projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
radeonsi: DCC MSAA blending bug - include logic op, limit to Navi14 and older
[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 "sid.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 #include "util/fast_idiv_by_const.h"
36
37 struct gfx10_format {
38 unsigned img_format:9;
39
40 /* Various formats are only supported with workarounds for vertex fetch,
41 * and some 32_32_32 formats are supported natively, but only for buffers
42 * (possibly with some image support, actually, but no filtering). */
43 bool buffers_only:1;
44 };
45
46 #include "gfx10_format_table.h"
47
48 static unsigned si_map_swizzle(unsigned swizzle)
49 {
50 switch (swizzle) {
51 case PIPE_SWIZZLE_Y:
52 return V_008F0C_SQ_SEL_Y;
53 case PIPE_SWIZZLE_Z:
54 return V_008F0C_SQ_SEL_Z;
55 case PIPE_SWIZZLE_W:
56 return V_008F0C_SQ_SEL_W;
57 case PIPE_SWIZZLE_0:
58 return V_008F0C_SQ_SEL_0;
59 case PIPE_SWIZZLE_1:
60 return V_008F0C_SQ_SEL_1;
61 default: /* PIPE_SWIZZLE_X */
62 return V_008F0C_SQ_SEL_X;
63 }
64 }
65
66 /* 12.4 fixed-point */
67 static unsigned si_pack_float_12p4(float x)
68 {
69 return x <= 0 ? 0 :
70 x >= 4096 ? 0xffff : x * 16;
71 }
72
73 /*
74 * Inferred framebuffer and blender state.
75 *
76 * CB_TARGET_MASK is emitted here to avoid a hang with dual source blending
77 * if there is not enough PS outputs.
78 */
79 static void si_emit_cb_render_state(struct si_context *sctx)
80 {
81 struct radeon_cmdbuf *cs = sctx->gfx_cs;
82 struct si_state_blend *blend = sctx->queued.named.blend;
83 /* CB_COLORn_INFO.FORMAT=INVALID should disable unbound colorbuffers,
84 * but you never know. */
85 uint32_t cb_target_mask = sctx->framebuffer.colorbuf_enabled_4bit;
86 unsigned i;
87
88 if (blend)
89 cb_target_mask &= blend->cb_target_mask;
90
91 /* Avoid a hang that happens when dual source blending is enabled
92 * but there is not enough color outputs. This is undefined behavior,
93 * so disable color writes completely.
94 *
95 * Reproducible with Unigine Heaven 4.0 and drirc missing.
96 */
97 if (blend && blend->dual_src_blend &&
98 sctx->ps_shader.cso &&
99 (sctx->ps_shader.cso->info.colors_written & 0x3) != 0x3)
100 cb_target_mask = 0;
101
102 /* GFX9: Flush DFSM when CB_TARGET_MASK changes.
103 * I think we don't have to do anything between IBs.
104 */
105 if (sctx->screen->dpbb_allowed &&
106 sctx->last_cb_target_mask != cb_target_mask) {
107 sctx->last_cb_target_mask = cb_target_mask;
108
109 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
110 radeon_emit(cs, EVENT_TYPE(V_028A90_BREAK_BATCH) | EVENT_INDEX(0));
111 }
112
113 unsigned initial_cdw = cs->current.cdw;
114 radeon_opt_set_context_reg(sctx, R_028238_CB_TARGET_MASK,
115 SI_TRACKED_CB_TARGET_MASK, cb_target_mask);
116
117 if (sctx->chip_class >= GFX8) {
118 /* DCC MSAA workaround.
119 * Alternatively, we can set CB_COLORi_DCC_CONTROL.OVERWRITE_-
120 * COMBINER_DISABLE, but that would be more complicated.
121 */
122 bool oc_disable = blend &&
123 blend->dcc_msaa_corruption_4bit & cb_target_mask &&
124 sctx->framebuffer.nr_samples >= 2;
125 unsigned watermark = sctx->framebuffer.dcc_overwrite_combiner_watermark;
126
127 radeon_opt_set_context_reg(
128 sctx, R_028424_CB_DCC_CONTROL,
129 SI_TRACKED_CB_DCC_CONTROL,
130 S_028424_OVERWRITE_COMBINER_MRT_SHARING_DISABLE(sctx->chip_class <= GFX9) |
131 S_028424_OVERWRITE_COMBINER_WATERMARK(watermark) |
132 S_028424_OVERWRITE_COMBINER_DISABLE(oc_disable) |
133 S_028424_DISABLE_CONSTANT_ENCODE_REG(sctx->screen->has_dcc_constant_encode));
134 }
135
136 /* RB+ register settings. */
137 if (sctx->screen->rbplus_allowed) {
138 unsigned spi_shader_col_format =
139 sctx->ps_shader.cso ?
140 sctx->ps_shader.current->key.part.ps.epilog.spi_shader_col_format : 0;
141 unsigned sx_ps_downconvert = 0;
142 unsigned sx_blend_opt_epsilon = 0;
143 unsigned sx_blend_opt_control = 0;
144
145 for (i = 0; i < sctx->framebuffer.state.nr_cbufs; i++) {
146 struct si_surface *surf =
147 (struct si_surface*)sctx->framebuffer.state.cbufs[i];
148 unsigned format, swap, spi_format, colormask;
149 bool has_alpha, has_rgb;
150
151 if (!surf) {
152 /* If the color buffer is not set, the driver sets 32_R
153 * as the SPI color format, because the hw doesn't allow
154 * holes between color outputs, so also set this to
155 * enable RB+.
156 */
157 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_32_R << (i * 4);
158 continue;
159 }
160
161 format = G_028C70_FORMAT(surf->cb_color_info);
162 swap = G_028C70_COMP_SWAP(surf->cb_color_info);
163 spi_format = (spi_shader_col_format >> (i * 4)) & 0xf;
164 colormask = (cb_target_mask >> (i * 4)) & 0xf;
165
166 /* Set if RGB and A are present. */
167 has_alpha = !G_028C74_FORCE_DST_ALPHA_1(surf->cb_color_attrib);
168
169 if (format == V_028C70_COLOR_8 ||
170 format == V_028C70_COLOR_16 ||
171 format == V_028C70_COLOR_32)
172 has_rgb = !has_alpha;
173 else
174 has_rgb = true;
175
176 /* Check the colormask and export format. */
177 if (!(colormask & (PIPE_MASK_RGBA & ~PIPE_MASK_A)))
178 has_rgb = false;
179 if (!(colormask & PIPE_MASK_A))
180 has_alpha = false;
181
182 if (spi_format == V_028714_SPI_SHADER_ZERO) {
183 has_rgb = false;
184 has_alpha = false;
185 }
186
187 /* Disable value checking for disabled channels. */
188 if (!has_rgb)
189 sx_blend_opt_control |= S_02875C_MRT0_COLOR_OPT_DISABLE(1) << (i * 4);
190 if (!has_alpha)
191 sx_blend_opt_control |= S_02875C_MRT0_ALPHA_OPT_DISABLE(1) << (i * 4);
192
193 /* Enable down-conversion for 32bpp and smaller formats. */
194 switch (format) {
195 case V_028C70_COLOR_8:
196 case V_028C70_COLOR_8_8:
197 case V_028C70_COLOR_8_8_8_8:
198 /* For 1 and 2-channel formats, use the superset thereof. */
199 if (spi_format == V_028714_SPI_SHADER_FP16_ABGR ||
200 spi_format == V_028714_SPI_SHADER_UINT16_ABGR ||
201 spi_format == V_028714_SPI_SHADER_SINT16_ABGR) {
202 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_8_8_8_8 << (i * 4);
203 sx_blend_opt_epsilon |= V_028758_8BIT_FORMAT << (i * 4);
204 }
205 break;
206
207 case V_028C70_COLOR_5_6_5:
208 if (spi_format == V_028714_SPI_SHADER_FP16_ABGR) {
209 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_5_6_5 << (i * 4);
210 sx_blend_opt_epsilon |= V_028758_6BIT_FORMAT << (i * 4);
211 }
212 break;
213
214 case V_028C70_COLOR_1_5_5_5:
215 if (spi_format == V_028714_SPI_SHADER_FP16_ABGR) {
216 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_1_5_5_5 << (i * 4);
217 sx_blend_opt_epsilon |= V_028758_5BIT_FORMAT << (i * 4);
218 }
219 break;
220
221 case V_028C70_COLOR_4_4_4_4:
222 if (spi_format == V_028714_SPI_SHADER_FP16_ABGR) {
223 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_4_4_4_4 << (i * 4);
224 sx_blend_opt_epsilon |= V_028758_4BIT_FORMAT << (i * 4);
225 }
226 break;
227
228 case V_028C70_COLOR_32:
229 if (swap == V_028C70_SWAP_STD &&
230 spi_format == V_028714_SPI_SHADER_32_R)
231 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_32_R << (i * 4);
232 else if (swap == V_028C70_SWAP_ALT_REV &&
233 spi_format == V_028714_SPI_SHADER_32_AR)
234 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_32_A << (i * 4);
235 break;
236
237 case V_028C70_COLOR_16:
238 case V_028C70_COLOR_16_16:
239 /* For 1-channel formats, use the superset thereof. */
240 if (spi_format == V_028714_SPI_SHADER_UNORM16_ABGR ||
241 spi_format == V_028714_SPI_SHADER_SNORM16_ABGR ||
242 spi_format == V_028714_SPI_SHADER_UINT16_ABGR ||
243 spi_format == V_028714_SPI_SHADER_SINT16_ABGR) {
244 if (swap == V_028C70_SWAP_STD ||
245 swap == V_028C70_SWAP_STD_REV)
246 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_16_16_GR << (i * 4);
247 else
248 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_16_16_AR << (i * 4);
249 }
250 break;
251
252 case V_028C70_COLOR_10_11_11:
253 if (spi_format == V_028714_SPI_SHADER_FP16_ABGR)
254 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_10_11_11 << (i * 4);
255 break;
256
257 case V_028C70_COLOR_2_10_10_10:
258 if (spi_format == V_028714_SPI_SHADER_FP16_ABGR) {
259 sx_ps_downconvert |= V_028754_SX_RT_EXPORT_2_10_10_10 << (i * 4);
260 sx_blend_opt_epsilon |= V_028758_10BIT_FORMAT << (i * 4);
261 }
262 break;
263 }
264 }
265
266 /* If there are no color outputs, the first color export is
267 * always enabled as 32_R, so also set this to enable RB+.
268 */
269 if (!sx_ps_downconvert)
270 sx_ps_downconvert = V_028754_SX_RT_EXPORT_32_R;
271
272 /* SX_PS_DOWNCONVERT, SX_BLEND_OPT_EPSILON, SX_BLEND_OPT_CONTROL */
273 radeon_opt_set_context_reg3(sctx, R_028754_SX_PS_DOWNCONVERT,
274 SI_TRACKED_SX_PS_DOWNCONVERT,
275 sx_ps_downconvert, sx_blend_opt_epsilon,
276 sx_blend_opt_control);
277 }
278 if (initial_cdw != cs->current.cdw)
279 sctx->context_roll = true;
280 }
281
282 /*
283 * Blender functions
284 */
285
286 static uint32_t si_translate_blend_function(int blend_func)
287 {
288 switch (blend_func) {
289 case PIPE_BLEND_ADD:
290 return V_028780_COMB_DST_PLUS_SRC;
291 case PIPE_BLEND_SUBTRACT:
292 return V_028780_COMB_SRC_MINUS_DST;
293 case PIPE_BLEND_REVERSE_SUBTRACT:
294 return V_028780_COMB_DST_MINUS_SRC;
295 case PIPE_BLEND_MIN:
296 return V_028780_COMB_MIN_DST_SRC;
297 case PIPE_BLEND_MAX:
298 return V_028780_COMB_MAX_DST_SRC;
299 default:
300 PRINT_ERR("Unknown blend function %d\n", blend_func);
301 assert(0);
302 break;
303 }
304 return 0;
305 }
306
307 static uint32_t si_translate_blend_factor(int blend_fact)
308 {
309 switch (blend_fact) {
310 case PIPE_BLENDFACTOR_ONE:
311 return V_028780_BLEND_ONE;
312 case PIPE_BLENDFACTOR_SRC_COLOR:
313 return V_028780_BLEND_SRC_COLOR;
314 case PIPE_BLENDFACTOR_SRC_ALPHA:
315 return V_028780_BLEND_SRC_ALPHA;
316 case PIPE_BLENDFACTOR_DST_ALPHA:
317 return V_028780_BLEND_DST_ALPHA;
318 case PIPE_BLENDFACTOR_DST_COLOR:
319 return V_028780_BLEND_DST_COLOR;
320 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
321 return V_028780_BLEND_SRC_ALPHA_SATURATE;
322 case PIPE_BLENDFACTOR_CONST_COLOR:
323 return V_028780_BLEND_CONSTANT_COLOR;
324 case PIPE_BLENDFACTOR_CONST_ALPHA:
325 return V_028780_BLEND_CONSTANT_ALPHA;
326 case PIPE_BLENDFACTOR_ZERO:
327 return V_028780_BLEND_ZERO;
328 case PIPE_BLENDFACTOR_INV_SRC_COLOR:
329 return V_028780_BLEND_ONE_MINUS_SRC_COLOR;
330 case PIPE_BLENDFACTOR_INV_SRC_ALPHA:
331 return V_028780_BLEND_ONE_MINUS_SRC_ALPHA;
332 case PIPE_BLENDFACTOR_INV_DST_ALPHA:
333 return V_028780_BLEND_ONE_MINUS_DST_ALPHA;
334 case PIPE_BLENDFACTOR_INV_DST_COLOR:
335 return V_028780_BLEND_ONE_MINUS_DST_COLOR;
336 case PIPE_BLENDFACTOR_INV_CONST_COLOR:
337 return V_028780_BLEND_ONE_MINUS_CONSTANT_COLOR;
338 case PIPE_BLENDFACTOR_INV_CONST_ALPHA:
339 return V_028780_BLEND_ONE_MINUS_CONSTANT_ALPHA;
340 case PIPE_BLENDFACTOR_SRC1_COLOR:
341 return V_028780_BLEND_SRC1_COLOR;
342 case PIPE_BLENDFACTOR_SRC1_ALPHA:
343 return V_028780_BLEND_SRC1_ALPHA;
344 case PIPE_BLENDFACTOR_INV_SRC1_COLOR:
345 return V_028780_BLEND_INV_SRC1_COLOR;
346 case PIPE_BLENDFACTOR_INV_SRC1_ALPHA:
347 return V_028780_BLEND_INV_SRC1_ALPHA;
348 default:
349 PRINT_ERR("Bad blend factor %d not supported!\n", blend_fact);
350 assert(0);
351 break;
352 }
353 return 0;
354 }
355
356 static uint32_t si_translate_blend_opt_function(int blend_func)
357 {
358 switch (blend_func) {
359 case PIPE_BLEND_ADD:
360 return V_028760_OPT_COMB_ADD;
361 case PIPE_BLEND_SUBTRACT:
362 return V_028760_OPT_COMB_SUBTRACT;
363 case PIPE_BLEND_REVERSE_SUBTRACT:
364 return V_028760_OPT_COMB_REVSUBTRACT;
365 case PIPE_BLEND_MIN:
366 return V_028760_OPT_COMB_MIN;
367 case PIPE_BLEND_MAX:
368 return V_028760_OPT_COMB_MAX;
369 default:
370 return V_028760_OPT_COMB_BLEND_DISABLED;
371 }
372 }
373
374 static uint32_t si_translate_blend_opt_factor(int blend_fact, bool is_alpha)
375 {
376 switch (blend_fact) {
377 case PIPE_BLENDFACTOR_ZERO:
378 return V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_ALL;
379 case PIPE_BLENDFACTOR_ONE:
380 return V_028760_BLEND_OPT_PRESERVE_ALL_IGNORE_NONE;
381 case PIPE_BLENDFACTOR_SRC_COLOR:
382 return is_alpha ? V_028760_BLEND_OPT_PRESERVE_A1_IGNORE_A0
383 : V_028760_BLEND_OPT_PRESERVE_C1_IGNORE_C0;
384 case PIPE_BLENDFACTOR_INV_SRC_COLOR:
385 return is_alpha ? V_028760_BLEND_OPT_PRESERVE_A0_IGNORE_A1
386 : V_028760_BLEND_OPT_PRESERVE_C0_IGNORE_C1;
387 case PIPE_BLENDFACTOR_SRC_ALPHA:
388 return V_028760_BLEND_OPT_PRESERVE_A1_IGNORE_A0;
389 case PIPE_BLENDFACTOR_INV_SRC_ALPHA:
390 return V_028760_BLEND_OPT_PRESERVE_A0_IGNORE_A1;
391 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
392 return is_alpha ? V_028760_BLEND_OPT_PRESERVE_ALL_IGNORE_NONE
393 : V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_A0;
394 default:
395 return V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE;
396 }
397 }
398
399 static void si_blend_check_commutativity(struct si_screen *sscreen,
400 struct si_state_blend *blend,
401 enum pipe_blend_func func,
402 enum pipe_blendfactor src,
403 enum pipe_blendfactor dst,
404 unsigned chanmask)
405 {
406 /* Src factor is allowed when it does not depend on Dst */
407 static const uint32_t src_allowed =
408 (1u << PIPE_BLENDFACTOR_ONE) |
409 (1u << PIPE_BLENDFACTOR_SRC_COLOR) |
410 (1u << PIPE_BLENDFACTOR_SRC_ALPHA) |
411 (1u << PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE) |
412 (1u << PIPE_BLENDFACTOR_CONST_COLOR) |
413 (1u << PIPE_BLENDFACTOR_CONST_ALPHA) |
414 (1u << PIPE_BLENDFACTOR_SRC1_COLOR) |
415 (1u << PIPE_BLENDFACTOR_SRC1_ALPHA) |
416 (1u << PIPE_BLENDFACTOR_ZERO) |
417 (1u << PIPE_BLENDFACTOR_INV_SRC_COLOR) |
418 (1u << PIPE_BLENDFACTOR_INV_SRC_ALPHA) |
419 (1u << PIPE_BLENDFACTOR_INV_CONST_COLOR) |
420 (1u << PIPE_BLENDFACTOR_INV_CONST_ALPHA) |
421 (1u << PIPE_BLENDFACTOR_INV_SRC1_COLOR) |
422 (1u << PIPE_BLENDFACTOR_INV_SRC1_ALPHA);
423
424 if (dst == PIPE_BLENDFACTOR_ONE &&
425 (src_allowed & (1u << src))) {
426 /* Addition is commutative, but floating point addition isn't
427 * associative: subtle changes can be introduced via different
428 * rounding.
429 *
430 * Out-of-order is also non-deterministic, which means that
431 * this breaks OpenGL invariance requirements. So only enable
432 * out-of-order additive blending if explicitly allowed by a
433 * setting.
434 */
435 if (func == PIPE_BLEND_MAX || func == PIPE_BLEND_MIN ||
436 (func == PIPE_BLEND_ADD && sscreen->commutative_blend_add))
437 blend->commutative_4bit |= chanmask;
438 }
439 }
440
441 /**
442 * Get rid of DST in the blend factors by commuting the operands:
443 * func(src * DST, dst * 0) ---> func(src * 0, dst * SRC)
444 */
445 static void si_blend_remove_dst(unsigned *func, unsigned *src_factor,
446 unsigned *dst_factor, unsigned expected_dst,
447 unsigned replacement_src)
448 {
449 if (*src_factor == expected_dst &&
450 *dst_factor == PIPE_BLENDFACTOR_ZERO) {
451 *src_factor = PIPE_BLENDFACTOR_ZERO;
452 *dst_factor = replacement_src;
453
454 /* Commuting the operands requires reversing subtractions. */
455 if (*func == PIPE_BLEND_SUBTRACT)
456 *func = PIPE_BLEND_REVERSE_SUBTRACT;
457 else if (*func == PIPE_BLEND_REVERSE_SUBTRACT)
458 *func = PIPE_BLEND_SUBTRACT;
459 }
460 }
461
462 static bool si_blend_factor_uses_dst(unsigned factor)
463 {
464 return factor == PIPE_BLENDFACTOR_DST_COLOR ||
465 factor == PIPE_BLENDFACTOR_DST_ALPHA ||
466 factor == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE ||
467 factor == PIPE_BLENDFACTOR_INV_DST_ALPHA ||
468 factor == PIPE_BLENDFACTOR_INV_DST_COLOR;
469 }
470
471 static void *si_create_blend_state_mode(struct pipe_context *ctx,
472 const struct pipe_blend_state *state,
473 unsigned mode)
474 {
475 struct si_context *sctx = (struct si_context*)ctx;
476 struct si_state_blend *blend = CALLOC_STRUCT(si_state_blend);
477 struct si_pm4_state *pm4 = &blend->pm4;
478 uint32_t sx_mrt_blend_opt[8] = {0};
479 uint32_t color_control = 0;
480 bool logicop_enable = state->logicop_enable &&
481 state->logicop_func != PIPE_LOGICOP_COPY;
482
483 if (!blend)
484 return NULL;
485
486 blend->alpha_to_coverage = state->alpha_to_coverage;
487 blend->alpha_to_one = state->alpha_to_one;
488 blend->dual_src_blend = util_blend_state_is_dual(state, 0);
489 blend->logicop_enable = logicop_enable;
490
491 if (logicop_enable) {
492 color_control |= S_028808_ROP3(state->logicop_func | (state->logicop_func << 4));
493 } else {
494 color_control |= S_028808_ROP3(0xcc);
495 }
496
497 si_pm4_set_reg(pm4, R_028B70_DB_ALPHA_TO_MASK,
498 S_028B70_ALPHA_TO_MASK_ENABLE(state->alpha_to_coverage) |
499 S_028B70_ALPHA_TO_MASK_OFFSET0(3) |
500 S_028B70_ALPHA_TO_MASK_OFFSET1(1) |
501 S_028B70_ALPHA_TO_MASK_OFFSET2(0) |
502 S_028B70_ALPHA_TO_MASK_OFFSET3(2) |
503 S_028B70_OFFSET_ROUND(1));
504
505 if (state->alpha_to_coverage)
506 blend->need_src_alpha_4bit |= 0xf;
507
508 blend->cb_target_mask = 0;
509 blend->cb_target_enabled_4bit = 0;
510
511 for (int i = 0; i < 8; i++) {
512 /* state->rt entries > 0 only written if independent blending */
513 const int j = state->independent_blend_enable ? i : 0;
514
515 unsigned eqRGB = state->rt[j].rgb_func;
516 unsigned srcRGB = state->rt[j].rgb_src_factor;
517 unsigned dstRGB = state->rt[j].rgb_dst_factor;
518 unsigned eqA = state->rt[j].alpha_func;
519 unsigned srcA = state->rt[j].alpha_src_factor;
520 unsigned dstA = state->rt[j].alpha_dst_factor;
521
522 unsigned srcRGB_opt, dstRGB_opt, srcA_opt, dstA_opt;
523 unsigned blend_cntl = 0;
524
525 sx_mrt_blend_opt[i] =
526 S_028760_COLOR_COMB_FCN(V_028760_OPT_COMB_BLEND_DISABLED) |
527 S_028760_ALPHA_COMB_FCN(V_028760_OPT_COMB_BLEND_DISABLED);
528
529 /* Only set dual source blending for MRT0 to avoid a hang. */
530 if (i >= 1 && blend->dual_src_blend) {
531 /* Vulkan does this for dual source blending. */
532 if (i == 1)
533 blend_cntl |= S_028780_ENABLE(1);
534
535 si_pm4_set_reg(pm4, R_028780_CB_BLEND0_CONTROL + i * 4, blend_cntl);
536 continue;
537 }
538
539 /* Only addition and subtraction equations are supported with
540 * dual source blending.
541 */
542 if (blend->dual_src_blend &&
543 (eqRGB == PIPE_BLEND_MIN || eqRGB == PIPE_BLEND_MAX ||
544 eqA == PIPE_BLEND_MIN || eqA == PIPE_BLEND_MAX)) {
545 assert(!"Unsupported equation for dual source blending");
546 si_pm4_set_reg(pm4, R_028780_CB_BLEND0_CONTROL + i * 4, blend_cntl);
547 continue;
548 }
549
550 /* cb_render_state will disable unused ones */
551 blend->cb_target_mask |= (unsigned)state->rt[j].colormask << (4 * i);
552 if (state->rt[j].colormask)
553 blend->cb_target_enabled_4bit |= 0xf << (4 * i);
554
555 if (!state->rt[j].colormask || !state->rt[j].blend_enable) {
556 si_pm4_set_reg(pm4, R_028780_CB_BLEND0_CONTROL + i * 4, blend_cntl);
557 continue;
558 }
559
560 si_blend_check_commutativity(sctx->screen, blend,
561 eqRGB, srcRGB, dstRGB, 0x7 << (4 * i));
562 si_blend_check_commutativity(sctx->screen, blend,
563 eqA, srcA, dstA, 0x8 << (4 * i));
564
565 /* Blending optimizations for RB+.
566 * These transformations don't change the behavior.
567 *
568 * First, get rid of DST in the blend factors:
569 * func(src * DST, dst * 0) ---> func(src * 0, dst * SRC)
570 */
571 si_blend_remove_dst(&eqRGB, &srcRGB, &dstRGB,
572 PIPE_BLENDFACTOR_DST_COLOR,
573 PIPE_BLENDFACTOR_SRC_COLOR);
574 si_blend_remove_dst(&eqA, &srcA, &dstA,
575 PIPE_BLENDFACTOR_DST_COLOR,
576 PIPE_BLENDFACTOR_SRC_COLOR);
577 si_blend_remove_dst(&eqA, &srcA, &dstA,
578 PIPE_BLENDFACTOR_DST_ALPHA,
579 PIPE_BLENDFACTOR_SRC_ALPHA);
580
581 /* Look up the ideal settings from tables. */
582 srcRGB_opt = si_translate_blend_opt_factor(srcRGB, false);
583 dstRGB_opt = si_translate_blend_opt_factor(dstRGB, false);
584 srcA_opt = si_translate_blend_opt_factor(srcA, true);
585 dstA_opt = si_translate_blend_opt_factor(dstA, true);
586
587 /* Handle interdependencies. */
588 if (si_blend_factor_uses_dst(srcRGB))
589 dstRGB_opt = V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE;
590 if (si_blend_factor_uses_dst(srcA))
591 dstA_opt = V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE;
592
593 if (srcRGB == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE &&
594 (dstRGB == PIPE_BLENDFACTOR_ZERO ||
595 dstRGB == PIPE_BLENDFACTOR_SRC_ALPHA ||
596 dstRGB == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE))
597 dstRGB_opt = V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_A0;
598
599 /* Set the final value. */
600 sx_mrt_blend_opt[i] =
601 S_028760_COLOR_SRC_OPT(srcRGB_opt) |
602 S_028760_COLOR_DST_OPT(dstRGB_opt) |
603 S_028760_COLOR_COMB_FCN(si_translate_blend_opt_function(eqRGB)) |
604 S_028760_ALPHA_SRC_OPT(srcA_opt) |
605 S_028760_ALPHA_DST_OPT(dstA_opt) |
606 S_028760_ALPHA_COMB_FCN(si_translate_blend_opt_function(eqA));
607
608 /* Set blend state. */
609 blend_cntl |= S_028780_ENABLE(1);
610 blend_cntl |= S_028780_COLOR_COMB_FCN(si_translate_blend_function(eqRGB));
611 blend_cntl |= S_028780_COLOR_SRCBLEND(si_translate_blend_factor(srcRGB));
612 blend_cntl |= S_028780_COLOR_DESTBLEND(si_translate_blend_factor(dstRGB));
613
614 if (srcA != srcRGB || dstA != dstRGB || eqA != eqRGB) {
615 blend_cntl |= S_028780_SEPARATE_ALPHA_BLEND(1);
616 blend_cntl |= S_028780_ALPHA_COMB_FCN(si_translate_blend_function(eqA));
617 blend_cntl |= S_028780_ALPHA_SRCBLEND(si_translate_blend_factor(srcA));
618 blend_cntl |= S_028780_ALPHA_DESTBLEND(si_translate_blend_factor(dstA));
619 }
620 si_pm4_set_reg(pm4, R_028780_CB_BLEND0_CONTROL + i * 4, blend_cntl);
621
622 blend->blend_enable_4bit |= 0xfu << (i * 4);
623
624 if (sctx->family <= CHIP_NAVI14)
625 blend->dcc_msaa_corruption_4bit |= 0xfu << (i * 4);
626
627 /* This is only important for formats without alpha. */
628 if (srcRGB == PIPE_BLENDFACTOR_SRC_ALPHA ||
629 dstRGB == PIPE_BLENDFACTOR_SRC_ALPHA ||
630 srcRGB == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE ||
631 dstRGB == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE ||
632 srcRGB == PIPE_BLENDFACTOR_INV_SRC_ALPHA ||
633 dstRGB == PIPE_BLENDFACTOR_INV_SRC_ALPHA)
634 blend->need_src_alpha_4bit |= 0xfu << (i * 4);
635 }
636
637 if (sctx->family <= CHIP_NAVI14 && logicop_enable)
638 blend->dcc_msaa_corruption_4bit |= blend->cb_target_enabled_4bit;
639
640 if (blend->cb_target_mask) {
641 color_control |= S_028808_MODE(mode);
642 } else {
643 color_control |= S_028808_MODE(V_028808_CB_DISABLE);
644 }
645
646 if (sctx->screen->rbplus_allowed) {
647 /* Disable RB+ blend optimizations for dual source blending.
648 * Vulkan does this.
649 */
650 if (blend->dual_src_blend) {
651 for (int i = 0; i < 8; i++) {
652 sx_mrt_blend_opt[i] =
653 S_028760_COLOR_COMB_FCN(V_028760_OPT_COMB_NONE) |
654 S_028760_ALPHA_COMB_FCN(V_028760_OPT_COMB_NONE);
655 }
656 }
657
658 for (int i = 0; i < 8; i++)
659 si_pm4_set_reg(pm4, R_028760_SX_MRT0_BLEND_OPT + i * 4,
660 sx_mrt_blend_opt[i]);
661
662 /* RB+ doesn't work with dual source blending, logic op, and RESOLVE. */
663 if (blend->dual_src_blend || logicop_enable ||
664 mode == V_028808_CB_RESOLVE)
665 color_control |= S_028808_DISABLE_DUAL_QUAD(1);
666 }
667
668 si_pm4_set_reg(pm4, R_028808_CB_COLOR_CONTROL, color_control);
669 return blend;
670 }
671
672 static void *si_create_blend_state(struct pipe_context *ctx,
673 const struct pipe_blend_state *state)
674 {
675 return si_create_blend_state_mode(ctx, state, V_028808_CB_NORMAL);
676 }
677
678 static void si_bind_blend_state(struct pipe_context *ctx, void *state)
679 {
680 struct si_context *sctx = (struct si_context *)ctx;
681 struct si_state_blend *old_blend = sctx->queued.named.blend;
682 struct si_state_blend *blend = (struct si_state_blend *)state;
683
684 if (!state)
685 return;
686
687 si_pm4_bind_state(sctx, blend, state);
688
689 if (!old_blend ||
690 old_blend->cb_target_mask != blend->cb_target_mask ||
691 old_blend->dual_src_blend != blend->dual_src_blend ||
692 (old_blend->blend_enable_4bit != blend->blend_enable_4bit &&
693 sctx->framebuffer.nr_samples >= 2 &&
694 sctx->screen->dcc_msaa_allowed))
695 si_mark_atom_dirty(sctx, &sctx->atoms.s.cb_render_state);
696
697 if (!old_blend ||
698 old_blend->cb_target_mask != blend->cb_target_mask ||
699 old_blend->alpha_to_coverage != blend->alpha_to_coverage ||
700 old_blend->alpha_to_one != blend->alpha_to_one ||
701 old_blend->dual_src_blend != blend->dual_src_blend ||
702 old_blend->blend_enable_4bit != blend->blend_enable_4bit ||
703 old_blend->need_src_alpha_4bit != blend->need_src_alpha_4bit)
704 sctx->do_update_shaders = true;
705
706 if (sctx->screen->dpbb_allowed &&
707 (!old_blend ||
708 old_blend->alpha_to_coverage != blend->alpha_to_coverage ||
709 old_blend->blend_enable_4bit != blend->blend_enable_4bit ||
710 old_blend->cb_target_enabled_4bit != blend->cb_target_enabled_4bit))
711 si_mark_atom_dirty(sctx, &sctx->atoms.s.dpbb_state);
712
713 if (sctx->screen->has_out_of_order_rast &&
714 (!old_blend ||
715 (old_blend->blend_enable_4bit != blend->blend_enable_4bit ||
716 old_blend->cb_target_enabled_4bit != blend->cb_target_enabled_4bit ||
717 old_blend->commutative_4bit != blend->commutative_4bit ||
718 old_blend->logicop_enable != blend->logicop_enable)))
719 si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
720 }
721
722 static void si_delete_blend_state(struct pipe_context *ctx, void *state)
723 {
724 struct si_context *sctx = (struct si_context *)ctx;
725 si_pm4_delete_state(sctx, blend, (struct si_state_blend *)state);
726 }
727
728 static void si_set_blend_color(struct pipe_context *ctx,
729 const struct pipe_blend_color *state)
730 {
731 struct si_context *sctx = (struct si_context *)ctx;
732 static const struct pipe_blend_color zeros;
733
734 sctx->blend_color.state = *state;
735 sctx->blend_color.any_nonzeros = memcmp(state, &zeros, sizeof(*state)) != 0;
736 si_mark_atom_dirty(sctx, &sctx->atoms.s.blend_color);
737 }
738
739 static void si_emit_blend_color(struct si_context *sctx)
740 {
741 struct radeon_cmdbuf *cs = sctx->gfx_cs;
742
743 radeon_set_context_reg_seq(cs, R_028414_CB_BLEND_RED, 4);
744 radeon_emit_array(cs, (uint32_t*)sctx->blend_color.state.color, 4);
745 }
746
747 /*
748 * Clipping
749 */
750
751 static void si_set_clip_state(struct pipe_context *ctx,
752 const struct pipe_clip_state *state)
753 {
754 struct si_context *sctx = (struct si_context *)ctx;
755 struct pipe_constant_buffer cb;
756 static const struct pipe_clip_state zeros;
757
758 if (memcmp(&sctx->clip_state.state, state, sizeof(*state)) == 0)
759 return;
760
761 sctx->clip_state.state = *state;
762 sctx->clip_state.any_nonzeros = memcmp(state, &zeros, sizeof(*state)) != 0;
763 si_mark_atom_dirty(sctx, &sctx->atoms.s.clip_state);
764
765 cb.buffer = NULL;
766 cb.user_buffer = state->ucp;
767 cb.buffer_offset = 0;
768 cb.buffer_size = 4*4*8;
769 si_set_rw_buffer(sctx, SI_VS_CONST_CLIP_PLANES, &cb);
770 pipe_resource_reference(&cb.buffer, NULL);
771 }
772
773 static void si_emit_clip_state(struct si_context *sctx)
774 {
775 struct radeon_cmdbuf *cs = sctx->gfx_cs;
776
777 radeon_set_context_reg_seq(cs, R_0285BC_PA_CL_UCP_0_X, 6*4);
778 radeon_emit_array(cs, (uint32_t*)sctx->clip_state.state.ucp, 6*4);
779 }
780
781 static void si_emit_clip_regs(struct si_context *sctx)
782 {
783 struct si_shader *vs = si_get_vs_state(sctx);
784 struct si_shader_selector *vs_sel = vs->selector;
785 struct tgsi_shader_info *info = &vs_sel->info;
786 struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
787 unsigned window_space =
788 info->properties[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION];
789 unsigned clipdist_mask = vs_sel->clipdist_mask;
790 unsigned ucp_mask = clipdist_mask ? 0 : rs->clip_plane_enable & SIX_BITS;
791 unsigned culldist_mask = vs_sel->culldist_mask;
792 unsigned total_mask;
793
794 if (vs->key.opt.clip_disable) {
795 assert(!info->culldist_writemask);
796 clipdist_mask = 0;
797 culldist_mask = 0;
798 }
799 total_mask = clipdist_mask | culldist_mask;
800
801 /* Clip distances on points have no effect, so need to be implemented
802 * as cull distances. This applies for the clipvertex case as well.
