projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
radeonsi: don't re-upload the sample position constant buffer repeatedly
[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_cmdbuf *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_opt_set_context_reg(sctx, R_028238_CB_TARGET_MASK,
91 SI_TRACKED_CB_TARGET_MASK, cb_target_mask);
92
93 /* GFX9: Flush DFSM when CB_TARGET_MASK changes.
94 * I think we don't have to do anything between IBs.
95 */
96 if (sctx->screen->dfsm_allowed &&
97 sctx->last_cb_target_mask != cb_target_mask) {
98 sctx->last_cb_target_mask = cb_target_mask;
99
100 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
101 radeon_emit(cs, EVENT_TYPE(V_028A90_FLUSH_DFSM) | EVENT_INDEX(0));
102 }
103
104 if (sctx->chip_class >= VI) {
105 /* DCC MSAA workaround for blending.
106 * Alternatively, we can set CB_COLORi_DCC_CONTROL.OVERWRITE_-
107 * COMBINER_DISABLE, but that would be more complicated.
108 */
109 bool oc_disable = (sctx->chip_class == VI ||
110 sctx->chip_class == GFX9) &&
111 blend &&
112 blend->blend_enable_4bit & cb_target_mask &&
113 sctx->framebuffer.nr_samples >= 2;
114
115 radeon_opt_set_context_reg(
116 sctx, R_028424_CB_DCC_CONTROL,
117 SI_TRACKED_CB_DCC_CONTROL,
118 S_028424_OVERWRITE_COMBINER_MRT_SHARING_DISABLE(1) |
119 S_028424_OVERWRITE_COMBINER_WATERMARK(4) |
120 S_028424_OVERWRITE_COMBINER_DISABLE(oc_disable));
121 }
122
123 /* RB+ register settings. */
124 if (sctx->screen->rbplus_allowed) {
125 unsigned spi_shader_col_format =
126 sctx->ps_shader.cso ?
127 sctx->ps_shader.current->key.part.ps.epilog.spi_shader_col_format : 0;
128 unsigned sx_ps_downconvert = 0;
129 unsigned sx_blend_opt_epsilon = 0;
130 unsigned sx_blend_opt_control = 0;
131
132 for (i = 0; i < sctx->framebuffer.state.nr_cbufs; i++) {
133 struct si_surface *surf =
134 (struct si_surface*)sctx->framebuffer.state.cbufs[i];
135 unsigned format, swap, spi_format, colormask;
136 bool has_alpha, has_rgb;
137
138 if (!surf)
139 continue;
140
141 format = G_028C70_FORMAT(surf->cb_color_info);
142 swap = G_028C70_COMP_SWAP(surf->cb_color_info);
143 spi_format = (spi_shader_col_format >> (i * 4)) & 0xf;
144 colormask = (cb_target_mask >> (i * 4)) & 0xf;
145
146 /* Set if RGB and A are present. */
147 has_alpha = !G_028C74_FORCE_DST_ALPHA_1(surf->cb_color_attrib);
148
149 if (format == V_028C70_COLOR_8 ||
150 format == V_028C70_COLOR_16 ||
151 format == V_028C70_COLOR_32)
152 has_rgb = !has_alpha;
153 else
154 has_rgb = true;
155
156 /* Check the colormask and export format. */
157 if (!(colormask & (PIPE_MASK_RGBA & ~PIPE_MASK_A)))
158 has_rgb = false;
159 if (!(colormask & PIPE_MASK_A))
160 has_alpha = false;
161
162 if (spi_format == V_028714_SPI_SHADER_ZERO) {
163 has_rgb = false;
164 has_alpha = false;
165 }
166
167 /* Disable value checking for disabled channels. */
168 if (!has_rgb)
169 sx_blend_opt_control |= S_02875C_MRT0_COLOR_OPT_DISABLE(1) << (i * 4);
170 if (!has_alpha)
171 sx_blend_opt_control |= S_02875C_MRT0_ALPHA_OPT_DISABLE(1) << (i * 4);
172
173 /* Enable down-conversion for 32bpp and smaller formats. */
174 switch (format) {
175 case V_028C70_COLOR_8:
176 case V_028C70_COLOR_8_8:
177 case V_028C70_COLOR_8_8_8_8:
178 /* For 1 and 2-channel formats, use the superset thereof. */
179 if (spi_format == V_028714_SPI_SHADER_FP16_ABGR ||
180 spi_format == V_028714_SPI_SHADER_UINT16_ABGR ||
181 spi_format == V_028714_SPI_SHADER_SINT16_ABGR) {
182 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_8_8_8_8 << (i * 4);
183 sx_blend_opt_epsilon |= V_028758_8BIT_FORMAT << (i * 4);
184 }
185 break;
186
187 case V_028C70_COLOR_5_6_5:
188 if (spi_format == V_028714_SPI_SHADER_FP16_ABGR) {
189 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_5_6_5 << (i * 4);
190 sx_blend_opt_epsilon |= V_028758_6BIT_FORMAT << (i * 4);
191 }
192 break;
193
194 case V_028C70_COLOR_1_5_5_5:
195 if (spi_format == V_028714_SPI_SHADER_FP16_ABGR) {
196 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_1_5_5_5 << (i * 4);
197 sx_blend_opt_epsilon |= V_028758_5BIT_FORMAT << (i * 4);
198 }
199 break;
200
201 case V_028C70_COLOR_4_4_4_4:
202 if (spi_format == V_028714_SPI_SHADER_FP16_ABGR) {
203 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_4_4_4_4 << (i * 4);
204 sx_blend_opt_epsilon |= V_028758_4BIT_FORMAT << (i * 4);
205 }
206 break;
207
208 case V_028C70_COLOR_32:
209 if (swap == V_028C70_SWAP_STD &&
210 spi_format == V_028714_SPI_SHADER_32_R)
211 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_32_R << (i * 4);
212 else if (swap == V_028C70_SWAP_ALT_REV &&
213 spi_format == V_028714_SPI_SHADER_32_AR)
214 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_32_A << (i * 4);
215 break;
216
217 case V_028C70_COLOR_16:
218 case V_028C70_COLOR_16_16:
219 /* For 1-channel formats, use the superset thereof. */
220 if (spi_format == V_028714_SPI_SHADER_UNORM16_ABGR ||
221 spi_format == V_028714_SPI_SHADER_SNORM16_ABGR ||
222 spi_format == V_028714_SPI_SHADER_UINT16_ABGR ||
223 spi_format == V_028714_SPI_SHADER_SINT16_ABGR) {
224 if (swap == V_028C70_SWAP_STD ||
225 swap == V_028C70_SWAP_STD_REV)
226 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_16_16_GR << (i * 4);
227 else
228 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_16_16_AR << (i * 4);
229 }
230 break;
231
232 case V_028C70_COLOR_10_11_11:
233 if (spi_format == V_028714_SPI_SHADER_FP16_ABGR) {
234 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_10_11_11 << (i * 4);
235 sx_blend_opt_epsilon |= V_028758_11BIT_FORMAT << (i * 4);
236 }
237 break;
238
239 case V_028C70_COLOR_2_10_10_10:
240 if (spi_format == V_028714_SPI_SHADER_FP16_ABGR) {
241 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_2_10_10_10 << (i * 4);
242 sx_blend_opt_epsilon |= V_028758_10BIT_FORMAT << (i * 4);
243 }
244 break;
245 }
246 }
247
248 /* SX_PS_DOWNCONVERT, SX_BLEND_OPT_EPSILON, SX_BLEND_OPT_CONTROL */
249 radeon_opt_set_context_reg3(sctx, R_028754_SX_PS_DOWNCONVERT,
250 SI_TRACKED_SX_PS_DOWNCONVERT,
251 sx_ps_downconvert, sx_blend_opt_epsilon,
252 sx_blend_opt_control);
253 }
254 }
255
256 /*
257 * Blender functions
258 */
259
260 static uint32_t si_translate_blend_function(int blend_func)
261 {
262 switch (blend_func) {
263 case PIPE_BLEND_ADD:
264 return V_028780_COMB_DST_PLUS_SRC;
265 case PIPE_BLEND_SUBTRACT:
266 return V_028780_COMB_SRC_MINUS_DST;
267 case PIPE_BLEND_REVERSE_SUBTRACT:
268 return V_028780_COMB_DST_MINUS_SRC;
269 case PIPE_BLEND_MIN:
270 return V_028780_COMB_MIN_DST_SRC;
271 case PIPE_BLEND_MAX:
272 return V_028780_COMB_MAX_DST_SRC;
273 default:
274 PRINT_ERR("Unknown blend function %d\n", blend_func);
275 assert(0);
276 break;
277 }
278 return 0;
279 }
280
281 static uint32_t si_translate_blend_factor(int blend_fact)
282 {
283 switch (blend_fact) {
284 case PIPE_BLENDFACTOR_ONE:
285 return V_028780_BLEND_ONE;
286 case PIPE_BLENDFACTOR_SRC_COLOR:
287 return V_028780_BLEND_SRC_COLOR;
288 case PIPE_BLENDFACTOR_SRC_ALPHA:
289 return V_028780_BLEND_SRC_ALPHA;
290 case PIPE_BLENDFACTOR_DST_ALPHA:
291 return V_028780_BLEND_DST_ALPHA;
292 case PIPE_BLENDFACTOR_DST_COLOR:
293 return V_028780_BLEND_DST_COLOR;
294 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
295 return V_028780_BLEND_SRC_ALPHA_SATURATE;
296 case PIPE_BLENDFACTOR_CONST_COLOR:
297 return V_028780_BLEND_CONSTANT_COLOR;
298 case PIPE_BLENDFACTOR_CONST_ALPHA:
299 return V_028780_BLEND_CONSTANT_ALPHA;
300 case PIPE_BLENDFACTOR_ZERO:
301 return V_028780_BLEND_ZERO;
302 case PIPE_BLENDFACTOR_INV_SRC_COLOR:
303 return V_028780_BLEND_ONE_MINUS_SRC_COLOR;
304 case PIPE_BLENDFACTOR_INV_SRC_ALPHA:
305 return V_028780_BLEND_ONE_MINUS_SRC_ALPHA;
306 case PIPE_BLENDFACTOR_INV_DST_ALPHA:
307 return V_028780_BLEND_ONE_MINUS_DST_ALPHA;
308 case PIPE_BLENDFACTOR_INV_DST_COLOR:
309 return V_028780_BLEND_ONE_MINUS_DST_COLOR;
310 case PIPE_BLENDFACTOR_INV_CONST_COLOR:
311 return V_028780_BLEND_ONE_MINUS_CONSTANT_COLOR;
312 case PIPE_BLENDFACTOR_INV_CONST_ALPHA:
313 return V_028780_BLEND_ONE_MINUS_CONSTANT_ALPHA;
314 case PIPE_BLENDFACTOR_SRC1_COLOR:
315 return V_028780_BLEND_SRC1_COLOR;
316 case PIPE_BLENDFACTOR_SRC1_ALPHA:
317 return V_028780_BLEND_SRC1_ALPHA;
318 case PIPE_BLENDFACTOR_INV_SRC1_COLOR:
319 return V_028780_BLEND_INV_SRC1_COLOR;
320 case PIPE_BLENDFACTOR_INV_SRC1_ALPHA:
321 return V_028780_BLEND_INV_SRC1_ALPHA;
322 default:
323 PRINT_ERR("Bad blend factor %d not supported!\n", blend_fact);
324 assert(0);
325 break;
326 }
327 return 0;
328 }
329
330 static uint32_t si_translate_blend_opt_function(int blend_func)
331 {
332 switch (blend_func) {
333 case PIPE_BLEND_ADD:
334 return V_028760_OPT_COMB_ADD;
335 case PIPE_BLEND_SUBTRACT:
336 return V_028760_OPT_COMB_SUBTRACT;
337 case PIPE_BLEND_REVERSE_SUBTRACT:
338 return V_028760_OPT_COMB_REVSUBTRACT;
339 case PIPE_BLEND_MIN:
340 return V_028760_OPT_COMB_MIN;
341 case PIPE_BLEND_MAX:
342 return V_028760_OPT_COMB_MAX;
343 default:
344 return V_028760_OPT_COMB_BLEND_DISABLED;
345 }
346 }
347
348 static uint32_t si_translate_blend_opt_factor(int blend_fact, bool is_alpha)
349 {
350 switch (blend_fact) {
351 case PIPE_BLENDFACTOR_ZERO:
352 return V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_ALL;
353 case PIPE_BLENDFACTOR_ONE:
354 return V_028760_BLEND_OPT_PRESERVE_ALL_IGNORE_NONE;
355 case PIPE_BLENDFACTOR_SRC_COLOR:
356 return is_alpha ? V_028760_BLEND_OPT_PRESERVE_A1_IGNORE_A0
357 : V_028760_BLEND_OPT_PRESERVE_C1_IGNORE_C0;
358 case PIPE_BLENDFACTOR_INV_SRC_COLOR:
359 return is_alpha ? V_028760_BLEND_OPT_PRESERVE_A0_IGNORE_A1
360 : V_028760_BLEND_OPT_PRESERVE_C0_IGNORE_C1;
361 case PIPE_BLENDFACTOR_SRC_ALPHA:
362 return V_028760_BLEND_OPT_PRESERVE_A1_IGNORE_A0;
363 case PIPE_BLENDFACTOR_INV_SRC_ALPHA:
364 return V_028760_BLEND_OPT_PRESERVE_A0_IGNORE_A1;
365 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
366 return is_alpha ? V_028760_BLEND_OPT_PRESERVE_ALL_IGNORE_NONE
367 : V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_A0;
368 default:
369 return V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE;
370 }
371 }
372
373 static void si_blend_check_commutativity(struct si_screen *sscreen,
374 struct si_state_blend *blend,
375 enum pipe_blend_func func,
376 enum pipe_blendfactor src,
377 enum pipe_blendfactor dst,
378 unsigned chanmask)
379 {
380 /* Src factor is allowed when it does not depend on Dst */
381 static const uint32_t src_allowed =
382 (1u << PIPE_BLENDFACTOR_ONE) |
383 (1u << PIPE_BLENDFACTOR_SRC_COLOR) |
384 (1u << PIPE_BLENDFACTOR_SRC_ALPHA) |
385 (1u << PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE) |
386 (1u << PIPE_BLENDFACTOR_CONST_COLOR) |
387 (1u << PIPE_BLENDFACTOR_CONST_ALPHA) |
388 (1u << PIPE_BLENDFACTOR_SRC1_COLOR) |
389 (1u << PIPE_BLENDFACTOR_SRC1_ALPHA) |
390 (1u << PIPE_BLENDFACTOR_ZERO) |
391 (1u << PIPE_BLENDFACTOR_INV_SRC_COLOR) |
392 (1u << PIPE_BLENDFACTOR_INV_SRC_ALPHA) |
393 (1u << PIPE_BLENDFACTOR_INV_CONST_COLOR) |
394 (1u << PIPE_BLENDFACTOR_INV_CONST_ALPHA) |
395 (1u << PIPE_BLENDFACTOR_INV_SRC1_COLOR) |
396 (1u << PIPE_BLENDFACTOR_INV_SRC1_ALPHA);
397
398 if (dst == PIPE_BLENDFACTOR_ONE &&
399 (src_allowed & (1u << src))) {
400 /* Addition is commutative, but floating point addition isn't
401 * associative: subtle changes can be introduced via different
402 * rounding.
403 *
404 * Out-of-order is also non-deterministic, which means that
405 * this breaks OpenGL invariance requirements. So only enable
406 * out-of-order additive blending if explicitly allowed by a
407 * setting.
408 */
409 if (func == PIPE_BLEND_MAX || func == PIPE_BLEND_MIN ||
410 (func == PIPE_BLEND_ADD && sscreen->commutative_blend_add))
411 blend->commutative_4bit |= chanmask;
412 }
413 }
414
415 /**
416 * Get rid of DST in the blend factors by commuting the operands:
417 * func(src * DST, dst * 0) ---> func(src * 0, dst * SRC)
418 */
419 static void si_blend_remove_dst(unsigned *func, unsigned *src_factor,
420 unsigned *dst_factor, unsigned expected_dst,
421 unsigned replacement_src)
422 {
423 if (*src_factor == expected_dst &&
424 *dst_factor == PIPE_BLENDFACTOR_ZERO) {
425 *src_factor = PIPE_BLENDFACTOR_ZERO;
426 *dst_factor = replacement_src;
427
428 /* Commuting the operands requires reversing subtractions. */
429 if (*func == PIPE_BLEND_SUBTRACT)
430 *func = PIPE_BLEND_REVERSE_SUBTRACT;
431 else if (*func == PIPE_BLEND_REVERSE_SUBTRACT)
432 *func = PIPE_BLEND_SUBTRACT;
433 }
434 }
435
436 static bool si_blend_factor_uses_dst(unsigned factor)
437 {
438 return factor == PIPE_BLENDFACTOR_DST_COLOR ||
439 factor == PIPE_BLENDFACTOR_DST_ALPHA ||
440 factor == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE ||
441 factor == PIPE_BLENDFACTOR_INV_DST_ALPHA ||
442 factor == PIPE_BLENDFACTOR_INV_DST_COLOR;
443 }
444
445 static void *si_create_blend_state_mode(struct pipe_context *ctx,
446 const struct pipe_blend_state *state,
447 unsigned mode)
448 {
449 struct si_context *sctx = (struct si_context*)ctx;
450 struct si_state_blend *blend = CALLOC_STRUCT(si_state_blend);
451 struct si_pm4_state *pm4 = &blend->pm4;
452 uint32_t sx_mrt_blend_opt[8] = {0};
453 uint32_t color_control = 0;
454
455 if (!blend)
456 return NULL;
457
458 blend->alpha_to_coverage = state->alpha_to_coverage;
459 blend->alpha_to_one = state->alpha_to_one;
460 blend->dual_src_blend = util_blend_state_is_dual(state, 0);
461 blend->logicop_enable = state->logicop_enable;
462
463 if (state->logicop_enable) {
464 color_control |= S_028808_ROP3(state->logicop_func | (state->logicop_func << 4));
465 } else {
466 color_control |= S_028808_ROP3(0xcc);
467 }
468
469 si_pm4_set_reg(pm4, R_028B70_DB_ALPHA_TO_MASK,
470 S_028B70_ALPHA_TO_MASK_ENABLE(state->alpha_to_coverage) |
471 S_028B70_ALPHA_TO_MASK_OFFSET0(2) |
472 S_028B70_ALPHA_TO_MASK_OFFSET1(2) |
473 S_028B70_ALPHA_TO_MASK_OFFSET2(2) |
474 S_028B70_ALPHA_TO_MASK_OFFSET3(2));
475
476 if (state->alpha_to_coverage)
477 blend->need_src_alpha_4bit |= 0xf;
478
479 blend->cb_target_mask = 0;
480 blend->cb_target_enabled_4bit = 0;
481
482 for (int i = 0; i < 8; i++) {
483 /* state->rt entries > 0 only written if independent blending */
484 const int j = state->independent_blend_enable ? i : 0;
485
486 unsigned eqRGB = state->rt[j].rgb_func;
487 unsigned srcRGB = state->rt[j].rgb_src_factor;
488 unsigned dstRGB = state->rt[j].rgb_dst_factor;
489 unsigned eqA = state->rt[j].alpha_func;
490 unsigned srcA = state->rt[j].alpha_src_factor;
491 unsigned dstA = state->rt[j].alpha_dst_factor;
492
493 unsigned srcRGB_opt, dstRGB_opt, srcA_opt, dstA_opt;
494 unsigned blend_cntl = 0;
495
496 sx_mrt_blend_opt[i] =
497 S_028760_COLOR_COMB_FCN(V_028760_OPT_COMB_BLEND_DISABLED) |
498 S_028760_ALPHA_COMB_FCN(V_028760_OPT_COMB_BLEND_DISABLED);
499
500 /* Only set dual source blending for MRT0 to avoid a hang. */
501 if (i >= 1 && blend->dual_src_blend) {
502 /* Vulkan does this for dual source blending. */
503 if (i == 1)
504 blend_cntl |= S_028780_ENABLE(1);
505
506 si_pm4_set_reg(pm4, R_028780_CB_BLEND0_CONTROL + i * 4, blend_cntl);
507 continue;
508 }
509
510 /* Only addition and subtraction equations are supported with
511 * dual source blending.
512 */
513 if (blend->dual_src_blend &&
514 (eqRGB == PIPE_BLEND_MIN || eqRGB == PIPE_BLEND_MAX ||
515 eqA == PIPE_BLEND_MIN || eqA == PIPE_BLEND_MAX)) {
516 assert(!"Unsupported equation for dual source blending");
517 si_pm4_set_reg(pm4, R_028780_CB_BLEND0_CONTROL + i * 4, blend_cntl);
518 continue;
519 }
520
521 /* cb_render_state will disable unused ones */
522 blend->cb_target_mask |= (unsigned)state->rt[j].colormask << (4 * i);
523 if (state->rt[j].colormask)
524 blend->cb_target_enabled_4bit |= 0xf << (4 * i);
525
526 if (!state->rt[j].colormask || !state->rt[j].blend_enable) {
527 si_pm4_set_reg(pm4, R_028780_CB_BLEND0_CONTROL + i * 4, blend_cntl);
528 continue;
529 }
530
531 si_blend_check_commutativity(sctx->screen, blend,
532 eqRGB, srcRGB, dstRGB, 0x7 << (4 * i));
533 si_blend_check_commutativity(sctx->screen, blend,
534 eqA, srcA, dstA, 0x8 << (4 * i));
535
536 /* Blending optimizations for RB+.
537 * These transformations don't change the behavior.
538 *
539 * First, get rid of DST in the blend factors:
540 * func(src * DST, dst * 0) ---> func(src * 0, dst * SRC)
541 */
542 si_blend_remove_dst(&eqRGB, &srcRGB, &dstRGB,
543 PIPE_BLENDFACTOR_DST_COLOR,
544 PIPE_BLENDFACTOR_SRC_COLOR);
545 si_blend_remove_dst(&eqA, &srcA, &dstA,
546 PIPE_BLENDFACTOR_DST_COLOR,
547 PIPE_BLENDFACTOR_SRC_COLOR);
548 si_blend_remove_dst(&eqA, &srcA, &dstA,
549 PIPE_BLENDFACTOR_DST_ALPHA,
550 PIPE_BLENDFACTOR_SRC_ALPHA);
551
552 /* Look up the ideal settings from tables. */
553 srcRGB_opt = si_translate_blend_opt_factor(srcRGB, false);
554 dstRGB_opt = si_translate_blend_opt_factor(dstRGB, false);
555 srcA_opt = si_translate_blend_opt_factor(srcA, true);
556 dstA_opt = si_translate_blend_opt_factor(dstA, true);
557
558 /* Handle interdependencies. */
559 if (si_blend_factor_uses_dst(srcRGB))
560 dstRGB_opt = V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE;
561 if (si_blend_factor_uses_dst(srcA))
562 dstA_opt = V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE;
563
564 if (srcRGB == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE &&
565 (dstRGB == PIPE_BLENDFACTOR_ZERO ||
566 dstRGB == PIPE_BLENDFACTOR_SRC_ALPHA ||
567 dstRGB == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE))
568 dstRGB_opt = V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_A0;
569
570 /* Set the final value. */
571 sx_mrt_blend_opt[i] =
572 S_028760_COLOR_SRC_OPT(srcRGB_opt) |
573 S_028760_COLOR_DST_OPT(dstRGB_opt) |
574 S_028760_COLOR_COMB_FCN(si_translate_blend_opt_function(eqRGB)) |
575 S_028760_ALPHA_SRC_OPT(srcA_opt) |
576 S_028760_ALPHA_DST_OPT(dstA_opt) |
577 S_028760_ALPHA_COMB_FCN(si_translate_blend_opt_function(eqA));
578
579 /* Set blend state. */
580 blend_cntl |= S_028780_ENABLE(1);
581 blend_cntl |= S_028780_COLOR_COMB_FCN(si_translate_blend_function(eqRGB));
582 blend_cntl |= S_028780_COLOR_SRCBLEND(si_translate_blend_factor(srcRGB));
583 blend_cntl |= S_028780_COLOR_DESTBLEND(si_translate_blend_factor(dstRGB));
584
585 if (srcA != srcRGB || dstA != dstRGB || eqA != eqRGB) {
586 blend_cntl |= S_028780_SEPARATE_ALPHA_BLEND(1);
587 blend_cntl |= S_028780_ALPHA_COMB_FCN(si_translate_blend_function(eqA));
588 blend_cntl |= S_028780_ALPHA_SRCBLEND(si_translate_blend_factor(srcA));
589 blend_cntl |= S_028780_ALPHA_DESTBLEND(si_translate_blend_factor(dstA));
590 }
591 si_pm4_set_reg(pm4, R_028780_CB_BLEND0_CONTROL + i * 4, blend_cntl);
592
593 blend->blend_enable_4bit |= 0xfu << (i * 4);
594
595 /* This is only important for formats without alpha. */
596 if (srcRGB == PIPE_BLENDFACTOR_SRC_ALPHA ||
597 dstRGB == PIPE_BLENDFACTOR_SRC_ALPHA ||
598 srcRGB == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE ||
599 dstRGB == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE ||
600 srcRGB == PIPE_BLENDFACTOR_INV_SRC_ALPHA ||
601 dstRGB == PIPE_BLENDFACTOR_INV_SRC_ALPHA)
602 blend->need_src_alpha_4bit |= 0xfu << (i * 4);
603 }
604
605 if (blend->cb_target_mask) {
606 color_control |= S_028808_MODE(mode);
607 } else {
608 color_control |= S_028808_MODE(V_028808_CB_DISABLE);
609 }
610
611 if (sctx->screen->rbplus_allowed) {
612 /* Disable RB+ blend optimizations for dual source blending.
613 * Vulkan does this.
614 */
615 if (blend->dual_src_blend) {
616 for (int i = 0; i < 8; i++) {
617 sx_mrt_blend_opt[i] =
618 S_028760_COLOR_COMB_FCN(V_028760_OPT_COMB_NONE) |
619 S_028760_ALPHA_COMB_FCN(V_028760_OPT_COMB_NONE);
620 }
621 }
622
623 for (int i = 0; i < 8; i++)
624 si_pm4_set_reg(pm4, R_028760_SX_MRT0_BLEND_OPT + i * 4,
625 sx_mrt_blend_opt[i]);
626
627 /* RB+ doesn't work with dual source blending, logic op, and RESOLVE. */
628 if (blend->dual_src_blend || state->logicop_enable ||
629 mode == V_028808_CB_RESOLVE)
630 color_control |= S_028808_DISABLE_DUAL_QUAD(1);
631 }
632
633 si_pm4_set_reg(pm4, R_028808_CB_COLOR_CONTROL, color_control);
634 return blend;
635 }
636
637 static void *si_create_blend_state(struct pipe_context *ctx,
638 const struct pipe_blend_state *state)
639 {
640 return si_create_blend_state_mode(ctx, state, V_028808_CB_NORMAL);
641 }
642
643 static void si_bind_blend_state(struct pipe_context *ctx, void *state)
644 {
645 struct si_context *sctx = (struct si_context *)ctx;
646 struct si_state_blend *old_blend = sctx->queued.named.blend;
647 struct si_state_blend *blend = (struct si_state_blend *)state;
648
649 if (!state)
650 return;
651
652 si_pm4_bind_state(sctx, blend, state);
653
654 if (!old_blend ||
655 old_blend->cb_target_mask != blend->cb_target_mask ||
656 old_blend->dual_src_blend != blend->dual_src_blend ||
657 (old_blend->blend_enable_4bit != blend->blend_enable_4bit &&
658 sctx->framebuffer.nr_samples >= 2 &&
659 sctx->screen->dcc_msaa_allowed))
660 si_mark_atom_dirty(sctx, &sctx->atoms.s.cb_render_state);
661
662 if (!old_blend ||
663 old_blend->cb_target_mask != blend->cb_target_mask ||
664 old_blend->alpha_to_coverage != blend->alpha_to_coverage ||
665 old_blend->alpha_to_one != blend->alpha_to_one ||
666 old_blend->dual_src_blend != blend->dual_src_blend ||
667 old_blend->blend_enable_4bit != blend->blend_enable_4bit ||
668 old_blend->need_src_alpha_4bit != blend->need_src_alpha_4bit)
669 sctx->do_update_shaders = true;
670
671 if (sctx->screen->dpbb_allowed &&
672 (!old_blend ||
673 old_blend->alpha_to_coverage != blend->alpha_to_coverage ||
674 old_blend->blend_enable_4bit != blend->blend_enable_4bit ||
675 old_blend->cb_target_enabled_4bit != blend->cb_target_enabled_4bit))
676 si_mark_atom_dirty(sctx, &sctx->atoms.s.dpbb_state);
677
678 if (sctx->screen->has_out_of_order_rast &&
679 (!old_blend ||
680 (old_blend->blend_enable_4bit != blend->blend_enable_4bit ||
681 old_blend->cb_target_enabled_4bit != blend->cb_target_enabled_4bit ||
682 old_blend->commutative_4bit != blend->commutative_4bit ||
683 old_blend->logicop_enable != blend->logicop_enable)))
684 si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
685 }
686
687 static void si_delete_blend_state(struct pipe_context *ctx, void *state)
688 {
689 struct si_context *sctx = (struct si_context *)ctx;
690 si_pm4_delete_state(sctx, blend, (struct si_state_blend *)state);
691 }
692
693 static void si_set_blend_color(struct pipe_context *ctx,
694 const struct pipe_blend_color *state)
695 {
696 struct si_context *sctx = (struct si_context *)ctx;
697 static const struct pipe_blend_color zeros;
698
699 sctx->blend_color.state = *state;
700 sctx->blend_color.any_nonzeros = memcmp(state, &zeros, sizeof(*state)) != 0;
701 si_mark_atom_dirty(sctx, &sctx->atoms.s.blend_color);
702 }
703
704 static void si_emit_blend_color(struct si_context *sctx)
705 {
706 struct radeon_cmdbuf *cs = sctx->gfx_cs;
707
708 radeon_set_context_reg_seq(cs, R_028414_CB_BLEND_RED, 4);
709 radeon_emit_array(cs, (uint32_t*)sctx->blend_color.state.color, 4);
710 }
711
712 /*
713 * Clipping
714 */
715
716 static void si_set_clip_state(struct pipe_context *ctx,
717 const struct pipe_clip_state *state)
718 {
719 struct si_context *sctx = (struct si_context *)ctx;
720 struct pipe_constant_buffer cb;
721 static const struct pipe_clip_state zeros;
722
723 if (memcmp(&sctx->clip_state.state, state, sizeof(*state)) == 0)
724 return;
725
726 sctx->clip_state.state = *state;
727 sctx->clip_state.any_nonzeros = memcmp(state, &zeros, sizeof(*state)) != 0;
728 si_mark_atom_dirty(sctx, &sctx->atoms.s.clip_state);
729
730 cb.buffer = NULL;
731 cb.user_buffer = state->ucp;
732 cb.buffer_offset = 0;
733 cb.buffer_size = 4*4*8;
734 si_set_rw_buffer(sctx, SI_VS_CONST_CLIP_PLANES, &cb);
735 pipe_resource_reference(&cb.buffer, NULL);
736 }
737
738 static void si_emit_clip_state(struct si_context *sctx)
739 {
740 struct radeon_cmdbuf *cs = sctx->gfx_cs;
741
742 radeon_set_context_reg_seq(cs, R_0285BC_PA_CL_UCP_0_X, 6*4);
743 radeon_emit_array(cs, (uint32_t*)sctx->clip_state.state.ucp, 6*4);
744 }
745
746 static void si_emit_clip_regs(struct si_context *sctx)
747 {
748 struct si_shader *vs = si_get_vs_state(sctx);
749 struct si_shader_selector *vs_sel = vs->selector;
750 struct tgsi_shader_info *info = &vs_sel->info;
751 struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
752 unsigned window_space =
753 info->properties[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION];
754 unsigned clipdist_mask = vs_sel->clipdist_mask;
755 unsigned ucp_mask = clipdist_mask ? 0 : rs->clip_plane_enable & SIX_BITS;
756 unsigned culldist_mask = vs_sel->culldist_mask;
757 unsigned total_mask;
758
759 if (vs->key.opt.clip_disable) {
760 assert(!info->culldist_writemask);
761 clipdist_mask = 0;
762 culldist_mask = 0;
763 }
764 total_mask = clipdist_mask | culldist_mask;
765
766 /* Clip distances on points have no effect, so need to be implemented
767 * as cull distances. This applies for the clipvertex case as well.
