S_008F30_CLAMP_X(si_tex_wrap(state->wrap_s)) |
4235 S_008F30_CLAMP_Y(si_tex_wrap(state->wrap_t)) |
4236 S_008F30_CLAMP_Z(si_tex_wrap(state->wrap_r)) |
4237 S_008F30_MAX_ANISO_RATIO(max_aniso_ratio) |
4238 S_008F30_DEPTH_COMPARE_FUNC(si_tex_compare(state->compare_func)) |
4239 S_008F30_FORCE_UNNORMALIZED(!state->normalized_coords) |
4240 S_008F30_ANISO_THRESHOLD(max_aniso_ratio >> 1) |
4241 S_008F30_ANISO_BIAS(max_aniso_ratio) |
4242 S_008F30_DISABLE_CUBE_WRAP(!state->seamless_cube_map) |
4243 S_008F30_COMPAT_MODE(sctx->chip_class >= VI));
4244 rstate->val[1] = (S_008F34_MIN_LOD(S_FIXED(CLAMP(state->min_lod, 0, 15), 8)) |
4245 S_008F34_MAX_LOD(S_FIXED(CLAMP(state->max_lod, 0, 15), 8)) |
4246 S_008F34_PERF_MIP(max_aniso_ratio ? max_aniso_ratio + 6 : 0));
4247 rstate->val[2] = (S_008F38_LOD_BIAS(S_FIXED(CLAMP(state->lod_bias, -16, 16), 8)) |
4248 S_008F38_XY_MAG_FILTER(si_tex_filter(state->mag_img_filter, max_aniso)) |
4249 S_008F38_XY_MIN_FILTER(si_tex_filter(state->min_img_filter, max_aniso)) |
4250 S_008F38_MIP_FILTER(si_tex_mipfilter(state->min_mip_filter)) |
4251 S_008F38_MIP_POINT_PRECLAMP(0) |
4252 S_008F38_DISABLE_LSB_CEIL(sctx->chip_class <= VI) |
4253 S_008F38_FILTER_PREC_FIX(1) |
4254 S_008F38_ANISO_OVERRIDE(sctx->chip_class >= VI));
4255 rstate->val[3] = si_translate_border_color(sctx, state, &state->border_color, false);
4256
4257 /* Create sampler resource for integer textures. */
4258 memcpy(rstate->integer_val, rstate->val, sizeof(rstate->val));
4259 rstate->integer_val[3] = si_translate_border_color(sctx, state, &state->border_color, true);
4260
4261 /* Create sampler resource for upgraded depth textures. */
4262 memcpy(rstate->upgraded_depth_val, rstate->val, sizeof(rstate->val));
4263
4264 for (unsigned i = 0; i < 4; ++i) {
4265 /* Use channel 0 on purpose, so that we can use OPAQUE_WHITE
4266 * when the border color is 1.0. */
4267 clamped_border_color.f[i] = CLAMP(state->border_color.f[0], 0, 1);
4268 }
4269
4270 if (memcmp(&state->border_color, &clamped_border_color, sizeof(clamped_border_color)) == 0)
4271 rstate->upgraded_depth_val[3] |= S_008F3C_UPGRADED_DEPTH(1);
4272 else
4273 rstate->upgraded_depth_val[3] =
4274 si_translate_border_color(sctx, state, &clamped_border_color, false) |
4275 S_008F3C_UPGRADED_DEPTH(1);
4276
4277 return rstate;
4278 }
4279
4280 static void si_set_sample_mask(struct pipe_context *ctx, unsigned sample_mask)
4281 {
4282 struct si_context *sctx = (struct si_context *)ctx;
4283
4284 if (sctx->sample_mask == (uint16_t)sample_mask)
4285 return;
4286
4287 sctx->sample_mask = sample_mask;
4288 si_mark_atom_dirty(sctx, &sctx->atoms.s.sample_mask);
4289 }
4290
4291 static void si_emit_sample_mask(struct si_context *sctx)
4292 {
4293 struct radeon_winsys_cs *cs = sctx->gfx_cs;
4294 unsigned mask = sctx->sample_mask;
4295
4296 /* Needed for line and polygon smoothing as well as for the Polaris
4297 * small primitive filter. We expect the state tracker to take care of
4298 * this for us.