803 *
804 * Setting this for primitives other than points should have no adverse
805 * effects.
806 */
807 clipdist_mask &= rs->clip_plane_enable;
808 culldist_mask |= clipdist_mask;
809
810 unsigned initial_cdw = sctx->gfx_cs->current.cdw;
811 radeon_opt_set_context_reg(sctx, R_02881C_PA_CL_VS_OUT_CNTL,
812 SI_TRACKED_PA_CL_VS_OUT_CNTL,
813 vs_sel->pa_cl_vs_out_cntl |
814 S_02881C_VS_OUT_CCDIST0_VEC_ENA((total_mask & 0x0F) != 0) |
815 S_02881C_VS_OUT_CCDIST1_VEC_ENA((total_mask & 0xF0) != 0) |
816 clipdist_mask | (culldist_mask << 8));
817 radeon_opt_set_context_reg(sctx, R_028810_PA_CL_CLIP_CNTL,
818 SI_TRACKED_PA_CL_CLIP_CNTL,
819 rs->pa_cl_clip_cntl |
820 ucp_mask |
821 S_028810_CLIP_DISABLE(window_space));
822
823 if (initial_cdw != sctx->gfx_cs->current.cdw)
824 sctx->context_roll = true;
825 }
826
827 /*
828 * inferred state between framebuffer and rasterizer
829 */
830 static void si_update_poly_offset_state(struct si_context *sctx)
831 {
832 struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
833
834 if (!rs || !rs->uses_poly_offset || !sctx->framebuffer.state.zsbuf) {
835 si_pm4_bind_state(sctx, poly_offset, NULL);
836 return;
837 }
838
839 /* Use the user format, not db_render_format, so that the polygon
840 * offset behaves as expected by applications.
841 */
842 switch (sctx->framebuffer.state.zsbuf->texture->format) {
843 case PIPE_FORMAT_Z16_UNORM:
844 si_pm4_bind_state(sctx, poly_offset, &rs->pm4_poly_offset[0]);
845 break;
846 default: /* 24-bit */
847 si_pm4_bind_state(sctx, poly_offset, &rs->pm4_poly_offset[1]);
848 break;
849 case PIPE_FORMAT_Z32_FLOAT:
850 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
851 si_pm4_bind_state(sctx, poly_offset, &rs->pm4_poly_offset[2]);
852 break;
853 }
854 }
855
856 /*
857 * Rasterizer
858 */
859
860 static uint32_t si_translate_fill(uint32_t func)
861 {
862 switch(func) {
863 case PIPE_POLYGON_MODE_FILL:
864 return V_028814_X_DRAW_TRIANGLES;
865 case PIPE_POLYGON_MODE_LINE:
866 return V_028814_X_DRAW_LINES;
867 case PIPE_POLYGON_MODE_POINT:
868 return V_028814_X_DRAW_POINTS;
869 default:
870 assert(0);
871 return V_028814_X_DRAW_POINTS;
872 }
873 }
874
875 static void *si_create_rs_state(struct pipe_context *ctx,
876 const struct pipe_rasterizer_state *state)
877 {
878 struct si_screen *sscreen = ((struct si_context *)ctx)->screen;
879 struct si_state_rasterizer *rs = CALLOC_STRUCT(si_state_rasterizer);
880 struct si_pm4_state *pm4 = &rs->pm4;
881 unsigned tmp, i;
882 float psize_min, psize_max;
883
884 if (!rs) {
885 return NULL;
886 }
887
888 if (!state->front_ccw) {
889 rs->cull_front = !!(state->cull_face & PIPE_FACE_FRONT);
890 rs->cull_back = !!(state->cull_face & PIPE_FACE_BACK);
891 } else {
892 rs->cull_back = !!(state->cull_face & PIPE_FACE_FRONT);
893 rs->cull_front = !!(state->cull_face & PIPE_FACE_BACK);
894 }
895 rs->depth_clamp_any = !state->depth_clip_near || !state->depth_clip_far;
896 rs->provoking_vertex_first = state->flatshade_first;
897 rs->scissor_enable = state->scissor;
898 rs->clip_halfz = state->clip_halfz;
899 rs->two_side = state->light_twoside;
900 rs->multisample_enable = state->multisample;
901 rs->force_persample_interp = state->force_persample_interp;
902 rs->clip_plane_enable = state->clip_plane_enable;
903 rs->half_pixel_center = state->half_pixel_center;
904 rs->line_stipple_enable = state->line_stipple_enable;
905 rs->poly_stipple_enable = state->poly_stipple_enable;
906 rs->line_smooth = state->line_smooth;
907 rs->line_width = state->line_width;
908 rs->poly_smooth = state->poly_smooth;
909 rs->uses_poly_offset = state->offset_point || state->offset_line ||
910 state->offset_tri;
911 rs->clamp_fragment_color = state->clamp_fragment_color;
912 rs->clamp_vertex_color = state->clamp_vertex_color;
913 rs->flatshade = state->flatshade;
914 rs->flatshade_first = state->flatshade_first;
915 rs->sprite_coord_enable = state->sprite_coord_enable;
916 rs->rasterizer_discard = state->rasterizer_discard;
917 rs->pa_sc_line_stipple = state->line_stipple_enable ?
918 S_028A0C_LINE_PATTERN(state->line_stipple_pattern) |
919 S_028A0C_REPEAT_COUNT(state->line_stipple_factor) : 0;
920 rs->pa_cl_clip_cntl =
921 S_028810_DX_CLIP_SPACE_DEF(state->clip_halfz) |
922 S_028810_ZCLIP_NEAR_DISABLE(!state->depth_clip_near) |
923 S_028810_ZCLIP_FAR_DISABLE(!state->depth_clip_far) |
924 S_028810_DX_RASTERIZATION_KILL(state->rasterizer_discard) |
925 S_028810_DX_LINEAR_ATTR_CLIP_ENA(1);
926
927 si_pm4_set_reg(pm4, R_0286D4_SPI_INTERP_CONTROL_0,
928 S_0286D4_FLAT_SHADE_ENA(1) |
929 S_0286D4_PNT_SPRITE_ENA(state->point_quad_rasterization) |
930 S_0286D4_PNT_SPRITE_OVRD_X(V_0286D4_SPI_PNT_SPRITE_SEL_S) |
931 S_0286D4_PNT_SPRITE_OVRD_Y(V_0286D4_SPI_PNT_SPRITE_SEL_T) |
932 S_0286D4_PNT_SPRITE_OVRD_Z(V_0286D4_SPI_PNT_SPRITE_SEL_0) |
933 S_0286D4_PNT_SPRITE_OVRD_W(V_0286D4_SPI_PNT_SPRITE_SEL_1) |
934 S_0286D4_PNT_SPRITE_TOP_1(state->sprite_coord_mode != PIPE_SPRITE_COORD_UPPER_LEFT));
935
936 /* point size 12.4 fixed point */
937 tmp = (unsigned)(state->point_size * 8.0);
938 si_pm4_set_reg(pm4, R_028A00_PA_SU_POINT_SIZE, S_028A00_HEIGHT(tmp) | S_028A00_WIDTH(tmp));
939
940 if (state->point_size_per_vertex) {
941 psize_min = util_get_min_point_size(state);
942 psize_max = SI_MAX_POINT_SIZE;
943 } else {
944 /* Force the point size to be as if the vertex output was disabled. */
945 psize_min = state->point_size;
946 psize_max = state->point_size;
947 }
948 rs->max_point_size = psize_max;
949
950 /* Divide by two, because 0.5 = 1 pixel. */
951 si_pm4_set_reg(pm4, R_028A04_PA_SU_POINT_MINMAX,
952 S_028A04_MIN_SIZE(si_pack_float_12p4(psize_min/2)) |
953 S_028A04_MAX_SIZE(si_pack_float_12p4(psize_max/2)));
954
955 si_pm4_set_reg(pm4, R_028A08_PA_SU_LINE_CNTL,
956 S_028A08_WIDTH(si_pack_float_12p4(state->line_width/2)));
957 si_pm4_set_reg(pm4, R_028A48_PA_SC_MODE_CNTL_0,
958 S_028A48_LINE_STIPPLE_ENABLE(state->line_stipple_enable) |
959 S_028A48_MSAA_ENABLE(state->multisample ||
960 state->poly_smooth ||
961 state->line_smooth) |
962 S_028A48_VPORT_SCISSOR_ENABLE(1) |
963 S_028A48_ALTERNATE_RBS_PER_TILE(sscreen->info.chip_class >= GFX9));
964
965 si_pm4_set_reg(pm4, R_028B7C_PA_SU_POLY_OFFSET_CLAMP, fui(state->offset_clamp));
966 si_pm4_set_reg(pm4, R_028814_PA_SU_SC_MODE_CNTL,
967 S_028814_PROVOKING_VTX_LAST(!state->flatshade_first) |
968 S_028814_CULL_FRONT((state->cull_face & PIPE_FACE_FRONT) ? 1 : 0) |
969 S_028814_CULL_BACK((state->cull_face & PIPE_FACE_BACK) ? 1 : 0) |
970 S_028814_FACE(!state->front_ccw) |
971 S_028814_POLY_OFFSET_FRONT_ENABLE(util_get_offset(state, state->fill_front)) |
972 S_028814_POLY_OFFSET_BACK_ENABLE(util_get_offset(state, state->fill_back)) |
973 S_028814_POLY_OFFSET_PARA_ENABLE(state->offset_point || state->offset_line) |
974 S_028814_POLY_MODE(state->fill_front != PIPE_POLYGON_MODE_FILL ||
975 state->fill_back != PIPE_POLYGON_MODE_FILL) |
976 S_028814_POLYMODE_FRONT_PTYPE(si_translate_fill(state->fill_front)) |
977 S_028814_POLYMODE_BACK_PTYPE(si_translate_fill(state->fill_back)));
978
979 if (!rs->uses_poly_offset)
980 return rs;
981
982 rs->pm4_poly_offset = CALLOC(3, sizeof(struct si_pm4_state));
983 if (!rs->pm4_poly_offset) {
984 FREE(rs);
985 return NULL;
986 }
987
988 /* Precalculate polygon offset states for 16-bit, 24-bit, and 32-bit zbuffers. */
989 for (i = 0; i < 3; i++) {
990 struct si_pm4_state *pm4 = &rs->pm4_poly_offset[i];
991 float offset_units = state->offset_units;
992 float offset_scale = state->offset_scale * 16.0f;
993 uint32_t pa_su_poly_offset_db_fmt_cntl = 0;
994
995 if (!state->offset_units_unscaled) {
996 switch (i) {
997 case 0: /* 16-bit zbuffer */
998 offset_units *= 4.0f;
999 pa_su_poly_offset_db_fmt_cntl =
1000 S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-16);
1001 break;
1002 case 1: /* 24-bit zbuffer */
1003 offset_units *= 2.0f;
1004 pa_su_poly_offset_db_fmt_cntl =
1005 S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-24);
1006 break;
1007 case 2: /* 32-bit zbuffer */
1008 offset_units *= 1.0f;
1009 pa_su_poly_offset_db_fmt_cntl = S_028B78_POLY_OFFSET_NEG_NUM_DB_BITS(-23) |
1010 S_028B78_POLY_OFFSET_DB_IS_FLOAT_FMT(1);
1011 break;
1012 }
1013 }
1014
1015 si_pm4_set_reg(pm4, R_028B80_PA_SU_POLY_OFFSET_FRONT_SCALE,
1016 fui(offset_scale));
1017 si_pm4_set_reg(pm4, R_028B84_PA_SU_POLY_OFFSET_FRONT_OFFSET,
1018 fui(offset_units));
1019 si_pm4_set_reg(pm4, R_028B88_PA_SU_POLY_OFFSET_BACK_SCALE,
1020 fui(offset_scale));
1021 si_pm4_set_reg(pm4, R_028B8C_PA_SU_POLY_OFFSET_BACK_OFFSET,
1022 fui(offset_units));
1023 si_pm4_set_reg(pm4, R_028B78_PA_SU_POLY_OFFSET_DB_FMT_CNTL,
1024 pa_su_poly_offset_db_fmt_cntl);
1025 }
1026
1027 return rs;
1028 }
1029
1030 static void si_bind_rs_state(struct pipe_context *ctx, void *state)
1031 {
1032 struct si_context *sctx = (struct si_context *)ctx;
1033 struct si_state_rasterizer *old_rs =
1034 (struct si_state_rasterizer*)sctx->queued.named.rasterizer;
1035 struct si_state_rasterizer *rs = (struct si_state_rasterizer *)state;
1036
1037 if (!state)
1038 return;
1039
1040 if (!old_rs || old_rs->multisample_enable != rs->multisample_enable) {
1041 si_mark_atom_dirty(sctx, &sctx->atoms.s.db_render_state);
1042
1043 /* Update the small primitive filter workaround if necessary. */
1044 if (sctx->screen->has_msaa_sample_loc_bug &&
1045 sctx->framebuffer.nr_samples > 1)
1046 si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_sample_locs);
1047 }
1048
1049 sctx->current_vs_state &= C_VS_STATE_CLAMP_VERTEX_COLOR;
1050 sctx->current_vs_state |= S_VS_STATE_CLAMP_VERTEX_COLOR(rs->clamp_vertex_color);
1051
1052 si_pm4_bind_state(sctx, rasterizer, rs);
1053 si_update_poly_offset_state(sctx);
1054
1055 if (!old_rs ||
1056 old_rs->scissor_enable != rs->scissor_enable)
1057 si_mark_atom_dirty(sctx, &sctx->atoms.s.scissors);
1058
1059 if (!old_rs ||
1060 old_rs->line_width != rs->line_width ||
1061 old_rs->max_point_size != rs->max_point_size ||
1062 old_rs->half_pixel_center != rs->half_pixel_center)
1063 si_mark_atom_dirty(sctx, &sctx->atoms.s.guardband);
1064
1065 if (!old_rs ||
1066 old_rs->clip_halfz != rs->clip_halfz)
1067 si_mark_atom_dirty(sctx, &sctx->atoms.s.viewports);
1068
1069 if (!old_rs ||
1070 old_rs->clip_plane_enable != rs->clip_plane_enable ||
1071 old_rs->pa_cl_clip_cntl != rs->pa_cl_clip_cntl)
1072 si_mark_atom_dirty(sctx, &sctx->atoms.s.clip_regs);
1073
1074 sctx->ia_multi_vgt_param_key.u.line_stipple_enabled =
1075 rs->line_stipple_enable;
1076
1077 if (!old_rs ||
1078 old_rs->clip_plane_enable != rs->clip_plane_enable ||
1079 old_rs->rasterizer_discard != rs->rasterizer_discard ||
1080 old_rs->sprite_coord_enable != rs->sprite_coord_enable ||
1081 old_rs->flatshade != rs->flatshade ||
1082 old_rs->two_side != rs->two_side ||
1083 old_rs->multisample_enable != rs->multisample_enable ||
1084 old_rs->poly_stipple_enable != rs->poly_stipple_enable ||
1085 old_rs->poly_smooth != rs->poly_smooth ||
1086 old_rs->line_smooth != rs->line_smooth ||
1087 old_rs->clamp_fragment_color != rs->clamp_fragment_color ||
1088 old_rs->force_persample_interp != rs->force_persample_interp)
1089 sctx->do_update_shaders = true;
1090 }
1091
1092 static void si_delete_rs_state(struct pipe_context *ctx, void *state)
1093 {
1094 struct si_context *sctx = (struct si_context *)ctx;
1095 struct si_state_rasterizer *rs = (struct si_state_rasterizer *)state;
1096
1097 if (sctx->queued.named.rasterizer == state)
1098 si_pm4_bind_state(sctx, poly_offset, NULL);
1099
1100 FREE(rs->pm4_poly_offset);
1101 si_pm4_delete_state(sctx, rasterizer, rs);
1102 }
1103
1104 /*
1105 * infeered state between dsa and stencil ref
1106 */
1107 static void si_emit_stencil_ref(struct si_context *sctx)
1108 {
1109 struct radeon_cmdbuf *cs = sctx->gfx_cs;
1110 struct pipe_stencil_ref *ref = &sctx->stencil_ref.state;
1111 struct si_dsa_stencil_ref_part *dsa = &sctx->stencil_ref.dsa_part;
1112
1113 radeon_set_context_reg_seq(cs, R_028430_DB_STENCILREFMASK, 2);
1114 radeon_emit(cs, S_028430_STENCILTESTVAL(ref->ref_value[0]) |
1115 S_028430_STENCILMASK(dsa->valuemask[0]) |
1116 S_028430_STENCILWRITEMASK(dsa->writemask[0]) |
1117 S_028430_STENCILOPVAL(1));
1118 radeon_emit(cs, S_028434_STENCILTESTVAL_BF(ref->ref_value[1]) |
1119 S_028434_STENCILMASK_BF(dsa->valuemask[1]) |
1120 S_028434_STENCILWRITEMASK_BF(dsa->writemask[1]) |
1121 S_028434_STENCILOPVAL_BF(1));
1122 }
1123
1124 static void si_set_stencil_ref(struct pipe_context *ctx,
1125 const struct pipe_stencil_ref *state)
1126 {
1127 struct si_context *sctx = (struct si_context *)ctx;
1128
1129 if (memcmp(&sctx->stencil_ref.state, state, sizeof(*state)) == 0)
1130 return;
1131
1132 sctx->stencil_ref.state = *state;
1133 si_mark_atom_dirty(sctx, &sctx->atoms.s.stencil_ref);
1134 }
1135
1136
1137 /*
1138 * DSA
1139 */
1140
1141 static uint32_t si_translate_stencil_op(int s_op)
1142 {
1143 switch (s_op) {
1144 case PIPE_STENCIL_OP_KEEP:
1145 return V_02842C_STENCIL_KEEP;
1146 case PIPE_STENCIL_OP_ZERO:
1147 return V_02842C_STENCIL_ZERO;
1148 case PIPE_STENCIL_OP_REPLACE:
1149 return V_02842C_STENCIL_REPLACE_TEST;
1150 case PIPE_STENCIL_OP_INCR:
1151 return V_02842C_STENCIL_ADD_CLAMP;
1152 case PIPE_STENCIL_OP_DECR:
1153 return V_02842C_STENCIL_SUB_CLAMP;
1154 case PIPE_STENCIL_OP_INCR_WRAP:
1155 return V_02842C_STENCIL_ADD_WRAP;
1156 case PIPE_STENCIL_OP_DECR_WRAP:
1157 return V_02842C_STENCIL_SUB_WRAP;
1158 case PIPE_STENCIL_OP_INVERT:
1159 return V_02842C_STENCIL_INVERT;
1160 default:
1161 PRINT_ERR("Unknown stencil op %d", s_op);
1162 assert(0);
1163 break;
1164 }
1165 return 0;
1166 }
1167
1168 static bool si_dsa_writes_stencil(const struct pipe_stencil_state *s)
1169 {
1170 return s->enabled && s->writemask &&
1171 (s->fail_op != PIPE_STENCIL_OP_KEEP ||
1172 s->zfail_op != PIPE_STENCIL_OP_KEEP ||
1173 s->zpass_op != PIPE_STENCIL_OP_KEEP);
1174 }
1175
1176 static bool si_order_invariant_stencil_op(enum pipe_stencil_op op)
1177 {
1178 /* REPLACE is normally order invariant, except when the stencil
1179 * reference value is written by the fragment shader. Tracking this
1180 * interaction does not seem worth the effort, so be conservative. */
1181 return op != PIPE_STENCIL_OP_INCR &&
1182 op != PIPE_STENCIL_OP_DECR &&
1183 op != PIPE_STENCIL_OP_REPLACE;
1184 }
1185
1186 /* Compute whether, assuming Z writes are disabled, this stencil state is order
1187 * invariant in the sense that the set of passing fragments as well as the
1188 * final stencil buffer result does not depend on the order of fragments. */
1189 static bool si_order_invariant_stencil_state(const struct pipe_stencil_state *state)
1190 {
1191 return !state->enabled || !state->writemask ||
1192 /* The following assumes that Z writes are disabled. */
1193 (state->func == PIPE_FUNC_ALWAYS &&
1194 si_order_invariant_stencil_op(state->zpass_op) &&
1195 si_order_invariant_stencil_op(state->zfail_op)) ||
1196 (state->func == PIPE_FUNC_NEVER &&
1197 si_order_invariant_stencil_op(state->fail_op));
1198 }
1199
1200 static void *si_create_dsa_state(struct pipe_context *ctx,
1201 const struct pipe_depth_stencil_alpha_state *state)
1202 {
1203 struct si_context *sctx = (struct si_context *)ctx;
1204 struct si_state_dsa *dsa = CALLOC_STRUCT(si_state_dsa);
1205 struct si_pm4_state *pm4 = &dsa->pm4;
1206 unsigned db_depth_control;
1207 uint32_t db_stencil_control = 0;
1208
1209 if (!dsa) {
1210 return NULL;
1211 }
1212
1213 dsa->stencil_ref.valuemask[0] = state->stencil[0].valuemask;
1214 dsa->stencil_ref.valuemask[1] = state->stencil[1].valuemask;
1215 dsa->stencil_ref.writemask[0] = state->stencil[0].writemask;
1216 dsa->stencil_ref.writemask[1] = state->stencil[1].writemask;
1217
1218 db_depth_control = S_028800_Z_ENABLE(state->depth.enabled) |
1219 S_028800_Z_WRITE_ENABLE(state->depth.writemask) |
1220 S_028800_ZFUNC(state->depth.func) |
1221 S_028800_DEPTH_BOUNDS_ENABLE(state->depth.bounds_test);
1222
1223 /* stencil */
1224 if (state->stencil[0].enabled) {
1225 db_depth_control |= S_028800_STENCIL_ENABLE(1);
1226 db_depth_control |= S_028800_STENCILFUNC(state->stencil[0].func);
1227 db_stencil_control |= S_02842C_STENCILFAIL(si_translate_stencil_op(state->stencil[0].fail_op));
1228 db_stencil_control |= S_02842C_STENCILZPASS(si_translate_stencil_op(state->stencil[0].zpass_op));
1229 db_stencil_control |= S_02842C_STENCILZFAIL(si_translate_stencil_op(state->stencil[0].zfail_op));
1230
1231 if (state->stencil[1].enabled) {
1232 db_depth_control |= S_028800_BACKFACE_ENABLE(1);
1233 db_depth_control |= S_028800_STENCILFUNC_BF(state->stencil[1].func);
1234 db_stencil_control |= S_02842C_STENCILFAIL_BF(si_translate_stencil_op(state->stencil[1].fail_op));
1235 db_stencil_control |= S_02842C_STENCILZPASS_BF(si_translate_stencil_op(state->stencil[1].zpass_op));
1236 db_stencil_control |= S_02842C_STENCILZFAIL_BF(si_translate_stencil_op(state->stencil[1].zfail_op));
1237 }
1238 }
1239
1240 /* alpha */
1241 if (state->alpha.enabled) {
1242 dsa->alpha_func = state->alpha.func;
1243
1244 si_pm4_set_reg(pm4, R_00B030_SPI_SHADER_USER_DATA_PS_0 +
1245 SI_SGPR_ALPHA_REF * 4, fui(state->alpha.ref_value));
1246 } else {
1247 dsa->alpha_func = PIPE_FUNC_ALWAYS;
1248 }
1249
1250 si_pm4_set_reg(pm4, R_028800_DB_DEPTH_CONTROL, db_depth_control);
1251 if (state->stencil[0].enabled)
1252 si_pm4_set_reg(pm4, R_02842C_DB_STENCIL_CONTROL, db_stencil_control);
1253 if (state->depth.bounds_test) {
1254 si_pm4_set_reg(pm4, R_028020_DB_DEPTH_BOUNDS_MIN, fui(state->depth.bounds_min));
1255 si_pm4_set_reg(pm4, R_028024_DB_DEPTH_BOUNDS_MAX, fui(state->depth.bounds_max));
1256 }
1257
1258 dsa->depth_enabled = state->depth.enabled;
1259 dsa->depth_write_enabled = state->depth.enabled &&
1260 state->depth.writemask;
1261 dsa->stencil_enabled = state->stencil[0].enabled;
1262 dsa->stencil_write_enabled = state->stencil[0].enabled &&
1263 (si_dsa_writes_stencil(&state->stencil[0]) ||
1264 si_dsa_writes_stencil(&state->stencil[1]));
1265 dsa->db_can_write = dsa->depth_write_enabled ||
1266 dsa->stencil_write_enabled;
1267
1268 bool zfunc_is_ordered =
1269 state->depth.func == PIPE_FUNC_NEVER ||
1270 state->depth.func == PIPE_FUNC_LESS ||
1271 state->depth.func == PIPE_FUNC_LEQUAL ||
1272 state->depth.func == PIPE_FUNC_GREATER ||
1273 state->depth.func == PIPE_FUNC_GEQUAL;
1274
1275 bool nozwrite_and_order_invariant_stencil =
1276 !dsa->db_can_write ||
1277 (!dsa->depth_write_enabled &&
1278 si_order_invariant_stencil_state(&state->stencil[0]) &&
1279 si_order_invariant_stencil_state(&state->stencil[1]));
1280
1281 dsa->order_invariance[1].zs =
1282 nozwrite_and_order_invariant_stencil ||
1283 (!dsa->stencil_write_enabled && zfunc_is_ordered);
1284 dsa->order_invariance[0].zs = !dsa->depth_write_enabled || zfunc_is_ordered;
1285
1286 dsa->order_invariance[1].pass_set =
1287 nozwrite_and_order_invariant_stencil ||
1288 (!dsa->stencil_write_enabled &&
1289 (state->depth.func == PIPE_FUNC_ALWAYS ||
1290 state->depth.func == PIPE_FUNC_NEVER));
1291 dsa->order_invariance[0].pass_set =
1292 !dsa->depth_write_enabled ||
1293 (state->depth.func == PIPE_FUNC_ALWAYS ||
1294 state->depth.func == PIPE_FUNC_NEVER);
1295
1296 dsa->order_invariance[1].pass_last =
1297 sctx->screen->assume_no_z_fights &&
1298 !dsa->stencil_write_enabled &&
1299 dsa->depth_write_enabled && zfunc_is_ordered;
1300 dsa->order_invariance[0].pass_last =
1301 sctx->screen->assume_no_z_fights &&
1302 dsa->depth_write_enabled && zfunc_is_ordered;
1303
1304 return dsa;
1305 }
1306
1307 static void si_bind_dsa_state(struct pipe_context *ctx, void *state)
1308 {
1309 struct si_context *sctx = (struct si_context *)ctx;
1310 struct si_state_dsa *old_dsa = sctx->queued.named.dsa;
1311 struct si_state_dsa *dsa = state;
1312
1313 if (!state)
1314 return;
1315
1316 si_pm4_bind_state(sctx, dsa, dsa);
1317
1318 if (memcmp(&dsa->stencil_ref, &sctx->stencil_ref.dsa_part,
1319 sizeof(struct si_dsa_stencil_ref_part)) != 0) {
1320 sctx->stencil_ref.dsa_part = dsa->stencil_ref;
1321 si_mark_atom_dirty(sctx, &sctx->atoms.s.stencil_ref);
1322 }
1323
1324 if (!old_dsa || old_dsa->alpha_func != dsa->alpha_func)
1325 sctx->do_update_shaders = true;
1326
1327 if (sctx->screen->dpbb_allowed &&
1328 (!old_dsa ||
1329 (old_dsa->depth_enabled != dsa->depth_enabled ||
1330 old_dsa->stencil_enabled != dsa->stencil_enabled ||
1331 old_dsa->db_can_write != dsa->db_can_write)))
1332 si_mark_atom_dirty(sctx, &sctx->atoms.s.dpbb_state);
1333
1334 if (sctx->screen->has_out_of_order_rast &&
1335 (!old_dsa ||
1336 memcmp(old_dsa->order_invariance, dsa->order_invariance,
1337 sizeof(old_dsa->order_invariance))))
1338 si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
1339 }
1340
1341 static void si_delete_dsa_state(struct pipe_context *ctx, void *state)
1342 {
1343 struct si_context *sctx = (struct si_context *)ctx;
1344 si_pm4_delete_state(sctx, dsa, (struct si_state_dsa *)state);
1345 }
1346
1347 static void *si_create_db_flush_dsa(struct si_context *sctx)
1348 {
1349 struct pipe_depth_stencil_alpha_state dsa = {};
1350
1351 return sctx->b.create_depth_stencil_alpha_state(&sctx->b, &dsa);
1352 }
1353
1354 /* DB RENDER STATE */
1355
1356 static void si_set_active_query_state(struct pipe_context *ctx, bool enable)
1357 {
1358 struct si_context *sctx = (struct si_context*)ctx;
1359
1360 /* Pipeline stat & streamout queries. */
1361 if (enable) {
1362 sctx->flags &= ~SI_CONTEXT_STOP_PIPELINE_STATS;
1363 sctx->flags |= SI_CONTEXT_START_PIPELINE_STATS;
1364 } else {
1365 sctx->flags &= ~SI_CONTEXT_START_PIPELINE_STATS;
1366 sctx->flags |= SI_CONTEXT_STOP_PIPELINE_STATS;
1367 }
1368
1369 /* Occlusion queries. */
1370 if (sctx->occlusion_queries_disabled != !enable) {
1371 sctx->occlusion_queries_disabled = !enable;
1372 si_mark_atom_dirty(sctx, &sctx->atoms.s.db_render_state);
1373 }
1374 }
1375
1376 void si_set_occlusion_query_state(struct si_context *sctx,
1377 bool old_perfect_enable)
1378 {
1379 si_mark_atom_dirty(sctx, &sctx->atoms.s.db_render_state);
1380
1381 bool perfect_enable = sctx->num_perfect_occlusion_queries != 0;
1382
1383 if (perfect_enable != old_perfect_enable)
1384 si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
1385 }
1386
1387 void si_save_qbo_state(struct si_context *sctx, struct si_qbo_state *st)
1388 {
1389 st->saved_compute = sctx->cs_shader_state.program;
1390
1391 si_get_pipe_constant_buffer(sctx, PIPE_SHADER_COMPUTE, 0, &st->saved_const0);
1392 si_get_shader_buffers(sctx, PIPE_SHADER_COMPUTE, 0, 3, st->saved_ssbo);
1393
1394 st->saved_ssbo_writable_mask = 0;
1395
1396 for (unsigned i = 0; i < 3; i++) {
1397 if (sctx->const_and_shader_buffers[PIPE_SHADER_COMPUTE].writable_mask &
1398 (1u << si_get_shaderbuf_slot(i)))
1399 st->saved_ssbo_writable_mask |= 1 << i;
1400 }
1401 }
1402
1403 void si_restore_qbo_state(struct si_context *sctx, struct si_qbo_state *st)
1404 {
1405 sctx->b.bind_compute_state(&sctx->b, st->saved_compute);
1406
1407 sctx->b.set_constant_buffer(&sctx->b, PIPE_SHADER_COMPUTE, 0, &st->saved_const0);
1408 pipe_resource_reference(&st->saved_const0.buffer, NULL);
1409
1410 sctx->b.set_shader_buffers(&sctx->b, PIPE_SHADER_COMPUTE, 0, 3, st->saved_ssbo,
1411 st->saved_ssbo_writable_mask);
1412 for (unsigned i = 0; i < 3; ++i)
1413 pipe_resource_reference(&st->saved_ssbo[i].buffer, NULL);
1414 }
1415
1416 static void si_emit_db_render_state(struct si_context *sctx)
1417 {
1418 struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
1419 unsigned db_shader_control, db_render_control, db_count_control;
1420 unsigned initial_cdw = sctx->gfx_cs->current.cdw;
1421
1422 /* DB_RENDER_CONTROL */
1423 if (sctx->dbcb_depth_copy_enabled ||
1424 sctx->dbcb_stencil_copy_enabled) {
1425 db_render_control =
1426 S_028000_DEPTH_COPY(sctx->dbcb_depth_copy_enabled) |
1427 S_028000_STENCIL_COPY(sctx->dbcb_stencil_copy_enabled) |
1428 S_028000_COPY_CENTROID(1) |
1429 S_028000_COPY_SAMPLE(sctx->dbcb_copy_sample);
1430 } else if (sctx->db_flush_depth_inplace || sctx->db_flush_stencil_inplace) {
1431 db_render_control =
1432 S_028000_DEPTH_COMPRESS_DISABLE(sctx->db_flush_depth_inplace) |
1433 S_028000_STENCIL_COMPRESS_DISABLE(sctx->db_flush_stencil_inplace);
1434 } else {
1435 db_render_control =
1436 S_028000_DEPTH_CLEAR_ENABLE(sctx->db_depth_clear) |
1437 S_028000_STENCIL_CLEAR_ENABLE(sctx->db_stencil_clear);
1438 }
1439
1440 /* DB_COUNT_CONTROL (occlusion queries) */
1441 if (sctx->num_occlusion_queries > 0 &&
1442 !sctx->occlusion_queries_disabled) {
1443 bool perfect = sctx->num_perfect_occlusion_queries > 0;
1444 bool gfx10_perfect = sctx->chip_class >= GFX10 && perfect;
1445
1446 if (sctx->chip_class >= GFX7) {
1447 unsigned log_sample_rate = sctx->framebuffer.log_samples;
1448
1449 /* Stoney doesn't increment occlusion query counters
1450 * if the sample rate is 16x. Use 8x sample rate instead.