768 *
769 * Setting this for primitives other than points should have no adverse
770 * effects.
771 */
772 clipdist_mask &= rs->clip_plane_enable;
773 culldist_mask |= clipdist_mask;
774
775 radeon_opt_set_context_reg(sctx, R_02881C_PA_CL_VS_OUT_CNTL,
776 SI_TRACKED_PA_CL_VS_OUT_CNTL,
777 vs_sel->pa_cl_vs_out_cntl |
778 S_02881C_VS_OUT_CCDIST0_VEC_ENA((total_mask & 0x0F) != 0) |
779 S_02881C_VS_OUT_CCDIST1_VEC_ENA((total_mask & 0xF0) != 0) |
780 clipdist_mask | (culldist_mask << 8));
781 radeon_opt_set_context_reg(sctx, R_028810_PA_CL_CLIP_CNTL,
782 SI_TRACKED_PA_CL_CLIP_CNTL,
783 rs->pa_cl_clip_cntl |
784 ucp_mask |
785 S_028810_CLIP_DISABLE(window_space));
786 }
787
788 /*
789 * inferred state between framebuffer and rasterizer
790 */
791 static void si_update_poly_offset_state(struct si_context *sctx)
792 {
793 struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
794
795 if (!rs || !rs->uses_poly_offset || !sctx->framebuffer.state.zsbuf) {
796 si_pm4_bind_state(sctx, poly_offset, NULL);
797 return;
798 }
799
800 /* Use the user format, not db_render_format, so that the polygon
801 * offset behaves as expected by applications.
802 */
803 switch (sctx->framebuffer.state.zsbuf->texture->format) {
804 case PIPE_FORMAT_Z16_UNORM:
805 si_pm4_bind_state(sctx, poly_offset, &rs->pm4_poly_offset[0]);
806 break;
807 default: /* 24-bit */
808 si_pm4_bind_state(sctx, poly_offset, &rs->pm4_poly_offset[1]);
809 break;
810 case PIPE_FORMAT_Z32_FLOAT:
811 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
812 si_pm4_bind_state(sctx, poly_offset, &rs->pm4_poly_offset[2]);
813 break;
814 }
815 }
816
817 /*
818 * Rasterizer
819 */
820
821 static uint32_t si_translate_fill(uint32_t func)
822 {
823 switch(func) {
824 case PIPE_POLYGON_MODE_FILL:
825 return V_028814_X_DRAW_TRIANGLES;
826 case PIPE_POLYGON_MODE_LINE:
827 return V_028814_X_DRAW_LINES;
828 case PIPE_POLYGON_MODE_POINT:
829 return V_028814_X_DRAW_POINTS;
830 default:
831 assert(0);
832 return V_028814_X_DRAW_POINTS;
833 }
834 }
835
836 static void *si_create_rs_state(struct pipe_context *ctx,
837 const struct pipe_rasterizer_state *state)
838 {
839 struct si_screen *sscreen = ((struct si_context *)ctx)->screen;
840 struct si_state_rasterizer *rs = CALLOC_STRUCT(si_state_rasterizer);
841 struct si_pm4_state *pm4 = &rs->pm4;
842 unsigned tmp, i;
843 float psize_min, psize_max;
844
845 if (!rs) {
846 return NULL;
847 }
848
849 rs->scissor_enable = state->scissor;
850 rs->clip_halfz = state->clip_halfz;
851 rs->two_side = state->light_twoside;
852 rs->multisample_enable = state->multisample;
853 rs->force_persample_interp = state->force_persample_interp;
854 rs->clip_plane_enable = state->clip_plane_enable;
855 rs->line_stipple_enable = state->line_stipple_enable;
856 rs->poly_stipple_enable = state->poly_stipple_enable;
857 rs->line_smooth = state->line_smooth;
858 rs->line_width = state->line_width;
859 rs->poly_smooth = state->poly_smooth;
860 rs->uses_poly_offset = state->offset_point || state->offset_line ||
861 state->offset_tri;
862 rs->clamp_fragment_color = state->clamp_fragment_color;
863 rs->clamp_vertex_color = state->clamp_vertex_color;
864 rs->flatshade = state->flatshade;
865 rs->sprite_coord_enable = state->sprite_coord_enable;
866 rs->rasterizer_discard = state->rasterizer_discard;
867 rs->pa_sc_line_stipple = state->line_stipple_enable ?
868 S_028A0C_LINE_PATTERN(state->line_stipple_pattern) |
869 S_028A0C_REPEAT_COUNT(state->line_stipple_factor) : 0;
870 rs->pa_cl_clip_cntl =
871 S_028810_DX_CLIP_SPACE_DEF(state->clip_halfz) |
872 S_028810_ZCLIP_NEAR_DISABLE(!state->depth_clip_near) |
873 S_028810_ZCLIP_FAR_DISABLE(!state->depth_clip_far) |
874 S_028810_DX_RASTERIZATION_KILL(state->rasterizer_discard) |
875 S_028810_DX_LINEAR_ATTR_CLIP_ENA(1);
876
877 si_pm4_set_reg(pm4, R_0286D4_SPI_INTERP_CONTROL_0,
878 S_0286D4_FLAT_SHADE_ENA(1) |
879 S_0286D4_PNT_SPRITE_ENA(state->point_quad_rasterization) |
880 S_0286D4_PNT_SPRITE_OVRD_X(V_0286D4_SPI_PNT_SPRITE_SEL_S) |
881 S_0286D4_PNT_SPRITE_OVRD_Y(V_0286D4_SPI_PNT_SPRITE_SEL_T) |
882 S_0286D4_PNT_SPRITE_OVRD_Z(V_0286D4_SPI_PNT_SPRITE_SEL_0) |
883 S_0286D4_PNT_SPRITE_OVRD_W(V_0286D4_SPI_PNT_SPRITE_SEL_1) |
884 S_0286D4_PNT_SPRITE_TOP_1(state->sprite_coord_mode != PIPE_SPRITE_COORD_UPPER_LEFT));
885
886 /* point size 12.4 fixed point */
887 tmp = (unsigned)(state->point_size * 8.0);
888 si_pm4_set_reg(pm4, R_028A00_PA_SU_POINT_SIZE, S_028A00_HEIGHT(tmp) | S_028A00_WIDTH(tmp));
889
890 if (state->point_size_per_vertex) {
891 psize_min = util_get_min_point_size(state);
892 psize_max = 8192;
893 } else {
894 /* Force the point size to be as if the vertex output was disabled. */
895 psize_min = state->point_size;
896 psize_max = state->point_size;
897 }
898 rs->max_point_size = psize_max;
899
900 /* Divide by two, because 0.5 = 1 pixel. */
901 si_pm4_set_reg(pm4, R_028A04_PA_SU_POINT_MINMAX,
902 S_028A04_MIN_SIZE(si_pack_float_12p4(psize_min/2)) |
903 S_028A04_MAX_SIZE(si_pack_float_12p4(psize_max/2)));
904
905 si_pm4_set_reg(pm4, R_028A08_PA_SU_LINE_CNTL,
906 S_028A08_WIDTH(si_pack_float_12p4(state->line_width/2)));
907 si_pm4_set_reg(pm4, R_028A48_PA_SC_MODE_CNTL_0,
908 S_028A48_LINE_STIPPLE_ENABLE(state->line_stipple_enable) |
909 S_028A48_MSAA_ENABLE(state->multisample ||
910 state->poly_smooth ||
911 state->line_smooth) |
912 S_028A48_VPORT_SCISSOR_ENABLE(1) |
913 S_028A48_ALTERNATE_RBS_PER_TILE(sscreen->info.chip_class >= GFX9));
914
915 si_pm4_set_reg(pm4, R_028BE4_PA_SU_VTX_CNTL,
916 S_028BE4_PIX_CENTER(state->half_pixel_center) |
917 S_028BE4_QUANT_MODE(V_028BE4_X_16_8_FIXED_POINT_1_256TH));
918
919 si_pm4_set_reg(pm4, R_028B7C_PA_SU_POLY_OFFSET_CLAMP, fui(state->offset_clamp));
920 si_pm4_set_reg(pm4, R_028814_PA_SU_SC_MODE_CNTL,
921 S_028814_PROVOKING_VTX_LAST(!state->flatshade_first) |
922 S_028814_CULL_FRONT((state->cull_face & PIPE_FACE_FRONT) ? 1 : 0) |
923 S_028814_CULL_BACK((state->cull_face & PIPE_FACE_BACK) ? 1 : 0) |
924 S_028814_FACE(!state->front_ccw) |
925 S_028814_POLY_OFFSET_FRONT_ENABLE(util_get_offset(state, state->fill_front)) |
926 S_028814_POLY_OFFSET_BACK_ENABLE(util_get_offset(state, state->fill_back)) |
927 S_028814_POLY_OFFSET_PARA_ENABLE(state->offset_point || state->offset_line) |
928 S_028814_POLY_MODE(state->fill_front != PIPE_POLYGON_MODE_FILL ||
929 state->fill_back != PIPE_POLYGON_MODE_FILL) |
930 S_028814_POLYMODE_FRONT_PTYPE(si_translate_fill(state->fill_front)) |
931 S_028814_POLYMODE_BACK_PTYPE(si_translate_fill(state->fill_back)));
932
933 if (!rs->uses_poly_offset)
934 return rs;
935
936 rs->pm4_poly_offset = CALLOC(3, sizeof(struct si_pm4_state));
937 if (!rs->pm4_poly_offset) {
938 FREE(rs);
939 return NULL;
940 }
941
942 /* Precalculate polygon offset states for 16-bit, 24-bit, and 32-bit zbuffers. */
943 for (i = 0; i < 3; i++) {
944 struct si_pm4_state *pm4 = &rs->pm4_poly_offset[i];
945 float offset_units = state->offset_units;
946 float offset_scale = state->offset_scale * 16.0f;
947 uint32_t pa_su_poly_offset_db_fmt_cntl = 0;
948
949 if (!state->offset_units_unscaled) {
950 switch (i) {
951 case 0: /* 16-bit zbuffer */
952 offset_units *= 4.0f;
953 pa_su_poly_offset_db_fmt_cntl =
954 S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-16);
955 break;
956 case 1: /* 24-bit zbuffer */
957 offset_units *= 2.0f;
958 pa_su_poly_offset_db_fmt_cntl =
959 S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-24);
960 break;
961 case 2: /* 32-bit zbuffer */
962 offset_units *= 1.0f;
963 pa_su_poly_offset_db_fmt_cntl = S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-23) |
964 S_028B78_POLY_OFFSET_DB_IS_FLOAT_FMT(1);
965 break;
966 }
967 }
968
969 si_pm4_set_reg(pm4, R_028B80_PA_SU_POLY_OFFSET_FRONT_SCALE,
970 fui(offset_scale));
971 si_pm4_set_reg(pm4, R_028B84_PA_SU_POLY_OFFSET_FRONT_OFFSET,
972 fui(offset_units));
973 si_pm4_set_reg(pm4, R_028B88_PA_SU_POLY_OFFSET_BACK_SCALE,
974 fui(offset_scale));
975 si_pm4_set_reg(pm4, R_028B8C_PA_SU_POLY_OFFSET_BACK_OFFSET,
976 fui(offset_units));
977 si_pm4_set_reg(pm4, R_028B78_PA_SU_POLY_OFFSET_DB_FMT_CNTL,
978 pa_su_poly_offset_db_fmt_cntl);
979 }
980
981 return rs;
982 }
983
984 static void si_bind_rs_state(struct pipe_context *ctx, void *state)
985 {
986 struct si_context *sctx = (struct si_context *)ctx;
987 struct si_state_rasterizer *old_rs =
988 (struct si_state_rasterizer*)sctx->queued.named.rasterizer;
989 struct si_state_rasterizer *rs = (struct si_state_rasterizer *)state;
990
991 if (!state)
992 return;
993
994 if (!old_rs || old_rs->multisample_enable != rs->multisample_enable) {
995 si_mark_atom_dirty(sctx, &sctx->atoms.s.db_render_state);
996
997 /* Update the small primitive filter workaround if necessary. */
998 if (sctx->screen->has_msaa_sample_loc_bug &&
999 sctx->framebuffer.nr_samples > 1)
1000 si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_sample_locs);
1001 }
1002
1003 sctx->current_vs_state &= C_VS_STATE_CLAMP_VERTEX_COLOR;
1004 sctx->current_vs_state |= S_VS_STATE_CLAMP_VERTEX_COLOR(rs->clamp_vertex_color);
1005
1006 si_pm4_bind_state(sctx, rasterizer, rs);
1007 si_update_poly_offset_state(sctx);
1008
1009 if (!old_rs ||
1010 old_rs->scissor_enable != rs->scissor_enable) {
1011 sctx->scissors.dirty_mask = (1 << SI_MAX_VIEWPORTS) - 1;
1012 si_mark_atom_dirty(sctx, &sctx->atoms.s.scissors);
1013 }
1014
1015 if (!old_rs ||
1016 old_rs->line_width != rs->line_width ||
1017 old_rs->max_point_size != rs->max_point_size)
1018 si_mark_atom_dirty(sctx, &sctx->atoms.s.guardband);
1019
1020 if (!old_rs ||
1021 old_rs->clip_halfz != rs->clip_halfz) {
1022 sctx->viewports.depth_range_dirty_mask = (1 << SI_MAX_VIEWPORTS) - 1;
1023 si_mark_atom_dirty(sctx, &sctx->atoms.s.viewports);
1024 }
1025
1026 if (!old_rs ||
1027 old_rs->clip_plane_enable != rs->clip_plane_enable ||
1028 old_rs->pa_cl_clip_cntl != rs->pa_cl_clip_cntl)
1029 si_mark_atom_dirty(sctx, &sctx->atoms.s.clip_regs);
1030
1031 sctx->ia_multi_vgt_param_key.u.line_stipple_enabled =
1032 rs->line_stipple_enable;
1033
1034 if (!old_rs ||
1035 old_rs->clip_plane_enable != rs->clip_plane_enable ||
1036 old_rs->rasterizer_discard != rs->rasterizer_discard ||
1037 old_rs->sprite_coord_enable != rs->sprite_coord_enable ||
1038 old_rs->flatshade != rs->flatshade ||
1039 old_rs->two_side != rs->two_side ||
1040 old_rs->multisample_enable != rs->multisample_enable ||
1041 old_rs->poly_stipple_enable != rs->poly_stipple_enable ||
1042 old_rs->poly_smooth != rs->poly_smooth ||
1043 old_rs->line_smooth != rs->line_smooth ||
1044 old_rs->clamp_fragment_color != rs->clamp_fragment_color ||
1045 old_rs->force_persample_interp != rs->force_persample_interp)
1046 sctx->do_update_shaders = true;
1047 }
1048
1049 static void si_delete_rs_state(struct pipe_context *ctx, void *state)
1050 {
1051 struct si_context *sctx = (struct si_context *)ctx;
1052 struct si_state_rasterizer *rs = (struct si_state_rasterizer *)state;
1053
1054 if (sctx->queued.named.rasterizer == state)
1055 si_pm4_bind_state(sctx, poly_offset, NULL);
1056
1057 FREE(rs->pm4_poly_offset);
1058 si_pm4_delete_state(sctx, rasterizer, rs);
1059 }
1060
1061 /*
1062 * infeered state between dsa and stencil ref
1063 */
1064 static void si_emit_stencil_ref(struct si_context *sctx)
1065 {
1066 struct radeon_cmdbuf *cs = sctx->gfx_cs;
1067 struct pipe_stencil_ref *ref = &sctx->stencil_ref.state;
1068 struct si_dsa_stencil_ref_part *dsa = &sctx->stencil_ref.dsa_part;
1069
1070 radeon_set_context_reg_seq(cs, R_028430_DB_STENCILREFMASK, 2);
1071 radeon_emit(cs, S_028430_STENCILTESTVAL(ref->ref_value[0]) |
1072 S_028430_STENCILMASK(dsa->valuemask[0]) |
1073 S_028430_STENCILWRITEMASK(dsa->writemask[0]) |
1074 S_028430_STENCILOPVAL(1));
1075 radeon_emit(cs, S_028434_STENCILTESTVAL_BF(ref->ref_value[1]) |
1076 S_028434_STENCILMASK_BF(dsa->valuemask[1]) |
1077 S_028434_STENCILWRITEMASK_BF(dsa->writemask[1]) |
1078 S_028434_STENCILOPVAL_BF(1));
1079 }
1080
1081 static void si_set_stencil_ref(struct pipe_context *ctx,
1082 const struct pipe_stencil_ref *state)
1083 {
1084 struct si_context *sctx = (struct si_context *)ctx;
1085
1086 if (memcmp(&sctx->stencil_ref.state, state, sizeof(*state)) == 0)
1087 return;
1088
1089 sctx->stencil_ref.state = *state;
1090 si_mark_atom_dirty(sctx, &sctx->atoms.s.stencil_ref);
1091 }
1092
1093
1094 /*
1095 * DSA
1096 */
1097
1098 static uint32_t si_translate_stencil_op(int s_op)
1099 {
1100 switch (s_op) {
1101 case PIPE_STENCIL_OP_KEEP:
1102 return V_02842C_STENCIL_KEEP;
1103 case PIPE_STENCIL_OP_ZERO:
1104 return V_02842C_STENCIL_ZERO;
1105 case PIPE_STENCIL_OP_REPLACE:
1106 return V_02842C_STENCIL_REPLACE_TEST;
1107 case PIPE_STENCIL_OP_INCR:
1108 return V_02842C_STENCIL_ADD_CLAMP;
1109 case PIPE_STENCIL_OP_DECR:
1110 return V_02842C_STENCIL_SUB_CLAMP;
1111 case PIPE_STENCIL_OP_INCR_WRAP:
1112 return V_02842C_STENCIL_ADD_WRAP;
1113 case PIPE_STENCIL_OP_DECR_WRAP:
1114 return V_02842C_STENCIL_SUB_WRAP;
1115 case PIPE_STENCIL_OP_INVERT:
1116 return V_02842C_STENCIL_INVERT;
1117 default:
1118 PRINT_ERR("Unknown stencil op %d", s_op);
1119 assert(0);
1120 break;
1121 }
1122 return 0;
1123 }
1124
1125 static bool si_dsa_writes_stencil(const struct pipe_stencil_state *s)
1126 {
1127 return s->enabled && s->writemask &&
1128 (s->fail_op != PIPE_STENCIL_OP_KEEP ||
1129 s->zfail_op != PIPE_STENCIL_OP_KEEP ||
1130 s->zpass_op != PIPE_STENCIL_OP_KEEP);
1131 }
1132
1133 static bool si_order_invariant_stencil_op(enum pipe_stencil_op op)
1134 {
1135 /* REPLACE is normally order invariant, except when the stencil
1136 * reference value is written by the fragment shader. Tracking this
1137 * interaction does not seem worth the effort, so be conservative. */
1138 return op != PIPE_STENCIL_OP_INCR &&
1139 op != PIPE_STENCIL_OP_DECR &&
1140 op != PIPE_STENCIL_OP_REPLACE;
1141 }
1142
1143 /* Compute whether, assuming Z writes are disabled, this stencil state is order
1144 * invariant in the sense that the set of passing fragments as well as the
1145 * final stencil buffer result does not depend on the order of fragments. */
1146 static bool si_order_invariant_stencil_state(const struct pipe_stencil_state *state)
1147 {
1148 return !state->enabled || !state->writemask ||
1149 /* The following assumes that Z writes are disabled. */
1150 (state->func == PIPE_FUNC_ALWAYS &&
1151 si_order_invariant_stencil_op(state->zpass_op) &&
1152 si_order_invariant_stencil_op(state->zfail_op)) ||
1153 (state->func == PIPE_FUNC_NEVER &&
1154 si_order_invariant_stencil_op(state->fail_op));
1155 }
1156
1157 static void *si_create_dsa_state(struct pipe_context *ctx,
1158 const struct pipe_depth_stencil_alpha_state *state)
1159 {
1160 struct si_context *sctx = (struct si_context *)ctx;
1161 struct si_state_dsa *dsa = CALLOC_STRUCT(si_state_dsa);
1162 struct si_pm4_state *pm4 = &dsa->pm4;
1163 unsigned db_depth_control;
1164 uint32_t db_stencil_control = 0;
1165
1166 if (!dsa) {
1167 return NULL;
1168 }
1169
1170 dsa->stencil_ref.valuemask[0] = state->stencil[0].valuemask;
1171 dsa->stencil_ref.valuemask[1] = state->stencil[1].valuemask;
1172 dsa->stencil_ref.writemask[0] = state->stencil[0].writemask;
1173 dsa->stencil_ref.writemask[1] = state->stencil[1].writemask;
1174
1175 db_depth_control = S_028800_Z_ENABLE(state->depth.enabled) |
1176 S_028800_Z_WRITE_ENABLE(state->depth.writemask) |
1177 S_028800_ZFUNC(state->depth.func) |
1178 S_028800_DEPTH_BOUNDS_ENABLE(state->depth.bounds_test);
1179
1180 /* stencil */
1181 if (state->stencil[0].enabled) {
1182 db_depth_control |= S_028800_STENCIL_ENABLE(1);
1183 db_depth_control |= S_028800_STENCILFUNC(state->stencil[0].func);
1184 db_stencil_control |= S_02842C_STENCILFAIL(si_translate_stencil_op(state->stencil[0].fail_op));
1185 db_stencil_control |= S_02842C_STENCILZPASS(si_translate_stencil_op(state->stencil[0].zpass_op));
1186 db_stencil_control |= S_02842C_STENCILZFAIL(si_translate_stencil_op(state->stencil[0].zfail_op));
1187
1188 if (state->stencil[1].enabled) {
1189 db_depth_control |= S_028800_BACKFACE_ENABLE(1);
1190 db_depth_control |= S_028800_STENCILFUNC_BF(state->stencil[1].func);
1191 db_stencil_control |= S_02842C_STENCILFAIL_BF(si_translate_stencil_op(state->stencil[1].fail_op));
1192 db_stencil_control |= S_02842C_STENCILZPASS_BF(si_translate_stencil_op(state->stencil[1].zpass_op));
1193 db_stencil_control |= S_02842C_STENCILZFAIL_BF(si_translate_stencil_op(state->stencil[1].zfail_op));
1194 }
1195 }
1196
1197 /* alpha */
1198 if (state->alpha.enabled) {
1199 dsa->alpha_func = state->alpha.func;
1200
1201 si_pm4_set_reg(pm4, R_00B030_SPI_SHADER_USER_DATA_PS_0 +
1202 SI_SGPR_ALPHA_REF * 4, fui(state->alpha.ref_value));
1203 } else {
1204 dsa->alpha_func = PIPE_FUNC_ALWAYS;
1205 }
1206
1207 si_pm4_set_reg(pm4, R_028800_DB_DEPTH_CONTROL, db_depth_control);
1208 if (state->stencil[0].enabled)
1209 si_pm4_set_reg(pm4, R_02842C_DB_STENCIL_CONTROL, db_stencil_control);
1210 if (state->depth.bounds_test) {
1211 si_pm4_set_reg(pm4, R_028020_DB_DEPTH_BOUNDS_MIN, fui(state->depth.bounds_min));
1212 si_pm4_set_reg(pm4, R_028024_DB_DEPTH_BOUNDS_MAX, fui(state->depth.bounds_max));
1213 }
1214
1215 dsa->depth_enabled = state->depth.enabled;
1216 dsa->depth_write_enabled = state->depth.enabled &&
1217 state->depth.writemask;
1218 dsa->stencil_enabled = state->stencil[0].enabled;
1219 dsa->stencil_write_enabled = state->stencil[0].enabled &&
1220 (si_dsa_writes_stencil(&state->stencil[0]) ||
1221 si_dsa_writes_stencil(&state->stencil[1]));
1222 dsa->db_can_write = dsa->depth_write_enabled ||
1223 dsa->stencil_write_enabled;
1224
1225 bool zfunc_is_ordered =
1226 state->depth.func == PIPE_FUNC_NEVER ||
1227 state->depth.func == PIPE_FUNC_LESS ||
1228 state->depth.func == PIPE_FUNC_LEQUAL ||
1229 state->depth.func == PIPE_FUNC_GREATER ||
1230 state->depth.func == PIPE_FUNC_GEQUAL;
1231
1232 bool nozwrite_and_order_invariant_stencil =
1233 !dsa->db_can_write ||
1234 (!dsa->depth_write_enabled &&
1235 si_order_invariant_stencil_state(&state->stencil[0]) &&
1236 si_order_invariant_stencil_state(&state->stencil[1]));
1237
1238 dsa->order_invariance[1].zs =
1239 nozwrite_and_order_invariant_stencil ||
1240 (!dsa->stencil_write_enabled && zfunc_is_ordered);
1241 dsa->order_invariance[0].zs = !dsa->depth_write_enabled || zfunc_is_ordered;
1242
1243 dsa->order_invariance[1].pass_set =
1244 nozwrite_and_order_invariant_stencil ||
1245 (!dsa->stencil_write_enabled &&
1246 (state->depth.func == PIPE_FUNC_ALWAYS ||
1247 state->depth.func == PIPE_FUNC_NEVER));
1248 dsa->order_invariance[0].pass_set =
1249 !dsa->depth_write_enabled ||
1250 (state->depth.func == PIPE_FUNC_ALWAYS ||
1251 state->depth.func == PIPE_FUNC_NEVER);
1252
1253 dsa->order_invariance[1].pass_last =
1254 sctx->screen->assume_no_z_fights &&
1255 !dsa->stencil_write_enabled &&
1256 dsa->depth_write_enabled && zfunc_is_ordered;
1257 dsa->order_invariance[0].pass_last =
1258 sctx->screen->assume_no_z_fights &&
1259 dsa->depth_write_enabled && zfunc_is_ordered;
1260
1261 return dsa;
1262 }
1263
1264 static void si_bind_dsa_state(struct pipe_context *ctx, void *state)
1265 {
1266 struct si_context *sctx = (struct si_context *)ctx;
1267 struct si_state_dsa *old_dsa = sctx->queued.named.dsa;
1268 struct si_state_dsa *dsa = state;
1269
1270 if (!state)
1271 return;
1272
1273 si_pm4_bind_state(sctx, dsa, dsa);
1274
1275 if (memcmp(&dsa->stencil_ref, &sctx->stencil_ref.dsa_part,
1276 sizeof(struct si_dsa_stencil_ref_part)) != 0) {
1277 sctx->stencil_ref.dsa_part = dsa->stencil_ref;
1278 si_mark_atom_dirty(sctx, &sctx->atoms.s.stencil_ref);
1279 }
1280
1281 if (!old_dsa || old_dsa->alpha_func != dsa->alpha_func)
1282 sctx->do_update_shaders = true;
1283
1284 if (sctx->screen->dpbb_allowed &&
1285 (!old_dsa ||
1286 (old_dsa->depth_enabled != dsa->depth_enabled ||
1287 old_dsa->stencil_enabled != dsa->stencil_enabled ||
1288 old_dsa->db_can_write != dsa->db_can_write)))
1289 si_mark_atom_dirty(sctx, &sctx->atoms.s.dpbb_state);
1290
1291 if (sctx->screen->has_out_of_order_rast &&
1292 (!old_dsa ||
1293 memcmp(old_dsa->order_invariance, dsa->order_invariance,
1294 sizeof(old_dsa->order_invariance))))
1295 si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
1296 }
1297
1298 static void si_delete_dsa_state(struct pipe_context *ctx, void *state)
1299 {
1300 struct si_context *sctx = (struct si_context *)ctx;
1301 si_pm4_delete_state(sctx, dsa, (struct si_state_dsa *)state);
1302 }
1303
1304 static void *si_create_db_flush_dsa(struct si_context *sctx)
1305 {
1306 struct pipe_depth_stencil_alpha_state dsa = {};
1307
1308 return sctx->b.create_depth_stencil_alpha_state(&sctx->b, &dsa);
1309 }
1310
1311 /* DB RENDER STATE */
1312
1313 static void si_set_active_query_state(struct pipe_context *ctx, boolean enable)
1314 {
1315 struct si_context *sctx = (struct si_context*)ctx;
1316
1317 /* Pipeline stat & streamout queries. */
1318 if (enable) {
1319 sctx->flags &= ~SI_CONTEXT_STOP_PIPELINE_STATS;
1320 sctx->flags |= SI_CONTEXT_START_PIPELINE_STATS;
1321 } else {
1322 sctx->flags &= ~SI_CONTEXT_START_PIPELINE_STATS;
1323 sctx->flags |= SI_CONTEXT_STOP_PIPELINE_STATS;
1324 }
1325
1326 /* Occlusion queries. */
1327 if (sctx->occlusion_queries_disabled != !enable) {
1328 sctx->occlusion_queries_disabled = !enable;
1329 si_mark_atom_dirty(sctx, &sctx->atoms.s.db_render_state);
1330 }
1331 }
1332
1333 void si_set_occlusion_query_state(struct si_context *sctx,
1334 bool old_perfect_enable)
1335 {
1336 si_mark_atom_dirty(sctx, &sctx->atoms.s.db_render_state);
1337
1338 bool perfect_enable = sctx->num_perfect_occlusion_queries != 0;
1339
1340 if (perfect_enable != old_perfect_enable)
1341 si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
1342 }
1343
1344 void si_save_qbo_state(struct si_context *sctx, struct si_qbo_state *st)
1345 {
1346 st->saved_compute = sctx->cs_shader_state.program;
1347
1348 si_get_pipe_constant_buffer(sctx, PIPE_SHADER_COMPUTE, 0, &st->saved_const0);
1349 si_get_shader_buffers(sctx, PIPE_SHADER_COMPUTE, 0, 3, st->saved_ssbo);
1350 }
1351
1352 static void si_emit_db_render_state(struct si_context *sctx)
1353 {
1354 struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
1355 unsigned db_shader_control, db_render_control, db_count_control;
1356
1357 /* DB_RENDER_CONTROL */
1358 if (sctx->dbcb_depth_copy_enabled ||
1359 sctx->dbcb_stencil_copy_enabled) {
1360 db_render_control =
1361 S_028000_DEPTH_COPY(sctx->dbcb_depth_copy_enabled) |
1362 S_028000_STENCIL_COPY(sctx->dbcb_stencil_copy_enabled) |
1363 S_028000_COPY_CENTROID(1) |
1364 S_028000_COPY_SAMPLE(sctx->dbcb_copy_sample);
1365 } else if (sctx->db_flush_depth_inplace || sctx->db_flush_stencil_inplace) {
1366 db_render_control =
1367 S_028000_DEPTH_COMPRESS_DISABLE(sctx->db_flush_depth_inplace) |
1368 S_028000_STENCIL_COMPRESS_DISABLE(sctx->db_flush_stencil_inplace);
1369 } else {
1370 db_render_control =
1371 S_028000_DEPTH_CLEAR_ENABLE(sctx->db_depth_clear) |
1372 S_028000_STENCIL_CLEAR_ENABLE(sctx->db_stencil_clear);
1373 }
1374
1375 /* DB_COUNT_CONTROL (occlusion queries) */
1376 if (sctx->num_occlusion_queries > 0 &&
1377 !sctx->occlusion_queries_disabled) {
1378 bool perfect = sctx->num_perfect_occlusion_queries > 0;
1379
1380 if (sctx->chip_class >= CIK) {
1381 unsigned log_sample_rate = sctx->framebuffer.log_samples;
1382
1383 /* Stoney doesn't increment occlusion query counters
1384 * if the sample rate is 16x. Use 8x sample rate instead.