4299 */
4300 assert(mask == 0xffff || sctx->framebuffer.nr_samples > 1 ||
4301 (mask & 1 && sctx->blitter->running));
4302
4303 radeon_set_context_reg_seq(cs, R_028C38_PA_SC_AA_MASK_X0Y0_X1Y0, 2);
4304 radeon_emit(cs, mask | (mask << 16));
4305 radeon_emit(cs, mask | (mask << 16));
4306 }
4307
4308 static void si_delete_sampler_state(struct pipe_context *ctx, void *state)
4309 {
4310 #ifdef DEBUG
4311 struct si_sampler_state *s = state;
4312
4313 assert(s->magic == SI_SAMPLER_STATE_MAGIC);
4314 s->magic = 0;
4315 #endif
4316 free(state);
4317 }
4318
4319 /*
4320 * Vertex elements & buffers
4321 */
4322
4323 static void *si_create_vertex_elements(struct pipe_context *ctx,
4324 unsigned count,
4325 const struct pipe_vertex_element *elements)
4326 {
4327 struct si_screen *sscreen = (struct si_screen*)ctx->screen;
4328 struct si_vertex_elements *v = CALLOC_STRUCT(si_vertex_elements);
4329 bool used[SI_NUM_VERTEX_BUFFERS] = {};
4330 int i;
4331
4332 assert(count <= SI_MAX_ATTRIBS);
4333 if (!v)
4334 return NULL;
4335
4336 v->count = count;
4337 v->desc_list_byte_size = align(count * 16, SI_CPDMA_ALIGNMENT);
4338
4339 for (i = 0; i < count; ++i) {
4340 const struct util_format_description *desc;
4341 const struct util_format_channel_description *channel;
4342 unsigned data_format, num_format;
4343 int first_non_void;
4344 unsigned vbo_index = elements[i].vertex_buffer_index;
4345 unsigned char swizzle[4];
4346
4347 if (vbo_index >= SI_NUM_VERTEX_BUFFERS) {
4348 FREE(v);
4349 return NULL;
4350 }
4351
4352 if (elements[i].instance_divisor) {
4353 v->uses_instance_divisors = true;
4354 v->instance_divisors[i] = elements[i].instance_divisor;
4355
4356 if (v->instance_divisors[i] == 1)
4357 v->instance_divisor_is_one |= 1u << i;
4358 else
4359 v->instance_divisor_is_fetched |= 1u << i;
4360 }
4361
4362 if (!used[vbo_index]) {
4363 v->first_vb_use_mask |= 1 << i;
4364 used[vbo_index] = true;
4365 }
4366
4367 desc = util_format_description(elements[i].src_format);
4368 first_non_void = util_format_get_first_non_void_channel(elements[i].src_format);
4369 data_format = si_translate_buffer_dataformat(ctx->screen, desc, first_non_void);
4370 num_format = si_translate_buffer_numformat(ctx->screen, desc, first_non_void);
4371 channel = first_non_void >= 0 ? &desc->channel[first_non_void] : NULL;
4372 memcpy(swizzle, desc->swizzle, sizeof(swizzle));
4373
4374 v->format_size[i] = desc->block.bits / 8;
4375 v->src_offset[i] = elements[i].src_offset;
4376 v->vertex_buffer_index[i] = vbo_index;
4377
4378 /* The hardware always treats the 2-bit alpha channel as
4379 * unsigned, so a shader workaround is needed. The affected
4380 * chips are VI and older except Stoney (GFX8.1).
4381 */
4382 if (data_format == V_008F0C_BUF_DATA_FORMAT_2_10_10_10 &&
4383 sscreen->info.chip_class <= VI &&
4384 sscreen->info.family != CHIP_STONEY) {
4385 if (num_format == V_008F0C_BUF_NUM_FORMAT_SNORM) {
4386 v->fix_fetch[i] = SI_FIX_FETCH_A2_SNORM;
4387 } else if (num_format == V_008F0C_BUF_NUM_FORMAT_SSCALED) {
4388 v->fix_fetch[i] = SI_FIX_FETCH_A2_SSCALED;
4389 } else if (num_format == V_008F0C_BUF_NUM_FORMAT_SINT) {
4390 /* This isn't actually used in OpenGL. */
4391 v->fix_fetch[i] = SI_FIX_FETCH_A2_SINT;
4392 }
4393 } else if (channel && channel->type == UTIL_FORMAT_TYPE_FIXED) {
4394 if (desc->swizzle[3] == PIPE_SWIZZLE_1)
4395 v->fix_fetch[i] = SI_FIX_FETCH_RGBX_32_FIXED;
4396 else
4397 v->fix_fetch[i] = SI_FIX_FETCH_RGBA_32_FIXED;
4398 } else if (channel && channel->size == 32 && !channel->pure_integer) {
4399 if (channel->type == UTIL_FORMAT_TYPE_SIGNED) {