1451 */
1452 if (sctx->family == CHIP_STONEY)
1453 log_sample_rate = MIN2(log_sample_rate, 3);
1454
1455 db_count_control =
1456 S_028004_PERFECT_ZPASS_COUNTS(perfect) |
1457 S_028004_DISABLE_CONSERVATIVE_ZPASS_COUNTS(gfx10_perfect) |
1458 S_028004_SAMPLE_RATE(log_sample_rate) |
1459 S_028004_ZPASS_ENABLE(1) |
1460 S_028004_SLICE_EVEN_ENABLE(1) |
1461 S_028004_SLICE_ODD_ENABLE(1);
1462 } else {
1463 db_count_control =
1464 S_028004_PERFECT_ZPASS_COUNTS(perfect) |
1465 S_028004_SAMPLE_RATE(sctx->framebuffer.log_samples);
1466 }
1467 } else {
1468 /* Disable occlusion queries. */
1469 if (sctx->chip_class >= GFX7) {
1470 db_count_control = 0;
1471 } else {
1472 db_count_control = S_028004_ZPASS_INCREMENT_DISABLE(1);
1473 }
1474 }
1475
1476 radeon_opt_set_context_reg2(sctx, R_028000_DB_RENDER_CONTROL,
1477 SI_TRACKED_DB_RENDER_CONTROL, db_render_control,
1478 db_count_control);
1479
1480 /* DB_RENDER_OVERRIDE2 */
1481 radeon_opt_set_context_reg(sctx, R_028010_DB_RENDER_OVERRIDE2,
1482 SI_TRACKED_DB_RENDER_OVERRIDE2,
1483 S_028010_DISABLE_ZMASK_EXPCLEAR_OPTIMIZATION(sctx->db_depth_disable_expclear) |
1484 S_028010_DISABLE_SMEM_EXPCLEAR_OPTIMIZATION(sctx->db_stencil_disable_expclear) |
1485 S_028010_DECOMPRESS_Z_ON_FLUSH(sctx->framebuffer.nr_samples >= 4));
1486
1487 db_shader_control = sctx->ps_db_shader_control;
1488
1489 /* Bug workaround for smoothing (overrasterization) on GFX6. */
1490 if (sctx->chip_class == GFX6 && sctx->smoothing_enabled) {
1491 db_shader_control &= C_02880C_Z_ORDER;
1492 db_shader_control |= S_02880C_Z_ORDER(V_02880C_LATE_Z);
1493 }
1494
1495 /* Disable the gl_SampleMask fragment shader output if MSAA is disabled. */
1496 if (!rs->multisample_enable)
1497 db_shader_control &= C_02880C_MASK_EXPORT_ENABLE;
1498
1499 if (sctx->screen->has_rbplus &&
1500 !sctx->screen->rbplus_allowed)
1501 db_shader_control |= S_02880C_DUAL_QUAD_DISABLE(1);
1502
1503 radeon_opt_set_context_reg(sctx, R_02880C_DB_SHADER_CONTROL,
1504 SI_TRACKED_DB_SHADER_CONTROL, db_shader_control);
1505
1506 if (initial_cdw != sctx->gfx_cs->current.cdw)
1507 sctx->context_roll = true;
1508 }
1509
1510 /*
1511 * format translation
1512 */
1513 static uint32_t si_translate_colorformat(enum pipe_format format)
1514 {
1515 const struct util_format_description *desc = util_format_description(format);
1516 if (!desc)
1517 return V_028C70_COLOR_INVALID;
1518
1519 #define HAS_SIZE(x,y,z,w) \
1520 (desc->channel[0].size == (x) && desc->channel[1].size == (y) && \
1521 desc->channel[2].size == (z) && desc->channel[3].size == (w))
1522
1523 if (format == PIPE_FORMAT_R11G11B10_FLOAT) /* isn't plain */
1524 return V_028C70_COLOR_10_11_11;
1525
1526 if (desc->layout != UTIL_FORMAT_LAYOUT_PLAIN)
1527 return V_028C70_COLOR_INVALID;
1528
1529 /* hw cannot support mixed formats (except depth/stencil, since
1530 * stencil is not written to). */
1531 if (desc->is_mixed && desc->colorspace != UTIL_FORMAT_COLORSPACE_ZS)
1532 return V_028C70_COLOR_INVALID;
1533
1534 switch (desc->nr_channels) {
1535 case 1:
1536 switch (desc->channel[0].size) {
1537 case 8:
1538 return V_028C70_COLOR_8;
1539 case 16:
1540 return V_028C70_COLOR_16;
1541 case 32:
1542 return V_028C70_COLOR_32;
1543 }
1544 break;
1545 case 2:
1546 if (desc->channel[0].size == desc->channel[1].size) {
1547 switch (desc->channel[0].size) {
1548 case 8:
1549 return V_028C70_COLOR_8_8;
1550 case 16:
1551 return V_028C70_COLOR_16_16;
1552 case 32:
1553 return V_028C70_COLOR_32_32;
1554 }
1555 } else if (HAS_SIZE(8,24,0,0)) {
1556 return V_028C70_COLOR_24_8;
1557 } else if (HAS_SIZE(24,8,0,0)) {
1558 return V_028C70_COLOR_8_24;
1559 }
1560 break;
1561 case 3:
1562 if (HAS_SIZE(5,6,5,0)) {
1563 return V_028C70_COLOR_5_6_5;
1564 } else if (HAS_SIZE(32,8,24,0)) {
1565 return V_028C70_COLOR_X24_8_32_FLOAT;
1566 }
1567 break;
1568 case 4:
1569 if (desc->channel[0].size == desc->channel[1].size &&
1570 desc->channel[0].size == desc->channel[2].size &&
1571 desc->channel[0].size == desc->channel[3].size) {
1572 switch (desc->channel[0].size) {
1573 case 4:
1574 return V_028C70_COLOR_4_4_4_4;
1575 case 8:
1576 return V_028C70_COLOR_8_8_8_8;
1577 case 16:
1578 return V_028C70_COLOR_16_16_16_16;
1579 case 32:
1580 return V_028C70_COLOR_32_32_32_32;
1581 }
1582 } else if (HAS_SIZE(5,5,5,1)) {
1583 return V_028C70_COLOR_1_5_5_5;
1584 } else if (HAS_SIZE(1,5,5,5)) {
1585 return V_028C70_COLOR_5_5_5_1;
1586 } else if (HAS_SIZE(10,10,10,2)) {
1587 return V_028C70_COLOR_2_10_10_10;
1588 }
1589 break;
1590 }
1591 return V_028C70_COLOR_INVALID;
1592 }
1593
1594 static uint32_t si_colorformat_endian_swap(uint32_t colorformat)
1595 {
1596 if (SI_BIG_ENDIAN) {
1597 switch(colorformat) {
1598 /* 8-bit buffers. */
1599 case V_028C70_COLOR_8:
1600 return V_028C70_ENDIAN_NONE;
1601
1602 /* 16-bit buffers. */
1603 case V_028C70_COLOR_5_6_5:
1604 case V_028C70_COLOR_1_5_5_5:
1605 case V_028C70_COLOR_4_4_4_4:
1606 case V_028C70_COLOR_16:
1607 case V_028C70_COLOR_8_8:
1608 return V_028C70_ENDIAN_8IN16;
1609
1610 /* 32-bit buffers. */
1611 case V_028C70_COLOR_8_8_8_8:
1612 case V_028C70_COLOR_2_10_10_10:
1613 case V_028C70_COLOR_8_24:
1614 case V_028C70_COLOR_24_8:
1615 case V_028C70_COLOR_16_16:
1616 return V_028C70_ENDIAN_8IN32;
1617
1618 /* 64-bit buffers. */
1619 case V_028C70_COLOR_16_16_16_16:
1620 return V_028C70_ENDIAN_8IN16;
1621
1622 case V_028C70_COLOR_32_32:
1623 return V_028C70_ENDIAN_8IN32;
1624
1625 /* 128-bit buffers. */
1626 case V_028C70_COLOR_32_32_32_32:
1627 return V_028C70_ENDIAN_8IN32;
1628 default:
1629 return V_028C70_ENDIAN_NONE; /* Unsupported. */
1630 }
1631 } else {
1632 return V_028C70_ENDIAN_NONE;
1633 }
1634 }
1635
1636 static uint32_t si_translate_dbformat(enum pipe_format format)
1637 {
1638 switch (format) {
1639 case PIPE_FORMAT_Z16_UNORM:
1640 return V_028040_Z_16;
1641 case PIPE_FORMAT_S8_UINT_Z24_UNORM:
1642 case PIPE_FORMAT_X8Z24_UNORM:
1643 case PIPE_FORMAT_Z24X8_UNORM:
1644 case PIPE_FORMAT_Z24_UNORM_S8_UINT:
1645 return V_028040_Z_24; /* deprecated on AMD GCN */
1646 case PIPE_FORMAT_Z32_FLOAT:
1647 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
1648 return V_028040_Z_32_FLOAT;
1649 default:
1650 return V_028040_Z_INVALID;
1651 }
1652 }
1653
1654 /*
1655 * Texture translation
1656 */
1657
1658 static uint32_t si_translate_texformat(struct pipe_screen *screen,
1659 enum pipe_format format,
1660 const struct util_format_description *desc,
1661 int first_non_void)
1662 {
1663 struct si_screen *sscreen = (struct si_screen*)screen;
1664 bool uniform = true;
1665 int i;
1666
1667 assert(sscreen->info.chip_class <= GFX9);
1668
1669 /* Colorspace (return non-RGB formats directly). */
1670 switch (desc->colorspace) {
1671 /* Depth stencil formats */
1672 case UTIL_FORMAT_COLORSPACE_ZS:
1673 switch (format) {
1674 case PIPE_FORMAT_Z16_UNORM:
1675 return V_008F14_IMG_DATA_FORMAT_16;
1676 case PIPE_FORMAT_X24S8_UINT:
1677 case PIPE_FORMAT_S8X24_UINT:
1678 /*
1679 * Implemented as an 8_8_8_8 data format to fix texture
1680 * gathers in stencil sampling. This affects at least
1681 * GL45-CTS.texture_cube_map_array.sampling on GFX8.
1682 */
1683 if (sscreen->info.chip_class <= GFX8)
1684 return V_008F14_IMG_DATA_FORMAT_8_8_8_8;
1685
1686 if (format == PIPE_FORMAT_X24S8_UINT)
1687 return V_008F14_IMG_DATA_FORMAT_8_24;
1688 else
1689 return V_008F14_IMG_DATA_FORMAT_24_8;
1690 case PIPE_FORMAT_Z24X8_UNORM:
1691 case PIPE_FORMAT_Z24_UNORM_S8_UINT:
1692 return V_008F14_IMG_DATA_FORMAT_8_24;
1693 case PIPE_FORMAT_X8Z24_UNORM:
1694 case PIPE_FORMAT_S8_UINT_Z24_UNORM:
1695 return V_008F14_IMG_DATA_FORMAT_24_8;
1696 case PIPE_FORMAT_S8_UINT:
1697 return V_008F14_IMG_DATA_FORMAT_8;
1698 case PIPE_FORMAT_Z32_FLOAT:
1699 return V_008F14_IMG_DATA_FORMAT_32;
1700 case PIPE_FORMAT_X32_S8X24_UINT:
1701 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
1702 return V_008F14_IMG_DATA_FORMAT_X24_8_32;
1703 default:
1704 goto out_unknown;
1705 }
1706
1707 case UTIL_FORMAT_COLORSPACE_YUV:
1708 goto out_unknown; /* TODO */
1709
1710 case UTIL_FORMAT_COLORSPACE_SRGB:
1711 if (desc->nr_channels != 4 && desc->nr_channels != 1)
1712 goto out_unknown;
1713 break;
1714
1715 default:
1716 break;
1717 }
1718
1719 if (desc->layout == UTIL_FORMAT_LAYOUT_RGTC) {
1720 if (!sscreen->info.has_format_bc1_through_bc7)
1721 goto out_unknown;
1722
1723 switch (format) {
1724 case PIPE_FORMAT_RGTC1_SNORM:
1725 case PIPE_FORMAT_LATC1_SNORM:
1726 case PIPE_FORMAT_RGTC1_UNORM:
1727 case PIPE_FORMAT_LATC1_UNORM:
1728 return V_008F14_IMG_DATA_FORMAT_BC4;
1729 case PIPE_FORMAT_RGTC2_SNORM:
1730 case PIPE_FORMAT_LATC2_SNORM:
1731 case PIPE_FORMAT_RGTC2_UNORM:
1732 case PIPE_FORMAT_LATC2_UNORM:
1733 return V_008F14_IMG_DATA_FORMAT_BC5;
1734 default:
1735 goto out_unknown;
1736 }
1737 }
1738
1739 if (desc->layout == UTIL_FORMAT_LAYOUT_ETC &&
1740 (sscreen->info.family == CHIP_STONEY ||
1741 sscreen->info.family == CHIP_VEGA10 ||
1742 sscreen->info.family == CHIP_RAVEN)) {
1743 switch (format) {
1744 case PIPE_FORMAT_ETC1_RGB8:
1745 case PIPE_FORMAT_ETC2_RGB8:
1746 case PIPE_FORMAT_ETC2_SRGB8:
1747 return V_008F14_IMG_DATA_FORMAT_ETC2_RGB;
1748 case PIPE_FORMAT_ETC2_RGB8A1:
1749 case PIPE_FORMAT_ETC2_SRGB8A1:
1750 return V_008F14_IMG_DATA_FORMAT_ETC2_RGBA1;
1751 case PIPE_FORMAT_ETC2_RGBA8:
1752 case PIPE_FORMAT_ETC2_SRGBA8:
1753 return V_008F14_IMG_DATA_FORMAT_ETC2_RGBA;
1754 case PIPE_FORMAT_ETC2_R11_UNORM:
1755 case PIPE_FORMAT_ETC2_R11_SNORM:
1756 return V_008F14_IMG_DATA_FORMAT_ETC2_R;
1757 case PIPE_FORMAT_ETC2_RG11_UNORM:
1758 case PIPE_FORMAT_ETC2_RG11_SNORM:
1759 return V_008F14_IMG_DATA_FORMAT_ETC2_RG;
1760 default:
1761 goto out_unknown;
1762 }
1763 }
1764
1765 if (desc->layout == UTIL_FORMAT_LAYOUT_BPTC) {
1766 if (!sscreen->info.has_format_bc1_through_bc7)
1767 goto out_unknown;
1768
1769 switch (format) {
1770 case PIPE_FORMAT_BPTC_RGBA_UNORM:
1771 case PIPE_FORMAT_BPTC_SRGBA:
1772 return V_008F14_IMG_DATA_FORMAT_BC7;
1773 case PIPE_FORMAT_BPTC_RGB_FLOAT:
1774 case PIPE_FORMAT_BPTC_RGB_UFLOAT:
1775 return V_008F14_IMG_DATA_FORMAT_BC6;
1776 default:
1777 goto out_unknown;
1778 }
1779 }
1780
1781 if (desc->layout == UTIL_FORMAT_LAYOUT_SUBSAMPLED) {
1782 switch (format) {
1783 case PIPE_FORMAT_R8G8_B8G8_UNORM:
1784 case PIPE_FORMAT_G8R8_B8R8_UNORM:
1785 return V_008F14_IMG_DATA_FORMAT_GB_GR;
1786 case PIPE_FORMAT_G8R8_G8B8_UNORM:
1787 case PIPE_FORMAT_R8G8_R8B8_UNORM:
1788 return V_008F14_IMG_DATA_FORMAT_BG_RG;
1789 default:
1790 goto out_unknown;
1791 }
1792 }
1793
1794 if (desc->layout == UTIL_FORMAT_LAYOUT_S3TC) {
1795 if (!sscreen->info.has_format_bc1_through_bc7)
1796 goto out_unknown;
1797
1798 switch (format) {
1799 case PIPE_FORMAT_DXT1_RGB:
1800 case PIPE_FORMAT_DXT1_RGBA:
1801 case PIPE_FORMAT_DXT1_SRGB:
1802 case PIPE_FORMAT_DXT1_SRGBA:
1803 return V_008F14_IMG_DATA_FORMAT_BC1;
1804 case PIPE_FORMAT_DXT3_RGBA:
1805 case PIPE_FORMAT_DXT3_SRGBA:
1806 return V_008F14_IMG_DATA_FORMAT_BC2;
1807 case PIPE_FORMAT_DXT5_RGBA:
1808 case PIPE_FORMAT_DXT5_SRGBA:
1809 return V_008F14_IMG_DATA_FORMAT_BC3;
1810 default:
1811 goto out_unknown;
1812 }
1813 }
1814
1815 if (format == PIPE_FORMAT_R9G9B9E5_FLOAT) {
1816 return V_008F14_IMG_DATA_FORMAT_5_9_9_9;
1817 } else if (format == PIPE_FORMAT_R11G11B10_FLOAT) {
1818 return V_008F14_IMG_DATA_FORMAT_10_11_11;
1819 }
1820
1821 /* R8G8Bx_SNORM - TODO CxV8U8 */
1822
1823 /* hw cannot support mixed formats (except depth/stencil, since only
1824 * depth is read).*/
1825 if (desc->is_mixed && desc->colorspace != UTIL_FORMAT_COLORSPACE_ZS)
1826 goto out_unknown;
1827
1828 /* See whether the components are of the same size. */
1829 for (i = 1; i < desc->nr_channels; i++) {
1830 uniform = uniform && desc->channel[0].size == desc->channel[i].size;
1831 }
1832
1833 /* Non-uniform formats. */
1834 if (!uniform) {
1835 switch(desc->nr_channels) {
1836 case 3:
1837 if (desc->channel[0].size == 5 &&
1838 desc->channel[1].size == 6 &&
1839 desc->channel[2].size == 5) {
1840 return V_008F14_IMG_DATA_FORMAT_5_6_5;
1841 }
1842 goto out_unknown;
1843 case 4:
1844 if (desc->channel[0].size == 5 &&
1845 desc->channel[1].size == 5 &&
1846 desc->channel[2].size == 5 &&
1847 desc->channel[3].size == 1) {
1848 return V_008F14_IMG_DATA_FORMAT_1_5_5_5;
1849 }
1850 if (desc->channel[0].size == 1 &&
1851 desc->channel[1].size == 5 &&
1852 desc->channel[2].size == 5 &&
1853 desc->channel[3].size == 5) {
1854 return V_008F14_IMG_DATA_FORMAT_5_5_5_1;
1855 }
1856 if (desc->channel[0].size == 10 &&
1857 desc->channel[1].size == 10 &&
1858 desc->channel[2].size == 10 &&
1859 desc->channel[3].size == 2) {
1860 return V_008F14_IMG_DATA_FORMAT_2_10_10_10;
1861 }
1862 goto out_unknown;
1863 }
1864 goto out_unknown;
1865 }
1866
1867 if (first_non_void < 0 || first_non_void > 3)
1868 goto out_unknown;
1869
1870 /* uniform formats */
1871 switch (desc->channel[first_non_void].size) {
1872 case 4:
1873 switch (desc->nr_channels) {
1874 #if 0 /* Not supported for render targets */
1875 case 2:
1876 return V_008F14_IMG_DATA_FORMAT_4_4;
1877 #endif
1878 case 4:
1879 return V_008F14_IMG_DATA_FORMAT_4_4_4_4;
1880 }
1881 break;
1882 case 8:
1883 switch (desc->nr_channels) {
1884 case 1:
1885 return V_008F14_IMG_DATA_FORMAT_8;
1886 case 2:
1887 return V_008F14_IMG_DATA_FORMAT_8_8;
1888 case 4:
1889 return V_008F14_IMG_DATA_FORMAT_8_8_8_8;
1890 }
1891 break;
1892 case 16:
1893 switch (desc->nr_channels) {
1894 case 1:
1895 return V_008F14_IMG_DATA_FORMAT_16;
1896 case 2:
1897 return V_008F14_IMG_DATA_FORMAT_16_16;
1898 case 4:
1899 return V_008F14_IMG_DATA_FORMAT_16_16_16_16;
1900 }
1901 break;
1902 case 32:
1903 switch (desc->nr_channels) {
1904 case 1:
1905 return V_008F14_IMG_DATA_FORMAT_32;
1906 case 2:
1907 return V_008F14_IMG_DATA_FORMAT_32_32;
1908 #if 0 /* Not supported for render targets */
1909 case 3:
1910 return V_008F14_IMG_DATA_FORMAT_32_32_32;
1911 #endif
1912 case 4:
1913 return V_008F14_IMG_DATA_FORMAT_32_32_32_32;
1914 }
1915 }
1916
1917 out_unknown:
1918 return ~0;
1919 }
1920
1921 static unsigned si_tex_wrap(unsigned wrap)
1922 {
1923 switch (wrap) {
1924 default:
1925 case PIPE_TEX_WRAP_REPEAT:
1926 return V_008F30_SQ_TEX_WRAP;
1927 case PIPE_TEX_WRAP_CLAMP:
1928 return V_008F30_SQ_TEX_CLAMP_HALF_BORDER;
1929 case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
1930 return V_008F30_SQ_TEX_CLAMP_LAST_TEXEL;
1931 case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
1932 return V_008F30_SQ_TEX_CLAMP_BORDER;
1933 case PIPE_TEX_WRAP_MIRROR_REPEAT:
1934 return V_008F30_SQ_TEX_MIRROR;
1935 case PIPE_TEX_WRAP_MIRROR_CLAMP:
1936 return V_008F30_SQ_TEX_MIRROR_ONCE_HALF_BORDER;
1937 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
1938 return V_008F30_SQ_TEX_MIRROR_ONCE_LAST_TEXEL;
1939 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
1940 return V_008F30_SQ_TEX_MIRROR_ONCE_BORDER;
1941 }
1942 }
1943
1944 static unsigned si_tex_mipfilter(unsigned filter)
1945 {
1946 switch (filter) {
1947 case PIPE_TEX_MIPFILTER_NEAREST:
1948 return V_008F38_SQ_TEX_Z_FILTER_POINT;
1949 case PIPE_TEX_MIPFILTER_LINEAR:
1950 return V_008F38_SQ_TEX_Z_FILTER_LINEAR;
1951 default:
1952 case PIPE_TEX_MIPFILTER_NONE:
1953 return V_008F38_SQ_TEX_Z_FILTER_NONE;
1954 }
1955 }
1956
1957 static unsigned si_tex_compare(unsigned compare)
1958 {
1959 switch (compare) {
1960 default:
1961 case PIPE_FUNC_NEVER:
1962 return V_008F30_SQ_TEX_DEPTH_COMPARE_NEVER;
1963 case PIPE_FUNC_LESS:
1964 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESS;
1965 case PIPE_FUNC_EQUAL:
1966 return V_008F30_SQ_TEX_DEPTH_COMPARE_EQUAL;
1967 case PIPE_FUNC_LEQUAL:
1968 return V_008F30_SQ_TEX_DEPTH_COMPARE_LESSEQUAL;
1969 case PIPE_FUNC_GREATER:
1970 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATER;
1971 case PIPE_FUNC_NOTEQUAL:
1972 return V_008F30_SQ_TEX_DEPTH_COMPARE_NOTEQUAL;
1973 case PIPE_FUNC_GEQUAL:
1974 return V_008F30_SQ_TEX_DEPTH_COMPARE_GREATEREQUAL;
1975 case PIPE_FUNC_ALWAYS:
1976 return V_008F30_SQ_TEX_DEPTH_COMPARE_ALWAYS;
1977 }
1978 }
1979
1980 static unsigned si_tex_dim(struct si_screen *sscreen, struct si_texture *tex,
1981 unsigned view_target, unsigned nr_samples)
1982 {
1983 unsigned res_target = tex->buffer.b.b.target;
1984
1985 if (view_target == PIPE_TEXTURE_CUBE ||
1986 view_target == PIPE_TEXTURE_CUBE_ARRAY)
1987 res_target = view_target;
1988 /* If interpreting cubemaps as something else, set 2D_ARRAY. */
1989 else if (res_target == PIPE_TEXTURE_CUBE ||
1990 res_target == PIPE_TEXTURE_CUBE_ARRAY)
1991 res_target = PIPE_TEXTURE_2D_ARRAY;
1992
1993 /* GFX9 allocates 1D textures as 2D. */
1994 if ((res_target == PIPE_TEXTURE_1D ||
1995 res_target == PIPE_TEXTURE_1D_ARRAY) &&
1996 sscreen->info.chip_class == GFX9 &&
1997 tex->surface.u.gfx9.resource_type == RADEON_RESOURCE_2D) {
1998 if (res_target == PIPE_TEXTURE_1D)
1999 res_target = PIPE_TEXTURE_2D;
2000 else
2001 res_target = PIPE_TEXTURE_2D_ARRAY;
2002 }
2003
2004 switch (res_target) {
2005 default:
2006 case PIPE_TEXTURE_1D:
2007 return V_008F1C_SQ_RSRC_IMG_1D;
2008 case PIPE_TEXTURE_1D_ARRAY:
2009 return V_008F1C_SQ_RSRC_IMG_1D_ARRAY;
2010 case PIPE_TEXTURE_2D:
2011 case PIPE_TEXTURE_RECT:
2012 return nr_samples > 1 ? V_008F1C_SQ_RSRC_IMG_2D_MSAA :
2013 V_008F1C_SQ_RSRC_IMG_2D;
2014 case PIPE_TEXTURE_2D_ARRAY:
2015 return nr_samples > 1 ? V_008F1C_SQ_RSRC_IMG_2D_MSAA_ARRAY :
2016 V_008F1C_SQ_RSRC_IMG_2D_ARRAY;
2017 case PIPE_TEXTURE_3D:
2018 return V_008F1C_SQ_RSRC_IMG_3D;
2019 case PIPE_TEXTURE_CUBE:
2020 case PIPE_TEXTURE_CUBE_ARRAY:
2021 return V_008F1C_SQ_RSRC_IMG_CUBE;
2022 }
2023 }
2024
2025 /*
2026 * Format support testing
2027 */
2028
2029 static bool si_is_sampler_format_supported(struct pipe_screen *screen, enum pipe_format format)
2030 {
2031 struct si_screen *sscreen = (struct si_screen *)screen;
2032
2033 if (sscreen->info.chip_class >= GFX10) {
2034 const struct gfx10_format *fmt = &gfx10_format_table[format];
2035 if (!fmt->img_format || fmt->buffers_only)
2036 return false;
2037 return true;
2038 }
2039
2040 const struct util_format_description *desc = util_format_description(format);
2041 if (!desc)
2042 return false;
2043
2044 return si_translate_texformat(screen, format, desc,
2045 util_format_get_first_non_void_channel(format)) != ~0U;
2046 }
2047
2048 static uint32_t si_translate_buffer_dataformat(struct pipe_screen *screen,
2049 const struct util_format_description *desc,
2050 int first_non_void)
2051 {
2052 int i;
2053
2054 assert(((struct si_screen *)screen)->info.chip_class <= GFX9);
2055
2056 if (desc->format == PIPE_FORMAT_R11G11B10_FLOAT)
2057 return V_008F0C_BUF_DATA_FORMAT_10_11_11;
2058
2059 assert(first_non_void >= 0);
2060
2061 if (desc->nr_channels == 4 &&
2062 desc->channel[0].size == 10 &&
2063 desc->channel[1].size == 10 &&
2064 desc->channel[2].size == 10 &&
2065 desc->channel[3].size == 2)
2066 return V_008F0C_BUF_DATA_FORMAT_2_10_10_10;
2067
2068 /* See whether the components are of the same size. */
2069 for (i = 0; i < desc->nr_channels; i++) {
2070 if (desc->channel[first_non_void].size != desc->channel[i].size)
2071 return V_008F0C_BUF_DATA_FORMAT_INVALID;
2072 }
2073
2074 switch (desc->channel[first_non_void].size) {
2075 case 8:
2076 switch (desc->nr_channels) {
2077 case 1:
2078 case 3: /* 3 loads */
2079 return V_008F0C_BUF_DATA_FORMAT_8;
2080 case 2:
2081 return V_008F0C_BUF_DATA_FORMAT_8_8;
2082 case 4:
2083 return V_008F0C_BUF_DATA_FORMAT_8_8_8_8;
2084 }
2085 break;
2086 case 16:
2087 switch (desc->nr_channels) {
2088 case 1:
2089 case 3: /* 3 loads */
2090 return V_008F0C_BUF_DATA_FORMAT_16;
2091 case 2:
2092 return V_008F0C_BUF_DATA_FORMAT_16_16;
2093 case 4:
2094 return V_008F0C_BUF_DATA_FORMAT_16_16_16_16;
2095 }
2096 break;
2097 case 32:
2098 switch (desc->nr_channels) {
2099 case 1:
2100 return V_008F0C_BUF_DATA_FORMAT_32;
2101 case 2:
2102 return V_008F0C_BUF_DATA_FORMAT_32_32;
2103 case 3:
2104 return V_008F0C_BUF_DATA_FORMAT_32_32_32;
2105 case 4:
2106 return V_008F0C_BUF_DATA_FORMAT_32_32_32_32;
2107 }
2108 break;
2109 case 64:
2110 /* Legacy double formats. */
2111 switch (desc->nr_channels) {
2112 case 1: /* 1 load */
2113 return V_008F0C_BUF_DATA_FORMAT_32_32;
2114 case 2: /* 1 load */
2115 return V_008F0C_BUF_DATA_FORMAT_32_32_32_32;
2116 case 3: /* 3 loads */
2117 return V_008F0C_BUF_DATA_FORMAT_32_32;
2118 case 4: /* 2 loads */
2119 return V_008F0C_BUF_DATA_FORMAT_32_32_32_32;
2120 }
2121 break;
2122 }
2123
2124 return V_008F0C_BUF_DATA_FORMAT_INVALID;
2125 }
2126
2127 static uint32_t si_translate_buffer_numformat(struct pipe_screen *screen,
2128 const struct util_format_description *desc,
2129 int first_non_void)
2130 {
2131 assert(((struct si_screen *)screen)->info.chip_class <= GFX9);
2132
2133 if (desc->format == PIPE_FORMAT_R11G11B10_FLOAT)
2134 return V_008F0C_BUF_NUM_FORMAT_FLOAT;
2135
2136 assert(first_non_void >= 0);
2137
2138 switch (desc->channel[first_non_void].type) {
2139 case UTIL_FORMAT_TYPE_SIGNED:
2140 case UTIL_FORMAT_TYPE_FIXED:
2141 if (desc->channel[first_non_void].size >= 32 ||
2142 desc->channel[first_non_void].pure_integer)
2143 return V_008F0C_BUF_NUM_FORMAT_SINT;
2144 else if (desc->channel[first_non_void].normalized)
2145 return V_008F0C_BUF_NUM_FORMAT_SNORM;
2146 else
2147 return V_008F0C_BUF_NUM_FORMAT_SSCALED;
2148 break;
2149 case UTIL_FORMAT_TYPE_UNSIGNED:
2150 if (desc->channel[first_non_void].size >= 32 ||
2151 desc->channel[first_non_void].pure_integer)
2152 return V_008F0C_BUF_NUM_FORMAT_UINT;
2153 else if (desc->channel[first_non_void].normalized)
2154 return V_008F0C_BUF_NUM_FORMAT_UNORM;
2155 else
2156 return V_008F0C_BUF_NUM_FORMAT_USCALED;
2157 break;
2158 case UTIL_FORMAT_TYPE_FLOAT:
2159 default:
2160 return V_008F0C_BUF_NUM_FORMAT_FLOAT;
2161 }
2162 }
2163
2164 static unsigned si_is_vertex_format_supported(struct pipe_screen *screen,
2165 enum pipe_format format,
2166 unsigned usage)
2167 {
2168 struct si_screen *sscreen = (struct si_screen *)screen;
2169 const struct util_format_description *desc;
2170 int first_non_void;
2171 unsigned data_format;
2172
2173 assert((usage & ~(PIPE_BIND_SHADER_IMAGE |
2174 PIPE_BIND_SAMPLER_VIEW |
2175 PIPE_BIND_VERTEX_BUFFER)) == 0);
2176
2177 desc = util_format_description(format);
2178 if (!desc)
2179 return 0;
2180
2181 /* There are no native 8_8_8 or 16_16_16 data formats, and we currently
2182 * select 8_8_8_8 and 16_16_16_16 instead. This works reasonably well
2183 * for read-only access (with caveats surrounding bounds checks), but
2184 * obviously fails for write access which we have to implement for
2185 * shader images. Luckily, OpenGL doesn't expect this to be supported
2186 * anyway, and so the only impact is on PBO uploads / downloads, which
2187 * shouldn't be expected to be fast for GL_RGB anyway.
2188 */
2189 if (desc->block.bits == 3 * 8 ||
2190 desc->block.bits == 3 * 16) {
2191 if (usage & (PIPE_BIND_SHADER_IMAGE | PIPE_BIND_SAMPLER_VIEW)) {
2192 usage &= ~(PIPE_BIND_SHADER_IMAGE | PIPE_BIND_SAMPLER_VIEW);
2193 if (!usage)
2194 return 0;
2195 }
2196 }
2197
2198 if (sscreen->info.chip_class >= GFX10) {
2199 const struct gfx10_format *fmt = &gfx10_format_table[format];
2200 if (!fmt->img_format || fmt->img_format >= 128)
2201 return 0;
2202 return usage;
2203 }
2204
2205 first_non_void = util_format_get_first_non_void_channel(format);
2206 data_format = si_translate_buffer_dataformat(screen, desc, first_non_void);
2207 if (data_format == V_008F0C_BUF_DATA_FORMAT_INVALID)
2208 return 0;
2209
2210 return usage;
2211 }
2212
2213 static bool si_is_colorbuffer_format_supported(enum pipe_format format)
2214 {
2215 return si_translate_colorformat(format) != V_028C70_COLOR_INVALID &&
2216 si_translate_colorswap(format, false) != ~0U;
2217 }
2218
2219 static bool si_is_zs_format_supported(enum pipe_format format)
2220 {
2221 return si_translate_dbformat(format) != V_028040_Z_INVALID;
2222 }
2223
2224 static bool si_is_format_supported(struct pipe_screen *screen,
2225 enum pipe_format format,
2226 enum pipe_texture_target target,
2227 unsigned sample_count,
2228 unsigned storage_sample_count,
2229 unsigned usage)
2230 {
2231 struct si_screen *sscreen = (struct si_screen *)screen;
2232 unsigned retval = 0;
2233
2234 if (target >= PIPE_MAX_TEXTURE_TYPES) {
2235 PRINT_ERR("radeonsi: unsupported texture type %d\n", target);
2236 return false;
2237 }
2238
2239 if (MAX2(1, sample_count) < MAX2(1, storage_sample_count))
2240 return false;
2241
2242 if (sample_count > 1) {
2243 if (!screen->get_param(screen, PIPE_CAP_TEXTURE_MULTISAMPLE))
2244 return false;
2245
2246 if (usage & PIPE_BIND_SHADER_IMAGE)
2247 return false;
2248
2249 /* Only power-of-two sample counts are supported. */
2250 if (!util_is_power_of_two_or_zero(sample_count) ||
2251 !util_is_power_of_two_or_zero(storage_sample_count))
2252 return false;
2253
2254 /* MSAA support without framebuffer attachments. */
2255 if (format == PIPE_FORMAT_NONE && sample_count <= 16)
2256 return true;
2257
2258 if (!sscreen->info.has_eqaa_surface_allocator ||
2259 util_format_is_depth_or_stencil(format)) {
2260 /* Color without EQAA or depth/stencil. */
2261 if (sample_count > 8 ||
2262 sample_count != storage_sample_count)
2263 return false;
2264 } else {
2265 /* Color with EQAA. */
2266 if (sample_count > 16 ||
2267 storage_sample_count > 8)
2268 return false;
2269 }
2270 }
2271
2272 if (usage & (PIPE_BIND_SAMPLER_VIEW |
2273 PIPE_BIND_SHADER_IMAGE)) {
2274 if (target == PIPE_BUFFER) {
2275 retval |= si_is_vertex_format_supported(
2276 screen, format, usage & (PIPE_BIND_SAMPLER_VIEW |
2277 PIPE_BIND_SHADER_IMAGE));
2278 } else {
2279 if (si_is_sampler_format_supported(screen, format))
2280 retval |= usage & (PIPE_BIND_SAMPLER_VIEW |
2281 PIPE_BIND_SHADER_IMAGE);
2282 }
2283 }
2284
2285 if ((usage & (PIPE_BIND_RENDER_TARGET |
2286 PIPE_BIND_DISPLAY_TARGET |
2287 PIPE_BIND_SCANOUT |
2288 PIPE_BIND_SHARED |
2289 PIPE_BIND_BLENDABLE)) &&
2290 si_is_colorbuffer_format_supported(format)) {
2291 retval |= usage &
2292 (PIPE_BIND_RENDER_TARGET |
2293 PIPE_BIND_DISPLAY_TARGET |
2294 PIPE_BIND_SCANOUT |
2295 PIPE_BIND_SHARED);
2296 if (!util_format_is_pure_integer(format) &&
2297 !util_format_is_depth_or_stencil(format))
2298 retval |= usage & PIPE_BIND_BLENDABLE;
2299 }
2300
2301 if ((usage & PIPE_BIND_DEPTH_STENCIL) &&
2302 si_is_zs_format_supported(format)) {
2303 retval |= PIPE_BIND_DEPTH_STENCIL;
2304 }
2305
2306 if (usage & PIPE_BIND_VERTEX_BUFFER) {
2307 retval |= si_is_vertex_format_supported(screen, format,
2308 PIPE_BIND_VERTEX_BUFFER);
2309 }
2310
2311 if ((usage & PIPE_BIND_LINEAR) &&
2312 !util_format_is_compressed(format) &&
2313 !(usage & PIPE_BIND_DEPTH_STENCIL))
2314 retval |= PIPE_BIND_LINEAR;
2315
2316 return retval == usage;
2317 }
2318
2319 /*
2320 * framebuffer handling
2321 */
2322
2323 static void si_choose_spi_color_formats(struct si_surface *surf,
2324 unsigned format, unsigned swap,
2325 unsigned ntype, bool is_depth)
2326 {
2327 /* Alpha is needed for alpha-to-coverage.