1385 */
1386 if (sctx->family == CHIP_STONEY)
1387 log_sample_rate = MIN2(log_sample_rate, 3);
1388
1389 db_count_control =
1390 S_028004_PERFECT_ZPASS_COUNTS(perfect) |
1391 S_028004_SAMPLE_RATE(log_sample_rate) |
1392 S_028004_ZPASS_ENABLE(1) |
1393 S_028004_SLICE_EVEN_ENABLE(1) |
1394 S_028004_SLICE_ODD_ENABLE(1);
1395 } else {
1396 db_count_control =
1397 S_028004_PERFECT_ZPASS_COUNTS(perfect) |
1398 S_028004_SAMPLE_RATE(sctx->framebuffer.log_samples);
1399 }
1400 } else {
1401 /* Disable occlusion queries. */
1402 if (sctx->chip_class >= CIK) {
1403 db_count_control = 0;
1404 } else {
1405 db_count_control = S_028004_ZPASS_INCREMENT_DISABLE(1);
1406 }
1407 }
1408
1409 radeon_opt_set_context_reg2(sctx, R_028000_DB_RENDER_CONTROL,
1410 SI_TRACKED_DB_RENDER_CONTROL, db_render_control,
1411 db_count_control);
1412
1413 /* DB_RENDER_OVERRIDE2 */
1414 radeon_opt_set_context_reg(sctx, R_028010_DB_RENDER_OVERRIDE2,
1415 SI_TRACKED_DB_RENDER_OVERRIDE2,
1416 S_028010_DISABLE_ZMASK_EXPCLEAR_OPTIMIZATION(sctx->db_depth_disable_expclear) |
1417 S_028010_DISABLE_SMEM_EXPCLEAR_OPTIMIZATION(sctx->db_stencil_disable_expclear) |
1418 S_028010_DECOMPRESS_Z_ON_FLUSH(sctx->framebuffer.nr_samples >= 4));
1419
1420 db_shader_control = sctx->ps_db_shader_control;
1421
1422 /* Bug workaround for smoothing (overrasterization) on SI. */
1423 if (sctx->chip_class == SI && sctx->smoothing_enabled) {
1424 db_shader_control &= C_02880C_Z_ORDER;
1425 db_shader_control |= S_02880C_Z_ORDER(V_02880C_LATE_Z);
1426 }
1427
1428 /* Disable the gl_SampleMask fragment shader output if MSAA is disabled. */
1429 if (!rs->multisample_enable)
1430 db_shader_control &= C_02880C_MASK_EXPORT_ENABLE;
1431
1432 if (sctx->screen->has_rbplus &&
1433 !sctx->screen->rbplus_allowed)
1434 db_shader_control |= S_02880C_DUAL_QUAD_DISABLE(1);
1435
1436 radeon_opt_set_context_reg(sctx, R_02880C_DB_SHADER_CONTROL,
1437 SI_TRACKED_DB_SHADER_CONTROL, db_shader_control);
1438 }
1439
1440 /*
1441 * format translation
1442 */
1443 static uint32_t si_translate_colorformat(enum pipe_format format)
1444 {
1445 const struct util_format_description *desc = util_format_description(format);
1446 if (!desc)
1447 return V_028C70_COLOR_INVALID;
1448
1449 #define HAS_SIZE(x,y,z,w) \
1450 (desc->channel[0].size == (x) && desc->channel[1].size == (y) && \
1451 desc->channel[2].size == (z) && desc->channel[3].size == (w))
1452
1453 if (format == PIPE_FORMAT_R11G11B10_FLOAT) /* isn't plain */
1454 return V_028C70_COLOR_10_11_11;
1455
1456 if (desc->layout != UTIL_FORMAT_LAYOUT_PLAIN)
1457 return V_028C70_COLOR_INVALID;
1458
1459 /* hw cannot support mixed formats (except depth/stencil, since
1460 * stencil is not written to). */
1461 if (desc->is_mixed && desc->colorspace != UTIL_FORMAT_COLORSPACE_ZS)
1462 return V_028C70_COLOR_INVALID;
1463
1464 switch (desc->nr_channels) {
1465 case 1:
1466 switch (desc->channel[0].size) {
1467 case 8:
1468 return V_028C70_COLOR_8;
1469 case 16:
1470 return V_028C70_COLOR_16;
1471 case 32:
1472 return V_028C70_COLOR_32;
1473 }
1474 break;
1475 case 2:
1476 if (desc->channel[0].size == desc->channel[1].size) {
1477 switch (desc->channel[0].size) {
1478 case 8:
1479 return V_028C70_COLOR_8_8;
1480 case 16:
1481 return V_028C70_COLOR_16_16;
1482 case 32:
1483 return V_028C70_COLOR_32_32;
1484 }
1485 } else if (HAS_SIZE(8,24,0,0)) {
1486 return V_028C70_COLOR_24_8;
1487 } else if (HAS_SIZE(24,8,0,0)) {
1488 return V_028C70_COLOR_8_24;
1489 }
1490 break;
1491 case 3:
1492 if (HAS_SIZE(5,6,5,0)) {
1493 return V_028C70_COLOR_5_6_5;
1494 } else if (HAS_SIZE(32,8,24,0)) {
1495 return V_028C70_COLOR_X24_8_32_FLOAT;
1496 }
1497 break;
1498 case 4:
1499 if (desc->channel[0].size == desc->channel[1].size &&
1500 desc->channel[0].size == desc->channel[2].size &&
1501 desc->channel[0].size == desc->channel[3].size) {
1502 switch (desc->channel[0].size) {
1503 case 4:
1504 return V_028C70_COLOR_4_4_4_4;
1505 case 8:
1506 return V_028C70_COLOR_8_8_8_8;
1507 case 16:
1508 return V_028C70_COLOR_16_16_16_16;
1509 case 32:
1510 return V_028C70_COLOR_32_32_32_32;
1511 }
1512 } else if (HAS_SIZE(5,5,5,1)) {
1513 return V_028C70_COLOR_1_5_5_5;
1514 } else if (HAS_SIZE(1,5,5,5)) {
1515 return V_028C70_COLOR_5_5_5_1;
1516 } else if (HAS_SIZE(10,10,10,2)) {
1517 return V_028C70_COLOR_2_10_10_10;
1518 }
1519 break;
1520 }
1521 return V_028C70_COLOR_INVALID;
1522 }
1523
1524 static uint32_t si_colorformat_endian_swap(uint32_t colorformat)
1525 {
1526 if (SI_BIG_ENDIAN) {
1527 switch(colorformat) {
1528 /* 8-bit buffers. */
1529 case V_028C70_COLOR_8:
1530 return V_028C70_ENDIAN_NONE;
1531
1532 /* 16-bit buffers. */
1533 case V_028C70_COLOR_5_6_5:
1534 case V_028C70_COLOR_1_5_5_5:
1535 case V_028C70_COLOR_4_4_4_4:
1536 case V_028C70_COLOR_16:
1537 case V_028C70_COLOR_8_8:
1538 return V_028C70_ENDIAN_8IN16;
1539
1540 /* 32-bit buffers. */
1541 case V_028C70_COLOR_8_8_8_8:
1542 case V_028C70_COLOR_2_10_10_10:
1543 case V_028C70_COLOR_8_24:
1544 case V_028C70_COLOR_24_8:
1545 case V_028C70_COLOR_16_16:
1546 return V_028C70_ENDIAN_8IN32;
1547
1548 /* 64-bit buffers. */
1549 case V_028C70_COLOR_16_16_16_16:
1550 return V_028C70_ENDIAN_8IN16;
1551
1552 case V_028C70_COLOR_32_32:
1553 return V_028C70_ENDIAN_8IN32;
1554
1555 /* 128-bit buffers. */
1556 case V_028C70_COLOR_32_32_32_32:
1557 return V_028C70_ENDIAN_8IN32;
1558 default:
1559 return V_028C70_ENDIAN_NONE; /* Unsupported. */
1560 }
1561 } else {
1562 return V_028C70_ENDIAN_NONE;
1563 }
1564 }
1565
1566 static uint32_t si_translate_dbformat(enum pipe_format format)
1567 {
1568 switch (format) {
1569 case PIPE_FORMAT_Z16_UNORM:
1570 return V_028040_Z_16;
1571 case PIPE_FORMAT_S8_UINT_Z24_UNORM:
1572 case PIPE_FORMAT_X8Z24_UNORM:
1573 case PIPE_FORMAT_Z24X8_UNORM:
1574 case PIPE_FORMAT_Z24_UNORM_S8_UINT:
1575 return V_028040_Z_24; /* deprecated on SI */
1576 case PIPE_FORMAT_Z32_FLOAT:
1577 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
1578 return V_028040_Z_32_FLOAT;
1579 default:
1580 return V_028040_Z_INVALID;
1581 }
1582 }
1583
1584 /*
1585 * Texture translation
1586 */
1587
1588 static uint32_t si_translate_texformat(struct pipe_screen *screen,
1589 enum pipe_format format,
1590 const struct util_format_description *desc,
1591 int first_non_void)
1592 {
1593 struct si_screen *sscreen = (struct si_screen*)screen;
1594 bool uniform = true;
1595 int i;
1596
1597 /* Colorspace (return non-RGB formats directly). */
1598 switch (desc->colorspace) {
1599 /* Depth stencil formats */
1600 case UTIL_FORMAT_COLORSPACE_ZS:
1601 switch (format) {
1602 case PIPE_FORMAT_Z16_UNORM:
1603 return V_008F14_IMG_DATA_FORMAT_16;
1604 case PIPE_FORMAT_X24S8_UINT:
1605 case PIPE_FORMAT_S8X24_UINT:
1606 /*
1607 * Implemented as an 8_8_8_8 data format to fix texture
1608 * gathers in stencil sampling. This affects at least
1609 * GL45-CTS.texture_cube_map_array.sampling on VI.
1610 */
1611 if (sscreen->info.chip_class <= VI)
1612 return V_008F14_IMG_DATA_FORMAT_8_8_8_8;
1613
1614 if (format == PIPE_FORMAT_X24S8_UINT)
1615 return V_008F14_IMG_DATA_FORMAT_8_24;
1616 else
1617 return V_008F14_IMG_DATA_FORMAT_24_8;
1618 case PIPE_FORMAT_Z24X8_UNORM:
1619 case PIPE_FORMAT_Z24_UNORM_S8_UINT:
1620 return V_008F14_IMG_DATA_FORMAT_8_24;
1621 case PIPE_FORMAT_X8Z24_UNORM:
1622 case PIPE_FORMAT_S8_UINT_Z24_UNORM:
1623 return V_008F14_IMG_DATA_FORMAT_24_8;
1624 case PIPE_FORMAT_S8_UINT:
1625 return V_008F14_IMG_DATA_FORMAT_8;
1626 case PIPE_FORMAT_Z32_FLOAT:
1627 return V_008F14_IMG_DATA_FORMAT_32;
1628 case PIPE_FORMAT_X32_S8X24_UINT:
1629 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
1630 return V_008F14_IMG_DATA_FORMAT_X24_8_32;
1631 default:
1632 goto out_unknown;
1633 }
1634
1635 case UTIL_FORMAT_COLORSPACE_YUV:
1636 goto out_unknown; /* TODO */
1637
1638 case UTIL_FORMAT_COLORSPACE_SRGB:
1639 if (desc->nr_channels != 4 && desc->nr_channels != 1)
1640 goto out_unknown;
1641 break;
1642
1643 default:
1644 break;
1645 }
1646
1647 if (desc->layout == UTIL_FORMAT_LAYOUT_RGTC) {
1648 if (!sscreen->info.has_format_bc1_through_bc7)
1649 goto out_unknown;
1650
1651 switch (format) {
1652 case PIPE_FORMAT_RGTC1_SNORM:
1653 case PIPE_FORMAT_LATC1_SNORM:
1654 case PIPE_FORMAT_RGTC1_UNORM:
1655 case PIPE_FORMAT_LATC1_UNORM:
1656 return V_008F14_IMG_DATA_FORMAT_BC4;
1657 case PIPE_FORMAT_RGTC2_SNORM:
1658 case PIPE_FORMAT_LATC2_SNORM:
1659 case PIPE_FORMAT_RGTC2_UNORM:
1660 case PIPE_FORMAT_LATC2_UNORM:
1661 return V_008F14_IMG_DATA_FORMAT_BC5;
1662 default:
1663 goto out_unknown;
1664 }
1665 }
1666
1667 if (desc->layout == UTIL_FORMAT_LAYOUT_ETC &&
1668 (sscreen->info.family == CHIP_STONEY ||
1669 sscreen->info.family == CHIP_VEGA10 ||
1670 sscreen->info.family == CHIP_RAVEN)) {
1671 switch (format) {
1672 case PIPE_FORMAT_ETC1_RGB8:
1673 case PIPE_FORMAT_ETC2_RGB8:
1674 case PIPE_FORMAT_ETC2_SRGB8:
1675 return V_008F14_IMG_DATA_FORMAT_ETC2_RGB;
1676 case PIPE_FORMAT_ETC2_RGB8A1:
1677 case PIPE_FORMAT_ETC2_SRGB8A1:
1678 return V_008F14_IMG_DATA_FORMAT_ETC2_RGBA1;
1679 case PIPE_FORMAT_ETC2_RGBA8:
1680 case PIPE_FORMAT_ETC2_SRGBA8:
1681 return V_008F14_IMG_DATA_FORMAT_ETC2_RGBA;
1682 case PIPE_FORMAT_ETC2_R11_UNORM:
1683 case PIPE_FORMAT_ETC2_R11_SNORM:
1684 return V_008F14_IMG_DATA_FORMAT_ETC2_R;
1685 case PIPE_FORMAT_ETC2_RG11_UNORM:
1686 case PIPE_FORMAT_ETC2_RG11_SNORM:
1687 return V_008F14_IMG_DATA_FORMAT_ETC2_RG;
1688 default:
1689 goto out_unknown;
1690 }
1691 }
1692
1693 if (desc->layout == UTIL_FORMAT_LAYOUT_BPTC) {
1694 if (!sscreen->info.has_format_bc1_through_bc7)
1695 goto out_unknown;
1696
1697 switch (format) {
1698 case PIPE_FORMAT_BPTC_RGBA_UNORM:
1699 case PIPE_FORMAT_BPTC_SRGBA:
1700 return V_008F14_IMG_DATA_FORMAT_BC7;
1701 case PIPE_FORMAT_BPTC_RGB_FLOAT:
1702 case PIPE_FORMAT_BPTC_RGB_UFLOAT:
1703 return V_008F14_IMG_DATA_FORMAT_BC6;
1704 default:
1705 goto out_unknown;
1706 }
1707 }
1708
1709 if (desc->layout == UTIL_FORMAT_LAYOUT_SUBSAMPLED) {
1710 switch (format) {
1711 case PIPE_FORMAT_R8G8_B8G8_UNORM:
1712 case PIPE_FORMAT_G8R8_B8R8_UNORM:
1713 return V_008F14_IMG_DATA_FORMAT_GB_GR;
1714 case PIPE_FORMAT_G8R8_G8B8_UNORM:
1715 case PIPE_FORMAT_R8G8_R8B8_UNORM:
1716 return V_008F14_IMG_DATA_FORMAT_BG_RG;
1717 default:
1718 goto out_unknown;
1719 }
1720 }
1721
1722 if (desc->layout == UTIL_FORMAT_LAYOUT_S3TC) {
1723 if (!sscreen->info.has_format_bc1_through_bc7)
1724 goto out_unknown;
1725
1726 switch (format) {
1727 case PIPE_FORMAT_DXT1_RGB:
1728 case PIPE_FORMAT_DXT1_RGBA:
1729 case PIPE_FORMAT_DXT1_SRGB:
1730 case PIPE_FORMAT_DXT1_SRGBA:
1731 return V_008F14_IMG_DATA_FORMAT_BC1;
1732 case PIPE_FORMAT_DXT3_RGBA:
1733 case PIPE_FORMAT_DXT3_SRGBA:
1734 return V_008F14_IMG_DATA_FORMAT_BC2;
1735 case PIPE_FORMAT_DXT5_RGBA:
1736 case PIPE_FORMAT_DXT5_SRGBA:
1737 return V_008F14_IMG_DATA_FORMAT_BC3;
1738 default:
1739 goto out_unknown;
1740 }
1741 }
1742
1743 if (format == PIPE_FORMAT_R9G9B9E5_FLOAT) {
1744 return V_008F14_IMG_DATA_FORMAT_5_9_9_9;
1745 } else if (format == PIPE_FORMAT_R11G11B10_FLOAT) {
1746 return V_008F14_IMG_DATA_FORMAT_10_11_11;
1747 }
1748
1749 /* R8G8Bx_SNORM - TODO CxV8U8 */
1750
1751 /* hw cannot support mixed formats (except depth/stencil, since only
1752 * depth is read).*/
1753 if (desc->is_mixed && desc->colorspace != UTIL_FORMAT_COLORSPACE_ZS)
1754 goto out_unknown;
1755
1756 /* See whether the components are of the same size. */
1757 for (i = 1; i < desc->nr_channels; i++) {
1758 uniform = uniform && desc->channel[0].size == desc->channel[i].size;
1759 }
1760
1761 /* Non-uniform formats. */
1762 if (!uniform) {
1763 switch(desc->nr_channels) {
1764 case 3:
1765 if (desc->channel[0].size == 5 &&
1766 desc->channel[1].size == 6 &&
1767 desc->channel[2].size == 5) {
1768 return V_008F14_IMG_DATA_FORMAT_5_6_5;
1769 }
1770 goto out_unknown;
1771 case 4:
1772 if (desc->channel[0].size == 5 &&
1773 desc->channel[1].size == 5 &&
1774 desc->channel[2].size == 5 &&
1775 desc->channel[3].size == 1) {
1776 return V_008F14_IMG_DATA_FORMAT_1_5_5_5;
1777 }
1778 if (desc->channel[0].size == 1 &&
1779 desc->channel[1].size == 5 &&
1780 desc->channel[2].size == 5 &&
1781 desc->channel[3].size == 5) {
1782 return V_008F14_IMG_DATA_FORMAT_5_5_5_1;
1783 }
1784 if (desc->channel[0].size == 10 &&
1785 desc->channel[1].size == 10 &&
1786 desc->channel[2].size == 10 &&
1787 desc->channel[3].size == 2) {
1788 return V_008F14_IMG_DATA_FORMAT_2_10_10_10;
1789 }
1790 goto out_unknown;
1791 }
1792 goto out_unknown;
1793 }
1794
1795 if (first_non_void < 0 || first_non_void > 3)
1796 goto out_unknown;
1797
1798 /* uniform formats */
1799 switch (desc->channel[first_non_void].size) {
1800 case 4:
1801 switch (desc->nr_channels) {
1802 #if 0 /* Not supported for render targets */
1803 case 2:
1804 return V_008F14_IMG_DATA_FORMAT_4_4;
1805 #endif
1806 case 4:
1807 return V_008F14_IMG_DATA_FORMAT_4_4_4_4;
1808 }
1809 break;
1810 case 8:
1811 switch (desc->nr_channels) {
1812 case 1:
1813 return V_008F14_IMG_DATA_FORMAT_8;
1814 case 2:
1815 return V_008F14_IMG_DATA_FORMAT_8_8;
1816 case 4:
1817 return V_008F14_IMG_DATA_FORMAT_8_8_8_8;
1818 }
1819 break;
1820 case 16:
1821 switch (desc->nr_channels) {
1822 case 1:
1823 return V_008F14_IMG_DATA_FORMAT_16;
1824 case 2:
1825 return V_008F14_IMG_DATA_FORMAT_16_16;
1826 case 4:
1827 return V_008F14_IMG_DATA_FORMAT_16_16_16_16;
1828 }
1829 break;
1830 case 32:
1831 switch (desc->nr_channels) {
1832 case 1:
1833 return V_008F14_IMG_DATA_FORMAT_32;
1834 case 2:
1835 return V_008F14_IMG_DATA_FORMAT_32_32;
1836 #if 0 /* Not supported for render targets */
1837 case 3:
1838 return V_008F14_IMG_DATA_FORMAT_32_32_32;
1839 #endif
1840 case 4:
1841 return V_008F14_IMG_DATA_FORMAT_32_32_32_32;
1842 }
1843 }
1844
1845 out_unknown:
1846 return ~0;
1847 }
1848
1849 static unsigned si_tex_wrap(unsigned wrap)
1850 {
1851 switch (wrap) {
1852 default:
1853 case PIPE_TEX_WRAP_REPEAT:
1854 return V_008F30_SQ_TEX_WRAP;
1855 case PIPE_TEX_WRAP_CLAMP:
1856 return V_008F30_SQ_TEX_CLAMP_HALF_BORDER;
1857 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
1858 return V_008F30_SQ_TEX_CLAMP_LAST_TEXEL;
1859 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
1860 return V_008F30_SQ_TEX_CLAMP_BORDER;
1861 case PIPE_TEX_WRAP_MIRROR_REPEAT:
1862 return V_008F30_SQ_TEX_MIRROR;
1863 case PIPE_TEX_WRAP_MIRROR_CLAMP:
1864 return V_008F30_SQ_TEX_MIRROR_ONCE_HALF_BORDER;
1865 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
1866 return V_008F30_SQ_TEX_MIRROR_ONCE_LAST_TEXEL;
1867 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
1868 return V_008F30_SQ_TEX_MIRROR_ONCE_BORDER;
1869 }
1870 }
1871
1872 static unsigned si_tex_mipfilter(unsigned filter)
1873 {
1874 switch (filter) {
1875 case PIPE_TEX_MIPFILTER_NEAREST:
1876 return V_008F38_SQ_TEX_Z_FILTER_POINT;
1877 case PIPE_TEX_MIPFILTER_LINEAR:
1878 return V_008F38_SQ_TEX_Z_FILTER_LINEAR;
1879 default:
1880 case PIPE_TEX_MIPFILTER_NONE:
1881 return V_008F38_SQ_TEX_Z_FILTER_NONE;
1882 }
1883 }
1884
1885 static unsigned si_tex_compare(unsigned compare)
1886 {
1887 switch (compare) {
1888 default:
1889 case PIPE_FUNC_NEVER:
1890 return V_008F30_SQ_TEX_DEPTH_COMPARE_NEVER;
1891 case PIPE_FUNC_LESS:
1892 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESS;
1893 case PIPE_FUNC_EQUAL:
1894 return V_008F30_SQ_TEX_DEPTH_COMPARE_EQUAL;
1895 case PIPE_FUNC_LEQUAL:
1896 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESSEQUAL;
1897 case PIPE_FUNC_GREATER:
1898 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATER;
1899 case PIPE_FUNC_NOTEQUAL:
1900 return V_008F30_SQ_TEX_DEPTH_COMPARE_NOTEQUAL;
1901 case PIPE_FUNC_GEQUAL:
1902 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATEREQUAL;
1903 case PIPE_FUNC_ALWAYS:
1904 return V_008F30_SQ_TEX_DEPTH_COMPARE_ALWAYS;
1905 }
1906 }
1907
1908 static unsigned si_tex_dim(struct si_screen *sscreen, struct si_texture *tex,
1909 unsigned view_target, unsigned nr_samples)
1910 {
1911 unsigned res_target = tex->buffer.b.b.target;
1912
1913 if (view_target == PIPE_TEXTURE_CUBE ||
1914 view_target == PIPE_TEXTURE_CUBE_ARRAY)
1915 res_target = view_target;
1916 /* If interpreting cubemaps as something else, set 2D_ARRAY. */
1917 else if (res_target == PIPE_TEXTURE_CUBE ||
1918 res_target == PIPE_TEXTURE_CUBE_ARRAY)
1919 res_target = PIPE_TEXTURE_2D_ARRAY;
1920
1921 /* GFX9 allocates 1D textures as 2D. */
1922 if ((res_target == PIPE_TEXTURE_1D ||
1923 res_target == PIPE_TEXTURE_1D_ARRAY) &&
1924 sscreen->info.chip_class >= GFX9 &&
1925 tex->surface.u.gfx9.resource_type == RADEON_RESOURCE_2D) {
1926 if (res_target == PIPE_TEXTURE_1D)
1927 res_target = PIPE_TEXTURE_2D;
1928 else
1929 res_target = PIPE_TEXTURE_2D_ARRAY;
1930 }
1931
1932 switch (res_target) {
1933 default:
1934 case PIPE_TEXTURE_1D:
1935 return V_008F1C_SQ_RSRC_IMG_1D;
1936 case PIPE_TEXTURE_1D_ARRAY:
1937 return V_008F1C_SQ_RSRC_IMG_1D_ARRAY;
1938 case PIPE_TEXTURE_2D:
1939 case PIPE_TEXTURE_RECT:
1940 return nr_samples > 1 ? V_008F1C_SQ_RSRC_IMG_2D_MSAA :
1941 V_008F1C_SQ_RSRC_IMG_2D;
1942 case PIPE_TEXTURE_2D_ARRAY:
1943 return nr_samples > 1 ? V_008F1C_SQ_RSRC_IMG_2D_MSAA_ARRAY :
1944 V_008F1C_SQ_RSRC_IMG_2D_ARRAY;
1945 case PIPE_TEXTURE_3D:
1946 return V_008F1C_SQ_RSRC_IMG_3D;
1947 case PIPE_TEXTURE_CUBE:
1948 case PIPE_TEXTURE_CUBE_ARRAY:
1949 return V_008F1C_SQ_RSRC_IMG_CUBE;
1950 }
1951 }
1952
1953 /*
1954 * Format support testing
1955 */
1956
1957 static bool si_is_sampler_format_supported(struct pipe_screen *screen, enum pipe_format format)
1958 {
1959 const struct util_format_description *desc = util_format_description(format);
1960 if (!desc)
1961 return false;
1962
1963 return si_translate_texformat(screen, format, desc,
1964 util_format_get_first_non_void_channel(format)) != ~0U;
1965 }
1966
1967 static uint32_t si_translate_buffer_dataformat(struct pipe_screen *screen,
1968 const struct util_format_description *desc,
1969 int first_non_void)
1970 {
1971 int i;
1972
1973 if (desc->format == PIPE_FORMAT_R11G11B10_FLOAT)
1974 return V_008F0C_BUF_DATA_FORMAT_10_11_11;
1975
1976 assert(first_non_void >= 0);
1977
1978 if (desc->nr_channels == 4 &&
1979 desc->channel[0].size == 10 &&
1980 desc->channel[1].size == 10 &&
1981 desc->channel[2].size == 10 &&
1982 desc->channel[3].size == 2)
1983 return V_008F0C_BUF_DATA_FORMAT_2_10_10_10;
1984
1985 /* See whether the components are of the same size. */
1986 for (i = 0; i < desc->nr_channels; i++) {
1987 if (desc->channel[first_non_void].size != desc->channel[i].size)
1988 return V_008F0C_BUF_DATA_FORMAT_INVALID;
1989 }
1990
1991 switch (desc->channel[first_non_void].size) {
1992 case 8:
1993 switch (desc->nr_channels) {
1994 case 1:
1995 case 3: /* 3 loads */
1996 return V_008F0C_BUF_DATA_FORMAT_8;
1997 case 2:
1998 return V_008F0C_BUF_DATA_FORMAT_8_8;
1999 case 4:
2000 return V_008F0C_BUF_DATA_FORMAT_8_8_8_8;
2001 }
2002 break;
2003 case 16:
2004 switch (desc->nr_channels) {
2005 case 1:
2006 case 3: /* 3 loads */
2007 return V_008F0C_BUF_DATA_FORMAT_16;
2008 case 2:
2009 return V_008F0C_BUF_DATA_FORMAT_16_16;
2010 case 4:
2011 return V_008F0C_BUF_DATA_FORMAT_16_16_16_16;
2012 }
2013 break;
2014 case 32:
2015 switch (desc->nr_channels) {
2016 case 1:
2017 return V_008F0C_BUF_DATA_FORMAT_32;
2018 case 2:
2019 return V_008F0C_BUF_DATA_FORMAT_32_32;
2020 case 3:
2021 return V_008F0C_BUF_DATA_FORMAT_32_32_32;
2022 case 4:
2023 return V_008F0C_BUF_DATA_FORMAT_32_32_32_32;
2024 }
2025 break;
2026 case 64:
2027 /* Legacy double formats. */
2028 switch (desc->nr_channels) {
2029 case 1: /* 1 load */
2030 return V_008F0C_BUF_DATA_FORMAT_32_32;
2031 case 2: /* 1 load */
2032 return V_008F0C_BUF_DATA_FORMAT_32_32_32_32;
2033 case 3: /* 3 loads */
2034 return V_008F0C_BUF_DATA_FORMAT_32_32;
2035 case 4: /* 2 loads */
2036 return V_008F0C_BUF_DATA_FORMAT_32_32_32_32;
2037 }
2038 break;
2039 }
2040
2041 return V_008F0C_BUF_DATA_FORMAT_INVALID;
2042 }
2043
2044 static uint32_t si_translate_buffer_numformat(struct pipe_screen *screen,
2045 const struct util_format_description *desc,
2046 int first_non_void)
2047 {
2048 if (desc->format == PIPE_FORMAT_R11G11B10_FLOAT)
2049 return V_008F0C_BUF_NUM_FORMAT_FLOAT;
2050
2051 assert(first_non_void >= 0);
2052
2053 switch (desc->channel[first_non_void].type) {
2054 case UTIL_FORMAT_TYPE_SIGNED:
2055 case UTIL_FORMAT_TYPE_FIXED:
2056 if (desc->channel[first_non_void].size >= 32 ||
2057 desc->channel[first_non_void].pure_integer)
2058 return V_008F0C_BUF_NUM_FORMAT_SINT;
2059 else if (desc->channel[first_non_void].normalized)
2060 return V_008F0C_BUF_NUM_FORMAT_SNORM;
2061 else
2062 return V_008F0C_BUF_NUM_FORMAT_SSCALED;
2063 break;
2064 case UTIL_FORMAT_TYPE_UNSIGNED:
2065 if (desc->channel[first_non_void].size >= 32 ||
2066 desc->channel[first_non_void].pure_integer)
2067 return V_008F0C_BUF_NUM_FORMAT_UINT;
2068 else if (desc->channel[first_non_void].normalized)
2069 return V_008F0C_BUF_NUM_FORMAT_UNORM;
2070 else
2071 return V_008F0C_BUF_NUM_FORMAT_USCALED;
2072 break;
2073 case UTIL_FORMAT_TYPE_FLOAT:
2074 default:
2075 return V_008F0C_BUF_NUM_FORMAT_FLOAT;
2076 }
2077 }
2078
2079 static unsigned si_is_vertex_format_supported(struct pipe_screen *screen,
2080 enum pipe_format format,
2081 unsigned usage)
2082 {
2083 const struct util_format_description *desc;
2084 int first_non_void;
2085 unsigned data_format;
2086
2087 assert((usage & ~(PIPE_BIND_SHADER_IMAGE |
2088 PIPE_BIND_SAMPLER_VIEW |
2089 PIPE_BIND_VERTEX_BUFFER)) == 0);
2090
2091 desc = util_format_description(format);
2092 if (!desc)
2093 return 0;
2094
2095 /* There are no native 8_8_8 or 16_16_16 data formats, and we currently
2096 * select 8_8_8_8 and 16_16_16_16 instead. This works reasonably well
2097 * for read-only access (with caveats surrounding bounds checks), but
2098 * obviously fails for write access which we have to implement for
2099 * shader images. Luckily, OpenGL doesn't expect this to be supported
2100 * anyway, and so the only impact is on PBO uploads / downloads, which
2101 * shouldn't be expected to be fast for GL_RGB anyway.
