projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
radv: rely on shader's wavesize when computing NGG info
[mesa.git] / src / amd / vulkan / radv_pipeline.c
1 /*
2 * Copyright © 2016 Red Hat.
3 * Copyright © 2016 Bas Nieuwenhuizen
4 *
5 * based in part on anv driver which is:
6 * Copyright © 2015 Intel Corporation
7 *
8 * Permission is hereby granted, free of charge, to any person obtaining a
9 * copy of this software and associated documentation files (the "Software"),
10 * to deal in the Software without restriction, including without limitation
11 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
12 * and/or sell copies of the Software, and to permit persons to whom the
13 * Software is furnished to do so, subject to the following conditions:
14 *
15 * The above copyright notice and this permission notice (including the next
16 * paragraph) shall be included in all copies or substantial portions of the
17 * Software.
18 *
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
22 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
23 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
24 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
25 * IN THE SOFTWARE.
26 */
27
28 #include "util/disk_cache.h"
29 #include "util/mesa-sha1.h"
30 #include "util/u_atomic.h"
31 #include "radv_debug.h"
32 #include "radv_private.h"
33 #include "radv_cs.h"
34 #include "radv_shader.h"
35 #include "nir/nir.h"
36 #include "nir/nir_builder.h"
37 #include "nir/nir_xfb_info.h"
38 #include "spirv/nir_spirv.h"
39 #include "vk_util.h"
40
41 #include <llvm-c/Core.h>
42 #include <llvm-c/TargetMachine.h>
43
44 #include "sid.h"
45 #include "ac_binary.h"
46 #include "ac_llvm_util.h"
47 #include "ac_nir_to_llvm.h"
48 #include "vk_format.h"
49 #include "util/debug.h"
50 #include "ac_exp_param.h"
51 #include "ac_shader_util.h"
52 #include "main/menums.h"
53
54 struct radv_blend_state {
55 uint32_t blend_enable_4bit;
56 uint32_t need_src_alpha;
57
58 uint32_t cb_color_control;
59 uint32_t cb_target_mask;
60 uint32_t cb_target_enabled_4bit;
61 uint32_t sx_mrt_blend_opt[8];
62 uint32_t cb_blend_control[8];
63
64 uint32_t spi_shader_col_format;
65 uint32_t cb_shader_mask;
66 uint32_t db_alpha_to_mask;
67
68 uint32_t commutative_4bit;
69
70 bool single_cb_enable;
71 bool mrt0_is_dual_src;
72 };
73
74 struct radv_dsa_order_invariance {
75 /* Whether the final result in Z/S buffers is guaranteed to be
76 * invariant under changes to the order in which fragments arrive.
77 */
78 bool zs;
79
80 /* Whether the set of fragments that pass the combined Z/S test is
81 * guaranteed to be invariant under changes to the order in which
82 * fragments arrive.
83 */
84 bool pass_set;
85 };
86
87 struct radv_tessellation_state {
88 uint32_t ls_hs_config;
89 unsigned num_patches;
90 unsigned lds_size;
91 uint32_t tf_param;
92 };
93
94 bool radv_pipeline_has_ngg(const struct radv_pipeline *pipeline)
95 {
96 struct radv_shader_variant *variant = NULL;
97 if (pipeline->shaders[MESA_SHADER_GEOMETRY])
98 variant = pipeline->shaders[MESA_SHADER_GEOMETRY];
99 else if (pipeline->shaders[MESA_SHADER_TESS_EVAL])
100 variant = pipeline->shaders[MESA_SHADER_TESS_EVAL];
101 else if (pipeline->shaders[MESA_SHADER_VERTEX])
102 variant = pipeline->shaders[MESA_SHADER_VERTEX];
103 else
104 return false;
105 return variant->info.is_ngg;
106 }
107
108 bool radv_pipeline_has_gs_copy_shader(const struct radv_pipeline *pipeline)
109 {
110 if (!radv_pipeline_has_gs(pipeline))
111 return false;
112
113 /* The GS copy shader is required if the pipeline has GS on GFX6-GFX9.
114 * On GFX10, it might be required in rare cases if it's not possible to
115 * enable NGG.
116 */
117 if (radv_pipeline_has_ngg(pipeline))
118 return false;
119
120 assert(pipeline->gs_copy_shader);
121 return true;
122 }
123
124 static void
125 radv_pipeline_destroy(struct radv_device *device,
126 struct radv_pipeline *pipeline,
127 const VkAllocationCallbacks* allocator)
128 {
129 for (unsigned i = 0; i < MESA_SHADER_STAGES; ++i)
130 if (pipeline->shaders[i])
131 radv_shader_variant_destroy(device, pipeline->shaders[i]);
132
133 if (pipeline->gs_copy_shader)
134 radv_shader_variant_destroy(device, pipeline->gs_copy_shader);
135
136 if(pipeline->cs.buf)
137 free(pipeline->cs.buf);
138 vk_free2(&device->alloc, allocator, pipeline);
139 }
140
141 void radv_DestroyPipeline(
142 VkDevice _device,
143 VkPipeline _pipeline,
144 const VkAllocationCallbacks* pAllocator)
145 {
146 RADV_FROM_HANDLE(radv_device, device, _device);
147 RADV_FROM_HANDLE(radv_pipeline, pipeline, _pipeline);
148
149 if (!_pipeline)
150 return;
151
152 radv_pipeline_destroy(device, pipeline, pAllocator);
153 }
154
155 static uint32_t get_hash_flags(struct radv_device *device)
156 {
157 uint32_t hash_flags = 0;
158
159 if (device->instance->debug_flags & RADV_DEBUG_UNSAFE_MATH)
160 hash_flags |= RADV_HASH_SHADER_UNSAFE_MATH;
161 if (device->instance->debug_flags & RADV_DEBUG_NO_NGG)
162 hash_flags |= RADV_HASH_SHADER_NO_NGG;
163 if (device->instance->perftest_flags & RADV_PERFTEST_SISCHED)
164 hash_flags |= RADV_HASH_SHADER_SISCHED;
165 if (device->physical_device->cs_wave_size == 32)
166 hash_flags |= RADV_HASH_SHADER_CS_WAVE32;
167 if (device->physical_device->ps_wave_size == 32)
168 hash_flags |= RADV_HASH_SHADER_PS_WAVE32;
169 if (device->physical_device->ge_wave_size == 32)
170 hash_flags |= RADV_HASH_SHADER_GE_WAVE32;
171 if (device->physical_device->use_aco)
172 hash_flags |= RADV_HASH_SHADER_ACO;
173 return hash_flags;
174 }
175
176 static VkResult
177 radv_pipeline_scratch_init(struct radv_device *device,
178 struct radv_pipeline *pipeline)
179 {
180 unsigned scratch_bytes_per_wave = 0;
181 unsigned max_waves = 0;
182 unsigned min_waves = 1;
183
184 for (int i = 0; i < MESA_SHADER_STAGES; ++i) {
185 if (pipeline->shaders[i]) {
186 unsigned max_stage_waves = device->scratch_waves;
187
188 scratch_bytes_per_wave = MAX2(scratch_bytes_per_wave,
189 pipeline->shaders[i]->config.scratch_bytes_per_wave);
190
191 max_stage_waves = MIN2(max_stage_waves,
192 4 * device->physical_device->rad_info.num_good_compute_units *
193 (256 / pipeline->shaders[i]->config.num_vgprs));
194 max_waves = MAX2(max_waves, max_stage_waves);
195 }
196 }
197
198 if (pipeline->shaders[MESA_SHADER_COMPUTE]) {
199 unsigned group_size = pipeline->shaders[MESA_SHADER_COMPUTE]->info.cs.block_size[0] *
200 pipeline->shaders[MESA_SHADER_COMPUTE]->info.cs.block_size[1] *
201 pipeline->shaders[MESA_SHADER_COMPUTE]->info.cs.block_size[2];
202 min_waves = MAX2(min_waves, round_up_u32(group_size, 64));
203 }
204
205 if (scratch_bytes_per_wave)
206 max_waves = MIN2(max_waves, 0xffffffffu / scratch_bytes_per_wave);
207
208 if (scratch_bytes_per_wave && max_waves < min_waves) {
209 /* Not really true at this moment, but will be true on first
210 * execution. Avoid having hanging shaders. */
211 return vk_error(device->instance, VK_ERROR_OUT_OF_DEVICE_MEMORY);
212 }
213 pipeline->scratch_bytes_per_wave = scratch_bytes_per_wave;
214 pipeline->max_waves = max_waves;
215 return VK_SUCCESS;
216 }
217
218 static uint32_t si_translate_blend_logic_op(VkLogicOp op)
219 {
220 switch (op) {
221 case VK_LOGIC_OP_CLEAR:
222 return V_028808_ROP3_CLEAR;
223 case VK_LOGIC_OP_AND:
224 return V_028808_ROP3_AND;
225 case VK_LOGIC_OP_AND_REVERSE:
226 return V_028808_ROP3_AND_REVERSE;
227 case VK_LOGIC_OP_COPY:
228 return V_028808_ROP3_COPY;
229 case VK_LOGIC_OP_AND_INVERTED:
230 return V_028808_ROP3_AND_INVERTED;
231 case VK_LOGIC_OP_NO_OP:
232 return V_028808_ROP3_NO_OP;
233 case VK_LOGIC_OP_XOR:
234 return V_028808_ROP3_XOR;
235 case VK_LOGIC_OP_OR:
236 return V_028808_ROP3_OR;
237 case VK_LOGIC_OP_NOR:
238 return V_028808_ROP3_NOR;
239 case VK_LOGIC_OP_EQUIVALENT:
240 return V_028808_ROP3_EQUIVALENT;
241 case VK_LOGIC_OP_INVERT:
242 return V_028808_ROP3_INVERT;
243 case VK_LOGIC_OP_OR_REVERSE:
244 return V_028808_ROP3_OR_REVERSE;
245 case VK_LOGIC_OP_COPY_INVERTED:
246 return V_028808_ROP3_COPY_INVERTED;
247 case VK_LOGIC_OP_OR_INVERTED:
248 return V_028808_ROP3_OR_INVERTED;
249 case VK_LOGIC_OP_NAND:
250 return V_028808_ROP3_NAND;
251 case VK_LOGIC_OP_SET:
252 return V_028808_ROP3_SET;
253 default:
254 unreachable("Unhandled logic op");
255 }
256 }
257
258
259 static uint32_t si_translate_blend_function(VkBlendOp op)
260 {
261 switch (op) {
262 case VK_BLEND_OP_ADD:
263 return V_028780_COMB_DST_PLUS_SRC;
264 case VK_BLEND_OP_SUBTRACT:
265 return V_028780_COMB_SRC_MINUS_DST;
266 case VK_BLEND_OP_REVERSE_SUBTRACT:
267 return V_028780_COMB_DST_MINUS_SRC;
268 case VK_BLEND_OP_MIN:
269 return V_028780_COMB_MIN_DST_SRC;
270 case VK_BLEND_OP_MAX:
271 return V_028780_COMB_MAX_DST_SRC;
272 default:
273 return 0;
274 }
275 }
276
277 static uint32_t si_translate_blend_factor(VkBlendFactor factor)
278 {
279 switch (factor) {
280 case VK_BLEND_FACTOR_ZERO:
281 return V_028780_BLEND_ZERO;
282 case VK_BLEND_FACTOR_ONE:
283 return V_028780_BLEND_ONE;
284 case VK_BLEND_FACTOR_SRC_COLOR:
285 return V_028780_BLEND_SRC_COLOR;
286 case VK_BLEND_FACTOR_ONE_MINUS_SRC_COLOR:
287 return V_028780_BLEND_ONE_MINUS_SRC_COLOR;
288 case VK_BLEND_FACTOR_DST_COLOR:
289 return V_028780_BLEND_DST_COLOR;
290 case VK_BLEND_FACTOR_ONE_MINUS_DST_COLOR:
291 return V_028780_BLEND_ONE_MINUS_DST_COLOR;
292 case VK_BLEND_FACTOR_SRC_ALPHA:
293 return V_028780_BLEND_SRC_ALPHA;
294 case VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA:
295 return V_028780_BLEND_ONE_MINUS_SRC_ALPHA;
296 case VK_BLEND_FACTOR_DST_ALPHA:
297 return V_028780_BLEND_DST_ALPHA;
298 case VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA:
299 return V_028780_BLEND_ONE_MINUS_DST_ALPHA;
300 case VK_BLEND_FACTOR_CONSTANT_COLOR:
301 return V_028780_BLEND_CONSTANT_COLOR;
302 case VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_COLOR:
303 return V_028780_BLEND_ONE_MINUS_CONSTANT_COLOR;
304 case VK_BLEND_FACTOR_CONSTANT_ALPHA:
305 return V_028780_BLEND_CONSTANT_ALPHA;
306 case VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_ALPHA:
307 return V_028780_BLEND_ONE_MINUS_CONSTANT_ALPHA;
308 case VK_BLEND_FACTOR_SRC_ALPHA_SATURATE:
309 return V_028780_BLEND_SRC_ALPHA_SATURATE;
310 case VK_BLEND_FACTOR_SRC1_COLOR:
311 return V_028780_BLEND_SRC1_COLOR;
312 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_COLOR:
313 return V_028780_BLEND_INV_SRC1_COLOR;
314 case VK_BLEND_FACTOR_SRC1_ALPHA:
315 return V_028780_BLEND_SRC1_ALPHA;
316 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA:
317 return V_028780_BLEND_INV_SRC1_ALPHA;
318 default:
319 return 0;
320 }
321 }
322
323 static uint32_t si_translate_blend_opt_function(VkBlendOp op)
324 {
325 switch (op) {
326 case VK_BLEND_OP_ADD:
327 return V_028760_OPT_COMB_ADD;
328 case VK_BLEND_OP_SUBTRACT:
329 return V_028760_OPT_COMB_SUBTRACT;
330 case VK_BLEND_OP_REVERSE_SUBTRACT:
331 return V_028760_OPT_COMB_REVSUBTRACT;
332 case VK_BLEND_OP_MIN:
333 return V_028760_OPT_COMB_MIN;
334 case VK_BLEND_OP_MAX:
335 return V_028760_OPT_COMB_MAX;
336 default:
337 return V_028760_OPT_COMB_BLEND_DISABLED;
338 }
339 }
340
341 static uint32_t si_translate_blend_opt_factor(VkBlendFactor factor, bool is_alpha)
342 {
343 switch (factor) {
344 case VK_BLEND_FACTOR_ZERO:
345 return V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_ALL;
346 case VK_BLEND_FACTOR_ONE:
347 return V_028760_BLEND_OPT_PRESERVE_ALL_IGNORE_NONE;
348 case VK_BLEND_FACTOR_SRC_COLOR:
349 return is_alpha ? V_028760_BLEND_OPT_PRESERVE_A1_IGNORE_A0
350 : V_028760_BLEND_OPT_PRESERVE_C1_IGNORE_C0;
351 case VK_BLEND_FACTOR_ONE_MINUS_SRC_COLOR:
352 return is_alpha ? V_028760_BLEND_OPT_PRESERVE_A0_IGNORE_A1
353 : V_028760_BLEND_OPT_PRESERVE_C0_IGNORE_C1;
354 case VK_BLEND_FACTOR_SRC_ALPHA:
355 return V_028760_BLEND_OPT_PRESERVE_A1_IGNORE_A0;
356 case VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA:
357 return V_028760_BLEND_OPT_PRESERVE_A0_IGNORE_A1;
358 case VK_BLEND_FACTOR_SRC_ALPHA_SATURATE:
359 return is_alpha ? V_028760_BLEND_OPT_PRESERVE_ALL_IGNORE_NONE
360 : V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_A0;
361 default:
362 return V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE;
363 }
364 }
365
366 /**
367 * Get rid of DST in the blend factors by commuting the operands:
368 * func(src * DST, dst * 0) ---> func(src * 0, dst * SRC)
369 */
370 static void si_blend_remove_dst(unsigned *func, unsigned *src_factor,
371 unsigned *dst_factor, unsigned expected_dst,
372 unsigned replacement_src)
373 {
374 if (*src_factor == expected_dst &&
375 *dst_factor == VK_BLEND_FACTOR_ZERO) {
376 *src_factor = VK_BLEND_FACTOR_ZERO;
377 *dst_factor = replacement_src;
378
379 /* Commuting the operands requires reversing subtractions. */
380 if (*func == VK_BLEND_OP_SUBTRACT)
381 *func = VK_BLEND_OP_REVERSE_SUBTRACT;
382 else if (*func == VK_BLEND_OP_REVERSE_SUBTRACT)
383 *func = VK_BLEND_OP_SUBTRACT;
384 }
385 }
386
387 static bool si_blend_factor_uses_dst(unsigned factor)
388 {
389 return factor == VK_BLEND_FACTOR_DST_COLOR ||
390 factor == VK_BLEND_FACTOR_DST_ALPHA ||
391 factor == VK_BLEND_FACTOR_SRC_ALPHA_SATURATE ||
392 factor == VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA ||
393 factor == VK_BLEND_FACTOR_ONE_MINUS_DST_COLOR;
394 }
395
396 static bool is_dual_src(VkBlendFactor factor)
397 {
398 switch (factor) {
399 case VK_BLEND_FACTOR_SRC1_COLOR:
400 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_COLOR:
401 case VK_BLEND_FACTOR_SRC1_ALPHA:
402 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA:
403 return true;
404 default:
405 return false;
406 }
407 }
408
409 static unsigned si_choose_spi_color_format(VkFormat vk_format,
410 bool blend_enable,
411 bool blend_need_alpha)
412 {
413 const struct vk_format_description *desc = vk_format_description(vk_format);
414 unsigned format, ntype, swap;
415
416 /* Alpha is needed for alpha-to-coverage.