4400 if (channel->normalized) {
4401 if (desc->swizzle[3] == PIPE_SWIZZLE_1)
4402 v->fix_fetch[i] = SI_FIX_FETCH_RGBX_32_SNORM;
4403 else
4404 v->fix_fetch[i] = SI_FIX_FETCH_RGBA_32_SNORM;
4405 } else {
4406 v->fix_fetch[i] = SI_FIX_FETCH_RGBA_32_SSCALED;
4407 }
4408 } else if (channel->type == UTIL_FORMAT_TYPE_UNSIGNED) {
4409 if (channel->normalized) {
4410 if (desc->swizzle[3] == PIPE_SWIZZLE_1)
4411 v->fix_fetch[i] = SI_FIX_FETCH_RGBX_32_UNORM;
4412 else
4413 v->fix_fetch[i] = SI_FIX_FETCH_RGBA_32_UNORM;
4414 } else {
4415 v->fix_fetch[i] = SI_FIX_FETCH_RGBA_32_USCALED;
4416 }
4417 }
4418 } else if (channel && channel->size == 64 &&
4419 channel->type == UTIL_FORMAT_TYPE_FLOAT) {
4420 switch (desc->nr_channels) {
4421 case 1:
4422 case 2:
4423 v->fix_fetch[i] = SI_FIX_FETCH_RG_64_FLOAT;
4424 swizzle[0] = PIPE_SWIZZLE_X;
4425 swizzle[1] = PIPE_SWIZZLE_Y;
4426 swizzle[2] = desc->nr_channels == 2 ? PIPE_SWIZZLE_Z : PIPE_SWIZZLE_0;
4427 swizzle[3] = desc->nr_channels == 2 ? PIPE_SWIZZLE_W : PIPE_SWIZZLE_0;
4428 break;
4429 case 3:
4430 v->fix_fetch[i] = SI_FIX_FETCH_RGB_64_FLOAT;
4431 swizzle[0] = PIPE_SWIZZLE_X; /* 3 loads */
4432 swizzle[1] = PIPE_SWIZZLE_Y;
4433 swizzle[2] = PIPE_SWIZZLE_0;
4434 swizzle[3] = PIPE_SWIZZLE_0;
4435 break;
4436 case 4:
4437 v->fix_fetch[i] = SI_FIX_FETCH_RGBA_64_FLOAT;
4438 swizzle[0] = PIPE_SWIZZLE_X; /* 2 loads */
4439 swizzle[1] = PIPE_SWIZZLE_Y;
4440 swizzle[2] = PIPE_SWIZZLE_Z;
4441 swizzle[3] = PIPE_SWIZZLE_W;
4442 break;
4443 default:
4444 assert(0);
4445 }
4446 } else if (channel && desc->nr_channels == 3) {
4447 assert(desc->swizzle[0] == PIPE_SWIZZLE_X);
4448
4449 if (channel->size == 8) {
4450 if (channel->pure_integer)
4451 v->fix_fetch[i] = SI_FIX_FETCH_RGB_8_INT;
4452 else
4453 v->fix_fetch[i] = SI_FIX_FETCH_RGB_8;
4454 } else if (channel->size == 16) {
4455 if (channel->pure_integer)
4456 v->fix_fetch[i] = SI_FIX_FETCH_RGB_16_INT;
4457 else
4458 v->fix_fetch[i] = SI_FIX_FETCH_RGB_16;
4459 }
4460 }
4461
4462 v->rsrc_word3[i] = S_008F0C_DST_SEL_X(si_map_swizzle(swizzle[0])) |
4463 S_008F0C_DST_SEL_Y(si_map_swizzle(swizzle[1])) |
4464 S_008F0C_DST_SEL_Z(si_map_swizzle(swizzle[2])) |
4465 S_008F0C_DST_SEL_W(si_map_swizzle(swizzle[3])) |
4466 S_008F0C_NUM_FORMAT(num_format) |
4467 S_008F0C_DATA_FORMAT(data_format);
4468 }
4469 return v;
4470 }
4471
4472 static void si_bind_vertex_elements(struct pipe_context *ctx, void *state)
4473 {
4474 struct si_context *sctx = (struct si_context *)ctx;
4475 struct si_vertex_elements *old = sctx->vertex_elements;
4476 struct si_vertex_elements *v = (struct si_vertex_elements*)state;
4477
4478 sctx->vertex_elements = v;
4479 sctx->vertex_buffers_dirty = true;
4480
4481 if (v &&
4482 (!old ||
4483 old->count != v->count ||
4484 old->uses_instance_divisors != v->uses_instance_divisors ||
4485 v->uses_instance_divisors || /* we don't check which divisors changed */
4486 memcmp(old->fix_fetch, v->fix_fetch, sizeof(v->fix_fetch[0]) * v->count)))
4487 sctx->do_update_shaders = true;
4488
4489 if (v && v->instance_divisor_is_fetched) {
4490 struct pipe_constant_buffer cb;
4491
4492 cb.buffer = NULL;
4493 cb.user_buffer = v->instance_divisors;
4494 cb.buffer_offset = 0;
4495 cb.buffer_size = sizeof(uint32_t) * v->count;
4496 si_set_rw_buffer(sctx, SI_VS_CONST_INSTANCE_DIVISORS, &cb);
4497 }
4498 }
4499
4500 static void si_delete_vertex_element(struct pipe_context *ctx, void *state)
4501 {
4502 struct si_context *sctx = (struct si_context *)ctx;
4503
4504 if (sctx->vertex_elements == state)
4505 sctx->vertex_elements = NULL;
4506 FREE(state);
4507 }
4508