2102 */
2103 if (desc->block.bits == 3 * 8 ||
2104 desc->block.bits == 3 * 16) {
2105 if (usage & (PIPE_BIND_SHADER_IMAGE | PIPE_BIND_SAMPLER_VIEW)) {
2106 usage &= ~(PIPE_BIND_SHADER_IMAGE | PIPE_BIND_SAMPLER_VIEW);
2107 if (!usage)
2108 return 0;
2109 }
2110 }
2111
2112 first_non_void = util_format_get_first_non_void_channel(format);
2113 data_format = si_translate_buffer_dataformat(screen, desc, first_non_void);
2114 if (data_format == V_008F0C_BUF_DATA_FORMAT_INVALID)
2115 return 0;
2116
2117 return usage;
2118 }
2119
2120 static bool si_is_colorbuffer_format_supported(enum pipe_format format)
2121 {
2122 return si_translate_colorformat(format) != V_028C70_COLOR_INVALID &&
2123 si_translate_colorswap(format, false) != ~0U;
2124 }
2125
2126 static bool si_is_zs_format_supported(enum pipe_format format)
2127 {
2128 return si_translate_dbformat(format) != V_028040_Z_INVALID;
2129 }
2130
2131 static boolean si_is_format_supported(struct pipe_screen *screen,
2132 enum pipe_format format,
2133 enum pipe_texture_target target,
2134 unsigned sample_count,
2135 unsigned storage_sample_count,
2136 unsigned usage)
2137 {
2138 struct si_screen *sscreen = (struct si_screen *)screen;
2139 unsigned retval = 0;
2140
2141 if (target >= PIPE_MAX_TEXTURE_TYPES) {
2142 PRINT_ERR("r600: unsupported texture type %d\n", target);
2143 return false;
2144 }
2145
2146 if (MAX2(1, sample_count) < MAX2(1, storage_sample_count))
2147 return false;
2148
2149 if (sample_count > 1) {
2150 if (!screen->get_param(screen, PIPE_CAP_TEXTURE_MULTISAMPLE))
2151 return false;
2152
2153 if (usage & PIPE_BIND_SHADER_IMAGE)
2154 return false;
2155
2156 /* Only power-of-two sample counts are supported. */
2157 if (!util_is_power_of_two_or_zero(sample_count) ||
2158 !util_is_power_of_two_or_zero(storage_sample_count))
2159 return false;
2160
2161 /* MSAA support without framebuffer attachments. */
2162 if (format == PIPE_FORMAT_NONE && sample_count <= 16)
2163 return true;
2164
2165 if (!sscreen->info.has_eqaa_surface_allocator ||
2166 util_format_is_depth_or_stencil(format)) {
2167 /* Color without EQAA or depth/stencil. */
2168 if (sample_count > 8 ||
2169 sample_count != storage_sample_count)
2170 return false;
2171 } else {
2172 /* Color with EQAA. */
2173 if (sample_count > 16 ||
2174 storage_sample_count > 8)
2175 return false;
2176 }
2177 }
2178
2179 if (usage & (PIPE_BIND_SAMPLER_VIEW |
2180 PIPE_BIND_SHADER_IMAGE)) {
2181 if (target == PIPE_BUFFER) {
2182 retval |= si_is_vertex_format_supported(
2183 screen, format, usage & (PIPE_BIND_SAMPLER_VIEW |
2184 PIPE_BIND_SHADER_IMAGE));
2185 } else {
2186 if (si_is_sampler_format_supported(screen, format))
2187 retval |= usage & (PIPE_BIND_SAMPLER_VIEW |
2188 PIPE_BIND_SHADER_IMAGE);
2189 }
2190 }
2191
2192 if ((usage & (PIPE_BIND_RENDER_TARGET |
2193 PIPE_BIND_DISPLAY_TARGET |
2194 PIPE_BIND_SCANOUT |
2195 PIPE_BIND_SHARED |
2196 PIPE_BIND_BLENDABLE)) &&
2197 si_is_colorbuffer_format_supported(format)) {
2198 retval |= usage &
2199 (PIPE_BIND_RENDER_TARGET |
2200 PIPE_BIND_DISPLAY_TARGET |
2201 PIPE_BIND_SCANOUT |
2202 PIPE_BIND_SHARED);
2203 if (!util_format_is_pure_integer(format) &&
2204 !util_format_is_depth_or_stencil(format))
2205 retval |= usage & PIPE_BIND_BLENDABLE;
2206 }
2207
2208 if ((usage & PIPE_BIND_DEPTH_STENCIL) &&
2209 si_is_zs_format_supported(format)) {
2210 retval |= PIPE_BIND_DEPTH_STENCIL;
2211 }
2212
2213 if (usage & PIPE_BIND_VERTEX_BUFFER) {
2214 retval |= si_is_vertex_format_supported(screen, format,
2215 PIPE_BIND_VERTEX_BUFFER);
2216 }
2217
2218 if ((usage & PIPE_BIND_LINEAR) &&
2219 !util_format_is_compressed(format) &&
2220 !(usage & PIPE_BIND_DEPTH_STENCIL))
2221 retval |= PIPE_BIND_LINEAR;
2222
2223 return retval == usage;
2224 }
2225
2226 /*
2227 * framebuffer handling
2228 */
2229
2230 static void si_choose_spi_color_formats(struct si_surface *surf,
2231 unsigned format, unsigned swap,
2232 unsigned ntype, bool is_depth)
2233 {
2234 /* Alpha is needed for alpha-to-coverage.
2235 * Blending may be with or without alpha.
2236 */
2237 unsigned normal = 0; /* most optimal, may not support blending or export alpha */
2238 unsigned alpha = 0; /* exports alpha, but may not support blending */
2239 unsigned blend = 0; /* supports blending, but may not export alpha */
2240 unsigned blend_alpha = 0; /* least optimal, supports blending and exports alpha */
2241
2242 /* Choose the SPI color formats. These are required values for RB+.
2243 * Other chips have multiple choices, though they are not necessarily better.
2244 */
2245 switch (format) {
2246 case V_028C70_COLOR_5_6_5:
2247 case V_028C70_COLOR_1_5_5_5:
2248 case V_028C70_COLOR_5_5_5_1:
2249 case V_028C70_COLOR_4_4_4_4:
2250 case V_028C70_COLOR_10_11_11:
2251 case V_028C70_COLOR_11_11_10:
2252 case V_028C70_COLOR_8:
2253 case V_028C70_COLOR_8_8:
2254 case V_028C70_COLOR_8_8_8_8:
2255 case V_028C70_COLOR_10_10_10_2:
2256 case V_028C70_COLOR_2_10_10_10:
2257 if (ntype == V_028C70_NUMBER_UINT)
2258 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_UINT16_ABGR;
2259 else if (ntype == V_028C70_NUMBER_SINT)
2260 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_SINT16_ABGR;
2261 else
2262 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_FP16_ABGR;
2263 break;
2264
2265 case V_028C70_COLOR_16:
2266 case V_028C70_COLOR_16_16:
2267 case V_028C70_COLOR_16_16_16_16:
2268 if (ntype == V_028C70_NUMBER_UNORM ||
2269 ntype == V_028C70_NUMBER_SNORM) {
2270 /* UNORM16 and SNORM16 don't support blending */
2271 if (ntype == V_028C70_NUMBER_UNORM)
2272 normal = alpha = V_028714_SPI_SHADER_UNORM16_ABGR;
2273 else
2274 normal = alpha = V_028714_SPI_SHADER_SNORM16_ABGR;
2275
2276 /* Use 32 bits per channel for blending. */
2277 if (format == V_028C70_COLOR_16) {
2278 if (swap == V_028C70_SWAP_STD) { /* R */
2279 blend = V_028714_SPI_SHADER_32_R;
2280 blend_alpha = V_028714_SPI_SHADER_32_AR;
2281 } else if (swap == V_028C70_SWAP_ALT_REV) /* A */
2282 blend = blend_alpha = V_028714_SPI_SHADER_32_AR;
2283 else
2284 assert(0);
2285 } else if (format == V_028C70_COLOR_16_16) {
2286 if (swap == V_028C70_SWAP_STD) { /* RG */
2287 blend = V_028714_SPI_SHADER_32_GR;
2288 blend_alpha = V_028714_SPI_SHADER_32_ABGR;
2289 } else if (swap == V_028C70_SWAP_ALT) /* RA */
2290 blend = blend_alpha = V_028714_SPI_SHADER_32_AR;
2291 else
2292 assert(0);
2293 } else /* 16_16_16_16 */
2294 blend = blend_alpha = V_028714_SPI_SHADER_32_ABGR;
2295 } else if (ntype == V_028C70_NUMBER_UINT)
2296 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_UINT16_ABGR;
2297 else if (ntype == V_028C70_NUMBER_SINT)
2298 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_SINT16_ABGR;
2299 else if (ntype == V_028C70_NUMBER_FLOAT)
2300 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_FP16_ABGR;
2301 else
2302 assert(0);
2303 break;
2304
2305 case V_028C70_COLOR_32:
2306 if (swap == V_028C70_SWAP_STD) { /* R */
2307 blend = normal = V_028714_SPI_SHADER_32_R;
2308 alpha = blend_alpha = V_028714_SPI_SHADER_32_AR;
2309 } else if (swap == V_028C70_SWAP_ALT_REV) /* A */
2310 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_32_AR;
2311 else
2312 assert(0);
2313 break;
2314
2315 case V_028C70_COLOR_32_32:
2316 if (swap == V_028C70_SWAP_STD) { /* RG */
2317 blend = normal = V_028714_SPI_SHADER_32_GR;
2318 alpha = blend_alpha = V_028714_SPI_SHADER_32_ABGR;
2319 } else if (swap == V_028C70_SWAP_ALT) /* RA */
2320 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_32_AR;
2321 else
2322 assert(0);
2323 break;
2324
2325 case V_028C70_COLOR_32_32_32_32:
2326 case V_028C70_COLOR_8_24:
2327 case V_028C70_COLOR_24_8:
2328 case V_028C70_COLOR_X24_8_32_FLOAT:
2329 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_32_ABGR;
2330 break;
2331
2332 default:
2333 assert(0);
2334 return;
2335 }
2336
2337 /* The DB->CB copy needs 32_ABGR. */
2338 if (is_depth)
2339 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_32_ABGR;
2340
2341 surf->spi_shader_col_format = normal;
2342 surf->spi_shader_col_format_alpha = alpha;
2343 surf->spi_shader_col_format_blend = blend;
2344 surf->spi_shader_col_format_blend_alpha = blend_alpha;
2345 }
2346
2347 static void si_initialize_color_surface(struct si_context *sctx,
2348 struct si_surface *surf)
2349 {
2350 struct si_texture *tex = (struct si_texture*)surf->base.texture;
2351 unsigned color_info, color_attrib;
2352 unsigned format, swap, ntype, endian;
2353 const struct util_format_description *desc;
2354 int firstchan;
2355 unsigned blend_clamp = 0, blend_bypass = 0;
2356
2357 desc = util_format_description(surf->base.format);
2358 for (firstchan = 0; firstchan < 4; firstchan++) {
2359 if (desc->channel[firstchan].type != UTIL_FORMAT_TYPE_VOID) {
2360 break;
2361 }
2362 }
2363 if (firstchan == 4 || desc->channel[firstchan].type == UTIL_FORMAT_TYPE_FLOAT) {
2364 ntype = V_028C70_NUMBER_FLOAT;
2365 } else {
2366 ntype = V_028C70_NUMBER_UNORM;
2367 if (desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB)
2368 ntype = V_028C70_NUMBER_SRGB;
2369 else if (desc->channel[firstchan].type == UTIL_FORMAT_TYPE_SIGNED) {
2370 if (desc->channel[firstchan].pure_integer) {
2371 ntype = V_028C70_NUMBER_SINT;
2372 } else {
2373 assert(desc->channel[firstchan].normalized);
2374 ntype = V_028C70_NUMBER_SNORM;
2375 }
2376 } else if (desc->channel[firstchan].type == UTIL_FORMAT_TYPE_UNSIGNED) {
2377 if (desc->channel[firstchan].pure_integer) {
2378 ntype = V_028C70_NUMBER_UINT;
2379 } else {
2380 assert(desc->channel[firstchan].normalized);
2381 ntype = V_028C70_NUMBER_UNORM;
2382 }
2383 }
2384 }
2385
2386 format = si_translate_colorformat(surf->base.format);
2387 if (format == V_028C70_COLOR_INVALID) {
2388 PRINT_ERR("Invalid CB format: %d, disabling CB.\n", surf->base.format);
2389 }
2390 assert(format != V_028C70_COLOR_INVALID);
2391 swap = si_translate_colorswap(surf->base.format, false);
2392 endian = si_colorformat_endian_swap(format);
2393
2394 /* blend clamp should be set for all NORM/SRGB types */
2395 if (ntype == V_028C70_NUMBER_UNORM ||
2396 ntype == V_028C70_NUMBER_SNORM ||
2397 ntype == V_028C70_NUMBER_SRGB)
2398 blend_clamp = 1;
2399
2400 /* set blend bypass according to docs if SINT/UINT or
2401 8/24 COLOR variants */
2402 if (ntype == V_028C70_NUMBER_UINT || ntype == V_028C70_NUMBER_SINT ||
2403 format == V_028C70_COLOR_8_24 || format == V_028C70_COLOR_24_8 ||
2404 format == V_028C70_COLOR_X24_8_32_FLOAT) {
2405 blend_clamp = 0;
2406 blend_bypass = 1;
2407 }
2408
2409 if (ntype == V_028C70_NUMBER_UINT || ntype == V_028C70_NUMBER_SINT) {
2410 if (format == V_028C70_COLOR_8 ||
2411 format == V_028C70_COLOR_8_8 ||
2412 format == V_028C70_COLOR_8_8_8_8)
2413 surf->color_is_int8 = true;
2414 else if (format == V_028C70_COLOR_10_10_10_2 ||
2415 format == V_028C70_COLOR_2_10_10_10)
2416 surf->color_is_int10 = true;
2417 }
2418
2419 color_info = S_028C70_FORMAT(format) |
2420 S_028C70_COMP_SWAP(swap) |
2421 S_028C70_BLEND_CLAMP(blend_clamp) |
2422 S_028C70_BLEND_BYPASS(blend_bypass) |
2423 S_028C70_SIMPLE_FLOAT(1) |
2424 S_028C70_ROUND_MODE(ntype != V_028C70_NUMBER_UNORM &&
2425 ntype != V_028C70_NUMBER_SNORM &&
2426 ntype != V_028C70_NUMBER_SRGB &&
2427 format != V_028C70_COLOR_8_24 &&
2428 format != V_028C70_COLOR_24_8) |
2429 S_028C70_NUMBER_TYPE(ntype) |
2430 S_028C70_ENDIAN(endian);
2431
2432 /* Intensity is implemented as Red, so treat it that way. */
2433 color_attrib = S_028C74_FORCE_DST_ALPHA_1(desc->swizzle[3] == PIPE_SWIZZLE_1 ||
2434 util_format_is_intensity(surf->base.format));
2435
2436 if (tex->buffer.b.b.nr_samples > 1) {
2437 unsigned log_samples = util_logbase2(tex->buffer.b.b.nr_samples);
2438 unsigned log_fragments = util_logbase2(tex->buffer.b.b.nr_storage_samples);
2439
2440 color_attrib |= S_028C74_NUM_SAMPLES(log_samples) |
2441 S_028C74_NUM_FRAGMENTS(log_fragments);
2442
2443 if (tex->surface.fmask_size) {
2444 color_info |= S_028C70_COMPRESSION(1);
2445 unsigned fmask_bankh = util_logbase2(tex->surface.u.legacy.fmask.bankh);
2446
2447 if (sctx->chip_class == SI) {
2448 /* due to a hw bug, FMASK_BANK_HEIGHT must be set on SI too */
2449 color_attrib |= S_028C74_FMASK_BANK_HEIGHT(fmask_bankh);
2450 }
2451 }
2452 }
2453
2454 if (sctx->chip_class >= VI) {
2455 unsigned max_uncompressed_block_size = V_028C78_MAX_BLOCK_SIZE_256B;
2456 unsigned min_compressed_block_size = V_028C78_MIN_BLOCK_SIZE_32B;
2457
2458 /* amdvlk: [min-compressed-block-size] should be set to 32 for dGPU and
2459 64 for APU because all of our APUs to date use DIMMs which have
2460 a request granularity size of 64B while all other chips have a
2461 32B request size */
2462 if (!sctx->screen->info.has_dedicated_vram)
2463 min_compressed_block_size = V_028C78_MIN_BLOCK_SIZE_64B;
2464
2465 if (tex->buffer.b.b.nr_storage_samples > 1) {
2466 if (tex->surface.bpe == 1)
2467 max_uncompressed_block_size = V_028C78_MAX_BLOCK_SIZE_64B;
2468 else if (tex->surface.bpe == 2)
2469 max_uncompressed_block_size = V_028C78_MAX_BLOCK_SIZE_128B;
2470 }
2471
2472 surf->cb_dcc_control = S_028C78_MAX_UNCOMPRESSED_BLOCK_SIZE(max_uncompressed_block_size) |
2473 S_028C78_MIN_COMPRESSED_BLOCK_SIZE(min_compressed_block_size) |
2474 S_028C78_INDEPENDENT_64B_BLOCKS(1);
2475 }
2476
2477 /* This must be set for fast clear to work without FMASK. */
2478 if (!tex->surface.fmask_size && sctx->chip_class == SI) {
2479 unsigned bankh = util_logbase2(tex->surface.u.legacy.bankh);
2480 color_attrib |= S_028C74_FMASK_BANK_HEIGHT(bankh);
2481 }
2482
2483 unsigned color_view = S_028C6C_SLICE_START(surf->base.u.tex.first_layer) |
2484 S_028C6C_SLICE_MAX(surf->base.u.tex.last_layer);
2485
2486 if (sctx->chip_class >= GFX9) {
2487 unsigned mip0_depth = util_max_layer(&tex->buffer.b.b, 0);
2488
2489 color_view |= S_028C6C_MIP_LEVEL(surf->base.u.tex.level);
2490 color_attrib |= S_028C74_MIP0_DEPTH(mip0_depth) |
2491 S_028C74_RESOURCE_TYPE(tex->surface.u.gfx9.resource_type);
2492 surf->cb_color_attrib2 = S_028C68_MIP0_WIDTH(surf->width0 - 1) |
2493 S_028C68_MIP0_HEIGHT(surf->height0 - 1) |
2494 S_028C68_MAX_MIP(tex->buffer.b.b.last_level);
2495 }
2496
2497 surf->cb_color_view = color_view;
2498 surf->cb_color_info = color_info;
2499 surf->cb_color_attrib = color_attrib;
2500
2501 /* Determine pixel shader export format */
2502 si_choose_spi_color_formats(surf, format, swap, ntype, tex->is_depth);
2503
2504 surf->color_initialized = true;
2505 }
2506
2507 static void si_init_depth_surface(struct si_context *sctx,
2508 struct si_surface *surf)
2509 {
2510 struct si_texture *tex = (struct si_texture*)surf->base.texture;
2511 unsigned level = surf->base.u.tex.level;
2512 unsigned format, stencil_format;
2513 uint32_t z_info, s_info;
2514
2515 format = si_translate_dbformat(tex->db_render_format);
2516 stencil_format = tex->surface.has_stencil ?
2517 V_028044_STENCIL_8 : V_028044_STENCIL_INVALID;
2518
2519 assert(format != V_028040_Z_INVALID);
2520 if (format == V_028040_Z_INVALID)
2521 PRINT_ERR("Invalid DB format: %d, disabling DB.\n", tex->buffer.b.b.format);
2522
2523 surf->db_depth_view = S_028008_SLICE_START(surf->base.u.tex.first_layer) |
2524 S_028008_SLICE_MAX(surf->base.u.tex.last_layer);
2525 surf->db_htile_data_base = 0;
2526 surf->db_htile_surface = 0;
2527
2528 if (sctx->chip_class >= GFX9) {
2529 assert(tex->surface.u.gfx9.surf_offset == 0);
2530 surf->db_depth_base = tex->buffer.gpu_address >> 8;
2531 surf->db_stencil_base = (tex->buffer.gpu_address +
2532 tex->surface.u.gfx9.stencil_offset) >> 8;
2533 z_info = S_028038_FORMAT(format) |
2534 S_028038_NUM_SAMPLES(util_logbase2(tex->buffer.b.b.nr_samples)) |
2535 S_028038_SW_MODE(tex->surface.u.gfx9.surf.swizzle_mode) |
2536 S_028038_MAXMIP(tex->buffer.b.b.last_level);
2537 s_info = S_02803C_FORMAT(stencil_format) |
2538 S_02803C_SW_MODE(tex->surface.u.gfx9.stencil.swizzle_mode);
2539 surf->db_z_info2 = S_028068_EPITCH(tex->surface.u.gfx9.surf.epitch);
2540 surf->db_stencil_info2 = S_02806C_EPITCH(tex->surface.u.gfx9.stencil.epitch);
2541 surf->db_depth_view |= S_028008_MIPID(level);
2542 surf->db_depth_size = S_02801C_X_MAX(tex->buffer.b.b.width0 - 1) |
2543 S_02801C_Y_MAX(tex->buffer.b.b.height0 - 1);
2544
2545 if (si_htile_enabled(tex, level)) {
2546 z_info |= S_028038_TILE_SURFACE_ENABLE(1) |
2547 S_028038_ALLOW_EXPCLEAR(1);
2548
2549 if (tex->tc_compatible_htile) {
2550 unsigned max_zplanes = 4;
2551
2552 if (tex->db_render_format == PIPE_FORMAT_Z16_UNORM &&
2553 tex->buffer.b.b.nr_samples > 1)
2554 max_zplanes = 2;
2555
2556 z_info |= S_028038_DECOMPRESS_ON_N_ZPLANES(max_zplanes + 1) |
2557 S_028038_ITERATE_FLUSH(1);
2558 s_info |= S_02803C_ITERATE_FLUSH(1);
2559 }
2560
2561 if (tex->surface.has_stencil) {
2562 /* Stencil buffer workaround ported from the SI-CI-VI code.
2563 * See that for explanation.
2564 */
2565 s_info |= S_02803C_ALLOW_EXPCLEAR(tex->buffer.b.b.nr_samples <= 1);
2566 } else {
2567 /* Use all HTILE for depth if there's no stencil. */
2568 s_info |= S_02803C_TILE_STENCIL_DISABLE(1);
2569 }
2570
2571 surf->db_htile_data_base = (tex->buffer.gpu_address +
2572 tex->htile_offset) >> 8;
2573 surf->db_htile_surface = S_028ABC_FULL_CACHE(1) |
2574 S_028ABC_PIPE_ALIGNED(tex->surface.u.gfx9.htile.pipe_aligned) |
2575 S_028ABC_RB_ALIGNED(tex->surface.u.gfx9.htile.rb_aligned);
2576 }
2577 } else {
2578 /* SI-CI-VI */
2579 struct legacy_surf_level *levelinfo = &tex->surface.u.legacy.level[level];
2580
2581 assert(levelinfo->nblk_x % 8 == 0 && levelinfo->nblk_y % 8 == 0);
2582
2583 surf->db_depth_base = (tex->buffer.gpu_address +
2584 tex->surface.u.legacy.level[level].offset) >> 8;
2585 surf->db_stencil_base = (tex->buffer.gpu_address +
2586 tex->surface.u.legacy.stencil_level[level].offset) >> 8;
2587
2588 z_info = S_028040_FORMAT(format) |
2589 S_028040_NUM_SAMPLES(util_logbase2(tex->buffer.b.b.nr_samples));
2590 s_info = S_028044_FORMAT(stencil_format);
2591 surf->db_depth_info = S_02803C_ADDR5_SWIZZLE_MASK(!tex->tc_compatible_htile);
2592
2593 if (sctx->chip_class >= CIK) {
2594 struct radeon_info *info = &sctx->screen->info;
2595 unsigned index = tex->surface.u.legacy.tiling_index[level];
2596 unsigned stencil_index = tex->surface.u.legacy.stencil_tiling_index[level];
2597 unsigned macro_index = tex->surface.u.legacy.macro_tile_index;
2598 unsigned tile_mode = info->si_tile_mode_array[index];
2599 unsigned stencil_tile_mode = info->si_tile_mode_array[stencil_index];
2600 unsigned macro_mode = info->cik_macrotile_mode_array[macro_index];
2601
2602 surf->db_depth_info |=
2603 S_02803C_ARRAY_MODE(G_009910_ARRAY_MODE(tile_mode)) |
2604 S_02803C_PIPE_CONFIG(G_009910_PIPE_CONFIG(tile_mode)) |
2605 S_02803C_BANK_WIDTH(G_009990_BANK_WIDTH(macro_mode)) |
2606 S_02803C_BANK_HEIGHT(G_009990_BANK_HEIGHT(macro_mode)) |
2607 S_02803C_MACRO_TILE_ASPECT(G_009990_MACRO_TILE_ASPECT(macro_mode)) |
2608 S_02803C_NUM_BANKS(G_009990_NUM_BANKS(macro_mode));
2609 z_info |= S_028040_TILE_SPLIT(G_009910_TILE_SPLIT(tile_mode));
2610 s_info |= S_028044_TILE_SPLIT(G_009910_TILE_SPLIT(stencil_tile_mode));
2611 } else {
2612 unsigned tile_mode_index = si_tile_mode_index(tex, level, false);
2613 z_info |= S_028040_TILE_MODE_INDEX(tile_mode_index);
2614 tile_mode_index = si_tile_mode_index(tex, level, true);
2615 s_info |= S_028044_TILE_MODE_INDEX(tile_mode_index);
2616 }
2617
2618 surf->db_depth_size = S_028058_PITCH_TILE_MAX((levelinfo->nblk_x / 8) - 1) |
2619 S_028058_HEIGHT_TILE_MAX((levelinfo->nblk_y / 8) - 1);
2620 surf->db_depth_slice = S_02805C_SLICE_TILE_MAX((levelinfo->nblk_x *
2621 levelinfo->nblk_y) / 64 - 1);
2622
2623 if (si_htile_enabled(tex, level)) {
2624 z_info |= S_028040_TILE_SURFACE_ENABLE(1) |
2625 S_028040_ALLOW_EXPCLEAR(1);
2626
2627 if (tex->surface.has_stencil) {
2628 /* Workaround: For a not yet understood reason, the
2629 * combination of MSAA, fast stencil clear and stencil
2630 * decompress messes with subsequent stencil buffer
2631 * uses. Problem was reproduced on Verde, Bonaire,
2632 * Tonga, and Carrizo.
2633 *
2634 * Disabling EXPCLEAR works around the problem.
2635 *
2636 * Check piglit's arb_texture_multisample-stencil-clear
2637 * test if you want to try changing this.
2638 */
2639 if (tex->buffer.b.b.nr_samples <= 1)
2640 s_info |= S_028044_ALLOW_EXPCLEAR(1);
2641 } else if (!tex->tc_compatible_htile) {
2642 /* Use all of the htile_buffer for depth if there's no stencil.