417 * Blending may be with or without alpha.
418 */
419 unsigned normal = 0; /* most optimal, may not support blending or export alpha */
420 unsigned alpha = 0; /* exports alpha, but may not support blending */
421 unsigned blend = 0; /* supports blending, but may not export alpha */
422 unsigned blend_alpha = 0; /* least optimal, supports blending and exports alpha */
423
424 format = radv_translate_colorformat(vk_format);
425 ntype = radv_translate_color_numformat(vk_format, desc,
426 vk_format_get_first_non_void_channel(vk_format));
427 swap = radv_translate_colorswap(vk_format, false);
428
429 /* Choose the SPI color formats. These are required values for Stoney/RB+.
430 * Other chips have multiple choices, though they are not necessarily better.
431 */
432 switch (format) {
433 case V_028C70_COLOR_5_6_5:
434 case V_028C70_COLOR_1_5_5_5:
435 case V_028C70_COLOR_5_5_5_1:
436 case V_028C70_COLOR_4_4_4_4:
437 case V_028C70_COLOR_10_11_11:
438 case V_028C70_COLOR_11_11_10:
439 case V_028C70_COLOR_8:
440 case V_028C70_COLOR_8_8:
441 case V_028C70_COLOR_8_8_8_8:
442 case V_028C70_COLOR_10_10_10_2:
443 case V_028C70_COLOR_2_10_10_10:
444 if (ntype == V_028C70_NUMBER_UINT)
445 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_UINT16_ABGR;
446 else if (ntype == V_028C70_NUMBER_SINT)
447 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_SINT16_ABGR;
448 else
449 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_FP16_ABGR;
450 break;
451
452 case V_028C70_COLOR_16:
453 case V_028C70_COLOR_16_16:
454 case V_028C70_COLOR_16_16_16_16:
455 if (ntype == V_028C70_NUMBER_UNORM ||
456 ntype == V_028C70_NUMBER_SNORM) {
457 /* UNORM16 and SNORM16 don't support blending */
458 if (ntype == V_028C70_NUMBER_UNORM)
459 normal = alpha = V_028714_SPI_SHADER_UNORM16_ABGR;
460 else
461 normal = alpha = V_028714_SPI_SHADER_SNORM16_ABGR;
462
463 /* Use 32 bits per channel for blending. */
464 if (format == V_028C70_COLOR_16) {
465 if (swap == V_028C70_SWAP_STD) { /* R */
466 blend = V_028714_SPI_SHADER_32_R;
467 blend_alpha = V_028714_SPI_SHADER_32_AR;
468 } else if (swap == V_028C70_SWAP_ALT_REV) /* A */
469 blend = blend_alpha = V_028714_SPI_SHADER_32_AR;
470 else
471 assert(0);
472 } else if (format == V_028C70_COLOR_16_16) {
473 if (swap == V_028C70_SWAP_STD) { /* RG */
474 blend = V_028714_SPI_SHADER_32_GR;
475 blend_alpha = V_028714_SPI_SHADER_32_ABGR;
476 } else if (swap == V_028C70_SWAP_ALT) /* RA */
477 blend = blend_alpha = V_028714_SPI_SHADER_32_AR;
478 else
479 assert(0);
480 } else /* 16_16_16_16 */
481 blend = blend_alpha = V_028714_SPI_SHADER_32_ABGR;
482 } else if (ntype == V_028C70_NUMBER_UINT)
483 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_UINT16_ABGR;
484 else if (ntype == V_028C70_NUMBER_SINT)
485 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_SINT16_ABGR;
486 else if (ntype == V_028C70_NUMBER_FLOAT)
487 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_FP16_ABGR;
488 else
489 assert(0);
490 break;
491
492 case V_028C70_COLOR_32:
493 if (swap == V_028C70_SWAP_STD) { /* R */
494 blend = normal = V_028714_SPI_SHADER_32_R;
495 alpha = blend_alpha = V_028714_SPI_SHADER_32_AR;
496 } else if (swap == V_028C70_SWAP_ALT_REV) /* A */
497 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_32_AR;
498 else
499 assert(0);
500 break;
501
502 case V_028C70_COLOR_32_32:
503 if (swap == V_028C70_SWAP_STD) { /* RG */
504 blend = normal = V_028714_SPI_SHADER_32_GR;
505 alpha = blend_alpha = V_028714_SPI_SHADER_32_ABGR;
506 } else if (swap == V_028C70_SWAP_ALT) /* RA */
507 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_32_AR;
508 else
509 assert(0);
510 break;
511
512 case V_028C70_COLOR_32_32_32_32:
513 case V_028C70_COLOR_8_24:
514 case V_028C70_COLOR_24_8:
515 case V_028C70_COLOR_X24_8_32_FLOAT:
516 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_32_ABGR;
517 break;
518
519 default:
520 unreachable("unhandled blend format");
521 }
522
523 if (blend_enable && blend_need_alpha)
524 return blend_alpha;
525 else if(blend_need_alpha)
526 return alpha;
527 else if(blend_enable)
528 return blend;
529 else
530 return normal;
531 }
532
533 static void
534 radv_pipeline_compute_spi_color_formats(struct radv_pipeline *pipeline,
535 const VkGraphicsPipelineCreateInfo *pCreateInfo,
536 struct radv_blend_state *blend)
537 {
538 RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass);
539 struct radv_subpass *subpass = pass->subpasses + pCreateInfo->subpass;
540 unsigned col_format = 0;
541 unsigned num_targets;
542
543 for (unsigned i = 0; i < (blend->single_cb_enable ? 1 : subpass->color_count); ++i) {
544 unsigned cf;
545
546 if (subpass->color_attachments[i].attachment == VK_ATTACHMENT_UNUSED) {
547 cf = V_028714_SPI_SHADER_ZERO;
548 } else {
549 struct radv_render_pass_attachment *attachment = pass->attachments + subpass->color_attachments[i].attachment;
550 bool blend_enable =
551 blend->blend_enable_4bit & (0xfu << (i * 4));
552
553 cf = si_choose_spi_color_format(attachment->format,
554 blend_enable,
555 blend->need_src_alpha & (1 << i));
556 }
557
558 col_format |= cf << (4 * i);
559 }
560
561 if (!(col_format & 0xf) && blend->need_src_alpha & (1 << 0)) {
562 /* When a subpass doesn't have any color attachments, write the
563 * alpha channel of MRT0 when alpha coverage is enabled because
564 * the depth attachment needs it.
565 */
566 col_format |= V_028714_SPI_SHADER_32_AR;
567 }
568
569 /* If the i-th target format is set, all previous target formats must
570 * be non-zero to avoid hangs.
571 */
572 num_targets = (util_last_bit(col_format) + 3) / 4;
573 for (unsigned i = 0; i < num_targets; i++) {
574 if (!(col_format & (0xf << (i * 4)))) {
575 col_format |= V_028714_SPI_SHADER_32_R << (i * 4);
576 }
577 }
578
579 /* The output for dual source blending should have the same format as
580 * the first output.
581 */
582 if (blend->mrt0_is_dual_src)
583 col_format |= (col_format & 0xf) << 4;
584
585 blend->cb_shader_mask = ac_get_cb_shader_mask(col_format);
586 blend->spi_shader_col_format = col_format;
587 }
588
589 static bool
590 format_is_int8(VkFormat format)
591 {
592 const struct vk_format_description *desc = vk_format_description(format);
593 int channel = vk_format_get_first_non_void_channel(format);
594
595 return channel >= 0 && desc->channel[channel].pure_integer &&
596 desc->channel[channel].size == 8;
597 }
598
599 static bool
600 format_is_int10(VkFormat format)
601 {
602 const struct vk_format_description *desc = vk_format_description(format);
603
604 if (desc->nr_channels != 4)
605 return false;
606 for (unsigned i = 0; i < 4; i++) {
607 if (desc->channel[i].pure_integer && desc->channel[i].size == 10)
608 return true;
609 }
610 return false;
611 }
612
613 /*
614 * Ordered so that for each i,
615 * radv_format_meta_fs_key(radv_fs_key_format_exemplars[i]) == i.
616 */
617 const VkFormat radv_fs_key_format_exemplars[NUM_META_FS_KEYS] = {
618 VK_FORMAT_R32_SFLOAT,
619 VK_FORMAT_R32G32_SFLOAT,
620 VK_FORMAT_R8G8B8A8_UNORM,
621 VK_FORMAT_R16G16B16A16_UNORM,
622 VK_FORMAT_R16G16B16A16_SNORM,
623 VK_FORMAT_R16G16B16A16_UINT,
624 VK_FORMAT_R16G16B16A16_SINT,
625 VK_FORMAT_R32G32B32A32_SFLOAT,
626 VK_FORMAT_R8G8B8A8_UINT,
627 VK_FORMAT_R8G8B8A8_SINT,
628 VK_FORMAT_A2R10G10B10_UINT_PACK32,
629 VK_FORMAT_A2R10G10B10_SINT_PACK32,
630 };
631
632 unsigned radv_format_meta_fs_key(VkFormat format)
633 {
634 unsigned col_format = si_choose_spi_color_format(format, false, false);
635
636 assert(col_format != V_028714_SPI_SHADER_32_AR);
637 if (col_format >= V_028714_SPI_SHADER_32_AR)
638 --col_format; /* Skip V_028714_SPI_SHADER_32_AR since there is no such VkFormat */
639
640 --col_format; /* Skip V_028714_SPI_SHADER_ZERO */
641 bool is_int8 = format_is_int8(format);
642 bool is_int10 = format_is_int10(format);
643
644 return col_format + (is_int8 ? 3 : is_int10 ? 5 : 0);
645 }
646
647 static void
648 radv_pipeline_compute_get_int_clamp(const VkGraphicsPipelineCreateInfo *pCreateInfo,
649 unsigned *is_int8, unsigned *is_int10)
650 {
651 RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass);
652 struct radv_subpass *subpass = pass->subpasses + pCreateInfo->subpass;
653 *is_int8 = 0;
654 *is_int10 = 0;
655
656 for (unsigned i = 0; i < subpass->color_count; ++i) {
657 struct radv_render_pass_attachment *attachment;
658
659 if (subpass->color_attachments[i].attachment == VK_ATTACHMENT_UNUSED)
660 continue;
661
662 attachment = pass->attachments + subpass->color_attachments[i].attachment;
663
664 if (format_is_int8(attachment->format))
665 *is_int8 |= 1 << i;
666 if (format_is_int10(attachment->format))
667 *is_int10 |= 1 << i;
668 }
669 }
670
671 static void
672 radv_blend_check_commutativity(struct radv_blend_state *blend,
673 VkBlendOp op, VkBlendFactor src,
674 VkBlendFactor dst, unsigned chanmask)
675 {
676 /* Src factor is allowed when it does not depend on Dst. */
677 static const uint32_t src_allowed =
678 (1u << VK_BLEND_FACTOR_ONE) |
679 (1u << VK_BLEND_FACTOR_SRC_COLOR) |
680 (1u << VK_BLEND_FACTOR_SRC_ALPHA) |
681 (1u << VK_BLEND_FACTOR_SRC_ALPHA_SATURATE) |
682 (1u << VK_BLEND_FACTOR_CONSTANT_COLOR) |
683 (1u << VK_BLEND_FACTOR_CONSTANT_ALPHA) |
684 (1u << VK_BLEND_FACTOR_SRC1_COLOR) |
685 (1u << VK_BLEND_FACTOR_SRC1_ALPHA) |
686 (1u << VK_BLEND_FACTOR_ZERO) |
687 (1u << VK_BLEND_FACTOR_ONE_MINUS_SRC_COLOR) |
688 (1u << VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA) |
689 (1u << VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_COLOR) |
690 (1u << VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_ALPHA) |
691 (1u << VK_BLEND_FACTOR_ONE_MINUS_SRC1_COLOR) |
692 (1u << VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA);
693
694 if (dst == VK_BLEND_FACTOR_ONE &&
695 (src_allowed & (1u << src))) {
696 /* Addition is commutative, but floating point addition isn't
697 * associative: subtle changes can be introduced via different
698 * rounding. Be conservative, only enable for min and max.
699 */
700 if (op == VK_BLEND_OP_MAX || op == VK_BLEND_OP_MIN)
701 blend->commutative_4bit |= chanmask;
702 }
703 }
704
705 static struct radv_blend_state
706 radv_pipeline_init_blend_state(struct radv_pipeline *pipeline,
707 const VkGraphicsPipelineCreateInfo *pCreateInfo,
708 const struct radv_graphics_pipeline_create_info *extra)
709 {
710 const VkPipelineColorBlendStateCreateInfo *vkblend = pCreateInfo->pColorBlendState;
711 const VkPipelineMultisampleStateCreateInfo *vkms = pCreateInfo->pMultisampleState;
712 struct radv_blend_state blend = {0};
713 unsigned mode = V_028808_CB_NORMAL;
714 int i;
715
716 if (!vkblend)
717 return blend;
718
719 if (extra && extra->custom_blend_mode) {
720 blend.single_cb_enable = true;
721 mode = extra->custom_blend_mode;
722 }
723 blend.cb_color_control = 0;
724 if (vkblend->logicOpEnable)
725 blend.cb_color_control |= S_028808_ROP3(si_translate_blend_logic_op(vkblend->logicOp));
726 else
727 blend.cb_color_control |= S_028808_ROP3(V_028808_ROP3_COPY);
728
729 blend.db_alpha_to_mask = S_028B70_ALPHA_TO_MASK_OFFSET0(3) |
730 S_028B70_ALPHA_TO_MASK_OFFSET1(1) |
731 S_028B70_ALPHA_TO_MASK_OFFSET2(0) |
732 S_028B70_ALPHA_TO_MASK_OFFSET3(2) |
733 S_028B70_OFFSET_ROUND(1);
734
735 if (vkms && vkms->alphaToCoverageEnable) {
736 blend.db_alpha_to_mask |= S_028B70_ALPHA_TO_MASK_ENABLE(1);
737 blend.need_src_alpha |= 0x1;
738 }
739
740 blend.cb_target_mask = 0;
741 for (i = 0; i < vkblend->attachmentCount; i++) {
742 const VkPipelineColorBlendAttachmentState *att = &vkblend->pAttachments[i];
743 unsigned blend_cntl = 0;
744 unsigned srcRGB_opt, dstRGB_opt, srcA_opt, dstA_opt;
745 VkBlendOp eqRGB = att->colorBlendOp;
746 VkBlendFactor srcRGB = att->srcColorBlendFactor;
747 VkBlendFactor dstRGB = att->dstColorBlendFactor;
748 VkBlendOp eqA = att->alphaBlendOp;
749 VkBlendFactor srcA = att->srcAlphaBlendFactor;
750 VkBlendFactor dstA = att->dstAlphaBlendFactor;
751
752 blend.sx_mrt_blend_opt[i] = S_028760_COLOR_COMB_FCN(V_028760_OPT_COMB_BLEND_DISABLED) | S_028760_ALPHA_COMB_FCN(V_028760_OPT_COMB_BLEND_DISABLED);
753
754 if (!att->colorWriteMask)
755 continue;
756
757 blend.cb_target_mask |= (unsigned)att->colorWriteMask << (4 * i);
758 blend.cb_target_enabled_4bit |= 0xf << (4 * i);
759 if (!att->blendEnable) {
760 blend.cb_blend_control[i] = blend_cntl;
761 continue;
762 }
763
764 if (is_dual_src(srcRGB) || is_dual_src(dstRGB) || is_dual_src(srcA) || is_dual_src(dstA))
765 if (i == 0)
766 blend.mrt0_is_dual_src = true;
767
768 if (eqRGB == VK_BLEND_OP_MIN || eqRGB == VK_BLEND_OP_MAX) {
769 srcRGB = VK_BLEND_FACTOR_ONE;
770 dstRGB = VK_BLEND_FACTOR_ONE;
771 }
772 if (eqA == VK_BLEND_OP_MIN || eqA == VK_BLEND_OP_MAX) {
773 srcA = VK_BLEND_FACTOR_ONE;
774 dstA = VK_BLEND_FACTOR_ONE;
775 }
776
777 radv_blend_check_commutativity(&blend, eqRGB, srcRGB, dstRGB,
778 0x7 << (4 * i));
779 radv_blend_check_commutativity(&blend, eqA, srcA, dstA,
780 0x8 << (4 * i));
781
782 /* Blending optimizations for RB+.