2328 * Blending may be with or without alpha.
2329 */
2330 unsigned normal = 0; /* most optimal, may not support blending or export alpha */
2331 unsigned alpha = 0; /* exports alpha, but may not support blending */
2332 unsigned blend = 0; /* supports blending, but may not export alpha */
2333 unsigned blend_alpha = 0; /* least optimal, supports blending and exports alpha */
2334
2335 /* Choose the SPI color formats. These are required values for RB+.
2336 * Other chips have multiple choices, though they are not necessarily better.
2337 */
2338 switch (format) {
2339 case V_028C70_COLOR_5_6_5:
2340 case V_028C70_COLOR_1_5_5_5:
2341 case V_028C70_COLOR_5_5_5_1:
2342 case V_028C70_COLOR_4_4_4_4:
2343 case V_028C70_COLOR_10_11_11:
2344 case V_028C70_COLOR_11_11_10:
2345 case V_028C70_COLOR_8:
2346 case V_028C70_COLOR_8_8:
2347 case V_028C70_COLOR_8_8_8_8:
2348 case V_028C70_COLOR_10_10_10_2:
2349 case V_028C70_COLOR_2_10_10_10:
2350 if (ntype == V_028C70_NUMBER_UINT)
2351 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_UINT16_ABGR;
2352 else if (ntype == V_028C70_NUMBER_SINT)
2353 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_SINT16_ABGR;
2354 else
2355 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_FP16_ABGR;
2356 break;
2357
2358 case V_028C70_COLOR_16:
2359 case V_028C70_COLOR_16_16:
2360 case V_028C70_COLOR_16_16_16_16:
2361 if (ntype == V_028C70_NUMBER_UNORM ||
2362 ntype == V_028C70_NUMBER_SNORM) {
2363 /* UNORM16 and SNORM16 don't support blending */
2364 if (ntype == V_028C70_NUMBER_UNORM)
2365 normal = alpha = V_028714_SPI_SHADER_UNORM16_ABGR;
2366 else
2367 normal = alpha = V_028714_SPI_SHADER_SNORM16_ABGR;
2368
2369 /* Use 32 bits per channel for blending. */
2370 if (format == V_028C70_COLOR_16) {
2371 if (swap == V_028C70_SWAP_STD) { /* R */
2372 blend = V_028714_SPI_SHADER_32_R;
2373 blend_alpha = V_028714_SPI_SHADER_32_AR;
2374 } else if (swap == V_028C70_SWAP_ALT_REV) /* A */
2375 blend = blend_alpha = V_028714_SPI_SHADER_32_AR;
2376 else
2377 assert(0);
2378 } else if (format == V_028C70_COLOR_16_16) {
2379 if (swap == V_028C70_SWAP_STD) { /* RG */
2380 blend = V_028714_SPI_SHADER_32_GR;
2381 blend_alpha = V_028714_SPI_SHADER_32_ABGR;
2382 } else if (swap == V_028C70_SWAP_ALT) /* RA */
2383 blend = blend_alpha = V_028714_SPI_SHADER_32_AR;
2384 else
2385 assert(0);
2386 } else /* 16_16_16_16 */
2387 blend = blend_alpha = V_028714_SPI_SHADER_32_ABGR;
2388 } else if (ntype == V_028C70_NUMBER_UINT)
2389 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_UINT16_ABGR;
2390 else if (ntype == V_028C70_NUMBER_SINT)
2391 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_SINT16_ABGR;
2392 else if (ntype == V_028C70_NUMBER_FLOAT)
2393 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_FP16_ABGR;
2394 else
2395 assert(0);
2396 break;
2397
2398 case V_028C70_COLOR_32:
2399 if (swap == V_028C70_SWAP_STD) { /* R */
2400 blend = normal = V_028714_SPI_SHADER_32_R;
2401 alpha = blend_alpha = V_028714_SPI_SHADER_32_AR;
2402 } else if (swap == V_028C70_SWAP_ALT_REV) /* A */
2403 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_32_AR;
2404 else
2405 assert(0);
2406 break;
2407
2408 case V_028C70_COLOR_32_32:
2409 if (swap == V_028C70_SWAP_STD) { /* RG */
2410 blend = normal = V_028714_SPI_SHADER_32_GR;
2411 alpha = blend_alpha = V_028714_SPI_SHADER_32_ABGR;
2412 } else if (swap == V_028C70_SWAP_ALT) /* RA */
2413 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_32_AR;
2414 else
2415 assert(0);
2416 break;
2417
2418 case V_028C70_COLOR_32_32_32_32:
2419 case V_028C70_COLOR_8_24:
2420 case V_028C70_COLOR_24_8:
2421 case V_028C70_COLOR_X24_8_32_FLOAT:
2422 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_32_ABGR;
2423 break;
2424
2425 default:
2426 assert(0);
2427 return;
2428 }
2429
2430 /* The DB->CB copy needs 32_ABGR. */
2431 if (is_depth)
2432 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_32_ABGR;
2433
2434 surf->spi_shader_col_format = normal;
2435 surf->spi_shader_col_format_alpha = alpha;
2436 surf->spi_shader_col_format_blend = blend;
2437 surf->spi_shader_col_format_blend_alpha = blend_alpha;
2438 }
2439
2440 static void si_initialize_color_surface(struct si_context *sctx,
2441 struct si_surface *surf)
2442 {
2443 struct si_texture *tex = (struct si_texture*)surf->base.texture;
2444 unsigned color_info, color_attrib;
2445 unsigned format, swap, ntype, endian;
2446 const struct util_format_description *desc;
2447 int firstchan;
2448 unsigned blend_clamp = 0, blend_bypass = 0;
2449
2450 desc = util_format_description(surf->base.format);
2451 for (firstchan = 0; firstchan < 4; firstchan++) {
2452 if (desc->channel[firstchan].type != UTIL_FORMAT_TYPE_VOID) {
2453 break;
2454 }
2455 }
2456 if (firstchan == 4 || desc->channel[firstchan].type == UTIL_FORMAT_TYPE_FLOAT) {
2457 ntype = V_028C70_NUMBER_FLOAT;
2458 } else {
2459 ntype = V_028C70_NUMBER_UNORM;
2460 if (desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB)
2461 ntype = V_028C70_NUMBER_SRGB;
2462 else if (desc->channel[firstchan].type == UTIL_FORMAT_TYPE_SIGNED) {
2463 if (desc->channel[firstchan].pure_integer) {
2464 ntype = V_028C70_NUMBER_SINT;
2465 } else {
2466 assert(desc->channel[firstchan].normalized);
2467 ntype = V_028C70_NUMBER_SNORM;
2468 }
2469 } else if (desc->channel[firstchan].type == UTIL_FORMAT_TYPE_UNSIGNED) {
2470 if (desc->channel[firstchan].pure_integer) {
2471 ntype = V_028C70_NUMBER_UINT;
2472 } else {
2473 assert(desc->channel[firstchan].normalized);
2474 ntype = V_028C70_NUMBER_UNORM;
2475 }
2476 }
2477 }
2478
2479 format = si_translate_colorformat(surf->base.format);
2480 if (format == V_028C70_COLOR_INVALID) {
2481 PRINT_ERR("Invalid CB format: %d, disabling CB.\n", surf->base.format);
2482 }
2483 assert(format != V_028C70_COLOR_INVALID);
2484 swap = si_translate_colorswap(surf->base.format, false);
2485 endian = si_colorformat_endian_swap(format);
2486
2487 /* blend clamp should be set for all NORM/SRGB types */
2488 if (ntype == V_028C70_NUMBER_UNORM ||
2489 ntype == V_028C70_NUMBER_SNORM ||
2490 ntype == V_028C70_NUMBER_SRGB)
2491 blend_clamp = 1;
2492
2493 /* set blend bypass according to docs if SINT/UINT or
2494 8/24 COLOR variants */
2495 if (ntype == V_028C70_NUMBER_UINT || ntype == V_028C70_NUMBER_SINT ||
2496 format == V_028C70_COLOR_8_24 || format == V_028C70_COLOR_24_8 ||
2497 format == V_028C70_COLOR_X24_8_32_FLOAT) {
2498 blend_clamp = 0;
2499 blend_bypass = 1;
2500 }
2501
2502 if (ntype == V_028C70_NUMBER_UINT || ntype == V_028C70_NUMBER_SINT) {
2503 if (format == V_028C70_COLOR_8 ||
2504 format == V_028C70_COLOR_8_8 ||
2505 format == V_028C70_COLOR_8_8_8_8)
2506 surf->color_is_int8 = true;
2507 else if (format == V_028C70_COLOR_10_10_10_2 ||
2508 format == V_028C70_COLOR_2_10_10_10)
2509 surf->color_is_int10 = true;
2510 }
2511
2512 color_info = S_028C70_FORMAT(format) |
2513 S_028C70_COMP_SWAP(swap) |
2514 S_028C70_BLEND_CLAMP(blend_clamp) |
2515 S_028C70_BLEND_BYPASS(blend_bypass) |
2516 S_028C70_SIMPLE_FLOAT(1) |
2517 S_028C70_ROUND_MODE(ntype != V_028C70_NUMBER_UNORM &&
2518 ntype != V_028C70_NUMBER_SNORM &&
2519 ntype != V_028C70_NUMBER_SRGB &&
2520 format != V_028C70_COLOR_8_24 &&
2521 format != V_028C70_COLOR_24_8) |
2522 S_028C70_NUMBER_TYPE(ntype) |
2523 S_028C70_ENDIAN(endian);
2524
2525 /* Intensity is implemented as Red, so treat it that way. */
2526 color_attrib = S_028C74_FORCE_DST_ALPHA_1(desc->swizzle[3] == PIPE_SWIZZLE_1 ||
2527 util_format_is_intensity(surf->base.format));
2528
2529 if (tex->buffer.b.b.nr_samples > 1) {
2530 unsigned log_samples = util_logbase2(tex->buffer.b.b.nr_samples);
2531 unsigned log_fragments = util_logbase2(tex->buffer.b.b.nr_storage_samples);
2532
2533 color_attrib |= S_028C74_NUM_SAMPLES(log_samples) |
2534 S_028C74_NUM_FRAGMENTS(log_fragments);
2535
2536 if (tex->fmask_offset) {
2537 color_info |= S_028C70_COMPRESSION(1);
2538 unsigned fmask_bankh = util_logbase2(tex->surface.u.legacy.fmask.bankh);
2539
2540 if (sctx->chip_class == GFX6) {
2541 /* due to a hw bug, FMASK_BANK_HEIGHT must be set on GFX6 too */
2542 color_attrib |= S_028C74_FMASK_BANK_HEIGHT(fmask_bankh);
2543 }
2544 }
2545 }
2546
2547 if (sctx->chip_class >= GFX10) {
2548 unsigned min_compressed_block_size = V_028C78_MIN_BLOCK_SIZE_32B;
2549
2550 /* amdvlk: [min-compressed-block-size] should be set to 32 for dGPU and
2551 64 for APU because all of our APUs to date use DIMMs which have
2552 a request granularity size of 64B while all other chips have a
2553 32B request size */
2554 if (!sctx->screen->info.has_dedicated_vram)
2555 min_compressed_block_size = V_028C78_MIN_BLOCK_SIZE_64B;
2556
2557 surf->cb_dcc_control =
2558 S_028C78_MAX_UNCOMPRESSED_BLOCK_SIZE(V_028C78_MAX_BLOCK_SIZE_256B) |
2559 S_028C78_MAX_COMPRESSED_BLOCK_SIZE(V_028C78_MAX_BLOCK_SIZE_128B) |
2560 S_028C78_MIN_COMPRESSED_BLOCK_SIZE(min_compressed_block_size) |
2561 S_028C78_INDEPENDENT_64B_BLOCKS(0) |
2562 S_028C78_INDEPENDENT_128B_BLOCKS(1);
2563 } else if (sctx->chip_class >= GFX8) {
2564 unsigned max_uncompressed_block_size = V_028C78_MAX_BLOCK_SIZE_256B;
2565 unsigned min_compressed_block_size = V_028C78_MIN_BLOCK_SIZE_32B;
2566
2567 /* amdvlk: [min-compressed-block-size] should be set to 32 for dGPU and
2568 64 for APU because all of our APUs to date use DIMMs which have
2569 a request granularity size of 64B while all other chips have a
2570 32B request size */
2571 if (!sctx->screen->info.has_dedicated_vram)
2572 min_compressed_block_size = V_028C78_MIN_BLOCK_SIZE_64B;
2573
2574 if (tex->buffer.b.b.nr_storage_samples > 1) {
2575 if (tex->surface.bpe == 1)
2576 max_uncompressed_block_size = V_028C78_MAX_BLOCK_SIZE_64B;
2577 else if (tex->surface.bpe == 2)
2578 max_uncompressed_block_size = V_028C78_MAX_BLOCK_SIZE_128B;
2579 }
2580
2581 surf->cb_dcc_control = S_028C78_MAX_UNCOMPRESSED_BLOCK_SIZE(max_uncompressed_block_size) |
2582 S_028C78_MIN_COMPRESSED_BLOCK_SIZE(min_compressed_block_size) |
2583 S_028C78_INDEPENDENT_64B_BLOCKS(1);
2584 }
2585
2586 /* This must be set for fast clear to work without FMASK. */
2587 if (!tex->surface.fmask_size && sctx->chip_class == GFX6) {
2588 unsigned bankh = util_logbase2(tex->surface.u.legacy.bankh);
2589 color_attrib |= S_028C74_FMASK_BANK_HEIGHT(bankh);
2590 }
2591
2592 /* GFX10 field has the same base shift as the GFX6 field */
2593 unsigned color_view = S_028C6C_SLICE_START(surf->base.u.tex.first_layer) |
2594 S_028C6C_SLICE_MAX_GFX10(surf->base.u.tex.last_layer);
2595 unsigned mip0_depth = util_max_layer(&tex->buffer.b.b, 0);
2596
2597 if (sctx->chip_class >= GFX10) {
2598 color_view |= S_028C6C_MIP_LEVEL_GFX10(surf->base.u.tex.level);
2599
2600 surf->cb_color_attrib3 = S_028EE0_MIP0_DEPTH(mip0_depth) |
2601 S_028EE0_RESOURCE_TYPE(tex->surface.u.gfx9.resource_type) |
2602 S_028EE0_RESOURCE_LEVEL(1);
2603 } else if (sctx->chip_class == GFX9) {
2604 color_view |= S_028C6C_MIP_LEVEL_GFX9(surf->base.u.tex.level);
2605 color_attrib |= S_028C74_MIP0_DEPTH(mip0_depth) |
2606 S_028C74_RESOURCE_TYPE(tex->surface.u.gfx9.resource_type);
2607 }
2608
2609 if (sctx->chip_class >= GFX9) {
2610 surf->cb_color_attrib2 = S_028C68_MIP0_WIDTH(surf->width0 - 1) |
2611 S_028C68_MIP0_HEIGHT(surf->height0 - 1) |
2612 S_028C68_MAX_MIP(tex->buffer.b.b.last_level);
2613 }
2614
2615 surf->cb_color_view = color_view;
2616 surf->cb_color_info = color_info;
2617 surf->cb_color_attrib = color_attrib;
2618
2619 /* Determine pixel shader export format */
2620 si_choose_spi_color_formats(surf, format, swap, ntype, tex->is_depth);
2621
2622 surf->color_initialized = true;
2623 }
2624
2625 static void si_init_depth_surface(struct si_context *sctx,
2626 struct si_surface *surf)
2627 {
2628 struct si_texture *tex = (struct si_texture*)surf->base.texture;
2629 unsigned level = surf->base.u.tex.level;
2630 unsigned format, stencil_format;
2631 uint32_t z_info, s_info;
2632
2633 format = si_translate_dbformat(tex->db_render_format);
2634 stencil_format = tex->surface.has_stencil ?
2635 V_028044_STENCIL_8 : V_028044_STENCIL_INVALID;
2636
2637 assert(format != V_028040_Z_INVALID);
2638 if (format == V_028040_Z_INVALID)
2639 PRINT_ERR("Invalid DB format: %d, disabling DB.\n", tex->buffer.b.b.format);
2640
2641 surf->db_depth_view = S_028008_SLICE_START(surf->base.u.tex.first_layer) |
2642 S_028008_SLICE_MAX(surf->base.u.tex.last_layer);
2643 surf->db_htile_data_base = 0;
2644 surf->db_htile_surface = 0;
2645
2646 if (sctx->chip_class >= GFX10) {
2647 surf->db_depth_view |= S_028008_SLICE_START_HI(surf->base.u.tex.first_layer >> 11) |
2648 S_028008_SLICE_MAX_HI(surf->base.u.tex.last_layer >> 11);
2649 }
2650
2651 if (sctx->chip_class >= GFX9) {
2652 assert(tex->surface.u.gfx9.surf_offset == 0);
2653 surf->db_depth_base = tex->buffer.gpu_address >> 8;
2654 surf->db_stencil_base = (tex->buffer.gpu_address +
2655 tex->surface.u.gfx9.stencil_offset) >> 8;
2656 z_info = S_028038_FORMAT(format) |
2657 S_028038_NUM_SAMPLES(util_logbase2(tex->buffer.b.b.nr_samples)) |
2658 S_028038_SW_MODE(tex->surface.u.gfx9.surf.swizzle_mode) |
2659 S_028038_MAXMIP(tex->buffer.b.b.last_level);
2660 s_info = S_02803C_FORMAT(stencil_format) |
2661 S_02803C_SW_MODE(tex->surface.u.gfx9.stencil.swizzle_mode);
2662
2663 if (sctx->chip_class == GFX9) {
2664 surf->db_z_info2 = S_028068_EPITCH(tex->surface.u.gfx9.surf.epitch);
2665 surf->db_stencil_info2 = S_02806C_EPITCH(tex->surface.u.gfx9.stencil.epitch);
2666 }
2667 surf->db_depth_view |= S_028008_MIPID(level);
2668 surf->db_depth_size = S_02801C_X_MAX(tex->buffer.b.b.width0 - 1) |
2669 S_02801C_Y_MAX(tex->buffer.b.b.height0 - 1);
2670
2671 if (si_htile_enabled(tex, level, PIPE_MASK_ZS)) {
2672 z_info |= S_028038_TILE_SURFACE_ENABLE(1) |
2673 S_028038_ALLOW_EXPCLEAR(1);
2674
2675 if (tex->tc_compatible_htile) {
2676 unsigned max_zplanes = 4;
2677
2678 if (tex->db_render_format == PIPE_FORMAT_Z16_UNORM &&
2679 tex->buffer.b.b.nr_samples > 1)
2680 max_zplanes = 2;
2681
2682 z_info |= S_028038_DECOMPRESS_ON_N_ZPLANES(max_zplanes + 1);
2683
2684 if (sctx->chip_class >= GFX10) {
2685 z_info |= S_028040_ITERATE_FLUSH(1);
2686 s_info |= S_028044_ITERATE_FLUSH(!tex->htile_stencil_disabled);
2687 } else {
2688 z_info |= S_028038_ITERATE_FLUSH(1);
2689 s_info |= S_02803C_ITERATE_FLUSH(1);
2690 }
2691 }
2692
2693 if (tex->surface.has_stencil && !tex->htile_stencil_disabled) {
2694 /* Stencil buffer workaround ported from the GFX6-GFX8 code.
2695 * See that for explanation.
2696 */
2697 s_info |= S_02803C_ALLOW_EXPCLEAR(tex->buffer.b.b.nr_samples <= 1);
2698 } else {
2699 /* Use all HTILE for depth if there's no stencil. */
2700 s_info |= S_02803C_TILE_STENCIL_DISABLE(1);
2701 }
2702
2703 surf->db_htile_data_base = (tex->buffer.gpu_address +
2704 tex->htile_offset) >> 8;
2705 surf->db_htile_surface = S_028ABC_FULL_CACHE(1) |
2706 S_028ABC_PIPE_ALIGNED(tex->surface.u.gfx9.htile.pipe_aligned);
2707 if (sctx->chip_class == GFX9) {
2708 surf->db_htile_surface |=
2709 S_028ABC_RB_ALIGNED(tex->surface.u.gfx9.htile.rb_aligned);
2710 }
2711 }
2712 } else {
2713 /* GFX6-GFX8 */
2714 struct legacy_surf_level *levelinfo = &tex->surface.u.legacy.level[level];
2715
2716 assert(levelinfo->nblk_x % 8 == 0 && levelinfo->nblk_y % 8 == 0);
2717
2718 surf->db_depth_base = (tex->buffer.gpu_address +
2719 tex->surface.u.legacy.level[level].offset) >> 8;
2720 surf->db_stencil_base = (tex->buffer.gpu_address +
2721 tex->surface.u.legacy.stencil_level[level].offset) >> 8;
2722
2723 z_info = S_028040_FORMAT(format) |
2724 S_028040_NUM_SAMPLES(util_logbase2(tex->buffer.b.b.nr_samples));
2725 s_info = S_028044_FORMAT(stencil_format);
2726 surf->db_depth_info = S_02803C_ADDR5_SWIZZLE_MASK(!tex->tc_compatible_htile);
2727
2728 if (sctx->chip_class >= GFX7) {
2729 struct radeon_info *info = &sctx->screen->info;
2730 unsigned index = tex->surface.u.legacy.tiling_index[level];
2731 unsigned stencil_index = tex->surface.u.legacy.stencil_tiling_index[level];
2732 unsigned macro_index = tex->surface.u.legacy.macro_tile_index;
2733 unsigned tile_mode = info->si_tile_mode_array[index];
2734 unsigned stencil_tile_mode = info->si_tile_mode_array[stencil_index];
2735 unsigned macro_mode = info->cik_macrotile_mode_array[macro_index];
2736
2737 surf->db_depth_info |=
2738 S_02803C_ARRAY_MODE(G_009910_ARRAY_MODE(tile_mode)) |
2739 S_02803C_PIPE_CONFIG(G_009910_PIPE_CONFIG(tile_mode)) |
2740 S_02803C_BANK_WIDTH(G_009990_BANK_WIDTH(macro_mode)) |
2741 S_02803C_BANK_HEIGHT(G_009990_BANK_HEIGHT(macro_mode)) |
2742 S_02803C_MACRO_TILE_ASPECT(G_009990_MACRO_TILE_ASPECT(macro_mode)) |
2743 S_02803C_NUM_BANKS(G_009990_NUM_BANKS(macro_mode));
2744 z_info |= S_028040_TILE_SPLIT(G_009910_TILE_SPLIT(tile_mode));
2745 s_info |= S_028044_TILE_SPLIT(G_009910_TILE_SPLIT(stencil_tile_mode));
2746 } else {
2747 unsigned tile_mode_index = si_tile_mode_index(tex, level, false);
2748 z_info |= S_028040_TILE_MODE_INDEX(tile_mode_index);
2749 tile_mode_index = si_tile_mode_index(tex, level, true);
2750 s_info |= S_028044_TILE_MODE_INDEX(tile_mode_index);
2751 }
2752
2753 surf->db_depth_size = S_028058_PITCH_TILE_MAX((levelinfo->nblk_x / 8) - 1) |
2754 S_028058_HEIGHT_TILE_MAX((levelinfo->nblk_y / 8) - 1);
2755 surf->db_depth_slice = S_02805C_SLICE_TILE_MAX((levelinfo->nblk_x *
2756 levelinfo->nblk_y) / 64 - 1);
2757
2758 if (si_htile_enabled(tex, level, PIPE_MASK_ZS)) {
2759 z_info |= S_028040_TILE_SURFACE_ENABLE(1) |
2760 S_028040_ALLOW_EXPCLEAR(1);
2761
2762 if (tex->surface.has_stencil) {
2763 /* Workaround: For a not yet understood reason, the
2764 * combination of MSAA, fast stencil clear and stencil
2765 * decompress messes with subsequent stencil buffer
2766 * uses. Problem was reproduced on Verde, Bonaire,
2767 * Tonga, and Carrizo.
2768 *
2769 * Disabling EXPCLEAR works around the problem.
2770 *
2771 * Check piglit's arb_texture_multisample-stencil-clear
2772 * test if you want to try changing this.
2773 */
2774 if (tex->buffer.b.b.nr_samples <= 1)
2775 s_info |= S_028044_ALLOW_EXPCLEAR(1);
2776 } else if (!tex->tc_compatible_htile) {
2777 /* Use all of the htile_buffer for depth if there's no stencil.
2778 * This must not be set when TC-compatible HTILE is enabled
2779 * due to a hw bug.
2780 */
2781 s_info |= S_028044_TILE_STENCIL_DISABLE(1);
2782 }
2783
2784 surf->db_htile_data_base = (tex->buffer.gpu_address +
2785 tex->htile_offset) >> 8;
2786 surf->db_htile_surface = S_028ABC_FULL_CACHE(1);
2787
2788 if (tex->tc_compatible_htile) {
2789 surf->db_htile_surface |= S_028ABC_TC_COMPATIBLE(1);
2790
2791 /* 0 = full compression. N = only compress up to N-1 Z planes. */
2792 if (tex->buffer.b.b.nr_samples <= 1)
2793 z_info |= S_028040_DECOMPRESS_ON_N_ZPLANES(5);
2794 else if (tex->buffer.b.b.nr_samples <= 4)
2795 z_info |= S_028040_DECOMPRESS_ON_N_ZPLANES(3);
2796 else
2797 z_info |= S_028040_DECOMPRESS_ON_N_ZPLANES(2);
2798 }
2799 }
2800 }
2801
2802 surf->db_z_info = z_info;
2803 surf->db_stencil_info = s_info;
2804
2805 surf->depth_initialized = true;
2806 }
2807
2808 void si_update_fb_dirtiness_after_rendering(struct si_context *sctx)
2809 {
2810 if (sctx->decompression_enabled)
2811 return;
2812
2813 if (sctx->framebuffer.state.zsbuf) {
2814 struct pipe_surface *surf = sctx->framebuffer.state.zsbuf;
2815 struct si_texture *tex = (struct si_texture *)surf->texture;
2816
2817 tex->dirty_level_mask |= 1 << surf->u.tex.level;
2818
2819 if (tex->surface.has_stencil)
2820 tex->stencil_dirty_level_mask |= 1 << surf->u.tex.level;
2821 }
2822
2823 unsigned compressed_cb_mask = sctx->framebuffer.compressed_cb_mask;
2824 while (compressed_cb_mask) {
2825 unsigned i = u_bit_scan(&compressed_cb_mask);
2826 struct pipe_surface *surf = sctx->framebuffer.state.cbufs[i];
2827 struct si_texture *tex = (struct si_texture*)surf->texture;
2828
2829 if (tex->fmask_offset)
2830 tex->dirty_level_mask |= 1 << surf->u.tex.level;
2831 if (tex->dcc_gather_statistics)
2832 tex->separate_dcc_dirty = true;
2833 }
2834 }
2835
2836 static void si_dec_framebuffer_counters(const struct pipe_framebuffer_state *state)
2837 {
2838 for (int i = 0; i < state->nr_cbufs; ++i) {
2839 struct si_surface *surf = NULL;
2840 struct si_texture *tex;
2841
2842 if (!state->cbufs[i])
2843 continue;
2844 surf = (struct si_surface*)state->cbufs[i];
2845 tex = (struct si_texture*)surf->base.texture;
2846
2847 p_atomic_dec(&tex->framebuffers_bound);
2848 }
2849 }
2850
2851 static void si_set_framebuffer_state(struct pipe_context *ctx,
2852 const struct pipe_framebuffer_state *state)
2853 {
2854 struct si_context *sctx = (struct si_context *)ctx;
2855 struct si_surface *surf = NULL;
2856 struct si_texture *tex;
2857 bool old_any_dst_linear = sctx->framebuffer.any_dst_linear;
2858 unsigned old_nr_samples = sctx->framebuffer.nr_samples;
2859 unsigned old_colorbuf_enabled_4bit = sctx->framebuffer.colorbuf_enabled_4bit;
2860 bool old_has_zsbuf = !!sctx->framebuffer.state.zsbuf;
2861 bool old_has_stencil =
2862 old_has_zsbuf &&
2863 ((struct si_texture*)sctx->framebuffer.state.zsbuf->texture)->surface.has_stencil;
2864 bool unbound = false;
2865 int i;
2866
2867 /* Reject zero-sized framebuffers due to a hw bug on GFX6 that occurs
2868 * when PA_SU_HARDWARE_SCREEN_OFFSET != 0 and any_scissor.BR_X/Y <= 0.
2869 * We could implement the full workaround here, but it's a useless case.
2870 */
2871 if ((!state->width || !state->height) && (state->nr_cbufs || state->zsbuf)) {
2872 unreachable("the framebuffer shouldn't have zero area");
2873 return;
2874 }
2875
2876 si_update_fb_dirtiness_after_rendering(sctx);
2877
2878 for (i = 0; i < sctx->framebuffer.state.nr_cbufs; i++) {
2879 if (!sctx->framebuffer.state.cbufs[i])
2880 continue;
2881
2882 tex = (struct si_texture*)sctx->framebuffer.state.cbufs[i]->texture;
2883 if (tex->dcc_gather_statistics)
2884 vi_separate_dcc_stop_query(sctx, tex);
2885 }
2886
2887 /* Disable DCC if the formats are incompatible. */
2888 for (i = 0; i < state->nr_cbufs; i++) {
2889 if (!state->cbufs[i])
2890 continue;
2891
2892 surf = (struct si_surface*)state->cbufs[i];
2893 tex = (struct si_texture*)surf->base.texture;
2894
2895 if (!surf->dcc_incompatible)
2896 continue;
2897
2898 /* Since the DCC decompression calls back into set_framebuffer-
2899 * _state, we need to unbind the framebuffer, so that
2900 * vi_separate_dcc_stop_query isn't called twice with the same
2901 * color buffer.
2902 */
2903 if (!unbound) {
2904 util_copy_framebuffer_state(&sctx->framebuffer.state, NULL);
2905 unbound = true;
2906 }
2907
2908 if (vi_dcc_enabled(tex, surf->base.u.tex.level))
2909 if (!si_texture_disable_dcc(sctx, tex))
2910 si_decompress_dcc(sctx, tex);
2911
2912 surf->dcc_incompatible = false;
2913 }
2914
2915 /* Only flush TC when changing the framebuffer state, because
2916 * the only client not using TC that can change textures is
2917 * the framebuffer.
2918 *
2919 * Wait for compute shaders because of possible transitions:
2920 * - FB write -> shader read
2921 * - shader write -> FB read
2922 *
2923 * DB caches are flushed on demand (using si_decompress_textures).
2924 *
2925 * When MSAA is enabled, CB and TC caches are flushed on demand
2926 * (after FMASK decompression). Shader write -> FB read transitions
2927 * cannot happen for MSAA textures, because MSAA shader images are
2928 * not supported.
2929 *
2930 * Only flush and wait for CB if there is actually a bound color buffer.
2931 */
2932 if (sctx->framebuffer.uncompressed_cb_mask) {
2933 si_make_CB_shader_coherent(sctx, sctx->framebuffer.nr_samples,
2934 sctx->framebuffer.CB_has_shader_readable_metadata,
2935 sctx->framebuffer.all_DCC_pipe_aligned);
2936 }
2937
2938 sctx->flags |= SI_CONTEXT_CS_PARTIAL_FLUSH;
2939
2940 /* u_blitter doesn't invoke depth decompression when it does multiple
2941 * blits in a row, but the only case when it matters for DB is when
2942 * doing generate_mipmap. So here we flush DB manually between
2943 * individual generate_mipmap blits.
2944 * Note that lower mipmap levels aren't compressed.
2945 */
2946 if (sctx->generate_mipmap_for_depth) {
2947 si_make_DB_shader_coherent(sctx, 1, false,
2948 sctx->framebuffer.DB_has_shader_readable_metadata);
2949 } else if (sctx->chip_class == GFX9) {
2950 /* It appears that DB metadata "leaks" in a sequence of:
2951 * - depth clear
2952 * - DCC decompress for shader image writes (with DB disabled)
2953 * - render with DEPTH_BEFORE_SHADER=1
2954 * Flushing DB metadata works around the problem.
2955 */
2956 sctx->flags |= SI_CONTEXT_FLUSH_AND_INV_DB_META;
2957 }
2958
2959 /* Take the maximum of the old and new count. If the new count is lower,
2960 * dirtying is needed to disable the unbound colorbuffers.
2961 */
2962 sctx->framebuffer.dirty_cbufs |=
2963 (1 << MAX2(sctx->framebuffer.state.nr_cbufs, state->nr_cbufs)) - 1;
2964 sctx->framebuffer.dirty_zsbuf |= sctx->framebuffer.state.zsbuf != state->zsbuf;
2965
2966 si_dec_framebuffer_counters(&sctx->framebuffer.state);
2967 util_copy_framebuffer_state(&sctx->framebuffer.state, state);
2968
2969 sctx->framebuffer.colorbuf_enabled_4bit = 0;
2970 sctx->framebuffer.spi_shader_col_format = 0;
2971 sctx->framebuffer.spi_shader_col_format_alpha = 0;
2972 sctx->framebuffer.spi_shader_col_format_blend = 0;
2973 sctx->framebuffer.spi_shader_col_format_blend_alpha = 0;
2974 sctx->framebuffer.color_is_int8 = 0;
2975 sctx->framebuffer.color_is_int10 = 0;
2976
2977 sctx->framebuffer.compressed_cb_mask = 0;
2978 sctx->framebuffer.uncompressed_cb_mask = 0;
2979 sctx->framebuffer.nr_samples = util_framebuffer_get_num_samples(state);
2980 sctx->framebuffer.nr_color_samples = sctx->framebuffer.nr_samples;
2981 sctx->framebuffer.log_samples = util_logbase2(sctx->framebuffer.nr_samples);
2982 sctx->framebuffer.any_dst_linear = false;
2983 sctx->framebuffer.CB_has_shader_readable_metadata = false;
2984 sctx->framebuffer.DB_has_shader_readable_metadata = false;
2985 sctx->framebuffer.all_DCC_pipe_aligned = true;
2986 sctx->framebuffer.min_bytes_per_pixel = 0;
2987
2988 for (i = 0; i < state->nr_cbufs; i++) {
2989 if (!state->cbufs[i])
2990 continue;
2991
2992 surf = (struct si_surface*)state->cbufs[i];
2993 tex = (struct si_texture*)surf->base.texture;
2994
2995 if (!surf->color_initialized) {
2996 si_initialize_color_surface(sctx, surf);
2997 }
2998
2999 sctx->framebuffer.colorbuf_enabled_4bit |= 0xf << (i * 4);
3000 sctx->framebuffer.spi_shader_col_format |=
3001 surf->spi_shader_col_format << (i * 4);
3002 sctx->framebuffer.spi_shader_col_format_alpha |=
3003 surf->spi_shader_col_format_alpha << (i * 4);
3004 sctx->framebuffer.spi_shader_col_format_blend |=
3005 surf->spi_shader_col_format_blend << (i * 4);
3006 sctx->framebuffer.spi_shader_col_format_blend_alpha |=
3007 surf->spi_shader_col_format_blend_alpha << (i * 4);
3008
3009 if (surf->color_is_int8)
3010 sctx->framebuffer.color_is_int8 |= 1 << i;
3011 if (surf->color_is_int10)
3012 sctx->framebuffer.color_is_int10 |= 1 << i;
3013
3014 if (tex->fmask_offset)
3015 sctx->framebuffer.compressed_cb_mask |= 1 << i;
3016 else
3017 sctx->framebuffer.uncompressed_cb_mask |= 1 << i;
3018
3019 /* Don't update nr_color_samples for non-AA buffers.