4509 static void si_set_vertex_buffers(struct pipe_context *ctx,
4510 unsigned start_slot, unsigned count,
4511 const struct pipe_vertex_buffer *buffers)
4512 {
4513 struct si_context *sctx = (struct si_context *)ctx;
4514 struct pipe_vertex_buffer *dst = sctx->vertex_buffer + start_slot;
4515 int i;
4516
4517 assert(start_slot + count <= ARRAY_SIZE(sctx->vertex_buffer));
4518
4519 if (buffers) {
4520 for (i = 0; i < count; i++) {
4521 const struct pipe_vertex_buffer *src = buffers + i;
4522 struct pipe_vertex_buffer *dsti = dst + i;
4523 struct pipe_resource *buf = src->buffer.resource;
4524
4525 pipe_resource_reference(&dsti->buffer.resource, buf);
4526 dsti->buffer_offset = src->buffer_offset;
4527 dsti->stride = src->stride;
4528 si_context_add_resource_size(sctx, buf);
4529 if (buf)
4530 r600_resource(buf)->bind_history |= PIPE_BIND_VERTEX_BUFFER;
4531 }
4532 } else {
4533 for (i = 0; i < count; i++) {
4534 pipe_resource_reference(&dst[i].buffer.resource, NULL);
4535 }
4536 }
4537 sctx->vertex_buffers_dirty = true;
4538 }
4539
4540 /*
4541 * Misc
4542 */
4543
4544 static void si_set_tess_state(struct pipe_context *ctx,
4545 const float default_outer_level[4],
4546 const float default_inner_level[2])
4547 {
4548 struct si_context *sctx = (struct si_context *)ctx;
4549 struct pipe_constant_buffer cb;
4550 float array[8];
4551
4552 memcpy(array, default_outer_level, sizeof(float) * 4);
4553 memcpy(array+4, default_inner_level, sizeof(float) * 2);
4554
4555 cb.buffer = NULL;
4556 cb.user_buffer = NULL;
4557 cb.buffer_size = sizeof(array);
4558
4559 si_upload_const_buffer(sctx, (struct r600_resource**)&cb.buffer,
4560 (void*)array, sizeof(array),
4561 &cb.buffer_offset);
4562
4563 si_set_rw_buffer(sctx, SI_HS_CONST_DEFAULT_TESS_LEVELS, &cb);
4564 pipe_resource_reference(&cb.buffer, NULL);
4565 }
4566
4567 static void si_texture_barrier(struct pipe_context *ctx, unsigned flags)
4568 {
4569 struct si_context *sctx = (struct si_context *)ctx;
4570
4571 si_update_fb_dirtiness_after_rendering(sctx);
4572
4573 /* Multisample surfaces are flushed in si_decompress_textures. */
4574 if (sctx->framebuffer.uncompressed_cb_mask)
4575 si_make_CB_shader_coherent(sctx, sctx->framebuffer.nr_samples,
4576 sctx->framebuffer.CB_has_shader_readable_metadata);
4577 }
4578
4579 /* This only ensures coherency for shader image/buffer stores. */
4580 static void si_memory_barrier(struct pipe_context *ctx, unsigned flags)
4581 {
4582 struct si_context *sctx = (struct si_context *)ctx;
4583
4584 /* Subsequent commands must wait for all shader invocations to
4585 * complete. */
4586 sctx->flags |= SI_CONTEXT_PS_PARTIAL_FLUSH |
4587 SI_CONTEXT_CS_PARTIAL_FLUSH;
4588
4589 if (flags & PIPE_BARRIER_CONSTANT_BUFFER)
4590 sctx->flags |= SI_CONTEXT_INV_SMEM_L1 |
4591 SI_CONTEXT_INV_VMEM_L1;
4592
4593 if (flags & (PIPE_BARRIER_VERTEX_BUFFER |
4594 PIPE_BARRIER_SHADER_BUFFER |
4595 PIPE_BARRIER_TEXTURE |
4596 PIPE_BARRIER_IMAGE |
4597 PIPE_BARRIER_STREAMOUT_BUFFER |
4598 PIPE_BARRIER_GLOBAL_BUFFER)) {
4599 /* As far as I can tell, L1 contents are written back to L2
4600 * automatically at end of shader, but the contents of other
4601 * L1 caches might still be stale. */
4602 sctx->flags |= SI_CONTEXT_INV_VMEM_L1;
4603 }
4604
4605 if (flags & PIPE_BARRIER_INDEX_BUFFER) {
4606 /* Indices are read through TC L2 since VI.
4607 * L1 isn't used.
4608 */
4609 if (sctx->screen->info.chip_class <= CIK)
4610 sctx->flags |= SI_CONTEXT_WRITEBACK_GLOBAL_L2;
4611 }
4612
4613 /* MSAA color, any depth and any stencil are flushed in
4614 * si_decompress_textures when needed.