783 * These transformations don't change the behavior.
784 *
785 * First, get rid of DST in the blend factors:
786 * func(src * DST, dst * 0) ---> func(src * 0, dst * SRC)
787 */
788 si_blend_remove_dst(&eqRGB, &srcRGB, &dstRGB,
789 VK_BLEND_FACTOR_DST_COLOR,
790 VK_BLEND_FACTOR_SRC_COLOR);
791
792 si_blend_remove_dst(&eqA, &srcA, &dstA,
793 VK_BLEND_FACTOR_DST_COLOR,
794 VK_BLEND_FACTOR_SRC_COLOR);
795
796 si_blend_remove_dst(&eqA, &srcA, &dstA,
797 VK_BLEND_FACTOR_DST_ALPHA,
798 VK_BLEND_FACTOR_SRC_ALPHA);
799
800 /* Look up the ideal settings from tables. */
801 srcRGB_opt = si_translate_blend_opt_factor(srcRGB, false);
802 dstRGB_opt = si_translate_blend_opt_factor(dstRGB, false);
803 srcA_opt = si_translate_blend_opt_factor(srcA, true);
804 dstA_opt = si_translate_blend_opt_factor(dstA, true);
805
806 /* Handle interdependencies. */
807 if (si_blend_factor_uses_dst(srcRGB))
808 dstRGB_opt = V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE;
809 if (si_blend_factor_uses_dst(srcA))
810 dstA_opt = V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE;
811
812 if (srcRGB == VK_BLEND_FACTOR_SRC_ALPHA_SATURATE &&
813 (dstRGB == VK_BLEND_FACTOR_ZERO ||
814 dstRGB == VK_BLEND_FACTOR_SRC_ALPHA ||
815 dstRGB == VK_BLEND_FACTOR_SRC_ALPHA_SATURATE))
816 dstRGB_opt = V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_A0;
817
818 /* Set the final value. */
819 blend.sx_mrt_blend_opt[i] =
820 S_028760_COLOR_SRC_OPT(srcRGB_opt) |
821 S_028760_COLOR_DST_OPT(dstRGB_opt) |
822 S_028760_COLOR_COMB_FCN(si_translate_blend_opt_function(eqRGB)) |
823 S_028760_ALPHA_SRC_OPT(srcA_opt) |
824 S_028760_ALPHA_DST_OPT(dstA_opt) |
825 S_028760_ALPHA_COMB_FCN(si_translate_blend_opt_function(eqA));
826 blend_cntl |= S_028780_ENABLE(1);
827
828 blend_cntl |= S_028780_COLOR_COMB_FCN(si_translate_blend_function(eqRGB));
829 blend_cntl |= S_028780_COLOR_SRCBLEND(si_translate_blend_factor(srcRGB));
830 blend_cntl |= S_028780_COLOR_DESTBLEND(si_translate_blend_factor(dstRGB));
831 if (srcA != srcRGB || dstA != dstRGB || eqA != eqRGB) {
832 blend_cntl |= S_028780_SEPARATE_ALPHA_BLEND(1);
833 blend_cntl |= S_028780_ALPHA_COMB_FCN(si_translate_blend_function(eqA));
834 blend_cntl |= S_028780_ALPHA_SRCBLEND(si_translate_blend_factor(srcA));
835 blend_cntl |= S_028780_ALPHA_DESTBLEND(si_translate_blend_factor(dstA));
836 }
837 blend.cb_blend_control[i] = blend_cntl;
838
839 blend.blend_enable_4bit |= 0xfu << (i * 4);
840
841 if (srcRGB == VK_BLEND_FACTOR_SRC_ALPHA ||
842 dstRGB == VK_BLEND_FACTOR_SRC_ALPHA ||
843 srcRGB == VK_BLEND_FACTOR_SRC_ALPHA_SATURATE ||
844 dstRGB == VK_BLEND_FACTOR_SRC_ALPHA_SATURATE ||
845 srcRGB == VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA ||
846 dstRGB == VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA)
847 blend.need_src_alpha |= 1 << i;
848 }
849 for (i = vkblend->attachmentCount; i < 8; i++) {
850 blend.cb_blend_control[i] = 0;
851 blend.sx_mrt_blend_opt[i] = S_028760_COLOR_COMB_FCN(V_028760_OPT_COMB_BLEND_DISABLED) | S_028760_ALPHA_COMB_FCN(V_028760_OPT_COMB_BLEND_DISABLED);
852 }
853
854 if (pipeline->device->physical_device->rad_info.has_rbplus) {
855 /* Disable RB+ blend optimizations for dual source blending. */
856 if (blend.mrt0_is_dual_src) {
857 for (i = 0; i < 8; i++) {
858 blend.sx_mrt_blend_opt[i] =
859 S_028760_COLOR_COMB_FCN(V_028760_OPT_COMB_NONE) |
860 S_028760_ALPHA_COMB_FCN(V_028760_OPT_COMB_NONE);
861 }
862 }
863
864 /* RB+ doesn't work with dual source blending, logic op and
865 * RESOLVE.
866 */
867 if (blend.mrt0_is_dual_src || vkblend->logicOpEnable ||
868 mode == V_028808_CB_RESOLVE)
869 blend.cb_color_control |= S_028808_DISABLE_DUAL_QUAD(1);
870 }
871
872 if (blend.cb_target_mask)
873 blend.cb_color_control |= S_028808_MODE(mode);
874 else
875 blend.cb_color_control |= S_028808_MODE(V_028808_CB_DISABLE);
876
877 radv_pipeline_compute_spi_color_formats(pipeline, pCreateInfo, &blend);
878 return blend;
879 }
880
881 static uint32_t si_translate_stencil_op(enum VkStencilOp op)
882 {
883 switch (op) {
884 case VK_STENCIL_OP_KEEP:
885 return V_02842C_STENCIL_KEEP;
886 case VK_STENCIL_OP_ZERO:
887 return V_02842C_STENCIL_ZERO;
888 case VK_STENCIL_OP_REPLACE:
889 return V_02842C_STENCIL_REPLACE_TEST;
890 case VK_STENCIL_OP_INCREMENT_AND_CLAMP:
891 return V_02842C_STENCIL_ADD_CLAMP;
892 case VK_STENCIL_OP_DECREMENT_AND_CLAMP:
893 return V_02842C_STENCIL_SUB_CLAMP;
894 case VK_STENCIL_OP_INVERT:
895 return V_02842C_STENCIL_INVERT;
896 case VK_STENCIL_OP_INCREMENT_AND_WRAP:
897 return V_02842C_STENCIL_ADD_WRAP;
898 case VK_STENCIL_OP_DECREMENT_AND_WRAP:
899 return V_02842C_STENCIL_SUB_WRAP;
900 default:
901 return 0;
902 }
903 }
904
905 static uint32_t si_translate_fill(VkPolygonMode func)
906 {
907 switch(func) {
908 case VK_POLYGON_MODE_FILL:
909 return V_028814_X_DRAW_TRIANGLES;
910 case VK_POLYGON_MODE_LINE:
911 return V_028814_X_DRAW_LINES;
912 case VK_POLYGON_MODE_POINT:
913 return V_028814_X_DRAW_POINTS;
914 default:
915 assert(0);
916 return V_028814_X_DRAW_POINTS;
917 }
918 }
919
920 static uint8_t radv_pipeline_get_ps_iter_samples(const VkPipelineMultisampleStateCreateInfo *vkms)
921 {
922 uint32_t num_samples = vkms->rasterizationSamples;
923 uint32_t ps_iter_samples = 1;
924
925 if (vkms->sampleShadingEnable) {
926 ps_iter_samples = ceil(vkms->minSampleShading * num_samples);
927 ps_iter_samples = util_next_power_of_two(ps_iter_samples);
928 }
929 return ps_iter_samples;
930 }
931
932 static bool
933 radv_is_depth_write_enabled(const VkPipelineDepthStencilStateCreateInfo *pCreateInfo)
934 {
935 return pCreateInfo->depthTestEnable &&
936 pCreateInfo->depthWriteEnable &&
937 pCreateInfo->depthCompareOp != VK_COMPARE_OP_NEVER;
938 }
939
940 static bool
941 radv_writes_stencil(const VkStencilOpState *state)
942 {
943 return state->writeMask &&
944 (state->failOp != VK_STENCIL_OP_KEEP ||
945 state->passOp != VK_STENCIL_OP_KEEP ||
946 state->depthFailOp != VK_STENCIL_OP_KEEP);
947 }
948
949 static bool
950 radv_is_stencil_write_enabled(const VkPipelineDepthStencilStateCreateInfo *pCreateInfo)
951 {
952 return pCreateInfo->stencilTestEnable &&
953 (radv_writes_stencil(&pCreateInfo->front) ||
954 radv_writes_stencil(&pCreateInfo->back));
955 }
956
957 static bool
958 radv_is_ds_write_enabled(const VkPipelineDepthStencilStateCreateInfo *pCreateInfo)
959 {
960 return radv_is_depth_write_enabled(pCreateInfo) ||
961 radv_is_stencil_write_enabled(pCreateInfo);
962 }
963
964 static bool
965 radv_order_invariant_stencil_op(VkStencilOp op)
966 {
967 /* REPLACE is normally order invariant, except when the stencil
968 * reference value is written by the fragment shader. Tracking this
969 * interaction does not seem worth the effort, so be conservative.
970 */
971 return op != VK_STENCIL_OP_INCREMENT_AND_CLAMP &&
972 op != VK_STENCIL_OP_DECREMENT_AND_CLAMP &&
973 op != VK_STENCIL_OP_REPLACE;
974 }
975
976 static bool
977 radv_order_invariant_stencil_state(const VkStencilOpState *state)
978 {
979 /* Compute whether, assuming Z writes are disabled, this stencil state
980 * is order invariant in the sense that the set of passing fragments as
981 * well as the final stencil buffer result does not depend on the order
982 * of fragments.
983 */
984 return !state->writeMask ||
985 /* The following assumes that Z writes are disabled. */
986 (state->compareOp == VK_COMPARE_OP_ALWAYS &&
987 radv_order_invariant_stencil_op(state->passOp) &&
988 radv_order_invariant_stencil_op(state->depthFailOp)) ||
989 (state->compareOp == VK_COMPARE_OP_NEVER &&
990 radv_order_invariant_stencil_op(state->failOp));
991 }
992
993 static bool
994 radv_pipeline_out_of_order_rast(struct radv_pipeline *pipeline,
995 struct radv_blend_state *blend,
996 const VkGraphicsPipelineCreateInfo *pCreateInfo)
997 {
998 RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass);
999 struct radv_subpass *subpass = pass->subpasses + pCreateInfo->subpass;
1000 unsigned colormask = blend->cb_target_enabled_4bit;
1001
1002 if (!pipeline->device->physical_device->out_of_order_rast_allowed)
1003 return false;
1004
1005 /* Be conservative if a logic operation is enabled with color buffers. */
1006 if (colormask && pCreateInfo->pColorBlendState->logicOpEnable)
1007 return false;
1008
1009 /* Default depth/stencil invariance when no attachment is bound. */
1010 struct radv_dsa_order_invariance dsa_order_invariant = {
1011 .zs = true, .pass_set = true
1012 };
1013
1014 if (pCreateInfo->pDepthStencilState &&
1015 subpass->depth_stencil_attachment) {
1016 const VkPipelineDepthStencilStateCreateInfo *vkds =
1017 pCreateInfo->pDepthStencilState;
1018 struct radv_render_pass_attachment *attachment =
1019 pass->attachments + subpass->depth_stencil_attachment->attachment;
1020 bool has_stencil = vk_format_is_stencil(attachment->format);
1021 struct radv_dsa_order_invariance order_invariance[2];
1022 struct radv_shader_variant *ps =
1023 pipeline->shaders[MESA_SHADER_FRAGMENT];
1024
1025 /* Compute depth/stencil order invariance in order to know if
1026 * it's safe to enable out-of-order.
1027 */
1028 bool zfunc_is_ordered =
1029 vkds->depthCompareOp == VK_COMPARE_OP_NEVER ||
1030 vkds->depthCompareOp == VK_COMPARE_OP_LESS ||
1031 vkds->depthCompareOp == VK_COMPARE_OP_LESS_OR_EQUAL ||
1032 vkds->depthCompareOp == VK_COMPARE_OP_GREATER ||
1033 vkds->depthCompareOp == VK_COMPARE_OP_GREATER_OR_EQUAL;
1034
1035 bool nozwrite_and_order_invariant_stencil =
1036 !radv_is_ds_write_enabled(vkds) ||
1037 (!radv_is_depth_write_enabled(vkds) &&
1038 radv_order_invariant_stencil_state(&vkds->front) &&
1039 radv_order_invariant_stencil_state(&vkds->back));
1040
1041 order_invariance[1].zs =
1042 nozwrite_and_order_invariant_stencil ||
1043 (!radv_is_stencil_write_enabled(vkds) &&
1044 zfunc_is_ordered);
1045 order_invariance[0].zs =
1046 !radv_is_depth_write_enabled(vkds) || zfunc_is_ordered;
1047
1048 order_invariance[1].pass_set =
1049 nozwrite_and_order_invariant_stencil ||
1050 (!radv_is_stencil_write_enabled(vkds) &&
1051 (vkds->depthCompareOp == VK_COMPARE_OP_ALWAYS ||
1052 vkds->depthCompareOp == VK_COMPARE_OP_NEVER));
1053 order_invariance[0].pass_set =
1054 !radv_is_depth_write_enabled(vkds) ||
1055 (vkds->depthCompareOp == VK_COMPARE_OP_ALWAYS ||
1056 vkds->depthCompareOp == VK_COMPARE_OP_NEVER);
1057
1058 dsa_order_invariant = order_invariance[has_stencil];
1059 if (!dsa_order_invariant.zs)
1060 return false;
1061
1062 /* The set of PS invocations is always order invariant,
1063 * except when early Z/S tests are requested.
1064 */
1065 if (ps &&
1066 ps->info.ps.writes_memory &&
1067 ps->info.ps.early_fragment_test &&
1068 !dsa_order_invariant.pass_set)
1069 return false;
1070
1071 /* Determine if out-of-order rasterization should be disabled
1072 * when occlusion queries are used.
1073 */
1074 pipeline->graphics.disable_out_of_order_rast_for_occlusion =
1075 !dsa_order_invariant.pass_set;
1076 }
1077
1078 /* No color buffers are enabled for writing. */
1079 if (!colormask)
1080 return true;
1081
1082 unsigned blendmask = colormask & blend->blend_enable_4bit;
1083
1084 if (blendmask) {
1085 /* Only commutative blending. */
1086 if (blendmask & ~blend->commutative_4bit)
1087 return false;
1088
1089 if (!dsa_order_invariant.pass_set)
1090 return false;
1091 }
1092
1093 if (colormask & ~blendmask)
1094 return false;
1095
1096 return true;
1097 }
1098
1099 static void
1100 radv_pipeline_init_multisample_state(struct radv_pipeline *pipeline,
1101 struct radv_blend_state *blend,
1102 const VkGraphicsPipelineCreateInfo *pCreateInfo)
1103 {
1104 const VkPipelineMultisampleStateCreateInfo *vkms = pCreateInfo->pMultisampleState;
1105 struct radv_multisample_state *ms = &pipeline->graphics.ms;
1106 unsigned num_tile_pipes = pipeline->device->physical_device->rad_info.num_tile_pipes;
1107 bool out_of_order_rast = false;
1108 int ps_iter_samples = 1;
1109 uint32_t mask = 0xffff;
1110
1111 if (vkms)
1112 ms->num_samples = vkms->rasterizationSamples;
1113 else
1114 ms->num_samples = 1;
1115
1116 if (vkms)
1117 ps_iter_samples = radv_pipeline_get_ps_iter_samples(vkms);
1118 if (vkms && !vkms->sampleShadingEnable && pipeline->shaders[MESA_SHADER_FRAGMENT]->info.ps.force_persample) {
1119 ps_iter_samples = ms->num_samples;
1120 }
1121
1122 const struct VkPipelineRasterizationStateRasterizationOrderAMD *raster_order =
1123 vk_find_struct_const(pCreateInfo->pRasterizationState->pNext, PIPELINE_RASTERIZATION_STATE_RASTERIZATION_ORDER_AMD);
1124 if (raster_order && raster_order->rasterizationOrder == VK_RASTERIZATION_ORDER_RELAXED_AMD) {
1125 /* Out-of-order rasterization is explicitly enabled by the
1126 * application.