3020 * (e.g. destination of MSAA resolve)
3021 */
3022 if (tex->buffer.b.b.nr_samples >= 2 &&
3023 tex->buffer.b.b.nr_storage_samples < tex->buffer.b.b.nr_samples) {
3024 sctx->framebuffer.nr_color_samples =
3025 MIN2(sctx->framebuffer.nr_color_samples,
3026 tex->buffer.b.b.nr_storage_samples);
3027 sctx->framebuffer.nr_color_samples =
3028 MAX2(1, sctx->framebuffer.nr_color_samples);
3029 }
3030
3031 if (tex->surface.is_linear)
3032 sctx->framebuffer.any_dst_linear = true;
3033
3034 if (vi_dcc_enabled(tex, surf->base.u.tex.level)) {
3035 sctx->framebuffer.CB_has_shader_readable_metadata = true;
3036
3037 if (sctx->chip_class >= GFX9 &&
3038 !tex->surface.u.gfx9.dcc.pipe_aligned)
3039 sctx->framebuffer.all_DCC_pipe_aligned = false;
3040 }
3041
3042 si_context_add_resource_size(sctx, surf->base.texture);
3043
3044 p_atomic_inc(&tex->framebuffers_bound);
3045
3046 if (tex->dcc_gather_statistics) {
3047 /* Dirty tracking must be enabled for DCC usage analysis. */
3048 sctx->framebuffer.compressed_cb_mask |= 1 << i;
3049 vi_separate_dcc_start_query(sctx, tex);
3050 }
3051
3052 /* Update the minimum but don't keep 0. */
3053 if (!sctx->framebuffer.min_bytes_per_pixel ||
3054 tex->surface.bpe < sctx->framebuffer.min_bytes_per_pixel)
3055 sctx->framebuffer.min_bytes_per_pixel = tex->surface.bpe;
3056 }
3057
3058 /* For optimal DCC performance. */
3059 if (sctx->chip_class >= GFX10)
3060 sctx->framebuffer.dcc_overwrite_combiner_watermark = 6;
3061 else
3062 sctx->framebuffer.dcc_overwrite_combiner_watermark = 4;
3063
3064 struct si_texture *zstex = NULL;
3065
3066 if (state->zsbuf) {
3067 surf = (struct si_surface*)state->zsbuf;
3068 zstex = (struct si_texture*)surf->base.texture;
3069
3070 if (!surf->depth_initialized) {
3071 si_init_depth_surface(sctx, surf);
3072 }
3073
3074 if (vi_tc_compat_htile_enabled(zstex, surf->base.u.tex.level,
3075 PIPE_MASK_ZS))
3076 sctx->framebuffer.DB_has_shader_readable_metadata = true;
3077
3078 si_context_add_resource_size(sctx, surf->base.texture);
3079
3080 /* Update the minimum but don't keep 0. */
3081 if (!sctx->framebuffer.min_bytes_per_pixel ||
3082 zstex->surface.bpe < sctx->framebuffer.min_bytes_per_pixel)
3083 sctx->framebuffer.min_bytes_per_pixel = zstex->surface.bpe;
3084 }
3085
3086 si_update_ps_colorbuf0_slot(sctx);
3087 si_update_poly_offset_state(sctx);
3088 si_mark_atom_dirty(sctx, &sctx->atoms.s.cb_render_state);
3089 si_mark_atom_dirty(sctx, &sctx->atoms.s.framebuffer);
3090
3091 if (sctx->screen->dpbb_allowed)
3092 si_mark_atom_dirty(sctx, &sctx->atoms.s.dpbb_state);
3093
3094 if (sctx->framebuffer.any_dst_linear != old_any_dst_linear)
3095 si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
3096
3097 if (sctx->screen->has_out_of_order_rast &&
3098 (sctx->framebuffer.colorbuf_enabled_4bit != old_colorbuf_enabled_4bit ||
3099 !!sctx->framebuffer.state.zsbuf != old_has_zsbuf ||
3100 (zstex && zstex->surface.has_stencil != old_has_stencil)))
3101 si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
3102
3103 if (sctx->framebuffer.nr_samples != old_nr_samples) {
3104 struct pipe_constant_buffer constbuf = {0};
3105
3106 si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
3107 si_mark_atom_dirty(sctx, &sctx->atoms.s.db_render_state);
3108
3109 constbuf.buffer = sctx->sample_pos_buffer;
3110
3111 /* Set sample locations as fragment shader constants. */
3112 switch (sctx->framebuffer.nr_samples) {
3113 case 1:
3114 constbuf.buffer_offset = 0;
3115 break;
3116 case 2:
3117 constbuf.buffer_offset = (ubyte*)sctx->sample_positions.x2 -
3118 (ubyte*)sctx->sample_positions.x1;
3119 break;
3120 case 4:
3121 constbuf.buffer_offset = (ubyte*)sctx->sample_positions.x4 -
3122 (ubyte*)sctx->sample_positions.x1;
3123 break;
3124 case 8:
3125 constbuf.buffer_offset = (ubyte*)sctx->sample_positions.x8 -
3126 (ubyte*)sctx->sample_positions.x1;
3127 break;
3128 case 16:
3129 constbuf.buffer_offset = (ubyte*)sctx->sample_positions.x16 -
3130 (ubyte*)sctx->sample_positions.x1;
3131 break;
3132 default:
3133 PRINT_ERR("Requested an invalid number of samples %i.\n",
3134 sctx->framebuffer.nr_samples);
3135 assert(0);
3136 }
3137 constbuf.buffer_size = sctx->framebuffer.nr_samples * 2 * 4;
3138 si_set_rw_buffer(sctx, SI_PS_CONST_SAMPLE_POSITIONS, &constbuf);
3139
3140 si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_sample_locs);
3141 }
3142
3143 sctx->do_update_shaders = true;
3144
3145 if (!sctx->decompression_enabled) {
3146 /* Prevent textures decompression when the framebuffer state
3147 * changes come from the decompression passes themselves.
3148 */
3149 sctx->need_check_render_feedback = true;
3150 }
3151 }
3152
3153 static void si_emit_framebuffer_state(struct si_context *sctx)
3154 {
3155 struct radeon_cmdbuf *cs = sctx->gfx_cs;
3156 struct pipe_framebuffer_state *state = &sctx->framebuffer.state;
3157 unsigned i, nr_cbufs = state->nr_cbufs;
3158 struct si_texture *tex = NULL;
3159 struct si_surface *cb = NULL;
3160 unsigned cb_color_info = 0;
3161
3162 /* Colorbuffers. */
3163 for (i = 0; i < nr_cbufs; i++) {
3164 uint64_t cb_color_base, cb_color_fmask, cb_color_cmask, cb_dcc_base;
3165 unsigned cb_color_attrib;
3166
3167 if (!(sctx->framebuffer.dirty_cbufs & (1 << i)))
3168 continue;
3169
3170 cb = (struct si_surface*)state->cbufs[i];
3171 if (!cb) {
3172 radeon_set_context_reg(cs, R_028C70_CB_COLOR0_INFO + i * 0x3C,
3173 S_028C70_FORMAT(V_028C70_COLOR_INVALID));
3174 continue;
3175 }
3176
3177 tex = (struct si_texture *)cb->base.texture;
3178 radeon_add_to_buffer_list(sctx, sctx->gfx_cs,
3179 &tex->buffer, RADEON_USAGE_READWRITE,
3180 tex->buffer.b.b.nr_samples > 1 ?
3181 RADEON_PRIO_COLOR_BUFFER_MSAA :
3182 RADEON_PRIO_COLOR_BUFFER);
3183
3184 if (tex->cmask_buffer && tex->cmask_buffer != &tex->buffer) {
3185 radeon_add_to_buffer_list(sctx, sctx->gfx_cs,
3186 tex->cmask_buffer, RADEON_USAGE_READWRITE,
3187 RADEON_PRIO_SEPARATE_META);
3188 }
3189
3190 if (tex->dcc_separate_buffer)
3191 radeon_add_to_buffer_list(sctx, sctx->gfx_cs,
3192 tex->dcc_separate_buffer,
3193 RADEON_USAGE_READWRITE,
3194 RADEON_PRIO_SEPARATE_META);
3195
3196 /* Compute mutable surface parameters. */
3197 cb_color_base = tex->buffer.gpu_address >> 8;
3198 cb_color_fmask = 0;
3199 cb_color_cmask = tex->cmask_base_address_reg;
3200 cb_dcc_base = 0;
3201 cb_color_info = cb->cb_color_info | tex->cb_color_info;
3202 cb_color_attrib = cb->cb_color_attrib;
3203
3204 if (cb->base.u.tex.level > 0)
3205 cb_color_info &= C_028C70_FAST_CLEAR;
3206
3207 if (tex->fmask_offset) {
3208 cb_color_fmask = (tex->buffer.gpu_address + tex->fmask_offset) >> 8;
3209 cb_color_fmask |= tex->surface.fmask_tile_swizzle;
3210 }
3211
3212 /* Set up DCC. */
3213 if (vi_dcc_enabled(tex, cb->base.u.tex.level)) {
3214 bool is_msaa_resolve_dst = state->cbufs[0] &&
3215 state->cbufs[0]->texture->nr_samples > 1 &&
3216 state->cbufs[1] == &cb->base &&
3217 state->cbufs[1]->texture->nr_samples <= 1;
3218
3219 if (!is_msaa_resolve_dst)
3220 cb_color_info |= S_028C70_DCC_ENABLE(1);
3221
3222 cb_dcc_base = ((!tex->dcc_separate_buffer ? tex->buffer.gpu_address : 0) +
3223 tex->dcc_offset) >> 8;
3224
3225 unsigned dcc_tile_swizzle = tex->surface.tile_swizzle;
3226 dcc_tile_swizzle &= (tex->surface.dcc_alignment - 1) >> 8;
3227 cb_dcc_base |= dcc_tile_swizzle;
3228 }
3229
3230 if (sctx->chip_class >= GFX10) {
3231 unsigned cb_color_attrib3;
3232
3233 /* Set mutable surface parameters. */
3234 cb_color_base += tex->surface.u.gfx9.surf_offset >> 8;
3235 cb_color_base |= tex->surface.tile_swizzle;
3236 if (!tex->fmask_offset)
3237 cb_color_fmask = cb_color_base;
3238 if (cb->base.u.tex.level > 0)
3239 cb_color_cmask = cb_color_base;
3240
3241 cb_color_attrib3 = cb->cb_color_attrib3 |
3242 S_028EE0_COLOR_SW_MODE(tex->surface.u.gfx9.surf.swizzle_mode) |
3243 S_028EE0_FMASK_SW_MODE(tex->surface.u.gfx9.fmask.swizzle_mode) |
3244 S_028EE0_CMASK_PIPE_ALIGNED(tex->surface.u.gfx9.cmask.pipe_aligned) |
3245 S_028EE0_DCC_PIPE_ALIGNED(tex->surface.u.gfx9.dcc.pipe_aligned);
3246
3247 radeon_set_context_reg_seq(cs, R_028C60_CB_COLOR0_BASE + i * 0x3C, 14);
3248 radeon_emit(cs, cb_color_base); /* CB_COLOR0_BASE */
3249 radeon_emit(cs, 0); /* hole */
3250 radeon_emit(cs, 0); /* hole */
3251 radeon_emit(cs, cb->cb_color_view); /* CB_COLOR0_VIEW */
3252 radeon_emit(cs, cb_color_info); /* CB_COLOR0_INFO */
3253 radeon_emit(cs, cb_color_attrib); /* CB_COLOR0_ATTRIB */
3254 radeon_emit(cs, cb->cb_dcc_control); /* CB_COLOR0_DCC_CONTROL */
3255 radeon_emit(cs, cb_color_cmask); /* CB_COLOR0_CMASK */
3256 radeon_emit(cs, 0); /* hole */
3257 radeon_emit(cs, cb_color_fmask); /* CB_COLOR0_FMASK */
3258 radeon_emit(cs, 0); /* hole */
3259 radeon_emit(cs, tex->color_clear_value[0]); /* CB_COLOR0_CLEAR_WORD0 */
3260 radeon_emit(cs, tex->color_clear_value[1]); /* CB_COLOR0_CLEAR_WORD1 */
3261 radeon_emit(cs, cb_dcc_base); /* CB_COLOR0_DCC_BASE */
3262
3263 radeon_set_context_reg(cs, R_028E40_CB_COLOR0_BASE_EXT + i * 4,
3264 cb_color_base >> 32);
3265 radeon_set_context_reg(cs, R_028E60_CB_COLOR0_CMASK_BASE_EXT + i * 4,
3266 cb_color_cmask >> 32);
3267 radeon_set_context_reg(cs, R_028E80_CB_COLOR0_FMASK_BASE_EXT + i * 4,
3268 cb_color_fmask >> 32);
3269 radeon_set_context_reg(cs, R_028EA0_CB_COLOR0_DCC_BASE_EXT + i * 4,
3270 cb_dcc_base >> 32);
3271 radeon_set_context_reg(cs, R_028EC0_CB_COLOR0_ATTRIB2 + i * 4,
3272 cb->cb_color_attrib2);
3273 radeon_set_context_reg(cs, R_028EE0_CB_COLOR0_ATTRIB3 + i * 4,
3274 cb_color_attrib3);
3275 } else if (sctx->chip_class == GFX9) {
3276 struct gfx9_surf_meta_flags meta;
3277
3278 if (tex->dcc_offset)
3279 meta = tex->surface.u.gfx9.dcc;
3280 else
3281 meta = tex->surface.u.gfx9.cmask;
3282
3283 /* Set mutable surface parameters. */
3284 cb_color_base += tex->surface.u.gfx9.surf_offset >> 8;
3285 cb_color_base |= tex->surface.tile_swizzle;
3286 if (!tex->fmask_offset)
3287 cb_color_fmask = cb_color_base;
3288 if (cb->base.u.tex.level > 0)
3289 cb_color_cmask = cb_color_base;
3290 cb_color_attrib |= S_028C74_COLOR_SW_MODE(tex->surface.u.gfx9.surf.swizzle_mode) |
3291 S_028C74_FMASK_SW_MODE(tex->surface.u.gfx9.fmask.swizzle_mode) |
3292 S_028C74_RB_ALIGNED(meta.rb_aligned) |
3293 S_028C74_PIPE_ALIGNED(meta.pipe_aligned);
3294
3295 radeon_set_context_reg_seq(cs, R_028C60_CB_COLOR0_BASE + i * 0x3C, 15);
3296 radeon_emit(cs, cb_color_base); /* CB_COLOR0_BASE */
3297 radeon_emit(cs, S_028C64_BASE_256B(cb_color_base >> 32)); /* CB_COLOR0_BASE_EXT */
3298 radeon_emit(cs, cb->cb_color_attrib2); /* CB_COLOR0_ATTRIB2 */
3299 radeon_emit(cs, cb->cb_color_view); /* CB_COLOR0_VIEW */
3300 radeon_emit(cs, cb_color_info); /* CB_COLOR0_INFO */
3301 radeon_emit(cs, cb_color_attrib); /* CB_COLOR0_ATTRIB */
3302 radeon_emit(cs, cb->cb_dcc_control); /* CB_COLOR0_DCC_CONTROL */
3303 radeon_emit(cs, cb_color_cmask); /* CB_COLOR0_CMASK */
3304 radeon_emit(cs, S_028C80_BASE_256B(cb_color_cmask >> 32)); /* CB_COLOR0_CMASK_BASE_EXT */
3305 radeon_emit(cs, cb_color_fmask); /* CB_COLOR0_FMASK */
3306 radeon_emit(cs, S_028C88_BASE_256B(cb_color_fmask >> 32)); /* CB_COLOR0_FMASK_BASE_EXT */
3307 radeon_emit(cs, tex->color_clear_value[0]); /* CB_COLOR0_CLEAR_WORD0 */
3308 radeon_emit(cs, tex->color_clear_value[1]); /* CB_COLOR0_CLEAR_WORD1 */
3309 radeon_emit(cs, cb_dcc_base); /* CB_COLOR0_DCC_BASE */
3310 radeon_emit(cs, S_028C98_BASE_256B(cb_dcc_base >> 32)); /* CB_COLOR0_DCC_BASE_EXT */
3311
3312 radeon_set_context_reg(cs, R_0287A0_CB_MRT0_EPITCH + i * 4,
3313 S_0287A0_EPITCH(tex->surface.u.gfx9.surf.epitch));
3314 } else {
3315 /* Compute mutable surface parameters (GFX6-GFX8). */
3316 const struct legacy_surf_level *level_info =
3317 &tex->surface.u.legacy.level[cb->base.u.tex.level];
3318 unsigned pitch_tile_max, slice_tile_max, tile_mode_index;
3319 unsigned cb_color_pitch, cb_color_slice, cb_color_fmask_slice;
3320
3321 cb_color_base += level_info->offset >> 8;
3322 /* Only macrotiled modes can set tile swizzle. */
3323 if (level_info->mode == RADEON_SURF_MODE_2D)
3324 cb_color_base |= tex->surface.tile_swizzle;
3325
3326 if (!tex->fmask_offset)
3327 cb_color_fmask = cb_color_base;
3328 if (cb->base.u.tex.level > 0)
3329 cb_color_cmask = cb_color_base;
3330 if (cb_dcc_base)
3331 cb_dcc_base += level_info->dcc_offset >> 8;
3332
3333 pitch_tile_max = level_info->nblk_x / 8 - 1;
3334 slice_tile_max = level_info->nblk_x *
3335 level_info->nblk_y / 64 - 1;
3336 tile_mode_index = si_tile_mode_index(tex, cb->base.u.tex.level, false);
3337
3338 cb_color_attrib |= S_028C74_TILE_MODE_INDEX(tile_mode_index);
3339 cb_color_pitch = S_028C64_TILE_MAX(pitch_tile_max);
3340 cb_color_slice = S_028C68_TILE_MAX(slice_tile_max);
3341
3342 if (tex->fmask_offset) {
3343 if (sctx->chip_class >= GFX7)
3344 cb_color_pitch |= S_028C64_FMASK_TILE_MAX(tex->surface.u.legacy.fmask.pitch_in_pixels / 8 - 1);
3345 cb_color_attrib |= S_028C74_FMASK_TILE_MODE_INDEX(tex->surface.u.legacy.fmask.tiling_index);
3346 cb_color_fmask_slice = S_028C88_TILE_MAX(tex->surface.u.legacy.fmask.slice_tile_max);
3347 } else {
3348 /* This must be set for fast clear to work without FMASK. */
3349 if (sctx->chip_class >= GFX7)
3350 cb_color_pitch |= S_028C64_FMASK_TILE_MAX(pitch_tile_max);
3351 cb_color_attrib |= S_028C74_FMASK_TILE_MODE_INDEX(tile_mode_index);
3352 cb_color_fmask_slice = S_028C88_TILE_MAX(slice_tile_max);
3353 }
3354
3355 radeon_set_context_reg_seq(cs, R_028C60_CB_COLOR0_BASE + i * 0x3C,
3356 sctx->chip_class >= GFX8 ? 14 : 13);
3357 radeon_emit(cs, cb_color_base); /* CB_COLOR0_BASE */
3358 radeon_emit(cs, cb_color_pitch); /* CB_COLOR0_PITCH */
3359 radeon_emit(cs, cb_color_slice); /* CB_COLOR0_SLICE */
3360 radeon_emit(cs, cb->cb_color_view); /* CB_COLOR0_VIEW */
3361 radeon_emit(cs, cb_color_info); /* CB_COLOR0_INFO */
3362 radeon_emit(cs, cb_color_attrib); /* CB_COLOR0_ATTRIB */
3363 radeon_emit(cs, cb->cb_dcc_control); /* CB_COLOR0_DCC_CONTROL */
3364 radeon_emit(cs, cb_color_cmask); /* CB_COLOR0_CMASK */
3365 radeon_emit(cs, tex->surface.u.legacy.cmask_slice_tile_max); /* CB_COLOR0_CMASK_SLICE */
3366 radeon_emit(cs, cb_color_fmask); /* CB_COLOR0_FMASK */
3367 radeon_emit(cs, cb_color_fmask_slice); /* CB_COLOR0_FMASK_SLICE */
3368 radeon_emit(cs, tex->color_clear_value[0]); /* CB_COLOR0_CLEAR_WORD0 */
3369 radeon_emit(cs, tex->color_clear_value[1]); /* CB_COLOR0_CLEAR_WORD1 */
3370
3371 if (sctx->chip_class >= GFX8) /* R_028C94_CB_COLOR0_DCC_BASE */
3372 radeon_emit(cs, cb_dcc_base);
3373 }
3374 }
3375 for (; i < 8 ; i++)
3376 if (sctx->framebuffer.dirty_cbufs & (1 << i))
3377 radeon_set_context_reg(cs, R_028C70_CB_COLOR0_INFO + i * 0x3C, 0);
3378
3379 /* ZS buffer. */
3380 if (state->zsbuf && sctx->framebuffer.dirty_zsbuf) {
3381 struct si_surface *zb = (struct si_surface*)state->zsbuf;
3382 struct si_texture *tex = (struct si_texture*)zb->base.texture;
3383
3384 radeon_add_to_buffer_list(sctx, sctx->gfx_cs,
3385 &tex->buffer, RADEON_USAGE_READWRITE,
3386 zb->base.texture->nr_samples > 1 ?
3387 RADEON_PRIO_DEPTH_BUFFER_MSAA :
3388 RADEON_PRIO_DEPTH_BUFFER);
3389
3390 if (sctx->chip_class >= GFX10) {
3391 radeon_set_context_reg(cs, R_028014_DB_HTILE_DATA_BASE, zb->db_htile_data_base);
3392 radeon_set_context_reg(cs, R_02801C_DB_DEPTH_SIZE_XY, zb->db_depth_size);
3393
3394 radeon_set_context_reg_seq(cs, R_02803C_DB_DEPTH_INFO, 7);
3395 radeon_emit(cs, S_02803C_RESOURCE_LEVEL(1)); /* DB_DEPTH_INFO */
3396 radeon_emit(cs, zb->db_z_info | /* DB_Z_INFO */
3397 S_028038_ZRANGE_PRECISION(tex->depth_clear_value != 0));
3398 radeon_emit(cs, zb->db_stencil_info); /* DB_STENCIL_INFO */
3399 radeon_emit(cs, zb->db_depth_base); /* DB_Z_READ_BASE */
3400 radeon_emit(cs, zb->db_stencil_base); /* DB_STENCIL_READ_BASE */
3401 radeon_emit(cs, zb->db_depth_base); /* DB_Z_WRITE_BASE */
3402 radeon_emit(cs, zb->db_stencil_base); /* DB_STENCIL_WRITE_BASE */
3403
3404 radeon_set_context_reg_seq(cs, R_028068_DB_Z_READ_BASE_HI, 5);
3405 radeon_emit(cs, zb->db_depth_base >> 32); /* DB_Z_READ_BASE_HI */
3406 radeon_emit(cs, zb->db_stencil_base >> 32); /* DB_STENCIL_READ_BASE_HI */
3407 radeon_emit(cs, zb->db_depth_base >> 32); /* DB_Z_WRITE_BASE_HI */
3408 radeon_emit(cs, zb->db_stencil_base >> 32); /* DB_STENCIL_WRITE_BASE_HI */
3409 radeon_emit(cs, zb->db_htile_data_base >> 32); /* DB_HTILE_DATA_BASE_HI */
3410 } else if (sctx->chip_class == GFX9) {
3411 radeon_set_context_reg_seq(cs, R_028014_DB_HTILE_DATA_BASE, 3);
3412 radeon_emit(cs, zb->db_htile_data_base); /* DB_HTILE_DATA_BASE */
3413 radeon_emit(cs, S_028018_BASE_HI(zb->db_htile_data_base >> 32)); /* DB_HTILE_DATA_BASE_HI */
3414 radeon_emit(cs, zb->db_depth_size); /* DB_DEPTH_SIZE */
3415
3416 radeon_set_context_reg_seq(cs, R_028038_DB_Z_INFO, 10);
3417 radeon_emit(cs, zb->db_z_info | /* DB_Z_INFO */
3418 S_028038_ZRANGE_PRECISION(tex->depth_clear_value != 0));
3419 radeon_emit(cs, zb->db_stencil_info); /* DB_STENCIL_INFO */
3420 radeon_emit(cs, zb->db_depth_base); /* DB_Z_READ_BASE */
3421 radeon_emit(cs, S_028044_BASE_HI(zb->db_depth_base >> 32)); /* DB_Z_READ_BASE_HI */
3422 radeon_emit(cs, zb->db_stencil_base); /* DB_STENCIL_READ_BASE */
3423 radeon_emit(cs, S_02804C_BASE_HI(zb->db_stencil_base >> 32)); /* DB_STENCIL_READ_BASE_HI */
3424 radeon_emit(cs, zb->db_depth_base); /* DB_Z_WRITE_BASE */
3425 radeon_emit(cs, S_028054_BASE_HI(zb->db_depth_base >> 32)); /* DB_Z_WRITE_BASE_HI */
3426 radeon_emit(cs, zb->db_stencil_base); /* DB_STENCIL_WRITE_BASE */
3427 radeon_emit(cs, S_02805C_BASE_HI(zb->db_stencil_base >> 32)); /* DB_STENCIL_WRITE_BASE_HI */
3428
3429 radeon_set_context_reg_seq(cs, R_028068_DB_Z_INFO2, 2);
3430 radeon_emit(cs, zb->db_z_info2); /* DB_Z_INFO2 */
3431 radeon_emit(cs, zb->db_stencil_info2); /* DB_STENCIL_INFO2 */
3432 } else {
3433 radeon_set_context_reg(cs, R_028014_DB_HTILE_DATA_BASE, zb->db_htile_data_base);
3434
3435 radeon_set_context_reg_seq(cs, R_02803C_DB_DEPTH_INFO, 9);
3436 radeon_emit(cs, zb->db_depth_info); /* DB_DEPTH_INFO */
3437 radeon_emit(cs, zb->db_z_info | /* DB_Z_INFO */
3438 S_028040_ZRANGE_PRECISION(tex->depth_clear_value != 0));
3439 radeon_emit(cs, zb->db_stencil_info); /* DB_STENCIL_INFO */
3440 radeon_emit(cs, zb->db_depth_base); /* DB_Z_READ_BASE */
3441 radeon_emit(cs, zb->db_stencil_base); /* DB_STENCIL_READ_BASE */
3442 radeon_emit(cs, zb->db_depth_base); /* DB_Z_WRITE_BASE */
3443 radeon_emit(cs, zb->db_stencil_base); /* DB_STENCIL_WRITE_BASE */
3444 radeon_emit(cs, zb->db_depth_size); /* DB_DEPTH_SIZE */
3445 radeon_emit(cs, zb->db_depth_slice); /* DB_DEPTH_SLICE */
3446 }
3447
3448 radeon_set_context_reg_seq(cs, R_028028_DB_STENCIL_CLEAR, 2);
3449 radeon_emit(cs, tex->stencil_clear_value); /* R_028028_DB_STENCIL_CLEAR */
3450 radeon_emit(cs, fui(tex->depth_clear_value)); /* R_02802C_DB_DEPTH_CLEAR */
3451
3452 radeon_set_context_reg(cs, R_028008_DB_DEPTH_VIEW, zb->db_depth_view);
3453 radeon_set_context_reg(cs, R_028ABC_DB_HTILE_SURFACE, zb->db_htile_surface);
3454 } else if (sctx->framebuffer.dirty_zsbuf) {
3455 if (sctx->chip_class == GFX9)
3456 radeon_set_context_reg_seq(cs, R_028038_DB_Z_INFO, 2);
3457 else
3458 radeon_set_context_reg_seq(cs, R_028040_DB_Z_INFO, 2);
3459
3460 radeon_emit(cs, S_028040_FORMAT(V_028040_Z_INVALID)); /* DB_Z_INFO */
3461 radeon_emit(cs, S_028044_FORMAT(V_028044_STENCIL_INVALID)); /* DB_STENCIL_INFO */
3462 }
3463
3464 /* Framebuffer dimensions. */
3465 /* PA_SC_WINDOW_SCISSOR_TL is set in si_init_config() */
3466 radeon_set_context_reg(cs, R_028208_PA_SC_WINDOW_SCISSOR_BR,
3467 S_028208_BR_X(state->width) | S_028208_BR_Y(state->height));
3468
3469 if (sctx->screen->dfsm_allowed) {
3470 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
3471 radeon_emit(cs, EVENT_TYPE(V_028A90_BREAK_BATCH) | EVENT_INDEX(0));
3472 }
3473
3474 sctx->framebuffer.dirty_cbufs = 0;
3475 sctx->framebuffer.dirty_zsbuf = false;
3476 }
3477
3478 static void si_emit_msaa_sample_locs(struct si_context *sctx)
3479 {
3480 struct radeon_cmdbuf *cs = sctx->gfx_cs;
3481 struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
3482 unsigned nr_samples = sctx->framebuffer.nr_samples;
3483 bool has_msaa_sample_loc_bug = sctx->screen->has_msaa_sample_loc_bug;
3484
3485 /* Smoothing (only possible with nr_samples == 1) uses the same
3486 * sample locations as the MSAA it simulates.
3487 */
3488 if (nr_samples <= 1 && sctx->smoothing_enabled)
3489 nr_samples = SI_NUM_SMOOTH_AA_SAMPLES;
3490
3491 /* On Polaris, the small primitive filter uses the sample locations
3492 * even when MSAA is off, so we need to make sure they're set to 0.
3493 *
3494 * GFX10 uses sample locations unconditionally, so they always need
3495 * to be set up.
3496 */
3497 if ((nr_samples >= 2 || has_msaa_sample_loc_bug ||
3498 sctx->chip_class >= GFX10) &&
3499 nr_samples != sctx->sample_locs_num_samples) {
3500 sctx->sample_locs_num_samples = nr_samples;
3501 si_emit_sample_locations(cs, nr_samples);
3502 }
3503
3504 if (sctx->family >= CHIP_POLARIS10) {
3505 unsigned small_prim_filter_cntl =
3506 S_028830_SMALL_PRIM_FILTER_ENABLE(1) |
3507 /* line bug */
3508 S_028830_LINE_FILTER_DISABLE(sctx->family <= CHIP_POLARIS12);
3509
3510 /* The alternative of setting sample locations to 0 would
3511 * require a DB flush to avoid Z errors, see
3512 * https://bugs.freedesktop.org/show_bug.cgi?id=96908
3513 */
3514 if (has_msaa_sample_loc_bug &&
3515 sctx->framebuffer.nr_samples > 1 &&
3516 !rs->multisample_enable)
3517 small_prim_filter_cntl &= C_028830_SMALL_PRIM_FILTER_ENABLE;
3518
3519 radeon_opt_set_context_reg(sctx,
3520 R_028830_PA_SU_SMALL_PRIM_FILTER_CNTL,
3521 SI_TRACKED_PA_SU_SMALL_PRIM_FILTER_CNTL,
3522 small_prim_filter_cntl);
3523 }
3524
3525 /* The exclusion bits can be set to improve rasterization efficiency
3526 * if no sample lies on the pixel boundary (-8 sample offset).
3527 */
3528 bool exclusion = sctx->chip_class >= GFX7 &&
3529 (!rs->multisample_enable || nr_samples != 16);
3530 radeon_opt_set_context_reg(sctx, R_02882C_PA_SU_PRIM_FILTER_CNTL,
3531 SI_TRACKED_PA_SU_PRIM_FILTER_CNTL,
3532 S_02882C_XMAX_RIGHT_EXCLUSION(exclusion) |
3533 S_02882C_YMAX_BOTTOM_EXCLUSION(exclusion));
3534 }
3535
3536 static bool si_out_of_order_rasterization(struct si_context *sctx)
3537 {
3538 struct si_state_blend *blend = sctx->queued.named.blend;
3539 struct si_state_dsa *dsa = sctx->queued.named.dsa;
3540
3541 if (!sctx->screen->has_out_of_order_rast)
3542 return false;
3543
3544 unsigned colormask = sctx->framebuffer.colorbuf_enabled_4bit;
3545
3546 if (blend) {
3547 colormask &= blend->cb_target_enabled_4bit;
3548 } else {
3549 colormask = 0;
3550 }
3551
3552 /* Conservative: No logic op. */
3553 if (colormask && blend->logicop_enable)
3554 return false;
3555
3556 struct si_dsa_order_invariance dsa_order_invariant = {
3557 .zs = true, .pass_set = true, .pass_last = false
3558 };
3559
3560 if (sctx->framebuffer.state.zsbuf) {
3561 struct si_texture *zstex =
3562 (struct si_texture*)sctx->framebuffer.state.zsbuf->texture;
3563 bool has_stencil = zstex->surface.has_stencil;
3564 dsa_order_invariant = dsa->order_invariance[has_stencil];
3565 if (!dsa_order_invariant.zs)
3566 return false;
3567
3568 /* The set of PS invocations is always order invariant,
3569 * except when early Z/S tests are requested. */
3570 if (sctx->ps_shader.cso &&
3571 sctx->ps_shader.cso->info.writes_memory &&
3572 sctx->ps_shader.cso->info.properties[TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL] &&
3573 !dsa_order_invariant.pass_set)
3574 return false;
3575
3576 if (sctx->num_perfect_occlusion_queries != 0 &&
3577 !dsa_order_invariant.pass_set)
3578 return false;
3579 }
3580
3581 if (!colormask)
3582 return true;
3583
3584 unsigned blendmask = colormask & blend->blend_enable_4bit;
3585
3586 if (blendmask) {
3587 /* Only commutative blending. */
3588 if (blendmask & ~blend->commutative_4bit)
3589 return false;
3590
3591 if (!dsa_order_invariant.pass_set)
3592 return false;
3593 }
3594
3595 if (colormask & ~blendmask) {
3596 if (!dsa_order_invariant.pass_last)
3597 return false;
3598 }
3599
3600 return true;
3601 }
3602
3603 static void si_emit_msaa_config(struct si_context *sctx)
3604 {
3605 struct radeon_cmdbuf *cs = sctx->gfx_cs;
3606 unsigned num_tile_pipes = sctx->screen->info.num_tile_pipes;
3607 /* 33% faster rendering to linear color buffers */
3608 bool dst_is_linear = sctx->framebuffer.any_dst_linear;
3609 bool out_of_order_rast = si_out_of_order_rasterization(sctx);
3610 unsigned sc_mode_cntl_1 =
3611 S_028A4C_WALK_SIZE(dst_is_linear) |
3612 S_028A4C_WALK_FENCE_ENABLE(!dst_is_linear) |
3613 S_028A4C_WALK_FENCE_SIZE(num_tile_pipes == 2 ? 2 : 3) |
3614 S_028A4C_OUT_OF_ORDER_PRIMITIVE_ENABLE(out_of_order_rast) |
3615 S_028A4C_OUT_OF_ORDER_WATER_MARK(0x7) |
3616 /* always 1: */
3617 S_028A4C_WALK_ALIGN8_PRIM_FITS_ST(1) |
3618 S_028A4C_SUPERTILE_WALK_ORDER_ENABLE(1) |
3619 S_028A4C_TILE_WALK_ORDER_ENABLE(1) |
3620 S_028A4C_MULTI_SHADER_ENGINE_PRIM_DISCARD_ENABLE(1) |
3621 S_028A4C_FORCE_EOV_CNTDWN_ENABLE(1) |
3622 S_028A4C_FORCE_EOV_REZ_ENABLE(1);
3623 unsigned db_eqaa = S_028804_HIGH_QUALITY_INTERSECTIONS(1) |
3624 S_028804_INCOHERENT_EQAA_READS(1) |
3625 S_028804_INTERPOLATE_COMP_Z(1) |
3626 S_028804_STATIC_ANCHOR_ASSOCIATIONS(1);
3627 unsigned coverage_samples, color_samples, z_samples;
3628 struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
3629
3630 /* S: Coverage samples (up to 16x):
3631 * - Scan conversion samples (PA_SC_AA_CONFIG.MSAA_NUM_SAMPLES)
3632 * - CB FMASK samples (CB_COLORi_ATTRIB.NUM_SAMPLES)
3633 *
3634 * Z: Z/S samples (up to 8x, must be <= coverage samples and >= color samples):
3635 * - Value seen by DB (DB_Z_INFO.NUM_SAMPLES)
3636 * - Value seen by CB, must be correct even if Z/S is unbound (DB_EQAA.MAX_ANCHOR_SAMPLES)
3637 * # Missing samples are derived from Z planes if Z is compressed (up to 16x quality), or
3638 * # from the closest defined sample if Z is uncompressed (same quality as the number of
3639 * # Z samples).