4615 */
4616 if (flags & PIPE_BARRIER_FRAMEBUFFER &&
4617 sctx->framebuffer.uncompressed_cb_mask) {
4618 sctx->flags |= SI_CONTEXT_FLUSH_AND_INV_CB;
4619
4620 if (sctx->chip_class <= VI)
4621 sctx->flags |= SI_CONTEXT_WRITEBACK_GLOBAL_L2;
4622 }
4623
4624 /* Indirect buffers use TC L2 on GFX9, but not older hw. */
4625 if (sctx->screen->info.chip_class <= VI &&
4626 flags & PIPE_BARRIER_INDIRECT_BUFFER)
4627 sctx->flags |= SI_CONTEXT_WRITEBACK_GLOBAL_L2;
4628 }
4629
4630 static void *si_create_blend_custom(struct si_context *sctx, unsigned mode)
4631 {
4632 struct pipe_blend_state blend;
4633
4634 memset(&blend, 0, sizeof(blend));
4635 blend.independent_blend_enable = true;
4636 blend.rt[0].colormask = 0xf;
4637 return si_create_blend_state_mode(&sctx->b, &blend, mode);
4638 }
4639
4640 static void si_init_config(struct si_context *sctx);
4641
4642 void si_init_state_functions(struct si_context *sctx)
4643 {
4644 sctx->atoms.s.framebuffer.emit = si_emit_framebuffer_state;
4645 sctx->atoms.s.msaa_sample_locs.emit = si_emit_msaa_sample_locs;
4646 sctx->atoms.s.db_render_state.emit = si_emit_db_render_state;
4647 sctx->atoms.s.dpbb_state.emit = si_emit_dpbb_state;
4648 sctx->atoms.s.msaa_config.emit = si_emit_msaa_config;
4649 sctx->atoms.s.sample_mask.emit = si_emit_sample_mask;
4650 sctx->atoms.s.cb_render_state.emit = si_emit_cb_render_state;
4651 sctx->atoms.s.blend_color.emit = si_emit_blend_color;
4652 sctx->atoms.s.clip_regs.emit = si_emit_clip_regs;
4653 sctx->atoms.s.clip_state.emit = si_emit_clip_state;
4654 sctx->atoms.s.stencil_ref.emit = si_emit_stencil_ref;
4655
4656 sctx->b.create_blend_state = si_create_blend_state;
4657 sctx->b.bind_blend_state = si_bind_blend_state;
4658 sctx->b.delete_blend_state = si_delete_blend_state;
4659 sctx->b.set_blend_color = si_set_blend_color;
4660
4661 sctx->b.create_rasterizer_state = si_create_rs_state;
4662 sctx->b.bind_rasterizer_state = si_bind_rs_state;
4663 sctx->b.delete_rasterizer_state = si_delete_rs_state;
4664
4665 sctx->b.create_depth_stencil_alpha_state = si_create_dsa_state;
4666 sctx->b.bind_depth_stencil_alpha_state = si_bind_dsa_state;
4667 sctx->b.delete_depth_stencil_alpha_state = si_delete_dsa_state;
4668
4669 sctx->custom_dsa_flush = si_create_db_flush_dsa(sctx);
4670 sctx->custom_blend_resolve = si_create_blend_custom(sctx, V_028808_CB_RESOLVE);
4671 sctx->custom_blend_fmask_decompress = si_create_blend_custom(sctx, V_028808_CB_FMASK_DECOMPRESS);
4672 sctx->custom_blend_eliminate_fastclear = si_create_blend_custom(sctx, V_028808_CB_ELIMINATE_FAST_CLEAR);
4673 sctx->custom_blend_dcc_decompress = si_create_blend_custom(sctx, V_028808_CB_DCC_DECOMPRESS);
4674
4675 sctx->b.set_clip_state = si_set_clip_state;
4676 sctx->b.set_stencil_ref = si_set_stencil_ref;
4677
4678 sctx->b.set_framebuffer_state = si_set_framebuffer_state;
4679
4680 sctx->b.create_sampler_state = si_create_sampler_state;
4681 sctx->b.delete_sampler_state = si_delete_sampler_state;
4682
4683 sctx->b.create_sampler_view = si_create_sampler_view;
4684 sctx->b.sampler_view_destroy = si_sampler_view_destroy;
4685
4686 sctx->b.set_sample_mask = si_set_sample_mask;
4687
4688 sctx->b.create_vertex_elements_state = si_create_vertex_elements;
4689 sctx->b.bind_vertex_elements_state = si_bind_vertex_elements;
4690 sctx->b.delete_vertex_elements_state = si_delete_vertex_element;
4691 sctx->b.set_vertex_buffers = si_set_vertex_buffers;
4692
4693 sctx->b.texture_barrier = si_texture_barrier;
4694 sctx->b.memory_barrier = si_memory_barrier;
4695 sctx->b.set_min_samples = si_set_min_samples;
4696 sctx->b.set_tess_state = si_set_tess_state;
4697
4698 sctx->b.set_active_query_state = si_set_active_query_state;
4699
4700 sctx->b.draw_vbo = si_draw_vbo;
4701
4702 si_init_config(sctx);
4703 }
4704
4705 void si_init_screen_state_functions(struct si_screen *sscreen)
4706 {
4707 sscreen->b.is_format_supported = si_is_format_supported;
4708 }
4709
4710 static void si_set_grbm_gfx_index(struct si_context *sctx,
4711 struct si_pm4_state *pm4, unsigned value)
4712 {
4713 unsigned reg = sctx->chip_class >= CIK ? R_030800_GRBM_GFX_INDEX :