1127 */
1128 out_of_order_rast = true;
1129 } else {
1130 /* Determine if the driver can enable out-of-order
1131 * rasterization internally.
1132 */
1133 out_of_order_rast =
1134 radv_pipeline_out_of_order_rast(pipeline, blend, pCreateInfo);
1135 }
1136
1137 ms->pa_sc_line_cntl = S_028BDC_DX10_DIAMOND_TEST_ENA(1);
1138 ms->pa_sc_aa_config = 0;
1139 ms->db_eqaa = S_028804_HIGH_QUALITY_INTERSECTIONS(1) |
1140 S_028804_INCOHERENT_EQAA_READS(1) |
1141 S_028804_INTERPOLATE_COMP_Z(1) |
1142 S_028804_STATIC_ANCHOR_ASSOCIATIONS(1);
1143 ms->pa_sc_mode_cntl_1 =
1144 S_028A4C_WALK_FENCE_ENABLE(1) | //TODO linear dst fixes
1145 S_028A4C_WALK_FENCE_SIZE(num_tile_pipes == 2 ? 2 : 3) |
1146 S_028A4C_OUT_OF_ORDER_PRIMITIVE_ENABLE(out_of_order_rast) |
1147 S_028A4C_OUT_OF_ORDER_WATER_MARK(0x7) |
1148 /* always 1: */
1149 S_028A4C_WALK_ALIGN8_PRIM_FITS_ST(1) |
1150 S_028A4C_SUPERTILE_WALK_ORDER_ENABLE(1) |
1151 S_028A4C_TILE_WALK_ORDER_ENABLE(1) |
1152 S_028A4C_MULTI_SHADER_ENGINE_PRIM_DISCARD_ENABLE(1) |
1153 S_028A4C_FORCE_EOV_CNTDWN_ENABLE(1) |
1154 S_028A4C_FORCE_EOV_REZ_ENABLE(1);
1155 ms->pa_sc_mode_cntl_0 = S_028A48_ALTERNATE_RBS_PER_TILE(pipeline->device->physical_device->rad_info.chip_class >= GFX9) |
1156 S_028A48_VPORT_SCISSOR_ENABLE(1);
1157
1158 if (ms->num_samples > 1) {
1159 unsigned log_samples = util_logbase2(ms->num_samples);
1160 unsigned log_ps_iter_samples = util_logbase2(ps_iter_samples);
1161 ms->pa_sc_mode_cntl_0 |= S_028A48_MSAA_ENABLE(1);
1162 ms->pa_sc_line_cntl |= S_028BDC_EXPAND_LINE_WIDTH(1); /* CM_R_028BDC_PA_SC_LINE_CNTL */
1163 ms->db_eqaa |= S_028804_MAX_ANCHOR_SAMPLES(log_samples) |
1164 S_028804_PS_ITER_SAMPLES(log_ps_iter_samples) |
1165 S_028804_MASK_EXPORT_NUM_SAMPLES(log_samples) |
1166 S_028804_ALPHA_TO_MASK_NUM_SAMPLES(log_samples);
1167 ms->pa_sc_aa_config |= S_028BE0_MSAA_NUM_SAMPLES(log_samples) |
1168 S_028BE0_MAX_SAMPLE_DIST(radv_get_default_max_sample_dist(log_samples)) |
1169 S_028BE0_MSAA_EXPOSED_SAMPLES(log_samples); /* CM_R_028BE0_PA_SC_AA_CONFIG */
1170 ms->pa_sc_mode_cntl_1 |= S_028A4C_PS_ITER_SAMPLE(ps_iter_samples > 1);
1171 if (ps_iter_samples > 1)
1172 pipeline->graphics.spi_baryc_cntl |= S_0286E0_POS_FLOAT_LOCATION(2);
1173 }
1174
1175 if (vkms && vkms->pSampleMask) {
1176 mask = vkms->pSampleMask[0] & 0xffff;
1177 }
1178
1179 ms->pa_sc_aa_mask[0] = mask | (mask << 16);
1180 ms->pa_sc_aa_mask[1] = mask | (mask << 16);
1181 }
1182
1183 static bool
1184 radv_prim_can_use_guardband(enum VkPrimitiveTopology topology)
1185 {
1186 switch (topology) {
1187 case VK_PRIMITIVE_TOPOLOGY_POINT_LIST:
1188 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST:
1189 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP:
1190 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY:
1191 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY:
1192 return false;
1193 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST:
1194 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP:
1195 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN:
1196 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY:
1197 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY:
1198 case VK_PRIMITIVE_TOPOLOGY_PATCH_LIST:
1199 return true;
1200 default:
1201 unreachable("unhandled primitive type");
1202 }
1203 }
1204
1205 static uint32_t
1206 si_translate_prim(enum VkPrimitiveTopology topology)
1207 {
1208 switch (topology) {
1209 case VK_PRIMITIVE_TOPOLOGY_POINT_LIST:
1210 return V_008958_DI_PT_POINTLIST;
1211 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST:
1212 return V_008958_DI_PT_LINELIST;
1213 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP:
1214 return V_008958_DI_PT_LINESTRIP;
1215 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST:
1216 return V_008958_DI_PT_TRILIST;
1217 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP:
1218 return V_008958_DI_PT_TRISTRIP;
1219 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN:
1220 return V_008958_DI_PT_TRIFAN;
1221 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY:
1222 return V_008958_DI_PT_LINELIST_ADJ;
1223 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY:
1224 return V_008958_DI_PT_LINESTRIP_ADJ;
1225 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY:
1226 return V_008958_DI_PT_TRILIST_ADJ;
1227 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY:
1228 return V_008958_DI_PT_TRISTRIP_ADJ;
1229 case VK_PRIMITIVE_TOPOLOGY_PATCH_LIST:
1230 return V_008958_DI_PT_PATCH;
1231 default:
1232 assert(0);
1233 return 0;
1234 }
1235 }
1236
1237 static uint32_t
1238 si_conv_gl_prim_to_gs_out(unsigned gl_prim)
1239 {
1240 switch (gl_prim) {
1241 case 0: /* GL_POINTS */
1242 return V_028A6C_OUTPRIM_TYPE_POINTLIST;
1243 case 1: /* GL_LINES */
1244 case 3: /* GL_LINE_STRIP */
1245 case 0xA: /* GL_LINE_STRIP_ADJACENCY_ARB */
1246 case 0x8E7A: /* GL_ISOLINES */
1247 return V_028A6C_OUTPRIM_TYPE_LINESTRIP;
1248
1249 case 4: /* GL_TRIANGLES */
1250 case 0xc: /* GL_TRIANGLES_ADJACENCY_ARB */
1251 case 5: /* GL_TRIANGLE_STRIP */
1252 case 7: /* GL_QUADS */
1253 return V_028A6C_OUTPRIM_TYPE_TRISTRIP;
1254 default:
1255 assert(0);
1256 return 0;
1257 }
1258 }
1259
1260 static uint32_t
1261 si_conv_prim_to_gs_out(enum VkPrimitiveTopology topology)
1262 {
1263 switch (topology) {
1264 case VK_PRIMITIVE_TOPOLOGY_POINT_LIST:
1265 case VK_PRIMITIVE_TOPOLOGY_PATCH_LIST:
1266 return V_028A6C_OUTPRIM_TYPE_POINTLIST;
1267 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST:
1268 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP:
1269 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY:
1270 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY:
1271 return V_028A6C_OUTPRIM_TYPE_LINESTRIP;
1272 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST:
1273 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP:
1274 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN:
1275 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY:
1276 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY:
1277 return V_028A6C_OUTPRIM_TYPE_TRISTRIP;
1278 default:
1279 assert(0);
1280 return 0;
1281 }
1282 }
1283
1284 static unsigned radv_dynamic_state_mask(VkDynamicState state)
1285 {
1286 switch(state) {
1287 case VK_DYNAMIC_STATE_VIEWPORT:
1288 return RADV_DYNAMIC_VIEWPORT;
1289 case VK_DYNAMIC_STATE_SCISSOR:
1290 return RADV_DYNAMIC_SCISSOR;
1291 case VK_DYNAMIC_STATE_LINE_WIDTH:
1292 return RADV_DYNAMIC_LINE_WIDTH;
1293 case VK_DYNAMIC_STATE_DEPTH_BIAS:
1294 return RADV_DYNAMIC_DEPTH_BIAS;
1295 case VK_DYNAMIC_STATE_BLEND_CONSTANTS:
1296 return RADV_DYNAMIC_BLEND_CONSTANTS;
1297 case VK_DYNAMIC_STATE_DEPTH_BOUNDS:
1298 return RADV_DYNAMIC_DEPTH_BOUNDS;
1299 case VK_DYNAMIC_STATE_STENCIL_COMPARE_MASK:
1300 return RADV_DYNAMIC_STENCIL_COMPARE_MASK;
1301 case VK_DYNAMIC_STATE_STENCIL_WRITE_MASK:
1302 return RADV_DYNAMIC_STENCIL_WRITE_MASK;
1303 case VK_DYNAMIC_STATE_STENCIL_REFERENCE:
1304 return RADV_DYNAMIC_STENCIL_REFERENCE;
1305 case VK_DYNAMIC_STATE_DISCARD_RECTANGLE_EXT:
1306 return RADV_DYNAMIC_DISCARD_RECTANGLE;
1307 case VK_DYNAMIC_STATE_SAMPLE_LOCATIONS_EXT:
1308 return RADV_DYNAMIC_SAMPLE_LOCATIONS;
1309 default:
1310 unreachable("Unhandled dynamic state");
1311 }
1312 }
1313
1314 static uint32_t radv_pipeline_needed_dynamic_state(const VkGraphicsPipelineCreateInfo *pCreateInfo)
1315 {
1316 uint32_t states = RADV_DYNAMIC_ALL;
1317
1318 /* If rasterization is disabled we do not care about any of the dynamic states,
1319 * since they are all rasterization related only. */
1320 if (pCreateInfo->pRasterizationState->rasterizerDiscardEnable)
1321 return 0;
1322
1323 if (!pCreateInfo->pRasterizationState->depthBiasEnable)
1324 states &= ~RADV_DYNAMIC_DEPTH_BIAS;
1325
1326 if (!pCreateInfo->pDepthStencilState ||
1327 !pCreateInfo->pDepthStencilState->depthBoundsTestEnable)
1328 states &= ~RADV_DYNAMIC_DEPTH_BOUNDS;
1329
1330 if (!pCreateInfo->pDepthStencilState ||
1331 !pCreateInfo->pDepthStencilState->stencilTestEnable)
1332 states &= ~(RADV_DYNAMIC_STENCIL_COMPARE_MASK |
1333 RADV_DYNAMIC_STENCIL_WRITE_MASK |
1334 RADV_DYNAMIC_STENCIL_REFERENCE);
1335
1336 if (!vk_find_struct_const(pCreateInfo->pNext, PIPELINE_DISCARD_RECTANGLE_STATE_CREATE_INFO_EXT))
1337 states &= ~RADV_DYNAMIC_DISCARD_RECTANGLE;
1338
1339 if (!pCreateInfo->pMultisampleState ||
1340 !vk_find_struct_const(pCreateInfo->pMultisampleState->pNext,
1341 PIPELINE_SAMPLE_LOCATIONS_STATE_CREATE_INFO_EXT))
1342 states &= ~RADV_DYNAMIC_SAMPLE_LOCATIONS;
1343
1344 /* TODO: blend constants & line width. */
1345
1346 return states;
1347 }
1348
1349
1350 static void
1351 radv_pipeline_init_dynamic_state(struct radv_pipeline *pipeline,
1352 const VkGraphicsPipelineCreateInfo *pCreateInfo)
1353 {
1354 uint32_t needed_states = radv_pipeline_needed_dynamic_state(pCreateInfo);
1355 uint32_t states = needed_states;
1356 RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass);
1357 struct radv_subpass *subpass = &pass->subpasses[pCreateInfo->subpass];
1358
1359 pipeline->dynamic_state = default_dynamic_state;
1360 pipeline->graphics.needed_dynamic_state = needed_states;
1361
1362 if (pCreateInfo->pDynamicState) {
1363 /* Remove all of the states that are marked as dynamic */
1364 uint32_t count = pCreateInfo->pDynamicState->dynamicStateCount;
1365 for (uint32_t s = 0; s < count; s++)
1366 states &= ~radv_dynamic_state_mask(pCreateInfo->pDynamicState->pDynamicStates[s]);
1367 }
1368
1369 struct radv_dynamic_state *dynamic = &pipeline->dynamic_state;
1370
1371 if (needed_states & RADV_DYNAMIC_VIEWPORT) {
1372 assert(pCreateInfo->pViewportState);
1373
1374 dynamic->viewport.count = pCreateInfo->pViewportState->viewportCount;
1375 if (states & RADV_DYNAMIC_VIEWPORT) {
1376 typed_memcpy(dynamic->viewport.viewports,
1377 pCreateInfo->pViewportState->pViewports,
1378 pCreateInfo->pViewportState->viewportCount);
1379 }
1380 }
1381
1382 if (needed_states & RADV_DYNAMIC_SCISSOR) {
1383 dynamic->scissor.count = pCreateInfo->pViewportState->scissorCount;
1384 if (states & RADV_DYNAMIC_SCISSOR) {
1385 typed_memcpy(dynamic->scissor.scissors,
1386 pCreateInfo->pViewportState->pScissors,
1387 pCreateInfo->pViewportState->scissorCount);
1388 }
1389 }
1390
1391 if (states & RADV_DYNAMIC_LINE_WIDTH) {
1392 assert(pCreateInfo->pRasterizationState);
1393 dynamic->line_width = pCreateInfo->pRasterizationState->lineWidth;
1394 }
1395
1396 if (states & RADV_DYNAMIC_DEPTH_BIAS) {
1397 assert(pCreateInfo->pRasterizationState);
1398 dynamic->depth_bias.bias =
1399 pCreateInfo->pRasterizationState->depthBiasConstantFactor;
1400 dynamic->depth_bias.clamp =
1401 pCreateInfo->pRasterizationState->depthBiasClamp;
1402 dynamic->depth_bias.slope =
1403 pCreateInfo->pRasterizationState->depthBiasSlopeFactor;
1404 }
1405
1406 /* Section 9.2 of the Vulkan 1.0.15 spec says:
1407 *
1408 * pColorBlendState is [...] NULL if the pipeline has rasterization
1409 * disabled or if the subpass of the render pass the pipeline is
1410 * created against does not use any color attachments.
1411 */
1412 if (subpass->has_color_att && states & RADV_DYNAMIC_BLEND_CONSTANTS) {
1413 assert(pCreateInfo->pColorBlendState);
1414 typed_memcpy(dynamic->blend_constants,
1415 pCreateInfo->pColorBlendState->blendConstants, 4);
1416 }
1417
1418 /* If there is no depthstencil attachment, then don't read
1419 * pDepthStencilState. The Vulkan spec states that pDepthStencilState may
1420 * be NULL in this case. Even if pDepthStencilState is non-NULL, there is
1421 * no need to override the depthstencil defaults in
1422 * radv_pipeline::dynamic_state when there is no depthstencil attachment.
1423 *
1424 * Section 9.2 of the Vulkan 1.0.15 spec says:
1425 *
1426 * pDepthStencilState is [...] NULL if the pipeline has rasterization
1427 * disabled or if the subpass of the render pass the pipeline is created
1428 * against does not use a depth/stencil attachment.