3640 *
3641 * F: Color samples (up to 8x, must be <= coverage samples):
3642 * - CB color samples (CB_COLORi_ATTRIB.NUM_FRAGMENTS)
3643 * - PS iter samples (DB_EQAA.PS_ITER_SAMPLES)
3644 *
3645 * Can be anything between coverage and color samples:
3646 * - SampleMaskIn samples (PA_SC_AA_CONFIG.MSAA_EXPOSED_SAMPLES)
3647 * - SampleMaskOut samples (DB_EQAA.MASK_EXPORT_NUM_SAMPLES)
3648 * - Alpha-to-coverage samples (DB_EQAA.ALPHA_TO_MASK_NUM_SAMPLES)
3649 * - Occlusion query samples (DB_COUNT_CONTROL.SAMPLE_RATE)
3650 * # All are currently set the same as coverage samples.
3651 *
3652 * If color samples < coverage samples, FMASK has a higher bpp to store an "unknown"
3653 * flag for undefined color samples. A shader-based resolve must handle unknowns
3654 * or mask them out with AND. Unknowns can also be guessed from neighbors via
3655 * an edge-detect shader-based resolve, which is required to make "color samples = 1"
3656 * useful. The CB resolve always drops unknowns.
3657 *
3658 * Sensible AA configurations:
3659 * EQAA 16s 8z 8f - might look the same as 16x MSAA if Z is compressed
3660 * EQAA 16s 8z 4f - might look the same as 16x MSAA if Z is compressed
3661 * EQAA 16s 4z 4f - might look the same as 16x MSAA if Z is compressed
3662 * EQAA 8s 8z 8f = 8x MSAA
3663 * EQAA 8s 8z 4f - might look the same as 8x MSAA
3664 * EQAA 8s 8z 2f - might look the same as 8x MSAA with low-density geometry
3665 * EQAA 8s 4z 4f - might look the same as 8x MSAA if Z is compressed
3666 * EQAA 8s 4z 2f - might look the same as 8x MSAA with low-density geometry if Z is compressed
3667 * EQAA 4s 4z 4f = 4x MSAA
3668 * EQAA 4s 4z 2f - might look the same as 4x MSAA with low-density geometry
3669 * EQAA 2s 2z 2f = 2x MSAA
3670 */
3671 if (sctx->framebuffer.nr_samples > 1 && rs->multisample_enable) {
3672 coverage_samples = sctx->framebuffer.nr_samples;
3673 color_samples = sctx->framebuffer.nr_color_samples;
3674
3675 if (sctx->framebuffer.state.zsbuf) {
3676 z_samples = sctx->framebuffer.state.zsbuf->texture->nr_samples;
3677 z_samples = MAX2(1, z_samples);
3678 } else {
3679 z_samples = coverage_samples;
3680 }
3681 } else if (sctx->smoothing_enabled) {
3682 coverage_samples = color_samples = z_samples = SI_NUM_SMOOTH_AA_SAMPLES;
3683 } else {
3684 coverage_samples = color_samples = z_samples = 1;
3685 }
3686
3687 /* Required by OpenGL line rasterization.
3688 *
3689 * TODO: We should also enable perpendicular endcaps for AA lines,
3690 * but that requires implementing line stippling in the pixel
3691 * shader. SC can only do line stippling with axis-aligned
3692 * endcaps.
3693 */
3694 unsigned sc_line_cntl = S_028BDC_DX10_DIAMOND_TEST_ENA(1);
3695 unsigned sc_aa_config = 0;
3696
3697 if (coverage_samples > 1) {
3698 /* distance from the pixel center, indexed by log2(nr_samples) */
3699 static unsigned max_dist[] = {
3700 0, /* unused */
3701 4, /* 2x MSAA */
3702 6, /* 4x MSAA */
3703 7, /* 8x MSAA */
3704 8, /* 16x MSAA */
3705 };
3706 unsigned log_samples = util_logbase2(coverage_samples);
3707 unsigned log_z_samples = util_logbase2(z_samples);
3708 unsigned ps_iter_samples = si_get_ps_iter_samples(sctx);
3709 unsigned log_ps_iter_samples = util_logbase2(ps_iter_samples);
3710
3711 sc_line_cntl |= S_028BDC_EXPAND_LINE_WIDTH(1);
3712 sc_aa_config = S_028BE0_MSAA_NUM_SAMPLES(log_samples) |
3713 S_028BE0_MAX_SAMPLE_DIST(max_dist[log_samples]) |
3714 S_028BE0_MSAA_EXPOSED_SAMPLES(log_samples);
3715
3716 if (sctx->framebuffer.nr_samples > 1) {
3717 db_eqaa |= S_028804_MAX_ANCHOR_SAMPLES(log_z_samples) |
3718 S_028804_PS_ITER_SAMPLES(log_ps_iter_samples) |
3719 S_028804_MASK_EXPORT_NUM_SAMPLES(log_samples) |
3720 S_028804_ALPHA_TO_MASK_NUM_SAMPLES(log_samples);
3721 sc_mode_cntl_1 |= S_028A4C_PS_ITER_SAMPLE(ps_iter_samples > 1);
3722 } else if (sctx->smoothing_enabled) {
3723 db_eqaa |= S_028804_OVERRASTERIZATION_AMOUNT(log_samples);
3724 }
3725 }
3726
3727 unsigned initial_cdw = cs->current.cdw;
3728
3729 /* R_028BDC_PA_SC_LINE_CNTL, R_028BE0_PA_SC_AA_CONFIG */
3730 radeon_opt_set_context_reg2(sctx, R_028BDC_PA_SC_LINE_CNTL,
3731 SI_TRACKED_PA_SC_LINE_CNTL, sc_line_cntl,
3732 sc_aa_config);
3733 /* R_028804_DB_EQAA */
3734 radeon_opt_set_context_reg(sctx, R_028804_DB_EQAA, SI_TRACKED_DB_EQAA,
3735 db_eqaa);
3736 /* R_028A4C_PA_SC_MODE_CNTL_1 */
3737 radeon_opt_set_context_reg(sctx, R_028A4C_PA_SC_MODE_CNTL_1,
3738 SI_TRACKED_PA_SC_MODE_CNTL_1, sc_mode_cntl_1);
3739
3740 if (initial_cdw != cs->current.cdw) {
3741 sctx->context_roll = true;
3742
3743 /* GFX9: Flush DFSM when the AA mode changes. */
3744 if (sctx->screen->dfsm_allowed) {
3745 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
3746 radeon_emit(cs, EVENT_TYPE(V_028A90_FLUSH_DFSM) | EVENT_INDEX(0));
3747 }
3748 }
3749 }
3750
3751 void si_update_ps_iter_samples(struct si_context *sctx)
3752 {
3753 if (sctx->framebuffer.nr_samples > 1)
3754 si_mark_atom_dirty(sctx, &sctx->atoms.s.msaa_config);
3755 if (sctx->screen->dpbb_allowed)
3756 si_mark_atom_dirty(sctx, &sctx->atoms.s.dpbb_state);
3757 }
3758
3759 static void si_set_min_samples(struct pipe_context *ctx, unsigned min_samples)
3760 {
3761 struct si_context *sctx = (struct si_context *)ctx;
3762
3763 /* The hardware can only do sample shading with 2^n samples. */
3764 min_samples = util_next_power_of_two(min_samples);
3765
3766 if (sctx->ps_iter_samples == min_samples)
3767 return;
3768
3769 sctx->ps_iter_samples = min_samples;
3770 sctx->do_update_shaders = true;
3771
3772 si_update_ps_iter_samples(sctx);
3773 }
3774
3775 /*
3776 * Samplers
3777 */
3778
3779 /**
3780 * Build the sampler view descriptor for a buffer texture.
3781 * @param state 256-bit descriptor; only the high 128 bits are filled in
3782 */
3783 void
3784 si_make_buffer_descriptor(struct si_screen *screen, struct si_resource *buf,
3785 enum pipe_format format,
3786 unsigned offset, unsigned size,
3787 uint32_t *state)
3788 {
3789 const struct util_format_description *desc;
3790 unsigned stride;
3791 unsigned num_records;
3792
3793 desc = util_format_description(format);
3794 stride = desc->block.bits / 8;
3795
3796 num_records = size / stride;
3797 num_records = MIN2(num_records, (buf->b.b.width0 - offset) / stride);
3798
3799 /* The NUM_RECORDS field has a different meaning depending on the chip,
3800 * instruction type, STRIDE, and SWIZZLE_ENABLE.
3801 *
3802 * GFX6-7,10:
3803 * - If STRIDE == 0, it's in byte units.
3804 * - If STRIDE != 0, it's in units of STRIDE, used with inst.IDXEN.
3805 *
3806 * GFX8:
3807 * - For SMEM and STRIDE == 0, it's in byte units.
3808 * - For SMEM and STRIDE != 0, it's in units of STRIDE.
3809 * - For VMEM and STRIDE == 0 or SWIZZLE_ENABLE == 0, it's in byte units.
3810 * - For VMEM and STRIDE != 0 and SWIZZLE_ENABLE == 1, it's in units of STRIDE.
3811 * NOTE: There is incompatibility between VMEM and SMEM opcodes due to SWIZZLE_-
3812 * ENABLE. The workaround is to set STRIDE = 0 if SWIZZLE_ENABLE == 0 when
3813 * using SMEM. This can be done in the shader by clearing STRIDE with s_and.
3814 * That way the same descriptor can be used by both SMEM and VMEM.
3815 *
3816 * GFX9:
3817 * - For SMEM and STRIDE == 0, it's in byte units.
3818 * - For SMEM and STRIDE != 0, it's in units of STRIDE.
3819 * - For VMEM and inst.IDXEN == 0 or STRIDE == 0, it's in byte units.
3820 * - For VMEM and inst.IDXEN == 1 and STRIDE != 0, it's in units of STRIDE.
3821 */
3822 if (screen->info.chip_class == GFX9 && HAVE_LLVM < 0x0800)
3823 /* When vindex == 0, LLVM < 8.0 sets IDXEN = 0, thus changing units
3824 * from STRIDE to bytes. This works around it by setting
3825 * NUM_RECORDS to at least the size of one element, so that
3826 * the first element is readable when IDXEN == 0.
3827 */
3828 num_records = num_records ? MAX2(num_records, stride) : 0;
3829 else if (screen->info.chip_class == GFX8)
3830 num_records *= stride;
3831
3832 state[4] = 0;
3833 state[5] = S_008F04_STRIDE(stride);
3834 state[6] = num_records;
3835 state[7] = S_008F0C_DST_SEL_X(si_map_swizzle(desc->swizzle[0])) |
3836 S_008F0C_DST_SEL_Y(si_map_swizzle(desc->swizzle[1])) |
3837 S_008F0C_DST_SEL_Z(si_map_swizzle(desc->swizzle[2])) |
3838 S_008F0C_DST_SEL_W(si_map_swizzle(desc->swizzle[3]));
3839
3840 if (screen->info.chip_class >= GFX10) {
3841 const struct gfx10_format *fmt = &gfx10_format_table[format];
3842
3843 /* OOB_SELECT chooses the out-of-bounds check:
3844 * - 0: (index >= NUM_RECORDS) || (offset >= STRIDE)
3845 * - 1: index >= NUM_RECORDS
3846 * - 2: NUM_RECORDS == 0
3847 * - 3: if SWIZZLE_ENABLE == 0: offset >= NUM_RECORDS
3848 * else: swizzle_address >= NUM_RECORDS
3849 */
3850 state[7] |= S_008F0C_FORMAT(fmt->img_format) |
3851 S_008F0C_OOB_SELECT(0) |
3852 S_008F0C_RESOURCE_LEVEL(1);
3853 } else {
3854 int first_non_void;
3855 unsigned num_format, data_format;
3856
3857 first_non_void = util_format_get_first_non_void_channel(format);
3858 num_format = si_translate_buffer_numformat(&screen->b, desc, first_non_void);
3859 data_format = si_translate_buffer_dataformat(&screen->b, desc, first_non_void);
3860
3861 state[7] |= S_008F0C_NUM_FORMAT(num_format) |
3862 S_008F0C_DATA_FORMAT(data_format);
3863 }
3864 }
3865
3866 static unsigned gfx9_border_color_swizzle(const unsigned char swizzle[4])
3867 {
3868 unsigned bc_swizzle = V_008F20_BC_SWIZZLE_XYZW;
3869
3870 if (swizzle[3] == PIPE_SWIZZLE_X) {
3871 /* For the pre-defined border color values (white, opaque
3872 * black, transparent black), the only thing that matters is
3873 * that the alpha channel winds up in the correct place
3874 * (because the RGB channels are all the same) so either of
3875 * these enumerations will work.
3876 */
3877 if (swizzle[2] == PIPE_SWIZZLE_Y)
3878 bc_swizzle = V_008F20_BC_SWIZZLE_WZYX;
3879 else
3880 bc_swizzle = V_008F20_BC_SWIZZLE_WXYZ;
3881 } else if (swizzle[0] == PIPE_SWIZZLE_X) {
3882 if (swizzle[1] == PIPE_SWIZZLE_Y)
3883 bc_swizzle = V_008F20_BC_SWIZZLE_XYZW;
3884 else
3885 bc_swizzle = V_008F20_BC_SWIZZLE_XWYZ;
3886 } else if (swizzle[1] == PIPE_SWIZZLE_X) {
3887 bc_swizzle = V_008F20_BC_SWIZZLE_YXWZ;
3888 } else if (swizzle[2] == PIPE_SWIZZLE_X) {
3889 bc_swizzle = V_008F20_BC_SWIZZLE_ZYXW;
3890 }
3891
3892 return bc_swizzle;
3893 }
3894
3895 /**
3896 * Build the sampler view descriptor for a texture.
3897 */
3898 static void
3899 gfx10_make_texture_descriptor(struct si_screen *screen,
3900 struct si_texture *tex,
3901 bool sampler,
3902 enum pipe_texture_target target,
3903 enum pipe_format pipe_format,
3904 const unsigned char state_swizzle[4],
3905 unsigned first_level, unsigned last_level,
3906 unsigned first_layer, unsigned last_layer,
3907 unsigned width, unsigned height, unsigned depth,
3908 uint32_t *state,
3909 uint32_t *fmask_state)
3910 {
3911 struct pipe_resource *res = &tex->buffer.b.b;
3912 const struct util_format_description *desc;
3913 unsigned img_format;
3914 unsigned char swizzle[4];
3915 unsigned type;
3916 uint64_t va;
3917
3918 desc = util_format_description(pipe_format);
3919 img_format = gfx10_format_table[pipe_format].img_format;
3920
3921 if (desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS) {
3922 const unsigned char swizzle_xxxx[4] = {0, 0, 0, 0};
3923 const unsigned char swizzle_yyyy[4] = {1, 1, 1, 1};
3924 const unsigned char swizzle_wwww[4] = {3, 3, 3, 3};
3925 bool is_stencil = false;
3926
3927 switch (pipe_format) {
3928 case PIPE_FORMAT_S8_UINT_Z24_UNORM:
3929 case PIPE_FORMAT_X32_S8X24_UINT:
3930 case PIPE_FORMAT_X8Z24_UNORM:
3931 util_format_compose_swizzles(swizzle_yyyy, state_swizzle, swizzle);
3932 is_stencil = true;
3933 break;
3934 case PIPE_FORMAT_X24S8_UINT:
3935 /*
3936 * X24S8 is implemented as an 8_8_8_8 data format, to
3937 * fix texture gathers. This affects at least
3938 * GL45-CTS.texture_cube_map_array.sampling on GFX8.
3939 */
3940 util_format_compose_swizzles(swizzle_wwww, state_swizzle, swizzle);
3941 is_stencil = true;
3942 break;
3943 default:
3944 util_format_compose_swizzles(swizzle_xxxx, state_swizzle, swizzle);
3945 is_stencil = pipe_format == PIPE_FORMAT_S8_UINT;
3946 }
3947
3948 if (tex->upgraded_depth && !is_stencil) {
3949 assert(img_format == V_008F0C_IMG_FORMAT_32_FLOAT);
3950 img_format = V_008F0C_IMG_FORMAT_32_FLOAT_CLAMP;
3951 }
3952 } else {
3953 util_format_compose_swizzles(desc->swizzle, state_swizzle, swizzle);
3954 }
3955
3956 if (!sampler &&
3957 (res->target == PIPE_TEXTURE_CUBE ||
3958 res->target == PIPE_TEXTURE_CUBE_ARRAY)) {
3959 /* For the purpose of shader images, treat cube maps as 2D
3960 * arrays.
3961 */
3962 type = V_008F1C_SQ_RSRC_IMG_2D_ARRAY;
3963 } else {
3964 type = si_tex_dim(screen, tex, target, res->nr_samples);
3965 }
3966
3967 if (type == V_008F1C_SQ_RSRC_IMG_1D_ARRAY) {
3968 height = 1;
3969 depth = res->array_size;
3970 } else if (type == V_008F1C_SQ_RSRC_IMG_2D_ARRAY ||
3971 type == V_008F1C_SQ_RSRC_IMG_2D_MSAA_ARRAY) {
3972 if (sampler || res->target != PIPE_TEXTURE_3D)
3973 depth = res->array_size;
3974 } else if (type == V_008F1C_SQ_RSRC_IMG_CUBE)
3975 depth = res->array_size / 6;
3976
3977 state[0] = 0;
3978 state[1] = S_00A004_FORMAT(img_format) |
3979 S_00A004_WIDTH_LO(width - 1);
3980 state[2] = S_00A008_WIDTH_HI((width - 1) >> 2) |
3981 S_00A008_HEIGHT(height - 1) |
3982 S_00A008_RESOURCE_LEVEL(1);
3983 state[3] = S_00A00C_DST_SEL_X(si_map_swizzle(swizzle[0])) |
3984 S_00A00C_DST_SEL_Y(si_map_swizzle(swizzle[1])) |
3985 S_00A00C_DST_SEL_Z(si_map_swizzle(swizzle[2])) |
3986 S_00A00C_DST_SEL_W(si_map_swizzle(swizzle[3])) |
3987 S_00A00C_BASE_LEVEL(res->nr_samples > 1 ?
3988 0 : first_level) |
3989 S_00A00C_LAST_LEVEL(res->nr_samples > 1 ?
3990 util_logbase2(res->nr_samples) :
3991 last_level) |
3992 S_00A00C_BC_SWIZZLE(gfx9_border_color_swizzle(desc->swizzle)) |
3993 S_00A00C_TYPE(type);
3994 /* Depth is the the last accessible layer on gfx9+. The hw doesn't need
3995 * to know the total number of layers.
3996 */
3997 state[4] = S_00A010_DEPTH((type == V_008F1C_SQ_RSRC_IMG_3D && sampler)
3998 ? depth - 1 : last_layer) |
3999 S_00A010_BASE_ARRAY(first_layer);
4000 state[5] = S_00A014_ARRAY_PITCH(!!(type == V_008F1C_SQ_RSRC_IMG_3D && !sampler)) |
4001 S_00A014_MAX_MIP(res->nr_samples > 1 ?
4002 util_logbase2(res->nr_samples) :
4003 tex->buffer.b.b.last_level) |
4004 S_00A014_PERF_MOD(4);
4005 state[6] = 0;
4006 state[7] = 0;
4007
4008 if (tex->dcc_offset) {
4009 state[6] |= S_00A018_MAX_UNCOMPRESSED_BLOCK_SIZE(V_028C78_MAX_BLOCK_SIZE_256B) |
4010 S_00A018_MAX_COMPRESSED_BLOCK_SIZE(V_028C78_MAX_BLOCK_SIZE_128B) |
4011 S_00A018_ALPHA_IS_ON_MSB(vi_alpha_is_on_msb(screen, pipe_format));
4012 }
4013
4014 /* Initialize the sampler view for FMASK. */
4015 if (tex->fmask_offset) {
4016 uint32_t format;
4017
4018 va = tex->buffer.gpu_address + tex->fmask_offset;
4019
4020 #define FMASK(s,f) (((unsigned)(MAX2(1, s)) * 16) + (MAX2(1, f)))
4021 switch (FMASK(res->nr_samples, res->nr_storage_samples)) {
4022 case FMASK(2,1):
4023 format = V_008F0C_IMG_FORMAT_FMASK8_S2_F1;
4024 break;
4025 case FMASK(2,2):
4026 format = V_008F0C_IMG_FORMAT_FMASK8_S2_F2;
4027 break;
4028 case FMASK(4,1):
4029 format = V_008F0C_IMG_FORMAT_FMASK8_S4_F1;
4030 break;
4031 case FMASK(4,2):
4032 format = V_008F0C_IMG_FORMAT_FMASK8_S4_F2;
4033 break;
4034 case FMASK(4,4):
4035 format = V_008F0C_IMG_FORMAT_FMASK8_S4_F4;
4036 break;
4037 case FMASK(8,1):
4038 format = V_008F0C_IMG_FORMAT_FMASK8_S8_F1;
4039 break;
4040 case FMASK(8,2):
4041 format = V_008F0C_IMG_FORMAT_FMASK16_S8_F2;
4042 break;
4043 case FMASK(8,4):
4044 format = V_008F0C_IMG_FORMAT_FMASK32_S8_F4;
4045 break;
4046 case FMASK(8,8):
4047 format = V_008F0C_IMG_FORMAT_FMASK32_S8_F8;
4048 break;
4049 case FMASK(16,1):
4050 format = V_008F0C_IMG_FORMAT_FMASK16_S16_F1;
4051 break;
4052 case FMASK(16,2):
4053 format = V_008F0C_IMG_FORMAT_FMASK32_S16_F2;
4054 break;
4055 case FMASK(16,4):
4056 format = V_008F0C_IMG_FORMAT_FMASK64_S16_F4;
4057 break;
4058 case FMASK(16,8):
4059 format = V_008F0C_IMG_FORMAT_FMASK64_S16_F8;
4060 break;
4061 default:
4062 unreachable("invalid nr_samples");
4063 }
4064 #undef FMASK
4065 fmask_state[0] = (va >> 8) | tex->surface.fmask_tile_swizzle;
4066 fmask_state[1] = S_00A004_BASE_ADDRESS_HI(va >> 40) |
4067 S_00A004_FORMAT(format) |
4068 S_00A004_WIDTH_LO(width - 1);
4069 fmask_state[2] = S_00A008_WIDTH_HI((width - 1) >> 2) |
4070 S_00A008_HEIGHT(height - 1) |
4071 S_00A008_RESOURCE_LEVEL(1);
4072 fmask_state[3] = S_00A00C_DST_SEL_X(V_008F1C_SQ_SEL_X) |
4073 S_00A00C_DST_SEL_Y(V_008F1C_SQ_SEL_X) |
4074 S_00A00C_DST_SEL_Z(V_008F1C_SQ_SEL_X) |
4075 S_00A00C_DST_SEL_W(V_008F1C_SQ_SEL_X) |
4076 S_00A00C_SW_MODE(tex->surface.u.gfx9.fmask.swizzle_mode) |
4077 S_00A00C_TYPE(si_tex_dim(screen, tex, target, 0));
4078 fmask_state[4] = S_00A010_DEPTH(last_layer) |
4079 S_00A010_BASE_ARRAY(first_layer);
4080 fmask_state[5] = 0;
4081 fmask_state[6] = S_00A018_META_PIPE_ALIGNED(tex->surface.u.gfx9.cmask.pipe_aligned);
4082 fmask_state[7] = 0;
4083 }
4084 }
4085
4086 /**
4087 * Build the sampler view descriptor for a texture (SI-GFX9).
4088 */
4089 static void
4090 si_make_texture_descriptor(struct si_screen *screen,
4091 struct si_texture *tex,
4092 bool sampler,
4093 enum pipe_texture_target target,
4094 enum pipe_format pipe_format,
4095 const unsigned char state_swizzle[4],
4096 unsigned first_level, unsigned last_level,
4097 unsigned first_layer, unsigned last_layer,
4098 unsigned width, unsigned height, unsigned depth,
4099 uint32_t *state,
4100 uint32_t *fmask_state)
4101 {
4102 struct pipe_resource *res = &tex->buffer.b.b;
4103 const struct util_format_description *desc;
4104 unsigned char swizzle[4];
4105 int first_non_void;
4106 unsigned num_format, data_format, type, num_samples;
4107 uint64_t va;
4108
4109 desc = util_format_description(pipe_format);
4110
4111 num_samples = desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS ?
4112 MAX2(1, res->nr_samples) :
4113 MAX2(1, res->nr_storage_samples);
4114
4115 if (desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS) {
4116 const unsigned char swizzle_xxxx[4] = {0, 0, 0, 0};
4117 const unsigned char swizzle_yyyy[4] = {1, 1, 1, 1};
4118 const unsigned char swizzle_wwww[4] = {3, 3, 3, 3};
4119
4120 switch (pipe_format) {
4121 case PIPE_FORMAT_S8_UINT_Z24_UNORM:
4122 case PIPE_FORMAT_X32_S8X24_UINT:
4123 case PIPE_FORMAT_X8Z24_UNORM:
4124 util_format_compose_swizzles(swizzle_yyyy, state_swizzle, swizzle);
4125 break;
4126 case PIPE_FORMAT_X24S8_UINT:
4127 /*
4128 * X24S8 is implemented as an 8_8_8_8 data format, to
4129 * fix texture gathers. This affects at least
4130 * GL45-CTS.texture_cube_map_array.sampling on GFX8.
4131 */
4132 if (screen->info.chip_class <= GFX8)
4133 util_format_compose_swizzles(swizzle_wwww, state_swizzle, swizzle);
4134 else
4135 util_format_compose_swizzles(swizzle_yyyy, state_swizzle, swizzle);
4136 break;
4137 default:
4138 util_format_compose_swizzles(swizzle_xxxx, state_swizzle, swizzle);
4139 }
4140 } else {
4141 util_format_compose_swizzles(desc->swizzle, state_swizzle, swizzle);
4142 }
4143
4144 first_non_void = util_format_get_first_non_void_channel(pipe_format);
4145
4146 switch (pipe_format) {
4147 case PIPE_FORMAT_S8_UINT_Z24_UNORM:
4148 num_format = V_008F14_IMG_NUM_FORMAT_UNORM;
4149 break;
4150 default:
4151 if (first_non_void < 0) {
4152 if (util_format_is_compressed(pipe_format)) {
4153 switch (pipe_format) {
4154 case PIPE_FORMAT_DXT1_SRGB:
4155 case PIPE_FORMAT_DXT1_SRGBA:
4156 case PIPE_FORMAT_DXT3_SRGBA:
4157 case PIPE_FORMAT_DXT5_SRGBA:
4158 case PIPE_FORMAT_BPTC_SRGBA:
4159 case PIPE_FORMAT_ETC2_SRGB8:
4160 case PIPE_FORMAT_ETC2_SRGB8A1:
4161 case PIPE_FORMAT_ETC2_SRGBA8:
4162 num_format = V_008F14_IMG_NUM_FORMAT_SRGB;
4163 break;
4164 case PIPE_FORMAT_RGTC1_SNORM:
4165 case PIPE_FORMAT_LATC1_SNORM:
4166 case PIPE_FORMAT_RGTC2_SNORM:
4167 case PIPE_FORMAT_LATC2_SNORM:
4168 case PIPE_FORMAT_ETC2_R11_SNORM:
4169 case PIPE_FORMAT_ETC2_RG11_SNORM:
4170 /* implies float, so use SNORM/UNORM to determine
4171 whether data is signed or not */
4172 case PIPE_FORMAT_BPTC_RGB_FLOAT:
4173 num_format = V_008F14_IMG_NUM_FORMAT_SNORM;
4174 break;
4175 default:
4176 num_format = V_008F14_IMG_NUM_FORMAT_UNORM;
4177 break;
4178 }
4179 } else if (desc->layout == UTIL_FORMAT_LAYOUT_SUBSAMPLED) {
4180 num_format = V_008F14_IMG_NUM_FORMAT_UNORM;
4181 } else {
4182 num_format = V_008F14_IMG_NUM_FORMAT_FLOAT;
4183 }
4184 } else if (desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
4185 num_format = V_008F14_IMG_NUM_FORMAT_SRGB;
4186 } else {
4187 num_format = V_008F14_IMG_NUM_FORMAT_UNORM;
4188
4189 switch (desc->channel[first_non_void].type) {
4190 case UTIL_FORMAT_TYPE_FLOAT:
4191 num_format = V_008F14_IMG_NUM_FORMAT_FLOAT;
4192 break;
4193 case UTIL_FORMAT_TYPE_SIGNED:
4194 if (desc->channel[first_non_void].normalized)
4195 num_format = V_008F14_IMG_NUM_FORMAT_SNORM;
4196 else if (desc->channel[first_non_void].pure_integer)
4197 num_format = V_008F14_IMG_NUM_FORMAT_SINT;
4198 else
4199 num_format = V_008F14_IMG_NUM_FORMAT_SSCALED;
4200 break;
4201 case UTIL_FORMAT_TYPE_UNSIGNED:
4202 if (desc->channel[first_non_void].normalized)
4203 num_format = V_008F14_IMG_NUM_FORMAT_UNORM;
4204 else if (desc->channel[first_non_void].pure_integer)
4205 num_format = V_008F14_IMG_NUM_FORMAT_UINT;
4206 else
4207 num_format = V_008F14_IMG_NUM_FORMAT_USCALED;
4208 }
4209 }
4210 }
4211
4212 data_format = si_translate_texformat(&screen->b, pipe_format, desc, first_non_void);
4213 if (data_format == ~0) {
4214 data_format = 0;
4215 }
4216
4217 /* S8 with Z32 HTILE needs a special format. */
4218 if (screen->info.chip_class == GFX9 &&
4219 pipe_format == PIPE_FORMAT_S8_UINT &&
4220 tex->tc_compatible_htile)
4221 data_format = V_008F14_IMG_DATA_FORMAT_S8_32;
4222
4223 if (!sampler &&
4224 (res->target == PIPE_TEXTURE_CUBE ||
4225 res->target == PIPE_TEXTURE_CUBE_ARRAY ||
4226 (screen->info.chip_class <= GFX8 &&
4227 res->target == PIPE_TEXTURE_3D))) {
4228 /* For the purpose of shader images, treat cube maps and 3D
4229 * textures as 2D arrays. For 3D textures, the address
4230 * calculations for mipmaps are different, so we rely on the
4231 * caller to effectively disable mipmaps.
4232 */
4233 type = V_008F1C_SQ_RSRC_IMG_2D_ARRAY;
4234
4235 assert(res->target != PIPE_TEXTURE_3D || (first_level == 0 && last_level == 0));
4236 } else {
4237 type = si_tex_dim(screen, tex, target, num_samples);
4238 }
4239
4240 if (type == V_008F1C_SQ_RSRC_IMG_1D_ARRAY) {
4241 height = 1;
4242 depth = res->array_size;
4243 } else if (type == V_008F1C_SQ_RSRC_IMG_2D_ARRAY ||
4244 type == V_008F1C_SQ_RSRC_IMG_2D_MSAA_ARRAY) {
4245 if (sampler || res->target != PIPE_TEXTURE_3D)
4246 depth = res->array_size;
4247 } else if (type == V_008F1C_SQ_RSRC_IMG_CUBE)
4248 depth = res->array_size / 6;
4249
4250 state[0] = 0;
4251 state[1] = (S_008F14_DATA_FORMAT(data_format) |
4252 S_008F14_NUM_FORMAT(num_format));
4253 state[2] = (S_008F18_WIDTH(width - 1) |
4254 S_008F18_HEIGHT(height - 1) |
4255 S_008F18_PERF_MOD(4));
4256 state[3] = (S_008F1C_DST_SEL_X(si_map_swizzle(swizzle[0])) |
4257 S_008F1C_DST_SEL_Y(si_map_swizzle(swizzle[1])) |
4258 S_008F1C_DST_SEL_Z(si_map_swizzle(swizzle[2])) |
4259 S_008F1C_DST_SEL_W(si_map_swizzle(swizzle[3])) |
4260 S_008F1C_BASE_LEVEL(num_samples > 1 ? 0 : first_level) |
4261 S_008F1C_LAST_LEVEL(num_samples > 1 ?
4262 util_logbase2(num_samples) :
4263 last_level) |
4264 S_008F1C_TYPE(type));
4265 state[4] = 0;
4266 state[5] = S_008F24_BASE_ARRAY(first_layer);
4267 state[6] = 0;
4268 state[7] = 0;
4269
4270 if (screen->info.chip_class == GFX9) {
4271 unsigned bc_swizzle = gfx9_border_color_swizzle(desc->swizzle);
4272
4273 /* Depth is the the last accessible layer on Gfx9.
4274 * The hw doesn't need to know the total number of layers.
4275 */
4276 if (type == V_008F1C_SQ_RSRC_IMG_3D)
4277 state[4] |= S_008F20_DEPTH(depth - 1);
4278 else
4279 state[4] |= S_008F20_DEPTH(last_layer);
4280
4281 state[4] |= S_008F20_BC_SWIZZLE(bc_swizzle);
4282 state[5] |= S_008F24_MAX_MIP(num_samples > 1 ?