4714 R_00802C_GRBM_GFX_INDEX;
4715 si_pm4_set_reg(pm4, reg, value);
4716 }
4717
4718 static void si_set_grbm_gfx_index_se(struct si_context *sctx,
4719 struct si_pm4_state *pm4, unsigned se)
4720 {
4721 assert(se == ~0 || se < sctx->screen->info.max_se);
4722 si_set_grbm_gfx_index(sctx, pm4,
4723 (se == ~0 ? S_030800_SE_BROADCAST_WRITES(1) :
4724 S_030800_SE_INDEX(se)) |
4725 S_030800_SH_BROADCAST_WRITES(1) |
4726 S_030800_INSTANCE_BROADCAST_WRITES(1));
4727 }
4728
4729 static void
4730 si_write_harvested_raster_configs(struct si_context *sctx,
4731 struct si_pm4_state *pm4,
4732 unsigned raster_config,
4733 unsigned raster_config_1)
4734 {
4735 unsigned num_se = MAX2(sctx->screen->info.max_se, 1);
4736 unsigned raster_config_se[4];
4737 unsigned se;
4738
4739 ac_get_harvested_configs(&sctx->screen->info,
4740 raster_config,
4741 &raster_config_1,
4742 raster_config_se);
4743
4744 for (se = 0; se < num_se; se++) {
4745 si_set_grbm_gfx_index_se(sctx, pm4, se);
4746 si_pm4_set_reg(pm4, R_028350_PA_SC_RASTER_CONFIG, raster_config_se[se]);
4747 }
4748 si_set_grbm_gfx_index(sctx, pm4, ~0);
4749
4750 if (sctx->chip_class >= CIK) {
4751 si_pm4_set_reg(pm4, R_028354_PA_SC_RASTER_CONFIG_1, raster_config_1);
4752 }
4753 }
4754
4755 static void si_set_raster_config(struct si_context *sctx, struct si_pm4_state *pm4)
4756 {
4757 unsigned num_rb = MIN2(sctx->screen->info.num_render_backends, 16);
4758 unsigned rb_mask = sctx->screen->info.enabled_rb_mask;
4759 unsigned raster_config, raster_config_1;
4760
4761 ac_get_raster_config(&sctx->screen->info,
4762 &raster_config,
4763 &raster_config_1);
4764
4765 if (!rb_mask || util_bitcount(rb_mask) >= num_rb) {
4766 /* Always use the default config when all backends are enabled
4767 * (or when we failed to determine the enabled backends).
4768 */
4769 si_pm4_set_reg(pm4, R_028350_PA_SC_RASTER_CONFIG,
4770 raster_config);
4771 if (sctx->chip_class >= CIK)
4772 si_pm4_set_reg(pm4, R_028354_PA_SC_RASTER_CONFIG_1,
4773 raster_config_1);
4774 } else {
4775 si_write_harvested_raster_configs(sctx, pm4, raster_config, raster_config_1);
4776 }
4777 }
4778
4779 static void si_init_config(struct si_context *sctx)
4780 {
4781 struct si_screen *sscreen = sctx->screen;
4782 uint64_t border_color_va = sctx->border_color_buffer->gpu_address;
4783 bool has_clear_state = sscreen->has_clear_state;
4784 struct si_pm4_state *pm4 = CALLOC_STRUCT(si_pm4_state);
4785
4786 /* Only SI can disable CLEAR_STATE for now. */
4787 assert(has_clear_state || sscreen->info.chip_class == SI);
4788
4789 if (!pm4)
4790 return;
4791
4792 si_pm4_cmd_begin(pm4, PKT3_CONTEXT_CONTROL);
4793 si_pm4_cmd_add(pm4, CONTEXT_CONTROL_LOAD_ENABLE(1));
4794 si_pm4_cmd_add(pm4, CONTEXT_CONTROL_SHADOW_ENABLE(1));
4795 si_pm4_cmd_end(pm4, false);
4796
4797 if (has_clear_state) {
4798 si_pm4_cmd_begin(pm4, PKT3_CLEAR_STATE);
4799 si_pm4_cmd_add(pm4, 0);
4800 si_pm4_cmd_end(pm4, false);
4801 }
4802
4803 if (sctx->chip_class <= VI)
4804 si_set_raster_config(sctx, pm4);
4805
4806 si_pm4_set_reg(pm4, R_028A18_VGT_HOS_MAX_TESS_LEVEL, fui(64));
4807 if (!has_clear_state)
4808 si_pm4_set_reg(pm4, R_028A1C_VGT_HOS_MIN_TESS_LEVEL, fui(0));
4809
4810 /* FIXME calculate these values somehow ??? */
4811 if (sctx->chip_class <= VI) {
4812 si_pm4_set_reg(pm4, R_028A54_VGT_GS_PER_ES, SI_GS_PER_ES);
4813 si_pm4_set_reg(pm4, R_028A58_VGT_ES_PER_GS, 0x40);
4814 }
4815
4816 if (!has_clear_state) {
4817 si_pm4_set_reg(pm4, R_028A5C_VGT_GS_PER_VS, 0x2);
4818 si_pm4_set_reg(pm4, R_028A8C_VGT_PRIMITIVEID_RESET, 0x0);
4819 si_pm4_set_reg(pm4, R_028B98_VGT_STRMOUT_BUFFER_CONFIG, 0x0);
4820 }
4821
4822 si_pm4_set_reg(pm4, R_028AA0_VGT_INSTANCE_STEP_RATE_0, 1);
4823 if (!has_clear_state)
4824 si_pm4_set_reg(pm4, R_028AB8_VGT_VTX_CNT_EN, 0x0);
4825 if (sctx->chip_class < CIK)
4826 si_pm4_set_reg(pm4, R_008A14_PA_CL_ENHANCE, S_008A14_NUM_CLIP_SEQ(3) |
4827 S_008A14_CLIP_VTX_REORDER_ENA(1));
4828
4829 if (!has_clear_state)
4830 si_pm4_set_reg(pm4, R_02882C_PA_SU_PRIM_FILTER_CNTL, 0);
4831
4832 /* CLEAR_STATE doesn't clear these correctly on certain generations.