1429 */
1430 if (needed_states && subpass->depth_stencil_attachment) {
1431 assert(pCreateInfo->pDepthStencilState);
1432
1433 if (states & RADV_DYNAMIC_DEPTH_BOUNDS) {
1434 dynamic->depth_bounds.min =
1435 pCreateInfo->pDepthStencilState->minDepthBounds;
1436 dynamic->depth_bounds.max =
1437 pCreateInfo->pDepthStencilState->maxDepthBounds;
1438 }
1439
1440 if (states & RADV_DYNAMIC_STENCIL_COMPARE_MASK) {
1441 dynamic->stencil_compare_mask.front =
1442 pCreateInfo->pDepthStencilState->front.compareMask;
1443 dynamic->stencil_compare_mask.back =
1444 pCreateInfo->pDepthStencilState->back.compareMask;
1445 }
1446
1447 if (states & RADV_DYNAMIC_STENCIL_WRITE_MASK) {
1448 dynamic->stencil_write_mask.front =
1449 pCreateInfo->pDepthStencilState->front.writeMask;
1450 dynamic->stencil_write_mask.back =
1451 pCreateInfo->pDepthStencilState->back.writeMask;
1452 }
1453
1454 if (states & RADV_DYNAMIC_STENCIL_REFERENCE) {
1455 dynamic->stencil_reference.front =
1456 pCreateInfo->pDepthStencilState->front.reference;
1457 dynamic->stencil_reference.back =
1458 pCreateInfo->pDepthStencilState->back.reference;
1459 }
1460 }
1461
1462 const VkPipelineDiscardRectangleStateCreateInfoEXT *discard_rectangle_info =
1463 vk_find_struct_const(pCreateInfo->pNext, PIPELINE_DISCARD_RECTANGLE_STATE_CREATE_INFO_EXT);
1464 if (needed_states & RADV_DYNAMIC_DISCARD_RECTANGLE) {
1465 dynamic->discard_rectangle.count = discard_rectangle_info->discardRectangleCount;
1466 if (states & RADV_DYNAMIC_DISCARD_RECTANGLE) {
1467 typed_memcpy(dynamic->discard_rectangle.rectangles,
1468 discard_rectangle_info->pDiscardRectangles,
1469 discard_rectangle_info->discardRectangleCount);
1470 }
1471 }
1472
1473 if (needed_states & RADV_DYNAMIC_SAMPLE_LOCATIONS) {
1474 const VkPipelineSampleLocationsStateCreateInfoEXT *sample_location_info =
1475 vk_find_struct_const(pCreateInfo->pMultisampleState->pNext,
1476 PIPELINE_SAMPLE_LOCATIONS_STATE_CREATE_INFO_EXT);
1477 /* If sampleLocationsEnable is VK_FALSE, the default sample
1478 * locations are used and the values specified in
1479 * sampleLocationsInfo are ignored.
1480 */
1481 if (sample_location_info->sampleLocationsEnable) {
1482 const VkSampleLocationsInfoEXT *pSampleLocationsInfo =
1483 &sample_location_info->sampleLocationsInfo;
1484
1485 assert(pSampleLocationsInfo->sampleLocationsCount <= MAX_SAMPLE_LOCATIONS);
1486
1487 dynamic->sample_location.per_pixel = pSampleLocationsInfo->sampleLocationsPerPixel;
1488 dynamic->sample_location.grid_size = pSampleLocationsInfo->sampleLocationGridSize;
1489 dynamic->sample_location.count = pSampleLocationsInfo->sampleLocationsCount;
1490 typed_memcpy(&dynamic->sample_location.locations[0],
1491 pSampleLocationsInfo->pSampleLocations,
1492 pSampleLocationsInfo->sampleLocationsCount);
1493 }
1494 }
1495
1496 pipeline->dynamic_state.mask = states;
1497 }
1498
1499 static void
1500 gfx9_get_gs_info(const struct radv_pipeline_key *key,
1501 const struct radv_pipeline *pipeline,
1502 nir_shader **nir,
1503 struct radv_shader_info *infos,
1504 struct gfx9_gs_info *out)
1505 {
1506 struct radv_shader_info *gs_info = &infos[MESA_SHADER_GEOMETRY];
1507 struct radv_es_output_info *es_info;
1508 if (pipeline->device->physical_device->rad_info.chip_class >= GFX9)
1509 es_info = nir[MESA_SHADER_TESS_CTRL] ? &gs_info->tes.es_info : &gs_info->vs.es_info;
1510 else
1511 es_info = nir[MESA_SHADER_TESS_CTRL] ?
1512 &infos[MESA_SHADER_TESS_EVAL].tes.es_info :
1513 &infos[MESA_SHADER_VERTEX].vs.es_info;
1514
1515 unsigned gs_num_invocations = MAX2(gs_info->gs.invocations, 1);
1516 bool uses_adjacency;
1517 switch(key->topology) {
1518 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY:
1519 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY:
1520 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY:
1521 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY:
1522 uses_adjacency = true;
1523 break;
1524 default:
1525 uses_adjacency = false;
1526 break;
1527 }
1528
1529 /* All these are in dwords: */
1530 /* We can't allow using the whole LDS, because GS waves compete with
1531 * other shader stages for LDS space. */
1532 const unsigned max_lds_size = 8 * 1024;
1533 const unsigned esgs_itemsize = es_info->esgs_itemsize / 4;
1534 unsigned esgs_lds_size;
1535
1536 /* All these are per subgroup: */
1537 const unsigned max_out_prims = 32 * 1024;
1538 const unsigned max_es_verts = 255;
1539 const unsigned ideal_gs_prims = 64;
1540 unsigned max_gs_prims, gs_prims;
1541 unsigned min_es_verts, es_verts, worst_case_es_verts;
1542
1543 if (uses_adjacency || gs_num_invocations > 1)
1544 max_gs_prims = 127 / gs_num_invocations;
1545 else
1546 max_gs_prims = 255;
1547
1548 /* MAX_PRIMS_PER_SUBGROUP = gs_prims * max_vert_out * gs_invocations.
1549 * Make sure we don't go over the maximum value.
1550 */
1551 if (gs_info->gs.vertices_out > 0) {
1552 max_gs_prims = MIN2(max_gs_prims,
1553 max_out_prims /
1554 (gs_info->gs.vertices_out * gs_num_invocations));
1555 }
1556 assert(max_gs_prims > 0);
1557
1558 /* If the primitive has adjacency, halve the number of vertices
1559 * that will be reused in multiple primitives.
1560 */
1561 min_es_verts = gs_info->gs.vertices_in / (uses_adjacency ? 2 : 1);
1562
1563 gs_prims = MIN2(ideal_gs_prims, max_gs_prims);
1564 worst_case_es_verts = MIN2(min_es_verts * gs_prims, max_es_verts);
1565
1566 /* Compute ESGS LDS size based on the worst case number of ES vertices
1567 * needed to create the target number of GS prims per subgroup.
1568 */
1569 esgs_lds_size = esgs_itemsize * worst_case_es_verts;
1570
1571 /* If total LDS usage is too big, refactor partitions based on ratio
1572 * of ESGS item sizes.
1573 */
1574 if (esgs_lds_size > max_lds_size) {
1575 /* Our target GS Prims Per Subgroup was too large. Calculate
1576 * the maximum number of GS Prims Per Subgroup that will fit
1577 * into LDS, capped by the maximum that the hardware can support.
1578 */
1579 gs_prims = MIN2((max_lds_size / (esgs_itemsize * min_es_verts)),
1580 max_gs_prims);
1581 assert(gs_prims > 0);
1582 worst_case_es_verts = MIN2(min_es_verts * gs_prims,
1583 max_es_verts);
1584
1585 esgs_lds_size = esgs_itemsize * worst_case_es_verts;
1586 assert(esgs_lds_size <= max_lds_size);
1587 }
1588
1589 /* Now calculate remaining ESGS information. */
1590 if (esgs_lds_size)
1591 es_verts = MIN2(esgs_lds_size / esgs_itemsize, max_es_verts);
1592 else
1593 es_verts = max_es_verts;
1594
1595 /* Vertices for adjacency primitives are not always reused, so restore
1596 * it for ES_VERTS_PER_SUBGRP.
1597 */
1598 min_es_verts = gs_info->gs.vertices_in;
1599
1600 /* For normal primitives, the VGT only checks if they are past the ES
1601 * verts per subgroup after allocating a full GS primitive and if they
1602 * are, kick off a new subgroup. But if those additional ES verts are
1603 * unique (e.g. not reused) we need to make sure there is enough LDS
1604 * space to account for those ES verts beyond ES_VERTS_PER_SUBGRP.
1605 */
1606 es_verts -= min_es_verts - 1;
1607
1608 uint32_t es_verts_per_subgroup = es_verts;
1609 uint32_t gs_prims_per_subgroup = gs_prims;
1610 uint32_t gs_inst_prims_in_subgroup = gs_prims * gs_num_invocations;
1611 uint32_t max_prims_per_subgroup = gs_inst_prims_in_subgroup * gs_info->gs.vertices_out;
1612 out->lds_size = align(esgs_lds_size, 128) / 128;
1613 out->vgt_gs_onchip_cntl = S_028A44_ES_VERTS_PER_SUBGRP(es_verts_per_subgroup) |
1614 S_028A44_GS_PRIMS_PER_SUBGRP(gs_prims_per_subgroup) |
1615 S_028A44_GS_INST_PRIMS_IN_SUBGRP(gs_inst_prims_in_subgroup);
1616 out->vgt_gs_max_prims_per_subgroup = S_028A94_MAX_PRIMS_PER_SUBGROUP(max_prims_per_subgroup);
1617 out->vgt_esgs_ring_itemsize = esgs_itemsize;
1618 assert(max_prims_per_subgroup <= max_out_prims);
1619 }
1620
1621 static void clamp_gsprims_to_esverts(unsigned *max_gsprims, unsigned max_esverts,
1622 unsigned min_verts_per_prim, bool use_adjacency)
1623 {
1624 unsigned max_reuse = max_esverts - min_verts_per_prim;
1625 if (use_adjacency)
1626 max_reuse /= 2;
1627 *max_gsprims = MIN2(*max_gsprims, 1 + max_reuse);
1628 }
1629
1630 static unsigned
1631 radv_get_num_input_vertices(nir_shader **nir)
1632 {
1633 if (nir[MESA_SHADER_GEOMETRY]) {
1634 nir_shader *gs = nir[MESA_SHADER_GEOMETRY];
1635
1636 return gs->info.gs.vertices_in;
1637 }
1638
1639 if (nir[MESA_SHADER_TESS_CTRL]) {
1640 nir_shader *tes = nir[MESA_SHADER_TESS_EVAL];
1641
1642 if (tes->info.tess.point_mode)
1643 return 1;
1644 if (tes->info.tess.primitive_mode == GL_ISOLINES)
1645 return 2;
1646 return 3;
1647 }
1648
1649 return 3;
1650 }
1651
1652 static void
1653 gfx10_get_ngg_info(const struct radv_pipeline_key *key,
1654 struct radv_pipeline *pipeline,
1655 nir_shader **nir,
1656 struct radv_shader_info *infos,
1657 struct gfx10_ngg_info *ngg)
1658 {
1659 struct radv_shader_info *gs_info = &infos[MESA_SHADER_GEOMETRY];
1660 struct radv_es_output_info *es_info =
1661 nir[MESA_SHADER_TESS_CTRL] ? &gs_info->tes.es_info : &gs_info->vs.es_info;
1662 unsigned gs_type = nir[MESA_SHADER_GEOMETRY] ? MESA_SHADER_GEOMETRY : MESA_SHADER_VERTEX;
1663 unsigned max_verts_per_prim = radv_get_num_input_vertices(nir);
1664 unsigned min_verts_per_prim =
1665 gs_type == MESA_SHADER_GEOMETRY ? max_verts_per_prim : 1;
1666 unsigned gs_num_invocations = nir[MESA_SHADER_GEOMETRY] ? MAX2(gs_info->gs.invocations, 1) : 1;
1667 bool uses_adjacency;
1668 switch(key->topology) {
1669 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY:
1670 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY:
1671 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY:
1672 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY:
1673 uses_adjacency = true;
1674 break;
1675 default:
1676 uses_adjacency = false;
1677 break;
1678 }
1679
1680 /* All these are in dwords: */
1681 /* We can't allow using the whole LDS, because GS waves compete with
1682 * other shader stages for LDS space.
1683 *
1684 * TODO: We should really take the shader's internal LDS use into
1685 * account. The linker will fail if the size is greater than
1686 * 8K dwords.
1687 */
1688 const unsigned max_lds_size = 8 * 1024 - 768;
1689 const unsigned target_lds_size = max_lds_size;
1690 unsigned esvert_lds_size = 0;
1691 unsigned gsprim_lds_size = 0;
1692
1693 /* All these are per subgroup: */
1694 bool max_vert_out_per_gs_instance = false;
1695 unsigned max_esverts_base = 256;
1696 unsigned max_gsprims_base = 128; /* default prim group size clamp */
1697
1698 /* Hardware has the following non-natural restrictions on the value
1699 * of GE_CNTL.VERT_GRP_SIZE based on based on the primitive type of
1700 * the draw:
1701 * - at most 252 for any line input primitive type
1702 * - at most 251 for any quad input primitive type
1703 * - at most 251 for triangle strips with adjacency (this happens to
1704 * be the natural limit for triangle *lists* with adjacency)
1705 */
1706 max_esverts_base = MIN2(max_esverts_base, 251 + max_verts_per_prim - 1);
1707
1708 if (gs_type == MESA_SHADER_GEOMETRY) {
1709 unsigned max_out_verts_per_gsprim =
1710 gs_info->gs.vertices_out * gs_num_invocations;
1711
1712 if (max_out_verts_per_gsprim <= 256) {
1713 if (max_out_verts_per_gsprim) {
1714 max_gsprims_base = MIN2(max_gsprims_base,
1715 256 / max_out_verts_per_gsprim);
1716 }
1717 } else {
1718 /* Use special multi-cycling mode in which each GS
1719 * instance gets its own subgroup. Does not work with
1720 * tessellation. */
1721 max_vert_out_per_gs_instance = true;
1722 max_gsprims_base = 1;
1723 max_out_verts_per_gsprim = gs_info->gs.vertices_out;
1724 }
1725
1726 esvert_lds_size = es_info->esgs_itemsize / 4;
1727 gsprim_lds_size = (gs_info->gs.gsvs_vertex_size / 4 + 1) * max_out_verts_per_gsprim;
1728 } else {
1729 /* VS and TES. */
1730 /* LDS size for passing data from GS to ES. */
1731 struct radv_streamout_info *so_info = nir[MESA_SHADER_TESS_CTRL]
1732 ? &infos[MESA_SHADER_TESS_EVAL].so
1733 : &infos[MESA_SHADER_VERTEX].so;
1734
1735 if (so_info->num_outputs)
1736 esvert_lds_size = 4 * so_info->num_outputs + 1;
1737
1738 /* GS stores Primitive IDs (one DWORD) into LDS at the address
1739 * corresponding to the ES thread of the provoking vertex. All
1740 * ES threads load and export PrimitiveID for their thread.
1741 */
1742 if (!nir[MESA_SHADER_TESS_CTRL] &&
1743 infos[MESA_SHADER_VERTEX].vs.outinfo.export_prim_id)
1744 esvert_lds_size = MAX2(esvert_lds_size, 1);
1745 }
1746
1747 unsigned max_gsprims = max_gsprims_base;
1748 unsigned max_esverts = max_esverts_base;
1749
1750 if (esvert_lds_size)
1751 max_esverts = MIN2(max_esverts, target_lds_size / esvert_lds_size);
1752 if (gsprim_lds_size)
1753 max_gsprims = MIN2(max_gsprims, target_lds_size / gsprim_lds_size);
1754
1755 max_esverts = MIN2(max_esverts, max_gsprims * max_verts_per_prim);
1756 clamp_gsprims_to_esverts(&max_gsprims, max_esverts, min_verts_per_prim, uses_adjacency);
1757 assert(max_esverts >= max_verts_per_prim && max_gsprims >= 1);
1758
1759 if (esvert_lds_size || gsprim_lds_size) {
1760 /* Now that we have a rough proportionality between esverts
1761 * and gsprims based on the primitive type, scale both of them
1762 * down simultaneously based on required LDS space.
1763 *
1764 * We could be smarter about this if we knew how much vertex
1765 * reuse to expect.