2643 * This must not be set when TC-compatible HTILE is enabled
2644 * due to a hw bug.
2645 */
2646 s_info |= S_028044_TILE_STENCIL_DISABLE(1);
2647 }
2648
2649 surf->db_htile_data_base = (tex->buffer.gpu_address +
2650 tex->htile_offset) >> 8;
2651 surf->db_htile_surface = S_028ABC_FULL_CACHE(1);
2652
2653 if (tex->tc_compatible_htile) {
2654 surf->db_htile_surface |= S_028ABC_TC_COMPATIBLE(1);
2655
2656 /* 0 = full compression. N = only compress up to N-1 Z planes. */
2657 if (tex->buffer.b.b.nr_samples <= 1)
2658 z_info |= S_028040_DECOMPRESS_ON_N_ZPLANES(5);
2659 else if (tex->buffer.b.b.nr_samples <= 4)
2660 z_info |= S_028040_DECOMPRESS_ON_N_ZPLANES(3);
2661 else
2662 z_info |= S_028040_DECOMPRESS_ON_N_ZPLANES(2);
2663 }
2664 }
2665 }
2666
2667 surf->db_z_info = z_info;
2668 surf->db_stencil_info = s_info;
2669
2670 surf->depth_initialized = true;
2671 }
2672
2673 void si_update_fb_dirtiness_after_rendering(struct si_context *sctx)
2674 {
2675 if (sctx->decompression_enabled)
2676 return;
2677
2678 if (sctx->framebuffer.state.zsbuf) {
2679 struct pipe_surface *surf = sctx->framebuffer.state.zsbuf;
2680 struct si_texture *tex = (struct si_texture *)surf->texture;
2681
2682 tex->dirty_level_mask |= 1 << surf->u.tex.level;
2683
2684 if (tex->surface.has_stencil)
2685 tex->stencil_dirty_level_mask |= 1 << surf->u.tex.level;
2686 }
2687
2688 unsigned compressed_cb_mask = sctx->framebuffer.compressed_cb_mask;
2689 while (compressed_cb_mask) {
2690 unsigned i = u_bit_scan(&compressed_cb_mask);
2691 struct pipe_surface *surf = sctx->framebuffer.state.cbufs[i];
2692 struct si_texture *tex = (struct si_texture*)surf->texture;
2693
2694 if (tex->surface.fmask_size)
2695 tex->dirty_level_mask |= 1 << surf->u.tex.level;
2696 if (tex->dcc_gather_statistics)
2697 tex->separate_dcc_dirty = true;
2698 }
2699 }
2700
2701 static void si_dec_framebuffer_counters(const struct pipe_framebuffer_state *state)
2702 {
2703 for (int i = 0; i < state->nr_cbufs; ++i) {
2704 struct si_surface *surf = NULL;
2705 struct si_texture *tex;
2706
2707 if (!state->cbufs[i])
2708 continue;
2709 surf = (struct si_surface*)state->cbufs[i];
2710 tex = (struct si_texture*)surf->base.texture;
2711
2712 p_atomic_dec(&tex->framebuffers_bound);
2713 }
2714 }
2715
2716 static void si_set_framebuffer_state(struct pipe_context *ctx,
2717 const struct pipe_framebuffer_state *state)
2718 {
2719 struct si_context *sctx = (struct si_context *)ctx;
2720 struct si_surface *surf = NULL;
2721 struct si_texture *tex;
2722 bool old_any_dst_linear = sctx->framebuffer.any_dst_linear;
2723 unsigned old_nr_samples = sctx->framebuffer.nr_samples;
2724 unsigned old_colorbuf_enabled_4bit = sctx->framebuffer.colorbuf_enabled_4bit;
2725 bool old_has_zsbuf = !!sctx->framebuffer.state.zsbuf;
2726 bool old_has_stencil =
2727 old_has_zsbuf &&
2728 ((struct si_texture*)sctx->framebuffer.state.zsbuf->texture)->surface.has_stencil;
2729 bool unbound = false;
2730 int i;
2731
2732 /* Reject zero-sized framebuffers due to a hw bug on SI that occurs
2733 * when PA_SU_HARDWARE_SCREEN_OFFSET != 0 and any_scissor.BR_X/Y <= 0.
2734 * We could implement the full workaround here, but it's a useless case.
2735 */
2736 if ((!state->width || !state->height) && (state->nr_cbufs || state->zsbuf)) {
2737 unreachable("the framebuffer shouldn't have zero area");
2738 return;
2739 }
2740
2741 si_update_fb_dirtiness_after_rendering(sctx);
2742
2743 for (i = 0; i < sctx->framebuffer.state.nr_cbufs; i++) {
2744 if (!sctx->framebuffer.state.cbufs[i])
2745 continue;
2746
2747 tex = (struct si_texture*)sctx->framebuffer.state.cbufs[i]->texture;
2748 if (tex->dcc_gather_statistics)
2749 vi_separate_dcc_stop_query(sctx, tex);
2750 }
2751
2752 /* Disable DCC if the formats are incompatible. */
2753 for (i = 0; i < state->nr_cbufs; i++) {
2754 if (!state->cbufs[i])
2755 continue;
2756
2757 surf = (struct si_surface*)state->cbufs[i];
2758 tex = (struct si_texture*)surf->base.texture;
2759
2760 if (!surf->dcc_incompatible)
2761 continue;
2762
2763 /* Since the DCC decompression calls back into set_framebuffer-
2764 * _state, we need to unbind the framebuffer, so that
2765 * vi_separate_dcc_stop_query isn't called twice with the same
2766 * color buffer.
2767 */
2768 if (!unbound) {
2769 util_copy_framebuffer_state(&sctx->framebuffer.state, NULL);
2770 unbound = true;
2771 }
2772
2773 if (vi_dcc_enabled(tex, surf->base.u.tex.level))
2774 if (!si_texture_disable_dcc(sctx, tex))
2775 si_decompress_dcc(sctx, tex);
2776
2777 surf->dcc_incompatible = false;
2778 }
2779
2780 /* Only flush TC when changing the framebuffer state, because
2781 * the only client not using TC that can change textures is
2782 * the framebuffer.
2783 *
2784 * Wait for compute shaders because of possible transitions:
2785 * - FB write -> shader read
2786 * - shader write -> FB read
2787 *
2788 * DB caches are flushed on demand (using si_decompress_textures).
2789 *
2790 * When MSAA is enabled, CB and TC caches are flushed on demand
2791 * (after FMASK decompression). Shader write -> FB read transitions
2792 * cannot happen for MSAA textures, because MSAA shader images are
2793 * not supported.
2794 *
2795 * Only flush and wait for CB if there is actually a bound color buffer.
2796 */
2797 if (sctx->framebuffer.uncompressed_cb_mask)
2798 si_make_CB_shader_coherent(sctx, sctx->framebuffer.nr_samples,
2799 sctx->framebuffer.CB_has_shader_readable_metadata);
2800
2801 sctx->flags |= SI_CONTEXT_CS_PARTIAL_FLUSH;
2802
2803 /* u_blitter doesn't invoke depth decompression when it does multiple
2804 * blits in a row, but the only case when it matters for DB is when
2805 * doing generate_mipmap. So here we flush DB manually between
2806 * individual generate_mipmap blits.
2807 * Note that lower mipmap levels aren't compressed.
2808 */
2809 if (sctx->generate_mipmap_for_depth) {
2810 si_make_DB_shader_coherent(sctx, 1, false,
2811 sctx->framebuffer.DB_has_shader_readable_metadata);
2812 } else if (sctx->chip_class == GFX9) {
2813 /* It appears that DB metadata "leaks" in a sequence of:
2814 * - depth clear
2815 * - DCC decompress for shader image writes (with DB disabled)
2816 * - render with DEPTH_BEFORE_SHADER=1
2817 * Flushing DB metadata works around the problem.
2818 */
2819 sctx->flags |= SI_CONTEXT_FLUSH_AND_INV_DB_META;
2820 }
2821
2822 /* Take the maximum of the old and new count. If the new count is lower,
2823 * dirtying is needed to disable the unbound colorbuffers.
2824 */
2825 sctx->framebuffer.dirty_cbufs |=
2826 (1 << MAX2(sctx->framebuffer.state.nr_cbufs, state->nr_cbufs)) - 1;
2827 sctx->framebuffer.dirty_zsbuf |= sctx->framebuffer.state.zsbuf != state->zsbuf;
2828
2829 si_dec_framebuffer_counters(&sctx->framebuffer.state);
2830 util_copy_framebuffer_state(&sctx->framebuffer.state, state);
2831
2832 sctx->framebuffer.colorbuf_enabled_4bit = 0;
2833 sctx->framebuffer.spi_shader_col_format = 0;
2834 sctx->framebuffer.spi_shader_col_format_alpha = 0;
2835 sctx->framebuffer.spi_shader_col_format_blend = 0;
2836 sctx->framebuffer.spi_shader_col_format_blend_alpha = 0;
2837 sctx->framebuffer.color_is_int8 = 0;
2838 sctx->framebuffer.color_is_int10 = 0;
2839
2840 sctx->framebuffer.compressed_cb_mask = 0;
2841 sctx->framebuffer.uncompressed_cb_mask = 0;
2842 sctx->framebuffer.nr_samples = util_framebuffer_get_num_samples(state);
2843 sctx->framebuffer.nr_color_samples = sctx->framebuffer.nr_samples;
2844 sctx->framebuffer.log_samples = util_logbase2(sctx->framebuffer.nr_samples);
2845 sctx->framebuffer.any_dst_linear = false;
2846 sctx->framebuffer.CB_has_shader_readable_metadata = false;
2847 sctx->framebuffer.DB_has_shader_readable_metadata = false;
2848
2849 for (i = 0; i < state->nr_cbufs; i++) {
2850 if (!state->cbufs[i])
2851 continue;
2852
2853 surf = (struct si_surface*)state->cbufs[i];
2854 tex = (struct si_texture*)surf->base.texture;
2855
2856 if (!surf->color_initialized) {
2857 si_initialize_color_surface(sctx, surf);
2858 }
2859
2860 sctx->framebuffer.colorbuf_enabled_4bit |= 0xf << (i * 4);
2861 sctx->framebuffer.spi_shader_col_format |=
2862 surf->spi_shader_col_format << (i * 4);
2863 sctx->framebuffer.spi_shader_col_format_alpha |=
2864 surf->spi_shader_col_format_alpha << (i * 4);
2865 sctx->framebuffer.spi_shader_col_format_blend |=
2866 surf->spi_shader_col_format_blend << (i * 4);
2867 sctx->framebuffer.spi_shader_col_format_blend_alpha |=
2868 surf->spi_shader_col_format_blend_alpha << (i * 4);
2869
2870 if (surf->color_is_int8)
2871 sctx->framebuffer.color_is_int8 |= 1 << i;
2872 if (surf->color_is_int10)
2873 sctx->framebuffer.color_is_int10 |= 1 << i;
2874
2875 if (tex->surface.fmask_size)
2876 sctx->framebuffer.compressed_cb_mask |= 1 << i;
2877 else
2878 sctx->framebuffer.uncompressed_cb_mask |= 1 << i;
2879
2880 /* Don't update nr_color_samples for non-AA buffers.
2881 * (e.g. destination of MSAA resolve)
2882 */
2883 if (tex->buffer.b.b.nr_samples >= 2 &&
2884 tex->buffer.b.b.nr_storage_samples < tex->buffer.b.b.nr_samples) {
2885 sctx->framebuffer.nr_color_samples =
2886 MIN2(sctx->framebuffer.nr_color_samples,
2887 tex->buffer.b.b.nr_storage_samples);
2888 sctx->framebuffer.nr_color_samples =
2889 MAX2(1, sctx->framebuffer.nr_color_samples);
2890 }
2891
2892 if (tex->surface.is_linear)
2893 sctx->framebuffer.any_dst_linear = true;
2894
2895 if (vi_dcc_enabled(tex, surf->base.u.tex.level))
2896 sctx->framebuffer.CB_has_shader_readable_metadata = true;
2897
2898 si_context_add_resource_size(sctx, surf->base.texture);
2899
2900 p_atomic_inc(&tex->framebuffers_bound);
2901
2902 if (tex->dcc_gather_statistics) {
2903 /* Dirty tracking must be enabled for DCC usage analysis. */
2904 sctx->framebuffer.compressed_cb_mask |= 1 << i;
2905 vi_separate_dcc_start_query(sctx, tex);
2906 }
2907 }
2908
2909 struct si_texture *zstex = NULL;
2910
2911 if (state->zsbuf) {
2912 surf = (struct si_surface*)state->zsbuf;
2913 zstex = (struct si_texture*)surf->base.texture;
2914
2915 if (!surf->depth_initialized) {
2916 si_init_depth_surface(sctx, surf);
2917 }
2918
2919 if (vi_tc_compat_htile_enabled(zstex, surf->base.u.tex.level))
2920 sctx->framebuffer.DB_has_shader_readable_metadata = true;
2921
2922 si_context_add_resource_size(sctx, surf->base.texture);
2923 }
2924
2925 si_update_ps_colorbuf0_slot(sctx);
2926 si_update_poly_offset_state(sctx);
2927 si_mark_atom_dirty(sctx, &sctx->atoms.s.cb_render_state);
2928 si_mark_atom_dirty(sctx, &sctx->atoms.s.framebuffer);
2929
2930 if (sctx->screen->dpbb_allowed)
2931 si_mark_atom_dirty(sctx, &sctx->atoms.s.dpbb_state);
2932
2933 if (sctx->framebuffer.any_dst_linear != old_any_dst_linear)
2934 si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
2935
2936 if (sctx->screen->has_out_of_order_rast &&
2937 (sctx->framebuffer.colorbuf_enabled_4bit != old_colorbuf_enabled_4bit ||
2938 !!sctx->framebuffer.state.zsbuf != old_has_zsbuf ||
2939 (zstex && zstex->surface.has_stencil != old_has_stencil)))
2940 si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
2941
2942 if (sctx->framebuffer.nr_samples != old_nr_samples) {
2943 struct pipe_constant_buffer constbuf = {0};
2944
2945 si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
2946 si_mark_atom_dirty(sctx, &sctx->atoms.s.db_render_state);
2947
2948 constbuf.buffer = sctx->sample_pos_buffer;
2949
2950 /* Set sample locations as fragment shader constants. */
2951 switch (sctx->framebuffer.nr_samples) {
2952 case 1:
2953 constbuf.buffer_offset = 0;
2954 break;
2955 case 2:
2956 constbuf.buffer_offset = (ubyte*)sctx->sample_positions.x2 -
2957 (ubyte*)sctx->sample_positions.x1;
2958 break;
2959 case 4:
2960 constbuf.buffer_offset = (ubyte*)sctx->sample_positions.x4 -
2961 (ubyte*)sctx->sample_positions.x1;
2962 break;
2963 case 8:
2964 constbuf.buffer_offset = (ubyte*)sctx->sample_positions.x8 -
2965 (ubyte*)sctx->sample_positions.x1;
2966 break;
2967 case 16:
2968 constbuf.buffer_offset = (ubyte*)sctx->sample_positions.x16 -
2969 (ubyte*)sctx->sample_positions.x1;
2970 break;
2971 default:
2972 PRINT_ERR("Requested an invalid number of samples %i.\n",
2973 sctx->framebuffer.nr_samples);
2974 assert(0);
2975 }
2976 constbuf.buffer_size = sctx->framebuffer.nr_samples * 2 * 4;
2977 si_set_rw_buffer(sctx, SI_PS_CONST_SAMPLE_POSITIONS, &constbuf);
2978
2979 si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_sample_locs);
2980 }
2981
2982 sctx->do_update_shaders = true;
2983
2984 if (!sctx->decompression_enabled) {
2985 /* Prevent textures decompression when the framebuffer state
2986 * changes come from the decompression passes themselves.
2987 */
2988 sctx->need_check_render_feedback = true;
2989 }
2990 }
2991
2992 static void si_emit_framebuffer_state(struct si_context *sctx)
2993 {
2994 struct radeon_cmdbuf *cs = sctx->gfx_cs;
2995 struct pipe_framebuffer_state *state = &sctx->framebuffer.state;
2996 unsigned i, nr_cbufs = state->nr_cbufs;
2997 struct si_texture *tex = NULL;
2998 struct si_surface *cb = NULL;
2999 unsigned cb_color_info = 0;
3000
3001 /* Colorbuffers. */
3002 for (i = 0; i < nr_cbufs; i++) {
3003 uint64_t cb_color_base, cb_color_fmask, cb_color_cmask, cb_dcc_base;
3004 unsigned cb_color_attrib;
3005
3006 if (!(sctx->framebuffer.dirty_cbufs & (1 << i)))
3007 continue;
3008
3009 cb = (struct si_surface*)state->cbufs[i];
3010 if (!cb) {
3011 radeon_set_context_reg(cs, R_028C70_CB_COLOR0_INFO + i * 0x3C,
3012 S_028C70_FORMAT(V_028C70_COLOR_INVALID));
3013 continue;
3014 }
3015
3016 tex = (struct si_texture *)cb->base.texture;
3017 radeon_add_to_buffer_list(sctx, sctx->gfx_cs,
3018 &tex->buffer, RADEON_USAGE_READWRITE,
3019 tex->buffer.b.b.nr_samples > 1 ?
3020 RADEON_PRIO_COLOR_BUFFER_MSAA :
3021 RADEON_PRIO_COLOR_BUFFER);
3022
3023 if (tex->cmask_buffer && tex->cmask_buffer != &tex->buffer) {
3024 radeon_add_to_buffer_list(sctx, sctx->gfx_cs,
3025 tex->cmask_buffer, RADEON_USAGE_READWRITE,
3026 RADEON_PRIO_SEPARATE_META);
3027 }
3028
3029 if (tex->dcc_separate_buffer)
3030 radeon_add_to_buffer_list(sctx, sctx->gfx_cs,
3031 tex->dcc_separate_buffer,
3032 RADEON_USAGE_READWRITE,
3033 RADEON_PRIO_SEPARATE_META);
3034
3035 /* Compute mutable surface parameters. */
3036 cb_color_base = tex->buffer.gpu_address >> 8;
3037 cb_color_fmask = 0;
3038 cb_color_cmask = tex->cmask_base_address_reg;
3039 cb_dcc_base = 0;
3040 cb_color_info = cb->cb_color_info | tex->cb_color_info;
3041 cb_color_attrib = cb->cb_color_attrib;
3042
3043 if (cb->base.u.tex.level > 0)
3044 cb_color_info &= C_028C70_FAST_CLEAR;
3045
3046 if (tex->surface.fmask_size) {
3047 cb_color_fmask = (tex->buffer.gpu_address + tex->fmask_offset) >> 8;
3048 cb_color_fmask |= tex->surface.fmask_tile_swizzle;
3049 }
3050
3051 /* Set up DCC. */
3052 if (vi_dcc_enabled(tex, cb->base.u.tex.level)) {
3053 bool is_msaa_resolve_dst = state->cbufs[0] &&
3054 state->cbufs[0]->texture->nr_samples > 1 &&
3055 state->cbufs[1] == &cb->base &&
3056 state->cbufs[1]->texture->nr_samples <= 1;
3057
3058 if (!is_msaa_resolve_dst)
3059 cb_color_info |= S_028C70_DCC_ENABLE(1);
3060
3061 cb_dcc_base = ((!tex->dcc_separate_buffer ? tex->buffer.gpu_address : 0) +
3062 tex->dcc_offset) >> 8;
3063 cb_dcc_base |= tex->surface.tile_swizzle;
3064 }
3065
3066 if (sctx->chip_class >= GFX9) {
3067 struct gfx9_surf_meta_flags meta;
3068
3069 if (tex->dcc_offset)
3070 meta = tex->surface.u.gfx9.dcc;
3071 else
3072 meta = tex->surface.u.gfx9.cmask;
3073
3074 /* Set mutable surface parameters. */
3075 cb_color_base += tex->surface.u.gfx9.surf_offset >> 8;
3076 cb_color_base |= tex->surface.tile_swizzle;
3077 if (!tex->surface.fmask_size)
3078 cb_color_fmask = cb_color_base;
3079 if (cb->base.u.tex.level > 0)
3080 cb_color_cmask = cb_color_base;
3081 cb_color_attrib |= S_028C74_COLOR_SW_MODE(tex->surface.u.gfx9.surf.swizzle_mode) |
3082 S_028C74_FMASK_SW_MODE(tex->surface.u.gfx9.fmask.swizzle_mode) |
3083 S_028C74_RB_ALIGNED(meta.rb_aligned) |
3084 S_028C74_PIPE_ALIGNED(meta.pipe_aligned);
3085
3086 radeon_set_context_reg_seq(cs, R_028C60_CB_COLOR0_BASE + i * 0x3C, 15);
3087 radeon_emit(cs, cb_color_base); /* CB_COLOR0_BASE */
3088 radeon_emit(cs, S_028C64_BASE_256B(cb_color_base >> 32)); /* CB_COLOR0_BASE_EXT */
3089 radeon_emit(cs, cb->cb_color_attrib2); /* CB_COLOR0_ATTRIB2 */
3090 radeon_emit(cs, cb->cb_color_view); /* CB_COLOR0_VIEW */
3091 radeon_emit(cs, cb_color_info); /* CB_COLOR0_INFO */
3092 radeon_emit(cs, cb_color_attrib); /* CB_COLOR0_ATTRIB */
3093 radeon_emit(cs, cb->cb_dcc_control); /* CB_COLOR0_DCC_CONTROL */
3094 radeon_emit(cs, cb_color_cmask); /* CB_COLOR0_CMASK */
3095 radeon_emit(cs, S_028C80_BASE_256B(cb_color_cmask >> 32)); /* CB_COLOR0_CMASK_BASE_EXT */
3096 radeon_emit(cs, cb_color_fmask); /* CB_COLOR0_FMASK */
3097 radeon_emit(cs, S_028C88_BASE_256B(cb_color_fmask >> 32)); /* CB_COLOR0_FMASK_BASE_EXT */
3098 radeon_emit(cs, tex->color_clear_value[0]); /* CB_COLOR0_CLEAR_WORD0 */
3099 radeon_emit(cs, tex->color_clear_value[1]); /* CB_COLOR0_CLEAR_WORD1 */
3100 radeon_emit(cs, cb_dcc_base); /* CB_COLOR0_DCC_BASE */
3101 radeon_emit(cs, S_028C98_BASE_256B(cb_dcc_base >> 32)); /* CB_COLOR0_DCC_BASE_EXT */
3102
3103 radeon_set_context_reg(cs, R_0287A0_CB_MRT0_EPITCH + i * 4,
3104 S_0287A0_EPITCH(tex->surface.u.gfx9.surf.epitch));
3105 } else {
3106 /* Compute mutable surface parameters (SI-CI-VI). */
3107 const struct legacy_surf_level *level_info =
3108 &tex->surface.u.legacy.level[cb->base.u.tex.level];
3109 unsigned pitch_tile_max, slice_tile_max, tile_mode_index;
3110 unsigned cb_color_pitch, cb_color_slice, cb_color_fmask_slice;
3111
3112 cb_color_base += level_info->offset >> 8;
3113 /* Only macrotiled modes can set tile swizzle. */
3114 if (level_info->mode == RADEON_SURF_MODE_2D)
3115 cb_color_base |= tex->surface.tile_swizzle;
3116
3117 if (!tex->surface.fmask_size)
3118 cb_color_fmask = cb_color_base;
3119 if (cb->base.u.tex.level > 0)
3120 cb_color_cmask = cb_color_base;
3121 if (cb_dcc_base)
3122 cb_dcc_base += level_info->dcc_offset >> 8;
3123
3124 pitch_tile_max = level_info->nblk_x / 8 - 1;
3125 slice_tile_max = level_info->nblk_x *
3126 level_info->nblk_y / 64 - 1;
3127 tile_mode_index = si_tile_mode_index(tex, cb->base.u.tex.level, false);
3128
3129 cb_color_attrib |= S_028C74_TILE_MODE_INDEX(tile_mode_index);
3130 cb_color_pitch = S_028C64_TILE_MAX(pitch_tile_max);
3131 cb_color_slice = S_028C68_TILE_MAX(slice_tile_max);
3132
3133 if (tex->surface.fmask_size) {
3134 if (sctx->chip_class >= CIK)
3135 cb_color_pitch |= S_028C64_FMASK_TILE_MAX(tex->surface.u.legacy.fmask.pitch_in_pixels / 8 - 1);
3136 cb_color_attrib |= S_028C74_FMASK_TILE_MODE_INDEX(tex->surface.u.legacy.fmask.tiling_index);
3137 cb_color_fmask_slice = S_028C88_TILE_MAX(tex->surface.u.legacy.fmask.slice_tile_max);
3138 } else {
3139 /* This must be set for fast clear to work without FMASK. */
3140 if (sctx->chip_class >= CIK)
3141 cb_color_pitch |= S_028C64_FMASK_TILE_MAX(pitch_tile_max);
3142 cb_color_attrib |= S_028C74_FMASK_TILE_MODE_INDEX(tile_mode_index);
3143 cb_color_fmask_slice = S_028C88_TILE_MAX(slice_tile_max);
3144 }
3145
3146 radeon_set_context_reg_seq(cs, R_028C60_CB_COLOR0_BASE + i * 0x3C,
3147 sctx->chip_class >= VI ? 14 : 13);
3148 radeon_emit(cs, cb_color_base); /* CB_COLOR0_BASE */
3149 radeon_emit(cs, cb_color_pitch); /* CB_COLOR0_PITCH */
3150 radeon_emit(cs, cb_color_slice); /* CB_COLOR0_SLICE */
3151 radeon_emit(cs, cb->cb_color_view); /* CB_COLOR0_VIEW */
3152 radeon_emit(cs, cb_color_info); /* CB_COLOR0_INFO */
3153 radeon_emit(cs, cb_color_attrib); /* CB_COLOR0_ATTRIB */
3154 radeon_emit(cs, cb->cb_dcc_control); /* CB_COLOR0_DCC_CONTROL */
3155 radeon_emit(cs, cb_color_cmask); /* CB_COLOR0_CMASK */
3156 radeon_emit(cs, tex->surface.u.legacy.cmask_slice_tile_max); /* CB_COLOR0_CMASK_SLICE */
3157 radeon_emit(cs, cb_color_fmask); /* CB_COLOR0_FMASK */
3158 radeon_emit(cs, cb_color_fmask_slice); /* CB_COLOR0_FMASK_SLICE */
3159 radeon_emit(cs, tex->color_clear_value[0]); /* CB_COLOR0_CLEAR_WORD0 */
3160 radeon_emit(cs, tex->color_clear_value[1]); /* CB_COLOR0_CLEAR_WORD1 */
3161
3162 if (sctx->chip_class >= VI) /* R_028C94_CB_COLOR0_DCC_BASE */
3163 radeon_emit(cs, cb_dcc_base);
3164 }
3165 }
3166 for (; i < 8 ; i++)
3167 if (sctx->framebuffer.dirty_cbufs & (1 << i))
3168 radeon_set_context_reg(cs, R_028C70_CB_COLOR0_INFO + i * 0x3C, 0);
3169
3170 /* ZS buffer. */
3171 if (state->zsbuf && sctx->framebuffer.dirty_zsbuf) {
3172 struct si_surface *zb = (struct si_surface*)state->zsbuf;
3173 struct si_texture *tex = (struct si_texture*)zb->base.texture;
3174
3175 radeon_add_to_buffer_list(sctx, sctx->gfx_cs,
3176 &tex->buffer, RADEON_USAGE_READWRITE,
3177 zb->base.texture->nr_samples > 1 ?
3178 RADEON_PRIO_DEPTH_BUFFER_MSAA :
3179 RADEON_PRIO_DEPTH_BUFFER);
3180
3181 if (sctx->chip_class >= GFX9) {
3182 radeon_set_context_reg_seq(cs, R_028014_DB_HTILE_DATA_BASE, 3);
3183 radeon_emit(cs, zb->db_htile_data_base); /* DB_HTILE_DATA_BASE */
3184 radeon_emit(cs, S_028018_BASE_HI(zb->db_htile_data_base >> 32)); /* DB_HTILE_DATA_BASE_HI */
3185 radeon_emit(cs, zb->db_depth_size); /* DB_DEPTH_SIZE */
3186
3187 radeon_set_context_reg_seq(cs, R_028038_DB_Z_INFO, 10);
3188 radeon_emit(cs, zb->db_z_info | /* DB_Z_INFO */
3189 S_028038_ZRANGE_PRECISION(tex->depth_clear_value != 0));
3190 radeon_emit(cs, zb->db_stencil_info); /* DB_STENCIL_INFO */
3191 radeon_emit(cs, zb->db_depth_base); /* DB_Z_READ_BASE */
3192 radeon_emit(cs, S_028044_BASE_HI(zb->db_depth_base >> 32)); /* DB_Z_READ_BASE_HI */
3193 radeon_emit(cs, zb->db_stencil_base); /* DB_STENCIL_READ_BASE */
3194 radeon_emit(cs, S_02804C_BASE_HI(zb->db_stencil_base >> 32)); /* DB_STENCIL_READ_BASE_HI */
3195 radeon_emit(cs, zb->db_depth_base); /* DB_Z_WRITE_BASE */
3196 radeon_emit(cs, S_028054_BASE_HI(zb->db_depth_base >> 32)); /* DB_Z_WRITE_BASE_HI */
3197 radeon_emit(cs, zb->db_stencil_base); /* DB_STENCIL_WRITE_BASE */
3198 radeon_emit(cs, S_02805C_BASE_HI(zb->db_stencil_base >> 32)); /* DB_STENCIL_WRITE_BASE_HI */
3199
3200 radeon_set_context_reg_seq(cs, R_028068_DB_Z_INFO2, 2);
3201 radeon_emit(cs, zb->db_z_info2); /* DB_Z_INFO2 */
3202 radeon_emit(cs, zb->db_stencil_info2); /* DB_STENCIL_INFO2 */
3203 } else {
3204 radeon_set_context_reg(cs, R_028014_DB_HTILE_DATA_BASE, zb->db_htile_data_base);
3205
3206 radeon_set_context_reg_seq(cs, R_02803C_DB_DEPTH_INFO, 9);
3207 radeon_emit(cs, zb->db_depth_info); /* DB_DEPTH_INFO */
3208 radeon_emit(cs, zb->db_z_info | /* DB_Z_INFO */
3209 S_028040_ZRANGE_PRECISION(tex->depth_clear_value != 0));
3210 radeon_emit(cs, zb->db_stencil_info); /* DB_STENCIL_INFO */
3211 radeon_emit(cs, zb->db_depth_base); /* DB_Z_READ_BASE */
3212 radeon_emit(cs, zb->db_stencil_base); /* DB_STENCIL_READ_BASE */
3213 radeon_emit(cs, zb->db_depth_base); /* DB_Z_WRITE_BASE */
3214 radeon_emit(cs, zb->db_stencil_base); /* DB_STENCIL_WRITE_BASE */
3215 radeon_emit(cs, zb->db_depth_size); /* DB_DEPTH_SIZE */
3216 radeon_emit(cs, zb->db_depth_slice); /* DB_DEPTH_SLICE */
3217 }
3218
3219 radeon_set_context_reg_seq(cs, R_028028_DB_STENCIL_CLEAR, 2);
3220 radeon_emit(cs, tex->stencil_clear_value); /* R_028028_DB_STENCIL_CLEAR */
3221 radeon_emit(cs, fui(tex->depth_clear_value)); /* R_02802C_DB_DEPTH_CLEAR */
3222
3223 radeon_set_context_reg(cs, R_028008_DB_DEPTH_VIEW, zb->db_depth_view);
3224 radeon_set_context_reg(cs, R_028ABC_DB_HTILE_SURFACE, zb->db_htile_surface);
3225 } else if (sctx->framebuffer.dirty_zsbuf) {
3226 if (sctx->chip_class >= GFX9)
3227 radeon_set_context_reg_seq(cs, R_028038_DB_Z_INFO, 2);
3228 else
3229 radeon_set_context_reg_seq(cs, R_028040_DB_Z_INFO, 2);
3230
3231 radeon_emit(cs, S_028040_FORMAT(V_028040_Z_INVALID)); /* DB_Z_INFO */
3232 radeon_emit(cs, S_028044_FORMAT(V_028044_STENCIL_INVALID)); /* DB_STENCIL_INFO */
3233 }
3234
3235 /* Framebuffer dimensions. */
3236 /* PA_SC_WINDOW_SCISSOR_TL is set in si_init_config() */
3237 radeon_set_context_reg(cs, R_028208_PA_SC_WINDOW_SCISSOR_BR,
3238 S_028208_BR_X(state->width) | S_028208_BR_Y(state->height));
3239
3240 if (sctx->screen->dfsm_allowed) {
3241 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
3242 radeon_emit(cs, EVENT_TYPE(V_028A90_BREAK_BATCH) | EVENT_INDEX(0));
3243 }
3244
3245 sctx->framebuffer.dirty_cbufs = 0;
3246 sctx->framebuffer.dirty_zsbuf = false;
3247 }
3248
3249 static void si_emit_msaa_sample_locs(struct si_context *sctx)
3250 {
3251 struct radeon_cmdbuf *cs = sctx->gfx_cs;
3252 struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
3253 unsigned nr_samples = sctx->framebuffer.nr_samples;
3254 bool has_msaa_sample_loc_bug = sctx->screen->has_msaa_sample_loc_bug;
3255
3256 /* Smoothing (only possible with nr_samples == 1) uses the same
3257 * sample locations as the MSAA it simulates.