4833 * I don't know why. Deduced by trial and error.
4834 */
4835 if (sctx->chip_class <= CIK) {
4836 si_pm4_set_reg(pm4, R_028B28_VGT_STRMOUT_DRAW_OPAQUE_OFFSET, 0);
4837 si_pm4_set_reg(pm4, R_028204_PA_SC_WINDOW_SCISSOR_TL, S_028204_WINDOW_OFFSET_DISABLE(1));
4838 si_pm4_set_reg(pm4, R_028240_PA_SC_GENERIC_SCISSOR_TL, S_028240_WINDOW_OFFSET_DISABLE(1));
4839 si_pm4_set_reg(pm4, R_028244_PA_SC_GENERIC_SCISSOR_BR,
4840 S_028244_BR_X(16384) | S_028244_BR_Y(16384));
4841 si_pm4_set_reg(pm4, R_028030_PA_SC_SCREEN_SCISSOR_TL, 0);
4842 si_pm4_set_reg(pm4, R_028034_PA_SC_SCREEN_SCISSOR_BR,
4843 S_028034_BR_X(16384) | S_028034_BR_Y(16384));
4844 }
4845
4846 if (!has_clear_state) {
4847 si_pm4_set_reg(pm4, R_02820C_PA_SC_CLIPRECT_RULE, 0xFFFF);
4848 si_pm4_set_reg(pm4, R_028230_PA_SC_EDGERULE,
4849 S_028230_ER_TRI(0xA) |
4850 S_028230_ER_POINT(0xA) |
4851 S_028230_ER_RECT(0xA) |
4852 /* Required by DX10_DIAMOND_TEST_ENA: */
4853 S_028230_ER_LINE_LR(0x1A) |
4854 S_028230_ER_LINE_RL(0x26) |
4855 S_028230_ER_LINE_TB(0xA) |
4856 S_028230_ER_LINE_BT(0xA));
4857 /* PA_SU_HARDWARE_SCREEN_OFFSET must be 0 due to hw bug on SI */
4858 si_pm4_set_reg(pm4, R_028234_PA_SU_HARDWARE_SCREEN_OFFSET, 0);
4859 si_pm4_set_reg(pm4, R_028820_PA_CL_NANINF_CNTL, 0);
4860 si_pm4_set_reg(pm4, R_028AC0_DB_SRESULTS_COMPARE_STATE0, 0x0);
4861 si_pm4_set_reg(pm4, R_028AC4_DB_SRESULTS_COMPARE_STATE1, 0x0);
4862 si_pm4_set_reg(pm4, R_028AC8_DB_PRELOAD_CONTROL, 0x0);
4863 si_pm4_set_reg(pm4, R_02800C_DB_RENDER_OVERRIDE, 0);
4864 }
4865
4866 if (sctx->chip_class >= GFX9) {
4867 si_pm4_set_reg(pm4, R_030920_VGT_MAX_VTX_INDX, ~0);
4868 si_pm4_set_reg(pm4, R_030924_VGT_MIN_VTX_INDX, 0);
4869 si_pm4_set_reg(pm4, R_030928_VGT_INDX_OFFSET, 0);
4870 } else {
4871 /* These registers, when written, also overwrite the CLEAR_STATE
4872 * context, so we can't rely on CLEAR_STATE setting them.
4873 * It would be an issue if there was another UMD changing them.
4874 */
4875 si_pm4_set_reg(pm4, R_028400_VGT_MAX_VTX_INDX, ~0);
4876 si_pm4_set_reg(pm4, R_028404_VGT_MIN_VTX_INDX, 0);
4877 si_pm4_set_reg(pm4, R_028408_VGT_INDX_OFFSET, 0);
4878 }
4879
4880 if (sctx->chip_class >= CIK) {
4881 if (sctx->chip_class >= GFX9) {
4882 si_pm4_set_reg(pm4, R_00B41C_SPI_SHADER_PGM_RSRC3_HS,
4883 S_00B41C_CU_EN(0xffff) | S_00B41C_WAVE_LIMIT(0x3F));
4884 } else {
4885 si_pm4_set_reg(pm4, R_00B51C_SPI_SHADER_PGM_RSRC3_LS,
4886 S_00B51C_CU_EN(0xffff) | S_00B51C_WAVE_LIMIT(0x3F));
4887 si_pm4_set_reg(pm4, R_00B41C_SPI_SHADER_PGM_RSRC3_HS,
4888 S_00B41C_WAVE_LIMIT(0x3F));
4889 si_pm4_set_reg(pm4, R_00B31C_SPI_SHADER_PGM_RSRC3_ES,
4890 S_00B31C_CU_EN(0xffff) | S_00B31C_WAVE_LIMIT(0x3F));
4891
4892 /* If this is 0, Bonaire can hang even if GS isn't being used.
4893 * Other chips are unaffected. These are suboptimal values,
4894 * but we don't use on-chip GS.