1766 */
1767 unsigned lds_total = max_esverts * esvert_lds_size +
1768 max_gsprims * gsprim_lds_size;
1769 if (lds_total > target_lds_size) {
1770 max_esverts = max_esverts * target_lds_size / lds_total;
1771 max_gsprims = max_gsprims * target_lds_size / lds_total;
1772
1773 max_esverts = MIN2(max_esverts, max_gsprims * max_verts_per_prim);
1774 clamp_gsprims_to_esverts(&max_gsprims, max_esverts,
1775 min_verts_per_prim, uses_adjacency);
1776 assert(max_esverts >= max_verts_per_prim && max_gsprims >= 1);
1777 }
1778 }
1779
1780 /* Round up towards full wave sizes for better ALU utilization. */
1781 if (!max_vert_out_per_gs_instance) {
1782 unsigned orig_max_esverts;
1783 unsigned orig_max_gsprims;
1784 unsigned wavesize;
1785
1786 if (gs_type == MESA_SHADER_GEOMETRY) {
1787 wavesize = gs_info->wave_size;
1788 } else {
1789 wavesize = nir[MESA_SHADER_TESS_CTRL]
1790 ? infos[MESA_SHADER_TESS_EVAL].wave_size
1791 : infos[MESA_SHADER_VERTEX].wave_size;
1792 }
1793
1794 do {
1795 orig_max_esverts = max_esverts;
1796 orig_max_gsprims = max_gsprims;
1797
1798 max_esverts = align(max_esverts, wavesize);
1799 max_esverts = MIN2(max_esverts, max_esverts_base);
1800 if (esvert_lds_size)
1801 max_esverts = MIN2(max_esverts,
1802 (max_lds_size - max_gsprims * gsprim_lds_size) /
1803 esvert_lds_size);
1804 max_esverts = MIN2(max_esverts, max_gsprims * max_verts_per_prim);
1805
1806 max_gsprims = align(max_gsprims, wavesize);
1807 max_gsprims = MIN2(max_gsprims, max_gsprims_base);
1808 if (gsprim_lds_size)
1809 max_gsprims = MIN2(max_gsprims,
1810 (max_lds_size - max_esverts * esvert_lds_size) /
1811 gsprim_lds_size);
1812 clamp_gsprims_to_esverts(&max_gsprims, max_esverts,
1813 min_verts_per_prim, uses_adjacency);
1814 assert(max_esverts >= max_verts_per_prim && max_gsprims >= 1);
1815 } while (orig_max_esverts != max_esverts || orig_max_gsprims != max_gsprims);
1816 }
1817
1818 /* Hardware restriction: minimum value of max_esverts */
1819 max_esverts = MAX2(max_esverts, 23 + max_verts_per_prim);
1820
1821 unsigned max_out_vertices =
1822 max_vert_out_per_gs_instance ? gs_info->gs.vertices_out :
1823 gs_type == MESA_SHADER_GEOMETRY ?
1824 max_gsprims * gs_num_invocations * gs_info->gs.vertices_out :
1825 max_esverts;
1826 assert(max_out_vertices <= 256);
1827
1828 unsigned prim_amp_factor = 1;
1829 if (gs_type == MESA_SHADER_GEOMETRY) {
1830 /* Number of output primitives per GS input primitive after
1831 * GS instancing. */
1832 prim_amp_factor = gs_info->gs.vertices_out;
1833 }
1834
1835 /* The GE only checks against the maximum number of ES verts after
1836 * allocating a full GS primitive. So we need to ensure that whenever
1837 * this check passes, there is enough space for a full primitive without
1838 * vertex reuse.
1839 */
1840 ngg->hw_max_esverts = max_esverts - max_verts_per_prim + 1;
1841 ngg->max_gsprims = max_gsprims;
1842 ngg->max_out_verts = max_out_vertices;
1843 ngg->prim_amp_factor = prim_amp_factor;
1844 ngg->max_vert_out_per_gs_instance = max_vert_out_per_gs_instance;
1845 ngg->ngg_emit_size = max_gsprims * gsprim_lds_size;
1846 ngg->esgs_ring_size = 4 * max_esverts * esvert_lds_size;
1847
1848 if (gs_type == MESA_SHADER_GEOMETRY) {
1849 ngg->vgt_esgs_ring_itemsize = es_info->esgs_itemsize / 4;
1850 } else {
1851 ngg->vgt_esgs_ring_itemsize = 1;
1852 }
1853
1854 pipeline->graphics.esgs_ring_size = ngg->esgs_ring_size;
1855
1856 assert(ngg->hw_max_esverts >= 24); /* HW limitation */
1857 }
1858
1859 static void
1860 calculate_gs_ring_sizes(struct radv_pipeline *pipeline,
1861 const struct gfx9_gs_info *gs)
1862 {
1863 struct radv_device *device = pipeline->device;
1864 unsigned num_se = device->physical_device->rad_info.max_se;
1865 unsigned wave_size = 64;
1866 unsigned max_gs_waves = 32 * num_se; /* max 32 per SE on GCN */
1867 /* On GFX6-GFX7, the value comes from VGT_GS_VERTEX_REUSE = 16.
1868 * On GFX8+, the value comes from VGT_VERTEX_REUSE_BLOCK_CNTL = 30 (+2).
1869 */
1870 unsigned gs_vertex_reuse =
1871 (device->physical_device->rad_info.chip_class >= GFX8 ? 32 : 16) * num_se;
1872 unsigned alignment = 256 * num_se;
1873 /* The maximum size is 63.999 MB per SE. */
1874 unsigned max_size = ((unsigned)(63.999 * 1024 * 1024) & ~255) * num_se;
1875 struct radv_shader_info *gs_info = &pipeline->shaders[MESA_SHADER_GEOMETRY]->info;
1876
1877 /* Calculate the minimum size. */
1878 unsigned min_esgs_ring_size = align(gs->vgt_esgs_ring_itemsize * 4 * gs_vertex_reuse *
1879 wave_size, alignment);
1880 /* These are recommended sizes, not minimum sizes. */
1881 unsigned esgs_ring_size = max_gs_waves * 2 * wave_size *
1882 gs->vgt_esgs_ring_itemsize * 4 * gs_info->gs.vertices_in;
1883 unsigned gsvs_ring_size = max_gs_waves * 2 * wave_size *
1884 gs_info->gs.max_gsvs_emit_size;
1885
1886 min_esgs_ring_size = align(min_esgs_ring_size, alignment);
1887 esgs_ring_size = align(esgs_ring_size, alignment);
1888 gsvs_ring_size = align(gsvs_ring_size, alignment);
1889
1890 if (pipeline->device->physical_device->rad_info.chip_class <= GFX8)
1891 pipeline->graphics.esgs_ring_size = CLAMP(esgs_ring_size, min_esgs_ring_size, max_size);
1892
1893 pipeline->graphics.gsvs_ring_size = MIN2(gsvs_ring_size, max_size);
1894 }
1895
1896 static void si_multiwave_lds_size_workaround(struct radv_device *device,
1897 unsigned *lds_size)
1898 {
1899 /* If tessellation is all offchip and on-chip GS isn't used, this
1900 * workaround is not needed.
1901 */
1902 return;
1903
1904 /* SPI barrier management bug:
1905 * Make sure we have at least 4k of LDS in use to avoid the bug.
1906 * It applies to workgroup sizes of more than one wavefront.
1907 */
1908 if (device->physical_device->rad_info.family == CHIP_BONAIRE ||
1909 device->physical_device->rad_info.family == CHIP_KABINI)
1910 *lds_size = MAX2(*lds_size, 8);
1911 }
1912
1913 struct radv_shader_variant *
1914 radv_get_shader(struct radv_pipeline *pipeline,
1915 gl_shader_stage stage)
1916 {
1917 if (stage == MESA_SHADER_VERTEX) {
1918 if (pipeline->shaders[MESA_SHADER_VERTEX])
1919 return pipeline->shaders[MESA_SHADER_VERTEX];
1920 if (pipeline->shaders[MESA_SHADER_TESS_CTRL])
1921 return pipeline->shaders[MESA_SHADER_TESS_CTRL];
1922 if (pipeline->shaders[MESA_SHADER_GEOMETRY])
1923 return pipeline->shaders[MESA_SHADER_GEOMETRY];
1924 } else if (stage == MESA_SHADER_TESS_EVAL) {
1925 if (!radv_pipeline_has_tess(pipeline))
1926 return NULL;
1927 if (pipeline->shaders[MESA_SHADER_TESS_EVAL])
1928 return pipeline->shaders[MESA_SHADER_TESS_EVAL];
1929 if (pipeline->shaders[MESA_SHADER_GEOMETRY])
1930 return pipeline->shaders[MESA_SHADER_GEOMETRY];
1931 }
1932 return pipeline->shaders[stage];
1933 }
1934
1935 static struct radv_tessellation_state
1936 calculate_tess_state(struct radv_pipeline *pipeline,
1937 const VkGraphicsPipelineCreateInfo *pCreateInfo)
1938 {
1939 unsigned num_tcs_input_cp;
1940 unsigned num_tcs_output_cp;
1941 unsigned lds_size;
1942 unsigned num_patches;
1943 struct radv_tessellation_state tess = {0};
1944
1945 num_tcs_input_cp = pCreateInfo->pTessellationState->patchControlPoints;
1946 num_tcs_output_cp = pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.tcs.tcs_vertices_out; //TCS VERTICES OUT
1947 num_patches = pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.tcs.num_patches;
1948
1949 lds_size = pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.tcs.lds_size;
1950
1951 if (pipeline->device->physical_device->rad_info.chip_class >= GFX7) {
1952 assert(lds_size <= 65536);
1953 lds_size = align(lds_size, 512) / 512;
1954 } else {
1955 assert(lds_size <= 32768);
1956 lds_size = align(lds_size, 256) / 256;
1957 }
1958 si_multiwave_lds_size_workaround(pipeline->device, &lds_size);
1959
1960 tess.lds_size = lds_size;
1961
1962 tess.ls_hs_config = S_028B58_NUM_PATCHES(num_patches) |
1963 S_028B58_HS_NUM_INPUT_CP(num_tcs_input_cp) |
1964 S_028B58_HS_NUM_OUTPUT_CP(num_tcs_output_cp);
1965 tess.num_patches = num_patches;
1966
1967 struct radv_shader_variant *tes = radv_get_shader(pipeline, MESA_SHADER_TESS_EVAL);
1968 unsigned type = 0, partitioning = 0, topology = 0, distribution_mode = 0;
1969
1970 switch (tes->info.tes.primitive_mode) {
1971 case GL_TRIANGLES:
1972 type = V_028B6C_TESS_TRIANGLE;
1973 break;
1974 case GL_QUADS:
1975 type = V_028B6C_TESS_QUAD;
1976 break;
1977 case GL_ISOLINES:
1978 type = V_028B6C_TESS_ISOLINE;
1979 break;
1980 }
1981
1982 switch (tes->info.tes.spacing) {
1983 case TESS_SPACING_EQUAL:
1984 partitioning = V_028B6C_PART_INTEGER;
1985 break;
1986 case TESS_SPACING_FRACTIONAL_ODD:
1987 partitioning = V_028B6C_PART_FRAC_ODD;
1988 break;
1989 case TESS_SPACING_FRACTIONAL_EVEN:
1990 partitioning = V_028B6C_PART_FRAC_EVEN;
1991 break;
1992 default:
1993 break;
1994 }
1995
1996 bool ccw = tes->info.tes.ccw;
1997 const VkPipelineTessellationDomainOriginStateCreateInfo *domain_origin_state =
1998 vk_find_struct_const(pCreateInfo->pTessellationState,
1999 PIPELINE_TESSELLATION_DOMAIN_ORIGIN_STATE_CREATE_INFO);
2000
2001 if (domain_origin_state && domain_origin_state->domainOrigin != VK_TESSELLATION_DOMAIN_ORIGIN_UPPER_LEFT)
2002 ccw = !ccw;
2003
2004 if (tes->info.tes.point_mode)
2005 topology = V_028B6C_OUTPUT_POINT;
2006 else if (tes->info.tes.primitive_mode == GL_ISOLINES)
2007 topology = V_028B6C_OUTPUT_LINE;
2008 else if (ccw)
2009 topology = V_028B6C_OUTPUT_TRIANGLE_CCW;
2010 else
2011 topology = V_028B6C_OUTPUT_TRIANGLE_CW;
2012
2013 if (pipeline->device->physical_device->rad_info.has_distributed_tess) {
2014 if (pipeline->device->physical_device->rad_info.family == CHIP_FIJI ||
2015 pipeline->device->physical_device->rad_info.family >= CHIP_POLARIS10)
2016 distribution_mode = V_028B6C_DISTRIBUTION_MODE_TRAPEZOIDS;
2017 else
2018 distribution_mode = V_028B6C_DISTRIBUTION_MODE_DONUTS;
2019 } else
2020 distribution_mode = V_028B6C_DISTRIBUTION_MODE_NO_DIST;
2021
2022 tess.tf_param = S_028B6C_TYPE(type) |
2023 S_028B6C_PARTITIONING(partitioning) |
2024 S_028B6C_TOPOLOGY(topology) |
2025 S_028B6C_DISTRIBUTION_MODE(distribution_mode);
2026
2027 return tess;
2028 }
2029
2030 static const struct radv_prim_vertex_count prim_size_table[] = {
2031 [V_008958_DI_PT_NONE] = {0, 0},
2032 [V_008958_DI_PT_POINTLIST] = {1, 1},
2033 [V_008958_DI_PT_LINELIST] = {2, 2},
2034 [V_008958_DI_PT_LINESTRIP] = {2, 1},
2035 [V_008958_DI_PT_TRILIST] = {3, 3},
2036 [V_008958_DI_PT_TRIFAN] = {3, 1},
2037 [V_008958_DI_PT_TRISTRIP] = {3, 1},
2038 [V_008958_DI_PT_LINELIST_ADJ] = {4, 4},
2039 [V_008958_DI_PT_LINESTRIP_ADJ] = {4, 1},
2040 [V_008958_DI_PT_TRILIST_ADJ] = {6, 6},
2041 [V_008958_DI_PT_TRISTRIP_ADJ] = {6, 2},
2042 [V_008958_DI_PT_RECTLIST] = {3, 3},
2043 [V_008958_DI_PT_LINELOOP] = {2, 1},
2044 [V_008958_DI_PT_POLYGON] = {3, 1},
2045 [V_008958_DI_PT_2D_TRI_STRIP] = {0, 0},
2046 };
2047
2048 static const struct radv_vs_output_info *get_vs_output_info(const struct radv_pipeline *pipeline)
2049 {
2050 if (radv_pipeline_has_gs(pipeline))
2051 if (radv_pipeline_has_ngg(pipeline))
2052 return &pipeline->shaders[MESA_SHADER_GEOMETRY]->info.vs.outinfo;
2053 else
2054 return &pipeline->gs_copy_shader->info.vs.outinfo;
2055 else if (radv_pipeline_has_tess(pipeline))
2056 return &pipeline->shaders[MESA_SHADER_TESS_EVAL]->info.tes.outinfo;
2057 else
2058 return &pipeline->shaders[MESA_SHADER_VERTEX]->info.vs.outinfo;
2059 }
2060
2061 static void