3258 */
3259 if (nr_samples <= 1 && sctx->smoothing_enabled)
3260 nr_samples = SI_NUM_SMOOTH_AA_SAMPLES;
3261
3262 /* On Polaris, the small primitive filter uses the sample locations
3263 * even when MSAA is off, so we need to make sure they're set to 0.
3264 */
3265 if (has_msaa_sample_loc_bug)
3266 nr_samples = MAX2(nr_samples, 1);
3267
3268 if (nr_samples != sctx->sample_locs_num_samples) {
3269 sctx->sample_locs_num_samples = nr_samples;
3270 si_emit_sample_locations(cs, nr_samples);
3271 }
3272
3273 if (sctx->family >= CHIP_POLARIS10) {
3274 unsigned small_prim_filter_cntl =
3275 S_028830_SMALL_PRIM_FILTER_ENABLE(1) |
3276 /* line bug */
3277 S_028830_LINE_FILTER_DISABLE(sctx->family <= CHIP_POLARIS12);
3278
3279 /* The alternative of setting sample locations to 0 would
3280 * require a DB flush to avoid Z errors, see
3281 * https://bugs.freedesktop.org/show_bug.cgi?id=96908
3282 */
3283 if (has_msaa_sample_loc_bug &&
3284 sctx->framebuffer.nr_samples > 1 &&
3285 !rs->multisample_enable)
3286 small_prim_filter_cntl &= C_028830_SMALL_PRIM_FILTER_ENABLE;
3287
3288 radeon_opt_set_context_reg(sctx,
3289 R_028830_PA_SU_SMALL_PRIM_FILTER_CNTL,
3290 SI_TRACKED_PA_SU_SMALL_PRIM_FILTER_CNTL,
3291 small_prim_filter_cntl);
3292 }
3293
3294 /* The exclusion bits can be set to improve rasterization efficiency
3295 * if no sample lies on the pixel boundary (-8 sample offset).
3296 */
3297 bool exclusion = sctx->chip_class >= CIK &&
3298 (!rs->multisample_enable || nr_samples != 16);
3299 radeon_opt_set_context_reg(sctx, R_02882C_PA_SU_PRIM_FILTER_CNTL,
3300 SI_TRACKED_PA_SU_PRIM_FILTER_CNTL,
3301 S_02882C_XMAX_RIGHT_EXCLUSION(exclusion) |
3302 S_02882C_YMAX_BOTTOM_EXCLUSION(exclusion));
3303 }
3304
3305 static bool si_out_of_order_rasterization(struct si_context *sctx)
3306 {
3307 struct si_state_blend *blend = sctx->queued.named.blend;
3308 struct si_state_dsa *dsa = sctx->queued.named.dsa;
3309
3310 if (!sctx->screen->has_out_of_order_rast)
3311 return false;
3312
3313 unsigned colormask = sctx->framebuffer.colorbuf_enabled_4bit;
3314
3315 if (blend) {
3316 colormask &= blend->cb_target_enabled_4bit;
3317 } else {
3318 colormask = 0;
3319 }
3320
3321 /* Conservative: No logic op. */
3322 if (colormask && blend->logicop_enable)
3323 return false;
3324
3325 struct si_dsa_order_invariance dsa_order_invariant = {
3326 .zs = true, .pass_set = true, .pass_last = false
3327 };
3328
3329 if (sctx->framebuffer.state.zsbuf) {
3330 struct si_texture *zstex =
3331 (struct si_texture*)sctx->framebuffer.state.zsbuf->texture;
3332 bool has_stencil = zstex->surface.has_stencil;
3333 dsa_order_invariant = dsa->order_invariance[has_stencil];
3334 if (!dsa_order_invariant.zs)
3335 return false;
3336
3337 /* The set of PS invocations is always order invariant,
3338 * except when early Z/S tests are requested. */
3339 if (sctx->ps_shader.cso &&
3340 sctx->ps_shader.cso->info.writes_memory &&
3341 sctx->ps_shader.cso->info.properties[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL] &&
3342 !dsa_order_invariant.pass_set)
3343 return false;
3344
3345 if (sctx->num_perfect_occlusion_queries != 0 &&
3346 !dsa_order_invariant.pass_set)
3347 return false;
3348 }
3349
3350 if (!colormask)
3351 return true;
3352
3353 unsigned blendmask = colormask & blend->blend_enable_4bit;
3354
3355 if (blendmask) {
3356 /* Only commutative blending. */
3357 if (blendmask & ~blend->commutative_4bit)
3358 return false;
3359
3360 if (!dsa_order_invariant.pass_set)
3361 return false;
3362 }
3363
3364 if (colormask & ~blendmask) {
3365 if (!dsa_order_invariant.pass_last)
3366 return false;
3367 }
3368
3369 return true;
3370 }
3371
3372 static void si_emit_msaa_config(struct si_context *sctx)
3373 {
3374 struct radeon_cmdbuf *cs = sctx->gfx_cs;
3375 unsigned num_tile_pipes = sctx->screen->info.num_tile_pipes;
3376 /* 33% faster rendering to linear color buffers */
3377 bool dst_is_linear = sctx->framebuffer.any_dst_linear;
3378 bool out_of_order_rast = si_out_of_order_rasterization(sctx);
3379 unsigned sc_mode_cntl_1 =
3380 S_028A4C_WALK_SIZE(dst_is_linear) |
3381 S_028A4C_WALK_FENCE_ENABLE(!dst_is_linear) |
3382 S_028A4C_WALK_FENCE_SIZE(num_tile_pipes == 2 ? 2 : 3) |
3383 S_028A4C_OUT_OF_ORDER_PRIMITIVE_ENABLE(out_of_order_rast) |
3384 S_028A4C_OUT_OF_ORDER_WATER_MARK(0x7) |
3385 /* always 1: */
3386 S_028A4C_WALK_ALIGN8_PRIM_FITS_ST(1) |
3387 S_028A4C_SUPERTILE_WALK_ORDER_ENABLE(1) |
3388 S_028A4C_TILE_WALK_ORDER_ENABLE(1) |
3389 S_028A4C_MULTI_SHADER_ENGINE_PRIM_DISCARD_ENABLE(1) |
3390 S_028A4C_FORCE_EOV_CNTDWN_ENABLE(1) |
3391 S_028A4C_FORCE_EOV_REZ_ENABLE(1);
3392 unsigned db_eqaa = S_028804_HIGH_QUALITY_INTERSECTIONS(1) |
3393 S_028804_INCOHERENT_EQAA_READS(1) |
3394 S_028804_INTERPOLATE_COMP_Z(1) |
3395 S_028804_STATIC_ANCHOR_ASSOCIATIONS(1);
3396 unsigned coverage_samples, color_samples, z_samples;
3397
3398 /* S: Coverage samples (up to 16x):
3399 * - Scan conversion samples (PA_SC_AA_CONFIG.MSAA_NUM_SAMPLES)
3400 * - CB FMASK samples (CB_COLORi_ATTRIB.NUM_SAMPLES)
3401 *
3402 * Z: Z/S samples (up to 8x, must be <= coverage samples and >= color samples):
3403 * - Value seen by DB (DB_Z_INFO.NUM_SAMPLES)
3404 * - Value seen by CB, must be correct even if Z/S is unbound (DB_EQAA.MAX_ANCHOR_SAMPLES)
3405 * # Missing samples are derived from Z planes if Z is compressed (up to 16x quality), or
3406 * # from the closest defined sample if Z is uncompressed (same quality as the number of
3407 * # Z samples).
3408 *
3409 * F: Color samples (up to 8x, must be <= coverage samples):
3410 * - CB color samples (CB_COLORi_ATTRIB.NUM_FRAGMENTS)
3411 * - PS iter samples (DB_EQAA.PS_ITER_SAMPLES)
3412 *
3413 * Can be anything between coverage and color samples:
3414 * - SampleMaskIn samples (PA_SC_AA_CONFIG.MSAA_EXPOSED_SAMPLES)
3415 * - SampleMaskOut samples (DB_EQAA.MASK_EXPORT_NUM_SAMPLES)
3416 * - Alpha-to-coverage samples (DB_EQAA.ALPHA_TO_MASK_NUM_SAMPLES)
3417 * - Occlusion query samples (DB_COUNT_CONTROL.SAMPLE_RATE)
3418 * # All are currently set the same as coverage samples.
3419 *
3420 * If color samples < coverage samples, FMASK has a higher bpp to store an "unknown"
3421 * flag for undefined color samples. A shader-based resolve must handle unknowns
3422 * or mask them out with AND. Unknowns can also be guessed from neighbors via
3423 * an edge-detect shader-based resolve, which is required to make "color samples = 1"
3424 * useful. The CB resolve always drops unknowns.
3425 *
3426 * Sensible AA configurations:
3427 * EQAA 16s 8z 8f - might look the same as 16x MSAA if Z is compressed
3428 * EQAA 16s 8z 4f - might look the same as 16x MSAA if Z is compressed
3429 * EQAA 16s 4z 4f - might look the same as 16x MSAA if Z is compressed
3430 * EQAA 8s 8z 8f = 8x MSAA
3431 * EQAA 8s 8z 4f - might look the same as 8x MSAA
3432 * EQAA 8s 8z 2f - might look the same as 8x MSAA with low-density geometry
3433 * EQAA 8s 4z 4f - might look the same as 8x MSAA if Z is compressed
3434 * EQAA 8s 4z 2f - might look the same as 8x MSAA with low-density geometry if Z is compressed
3435 * EQAA 4s 4z 4f = 4x MSAA
3436 * EQAA 4s 4z 2f - might look the same as 4x MSAA with low-density geometry
3437 * EQAA 2s 2z 2f = 2x MSAA
3438 */
3439 if (sctx->framebuffer.nr_samples > 1) {
3440 coverage_samples = sctx->framebuffer.nr_samples;
3441 color_samples = sctx->framebuffer.nr_color_samples;
3442
3443 if (sctx->framebuffer.state.zsbuf) {
3444 z_samples = sctx->framebuffer.state.zsbuf->texture->nr_samples;
3445 z_samples = MAX2(1, z_samples);
3446 } else {
3447 z_samples = coverage_samples;
3448 }
3449 } else if (sctx->smoothing_enabled) {
3450 coverage_samples = color_samples = z_samples = SI_NUM_SMOOTH_AA_SAMPLES;
3451 } else {
3452 coverage_samples = color_samples = z_samples = 1;
3453 }
3454
3455 /* Required by OpenGL line rasterization.
3456 *
3457 * TODO: We should also enable perpendicular endcaps for AA lines,
3458 * but that requires implementing line stippling in the pixel
3459 * shader. SC can only do line stippling with axis-aligned
3460 * endcaps.
3461 */
3462 unsigned sc_line_cntl = S_028BDC_DX10_DIAMOND_TEST_ENA(1);
3463 unsigned sc_aa_config = 0;
3464
3465 if (coverage_samples > 1) {
3466 /* distance from the pixel center, indexed by log2(nr_samples) */
3467 static unsigned max_dist[] = {
3468 0, /* unused */
3469 4, /* 2x MSAA */
3470 6, /* 4x MSAA */
3471 7, /* 8x MSAA */
3472 8, /* 16x MSAA */
3473 };
3474 unsigned log_samples = util_logbase2(coverage_samples);
3475 unsigned log_z_samples = util_logbase2(z_samples);
3476 unsigned ps_iter_samples = si_get_ps_iter_samples(sctx);
3477 unsigned log_ps_iter_samples = util_logbase2(ps_iter_samples);
3478
3479 sc_line_cntl |= S_028BDC_EXPAND_LINE_WIDTH(1);
3480 sc_aa_config = S_028BE0_MSAA_NUM_SAMPLES(log_samples) |
3481 S_028BE0_MAX_SAMPLE_DIST(max_dist[log_samples]) |
3482 S_028BE0_MSAA_EXPOSED_SAMPLES(log_samples);
3483
3484 if (sctx->framebuffer.nr_samples > 1) {
3485 db_eqaa |= S_028804_MAX_ANCHOR_SAMPLES(log_z_samples) |
3486 S_028804_PS_ITER_SAMPLES(log_ps_iter_samples) |
3487 S_028804_MASK_EXPORT_NUM_SAMPLES(log_samples) |
3488 S_028804_ALPHA_TO_MASK_NUM_SAMPLES(log_samples);
3489 sc_mode_cntl_1 |= S_028A4C_PS_ITER_SAMPLE(ps_iter_samples > 1);
3490 } else if (sctx->smoothing_enabled) {
3491 db_eqaa |= S_028804_OVERRASTERIZATION_AMOUNT(log_samples);
3492 }
3493 }
3494
3495 /* R_028BDC_PA_SC_LINE_CNTL, R_028BE0_PA_SC_AA_CONFIG */
3496 radeon_opt_set_context_reg2(sctx, R_028BDC_PA_SC_LINE_CNTL,
3497 SI_TRACKED_PA_SC_LINE_CNTL, sc_line_cntl,
3498 sc_aa_config);
3499 /* R_028804_DB_EQAA */
3500 radeon_opt_set_context_reg(sctx, R_028804_DB_EQAA, SI_TRACKED_DB_EQAA,
3501 db_eqaa);
3502 /* R_028A4C_PA_SC_MODE_CNTL_1 */
3503 radeon_opt_set_context_reg(sctx, R_028A4C_PA_SC_MODE_CNTL_1,
3504 SI_TRACKED_PA_SC_MODE_CNTL_1, sc_mode_cntl_1);
3505
3506 /* GFX9: Flush DFSM when the AA mode changes. */
3507 if (sctx->screen->dfsm_allowed) {
3508 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
3509 radeon_emit(cs, EVENT_TYPE(V_028A90_FLUSH_DFSM) | EVENT_INDEX(0));
3510 }
3511 }
3512
3513 void si_update_ps_iter_samples(struct si_context *sctx)
3514 {
3515 if (sctx->framebuffer.nr_samples > 1)
3516 si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
3517 if (sctx->screen->dpbb_allowed)
3518 si_mark_atom_dirty(sctx, &sctx->atoms.s.dpbb_state);
3519 }
3520
3521 static void si_set_min_samples(struct pipe_context *ctx, unsigned min_samples)
3522 {
3523 struct si_context *sctx = (struct si_context *)ctx;
3524
3525 /* The hardware can only do sample shading with 2^n samples. */
3526 min_samples = util_next_power_of_two(min_samples);
3527
3528 if (sctx->ps_iter_samples == min_samples)
3529 return;
3530
3531 sctx->ps_iter_samples = min_samples;
3532 sctx->do_update_shaders = true;
3533
3534 si_update_ps_iter_samples(sctx);
3535 }
3536
3537 /*
3538 * Samplers
3539 */
3540
3541 /**
3542 * Build the sampler view descriptor for a buffer texture.
3543 * @param state 256-bit descriptor; only the high 128 bits are filled in
3544 */
3545 void
3546 si_make_buffer_descriptor(struct si_screen *screen, struct r600_resource *buf,
3547 enum pipe_format format,
3548 unsigned offset, unsigned size,
3549 uint32_t *state)
3550 {
3551 const struct util_format_description *desc;
3552 int first_non_void;
3553 unsigned stride;
3554 unsigned num_records;
3555 unsigned num_format, data_format;
3556
3557 desc = util_format_description(format);
3558 first_non_void = util_format_get_first_non_void_channel(format);
3559 stride = desc->block.bits / 8;
3560 num_format = si_translate_buffer_numformat(&screen->b, desc, first_non_void);
3561 data_format = si_translate_buffer_dataformat(&screen->b, desc, first_non_void);
3562
3563 num_records = size / stride;
3564 num_records = MIN2(num_records, (buf->b.b.width0 - offset) / stride);
3565
3566 /* The NUM_RECORDS field has a different meaning depending on the chip,
3567 * instruction type, STRIDE, and SWIZZLE_ENABLE.
3568 *
3569 * SI-CIK:
3570 * - If STRIDE == 0, it's in byte units.
3571 * - If STRIDE != 0, it's in units of STRIDE, used with inst.IDXEN.
3572 *
3573 * VI:
3574 * - For SMEM and STRIDE == 0, it's in byte units.
3575 * - For SMEM and STRIDE != 0, it's in units of STRIDE.
3576 * - For VMEM and STRIDE == 0 or SWIZZLE_ENABLE == 0, it's in byte units.
3577 * - For VMEM and STRIDE != 0 and SWIZZLE_ENABLE == 1, it's in units of STRIDE.
3578 * NOTE: There is incompatibility between VMEM and SMEM opcodes due to SWIZZLE_-
3579 * ENABLE. The workaround is to set STRIDE = 0 if SWIZZLE_ENABLE == 0 when
3580 * using SMEM. This can be done in the shader by clearing STRIDE with s_and.
3581 * That way the same descriptor can be used by both SMEM and VMEM.
3582 *
3583 * GFX9:
3584 * - For SMEM and STRIDE == 0, it's in byte units.
3585 * - For SMEM and STRIDE != 0, it's in units of STRIDE.
3586 * - For VMEM and inst.IDXEN == 0 or STRIDE == 0, it's in byte units.
3587 * - For VMEM and inst.IDXEN == 1 and STRIDE != 0, it's in units of STRIDE.
3588 */
3589 if (screen->info.chip_class >= GFX9)
3590 /* When vindex == 0, LLVM sets IDXEN = 0, thus changing units
3591 * from STRIDE to bytes. This works around it by setting
3592 * NUM_RECORDS to at least the size of one element, so that
3593 * the first element is readable when IDXEN == 0.
3594 *
3595 * TODO: Fix this in LLVM, but do we need a new intrinsic where
3596 * IDXEN is enforced?
3597 */
3598 num_records = num_records ? MAX2(num_records, stride) : 0;
3599 else if (screen->info.chip_class == VI)
3600 num_records *= stride;
3601
3602 state[4] = 0;
3603 state[5] = S_008F04_STRIDE(stride);
3604 state[6] = num_records;
3605 state[7] = S_008F0C_DST_SEL_X(si_map_swizzle(desc->swizzle[0])) |
3606 S_008F0C_DST_SEL_Y(si_map_swizzle(desc->swizzle[1])) |
3607 S_008F0C_DST_SEL_Z(si_map_swizzle(desc->swizzle[2])) |
3608 S_008F0C_DST_SEL_W(si_map_swizzle(desc->swizzle[3])) |
3609 S_008F0C_NUM_FORMAT(num_format) |
3610 S_008F0C_DATA_FORMAT(data_format);
3611 }
3612
3613 static unsigned gfx9_border_color_swizzle(const unsigned char swizzle[4])
3614 {
3615 unsigned bc_swizzle = V_008F20_BC_SWIZZLE_XYZW;
3616
3617 if (swizzle[3] == PIPE_SWIZZLE_X) {
3618 /* For the pre-defined border color values (white, opaque
3619 * black, transparent black), the only thing that matters is
3620 * that the alpha channel winds up in the correct place
3621 * (because the RGB channels are all the same) so either of
3622 * these enumerations will work.
3623 */
3624 if (swizzle[2] == PIPE_SWIZZLE_Y)
3625 bc_swizzle = V_008F20_BC_SWIZZLE_WZYX;
3626 else
3627 bc_swizzle = V_008F20_BC_SWIZZLE_WXYZ;
3628 } else if (swizzle[0] == PIPE_SWIZZLE_X) {
3629 if (swizzle[1] == PIPE_SWIZZLE_Y)
3630 bc_swizzle = V_008F20_BC_SWIZZLE_XYZW;
3631 else
3632 bc_swizzle = V_008F20_BC_SWIZZLE_XWYZ;
3633 } else if (swizzle[1] == PIPE_SWIZZLE_X) {
3634 bc_swizzle = V_008F20_BC_SWIZZLE_YXWZ;
3635 } else if (swizzle[2] == PIPE_SWIZZLE_X) {
3636 bc_swizzle = V_008F20_BC_SWIZZLE_ZYXW;
3637 }
3638
3639 return bc_swizzle;
3640 }
3641
3642 /**
3643 * Build the sampler view descriptor for a texture.
3644 */
3645 void
3646 si_make_texture_descriptor(struct si_screen *screen,
3647 struct si_texture *tex,
3648 bool sampler,
3649 enum pipe_texture_target target,
3650 enum pipe_format pipe_format,
3651 const unsigned char state_swizzle[4],
3652 unsigned first_level, unsigned last_level,
3653 unsigned first_layer, unsigned last_layer,
3654 unsigned width, unsigned height, unsigned depth,
3655 uint32_t *state,
3656 uint32_t *fmask_state)
3657 {
3658 struct pipe_resource *res = &tex->buffer.b.b;
3659 const struct util_format_description *desc;
3660 unsigned char swizzle[4];
3661 int first_non_void;
3662 unsigned num_format, data_format, type, num_samples;
3663 uint64_t va;
3664
3665 desc = util_format_description(pipe_format);
3666
3667 num_samples = desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS ?
3668 MAX2(1, res->nr_samples) :
3669 MAX2(1, res->nr_storage_samples);
3670
3671 if (desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS) {
3672 const unsigned char swizzle_xxxx[4] = {0, 0, 0, 0};
3673 const unsigned char swizzle_yyyy[4] = {1, 1, 1, 1};
3674 const unsigned char swizzle_wwww[4] = {3, 3, 3, 3};
3675
3676 switch (pipe_format) {
3677 case PIPE_FORMAT_S8_UINT_Z24_UNORM:
3678 case PIPE_FORMAT_X32_S8X24_UINT:
3679 case PIPE_FORMAT_X8Z24_UNORM:
3680 util_format_compose_swizzles(swizzle_yyyy, state_swizzle, swizzle);
3681 break;
3682 case PIPE_FORMAT_X24S8_UINT:
3683 /*
3684 * X24S8 is implemented as an 8_8_8_8 data format, to
3685 * fix texture gathers. This affects at least
3686 * GL45-CTS.texture_cube_map_array.sampling on VI.
3687 */
3688 if (screen->info.chip_class <= VI)
3689 util_format_compose_swizzles(swizzle_wwww, state_swizzle, swizzle);
3690 else
3691 util_format_compose_swizzles(swizzle_yyyy, state_swizzle, swizzle);
3692 break;
3693 default:
3694 util_format_compose_swizzles(swizzle_xxxx, state_swizzle, swizzle);
3695 }
3696 } else {
3697 util_format_compose_swizzles(desc->swizzle, state_swizzle, swizzle);
3698 }
3699
3700 first_non_void = util_format_get_first_non_void_channel(pipe_format);
3701
3702 switch (pipe_format) {
3703 case PIPE_FORMAT_S8_UINT_Z24_UNORM:
3704 num_format = V_008F14_IMG_NUM_FORMAT_UNORM;
3705 break;
3706 default:
3707 if (first_non_void < 0) {
3708 if (util_format_is_compressed(pipe_format)) {
3709 switch (pipe_format) {
3710 case PIPE_FORMAT_DXT1_SRGB:
3711 case PIPE_FORMAT_DXT1_SRGBA:
3712 case PIPE_FORMAT_DXT3_SRGBA:
3713 case PIPE_FORMAT_DXT5_SRGBA:
3714 case PIPE_FORMAT_BPTC_SRGBA:
3715 case PIPE_FORMAT_ETC2_SRGB8:
3716 case PIPE_FORMAT_ETC2_SRGB8A1:
3717 case PIPE_FORMAT_ETC2_SRGBA8:
3718 num_format = V_008F14_IMG_NUM_FORMAT_SRGB;
3719 break;
3720 case PIPE_FORMAT_RGTC1_SNORM:
3721 case PIPE_FORMAT_LATC1_SNORM:
3722 case PIPE_FORMAT_RGTC2_SNORM:
3723 case PIPE_FORMAT_LATC2_SNORM:
3724 case PIPE_FORMAT_ETC2_R11_SNORM:
3725 case PIPE_FORMAT_ETC2_RG11_SNORM:
3726 /* implies float, so use SNORM/UNORM to determine
3727 whether data is signed or not */
3728 case PIPE_FORMAT_BPTC_RGB_FLOAT:
3729 num_format = V_008F14_IMG_NUM_FORMAT_SNORM;
3730 break;
3731 default:
3732 num_format = V_008F14_IMG_NUM_FORMAT_UNORM;
3733 break;
3734 }
3735 } else if (desc->layout == UTIL_FORMAT_LAYOUT_SUBSAMPLED) {
3736 num_format = V_008F14_IMG_NUM_FORMAT_UNORM;
3737 } else {
3738 num_format = V_008F14_IMG_NUM_FORMAT_FLOAT;
3739 }
3740 } else if (desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
3741 num_format = V_008F14_IMG_NUM_FORMAT_SRGB;
3742 } else {
3743 num_format = V_008F14_IMG_NUM_FORMAT_UNORM;
3744
3745 switch (desc->channel[first_non_void].type) {
3746 case UTIL_FORMAT_TYPE_FLOAT:
3747 num_format = V_008F14_IMG_NUM_FORMAT_FLOAT;
3748 break;
3749 case UTIL_FORMAT_TYPE_SIGNED:
3750 if (desc->channel[first_non_void].normalized)
3751 num_format = V_008F14_IMG_NUM_FORMAT_SNORM;
3752 else if (desc->channel[first_non_void].pure_integer)
3753 num_format = V_008F14_IMG_NUM_FORMAT_SINT;
3754 else
3755 num_format = V_008F14_IMG_NUM_FORMAT_SSCALED;
3756 break;
3757 case UTIL_FORMAT_TYPE_UNSIGNED:
3758 if (desc->channel[first_non_void].normalized)
3759 num_format = V_008F14_IMG_NUM_FORMAT_UNORM;
3760 else if (desc->channel[first_non_void].pure_integer)
3761 num_format = V_008F14_IMG_NUM_FORMAT_UINT;
3762 else
3763 num_format = V_008F14_IMG_NUM_FORMAT_USCALED;
3764 }
3765 }
3766 }
3767
3768 data_format = si_translate_texformat(&screen->b, pipe_format, desc, first_non_void);
3769 if (data_format == ~0) {
3770 data_format = 0;
3771 }
3772
3773 /* S8 with Z32 HTILE needs a special format. */
3774 if (screen->info.chip_class >= GFX9 &&
3775 pipe_format == PIPE_FORMAT_S8_UINT &&
3776 tex->tc_compatible_htile)
3777 data_format = V_008F14_IMG_DATA_FORMAT_S8_32;
3778
3779 if (!sampler &&
3780 (res->target == PIPE_TEXTURE_CUBE ||
3781 res->target == PIPE_TEXTURE_CUBE_ARRAY ||
3782 (screen->info.chip_class <= VI &&
3783 res->target == PIPE_TEXTURE_3D))) {
3784 /* For the purpose of shader images, treat cube maps and 3D
3785 * textures as 2D arrays. For 3D textures, the address
3786 * calculations for mipmaps are different, so we rely on the
3787 * caller to effectively disable mipmaps.