4895 */
4896 si_pm4_set_reg(pm4, R_028A44_VGT_GS_ONCHIP_CNTL,
4897 S_028A44_ES_VERTS_PER_SUBGRP(64) |
4898 S_028A44_GS_PRIMS_PER_SUBGRP(4));
4899 }
4900 si_pm4_set_reg(pm4, R_00B21C_SPI_SHADER_PGM_RSRC3_GS,
4901 S_00B21C_CU_EN(0xffff) | S_00B21C_WAVE_LIMIT(0x3F));
4902
4903 /* Compute LATE_ALLOC_VS.LIMIT. */
4904 unsigned num_cu_per_sh = sscreen->info.num_good_compute_units /
4905 (sscreen->info.max_se *
4906 sscreen->info.max_sh_per_se);
4907 unsigned late_alloc_limit; /* The limit is per SH. */
4908
4909 if (sctx->family == CHIP_KABINI) {
4910 late_alloc_limit = 0; /* Potential hang on Kabini. */
4911 } else if (num_cu_per_sh <= 4) {
4912 /* Too few available compute units per SH. Disallowing
4913 * VS to run on one CU could hurt us more than late VS
4914 * allocation would help.
4915 *
4916 * 2 is the highest safe number that allows us to keep
4917 * all CUs enabled.
4918 */
4919 late_alloc_limit = 2;
4920 } else {
4921 /* This is a good initial value, allowing 1 late_alloc
4922 * wave per SIMD on num_cu - 2.
4923 */
4924 late_alloc_limit = (num_cu_per_sh - 2) * 4;
4925
4926 /* The limit is 0-based, so 0 means 1. */
4927 assert(late_alloc_limit > 0 && late_alloc_limit <= 64);
4928 late_alloc_limit -= 1;
4929 }
4930
4931 /* VS can't execute on one CU if the limit is > 2. */
4932 si_pm4_set_reg(pm4, R_00B118_SPI_SHADER_PGM_RSRC3_VS,
4933 S_00B118_CU_EN(late_alloc_limit > 2 ? 0xfffe : 0xffff) |
4934 S_00B118_WAVE_LIMIT(0x3F));
4935 si_pm4_set_reg(pm4, R_00B11C_SPI_SHADER_LATE_ALLOC_VS,
4936 S_00B11C_LIMIT(late_alloc_limit));
4937 si_pm4_set_reg(pm4, R_00B01C_SPI_SHADER_PGM_RSRC3_PS,
4938 S_00B01C_CU_EN(0xffff) | S_00B01C_WAVE_LIMIT(0x3F));
4939 }
4940
4941 if (sctx->chip_class >= VI) {
4942 unsigned vgt_tess_distribution;
4943
4944 vgt_tess_distribution =
4945 S_028B50_ACCUM_ISOLINE(32) |
4946 S_028B50_ACCUM_TRI(11) |
4947 S_028B50_ACCUM_QUAD(11) |
4948 S_028B50_DONUT_SPLIT(16);
4949
4950 /* Testing with Unigine Heaven extreme tesselation yielded best results
4951 * with TRAP_SPLIT = 3.
4952 */
4953 if (sctx->family == CHIP_FIJI ||
4954 sctx->family >= CHIP_POLARIS10)
4955 vgt_tess_distribution |= S_028B50_TRAP_SPLIT(3);
4956
4957 si_pm4_set_reg(pm4, R_028B50_VGT_TESS_DISTRIBUTION, vgt_tess_distribution);
4958 } else if (!has_clear_state) {
4959 si_pm4_set_reg(pm4, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL, 14);
4960 si_pm4_set_reg(pm4, R_028C5C_VGT_OUT_DEALLOC_CNTL, 16);
4961 }
4962
4963 si_pm4_set_reg(pm4, R_028080_TA_BC_BASE_ADDR, border_color_va >> 8);
4964 if (sctx->chip_class >= CIK) {
4965 si_pm4_set_reg(pm4, R_028084_TA_BC_BASE_ADDR_HI,
4966 S_028084_ADDRESS(border_color_va >> 40));
4967 }
4968 si_pm4_add_bo(pm4, sctx->border_color_buffer, RADEON_USAGE_READ,
4969 RADEON_PRIO_BORDER_COLORS);
4970
4971 if (sctx->chip_class >= GFX9) {
4972 unsigned num_se = sscreen->info.max_se;
4973 unsigned pc_lines = 0;
4974
4975 switch (sctx->family) {
4976 case CHIP_VEGA10:
4977 case CHIP_VEGA12:
4978 pc_lines = 4096;
4979 break;
4980 case CHIP_RAVEN:
4981 pc_lines = 1024;
4982 break;
4983 default:
4984 assert(0);
4985 }
4986
4987 si_pm4_set_reg(pm4, R_028C48_PA_SC_BINNER_CNTL_1,
4988 S_028C48_MAX_ALLOC_COUNT(MIN2(128, pc_lines / (4 * num_se))) |
4989 S_028C48_MAX_PRIM_PER_BATCH(1023));
4990 si_pm4_set_reg(pm4, R_028C4C_PA_SC_CONSERVATIVE_RASTERIZATION_CNTL,
4991 S_028C4C_NULL_SQUAD_AA_MASK_ENABLE(1));
4992 si_pm4_set_reg(pm4, R_030968_VGT_INSTANCE_BASE_ID, 0);
4993 }
4994
4995 si_pm4_upload_indirect_buffer(sctx, pm4);
4996 sctx->init_config = pm4;
4997 }