2062 radv_link_shaders(struct radv_pipeline *pipeline, nir_shader **shaders)
2063 {
2064 nir_shader* ordered_shaders[MESA_SHADER_STAGES];
2065 int shader_count = 0;
2066
2067 if(shaders[MESA_SHADER_FRAGMENT]) {
2068 ordered_shaders[shader_count++] = shaders[MESA_SHADER_FRAGMENT];
2069 }
2070 if(shaders[MESA_SHADER_GEOMETRY]) {
2071 ordered_shaders[shader_count++] = shaders[MESA_SHADER_GEOMETRY];
2072 }
2073 if(shaders[MESA_SHADER_TESS_EVAL]) {
2074 ordered_shaders[shader_count++] = shaders[MESA_SHADER_TESS_EVAL];
2075 }
2076 if(shaders[MESA_SHADER_TESS_CTRL]) {
2077 ordered_shaders[shader_count++] = shaders[MESA_SHADER_TESS_CTRL];
2078 }
2079 if(shaders[MESA_SHADER_VERTEX]) {
2080 ordered_shaders[shader_count++] = shaders[MESA_SHADER_VERTEX];
2081 }
2082
2083 if (shader_count > 1) {
2084 unsigned first = ordered_shaders[shader_count - 1]->info.stage;
2085 unsigned last = ordered_shaders[0]->info.stage;
2086
2087 if (ordered_shaders[0]->info.stage == MESA_SHADER_FRAGMENT &&
2088 ordered_shaders[1]->info.has_transform_feedback_varyings)
2089 nir_link_xfb_varyings(ordered_shaders[1], ordered_shaders[0]);
2090
2091 for (int i = 0; i < shader_count; ++i) {
2092 nir_variable_mode mask = 0;
2093
2094 if (ordered_shaders[i]->info.stage != first)
2095 mask = mask | nir_var_shader_in;
2096
2097 if (ordered_shaders[i]->info.stage != last)
2098 mask = mask | nir_var_shader_out;
2099
2100 nir_lower_io_to_scalar_early(ordered_shaders[i], mask);
2101 radv_optimize_nir(ordered_shaders[i], false, false);
2102 }
2103 }
2104
2105 for (int i = 1; i < shader_count; ++i) {
2106 nir_lower_io_arrays_to_elements(ordered_shaders[i],
2107 ordered_shaders[i - 1]);
2108
2109 if (nir_link_opt_varyings(ordered_shaders[i],
2110 ordered_shaders[i - 1]))
2111 radv_optimize_nir(ordered_shaders[i - 1], false, false);
2112
2113 nir_remove_dead_variables(ordered_shaders[i],
2114 nir_var_shader_out);
2115 nir_remove_dead_variables(ordered_shaders[i - 1],
2116 nir_var_shader_in);
2117
2118 bool progress = nir_remove_unused_varyings(ordered_shaders[i],
2119 ordered_shaders[i - 1]);
2120
2121 nir_compact_varyings(ordered_shaders[i],
2122 ordered_shaders[i - 1], true);
2123
2124 if (progress) {
2125 if (nir_lower_global_vars_to_local(ordered_shaders[i])) {
2126 ac_lower_indirect_derefs(ordered_shaders[i],
2127 pipeline->device->physical_device->rad_info.chip_class);
2128 }
2129 radv_optimize_nir(ordered_shaders[i], false, false);
2130
2131 if (nir_lower_global_vars_to_local(ordered_shaders[i - 1])) {
2132 ac_lower_indirect_derefs(ordered_shaders[i - 1],
2133 pipeline->device->physical_device->rad_info.chip_class);
2134 }
2135 radv_optimize_nir(ordered_shaders[i - 1], false, false);
2136 }
2137 }
2138 }
2139
2140 static uint32_t
2141 radv_get_attrib_stride(const VkPipelineVertexInputStateCreateInfo *input_state,
2142 uint32_t attrib_binding)
2143 {
2144 for (uint32_t i = 0; i < input_state->vertexBindingDescriptionCount; i++) {
2145 const VkVertexInputBindingDescription *input_binding =
2146 &input_state->pVertexBindingDescriptions[i];
2147
2148 if (input_binding->binding == attrib_binding)
2149 return input_binding->stride;
2150 }
2151
2152 return 0;
2153 }
2154
2155 static struct radv_pipeline_key
2156 radv_generate_graphics_pipeline_key(struct radv_pipeline *pipeline,
2157 const VkGraphicsPipelineCreateInfo *pCreateInfo,
2158 const struct radv_blend_state *blend,
2159 bool has_view_index)
2160 {
2161 const VkPipelineVertexInputStateCreateInfo *input_state =
2162 pCreateInfo->pVertexInputState;
2163 const VkPipelineVertexInputDivisorStateCreateInfoEXT *divisor_state =
2164 vk_find_struct_const(input_state->pNext, PIPELINE_VERTEX_INPUT_DIVISOR_STATE_CREATE_INFO_EXT);
2165
2166 struct radv_pipeline_key key;
2167 memset(&key, 0, sizeof(key));
2168
2169 if (pCreateInfo->flags & VK_PIPELINE_CREATE_DISABLE_OPTIMIZATION_BIT)
2170 key.optimisations_disabled = 1;
2171
2172 key.has_multiview_view_index = has_view_index;
2173
2174 uint32_t binding_input_rate = 0;
2175 uint32_t instance_rate_divisors[MAX_VERTEX_ATTRIBS];
2176 for (unsigned i = 0; i < input_state->vertexBindingDescriptionCount; ++i) {
2177 if (input_state->pVertexBindingDescriptions[i].inputRate) {
2178 unsigned binding = input_state->pVertexBindingDescriptions[i].binding;
2179 binding_input_rate |= 1u << binding;
2180 instance_rate_divisors[binding] = 1;
2181 }
2182 }
2183 if (divisor_state) {
2184 for (unsigned i = 0; i < divisor_state->vertexBindingDivisorCount; ++i) {
2185 instance_rate_divisors[divisor_state->pVertexBindingDivisors[i].binding] =
2186 divisor_state->pVertexBindingDivisors[i].divisor;
2187 }
2188 }
2189
2190 for (unsigned i = 0; i < input_state->vertexAttributeDescriptionCount; ++i) {
2191 const VkVertexInputAttributeDescription *desc =
2192 &input_state->pVertexAttributeDescriptions[i];
2193 const struct vk_format_description *format_desc;
2194 unsigned location = desc->location;
2195 unsigned binding = desc->binding;
2196 unsigned num_format, data_format;
2197 int first_non_void;
2198
2199 if (binding_input_rate & (1u << binding)) {
2200 key.instance_rate_inputs |= 1u << location;
2201 key.instance_rate_divisors[location] = instance_rate_divisors[binding];
2202 }
2203
2204 format_desc = vk_format_description(desc->format);
2205 first_non_void = vk_format_get_first_non_void_channel(desc->format);
2206
2207 num_format = radv_translate_buffer_numformat(format_desc, first_non_void);
2208 data_format = radv_translate_buffer_dataformat(format_desc, first_non_void);
2209
2210 key.vertex_attribute_formats[location] = data_format | (num_format << 4);
2211 key.vertex_attribute_bindings[location] = desc->binding;
2212 key.vertex_attribute_offsets[location] = desc->offset;
2213 key.vertex_attribute_strides[location] = radv_get_attrib_stride(input_state, desc->binding);
2214
2215 if (pipeline->device->physical_device->rad_info.chip_class <= GFX8 &&
2216 pipeline->device->physical_device->rad_info.family != CHIP_STONEY) {
2217 VkFormat format = input_state->pVertexAttributeDescriptions[i].format;
2218 uint64_t adjust;
2219 switch(format) {
2220 case VK_FORMAT_A2R10G10B10_SNORM_PACK32:
2221 case VK_FORMAT_A2B10G10R10_SNORM_PACK32:
2222 adjust = RADV_ALPHA_ADJUST_SNORM;
2223 break;
2224 case VK_FORMAT_A2R10G10B10_SSCALED_PACK32:
2225 case VK_FORMAT_A2B10G10R10_SSCALED_PACK32:
2226 adjust = RADV_ALPHA_ADJUST_SSCALED;
2227 break;
2228 case VK_FORMAT_A2R10G10B10_SINT_PACK32:
2229 case VK_FORMAT_A2B10G10R10_SINT_PACK32:
2230 adjust = RADV_ALPHA_ADJUST_SINT;
2231 break;
2232 default:
2233 adjust = 0;
2234 break;
2235 }
2236 key.vertex_alpha_adjust |= adjust << (2 * location);
2237 }
2238
2239 switch (desc->format) {
2240 case VK_FORMAT_B8G8R8A8_UNORM:
2241 case VK_FORMAT_B8G8R8A8_SNORM:
2242 case VK_FORMAT_B8G8R8A8_USCALED:
2243 case VK_FORMAT_B8G8R8A8_SSCALED:
2244 case VK_FORMAT_B8G8R8A8_UINT:
2245 case VK_FORMAT_B8G8R8A8_SINT:
2246 case VK_FORMAT_B8G8R8A8_SRGB:
2247 case VK_FORMAT_A2R10G10B10_UNORM_PACK32:
2248 case VK_FORMAT_A2R10G10B10_SNORM_PACK32:
2249 case VK_FORMAT_A2R10G10B10_USCALED_PACK32:
2250 case VK_FORMAT_A2R10G10B10_SSCALED_PACK32:
2251 case VK_FORMAT_A2R10G10B10_UINT_PACK32:
2252 case VK_FORMAT_A2R10G10B10_SINT_PACK32:
2253 key.vertex_post_shuffle |= 1 << location;
2254 break;
2255 default:
2256 break;
2257 }
2258 }
2259
2260 if (pCreateInfo->pTessellationState)
2261 key.tess_input_vertices = pCreateInfo->pTessellationState->patchControlPoints;
2262
2263
2264 if (pCreateInfo->pMultisampleState &&
2265 pCreateInfo->pMultisampleState->rasterizationSamples > 1) {
2266 uint32_t num_samples = pCreateInfo->pMultisampleState->rasterizationSamples;
2267 uint32_t ps_iter_samples = radv_pipeline_get_ps_iter_samples(pCreateInfo->pMultisampleState);
2268 key.num_samples = num_samples;
2269 key.log2_ps_iter_samples = util_logbase2(ps_iter_samples);
2270 }
2271
2272 key.col_format = blend->spi_shader_col_format;
2273 if (pipeline->device->physical_device->rad_info.chip_class < GFX8)
2274 radv_pipeline_compute_get_int_clamp(pCreateInfo, &key.is_int8, &key.is_int10);
2275
2276 if (pipeline->device->physical_device->rad_info.chip_class >= GFX10)
2277 key.topology = pCreateInfo->pInputAssemblyState->topology;
2278
2279 return key;
2280 }
2281
2282 static bool
2283 radv_nir_stage_uses_xfb(const nir_shader *nir)
2284 {
2285 nir_xfb_info *xfb = nir_gather_xfb_info(nir, NULL);
2286 bool uses_xfb = !!xfb;
2287
2288 ralloc_free(xfb);
2289 return uses_xfb;
2290 }
2291
2292 static void
2293 radv_fill_shader_keys(struct radv_device *device,
2294 struct radv_shader_variant_key *keys,
2295 const struct radv_pipeline_key *key,
2296 nir_shader **nir)
2297 {
2298 keys[MESA_SHADER_VERTEX].vs.instance_rate_inputs = key->instance_rate_inputs;
2299 keys[MESA_SHADER_VERTEX].vs.alpha_adjust = key->vertex_alpha_adjust;
2300 keys[MESA_SHADER_VERTEX].vs.post_shuffle = key->vertex_post_shuffle;
2301 for (unsigned i = 0; i < MAX_VERTEX_ATTRIBS; ++i) {
2302 keys[MESA_SHADER_VERTEX].vs.instance_rate_divisors[i] = key->instance_rate_divisors[i];
2303 keys[MESA_SHADER_VERTEX].vs.vertex_attribute_formats[i] = key->vertex_attribute_formats[i];
2304 keys[MESA_SHADER_VERTEX].vs.vertex_attribute_bindings[i] = key->vertex_attribute_bindings[i];
2305 keys[MESA_SHADER_VERTEX].vs.vertex_attribute_offsets[i] = key->vertex_attribute_offsets[i];
2306 keys[MESA_SHADER_VERTEX].vs.vertex_attribute_strides[i] = key->vertex_attribute_strides[i];
2307 }
2308 keys[MESA_SHADER_VERTEX].vs.outprim = si_conv_prim_to_gs_out(key->topology);
2309
2310 if (nir[MESA_SHADER_TESS_CTRL]) {
2311 keys[MESA_SHADER_VERTEX].vs_common_out.as_ls = true;
2312 keys[MESA_SHADER_TESS_CTRL].tcs.num_inputs = 0;
2313 keys[MESA_SHADER_TESS_CTRL].tcs.input_vertices = key->tess_input_vertices;
2314 keys[MESA_SHADER_TESS_CTRL].tcs.primitive_mode = nir[MESA_SHADER_TESS_EVAL]->info.tess.primitive_mode;
2315
2316 keys[MESA_SHADER_TESS_CTRL].tcs.tes_reads_tess_factors = !!(nir[MESA_SHADER_TESS_EVAL]->info.inputs_read & (VARYING_BIT_TESS_LEVEL_INNER | VARYING_BIT_TESS_LEVEL_OUTER));
2317 }
2318
2319 if (nir[MESA_SHADER_GEOMETRY]) {
2320 if (nir[MESA_SHADER_TESS_CTRL])
2321 keys[MESA_SHADER_TESS_EVAL].vs_common_out.as_es = true;
2322 else
2323 keys[MESA_SHADER_VERTEX].vs_common_out.as_es = true;
2324 }
2325
2326 if (device->physical_device->use_ngg) {
2327 if (nir[MESA_SHADER_TESS_CTRL]) {
2328 keys[MESA_SHADER_TESS_EVAL].vs_common_out.as_ngg = true;
2329 } else {
2330 keys[MESA_SHADER_VERTEX].vs_common_out.as_ngg = true;
2331 }
2332
2333 if (nir[MESA_SHADER_TESS_CTRL] &&
2334 nir[MESA_SHADER_GEOMETRY] &&
2335 nir[MESA_SHADER_GEOMETRY]->info.gs.invocations *
2336 nir[MESA_SHADER_GEOMETRY]->info.gs.vertices_out > 256) {
2337 /* Fallback to the legacy path if tessellation is
2338 * enabled with extreme geometry because
2339 * EN_MAX_VERT_OUT_PER_GS_INSTANCE doesn't work and it
2340 * might hang.
2341 */
2342 keys[MESA_SHADER_TESS_EVAL].vs_common_out.as_ngg = false;
2343 }
2344
2345 /*
2346 * Disable NGG with geometry shaders. There are a bunch of
2347 * issues still:
2348 * * GS primitives in pipeline statistic queries do not get
2349 * updates. See dEQP-VK.query_pool.statistics_query.geometry_shader_primitives
2350 * * General issues with the last primitive missing/corrupt:
2351 * https://bugs.freedesktop.org/show_bug.cgi?id=111248
2352 *
2353 * Furthermore, XGL/AMDVLK also disables this as of 9b632ef.