4283 util_logbase2(num_samples) :
4284 tex->buffer.b.b.last_level);
4285 } else {
4286 state[3] |= S_008F1C_POW2_PAD(res->last_level > 0);
4287 state[4] |= S_008F20_DEPTH(depth - 1);
4288 state[5] |= S_008F24_LAST_ARRAY(last_layer);
4289 }
4290
4291 if (tex->dcc_offset) {
4292 state[6] = S_008F28_ALPHA_IS_ON_MSB(vi_alpha_is_on_msb(screen, pipe_format));
4293 } else {
4294 /* The last dword is unused by hw. The shader uses it to clear
4295 * bits in the first dword of sampler state.
4296 */
4297 if (screen->info.chip_class <= GFX7 && res->nr_samples <= 1) {
4298 if (first_level == last_level)
4299 state[7] = C_008F30_MAX_ANISO_RATIO;
4300 else
4301 state[7] = 0xffffffff;
4302 }
4303 }
4304
4305 /* Initialize the sampler view for FMASK. */
4306 if (tex->fmask_offset) {
4307 uint32_t data_format, num_format;
4308
4309 va = tex->buffer.gpu_address + tex->fmask_offset;
4310
4311 #define FMASK(s,f) (((unsigned)(MAX2(1, s)) * 16) + (MAX2(1, f)))
4312 if (screen->info.chip_class == GFX9) {
4313 data_format = V_008F14_IMG_DATA_FORMAT_FMASK;
4314 switch (FMASK(res->nr_samples, res->nr_storage_samples)) {
4315 case FMASK(2,1):
4316 num_format = V_008F14_IMG_FMASK_8_2_1;
4317 break;
4318 case FMASK(2,2):
4319 num_format = V_008F14_IMG_FMASK_8_2_2;
4320 break;
4321 case FMASK(4,1):
4322 num_format = V_008F14_IMG_FMASK_8_4_1;
4323 break;
4324 case FMASK(4,2):
4325 num_format = V_008F14_IMG_FMASK_8_4_2;
4326 break;
4327 case FMASK(4,4):
4328 num_format = V_008F14_IMG_FMASK_8_4_4;
4329 break;
4330 case FMASK(8,1):
4331 num_format = V_008F14_IMG_FMASK_8_8_1;
4332 break;
4333 case FMASK(8,2):
4334 num_format = V_008F14_IMG_FMASK_16_8_2;
4335 break;
4336 case FMASK(8,4):
4337 num_format = V_008F14_IMG_FMASK_32_8_4;
4338 break;
4339 case FMASK(8,8):
4340 num_format = V_008F14_IMG_FMASK_32_8_8;
4341 break;
4342 case FMASK(16,1):
4343 num_format = V_008F14_IMG_FMASK_16_16_1;
4344 break;
4345 case FMASK(16,2):
4346 num_format = V_008F14_IMG_FMASK_32_16_2;
4347 break;
4348 case FMASK(16,4):
4349 num_format = V_008F14_IMG_FMASK_64_16_4;
4350 break;
4351 case FMASK(16,8):
4352 num_format = V_008F14_IMG_FMASK_64_16_8;
4353 break;
4354 default:
4355 unreachable("invalid nr_samples");
4356 }
4357 } else {
4358 switch (FMASK(res->nr_samples, res->nr_storage_samples)) {
4359 case FMASK(2,1):
4360 data_format = V_008F14_IMG_DATA_FORMAT_FMASK8_S2_F1;
4361 break;
4362 case FMASK(2,2):
4363 data_format = V_008F14_IMG_DATA_FORMAT_FMASK8_S2_F2;
4364 break;
4365 case FMASK(4,1):
4366 data_format = V_008F14_IMG_DATA_FORMAT_FMASK8_S4_F1;
4367 break;
4368 case FMASK(4,2):
4369 data_format = V_008F14_IMG_DATA_FORMAT_FMASK8_S4_F2;
4370 break;
4371 case FMASK(4,4):
4372 data_format = V_008F14_IMG_DATA_FORMAT_FMASK8_S4_F4;
4373 break;
4374 case FMASK(8,1):
4375 data_format = V_008F14_IMG_DATA_FORMAT_FMASK8_S8_F1;
4376 break;
4377 case FMASK(8,2):
4378 data_format = V_008F14_IMG_DATA_FORMAT_FMASK16_S8_F2;
4379 break;
4380 case FMASK(8,4):
4381 data_format = V_008F14_IMG_DATA_FORMAT_FMASK32_S8_F4;
4382 break;
4383 case FMASK(8,8):
4384 data_format = V_008F14_IMG_DATA_FORMAT_FMASK32_S8_F8;
4385 break;
4386 case FMASK(16,1):
4387 data_format = V_008F14_IMG_DATA_FORMAT_FMASK16_S16_F1;
4388 break;
4389 case FMASK(16,2):
4390 data_format = V_008F14_IMG_DATA_FORMAT_FMASK32_S16_F2;
4391 break;
4392 case FMASK(16,4):
4393 data_format = V_008F14_IMG_DATA_FORMAT_FMASK64_S16_F4;
4394 break;
4395 case FMASK(16,8):
4396 data_format = V_008F14_IMG_DATA_FORMAT_FMASK64_S16_F8;
4397 break;
4398 default:
4399 unreachable("invalid nr_samples");
4400 }
4401 num_format = V_008F14_IMG_NUM_FORMAT_UINT;
4402 }
4403 #undef FMASK
4404
4405 fmask_state[0] = (va >> 8) | tex->surface.fmask_tile_swizzle;
4406 fmask_state[1] = S_008F14_BASE_ADDRESS_HI(va >> 40) |
4407 S_008F14_DATA_FORMAT(data_format) |
4408 S_008F14_NUM_FORMAT(num_format);
4409 fmask_state[2] = S_008F18_WIDTH(width - 1) |
4410 S_008F18_HEIGHT(height - 1);
4411 fmask_state[3] = S_008F1C_DST_SEL_X(V_008F1C_SQ_SEL_X) |
4412 S_008F1C_DST_SEL_Y(V_008F1C_SQ_SEL_X) |
4413 S_008F1C_DST_SEL_Z(V_008F1C_SQ_SEL_X) |
4414 S_008F1C_DST_SEL_W(V_008F1C_SQ_SEL_X) |
4415 S_008F1C_TYPE(si_tex_dim(screen, tex, target, 0));
4416 fmask_state[4] = 0;
4417 fmask_state[5] = S_008F24_BASE_ARRAY(first_layer);
4418 fmask_state[6] = 0;
4419 fmask_state[7] = 0;
4420
4421 if (screen->info.chip_class == GFX9) {
4422 fmask_state[3] |= S_008F1C_SW_MODE(tex->surface.u.gfx9.fmask.swizzle_mode);
4423 fmask_state[4] |= S_008F20_DEPTH(last_layer) |
4424 S_008F20_PITCH(tex->surface.u.gfx9.fmask.epitch);
4425 fmask_state[5] |= S_008F24_META_PIPE_ALIGNED(tex->surface.u.gfx9.cmask.pipe_aligned) |
4426 S_008F24_META_RB_ALIGNED(tex->surface.u.gfx9.cmask.rb_aligned);
4427 } else {
4428 fmask_state[3] |= S_008F1C_TILING_INDEX(tex->surface.u.legacy.fmask.tiling_index);
4429 fmask_state[4] |= S_008F20_DEPTH(depth - 1) |
4430 S_008F20_PITCH(tex->surface.u.legacy.fmask.pitch_in_pixels - 1);
4431 fmask_state[5] |= S_008F24_LAST_ARRAY(last_layer);
4432 }
4433 }
4434 }
4435
4436 /**
4437 * Create a sampler view.
4438 *
4439 * @param ctx context
4440 * @param texture texture
4441 * @param state sampler view template
4442 * @param width0 width0 override (for compressed textures as int)
4443 * @param height0 height0 override (for compressed textures as int)
4444 * @param force_level set the base address to the level (for compressed textures)
4445 */
4446 struct pipe_sampler_view *
4447 si_create_sampler_view_custom(struct pipe_context *ctx,
4448 struct pipe_resource *texture,
4449 const struct pipe_sampler_view *state,
4450 unsigned width0, unsigned height0,
4451 unsigned force_level)
4452 {
4453 struct si_context *sctx = (struct si_context*)ctx;
4454 struct si_sampler_view *view = CALLOC_STRUCT(si_sampler_view);
4455 struct si_texture *tex = (struct si_texture*)texture;
4456 unsigned base_level, first_level, last_level;
4457 unsigned char state_swizzle[4];
4458 unsigned height, depth, width;
4459 unsigned last_layer = state->u.tex.last_layer;
4460 enum pipe_format pipe_format;
4461 const struct legacy_surf_level *surflevel;
4462
4463 if (!view)
4464 return NULL;
4465
4466 /* initialize base object */
4467 view->base = *state;
4468 view->base.texture = NULL;
4469 view->base.reference.count = 1;
4470 view->base.context = ctx;
4471
4472 assert(texture);
4473 pipe_resource_reference(&view->base.texture, texture);
4474
4475 if (state->format == PIPE_FORMAT_X24S8_UINT ||
4476 state->format == PIPE_FORMAT_S8X24_UINT ||
4477 state->format == PIPE_FORMAT_X32_S8X24_UINT ||
4478 state->format == PIPE_FORMAT_S8_UINT)
4479 view->is_stencil_sampler = true;
4480
4481 /* Buffer resource. */
4482 if (texture->target == PIPE_BUFFER) {
4483 si_make_buffer_descriptor(sctx->screen,
4484 si_resource(texture),
4485 state->format,
4486 state->u.buf.offset,
4487 state->u.buf.size,
4488 view->state);
4489 return &view->base;
4490 }
4491
4492 state_swizzle[0] = state->swizzle_r;
4493 state_swizzle[1] = state->swizzle_g;
4494 state_swizzle[2] = state->swizzle_b;
4495 state_swizzle[3] = state->swizzle_a;
4496
4497 base_level = 0;
4498 first_level = state->u.tex.first_level;
4499 last_level = state->u.tex.last_level;
4500 width = width0;
4501 height = height0;
4502 depth = texture->depth0;
4503
4504 if (sctx->chip_class <= GFX8 && force_level) {
4505 assert(force_level == first_level &&
4506 force_level == last_level);
4507 base_level = force_level;
4508 first_level = 0;
4509 last_level = 0;
4510 width = u_minify(width, force_level);
4511 height = u_minify(height, force_level);
4512 depth = u_minify(depth, force_level);
4513 }
4514
4515 /* This is not needed if state trackers set last_layer correctly. */
4516 if (state->target == PIPE_TEXTURE_1D ||
4517 state->target == PIPE_TEXTURE_2D ||
4518 state->target == PIPE_TEXTURE_RECT ||
4519 state->target == PIPE_TEXTURE_CUBE)
4520 last_layer = state->u.tex.first_layer;
4521
4522 /* Texturing with separate depth and stencil. */
4523 pipe_format = state->format;
4524
4525 /* Depth/stencil texturing sometimes needs separate texture. */
4526 if (tex->is_depth && !si_can_sample_zs(tex, view->is_stencil_sampler)) {
4527 if (!tex->flushed_depth_texture &&
4528 !si_init_flushed_depth_texture(ctx, texture)) {
4529 pipe_resource_reference(&view->base.texture, NULL);
4530 FREE(view);
4531 return NULL;
4532 }
4533
4534 assert(tex->flushed_depth_texture);
4535
4536 /* Override format for the case where the flushed texture
4537 * contains only Z or only S.
4538 */
4539 if (tex->flushed_depth_texture->buffer.b.b.format != tex->buffer.b.b.format)
4540 pipe_format = tex->flushed_depth_texture->buffer.b.b.format;
4541
4542 tex = tex->flushed_depth_texture;
4543 }
4544
4545 surflevel = tex->surface.u.legacy.level;
4546
4547 if (tex->db_compatible) {
4548 if (!view->is_stencil_sampler)
4549 pipe_format = tex->db_render_format;
4550
4551 switch (pipe_format) {
4552 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
4553 pipe_format = PIPE_FORMAT_Z32_FLOAT;
4554 break;
4555 case PIPE_FORMAT_X8Z24_UNORM:
4556 case PIPE_FORMAT_S8_UINT_Z24_UNORM:
4557 /* Z24 is always stored like this for DB
4558 * compatibility.
4559 */
4560 pipe_format = PIPE_FORMAT_Z24X8_UNORM;
4561 break;
4562 case PIPE_FORMAT_X24S8_UINT:
4563 case PIPE_FORMAT_S8X24_UINT:
4564 case PIPE_FORMAT_X32_S8X24_UINT:
4565 pipe_format = PIPE_FORMAT_S8_UINT;
4566 surflevel = tex->surface.u.legacy.stencil_level;
4567 break;
4568 default:;
4569 }
4570 }
4571
4572 view->dcc_incompatible =
4573 vi_dcc_formats_are_incompatible(texture,
4574 state->u.tex.first_level,
4575 state->format);
4576
4577 sctx->screen->make_texture_descriptor(sctx->screen, tex, true,
4578 state->target, pipe_format, state_swizzle,
4579 first_level, last_level,
4580 state->u.tex.first_layer, last_layer,
4581 width, height, depth,
4582 view->state, view->fmask_state);
4583
4584 const struct util_format_description *desc = util_format_description(pipe_format);
4585 view->is_integer = false;
4586
4587 for (unsigned i = 0; i < desc->nr_channels; ++i) {
4588 if (desc->channel[i].type == UTIL_FORMAT_TYPE_VOID)
4589 continue;
4590
4591 /* Whether the number format is {U,S}{SCALED,INT} */
4592 view->is_integer =
4593 (desc->channel[i].type == UTIL_FORMAT_TYPE_UNSIGNED ||
4594 desc->channel[i].type == UTIL_FORMAT_TYPE_SIGNED) &&
4595 (desc->channel[i].pure_integer || !desc->channel[i].normalized);
4596 break;
4597 }
4598
4599 view->base_level_info = &surflevel[base_level];
4600 view->base_level = base_level;
4601 view->block_width = util_format_get_blockwidth(pipe_format);
4602 return &view->base;
4603 }
4604
4605 static struct pipe_sampler_view *
4606 si_create_sampler_view(struct pipe_context *ctx,
4607 struct pipe_resource *texture,
4608 const struct pipe_sampler_view *state)
4609 {
4610 return si_create_sampler_view_custom(ctx, texture, state,
4611 texture ? texture->width0 : 0,
4612 texture ? texture->height0 : 0, 0);
4613 }
4614
4615 static void si_sampler_view_destroy(struct pipe_context *ctx,
4616 struct pipe_sampler_view *state)
4617 {
4618 struct si_sampler_view *view = (struct si_sampler_view *)state;
4619
4620 pipe_resource_reference(&state->texture, NULL);
4621 FREE(view);
4622 }
4623
4624 static bool wrap_mode_uses_border_color(unsigned wrap, bool linear_filter)
4625 {
4626 return wrap == PIPE_TEX_WRAP_CLAMP_TO_BORDER ||
4627 wrap == PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER ||
4628 (linear_filter &&
4629 (wrap == PIPE_TEX_WRAP_CLAMP ||
4630 wrap == PIPE_TEX_WRAP_MIRROR_CLAMP));
4631 }
4632
4633 static uint32_t si_translate_border_color(struct si_context *sctx,
4634 const struct pipe_sampler_state *state,
4635 const union pipe_color_union *color,
4636 bool is_integer)
4637 {
4638 bool linear_filter = state->min_img_filter != PIPE_TEX_FILTER_NEAREST ||
4639 state->mag_img_filter != PIPE_TEX_FILTER_NEAREST;
4640
4641 if (!wrap_mode_uses_border_color(state->wrap_s, linear_filter) &&
4642 !wrap_mode_uses_border_color(state->wrap_t, linear_filter) &&
4643 !wrap_mode_uses_border_color(state->wrap_r, linear_filter))
4644 return S_008F3C_BORDER_COLOR_TYPE(V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK);
4645
4646 #define simple_border_types(elt) \
4647 do { \
4648 if (color->elt[0] == 0 && color->elt[1] == 0 && \
4649 color->elt[2] == 0 && color->elt[3] == 0) \
4650 return S_008F3C_BORDER_COLOR_TYPE(V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK); \
4651 if (color->elt[0] == 0 && color->elt[1] == 0 && \
4652 color->elt[2] == 0 && color->elt[3] == 1) \
4653 return S_008F3C_BORDER_COLOR_TYPE(V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_BLACK); \
4654 if (color->elt[0] == 1 && color->elt[1] == 1 && \
4655 color->elt[2] == 1 && color->elt[3] == 1) \
4656 return S_008F3C_BORDER_COLOR_TYPE(V_008F3C_SQ_TEX_BORDER_COLOR_OPAQUE_WHITE); \
4657 } while (false)
4658
4659 if (is_integer)
4660 simple_border_types(ui);
4661 else
4662 simple_border_types(f);
4663
4664 #undef simple_border_types
4665
4666 int i;
4667
4668 /* Check if the border has been uploaded already. */
4669 for (i = 0; i < sctx->border_color_count; i++)
4670 if (memcmp(&sctx->border_color_table[i], color,
4671 sizeof(*color)) == 0)
4672 break;
4673
4674 if (i >= SI_MAX_BORDER_COLORS) {
4675 /* Getting 4096 unique border colors is very unlikely. */
4676 fprintf(stderr, "radeonsi: The border color table is full. "
4677 "Any new border colors will be just black. "
4678 "Please file a bug.\n");
4679 return S_008F3C_BORDER_COLOR_TYPE(V_008F3C_SQ_TEX_BORDER_COLOR_TRANS_BLACK);
4680 }
4681
4682 if (i == sctx->border_color_count) {
4683 /* Upload a new border color. */
4684 memcpy(&sctx->border_color_table[i], color,
4685 sizeof(*color));
4686 util_memcpy_cpu_to_le32(&sctx->border_color_map[i],
4687 color, sizeof(*color));
4688 sctx->border_color_count++;
4689 }
4690
4691 return S_008F3C_BORDER_COLOR_PTR(i) |
4692 S_008F3C_BORDER_COLOR_TYPE(V_008F3C_SQ_TEX_BORDER_COLOR_REGISTER);
4693 }
4694
4695 static inline int S_FIXED(float value, unsigned frac_bits)
4696 {
4697 return value * (1 << frac_bits);
4698 }
4699
4700 static inline unsigned si_tex_filter(unsigned filter, unsigned max_aniso)
4701 {
4702 if (filter == PIPE_TEX_FILTER_LINEAR)
4703 return max_aniso > 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_BILINEAR
4704 : V_008F38_SQ_TEX_XY_FILTER_BILINEAR;
4705 else
4706 return max_aniso > 1 ? V_008F38_SQ_TEX_XY_FILTER_ANISO_POINT
4707 : V_008F38_SQ_TEX_XY_FILTER_POINT;
4708 }
4709
4710 static inline unsigned si_tex_aniso_filter(unsigned filter)
4711 {
4712 if (filter < 2)
4713 return 0;
4714 if (filter < 4)
4715 return 1;
4716 if (filter < 8)
4717 return 2;
4718 if (filter < 16)
4719 return 3;
4720 return 4;
4721 }
4722
4723 static void *si_create_sampler_state(struct pipe_context *ctx,
4724 const struct pipe_sampler_state *state)
4725 {
4726 struct si_context *sctx = (struct si_context *)ctx;
4727 struct si_screen *sscreen = sctx->screen;
4728 struct si_sampler_state *rstate = CALLOC_STRUCT(si_sampler_state);
4729 unsigned max_aniso = sscreen->force_aniso >= 0 ? sscreen->force_aniso
4730 : state->max_anisotropy;
4731 unsigned max_aniso_ratio = si_tex_aniso_filter(max_aniso);
4732 union pipe_color_union clamped_border_color;
4733
4734 if (!rstate) {
4735 return NULL;
4736 }
4737
4738 #ifndef NDEBUG
4739 rstate->magic = SI_SAMPLER_STATE_MAGIC;
4740 #endif
4741 rstate->val[0] = (S_008F30_CLAMP_X(si_tex_wrap(state->wrap_s)) |
4742 S_008F30_CLAMP_Y(si_tex_wrap(state->wrap_t)) |
4743 S_008F30_CLAMP_Z(si_tex_wrap(state->wrap_r)) |
4744 S_008F30_MAX_ANISO_RATIO(max_aniso_ratio) |
4745 S_008F30_DEPTH_COMPARE_FUNC(si_tex_compare(state->compare_func)) |
4746 S_008F30_FORCE_UNNORMALIZED(!state->normalized_coords) |
4747 S_008F30_ANISO_THRESHOLD(max_aniso_ratio >> 1) |
4748 S_008F30_ANISO_BIAS(max_aniso_ratio) |
4749 S_008F30_DISABLE_CUBE_WRAP(!state->seamless_cube_map) |
4750 S_008F30_COMPAT_MODE(sctx->chip_class == GFX8 || sctx->chip_class == GFX9));
4751 rstate->val[1] = (S_008F34_MIN_LOD(S_FIXED(CLAMP(state->min_lod, 0, 15), 8)) |
4752 S_008F34_MAX_LOD(S_FIXED(CLAMP(state->max_lod, 0, 15), 8)) |
4753 S_008F34_PERF_MIP(max_aniso_ratio ? max_aniso_ratio + 6 : 0));
4754 rstate->val[2] = (S_008F38_LOD_BIAS(S_FIXED(CLAMP(state->lod_bias, -16, 16), 8)) |
4755 S_008F38_XY_MAG_FILTER(si_tex_filter(state->mag_img_filter, max_aniso)) |
4756 S_008F38_XY_MIN_FILTER(si_tex_filter(state->min_img_filter, max_aniso)) |
4757 S_008F38_MIP_FILTER(si_tex_mipfilter(state->min_mip_filter)) |
4758 S_008F38_MIP_POINT_PRECLAMP(0));
4759 rstate->val[3] = si_translate_border_color(sctx, state, &state->border_color, false);
4760
4761 if (sscreen->info.chip_class >= GFX10) {
4762 rstate->val[2] |= S_008F38_ANISO_OVERRIDE_GFX10(1);
4763 } else {
4764 rstate->val[2] |= S_008F38_DISABLE_LSB_CEIL(sctx->chip_class <= GFX8) |
4765 S_008F38_FILTER_PREC_FIX(1) |
4766 S_008F38_ANISO_OVERRIDE_GFX6(sctx->chip_class >= GFX8);
4767 }
4768
4769 /* Create sampler resource for integer textures. */
4770 memcpy(rstate->integer_val, rstate->val, sizeof(rstate->val));
4771 rstate->integer_val[3] = si_translate_border_color(sctx, state, &state->border_color, true);
4772
4773 /* Create sampler resource for upgraded depth textures. */
4774 memcpy(rstate->upgraded_depth_val, rstate->val, sizeof(rstate->val));
4775
4776 for (unsigned i = 0; i < 4; ++i) {
4777 /* Use channel 0 on purpose, so that we can use OPAQUE_WHITE
4778 * when the border color is 1.0. */
4779 clamped_border_color.f[i] = CLAMP(state->border_color.f[0], 0, 1);
4780 }
4781
4782 if (memcmp(&state->border_color, &clamped_border_color, sizeof(clamped_border_color)) == 0) {
4783 if (sscreen->info.chip_class <= GFX9)
4784 rstate->upgraded_depth_val[3] |= S_008F3C_UPGRADED_DEPTH(1);
4785 } else {
4786 rstate->upgraded_depth_val[3] =
4787 si_translate_border_color(sctx, state, &clamped_border_color, false);
4788 }
4789
4790 return rstate;
4791 }
4792
4793 static void si_set_sample_mask(struct pipe_context *ctx, unsigned sample_mask)
4794 {
4795 struct si_context *sctx = (struct si_context *)ctx;
4796
4797 if (sctx->sample_mask == (uint16_t)sample_mask)
4798 return;
4799
4800 sctx->sample_mask = sample_mask;
4801 si_mark_atom_dirty(sctx, &sctx->atoms.s.sample_mask);
4802 }
4803
4804 static void si_emit_sample_mask(struct si_context *sctx)
4805 {
4806 struct radeon_cmdbuf *cs = sctx->gfx_cs;
4807 unsigned mask = sctx->sample_mask;
4808
4809 /* Needed for line and polygon smoothing as well as for the Polaris
4810 * small primitive filter. We expect the state tracker to take care of
4811 * this for us.
4812 */
4813 assert(mask == 0xffff || sctx->framebuffer.nr_samples > 1 ||
4814 (mask & 1 && sctx->blitter->running));
4815
4816 radeon_set_context_reg_seq(cs, R_028C38_PA_SC_AA_MASK_X0Y0_X1Y0, 2);
4817 radeon_emit(cs, mask | (mask << 16));
4818 radeon_emit(cs, mask | (mask << 16));
4819 }
4820
4821 static void si_delete_sampler_state(struct pipe_context *ctx, void *state)
4822 {
4823 #ifndef NDEBUG
4824 struct si_sampler_state *s = state;
4825
4826 assert(s->magic == SI_SAMPLER_STATE_MAGIC);
4827 s->magic = 0;
4828 #endif
4829 free(state);
4830 }
4831
4832 /*
4833 * Vertex elements & buffers
4834 */
4835
4836 struct si_fast_udiv_info32
4837 si_compute_fast_udiv_info32(uint32_t D, unsigned num_bits)
4838 {
4839 struct util_fast_udiv_info info =
4840 util_compute_fast_udiv_info(D, num_bits, 32);
4841
4842 struct si_fast_udiv_info32 result = {
4843 info.multiplier,
4844 info.pre_shift,
4845 info.post_shift,
4846 info.increment,
4847 };
4848 return result;
4849 }
4850
4851 static void *si_create_vertex_elements(struct pipe_context *ctx,
4852 unsigned count,
4853 const struct pipe_vertex_element *elements)
4854 {
4855 struct si_screen *sscreen = (struct si_screen*)ctx->screen;
4856 struct si_vertex_elements *v = CALLOC_STRUCT(si_vertex_elements);
4857 bool used[SI_NUM_VERTEX_BUFFERS] = {};
4858 struct si_fast_udiv_info32 divisor_factors[SI_MAX_ATTRIBS] = {};
4859 STATIC_ASSERT(sizeof(struct si_fast_udiv_info32) == 16);
4860 STATIC_ASSERT(sizeof(divisor_factors[0].multiplier) == 4);
4861 STATIC_ASSERT(sizeof(divisor_factors[0].pre_shift) == 4);
4862 STATIC_ASSERT(sizeof(divisor_factors[0].post_shift) == 4);
4863 STATIC_ASSERT(sizeof(divisor_factors[0].increment) == 4);
4864 int i;
4865
4866 assert(count <= SI_MAX_ATTRIBS);
4867 if (!v)
4868 return NULL;
4869
4870 v->count = count;
4871 v->desc_list_byte_size = align(count * 16, SI_CPDMA_ALIGNMENT);
4872
4873 for (i = 0; i < count; ++i) {
4874 const struct util_format_description *desc;
4875 const struct util_format_channel_description *channel;
4876 int first_non_void;
4877 unsigned vbo_index = elements[i].vertex_buffer_index;
4878
4879 if (vbo_index >= SI_NUM_VERTEX_BUFFERS) {
4880 FREE(v);
4881 return NULL;
4882 }
4883
4884 unsigned instance_divisor = elements[i].instance_divisor;
4885 if (instance_divisor) {
4886 v->uses_instance_divisors = true;
4887
4888 if (instance_divisor == 1) {
4889 v->instance_divisor_is_one |= 1u << i;
4890 } else {
4891 v->instance_divisor_is_fetched |= 1u << i;
4892 divisor_factors[i] =
4893 si_compute_fast_udiv_info32(instance_divisor, 32);
4894 }
4895 }
4896
4897 if (!used[vbo_index]) {
4898 v->first_vb_use_mask |= 1 << i;
4899 used[vbo_index] = true;
4900 }
4901
4902 desc = util_format_description(elements[i].src_format);
4903 first_non_void = util_format_get_first_non_void_channel(elements[i].src_format);
4904 channel = first_non_void >= 0 ? &desc->channel[first_non_void] : NULL;
4905
4906 v->format_size[i] = desc->block.bits / 8;
4907 v->src_offset[i] = elements[i].src_offset;
4908 v->vertex_buffer_index[i] = vbo_index;
4909
4910 bool always_fix = false;
4911 union si_vs_fix_fetch fix_fetch;
4912 unsigned log_hw_load_size; /* the load element size as seen by the hardware */
4913
4914 fix_fetch.bits = 0;
4915 log_hw_load_size = MIN2(2, util_logbase2(desc->block.bits) - 3);
4916
4917 if (channel) {
4918 switch (channel->type) {
4919 case UTIL_FORMAT_TYPE_FLOAT: fix_fetch.u.format = AC_FETCH_FORMAT_FLOAT; break;
4920 case UTIL_FORMAT_TYPE_FIXED: fix_fetch.u.format = AC_FETCH_FORMAT_FIXED; break;
4921 case UTIL_FORMAT_TYPE_SIGNED: {
4922 if (channel->pure_integer)
4923 fix_fetch.u.format = AC_FETCH_FORMAT_SINT;
4924 else if (channel->normalized)
4925 fix_fetch.u.format = AC_FETCH_FORMAT_SNORM;
4926 else
4927 fix_fetch.u.format = AC_FETCH_FORMAT_SSCALED;
4928 break;
4929 }
4930 case UTIL_FORMAT_TYPE_UNSIGNED: {
4931 if (channel->pure_integer)
4932 fix_fetch.u.format = AC_FETCH_FORMAT_UINT;
4933 else if (channel->normalized)
4934 fix_fetch.u.format = AC_FETCH_FORMAT_UNORM;
4935 else
4936 fix_fetch.u.format = AC_FETCH_FORMAT_USCALED;
4937 break;
4938 }
4939 default: unreachable("bad format type");
4940 }
4941 } else {
4942 switch (elements[i].src_format) {
4943 case PIPE_FORMAT_R11G11B10_FLOAT: fix_fetch.u.format = AC_FETCH_FORMAT_FLOAT; break;
4944 default: unreachable("bad other format");
4945 }
4946 }
4947
4948 if (desc->channel[0].size == 10) {
4949 fix_fetch.u.log_size = 3; /* special encoding for 2_10_10_10 */
4950 log_hw_load_size = 2;
4951
4952 /* The hardware always treats the 2-bit alpha channel as
4953 * unsigned, so a shader workaround is needed. The affected
4954 * chips are GFX8 and older except Stoney (GFX8.1).
4955 */
4956 always_fix = sscreen->info.chip_class <= GFX8 &&
4957 sscreen->info.family != CHIP_STONEY &&
4958 channel->type == UTIL_FORMAT_TYPE_SIGNED;
4959 } else if (elements[i].src_format == PIPE_FORMAT_R11G11B10_FLOAT) {
4960 fix_fetch.u.log_size = 3; /* special encoding */
4961 fix_fetch.u.format = AC_FETCH_FORMAT_FIXED;
4962 log_hw_load_size = 2;
4963 } else {
4964 fix_fetch.u.log_size = util_logbase2(channel->size) - 3;
4965 fix_fetch.u.num_channels_m1 = desc->nr_channels - 1;
4966
4967 /* Always fix up:
4968 * - doubles (multiple loads + truncate to float)
4969 * - 32-bit requiring a conversion
4970 */
4971 always_fix =
4972 (fix_fetch.u.log_size == 3) ||
4973 (fix_fetch.u.log_size == 2 &&
4974 fix_fetch.u.format != AC_FETCH_FORMAT_FLOAT &&
4975 fix_fetch.u.format != AC_FETCH_FORMAT_UINT &&
4976 fix_fetch.u.format != AC_FETCH_FORMAT_SINT);
4977
4978 /* Also fixup 8_8_8 and 16_16_16. */
4979 if (desc->nr_channels == 3 && fix_fetch.u.log_size <= 1) {
4980 always_fix = true;
4981 log_hw_load_size = fix_fetch.u.log_size;
4982 }
4983 }
4984
4985 if (desc->swizzle[0] != PIPE_SWIZZLE_X) {
4986 assert(desc->swizzle[0] == PIPE_SWIZZLE_Z &&
4987 (desc->swizzle[2] == PIPE_SWIZZLE_X || desc->swizzle[2] == PIPE_SWIZZLE_0));
4988 fix_fetch.u.reverse = 1;
4989 }
4990
4991 /* Force the workaround for unaligned access here already if the
4992 * offset relative to the vertex buffer base is unaligned.
4993 *
4994 * There is a theoretical case in which this is too conservative:
4995 * if the vertex buffer's offset is also unaligned in just the
4996 * right way, we end up with an aligned address after all.
4997 * However, this case should be extremely rare in practice (it
4998 * won't happen in well-behaved applications), and taking it
4999 * into account would complicate the fast path (where everything
5000 * is nicely aligned).
5001 */
5002 bool check_alignment =
5003 log_hw_load_size >= 1 &&
5004 (sscreen->info.chip_class == GFX6 || sscreen->info.chip_class == GFX10);
5005 bool opencode = sscreen->options.vs_fetch_always_opencode;
5006
5007 if (check_alignment &&
5008 (elements[i].src_offset & ((1 << log_hw_load_size) - 1)) != 0)
5009 opencode = true;
5010
5011 if (always_fix || check_alignment || opencode)
5012 v->fix_fetch[i] = fix_fetch.bits;
5013
5014 if (opencode)
5015 v->fix_fetch_opencode |= 1 << i;
5016 if (opencode || always_fix)
5017 v->fix_fetch_always |= 1 << i;
5018
5019 if (check_alignment && !opencode) {
5020 assert(log_hw_load_size == 1 || log_hw_load_size == 2);
5021
5022 v->fix_fetch_unaligned |= 1 << i;
5023 v->hw_load_is_dword |= (log_hw_load_size - 1) << i;
5024 v->vb_alignment_check_mask |= 1 << vbo_index;
5025 }
5026
5027 v->rsrc_word3[i] = S_008F0C_DST_SEL_X(si_map_swizzle(desc->swizzle[0])) |
5028 S_008F0C_DST_SEL_Y(si_map_swizzle(desc->swizzle[1])) |
5029 S_008F0C_DST_SEL_Z(si_map_swizzle(desc->swizzle[2])) |
5030 S_008F0C_DST_SEL_W(si_map_swizzle(desc->swizzle[3]));
5031
5032 if (sscreen->info.chip_class >= GFX10) {
5033 const struct gfx10_format *fmt =
5034 &gfx10_format_table[elements[i].src_format];
5035 assert(fmt->img_format != 0 && fmt->img_format < 128);
5036 v->rsrc_word3[i] |= S_008F0C_FORMAT(fmt->img_format) |
5037 S_008F0C_RESOURCE_LEVEL(1);
5038 } else {
5039 unsigned data_format, num_format;
5040 data_format = si_translate_buffer_dataformat(ctx->screen, desc, first_non_void);
5041 num_format = si_translate_buffer_numformat(ctx->screen, desc, first_non_void);
5042 v->rsrc_word3[i] |= S_008F0C_NUM_FORMAT(num_format) |
5043 S_008F0C_DATA_FORMAT(data_format);
5044 }
5045 }
5046
5047 if (v->instance_divisor_is_fetched) {
5048 unsigned num_divisors = util_last_bit(v->instance_divisor_is_fetched);
5049
5050 v->instance_divisor_factor_buffer =
5051 (struct si_resource*)
5052 pipe_buffer_create(&sscreen->b, 0, PIPE_USAGE_DEFAULT,
5053 num_divisors * sizeof(divisor_factors[0]));
5054 if (!v->instance_divisor_factor_buffer) {
5055 FREE(v);
5056 return NULL;
5057 }
5058 void *map = sscreen->ws->buffer_map(v->instance_divisor_factor_buffer->buf,
5059 NULL, PIPE_TRANSFER_WRITE);
5060 memcpy(map , divisor_factors, num_divisors * sizeof(divisor_factors[0]));
5061 }
5062 return v;
5063 }
5064
5065 static void si_bind_vertex_elements(struct pipe_context *ctx, void *state)
5066 {
5067 struct si_context *sctx = (struct si_context *)ctx;
5068 struct si_vertex_elements *old = sctx->vertex_elements;
5069 struct si_vertex_elements *v = (struct si_vertex_elements*)state;
5070
5071 sctx->vertex_elements = v;
5072 sctx->vertex_buffers_dirty = true;
5073
5074 if (v &&
5075 (!old ||
5076 old->count != v->count ||
5077 old->uses_instance_divisors != v->uses_instance_divisors ||
5078 /* we don't check which divisors changed */
5079 v->uses_instance_divisors ||
5080 (old->vb_alignment_check_mask ^ v->vb_alignment_check_mask) & sctx->vertex_buffer_unaligned ||
5081 ((v->vb_alignment_check_mask & sctx->vertex_buffer_unaligned) &&
5082 memcmp(old->vertex_buffer_index, v->vertex_buffer_index,
5083 sizeof(v->vertex_buffer_index[0]) * v->count)) ||
5084 /* fix_fetch_{always,opencode,unaligned} and hw_load_is_dword are
5085 * functions of fix_fetch and the src_offset alignment.