3788 */
3789 type = V_008F1C_SQ_RSRC_IMG_2D_ARRAY;
3790
3791 assert(res->target != PIPE_TEXTURE_3D || (first_level == 0 && last_level == 0));
3792 } else {
3793 type = si_tex_dim(screen, tex, target, num_samples);
3794 }
3795
3796 if (type == V_008F1C_SQ_RSRC_IMG_1D_ARRAY) {
3797 height = 1;
3798 depth = res->array_size;
3799 } else if (type == V_008F1C_SQ_RSRC_IMG_2D_ARRAY ||
3800 type == V_008F1C_SQ_RSRC_IMG_2D_MSAA_ARRAY) {
3801 if (sampler || res->target != PIPE_TEXTURE_3D)
3802 depth = res->array_size;
3803 } else if (type == V_008F1C_SQ_RSRC_IMG_CUBE)
3804 depth = res->array_size / 6;
3805
3806 state[0] = 0;
3807 state[1] = (S_008F14_DATA_FORMAT_GFX6(data_format) |
3808 S_008F14_NUM_FORMAT_GFX6(num_format));
3809 state[2] = (S_008F18_WIDTH(width - 1) |
3810 S_008F18_HEIGHT(height - 1) |
3811 S_008F18_PERF_MOD(4));
3812 state[3] = (S_008F1C_DST_SEL_X(si_map_swizzle(swizzle[0])) |
3813 S_008F1C_DST_SEL_Y(si_map_swizzle(swizzle[1])) |
3814 S_008F1C_DST_SEL_Z(si_map_swizzle(swizzle[2])) |
3815 S_008F1C_DST_SEL_W(si_map_swizzle(swizzle[3])) |
3816 S_008F1C_BASE_LEVEL(num_samples > 1 ? 0 : first_level) |
3817 S_008F1C_LAST_LEVEL(num_samples > 1 ?
3818 util_logbase2(num_samples) :
3819 last_level) |
3820 S_008F1C_TYPE(type));
3821 state[4] = 0;
3822 state[5] = S_008F24_BASE_ARRAY(first_layer);
3823 state[6] = 0;
3824 state[7] = 0;
3825
3826 if (screen->info.chip_class >= GFX9) {
3827 unsigned bc_swizzle = gfx9_border_color_swizzle(desc->swizzle);
3828
3829 /* Depth is the the last accessible layer on Gfx9.
3830 * The hw doesn't need to know the total number of layers.
3831 */
3832 if (type == V_008F1C_SQ_RSRC_IMG_3D)
3833 state[4] |= S_008F20_DEPTH(depth - 1);
3834 else
3835 state[4] |= S_008F20_DEPTH(last_layer);
3836
3837 state[4] |= S_008F20_BC_SWIZZLE(bc_swizzle);
3838 state[5] |= S_008F24_MAX_MIP(num_samples > 1 ?
3839 util_logbase2(num_samples) :
3840 tex->buffer.b.b.last_level);
3841 } else {
3842 state[3] |= S_008F1C_POW2_PAD(res->last_level > 0);
3843 state[4] |= S_008F20_DEPTH(depth - 1);
3844 state[5] |= S_008F24_LAST_ARRAY(last_layer);
3845 }
3846
3847 if (tex->dcc_offset) {
3848 state[6] = S_008F28_ALPHA_IS_ON_MSB(vi_alpha_is_on_msb(pipe_format));
3849 } else {
3850 /* The last dword is unused by hw. The shader uses it to clear
3851 * bits in the first dword of sampler state.
3852 */
3853 if (screen->info.chip_class <= CIK && res->nr_samples <= 1) {
3854 if (first_level == last_level)
3855 state[7] = C_008F30_MAX_ANISO_RATIO;
3856 else
3857 state[7] = 0xffffffff;
3858 }
3859 }
3860
3861 /* Initialize the sampler view for FMASK. */
3862 if (tex->surface.fmask_size) {
3863 uint32_t data_format, num_format;
3864
3865 va = tex->buffer.gpu_address + tex->fmask_offset;
3866
3867 #define FMASK(s,f) (((unsigned)(MAX2(1, s)) * 16) + (MAX2(1, f)))
3868 if (screen->info.chip_class >= GFX9) {
3869 data_format = V_008F14_IMG_DATA_FORMAT_FMASK;
3870 switch (FMASK(res->nr_samples, res->nr_storage_samples)) {
3871 case FMASK(2,1):
3872 num_format = V_008F14_IMG_FMASK_8_2_1;
3873 break;
3874 case FMASK(2,2):
3875 num_format = V_008F14_IMG_FMASK_8_2_2;
3876 break;
3877 case FMASK(4,1):
3878 num_format = V_008F14_IMG_FMASK_8_4_1;
3879 break;
3880 case FMASK(4,2):
3881 num_format = V_008F14_IMG_FMASK_8_4_2;
3882 break;
3883 case FMASK(4,4):
3884 num_format = V_008F14_IMG_FMASK_8_4_4;
3885 break;
3886 case FMASK(8,1):
3887 num_format = V_008F14_IMG_FMASK_8_8_1;
3888 break;
3889 case FMASK(8,2):
3890 num_format = V_008F14_IMG_FMASK_16_8_2;
3891 break;
3892 case FMASK(8,4):
3893 num_format = V_008F14_IMG_FMASK_32_8_4;
3894 break;
3895 case FMASK(8,8):
3896 num_format = V_008F14_IMG_FMASK_32_8_8;
3897 break;
3898 case FMASK(16,1):
3899 num_format = V_008F14_IMG_FMASK_16_16_1;
3900 break;
3901 case FMASK(16,2):
3902 num_format = V_008F14_IMG_FMASK_32_16_2;
3903 break;
3904 case FMASK(16,4):
3905 num_format = V_008F14_IMG_FMASK_64_16_4;
3906 break;
3907 case FMASK(16,8):
3908 num_format = V_008F14_IMG_FMASK_64_16_8;
3909 break;
3910 default:
3911 unreachable("invalid nr_samples");
3912 }
3913 } else {
3914 switch (FMASK(res->nr_samples, res->nr_storage_samples)) {
3915 case FMASK(2,1):
3916 data_format = V_008F14_IMG_DATA_FORMAT_FMASK8_S2_F1;
3917 break;
3918 case FMASK(2,2):
3919 data_format = V_008F14_IMG_DATA_FORMAT_FMASK8_S2_F2;
3920 break;
3921 case FMASK(4,1):
3922 data_format = V_008F14_IMG_DATA_FORMAT_FMASK8_S4_F1;
3923 break;
3924 case FMASK(4,2):
3925 data_format = V_008F14_IMG_DATA_FORMAT_FMASK8_S4_F2;
3926 break;
3927 case FMASK(4,4):
3928 data_format = V_008F14_IMG_DATA_FORMAT_FMASK8_S4_F4;
3929 break;
3930 case FMASK(8,1):
3931 data_format = V_008F14_IMG_DATA_FORMAT_FMASK8_S8_F1;
3932 break;
3933 case FMASK(8,2):
3934 data_format = V_008F14_IMG_DATA_FORMAT_FMASK16_S8_F2;
3935 break;
3936 case FMASK(8,4):
3937 data_format = V_008F14_IMG_DATA_FORMAT_FMASK32_S8_F4;
3938 break;
3939 case FMASK(8,8):
3940 data_format = V_008F14_IMG_DATA_FORMAT_FMASK32_S8_F8;
3941 break;
3942 case FMASK(16,1):
3943 data_format = V_008F14_IMG_DATA_FORMAT_FMASK16_S16_F1;
3944 break;
3945 case FMASK(16,2):
3946 data_format = V_008F14_IMG_DATA_FORMAT_FMASK32_S16_F2;
3947 break;
3948 case FMASK(16,4):
3949 data_format = V_008F14_IMG_DATA_FORMAT_FMASK64_S16_F4;
3950 break;
3951 case FMASK(16,8):
3952 data_format = V_008F14_IMG_DATA_FORMAT_FMASK64_S16_F8;
3953 break;
3954 default:
3955 unreachable("invalid nr_samples");
3956 }
3957 num_format = V_008F14_IMG_NUM_FORMAT_UINT;
3958 }
3959 #undef FMASK
3960
3961 fmask_state[0] = (va >> 8) | tex->surface.fmask_tile_swizzle;
3962 fmask_state[1] = S_008F14_BASE_ADDRESS_HI(va >> 40) |
3963 S_008F14_DATA_FORMAT_GFX6(data_format) |
3964 S_008F14_NUM_FORMAT_GFX6(num_format);
3965 fmask_state[2] = S_008F18_WIDTH(width - 1) |
3966 S_008F18_HEIGHT(height - 1);
3967 fmask_state[3] = S_008F1C_DST_SEL_X(V_008F1C_SQ_SEL_X) |
3968 S_008F1C_DST_SEL_Y(V_008F1C_SQ_SEL_X) |
3969 S_008F1C_DST_SEL_Z(V_008F1C_SQ_SEL_X) |
3970 S_008F1C_DST_SEL_W(V_008F1C_SQ_SEL_X) |
3971 S_008F1C_TYPE(si_tex_dim(screen, tex, target, 0));
3972 fmask_state[4] = 0;
3973 fmask_state[5] = S_008F24_BASE_ARRAY(first_layer);
3974 fmask_state[6] = 0;
3975 fmask_state[7] = 0;
3976
3977 if (screen->info.chip_class >= GFX9) {
3978 fmask_state[3] |= S_008F1C_SW_MODE(tex->surface.u.gfx9.fmask.swizzle_mode);
3979 fmask_state[4] |= S_008F20_DEPTH(last_layer) |
3980 S_008F20_PITCH_GFX9(tex->surface.u.gfx9.fmask.epitch);
3981 fmask_state[5] |= S_008F24_META_PIPE_ALIGNED(tex->surface.u.gfx9.cmask.pipe_aligned) |
3982 S_008F24_META_RB_ALIGNED(tex->surface.u.gfx9.cmask.rb_aligned);
3983 } else {
3984 fmask_state[3] |= S_008F1C_TILING_INDEX(tex->surface.u.legacy.fmask.tiling_index);
3985 fmask_state[4] |= S_008F20_DEPTH(depth - 1) |
3986 S_008F20_PITCH_GFX6(tex->surface.u.legacy.fmask.pitch_in_pixels - 1);
3987 fmask_state[5] |= S_008F24_LAST_ARRAY(last_layer);
3988 }
3989 }
3990 }
3991
3992 /**
3993 * Create a sampler view.
3994 *
3995 * @param ctx context
3996 * @param texture texture
3997 * @param state sampler view template
3998 * @param width0 width0 override (for compressed textures as int)
3999 * @param height0 height0 override (for compressed textures as int)
4000 * @param force_level set the base address to the level (for compressed textures)
4001 */
4002 struct pipe_sampler_view *
4003 si_create_sampler_view_custom(struct pipe_context *ctx,
4004 struct pipe_resource *texture,
4005 const struct pipe_sampler_view *state,
4006 unsigned width0, unsigned height0,
4007 unsigned force_level)
4008 {
4009 struct si_context *sctx = (struct si_context*)ctx;
4010 struct si_sampler_view *view = CALLOC_STRUCT(si_sampler_view);
4011 struct si_texture *tex = (struct si_texture*)texture;
4012 unsigned base_level, first_level, last_level;
4013 unsigned char state_swizzle[4];
4014 unsigned height, depth, width;
4015 unsigned last_layer = state->u.tex.last_layer;
4016 enum pipe_format pipe_format;
4017 const struct legacy_surf_level *surflevel;
4018
4019 if (!view)
4020 return NULL;
4021
4022 /* initialize base object */
4023 view->base = *state;
4024 view->base.texture = NULL;
4025 view->base.reference.count = 1;
4026 view->base.context = ctx;
4027
4028 assert(texture);
4029 pipe_resource_reference(&view->base.texture, texture);
4030
4031 if (state->format == PIPE_FORMAT_X24S8_UINT ||
4032 state->format == PIPE_FORMAT_S8X24_UINT ||
4033 state->format == PIPE_FORMAT_X32_S8X24_UINT ||
4034 state->format == PIPE_FORMAT_S8_UINT)
4035 view->is_stencil_sampler = true;
4036
4037 /* Buffer resource. */
4038 if (texture->target == PIPE_BUFFER) {
4039 si_make_buffer_descriptor(sctx->screen,
4040 r600_resource(texture),
4041 state->format,
4042 state->u.buf.offset,
4043 state->u.buf.size,
4044 view->state);
4045 return &view->base;
4046 }
4047
4048 state_swizzle[0] = state->swizzle_r;
4049 state_swizzle[1] = state->swizzle_g;
4050 state_swizzle[2] = state->swizzle_b;
4051 state_swizzle[3] = state->swizzle_a;
4052
4053 base_level = 0;
4054 first_level = state->u.tex.first_level;
4055 last_level = state->u.tex.last_level;
4056 width = width0;
4057 height = height0;
4058 depth = texture->depth0;
4059
4060 if (sctx->chip_class <= VI && force_level) {
4061 assert(force_level == first_level &&
4062 force_level == last_level);
4063 base_level = force_level;
4064 first_level = 0;
4065 last_level = 0;
4066 width = u_minify(width, force_level);
4067 height = u_minify(height, force_level);
4068 depth = u_minify(depth, force_level);
4069 }
4070
4071 /* This is not needed if state trackers set last_layer correctly. */
4072 if (state->target == PIPE_TEXTURE_1D ||
4073 state->target == PIPE_TEXTURE_2D ||
4074 state->target == PIPE_TEXTURE_RECT ||
4075 state->target == PIPE_TEXTURE_CUBE)
4076 last_layer = state->u.tex.first_layer;
4077
4078 /* Texturing with separate depth and stencil. */
4079 pipe_format = state->format;
4080
4081 /* Depth/stencil texturing sometimes needs separate texture. */
4082 if (tex->is_depth && !si_can_sample_zs(tex, view->is_stencil_sampler)) {
4083 if (!tex->flushed_depth_texture &&
4084 !si_init_flushed_depth_texture(ctx, texture, NULL)) {
4085 pipe_resource_reference(&view->base.texture, NULL);
4086 FREE(view);
4087 return NULL;
4088 }
4089
4090 assert(tex->flushed_depth_texture);
4091
4092 /* Override format for the case where the flushed texture
4093 * contains only Z or only S.
4094 */
4095 if (tex->flushed_depth_texture->buffer.b.b.format != tex->buffer.b.b.format)
4096 pipe_format = tex->flushed_depth_texture->buffer.b.b.format;
4097
4098 tex = tex->flushed_depth_texture;
4099 }
4100
4101 surflevel = tex->surface.u.legacy.level;
4102
4103 if (tex->db_compatible) {
4104 if (!view->is_stencil_sampler)
4105 pipe_format = tex->db_render_format;
4106
4107 switch (pipe_format) {
4108 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
4109 pipe_format = PIPE_FORMAT_Z32_FLOAT;
4110 break;
4111 case PIPE_FORMAT_X8Z24_UNORM:
4112 case PIPE_FORMAT_S8_UINT_Z24_UNORM:
4113 /* Z24 is always stored like this for DB
4114 * compatibility.
4115 */
4116 pipe_format = PIPE_FORMAT_Z24X8_UNORM;
4117 break;
4118 case PIPE_FORMAT_X24S8_UINT:
4119 case PIPE_FORMAT_S8X24_UINT:
4120 case PIPE_FORMAT_X32_S8X24_UINT:
4121 pipe_format = PIPE_FORMAT_S8_UINT;
4122 surflevel = tex->surface.u.legacy.stencil_level;
4123 break;
4124 default:;
4125 }
4126 }
4127
4128 view->dcc_incompatible =
4129 vi_dcc_formats_are_incompatible(texture,
4130 state->u.tex.first_level,
4131 state->format);
4132
4133 si_make_texture_descriptor(sctx->screen, tex, true,
4134 state->target, pipe_format, state_swizzle,
4135 first_level, last_level,
4136 state->u.tex.first_layer, last_layer,
4137 width, height, depth,
4138 view->state, view->fmask_state);
4139
4140 unsigned num_format = G_008F14_NUM_FORMAT_GFX6(view->state[1]);
4141 view->is_integer =
4142 num_format == V_008F14_IMG_NUM_FORMAT_USCALED ||
4143 num_format == V_008F14_IMG_NUM_FORMAT_SSCALED ||
4144 num_format == V_008F14_IMG_NUM_FORMAT_UINT ||
4145 num_format == V_008F14_IMG_NUM_FORMAT_SINT;
4146 view->base_level_info = &surflevel[base_level];
4147 view->base_level = base_level;
4148 view->block_width = util_format_get_blockwidth(pipe_format);
4149 return &view->base;
4150 }
4151
4152 static struct pipe_sampler_view *
4153 si_create_sampler_view(struct pipe_context *ctx,
4154 struct pipe_resource *texture,
4155 const struct pipe_sampler_view *state)
4156 {
4157 return si_create_sampler_view_custom(ctx, texture, state,
4158 texture ? texture->width0 : 0,
4159 texture ? texture->height0 : 0, 0);
4160 }
4161
4162 static void si_sampler_view_destroy(struct pipe_context *ctx,
4163 struct pipe_sampler_view *state)
4164 {
4165 struct si_sampler_view *view = (struct si_sampler_view *)state;
4166
4167 pipe_resource_reference(&state->texture, NULL);
4168 FREE(view);
4169 }
4170
4171 static bool wrap_mode_uses_border_color(unsigned wrap, bool linear_filter)
4172 {
4173 return wrap == PIPE_TEX_WRAP_CLAMP_TO_BORDER ||
4174 wrap == PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER ||
4175 (linear_filter &&
4176 (wrap == PIPE_TEX_WRAP_CLAMP ||
4177 wrap == PIPE_TEX_WRAP_MIRROR_CLAMP));
4178 }
4179
4180 static uint32_t si_translate_border_color(struct si_context *sctx,
4181 const struct pipe_sampler_state *state,
4182 const union pipe_color_union *color,
4183 bool is_integer)
4184 {
4185 bool linear_filter = state->min_img_filter != PIPE_TEX_FILTER_NEAREST ||
4186 state->mag_img_filter != PIPE_TEX_FILTER_NEAREST;
4187
4188 if (!wrap_mode_uses_border_color(state->wrap_s, linear_filter) &&
4189 !wrap_mode_uses_border_color(state->wrap_t, linear_filter) &&
4190 !wrap_mode_uses_border_color(state->wrap_r, linear_filter))
4191 return S_008F3C_BORDER_COLOR_TYPE(V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK);
4192
4193 #define simple_border_types(elt) \
4194 do { \
4195 if (color->elt[0] == 0 && color->elt[1] == 0 && \
4196 color->elt[2] == 0 && color->elt[3] == 0) \
4197 return S_008F3C_BORDER_COLOR_TYPE(V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK); \
4198 if (color->elt[0] == 0 && color->elt[1] == 0 && \
4199 color->elt[2] == 0 && color->elt[3] == 1) \
4200 return S_008F3C_BORDER_COLOR_TYPE(V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_BLACK); \
4201 if (color->elt[0] == 1 && color->elt[1] == 1 && \
4202 color->elt[2] == 1 && color->elt[3] == 1) \
4203 return S_008F3C_BORDER_COLOR_TYPE(V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_WHITE); \
4204 } while (false)
4205
4206 if (is_integer)
4207 simple_border_types(ui);
4208 else
4209 simple_border_types(f);
4210
4211 #undef simple_border_types
4212
4213 int i;
4214
4215 /* Check if the border has been uploaded already. */
4216 for (i = 0; i < sctx->border_color_count; i++)
4217 if (memcmp(&sctx->border_color_table[i], color,
4218 sizeof(*color)) == 0)
4219 break;
4220
4221 if (i >= SI_MAX_BORDER_COLORS) {
4222 /* Getting 4096 unique border colors is very unlikely. */
4223 fprintf(stderr, "radeonsi: The border color table is full. "
4224 "Any new border colors will be just black. "
4225 "Please file a bug.\n");
4226 return S_008F3C_BORDER_COLOR_TYPE(V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK);
4227 }
4228
4229 if (i == sctx->border_color_count) {
4230 /* Upload a new border color. */
4231 memcpy(&sctx->border_color_table[i], color,
4232 sizeof(*color));
4233 util_memcpy_cpu_to_le32(&sctx->border_color_map[i],
4234 color, sizeof(*color));
4235 sctx->border_color_count++;
4236 }
4237
4238 return S_008F3C_BORDER_COLOR_PTR(i) |
4239 S_008F3C_BORDER_COLOR_TYPE(V_008F3C_SQ_TEX_BORDER_COLOR_REGISTER);
4240 }
4241
4242 static inline int S_FIXED(float value, unsigned frac_bits)
4243 {
4244 return value * (1 << frac_bits);
4245 }
4246
4247 static inline unsigned si_tex_filter(unsigned filter, unsigned max_aniso)
4248 {
4249 if (filter == PIPE_TEX_FILTER_LINEAR)
4250 return max_aniso > 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_BILINEAR
4251 : V_008F38_SQ_TEX_XY_FILTER_BILINEAR;
4252 else
4253 return max_aniso > 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_POINT
4254 : V_008F38_SQ_TEX_XY_FILTER_POINT;
4255 }
4256
4257 static inline unsigned si_tex_aniso_filter(unsigned filter)
4258 {
4259 if (filter < 2)
4260 return 0;
4261 if (filter < 4)
4262 return 1;
4263 if (filter < 8)
4264 return 2;
4265 if (filter < 16)
4266 return 3;
4267 return 4;
4268 }
4269
4270 static void *si_create_sampler_state(struct pipe_context *ctx,
4271 const struct pipe_sampler_state *state)
4272 {
4273 struct si_context *sctx = (struct si_context *)ctx;
4274 struct si_screen *sscreen = sctx->screen;
4275 struct si_sampler_state *rstate = CALLOC_STRUCT(si_sampler_state);
4276 unsigned max_aniso = sscreen->force_aniso >= 0 ? sscreen->force_aniso
4277 : state->max_anisotropy;
4278 unsigned max_aniso_ratio = si_tex_aniso_filter(max_aniso);
4279 union pipe_color_union clamped_border_color;
4280
4281 if (!rstate) {
4282 return NULL;
4283 }
4284
4285 #ifdef DEBUG
4286 rstate->magic = SI_SAMPLER_STATE_MAGIC;
4287 #endif
4288 rstate->val[0] = (S_008F30_CLAMP_X(si_tex_wrap(state->wrap_s)) |
4289 S_008F30_CLAMP_Y(si_tex_wrap(state->wrap_t)) |
4290 S_008F30_CLAMP_Z(si_tex_wrap(state->wrap_r)) |
4291 S_008F30_MAX_ANISO_RATIO(max_aniso_ratio) |
4292 S_008F30_DEPTH_COMPARE_FUNC(si_tex_compare(state->compare_func)) |
4293 S_008F30_FORCE_UNNORMALIZED(!state->normalized_coords) |
4294 S_008F30_ANISO_THRESHOLD(max_aniso_ratio >> 1) |
4295 S_008F30_ANISO_BIAS(max_aniso_ratio) |
4296 S_008F30_DISABLE_CUBE_WRAP(!state->seamless_cube_map) |
4297 S_008F30_COMPAT_MODE(sctx->chip_class >= VI));
4298 rstate->val[1] = (S_008F34_MIN_LOD(S_FIXED(CLAMP(state->min_lod, 0, 15), 8)) |
4299 S_008F34_MAX_LOD(S_FIXED(CLAMP(state->max_lod, 0, 15), 8)) |
4300 S_008F34_PERF_MIP(max_aniso_ratio ? max_aniso_ratio + 6 : 0));
4301 rstate->val[2] = (S_008F38_LOD_BIAS(S_FIXED(CLAMP(state->lod_bias, -16, 16), 8)) |
4302 S_008F38_XY_MAG_FILTER(si_tex_filter(state->mag_img_filter, max_aniso)) |
4303 S_008F38_XY_MIN_FILTER(si_tex_filter(state->min_img_filter, max_aniso)) |
4304 S_008F38_MIP_FILTER(si_tex_mipfilter(state->min_mip_filter)) |
4305 S_008F38_MIP_POINT_PRECLAMP(0) |
4306 S_008F38_DISABLE_LSB_CEIL(sctx->chip_class <= VI) |
4307 S_008F38_FILTER_PREC_FIX(1) |
4308 S_008F38_ANISO_OVERRIDE(sctx->chip_class >= VI));
4309 rstate->val[3] = si_translate_border_color(sctx, state, &state->border_color, false);
4310
4311 /* Create sampler resource for integer textures. */
4312 memcpy(rstate->integer_val, rstate->val, sizeof(rstate->val));
4313 rstate->integer_val[3] = si_translate_border_color(sctx, state, &state->border_color, true);
4314
4315 /* Create sampler resource for upgraded depth textures. */
4316 memcpy(rstate->upgraded_depth_val, rstate->val, sizeof(rstate->val));
4317
4318 for (unsigned i = 0; i < 4; ++i) {
4319 /* Use channel 0 on purpose, so that we can use OPAQUE_WHITE
4320 * when the border color is 1.0. */
4321 clamped_border_color.f[i] = CLAMP(state->border_color.f[0], 0, 1);
4322 }
4323
4324 if (memcmp(&state->border_color, &clamped_border_color, sizeof(clamped_border_color)) == 0)
4325 rstate->upgraded_depth_val[3] |= S_008F3C_UPGRADED_DEPTH(1);
4326 else
4327 rstate->upgraded_depth_val[3] =
4328 si_translate_border_color(sctx, state, &clamped_border_color, false) |
4329 S_008F3C_UPGRADED_DEPTH(1);
4330
4331 return rstate;
4332 }
4333
4334 static void si_set_sample_mask(struct pipe_context *ctx, unsigned sample_mask)
4335 {
4336 struct si_context *sctx = (struct si_context *)ctx;
4337
4338 if (sctx->sample_mask == (uint16_t)sample_mask)
4339 return;
4340
4341 sctx->sample_mask = sample_mask;
4342 si_mark_atom_dirty(sctx, &sctx->atoms.s.sample_mask);
4343 }
4344
4345 static void si_emit_sample_mask(struct si_context *sctx)
4346 {
4347 struct radeon_cmdbuf *cs = sctx->gfx_cs;
4348 unsigned mask = sctx->sample_mask;
4349
4350 /* Needed for line and polygon smoothing as well as for the Polaris
4351 * small primitive filter. We expect the state tracker to take care of
4352 * this for us.
4353 */
4354 assert(mask == 0xffff || sctx->framebuffer.nr_samples > 1 ||
4355 (mask & 1 && sctx->blitter->running));
4356
4357 radeon_set_context_reg_seq(cs, R_028C38_PA_SC_AA_MASK_X0Y0_X1Y0, 2);
4358 radeon_emit(cs, mask | (mask << 16));
4359 radeon_emit(cs, mask | (mask << 16));
4360 }
4361
4362 static void si_delete_sampler_state(struct pipe_context *ctx, void *state)
4363 {
4364 #ifdef DEBUG
4365 struct si_sampler_state *s = state;
4366
4367 assert(s->magic == SI_SAMPLER_STATE_MAGIC);
4368 s->magic = 0;
4369 #endif
4370 free(state);
4371 }
4372
4373 /*
4374 * Vertex elements & buffers
4375 */
4376
4377 static void *si_create_vertex_elements(struct pipe_context *ctx,
4378 unsigned count,
4379 const struct pipe_vertex_element *elements)
4380 {
4381 struct si_screen *sscreen = (struct si_screen*)ctx->screen;
4382 struct si_vertex_elements *v = CALLOC_STRUCT(si_vertex_elements);
4383 bool used[SI_NUM_VERTEX_BUFFERS] = {};
4384 int i;
4385
4386 assert(count <= SI_MAX_ATTRIBS);
4387 if (!v)
4388 return NULL;
4389
4390 v->count = count;
4391 v->desc_list_byte_size = align(count * 16, SI_CPDMA_ALIGNMENT);
4392
4393 for (i = 0; i < count; ++i) {
4394 const struct util_format_description *desc;
4395 const struct util_format_channel_description *channel;
4396 unsigned data_format, num_format;
4397 int first_non_void;
4398 unsigned vbo_index = elements[i].vertex_buffer_index;
4399 unsigned char swizzle[4];
4400
4401 if (vbo_index >= SI_NUM_VERTEX_BUFFERS) {
4402 FREE(v);
4403 return NULL;
4404 }
4405
4406 if (elements[i].instance_divisor) {
4407 v->uses_instance_divisors = true;
4408 v->instance_divisors[i] = elements[i].instance_divisor;
4409
4410 if (v->instance_divisors[i] == 1)
4411 v->instance_divisor_is_one |= 1u << i;
4412 else
4413 v->instance_divisor_is_fetched |= 1u << i;
4414 }
4415
4416 if (!used[vbo_index]) {
4417 v->first_vb_use_mask |= 1 << i;
4418 used[vbo_index] = true;
4419 }
4420
4421 desc = util_format_description(elements[i].src_format);
4422 first_non_void = util_format_get_first_non_void_channel(elements[i].src_format);
4423 data_format = si_translate_buffer_dataformat(ctx->screen, desc, first_non_void);
4424 num_format = si_translate_buffer_numformat(ctx->screen, desc, first_non_void);
4425 channel = first_non_void >= 0 ? &desc->channel[first_non_void] : NULL;
4426 memcpy(swizzle, desc->swizzle, sizeof(swizzle));
4427
4428 v->format_size[i] = desc->block.bits / 8;
4429 v->src_offset[i] = elements[i].src_offset;
4430 v->vertex_buffer_index[i] = vbo_index;
4431
4432 /* The hardware always treats the 2-bit alpha channel as
4433 * unsigned, so a shader workaround is needed. The affected
4434 * chips are VI and older except Stoney (GFX8.1).