5086 * If they change and fix_fetch doesn't, it must be due to different
5087 * src_offset alignment, which is reflected in fix_fetch_opencode. */
5088 old->fix_fetch_opencode != v->fix_fetch_opencode ||
5089 memcmp(old->fix_fetch, v->fix_fetch, sizeof(v->fix_fetch[0]) * v->count)))
5090 sctx->do_update_shaders = true;
5091
5092 if (v && v->instance_divisor_is_fetched) {
5093 struct pipe_constant_buffer cb;
5094
5095 cb.buffer = &v->instance_divisor_factor_buffer->b.b;
5096 cb.user_buffer = NULL;
5097 cb.buffer_offset = 0;
5098 cb.buffer_size = 0xffffffff;
5099 si_set_rw_buffer(sctx, SI_VS_CONST_INSTANCE_DIVISORS, &cb);
5100 }
5101 }
5102
5103 static void si_delete_vertex_element(struct pipe_context *ctx, void *state)
5104 {
5105 struct si_context *sctx = (struct si_context *)ctx;
5106 struct si_vertex_elements *v = (struct si_vertex_elements*)state;
5107
5108 if (sctx->vertex_elements == state)
5109 sctx->vertex_elements = NULL;
5110 si_resource_reference(&v->instance_divisor_factor_buffer, NULL);
5111 FREE(state);
5112 }
5113
5114 static void si_set_vertex_buffers(struct pipe_context *ctx,
5115 unsigned start_slot, unsigned count,
5116 const struct pipe_vertex_buffer *buffers)
5117 {
5118 struct si_context *sctx = (struct si_context *)ctx;
5119 struct pipe_vertex_buffer *dst = sctx->vertex_buffer + start_slot;
5120 uint32_t orig_unaligned = sctx->vertex_buffer_unaligned;
5121 uint32_t unaligned = orig_unaligned;
5122 int i;
5123
5124 assert(start_slot + count <= ARRAY_SIZE(sctx->vertex_buffer));
5125
5126 if (buffers) {
5127 for (i = 0; i < count; i++) {
5128 const struct pipe_vertex_buffer *src = buffers + i;
5129 struct pipe_vertex_buffer *dsti = dst + i;
5130 struct pipe_resource *buf = src->buffer.resource;
5131
5132 pipe_resource_reference(&dsti->buffer.resource, buf);
5133 dsti->buffer_offset = src->buffer_offset;
5134 dsti->stride = src->stride;
5135 if (dsti->buffer_offset & 3 || dsti->stride & 3)
5136 unaligned |= 1 << (start_slot + i);
5137 else
5138 unaligned &= ~(1 << (start_slot + i));
5139
5140 si_context_add_resource_size(sctx, buf);
5141 if (buf)
5142 si_resource(buf)->bind_history |= PIPE_BIND_VERTEX_BUFFER;
5143 }
5144 } else {
5145 for (i = 0; i < count; i++) {
5146 pipe_resource_reference(&dst[i].buffer.resource, NULL);
5147 }
5148 unaligned &= ~u_bit_consecutive(start_slot, count);
5149 }
5150 sctx->vertex_buffers_dirty = true;
5151 sctx->vertex_buffer_unaligned = unaligned;
5152
5153 /* Check whether alignment may have changed in a way that requires
5154 * shader changes. This check is conservative: a vertex buffer can only
5155 * trigger a shader change if the misalignment amount changes (e.g.
5156 * from byte-aligned to short-aligned), but we only keep track of
5157 * whether buffers are at least dword-aligned, since that should always
5158 * be the case in well-behaved applications anyway.
5159 */
5160 if (sctx->vertex_elements &&
5161 (sctx->vertex_elements->vb_alignment_check_mask &
5162 (unaligned | orig_unaligned) & u_bit_consecutive(start_slot, count)))
5163 sctx->do_update_shaders = true;
5164 }
5165
5166 /*
5167 * Misc
5168 */
5169
5170 static void si_set_tess_state(struct pipe_context *ctx,
5171 const float default_outer_level[4],
5172 const float default_inner_level[2])
5173 {
5174 struct si_context *sctx = (struct si_context *)ctx;
5175 struct pipe_constant_buffer cb;
5176 float array[8];
5177
5178 memcpy(array, default_outer_level, sizeof(float) * 4);
5179 memcpy(array+4, default_inner_level, sizeof(float) * 2);
5180
5181 cb.buffer = NULL;
5182 cb.user_buffer = NULL;
5183 cb.buffer_size = sizeof(array);
5184
5185 si_upload_const_buffer(sctx, (struct si_resource**)&cb.buffer,
5186 (void*)array, sizeof(array),
5187 &cb.buffer_offset);
5188
5189 si_set_rw_buffer(sctx, SI_HS_CONST_DEFAULT_TESS_LEVELS, &cb);
5190 pipe_resource_reference(&cb.buffer, NULL);
5191 }
5192
5193 static void si_texture_barrier(struct pipe_context *ctx, unsigned flags)
5194 {
5195 struct si_context *sctx = (struct si_context *)ctx;
5196
5197 si_update_fb_dirtiness_after_rendering(sctx);
5198
5199 /* Multisample surfaces are flushed in si_decompress_textures. */
5200 if (sctx->framebuffer.uncompressed_cb_mask) {
5201 si_make_CB_shader_coherent(sctx, sctx->framebuffer.nr_samples,
5202 sctx->framebuffer.CB_has_shader_readable_metadata,
5203 sctx->framebuffer.all_DCC_pipe_aligned);
5204 }
5205 }
5206
5207 /* This only ensures coherency for shader image/buffer stores. */
5208 static void si_memory_barrier(struct pipe_context *ctx, unsigned flags)
5209 {
5210 struct si_context *sctx = (struct si_context *)ctx;
5211
5212 if (!(flags & ~PIPE_BARRIER_UPDATE))
5213 return;
5214
5215 /* Subsequent commands must wait for all shader invocations to
5216 * complete. */
5217 sctx->flags |= SI_CONTEXT_PS_PARTIAL_FLUSH |
5218 SI_CONTEXT_CS_PARTIAL_FLUSH;
5219
5220 if (flags & PIPE_BARRIER_CONSTANT_BUFFER)
5221 sctx->flags |= SI_CONTEXT_INV_SCACHE |
5222 SI_CONTEXT_INV_VCACHE;
5223
5224 if (flags & (PIPE_BARRIER_VERTEX_BUFFER |
5225 PIPE_BARRIER_SHADER_BUFFER |
5226 PIPE_BARRIER_TEXTURE |
5227 PIPE_BARRIER_IMAGE |
5228 PIPE_BARRIER_STREAMOUT_BUFFER |
5229 PIPE_BARRIER_GLOBAL_BUFFER)) {
5230 /* As far as I can tell, L1 contents are written back to L2
5231 * automatically at end of shader, but the contents of other
5232 * L1 caches might still be stale. */
5233 sctx->flags |= SI_CONTEXT_INV_VCACHE;
5234 }
5235
5236 if (flags & PIPE_BARRIER_INDEX_BUFFER) {
5237 /* Indices are read through TC L2 since GFX8.
5238 * L1 isn't used.
5239 */
5240 if (sctx->screen->info.chip_class <= GFX7)
5241 sctx->flags |= SI_CONTEXT_WB_L2;
5242 }
5243
5244 /* MSAA color, any depth and any stencil are flushed in
5245 * si_decompress_textures when needed.
5246 */
5247 if (flags & PIPE_BARRIER_FRAMEBUFFER &&
5248 sctx->framebuffer.uncompressed_cb_mask) {
5249 sctx->flags |= SI_CONTEXT_FLUSH_AND_INV_CB;
5250
5251 if (sctx->chip_class <= GFX8)
5252 sctx->flags |= SI_CONTEXT_WB_L2;
5253 }
5254
5255 /* Indirect buffers use TC L2 on GFX9, but not older hw. */
5256 if (sctx->screen->info.chip_class <= GFX8 &&
5257 flags & PIPE_BARRIER_INDIRECT_BUFFER)
5258 sctx->flags |= SI_CONTEXT_WB_L2;
5259 }
5260
5261 static void *si_create_blend_custom(struct si_context *sctx, unsigned mode)
5262 {
5263 struct pipe_blend_state blend;
5264
5265 memset(&blend, 0, sizeof(blend));
5266 blend.independent_blend_enable = true;
5267 blend.rt[0].colormask = 0xf;
5268 return si_create_blend_state_mode(&sctx->b, &blend, mode);
5269 }
5270
5271 static void si_init_config(struct si_context *sctx);
5272
5273 void si_init_state_compute_functions(struct si_context *sctx)
5274 {
5275 sctx->b.create_sampler_state = si_create_sampler_state;
5276 sctx->b.delete_sampler_state = si_delete_sampler_state;
5277 sctx->b.create_sampler_view = si_create_sampler_view;
5278 sctx->b.sampler_view_destroy = si_sampler_view_destroy;
5279 sctx->b.memory_barrier = si_memory_barrier;
5280 }
5281
5282 void si_init_state_functions(struct si_context *sctx)
5283 {
5284 sctx->atoms.s.framebuffer.emit = si_emit_framebuffer_state;
5285 sctx->atoms.s.msaa_sample_locs.emit = si_emit_msaa_sample_locs;
5286 sctx->atoms.s.db_render_state.emit = si_emit_db_render_state;
5287 sctx->atoms.s.dpbb_state.emit = si_emit_dpbb_state;
5288 sctx->atoms.s.msaa_config.emit = si_emit_msaa_config;
5289 sctx->atoms.s.sample_mask.emit = si_emit_sample_mask;
5290 sctx->atoms.s.cb_render_state.emit = si_emit_cb_render_state;
5291 sctx->atoms.s.blend_color.emit = si_emit_blend_color;
5292 sctx->atoms.s.clip_regs.emit = si_emit_clip_regs;
5293 sctx->atoms.s.clip_state.emit = si_emit_clip_state;
5294 sctx->atoms.s.stencil_ref.emit = si_emit_stencil_ref;
5295
5296 sctx->b.create_blend_state = si_create_blend_state;
5297 sctx->b.bind_blend_state = si_bind_blend_state;
5298 sctx->b.delete_blend_state = si_delete_blend_state;
5299 sctx->b.set_blend_color = si_set_blend_color;
5300
5301 sctx->b.create_rasterizer_state = si_create_rs_state;
5302 sctx->b.bind_rasterizer_state = si_bind_rs_state;
5303 sctx->b.delete_rasterizer_state = si_delete_rs_state;
5304
5305 sctx->b.create_depth_stencil_alpha_state = si_create_dsa_state;
5306 sctx->b.bind_depth_stencil_alpha_state = si_bind_dsa_state;
5307 sctx->b.delete_depth_stencil_alpha_state = si_delete_dsa_state;
5308
5309 sctx->custom_dsa_flush = si_create_db_flush_dsa(sctx);
5310 sctx->custom_blend_resolve = si_create_blend_custom(sctx, V_028808_CB_RESOLVE);
5311 sctx->custom_blend_fmask_decompress = si_create_blend_custom(sctx, V_028808_CB_FMASK_DECOMPRESS);
5312 sctx->custom_blend_eliminate_fastclear = si_create_blend_custom(sctx, V_028808_CB_ELIMINATE_FAST_CLEAR);
5313 sctx->custom_blend_dcc_decompress = si_create_blend_custom(sctx, V_028808_CB_DCC_DECOMPRESS);
5314
5315 sctx->b.set_clip_state = si_set_clip_state;
5316 sctx->b.set_stencil_ref = si_set_stencil_ref;
5317
5318 sctx->b.set_framebuffer_state = si_set_framebuffer_state;
5319
5320 sctx->b.set_sample_mask = si_set_sample_mask;
5321
5322 sctx->b.create_vertex_elements_state = si_create_vertex_elements;
5323 sctx->b.bind_vertex_elements_state = si_bind_vertex_elements;
5324 sctx->b.delete_vertex_elements_state = si_delete_vertex_element;
5325 sctx->b.set_vertex_buffers = si_set_vertex_buffers;
5326
5327 sctx->b.texture_barrier = si_texture_barrier;
5328 sctx->b.set_min_samples = si_set_min_samples;
5329 sctx->b.set_tess_state = si_set_tess_state;
5330
5331 sctx->b.set_active_query_state = si_set_active_query_state;
5332
5333 si_init_config(sctx);
5334 }
5335
5336 void si_init_screen_state_functions(struct si_screen *sscreen)
5337 {
5338 sscreen->b.is_format_supported = si_is_format_supported;
5339
5340 if (sscreen->info.chip_class >= GFX10) {
5341 sscreen->make_texture_descriptor = gfx10_make_texture_descriptor;
5342 } else {
5343 sscreen->make_texture_descriptor = si_make_texture_descriptor;
5344 }
5345 }
5346
5347 static void si_set_grbm_gfx_index(struct si_context *sctx,
5348 struct si_pm4_state *pm4, unsigned value)
5349 {
5350 unsigned reg = sctx->chip_class >= GFX7 ? R_030800_GRBM_GFX_INDEX :
5351 R_00802C_GRBM_GFX_INDEX;
5352 si_pm4_set_reg(pm4, reg, value);
5353 }
5354
5355 static void si_set_grbm_gfx_index_se(struct si_context *sctx,
5356 struct si_pm4_state *pm4, unsigned se)
5357 {
5358 assert(se == ~0 || se < sctx->screen->info.max_se);
5359 si_set_grbm_gfx_index(sctx, pm4,
5360 (se == ~0 ? S_030800_SE_BROADCAST_WRITES(1) :
5361 S_030800_SE_INDEX(se)) |
5362 S_030800_SH_BROADCAST_WRITES(1) |
5363 S_030800_INSTANCE_BROADCAST_WRITES(1));
5364 }
5365
5366 static void
5367 si_write_harvested_raster_configs(struct si_context *sctx,
5368 struct si_pm4_state *pm4,
5369 unsigned raster_config,
5370 unsigned raster_config_1)
5371 {
5372 unsigned num_se = MAX2(sctx->screen->info.max_se, 1);
5373 unsigned raster_config_se[4];
5374 unsigned se;
5375
5376 ac_get_harvested_configs(&sctx->screen->info,
5377 raster_config,
5378 &raster_config_1,
5379 raster_config_se);
5380
5381 for (se = 0; se < num_se; se++) {
5382 si_set_grbm_gfx_index_se(sctx, pm4, se);
5383 si_pm4_set_reg(pm4, R_028350_PA_SC_RASTER_CONFIG, raster_config_se[se]);
5384 }
5385 si_set_grbm_gfx_index(sctx, pm4, ~0);
5386
5387 if (sctx->chip_class >= GFX7) {
5388 si_pm4_set_reg(pm4, R_028354_PA_SC_RASTER_CONFIG_1, raster_config_1);
5389 }
5390 }
5391
5392 static void si_set_raster_config(struct si_context *sctx, struct si_pm4_state *pm4)
5393 {
5394 struct si_screen *sscreen = sctx->screen;
5395 unsigned num_rb = MIN2(sscreen->info.num_render_backends, 16);
5396 unsigned rb_mask = sscreen->info.enabled_rb_mask;
5397 unsigned raster_config = sscreen->pa_sc_raster_config;
5398 unsigned raster_config_1 = sscreen->pa_sc_raster_config_1;
5399
5400 if (!rb_mask || util_bitcount(rb_mask) >= num_rb) {
5401 /* Always use the default config when all backends are enabled
5402 * (or when we failed to determine the enabled backends).
5403 */
5404 si_pm4_set_reg(pm4, R_028350_PA_SC_RASTER_CONFIG,
5405 raster_config);
5406 if (sctx->chip_class >= GFX7)
5407 si_pm4_set_reg(pm4, R_028354_PA_SC_RASTER_CONFIG_1,
5408 raster_config_1);
5409 } else {
5410 si_write_harvested_raster_configs(sctx, pm4, raster_config, raster_config_1);
5411 }
5412 }
5413
5414 static void si_init_config(struct si_context *sctx)
5415 {
5416 struct si_screen *sscreen = sctx->screen;
5417 uint64_t border_color_va = sctx->border_color_buffer->gpu_address;
5418 bool has_clear_state = sscreen->has_clear_state;
5419 struct si_pm4_state *pm4 = CALLOC_STRUCT(si_pm4_state);
5420
5421 if (!pm4)
5422 return;
5423
5424 si_pm4_cmd_begin(pm4, PKT3_CONTEXT_CONTROL);
5425 si_pm4_cmd_add(pm4, CONTEXT_CONTROL_LOAD_ENABLE(1));
5426 si_pm4_cmd_add(pm4, CONTEXT_CONTROL_SHADOW_ENABLE(1));
5427 si_pm4_cmd_end(pm4, false);
5428
5429 if (has_clear_state) {
5430 si_pm4_cmd_begin(pm4, PKT3_CLEAR_STATE);
5431 si_pm4_cmd_add(pm4, 0);
5432 si_pm4_cmd_end(pm4, false);
5433 }
5434
5435 if (sctx->chip_class <= GFX8)
5436 si_set_raster_config(sctx, pm4);
5437
5438 si_pm4_set_reg(pm4, R_028A18_VGT_HOS_MAX_TESS_LEVEL, fui(64));
5439 if (!has_clear_state)
5440 si_pm4_set_reg(pm4, R_028A1C_VGT_HOS_MIN_TESS_LEVEL, fui(0));
5441
5442 /* FIXME calculate these values somehow ??? */
5443 if (sctx->chip_class <= GFX8) {
5444 si_pm4_set_reg(pm4, R_028A54_VGT_GS_PER_ES, SI_GS_PER_ES);
5445 si_pm4_set_reg(pm4, R_028A58_VGT_ES_PER_GS, 0x40);
5446 }
5447
5448 if (!has_clear_state) {
5449 si_pm4_set_reg(pm4, R_028A5C_VGT_GS_PER_VS, 0x2);
5450 si_pm4_set_reg(pm4, R_028A8C_VGT_PRIMITIVEID_RESET, 0x0);
5451 si_pm4_set_reg(pm4, R_028B98_VGT_STRMOUT_BUFFER_CONFIG, 0x0);
5452 }
5453
5454 si_pm4_set_reg(pm4, R_028AA0_VGT_INSTANCE_STEP_RATE_0, 1);
5455 if (!has_clear_state)
5456 si_pm4_set_reg(pm4, R_028AB8_VGT_VTX_CNT_EN, 0x0);
5457 if (sctx->chip_class < GFX7)
5458 si_pm4_set_reg(pm4, R_008A14_PA_CL_ENHANCE, S_008A14_NUM_CLIP_SEQ(3) |
5459 S_008A14_CLIP_VTX_REORDER_ENA(1));
5460
5461 /* CLEAR_STATE doesn't clear these correctly on certain generations.
5462 * I don't know why. Deduced by trial and error.
5463 */
5464 if (sctx->chip_class <= GFX7 || !has_clear_state) {
5465 si_pm4_set_reg(pm4, R_028B28_VGT_STRMOUT_DRAW_OPAQUE_OFFSET, 0);
5466 si_pm4_set_reg(pm4, R_028204_PA_SC_WINDOW_SCISSOR_TL, S_028204_WINDOW_OFFSET_DISABLE(1));
5467 si_pm4_set_reg(pm4, R_028240_PA_SC_GENERIC_SCISSOR_TL, S_028240_WINDOW_OFFSET_DISABLE(1));
5468 si_pm4_set_reg(pm4, R_028244_PA_SC_GENERIC_SCISSOR_BR,
5469 S_028244_BR_X(16384) | S_028244_BR_Y(16384));
5470 si_pm4_set_reg(pm4, R_028030_PA_SC_SCREEN_SCISSOR_TL, 0);
5471 si_pm4_set_reg(pm4, R_028034_PA_SC_SCREEN_SCISSOR_BR,
5472 S_028034_BR_X(16384) | S_028034_BR_Y(16384));
5473 }
5474
5475 if (!has_clear_state) {
5476 si_pm4_set_reg(pm4, R_028230_PA_SC_EDGERULE,
5477 S_028230_ER_TRI(0xA) |
5478 S_028230_ER_POINT(0xA) |
5479 S_028230_ER_RECT(0xA) |
5480 /* Required by DX10_DIAMOND_TEST_ENA: */
5481 S_028230_ER_LINE_LR(0x1A) |
5482 S_028230_ER_LINE_RL(0x26) |
5483 S_028230_ER_LINE_TB(0xA) |
5484 S_028230_ER_LINE_BT(0xA));
5485 si_pm4_set_reg(pm4, R_028820_PA_CL_NANINF_CNTL, 0);
5486 si_pm4_set_reg(pm4, R_028AC0_DB_SRESULTS_COMPARE_STATE0, 0x0);
5487 si_pm4_set_reg(pm4, R_028AC4_DB_SRESULTS_COMPARE_STATE1, 0x0);
5488 si_pm4_set_reg(pm4, R_028AC8_DB_PRELOAD_CONTROL, 0x0);
5489 si_pm4_set_reg(pm4, R_02800C_DB_RENDER_OVERRIDE, 0);
5490 }
5491
5492 if (sctx->chip_class >= GFX10) {
5493 si_pm4_set_reg(pm4, R_028A98_VGT_DRAW_PAYLOAD_CNTL, 0);
5494 si_pm4_set_reg(pm4, R_030964_GE_MAX_VTX_INDX, ~0);
5495 si_pm4_set_reg(pm4, R_030924_GE_MIN_VTX_INDX, 0);
5496 si_pm4_set_reg(pm4, R_030928_GE_INDX_OFFSET, 0);
5497 si_pm4_set_reg(pm4, R_03097C_GE_STEREO_CNTL, 0);
5498 si_pm4_set_reg(pm4, R_030988_GE_USER_VGPR_EN, 0);
5499 } else if (sctx->chip_class == GFX9) {
5500 si_pm4_set_reg(pm4, R_030920_VGT_MAX_VTX_INDX, ~0);
5501 si_pm4_set_reg(pm4, R_030924_VGT_MIN_VTX_INDX, 0);
5502 si_pm4_set_reg(pm4, R_030928_VGT_INDX_OFFSET, 0);
5503 } else {
5504 /* These registers, when written, also overwrite the CLEAR_STATE
5505 * context, so we can't rely on CLEAR_STATE setting them.
5506 * It would be an issue if there was another UMD changing them.
5507 */
5508 si_pm4_set_reg(pm4, R_028400_VGT_MAX_VTX_INDX, ~0);
5509 si_pm4_set_reg(pm4, R_028404_VGT_MIN_VTX_INDX, 0);
5510 si_pm4_set_reg(pm4, R_028408_VGT_INDX_OFFSET, 0);
5511 }
5512
5513 if (sctx->chip_class >= GFX7) {
5514 if (sctx->chip_class >= GFX10) {
5515 /* Logical CUs 16 - 31 */
5516 si_pm4_set_reg(pm4, R_00B404_SPI_SHADER_PGM_RSRC4_HS,
5517 S_00B404_CU_EN(0xffff));
5518 si_pm4_set_reg(pm4, R_00B104_SPI_SHADER_PGM_RSRC4_VS,
5519 S_00B104_CU_EN(0xffff));
5520 si_pm4_set_reg(pm4, R_00B004_SPI_SHADER_PGM_RSRC4_PS,
5521 S_00B004_CU_EN(0xffff));
5522 }
5523
5524 if (sctx->chip_class >= GFX9) {
5525 si_pm4_set_reg(pm4, R_00B41C_SPI_SHADER_PGM_RSRC3_HS,
5526 S_00B41C_CU_EN(0xffff) | S_00B41C_WAVE_LIMIT(0x3F));
5527 } else {
5528 si_pm4_set_reg(pm4, R_00B51C_SPI_SHADER_PGM_RSRC3_LS,
5529 S_00B51C_CU_EN(0xffff) | S_00B51C_WAVE_LIMIT(0x3F));
5530 si_pm4_set_reg(pm4, R_00B41C_SPI_SHADER_PGM_RSRC3_HS,
5531 S_00B41C_WAVE_LIMIT(0x3F));
5532 si_pm4_set_reg(pm4, R_00B31C_SPI_SHADER_PGM_RSRC3_ES,
5533 S_00B31C_CU_EN(0xffff) | S_00B31C_WAVE_LIMIT(0x3F));
5534
5535 /* If this is 0, Bonaire can hang even if GS isn't being used.
5536 * Other chips are unaffected. These are suboptimal values,
5537 * but we don't use on-chip GS.
5538 */
5539 si_pm4_set_reg(pm4, R_028A44_VGT_GS_ONCHIP_CNTL,
5540 S_028A44_ES_VERTS_PER_SUBGRP(64) |
5541 S_028A44_GS_PRIMS_PER_SUBGRP(4));
5542 }
5543
5544 /* Compute LATE_ALLOC_VS.LIMIT. */
5545 unsigned num_cu_per_sh = sscreen->info.num_good_cu_per_sh;
5546 unsigned late_alloc_limit; /* The limit is per SH. */
5547
5548 if (sctx->family == CHIP_KABINI) {
5549 late_alloc_limit = 0; /* Potential hang on Kabini. */
5550 } else if (num_cu_per_sh <= 4) {
5551 /* Too few available compute units per SH. Disallowing
5552 * VS to run on one CU could hurt us more than late VS
5553 * allocation would help.
5554 *
5555 * 2 is the highest safe number that allows us to keep
5556 * all CUs enabled.
5557 */
5558 late_alloc_limit = 2;
5559 } else {
5560 /* This is a good initial value, allowing 1 late_alloc
5561 * wave per SIMD on num_cu - 2.
5562 */
5563 late_alloc_limit = (num_cu_per_sh - 2) * 4;
5564 }
5565
5566 unsigned cu_mask_vs = 0xffff;
5567 unsigned cu_mask_gs = 0xffff;
5568
5569 if (late_alloc_limit > 2) {
5570 if (sctx->chip_class >= GFX10) {
5571 /* CU2 & CU3 disabled because of the dual CU design */
5572 cu_mask_vs = 0xfff3;
5573 cu_mask_gs = 0xfff3; /* NGG only */
5574 } else {
5575 cu_mask_vs = 0xfffe; /* 1 CU disabled */
5576 }
5577 }
5578
5579 /* VS can't execute on one CU if the limit is > 2. */
5580 si_pm4_set_reg(pm4, R_00B118_SPI_SHADER_PGM_RSRC3_VS,
5581 S_00B118_CU_EN(cu_mask_vs) |
5582 S_00B118_WAVE_LIMIT(0x3F));
5583 si_pm4_set_reg(pm4, R_00B11C_SPI_SHADER_LATE_ALLOC_VS,
5584 S_00B11C_LIMIT(late_alloc_limit));
5585
5586 si_pm4_set_reg(pm4, R_00B21C_SPI_SHADER_PGM_RSRC3_GS,
5587 S_00B21C_CU_EN(cu_mask_gs) | S_00B21C_WAVE_LIMIT(0x3F));
5588
5589 if (sctx->chip_class >= GFX10) {
5590 si_pm4_set_reg(pm4, R_00B204_SPI_SHADER_PGM_RSRC4_GS,
5591 S_00B204_CU_EN(0xffff) |
5592 S_00B204_SPI_SHADER_LATE_ALLOC_GS_GFX10(late_alloc_limit));
5593 }
5594
5595 si_pm4_set_reg(pm4, R_00B01C_SPI_SHADER_PGM_RSRC3_PS,
5596 S_00B01C_CU_EN(0xffff) | S_00B01C_WAVE_LIMIT(0x3F));
5597 }
5598
5599 if (sctx->chip_class >= GFX10) {
5600 /* Break up a pixel wave if it contains deallocs for more than
5601 * half the parameter cache.
5602 *
5603 * To avoid a deadlock where pixel waves aren't launched
5604 * because they're waiting for more pixels while the frontend
5605 * is stuck waiting for PC space, the maximum allowed value is
5606 * the size of the PC minus the largest possible allocation for
5607 * a single primitive shader subgroup.
5608 */
5609 si_pm4_set_reg(pm4, R_028C50_PA_SC_NGG_MODE_CNTL,
5610 S_028C50_MAX_DEALLOCS_IN_WAVE(512));
5611 si_pm4_set_reg(pm4, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL, 14);
5612 si_pm4_set_reg(pm4, R_02835C_PA_SC_TILE_STEERING_OVERRIDE,
5613 sscreen->info.pa_sc_tile_steering_override);
5614
5615 si_pm4_set_reg(pm4, R_02807C_DB_RMI_L2_CACHE_CONTROL,
5616 S_02807C_Z_WR_POLICY(V_02807C_CACHE_STREAM_WR) |
5617 S_02807C_S_WR_POLICY(V_02807C_CACHE_STREAM_WR) |
5618 S_02807C_HTILE_WR_POLICY(V_02807C_CACHE_STREAM_WR) |
5619 S_02807C_ZPCPSD_WR_POLICY(V_02807C_CACHE_STREAM_WR) |
5620 S_02807C_Z_RD_POLICY(V_02807C_CACHE_NOA_RD) |
5621 S_02807C_S_RD_POLICY(V_02807C_CACHE_NOA_RD) |
5622 S_02807C_HTILE_RD_POLICY(V_02807C_CACHE_NOA_RD));
5623
5624 si_pm4_set_reg(pm4, R_028410_CB_RMI_GL2_CACHE_CONTROL,
5625 S_028410_CMASK_WR_POLICY(V_028410_CACHE_STREAM_WR) |
5626 S_028410_FMASK_WR_POLICY(V_028410_CACHE_STREAM_WR) |
5627 S_028410_DCC_WR_POLICY(V_028410_CACHE_STREAM_WR) |
5628 S_028410_COLOR_WR_POLICY(V_028410_CACHE_STREAM_WR) |
5629 S_028410_CMASK_RD_POLICY(V_028410_CACHE_NOA_RD) |
5630 S_028410_FMASK_RD_POLICY(V_028410_CACHE_NOA_RD) |
5631 S_028410_DCC_RD_POLICY(V_028410_CACHE_NOA_RD) |
5632 S_028410_COLOR_RD_POLICY(V_028410_CACHE_NOA_RD));
5633 si_pm4_set_reg(pm4, R_028428_CB_COVERAGE_OUT_CONTROL, 0);
5634
5635 si_pm4_set_reg(pm4, R_00B0C0_SPI_SHADER_REQ_CTRL_PS,
5636 S_00B0C0_SOFT_GROUPING_EN(1) |
5637 S_00B0C0_NUMBER_OF_REQUESTS_PER_CU(4 - 1));
5638 si_pm4_set_reg(pm4, R_00B1C0_SPI_SHADER_REQ_CTRL_VS, 0);
5639
5640 }
5641
5642 if (sctx->chip_class >= GFX8) {
5643 unsigned vgt_tess_distribution;
5644
5645 vgt_tess_distribution =
5646 S_028B50_ACCUM_ISOLINE(32) |
5647 S_028B50_ACCUM_TRI(11) |
5648 S_028B50_ACCUM_QUAD(11) |
5649 S_028B50_DONUT_SPLIT(16);
5650
5651 /* Testing with Unigine Heaven extreme tesselation yielded best results
5652 * with TRAP_SPLIT = 3.
5653 */
5654 if (sctx->family == CHIP_FIJI ||
5655 sctx->family >= CHIP_POLARIS10)
5656 vgt_tess_distribution |= S_028B50_TRAP_SPLIT(3);
5657
5658 si_pm4_set_reg(pm4, R_028B50_VGT_TESS_DISTRIBUTION, vgt_tess_distribution);
5659 } else if (!has_clear_state) {
5660 si_pm4_set_reg(pm4, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL, 14);
5661 si_pm4_set_reg(pm4, R_028C5C_VGT_OUT_DEALLOC_CNTL, 16);
5662 }
5663
5664 si_pm4_set_reg(pm4, R_028080_TA_BC_BASE_ADDR, border_color_va >> 8);
5665 if (sctx->chip_class >= GFX7) {
5666 si_pm4_set_reg(pm4, R_028084_TA_BC_BASE_ADDR_HI,
5667 S_028084_ADDRESS(border_color_va >> 40));
5668 }
5669 si_pm4_add_bo(pm4, sctx->border_color_buffer, RADEON_USAGE_READ,
5670 RADEON_PRIO_BORDER_COLORS);
5671
5672 if (sctx->chip_class >= GFX9) {
5673 unsigned num_se = sscreen->info.max_se;
5674 unsigned pc_lines = 0;
5675 unsigned max_alloc_count = 0;
5676
5677 switch (sctx->family) {
5678 case CHIP_VEGA10:
5679 case CHIP_VEGA12:
5680 case CHIP_VEGA20:
5681 pc_lines = 2048;
5682 break;
5683 case CHIP_RAVEN:
5684 case CHIP_RAVEN2:
5685 case CHIP_NAVI10:
5686 case CHIP_NAVI12:
5687 pc_lines = 1024;
5688 break;
5689 case CHIP_NAVI14:
5690 pc_lines = 512;
5691 break;
5692 default:
5693 assert(0);
5694 }
5695
5696 if (sctx->chip_class >= GFX10) {
5697 max_alloc_count = pc_lines / 3;
5698 } else {
5699 max_alloc_count = MIN2(128, pc_lines / (4 * num_se));
5700 }
5701
5702 si_pm4_set_reg(pm4, R_028C48_PA_SC_BINNER_CNTL_1,
5703 S_028C48_MAX_ALLOC_COUNT(max_alloc_count) |
5704 S_028C48_MAX_PRIM_PER_BATCH(1023));
5705 si_pm4_set_reg(pm4, R_028C4C_PA_SC_CONSERVATIVE_RASTERIZATION_CNTL,
5706 S_028C4C_NULL_SQUAD_AA_MASK_ENABLE(1));
5707 si_pm4_set_reg(pm4, R_030968_VGT_INSTANCE_BASE_ID, 0);
5708 }
5709
5710 si_pm4_upload_indirect_buffer(sctx, pm4);
5711 sctx->init_config = pm4;
5712 }