4435 */
4436 if (data_format == V_008F0C_BUF_DATA_FORMAT_2_10_10_10 &&
4437 sscreen->info.chip_class <= VI &&
4438 sscreen->info.family != CHIP_STONEY) {
4439 if (num_format == V_008F0C_BUF_NUM_FORMAT_SNORM) {
4440 v->fix_fetch[i] = SI_FIX_FETCH_A2_SNORM;
4441 } else if (num_format == V_008F0C_BUF_NUM_FORMAT_SSCALED) {
4442 v->fix_fetch[i] = SI_FIX_FETCH_A2_SSCALED;
4443 } else if (num_format == V_008F0C_BUF_NUM_FORMAT_SINT) {
4444 /* This isn't actually used in OpenGL. */
4445 v->fix_fetch[i] = SI_FIX_FETCH_A2_SINT;
4446 }
4447 } else if (channel && channel->type == UTIL_FORMAT_TYPE_FIXED) {
4448 if (desc->swizzle[3] == PIPE_SWIZZLE_1)
4449 v->fix_fetch[i] = SI_FIX_FETCH_RGBX_32_FIXED;
4450 else
4451 v->fix_fetch[i] = SI_FIX_FETCH_RGBA_32_FIXED;
4452 } else if (channel && channel->size == 32 && !channel->pure_integer) {
4453 if (channel->type == UTIL_FORMAT_TYPE_SIGNED) {
4454 if (channel->normalized) {
4455 if (desc->swizzle[3] == PIPE_SWIZZLE_1)
4456 v->fix_fetch[i] = SI_FIX_FETCH_RGBX_32_SNORM;
4457 else
4458 v->fix_fetch[i] = SI_FIX_FETCH_RGBA_32_SNORM;
4459 } else {
4460 v->fix_fetch[i] = SI_FIX_FETCH_RGBA_32_SSCALED;
4461 }
4462 } else if (channel->type == UTIL_FORMAT_TYPE_UNSIGNED) {
4463 if (channel->normalized) {
4464 if (desc->swizzle[3] == PIPE_SWIZZLE_1)
4465 v->fix_fetch[i] = SI_FIX_FETCH_RGBX_32_UNORM;
4466 else
4467 v->fix_fetch[i] = SI_FIX_FETCH_RGBA_32_UNORM;
4468 } else {
4469 v->fix_fetch[i] = SI_FIX_FETCH_RGBA_32_USCALED;
4470 }
4471 }
4472 } else if (channel && channel->size == 64 &&
4473 channel->type == UTIL_FORMAT_TYPE_FLOAT) {
4474 switch (desc->nr_channels) {
4475 case 1:
4476 case 2:
4477 v->fix_fetch[i] = SI_FIX_FETCH_RG_64_FLOAT;
4478 swizzle[0] = PIPE_SWIZZLE_X;
4479 swizzle[1] = PIPE_SWIZZLE_Y;
4480 swizzle[2] = desc->nr_channels == 2 ? PIPE_SWIZZLE_Z : PIPE_SWIZZLE_0;
4481 swizzle[3] = desc->nr_channels == 2 ? PIPE_SWIZZLE_W : PIPE_SWIZZLE_0;
4482 break;
4483 case 3:
4484 v->fix_fetch[i] = SI_FIX_FETCH_RGB_64_FLOAT;
4485 swizzle[0] = PIPE_SWIZZLE_X; /* 3 loads */
4486 swizzle[1] = PIPE_SWIZZLE_Y;
4487 swizzle[2] = PIPE_SWIZZLE_0;
4488 swizzle[3] = PIPE_SWIZZLE_0;
4489 break;
4490 case 4:
4491 v->fix_fetch[i] = SI_FIX_FETCH_RGBA_64_FLOAT;
4492 swizzle[0] = PIPE_SWIZZLE_X; /* 2 loads */
4493 swizzle[1] = PIPE_SWIZZLE_Y;
4494 swizzle[2] = PIPE_SWIZZLE_Z;
4495 swizzle[3] = PIPE_SWIZZLE_W;
4496 break;
4497 default:
4498 assert(0);
4499 }
4500 } else if (channel && desc->nr_channels == 3) {
4501 assert(desc->swizzle[0] == PIPE_SWIZZLE_X);
4502
4503 if (channel->size == 8) {
4504 if (channel->pure_integer)
4505 v->fix_fetch[i] = SI_FIX_FETCH_RGB_8_INT;
4506 else
4507 v->fix_fetch[i] = SI_FIX_FETCH_RGB_8;
4508 } else if (channel->size == 16) {
4509 if (channel->pure_integer)
4510 v->fix_fetch[i] = SI_FIX_FETCH_RGB_16_INT;
4511 else
4512 v->fix_fetch[i] = SI_FIX_FETCH_RGB_16;
4513 }
4514 }
4515
4516 v->rsrc_word3[i] = S_008F0C_DST_SEL_X(si_map_swizzle(swizzle[0])) |
4517 S_008F0C_DST_SEL_Y(si_map_swizzle(swizzle[1])) |
4518 S_008F0C_DST_SEL_Z(si_map_swizzle(swizzle[2])) |
4519 S_008F0C_DST_SEL_W(si_map_swizzle(swizzle[3])) |
4520 S_008F0C_NUM_FORMAT(num_format) |
4521 S_008F0C_DATA_FORMAT(data_format);
4522 }
4523 return v;
4524 }
4525
4526 static void si_bind_vertex_elements(struct pipe_context *ctx, void *state)
4527 {
4528 struct si_context *sctx = (struct si_context *)ctx;
4529 struct si_vertex_elements *old = sctx->vertex_elements;
4530 struct si_vertex_elements *v = (struct si_vertex_elements*)state;
4531
4532 sctx->vertex_elements = v;
4533 sctx->vertex_buffers_dirty = true;
4534
4535 if (v &&
4536 (!old ||
4537 old->count != v->count ||
4538 old->uses_instance_divisors != v->uses_instance_divisors ||
4539 v->uses_instance_divisors || /* we don't check which divisors changed */
4540 memcmp(old->fix_fetch, v->fix_fetch, sizeof(v->fix_fetch[0]) * v->count)))
4541 sctx->do_update_shaders = true;
4542
4543 if (v && v->instance_divisor_is_fetched) {
4544 struct pipe_constant_buffer cb;
4545
4546 cb.buffer = NULL;
4547 cb.user_buffer = v->instance_divisors;
4548 cb.buffer_offset = 0;
4549 cb.buffer_size = sizeof(uint32_t) * v->count;
4550 si_set_rw_buffer(sctx, SI_VS_CONST_INSTANCE_DIVISORS, &cb);
4551 }
4552 }
4553
4554 static void si_delete_vertex_element(struct pipe_context *ctx, void *state)
4555 {
4556 struct si_context *sctx = (struct si_context *)ctx;
4557
4558 if (sctx->vertex_elements == state)
4559 sctx->vertex_elements = NULL;
4560 FREE(state);
4561 }
4562
4563 static void si_set_vertex_buffers(struct pipe_context *ctx,
4564 unsigned start_slot, unsigned count,
4565 const struct pipe_vertex_buffer *buffers)
4566 {
4567 struct si_context *sctx = (struct si_context *)ctx;
4568 struct pipe_vertex_buffer *dst = sctx->vertex_buffer + start_slot;
4569 int i;
4570
4571 assert(start_slot + count <= ARRAY_SIZE(sctx->vertex_buffer));
4572
4573 if (buffers) {
4574 for (i = 0; i < count; i++) {
4575 const struct pipe_vertex_buffer *src = buffers + i;
4576 struct pipe_vertex_buffer *dsti = dst + i;
4577 struct pipe_resource *buf = src->buffer.resource;
4578
4579 pipe_resource_reference(&dsti->buffer.resource, buf);
4580 dsti->buffer_offset = src->buffer_offset;
4581 dsti->stride = src->stride;
4582 si_context_add_resource_size(sctx, buf);
4583 if (buf)
4584 r600_resource(buf)->bind_history |= PIPE_BIND_VERTEX_BUFFER;
4585 }
4586 } else {
4587 for (i = 0; i < count; i++) {
4588 pipe_resource_reference(&dst[i].buffer.resource, NULL);
4589 }
4590 }
4591 sctx->vertex_buffers_dirty = true;
4592 }
4593
4594 /*
4595 * Misc
4596 */
4597
4598 static void si_set_tess_state(struct pipe_context *ctx,
4599 const float default_outer_level[4],
4600 const float default_inner_level[2])
4601 {
4602 struct si_context *sctx = (struct si_context *)ctx;
4603 struct pipe_constant_buffer cb;
4604 float array[8];
4605
4606 memcpy(array, default_outer_level, sizeof(float) * 4);
4607 memcpy(array+4, default_inner_level, sizeof(float) * 2);
4608
4609 cb.buffer = NULL;
4610 cb.user_buffer = NULL;
4611 cb.buffer_size = sizeof(array);
4612
4613 si_upload_const_buffer(sctx, (struct r600_resource**)&cb.buffer,
4614 (void*)array, sizeof(array),
4615 &cb.buffer_offset);
4616
4617 si_set_rw_buffer(sctx, SI_HS_CONST_DEFAULT_TESS_LEVELS, &cb);
4618 pipe_resource_reference(&cb.buffer, NULL);
4619 }
4620
4621 static void si_texture_barrier(struct pipe_context *ctx, unsigned flags)
4622 {
4623 struct si_context *sctx = (struct si_context *)ctx;
4624
4625 si_update_fb_dirtiness_after_rendering(sctx);
4626
4627 /* Multisample surfaces are flushed in si_decompress_textures. */
4628 if (sctx->framebuffer.uncompressed_cb_mask)
4629 si_make_CB_shader_coherent(sctx, sctx->framebuffer.nr_samples,
4630 sctx->framebuffer.CB_has_shader_readable_metadata);
4631 }
4632
4633 /* This only ensures coherency for shader image/buffer stores. */
4634 static void si_memory_barrier(struct pipe_context *ctx, unsigned flags)
4635 {
4636 struct si_context *sctx = (struct si_context *)ctx;
4637
4638 /* Subsequent commands must wait for all shader invocations to
4639 * complete. */
4640 sctx->flags |= SI_CONTEXT_PS_PARTIAL_FLUSH |
4641 SI_CONTEXT_CS_PARTIAL_FLUSH;
4642
4643 if (flags & PIPE_BARRIER_CONSTANT_BUFFER)
4644 sctx->flags |= SI_CONTEXT_INV_SMEM_L1 |
4645 SI_CONTEXT_INV_VMEM_L1;
4646
4647 if (flags & (PIPE_BARRIER_VERTEX_BUFFER |
4648 PIPE_BARRIER_SHADER_BUFFER |
4649 PIPE_BARRIER_TEXTURE |
4650 PIPE_BARRIER_IMAGE |
4651 PIPE_BARRIER_STREAMOUT_BUFFER |
4652 PIPE_BARRIER_GLOBAL_BUFFER)) {
4653 /* As far as I can tell, L1 contents are written back to L2
4654 * automatically at end of shader, but the contents of other
4655 * L1 caches might still be stale. */
4656 sctx->flags |= SI_CONTEXT_INV_VMEM_L1;
4657 }
4658
4659 if (flags & PIPE_BARRIER_INDEX_BUFFER) {
4660 /* Indices are read through TC L2 since VI.
4661 * L1 isn't used.
4662 */
4663 if (sctx->screen->info.chip_class <= CIK)
4664 sctx->flags |= SI_CONTEXT_WRITEBACK_GLOBAL_L2;
4665 }
4666
4667 /* MSAA color, any depth and any stencil are flushed in
4668 * si_decompress_textures when needed.
4669 */
4670 if (flags & PIPE_BARRIER_FRAMEBUFFER &&
4671 sctx->framebuffer.uncompressed_cb_mask) {
4672 sctx->flags |= SI_CONTEXT_FLUSH_AND_INV_CB;
4673
4674 if (sctx->chip_class <= VI)
4675 sctx->flags |= SI_CONTEXT_WRITEBACK_GLOBAL_L2;
4676 }
4677
4678 /* Indirect buffers use TC L2 on GFX9, but not older hw. */
4679 if (sctx->screen->info.chip_class <= VI &&
4680 flags & PIPE_BARRIER_INDIRECT_BUFFER)
4681 sctx->flags |= SI_CONTEXT_WRITEBACK_GLOBAL_L2;
4682 }
4683
4684 static void *si_create_blend_custom(struct si_context *sctx, unsigned mode)
4685 {
4686 struct pipe_blend_state blend;
4687
4688 memset(&blend, 0, sizeof(blend));
4689 blend.independent_blend_enable = true;
4690 blend.rt[0].colormask = 0xf;
4691 return si_create_blend_state_mode(&sctx->b, &blend, mode);
4692 }
4693
4694 static void si_init_config(struct si_context *sctx);
4695
4696 void si_init_state_functions(struct si_context *sctx)
4697 {
4698 sctx->atoms.s.framebuffer.emit = si_emit_framebuffer_state;
4699 sctx->atoms.s.msaa_sample_locs.emit = si_emit_msaa_sample_locs;
4700 sctx->atoms.s.db_render_state.emit = si_emit_db_render_state;
4701 sctx->atoms.s.dpbb_state.emit = si_emit_dpbb_state;
4702 sctx->atoms.s.msaa_config.emit = si_emit_msaa_config;
4703 sctx->atoms.s.sample_mask.emit = si_emit_sample_mask;
4704 sctx->atoms.s.cb_render_state.emit = si_emit_cb_render_state;
4705 sctx->atoms.s.blend_color.emit = si_emit_blend_color;
4706 sctx->atoms.s.clip_regs.emit = si_emit_clip_regs;
4707 sctx->atoms.s.clip_state.emit = si_emit_clip_state;
4708 sctx->atoms.s.stencil_ref.emit = si_emit_stencil_ref;
4709
4710 sctx->b.create_blend_state = si_create_blend_state;
4711 sctx->b.bind_blend_state = si_bind_blend_state;
4712 sctx->b.delete_blend_state = si_delete_blend_state;
4713 sctx->b.set_blend_color = si_set_blend_color;
4714
4715 sctx->b.create_rasterizer_state = si_create_rs_state;
4716 sctx->b.bind_rasterizer_state = si_bind_rs_state;
4717 sctx->b.delete_rasterizer_state = si_delete_rs_state;
4718
4719 sctx->b.create_depth_stencil_alpha_state = si_create_dsa_state;
4720 sctx->b.bind_depth_stencil_alpha_state = si_bind_dsa_state;
4721 sctx->b.delete_depth_stencil_alpha_state = si_delete_dsa_state;
4722
4723 sctx->custom_dsa_flush = si_create_db_flush_dsa(sctx);
4724 sctx->custom_blend_resolve = si_create_blend_custom(sctx, V_028808_CB_RESOLVE);
4725 sctx->custom_blend_fmask_decompress = si_create_blend_custom(sctx, V_028808_CB_FMASK_DECOMPRESS);
4726 sctx->custom_blend_eliminate_fastclear = si_create_blend_custom(sctx, V_028808_CB_ELIMINATE_FAST_CLEAR);
4727 sctx->custom_blend_dcc_decompress = si_create_blend_custom(sctx, V_028808_CB_DCC_DECOMPRESS);
4728
4729 sctx->b.set_clip_state = si_set_clip_state;
4730 sctx->b.set_stencil_ref = si_set_stencil_ref;
4731
4732 sctx->b.set_framebuffer_state = si_set_framebuffer_state;
4733
4734 sctx->b.create_sampler_state = si_create_sampler_state;
4735 sctx->b.delete_sampler_state = si_delete_sampler_state;
4736
4737 sctx->b.create_sampler_view = si_create_sampler_view;
4738 sctx->b.sampler_view_destroy = si_sampler_view_destroy;
4739
4740 sctx->b.set_sample_mask = si_set_sample_mask;
4741
4742 sctx->b.create_vertex_elements_state = si_create_vertex_elements;
4743 sctx->b.bind_vertex_elements_state = si_bind_vertex_elements;
4744 sctx->b.delete_vertex_elements_state = si_delete_vertex_element;
4745 sctx->b.set_vertex_buffers = si_set_vertex_buffers;
4746
4747 sctx->b.texture_barrier = si_texture_barrier;
4748 sctx->b.memory_barrier = si_memory_barrier;
4749 sctx->b.set_min_samples = si_set_min_samples;
4750 sctx->b.set_tess_state = si_set_tess_state;
4751
4752 sctx->b.set_active_query_state = si_set_active_query_state;
4753
4754 sctx->b.draw_vbo = si_draw_vbo;
4755
4756 si_init_config(sctx);
4757 }
4758
4759 void si_init_screen_state_functions(struct si_screen *sscreen)
4760 {
4761 sscreen->b.is_format_supported = si_is_format_supported;
4762 }
4763
4764 static void si_set_grbm_gfx_index(struct si_context *sctx,
4765 struct si_pm4_state *pm4, unsigned value)
4766 {
4767 unsigned reg = sctx->chip_class >= CIK ? R_030800_GRBM_GFX_INDEX :
4768 R_00802C_GRBM_GFX_INDEX;
4769 si_pm4_set_reg(pm4, reg, value);
4770 }
4771
4772 static void si_set_grbm_gfx_index_se(struct si_context *sctx,
4773 struct si_pm4_state *pm4, unsigned se)
4774 {
4775 assert(se == ~0 || se < sctx->screen->info.max_se);
4776 si_set_grbm_gfx_index(sctx, pm4,
4777 (se == ~0 ? S_030800_SE_BROADCAST_WRITES(1) :
4778 S_030800_SE_INDEX(se)) |
4779 S_030800_SH_BROADCAST_WRITES(1) |
4780 S_030800_INSTANCE_BROADCAST_WRITES(1));
4781 }
4782
4783 static void
4784 si_write_harvested_raster_configs(struct si_context *sctx,
4785 struct si_pm4_state *pm4,
4786 unsigned raster_config,
4787 unsigned raster_config_1)
4788 {
4789 unsigned num_se = MAX2(sctx->screen->info.max_se, 1);
4790 unsigned raster_config_se[4];
4791 unsigned se;
4792
4793 ac_get_harvested_configs(&sctx->screen->info,
4794 raster_config,
4795 &raster_config_1,
4796 raster_config_se);
4797
4798 for (se = 0; se < num_se; se++) {
4799 si_set_grbm_gfx_index_se(sctx, pm4, se);
4800 si_pm4_set_reg(pm4, R_028350_PA_SC_RASTER_CONFIG, raster_config_se[se]);
4801 }
4802 si_set_grbm_gfx_index(sctx, pm4, ~0);
4803
4804 if (sctx->chip_class >= CIK) {
4805 si_pm4_set_reg(pm4, R_028354_PA_SC_RASTER_CONFIG_1, raster_config_1);
4806 }
4807 }
4808
4809 static void si_set_raster_config(struct si_context *sctx, struct si_pm4_state *pm4)
4810 {
4811 struct si_screen *sscreen = sctx->screen;
4812 unsigned num_rb = MIN2(sscreen->info.num_render_backends, 16);
4813 unsigned rb_mask = sscreen->info.enabled_rb_mask;
4814 unsigned raster_config = sscreen->pa_sc_raster_config;
4815 unsigned raster_config_1 = sscreen->pa_sc_raster_config_1;
4816
4817 if (!rb_mask || util_bitcount(rb_mask) >= num_rb) {
4818 /* Always use the default config when all backends are enabled
4819 * (or when we failed to determine the enabled backends).
4820 */
4821 si_pm4_set_reg(pm4, R_028350_PA_SC_RASTER_CONFIG,
4822 raster_config);
4823 if (sctx->chip_class >= CIK)
4824 si_pm4_set_reg(pm4, R_028354_PA_SC_RASTER_CONFIG_1,
4825 raster_config_1);
4826 } else {
4827 si_write_harvested_raster_configs(sctx, pm4, raster_config, raster_config_1);
4828 }
4829 }
4830
4831 static void si_init_config(struct si_context *sctx)
4832 {
4833 struct si_screen *sscreen = sctx->screen;
4834 uint64_t border_color_va = sctx->border_color_buffer->gpu_address;
4835 bool has_clear_state = sscreen->has_clear_state;
4836 struct si_pm4_state *pm4 = CALLOC_STRUCT(si_pm4_state);
4837
4838 /* Only SI can disable CLEAR_STATE for now. */
4839 assert(has_clear_state || sscreen->info.chip_class == SI);
4840
4841 if (!pm4)
4842 return;
4843
4844 si_pm4_cmd_begin(pm4, PKT3_CONTEXT_CONTROL);
4845 si_pm4_cmd_add(pm4, CONTEXT_CONTROL_LOAD_ENABLE(1));
4846 si_pm4_cmd_add(pm4, CONTEXT_CONTROL_SHADOW_ENABLE(1));
4847 si_pm4_cmd_end(pm4, false);
4848
4849 if (has_clear_state) {
4850 si_pm4_cmd_begin(pm4, PKT3_CLEAR_STATE);
4851 si_pm4_cmd_add(pm4, 0);
4852 si_pm4_cmd_end(pm4, false);
4853 }
4854
4855 if (sctx->chip_class <= VI)
4856 si_set_raster_config(sctx, pm4);
4857
4858 si_pm4_set_reg(pm4, R_028A18_VGT_HOS_MAX_TESS_LEVEL, fui(64));
4859 if (!has_clear_state)
4860 si_pm4_set_reg(pm4, R_028A1C_VGT_HOS_MIN_TESS_LEVEL, fui(0));
4861
4862 /* FIXME calculate these values somehow ??? */
4863 if (sctx->chip_class <= VI) {
4864 si_pm4_set_reg(pm4, R_028A54_VGT_GS_PER_ES, SI_GS_PER_ES);
4865 si_pm4_set_reg(pm4, R_028A58_VGT_ES_PER_GS, 0x40);
4866 }
4867
4868 if (!has_clear_state) {
4869 si_pm4_set_reg(pm4, R_028A5C_VGT_GS_PER_VS, 0x2);
4870 si_pm4_set_reg(pm4, R_028A8C_VGT_PRIMITIVEID_RESET, 0x0);
4871 si_pm4_set_reg(pm4, R_028B98_VGT_STRMOUT_BUFFER_CONFIG, 0x0);
4872 }
4873
4874 si_pm4_set_reg(pm4, R_028AA0_VGT_INSTANCE_STEP_RATE_0, 1);
4875 if (!has_clear_state)
4876 si_pm4_set_reg(pm4, R_028AB8_VGT_VTX_CNT_EN, 0x0);
4877 if (sctx->chip_class < CIK)
4878 si_pm4_set_reg(pm4, R_008A14_PA_CL_ENHANCE, S_008A14_NUM_CLIP_SEQ(3) |
4879 S_008A14_CLIP_VTX_REORDER_ENA(1));
4880
4881 /* CLEAR_STATE doesn't clear these correctly on certain generations.
4882 * I don't know why. Deduced by trial and error.
4883 */
4884 if (sctx->chip_class <= CIK) {
4885 si_pm4_set_reg(pm4, R_028B28_VGT_STRMOUT_DRAW_OPAQUE_OFFSET, 0);
4886 si_pm4_set_reg(pm4, R_028204_PA_SC_WINDOW_SCISSOR_TL, S_028204_WINDOW_OFFSET_DISABLE(1));
4887 si_pm4_set_reg(pm4, R_028240_PA_SC_GENERIC_SCISSOR_TL, S_028240_WINDOW_OFFSET_DISABLE(1));
4888 si_pm4_set_reg(pm4, R_028244_PA_SC_GENERIC_SCISSOR_BR,
4889 S_028244_BR_X(16384) | S_028244_BR_Y(16384));
4890 si_pm4_set_reg(pm4, R_028030_PA_SC_SCREEN_SCISSOR_TL, 0);
4891 si_pm4_set_reg(pm4, R_028034_PA_SC_SCREEN_SCISSOR_BR,
4892 S_028034_BR_X(16384) | S_028034_BR_Y(16384));
4893 }
4894
4895 if (!has_clear_state) {
4896 si_pm4_set_reg(pm4, R_028230_PA_SC_EDGERULE,
4897 S_028230_ER_TRI(0xA) |
4898 S_028230_ER_POINT(0xA) |
4899 S_028230_ER_RECT(0xA) |
4900 /* Required by DX10_DIAMOND_TEST_ENA: */
4901 S_028230_ER_LINE_LR(0x1A) |
4902 S_028230_ER_LINE_RL(0x26) |
4903 S_028230_ER_LINE_TB(0xA) |
4904 S_028230_ER_LINE_BT(0xA));
4905 si_pm4_set_reg(pm4, R_028820_PA_CL_NANINF_CNTL, 0);
4906 si_pm4_set_reg(pm4, R_028AC0_DB_SRESULTS_COMPARE_STATE0, 0x0);
4907 si_pm4_set_reg(pm4, R_028AC4_DB_SRESULTS_COMPARE_STATE1, 0x0);
4908 si_pm4_set_reg(pm4, R_028AC8_DB_PRELOAD_CONTROL, 0x0);
4909 si_pm4_set_reg(pm4, R_02800C_DB_RENDER_OVERRIDE, 0);
4910 }
4911
4912 if (sctx->chip_class >= GFX9) {
4913 si_pm4_set_reg(pm4, R_030920_VGT_MAX_VTX_INDX, ~0);
4914 si_pm4_set_reg(pm4, R_030924_VGT_MIN_VTX_INDX, 0);
4915 si_pm4_set_reg(pm4, R_030928_VGT_INDX_OFFSET, 0);
4916 } else {
4917 /* These registers, when written, also overwrite the CLEAR_STATE
4918 * context, so we can't rely on CLEAR_STATE setting them.
4919 * It would be an issue if there was another UMD changing them.
4920 */
4921 si_pm4_set_reg(pm4, R_028400_VGT_MAX_VTX_INDX, ~0);
4922 si_pm4_set_reg(pm4, R_028404_VGT_MIN_VTX_INDX, 0);
4923 si_pm4_set_reg(pm4, R_028408_VGT_INDX_OFFSET, 0);
4924 }
4925
4926 if (sctx->chip_class >= CIK) {
4927 if (sctx->chip_class >= GFX9) {
4928 si_pm4_set_reg(pm4, R_00B41C_SPI_SHADER_PGM_RSRC3_HS,
4929 S_00B41C_CU_EN(0xffff) | S_00B41C_WAVE_LIMIT(0x3F));
4930 } else {
4931 si_pm4_set_reg(pm4, R_00B51C_SPI_SHADER_PGM_RSRC3_LS,
4932 S_00B51C_CU_EN(0xffff) | S_00B51C_WAVE_LIMIT(0x3F));
4933 si_pm4_set_reg(pm4, R_00B41C_SPI_SHADER_PGM_RSRC3_HS,
4934 S_00B41C_WAVE_LIMIT(0x3F));
4935 si_pm4_set_reg(pm4, R_00B31C_SPI_SHADER_PGM_RSRC3_ES,
4936 S_00B31C_CU_EN(0xffff) | S_00B31C_WAVE_LIMIT(0x3F));
4937
4938 /* If this is 0, Bonaire can hang even if GS isn't being used.
4939 * Other chips are unaffected. These are suboptimal values,
4940 * but we don't use on-chip GS.
4941 */
4942 si_pm4_set_reg(pm4, R_028A44_VGT_GS_ONCHIP_CNTL,
4943 S_028A44_ES_VERTS_PER_SUBGRP(64) |
4944 S_028A44_GS_PRIMS_PER_SUBGRP(4));
4945 }
4946 si_pm4_set_reg(pm4, R_00B21C_SPI_SHADER_PGM_RSRC3_GS,
4947 S_00B21C_CU_EN(0xffff) | S_00B21C_WAVE_LIMIT(0x3F));
4948
4949 /* Compute LATE_ALLOC_VS.LIMIT. */
4950 unsigned num_cu_per_sh = sscreen->info.num_good_cu_per_sh;
4951 unsigned late_alloc_limit; /* The limit is per SH. */
4952
4953 if (sctx->family == CHIP_KABINI) {
4954 late_alloc_limit = 0; /* Potential hang on Kabini. */
4955 } else if (num_cu_per_sh <= 4) {
4956 /* Too few available compute units per SH. Disallowing
4957 * VS to run on one CU could hurt us more than late VS
4958 * allocation would help.
4959 *
4960 * 2 is the highest safe number that allows us to keep
4961 * all CUs enabled.
4962 */
4963 late_alloc_limit = 2;
4964 } else {
4965 /* This is a good initial value, allowing 1 late_alloc
4966 * wave per SIMD on num_cu - 2.
4967 */
4968 late_alloc_limit = (num_cu_per_sh - 2) * 4;
4969
4970 /* The limit is 0-based, so 0 means 1. */
4971 assert(late_alloc_limit > 0 && late_alloc_limit <= 64);
4972 late_alloc_limit -= 1;
4973 }
4974
4975 /* VS can't execute on one CU if the limit is > 2. */
4976 si_pm4_set_reg(pm4, R_00B118_SPI_SHADER_PGM_RSRC3_VS,
4977 S_00B118_CU_EN(late_alloc_limit > 2 ? 0xfffe : 0xffff) |
4978 S_00B118_WAVE_LIMIT(0x3F));
4979 si_pm4_set_reg(pm4, R_00B11C_SPI_SHADER_LATE_ALLOC_VS,
4980 S_00B11C_LIMIT(late_alloc_limit));
4981 si_pm4_set_reg(pm4, R_00B01C_SPI_SHADER_PGM_RSRC3_PS,
4982 S_00B01C_CU_EN(0xffff) | S_00B01C_WAVE_LIMIT(0x3F));
4983 }
4984
4985 if (sctx->chip_class >= VI) {
4986 unsigned vgt_tess_distribution;
4987
4988 vgt_tess_distribution =
4989 S_028B50_ACCUM_ISOLINE(32) |
4990 S_028B50_ACCUM_TRI(11) |
4991 S_028B50_ACCUM_QUAD(11) |
4992 S_028B50_DONUT_SPLIT(16);
4993
4994 /* Testing with Unigine Heaven extreme tesselation yielded best results
4995 * with TRAP_SPLIT = 3.
4996 */
4997 if (sctx->family == CHIP_FIJI ||
4998 sctx->family >= CHIP_POLARIS10)
4999 vgt_tess_distribution |= S_028B50_TRAP_SPLIT(3);
5000
5001 si_pm4_set_reg(pm4, R_028B50_VGT_TESS_DISTRIBUTION, vgt_tess_distribution);
5002 } else if (!has_clear_state) {
5003 si_pm4_set_reg(pm4, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL, 14);
5004 si_pm4_set_reg(pm4, R_028C5C_VGT_OUT_DEALLOC_CNTL, 16);
5005 }
5006
5007 si_pm4_set_reg(pm4, R_028080_TA_BC_BASE_ADDR, border_color_va >> 8);
5008 if (sctx->chip_class >= CIK) {
5009 si_pm4_set_reg(pm4, R_028084_TA_BC_BASE_ADDR_HI,
5010 S_028084_ADDRESS(border_color_va >> 40));
5011 }
5012 si_pm4_add_bo(pm4, sctx->border_color_buffer, RADEON_USAGE_READ,
5013 RADEON_PRIO_BORDER_COLORS);
5014
5015 if (sctx->chip_class >= GFX9) {
5016 unsigned num_se = sscreen->info.max_se;
5017 unsigned pc_lines = 0;
5018
5019 switch (sctx->family) {
5020 case CHIP_VEGA10:
5021 case CHIP_VEGA12:
5022 case CHIP_VEGA20:
5023 pc_lines = 4096;
5024 break;
5025 case CHIP_RAVEN:
5026 pc_lines = 1024;
5027 break;
5028 default:
5029 assert(0);
5030 }
5031
5032 si_pm4_set_reg(pm4, R_028C48_PA_SC_BINNER_CNTL_1,
5033 S_028C48_MAX_ALLOC_COUNT(MIN2(128, pc_lines / (4 * num_se))) |
5034 S_028C48_MAX_PRIM_PER_BATCH(1023));
5035 si_pm4_set_reg(pm4, R_028C4C_PA_SC_CONSERVATIVE_RASTERIZATION_CNTL,
5036 S_028C4C_NULL_SQUAD_AA_MASK_ENABLE(1));
5037 si_pm4_set_reg(pm4, R_030968_VGT_INSTANCE_BASE_ID, 0);
5038 }
5039
5040 si_pm4_upload_indirect_buffer(sctx, pm4);
5041 sctx->init_config = pm4;
5042 }