2354 */
2355 if (nir[MESA_SHADER_GEOMETRY]) {
2356 if (nir[MESA_SHADER_TESS_CTRL])
2357 keys[MESA_SHADER_TESS_EVAL].vs_common_out.as_ngg = false;
2358 else
2359 keys[MESA_SHADER_VERTEX].vs_common_out.as_ngg = false;
2360 }
2361
2362 if (!device->physical_device->use_ngg_streamout) {
2363 gl_shader_stage last_xfb_stage = MESA_SHADER_VERTEX;
2364
2365 for (int i = MESA_SHADER_VERTEX; i <= MESA_SHADER_GEOMETRY; i++) {
2366 if (nir[i])
2367 last_xfb_stage = i;
2368 }
2369
2370 if (nir[last_xfb_stage] &&
2371 radv_nir_stage_uses_xfb(nir[last_xfb_stage])) {
2372 if (nir[MESA_SHADER_TESS_CTRL])
2373 keys[MESA_SHADER_TESS_EVAL].vs_common_out.as_ngg = false;
2374 else
2375 keys[MESA_SHADER_VERTEX].vs_common_out.as_ngg = false;
2376 }
2377 }
2378 }
2379
2380 for(int i = 0; i < MESA_SHADER_STAGES; ++i)
2381 keys[i].has_multiview_view_index = key->has_multiview_view_index;
2382
2383 keys[MESA_SHADER_FRAGMENT].fs.col_format = key->col_format;
2384 keys[MESA_SHADER_FRAGMENT].fs.is_int8 = key->is_int8;
2385 keys[MESA_SHADER_FRAGMENT].fs.is_int10 = key->is_int10;
2386 keys[MESA_SHADER_FRAGMENT].fs.log2_ps_iter_samples = key->log2_ps_iter_samples;
2387 keys[MESA_SHADER_FRAGMENT].fs.num_samples = key->num_samples;
2388 }
2389
2390 static uint8_t
2391 radv_get_wave_size(struct radv_device *device,
2392 gl_shader_stage stage,
2393 const struct radv_shader_variant_key *key)
2394 {
2395 if (stage == MESA_SHADER_GEOMETRY && !key->vs_common_out.as_ngg)
2396 return 64;
2397 else if (stage == MESA_SHADER_COMPUTE)
2398 return device->physical_device->cs_wave_size;
2399 else if (stage == MESA_SHADER_FRAGMENT)
2400 return device->physical_device->ps_wave_size;
2401 else
2402 return device->physical_device->ge_wave_size;
2403 }
2404
2405 static void
2406 radv_fill_shader_info(struct radv_pipeline *pipeline,
2407 struct radv_shader_variant_key *keys,
2408 struct radv_shader_info *infos,
2409 nir_shader **nir)
2410 {
2411 unsigned active_stages = 0;
2412 unsigned filled_stages = 0;
2413
2414 for (int i = 0; i < MESA_SHADER_STAGES; i++) {
2415 if (nir[i])
2416 active_stages |= (1 << i);
2417 }
2418
2419 if (nir[MESA_SHADER_FRAGMENT]) {
2420 radv_nir_shader_info_init(&infos[MESA_SHADER_FRAGMENT]);
2421 radv_nir_shader_info_pass(nir[MESA_SHADER_FRAGMENT],
2422 pipeline->layout,
2423 &keys[MESA_SHADER_FRAGMENT],
2424 &infos[MESA_SHADER_FRAGMENT]);
2425
2426 /* TODO: These are no longer used as keys we should refactor this */
2427 keys[MESA_SHADER_VERTEX].vs_common_out.export_prim_id =
2428 infos[MESA_SHADER_FRAGMENT].ps.prim_id_input;
2429 keys[MESA_SHADER_VERTEX].vs_common_out.export_layer_id =
2430 infos[MESA_SHADER_FRAGMENT].ps.layer_input;
2431 keys[MESA_SHADER_VERTEX].vs_common_out.export_clip_dists =
2432 !!infos[MESA_SHADER_FRAGMENT].ps.num_input_clips_culls;
2433 keys[MESA_SHADER_TESS_EVAL].vs_common_out.export_prim_id =
2434 infos[MESA_SHADER_FRAGMENT].ps.prim_id_input;
2435 keys[MESA_SHADER_TESS_EVAL].vs_common_out.export_layer_id =
2436 infos[MESA_SHADER_FRAGMENT].ps.layer_input;
2437 keys[MESA_SHADER_TESS_EVAL].vs_common_out.export_clip_dists =
2438 !!infos[MESA_SHADER_FRAGMENT].ps.num_input_clips_culls;
2439
2440 filled_stages |= (1 << MESA_SHADER_FRAGMENT);
2441 }
2442
2443 if (pipeline->device->physical_device->rad_info.chip_class >= GFX9 &&
2444 nir[MESA_SHADER_TESS_CTRL]) {
2445 struct nir_shader *combined_nir[] = {nir[MESA_SHADER_VERTEX], nir[MESA_SHADER_TESS_CTRL]};
2446 struct radv_shader_variant_key key = keys[MESA_SHADER_TESS_CTRL];
2447 key.tcs.vs_key = keys[MESA_SHADER_VERTEX].vs;
2448
2449 radv_nir_shader_info_init(&infos[MESA_SHADER_TESS_CTRL]);
2450
2451 for (int i = 0; i < 2; i++) {
2452 radv_nir_shader_info_pass(combined_nir[i],
2453 pipeline->layout, &key,
2454 &infos[MESA_SHADER_TESS_CTRL]);
2455 }
2456
2457 keys[MESA_SHADER_TESS_EVAL].tes.num_patches =
2458 infos[MESA_SHADER_TESS_CTRL].tcs.num_patches;
2459 keys[MESA_SHADER_TESS_EVAL].tes.tcs_num_outputs =
2460 util_last_bit64(infos[MESA_SHADER_TESS_CTRL].tcs.outputs_written);
2461
2462 filled_stages |= (1 << MESA_SHADER_VERTEX);
2463 filled_stages |= (1 << MESA_SHADER_TESS_CTRL);
2464 }
2465
2466 if (pipeline->device->physical_device->rad_info.chip_class >= GFX9 &&
2467 nir[MESA_SHADER_GEOMETRY]) {
2468 gl_shader_stage pre_stage = nir[MESA_SHADER_TESS_EVAL] ? MESA_SHADER_TESS_EVAL : MESA_SHADER_VERTEX;
2469 struct nir_shader *combined_nir[] = {nir[pre_stage], nir[MESA_SHADER_GEOMETRY]};
2470
2471 radv_nir_shader_info_init(&infos[MESA_SHADER_GEOMETRY]);
2472
2473 for (int i = 0; i < 2; i++) {
2474 radv_nir_shader_info_pass(combined_nir[i],
2475 pipeline->layout,
2476 &keys[pre_stage],
2477 &infos[MESA_SHADER_GEOMETRY]);
2478 }
2479
2480 filled_stages |= (1 << pre_stage);
2481 filled_stages |= (1 << MESA_SHADER_GEOMETRY);
2482 }
2483
2484 active_stages ^= filled_stages;
2485 while (active_stages) {
2486 int i = u_bit_scan(&active_stages);
2487
2488 if (i == MESA_SHADER_TESS_CTRL) {
2489 keys[MESA_SHADER_TESS_CTRL].tcs.num_inputs =
2490 util_last_bit64(infos[MESA_SHADER_VERTEX].vs.ls_outputs_written);
2491 }
2492
2493 if (i == MESA_SHADER_TESS_EVAL) {
2494 keys[MESA_SHADER_TESS_EVAL].tes.num_patches =
2495 infos[MESA_SHADER_TESS_CTRL].tcs.num_patches;
2496 keys[MESA_SHADER_TESS_EVAL].tes.tcs_num_outputs =
2497 util_last_bit64(infos[MESA_SHADER_TESS_CTRL].tcs.outputs_written);
2498 }
2499
2500 radv_nir_shader_info_init(&infos[i]);
2501 radv_nir_shader_info_pass(nir[i], pipeline->layout,
2502 &keys[i], &infos[i]);
2503 }
2504
2505 for (int i = 0; i < MESA_SHADER_STAGES; i++) {
2506 if (nir[i])
2507 infos[i].wave_size =
2508 radv_get_wave_size(pipeline->device, i, &keys[i]);
2509 }
2510 }
2511
2512 static void
2513 merge_tess_info(struct shader_info *tes_info,
2514 const struct shader_info *tcs_info)
2515 {
2516 /* The Vulkan 1.0.38 spec, section 21.1 Tessellator says:
2517 *
2518 * "PointMode. Controls generation of points rather than triangles
2519 * or lines. This functionality defaults to disabled, and is
2520 * enabled if either shader stage includes the execution mode.
2521 *
2522 * and about Triangles, Quads, IsoLines, VertexOrderCw, VertexOrderCcw,
2523 * PointMode, SpacingEqual, SpacingFractionalEven, SpacingFractionalOdd,
2524 * and OutputVertices, it says:
2525 *
2526 * "One mode must be set in at least one of the tessellation
2527 * shader stages."
2528 *
2529 * So, the fields can be set in either the TCS or TES, but they must
2530 * agree if set in both. Our backend looks at TES, so bitwise-or in
2531 * the values from the TCS.
2532 */
2533 assert(tcs_info->tess.tcs_vertices_out == 0 ||
2534 tes_info->tess.tcs_vertices_out == 0 ||
2535 tcs_info->tess.tcs_vertices_out == tes_info->tess.tcs_vertices_out);
2536 tes_info->tess.tcs_vertices_out |= tcs_info->tess.tcs_vertices_out;
2537
2538 assert(tcs_info->tess.spacing == TESS_SPACING_UNSPECIFIED ||
2539 tes_info->tess.spacing == TESS_SPACING_UNSPECIFIED ||
2540 tcs_info->tess.spacing == tes_info->tess.spacing);
2541 tes_info->tess.spacing |= tcs_info->tess.spacing;
2542
2543 assert(tcs_info->tess.primitive_mode == 0 ||
2544 tes_info->tess.primitive_mode == 0 ||
2545 tcs_info->tess.primitive_mode == tes_info->tess.primitive_mode);
2546 tes_info->tess.primitive_mode |= tcs_info->tess.primitive_mode;
2547 tes_info->tess.ccw |= tcs_info->tess.ccw;
2548 tes_info->tess.point_mode |= tcs_info->tess.point_mode;
2549 }
2550
2551 static
2552 void radv_init_feedback(const VkPipelineCreationFeedbackCreateInfoEXT *ext)
2553 {
2554 if (!ext)
2555 return;
2556
2557 if (ext->pPipelineCreationFeedback) {
2558 ext->pPipelineCreationFeedback->flags = 0;
2559 ext->pPipelineCreationFeedback->duration = 0;
2560 }
2561
2562 for (unsigned i = 0; i < ext->pipelineStageCreationFeedbackCount; ++i) {
2563 ext->pPipelineStageCreationFeedbacks[i].flags = 0;
2564 ext->pPipelineStageCreationFeedbacks[i].duration = 0;
2565 }
2566 }
2567
2568 static
2569 void radv_start_feedback(VkPipelineCreationFeedbackEXT *feedback)
2570 {
2571 if (!feedback)
2572 return;
2573
2574 feedback->duration -= radv_get_current_time();
2575 feedback ->flags = VK_PIPELINE_CREATION_FEEDBACK_VALID_BIT_EXT;
2576 }
2577
2578 static
2579 void radv_stop_feedback(VkPipelineCreationFeedbackEXT *feedback, bool cache_hit)
2580 {
2581 if (!feedback)
2582 return;
2583
2584 feedback->duration += radv_get_current_time();
2585 feedback ->flags = VK_PIPELINE_CREATION_FEEDBACK_VALID_BIT_EXT |
2586 (cache_hit ? VK_PIPELINE_CREATION_FEEDBACK_APPLICATION_PIPELINE_CACHE_HIT_BIT_EXT : 0);
2587 }
2588
2589 static
2590 bool radv_aco_supported_stage(gl_shader_stage stage, bool has_gs, bool has_ts)
2591 {
2592 return (stage == MESA_SHADER_VERTEX && !has_gs && !has_ts) ||
2593 stage == MESA_SHADER_FRAGMENT ||
2594 stage == MESA_SHADER_COMPUTE;
2595 }
2596
2597 void radv_create_shaders(struct radv_pipeline *pipeline,
2598 struct radv_device *device,
2599 struct radv_pipeline_cache *cache,
2600 const struct radv_pipeline_key *key,
2601 const VkPipelineShaderStageCreateInfo **pStages,
2602 const VkPipelineCreateFlags flags,
2603 VkPipelineCreationFeedbackEXT *pipeline_feedback,
2604 VkPipelineCreationFeedbackEXT **stage_feedbacks)
2605 {
2606 struct radv_shader_module fs_m = {0};
2607 struct radv_shader_module *modules[MESA_SHADER_STAGES] = { 0, };
2608 nir_shader *nir[MESA_SHADER_STAGES] = {0};
2609 struct radv_shader_binary *binaries[MESA_SHADER_STAGES] = {NULL};
2610 struct radv_shader_variant_key keys[MESA_SHADER_STAGES] = {{{{{0}}}}};
2611 struct radv_shader_info infos[MESA_SHADER_STAGES] = {0};
2612 unsigned char hash[20], gs_copy_hash[20];
2613 bool keep_executable_info = (flags & VK_PIPELINE_CREATE_CAPTURE_INTERNAL_REPRESENTATIONS_BIT_KHR) || device->keep_shader_info;
2614
2615 radv_start_feedback(pipeline_feedback);
2616
2617 for (unsigned i = 0; i < MESA_SHADER_STAGES; ++i) {
2618 if (pStages[i]) {
2619 modules[i] = radv_shader_module_from_handle(pStages[i]->module);
2620 if (modules[i]->nir)
2621 _mesa_sha1_compute(modules[i]->nir->info.name,
2622 strlen(modules[i]->nir->info.name),
2623 modules[i]->sha1);
2624
2625 pipeline->active_stages |= mesa_to_vk_shader_stage(i);
2626 }
2627 }
2628
2629 radv_hash_shaders(hash, pStages, pipeline->layout, key, get_hash_flags(device));
2630 memcpy(gs_copy_hash, hash, 20);
2631 gs_copy_hash[0] ^= 1;
2632
2633 bool found_in_application_cache = true;
2634 if (modules[MESA_SHADER_GEOMETRY] && !keep_executable_info) {
2635 struct radv_shader_variant *variants[MESA_SHADER_STAGES] = {0};
2636 radv_create_shader_variants_from_pipeline_cache(device, cache, gs_copy_hash, variants,
2637 &found_in_application_cache);
2638 pipeline->gs_copy_shader = variants[MESA_SHADER_GEOMETRY];
2639 }
2640
2641 if (!keep_executable_info &&
2642 radv_create_shader_variants_from_pipeline_cache(device, cache, hash, pipeline->shaders,
2643 &found_in_application_cache) &&
2644 (!modules[MESA_SHADER_GEOMETRY] || pipeline->gs_copy_shader)) {
2645 radv_stop_feedback(pipeline_feedback, found_in_application_cache);
2646 return;
2647 }
2648
2649 if (!modules[MESA_SHADER_FRAGMENT] && !modules[MESA_SHADER_COMPUTE]) {
2650 nir_builder fs_b;
2651 nir_builder_init_simple_shader(&fs_b, NULL, MESA_SHADER_FRAGMENT, NULL);
2652 fs_b.shader->info.name = ralloc_strdup(fs_b.shader, "noop_fs");
2653 fs_m.nir = fs_b.shader;
2654 modules[MESA_SHADER_FRAGMENT] = &fs_m;
2655 }
2656
2657 bool has_gs = modules[MESA_SHADER_GEOMETRY];
2658 bool has_ts = modules[MESA_SHADER_TESS_CTRL] || modules[MESA_SHADER_TESS_EVAL];
2659 bool use_aco = device->physical_device->use_aco;
2660
2661 for (unsigned i = 0; i < MESA_SHADER_STAGES; ++i) {
2662 const VkPipelineShaderStageCreateInfo *stage = pStages[i];
2663
2664 if (!modules[i])
2665 continue;
2666
2667 radv_start_feedback(stage_feedbacks[i]);
2668
2669 bool aco = use_aco && radv_aco_supported_stage(i, has_gs, has_ts);
2670 nir[i] = radv_shader_compile_to_nir(device, modules[i],
2671 stage ? stage->pName : "main", i,
2672 stage ? stage->pSpecializationInfo : NULL,
2673 flags, pipeline->layout, aco);
2674
2675 /* We don't want to alter meta shaders IR directly so clone it
2676 * first.
2677 */
2678 if (nir[i]->info.name) {
2679 nir[i] = nir_shader_clone(NULL, nir[i]);
2680 }
2681
2682 radv_stop_feedback(stage_feedbacks[i], false);
2683 }
2684
2685 if (nir[MESA_SHADER_TESS_CTRL]) {
2686 nir_lower_patch_vertices(nir[MESA_SHADER_TESS_EVAL], nir[MESA_SHADER_TESS_CTRL]->info.tess.tcs_vertices_out, NULL);
2687 merge_tess_info(&nir[MESA_SHADER_TESS_EVAL]->info, &nir[MESA_SHADER_TESS_CTRL]->info);
2688 }
2689
2690 if (!(flags & VK_PIPELINE_CREATE_DISABLE_OPTIMIZATION_BIT))
2691 radv_link_shaders(pipeline, nir);
2692
2693 for (int i = 0; i < MESA_SHADER_STAGES; ++i) {
2694 if (nir[i]) {
2695 NIR_PASS_V(nir[i], nir_lower_non_uniform_access,
2696 nir_lower_non_uniform_ubo_access |
2697 nir_lower_non_uniform_ssbo_access |
2698 nir_lower_non_uniform_texture_access |
2699 nir_lower_non_uniform_image_access);
2700
2701 bool aco = use_aco && radv_aco_supported_stage(i, has_gs, has_ts);
2702 if (!aco)
2703 NIR_PASS_V(nir[i], nir_lower_bool_to_int32);
2704 }
2705
2706 if (radv_can_dump_shader(device, modules[i], false))
2707 nir_print_shader(nir[i], stderr);
2708 }
2709
2710 if (nir[MESA_SHADER_FRAGMENT])
2711 radv_lower_fs_io(nir[MESA_SHADER_FRAGMENT]);
2712
2713 radv_fill_shader_keys(device, keys, key, nir);
2714
2715 radv_fill_shader_info(pipeline, keys, infos, nir);
2716
2717 if ((nir[MESA_SHADER_VERTEX] &&
2718 keys[MESA_SHADER_VERTEX].vs_common_out.as_ngg) ||
2719 (nir[MESA_SHADER_TESS_EVAL] &&
2720 keys[MESA_SHADER_TESS_EVAL].vs_common_out.as_ngg)) {
2721 struct gfx10_ngg_info *ngg_info;
2722
2723 if (nir[MESA_SHADER_GEOMETRY])
2724 ngg_info = &infos[MESA_SHADER_GEOMETRY].ngg_info;
2725 else if (nir[MESA_SHADER_TESS_CTRL])
2726 ngg_info = &infos[MESA_SHADER_TESS_EVAL].ngg_info;
2727 else
2728 ngg_info = &infos[MESA_SHADER_VERTEX].ngg_info;
2729
2730 gfx10_get_ngg_info(key, pipeline, nir, infos, ngg_info);
2731 } else if (nir[MESA_SHADER_GEOMETRY]) {
2732 struct gfx9_gs_info *gs_info =
2733 &infos[MESA_SHADER_GEOMETRY].gs_ring_info;
2734
2735 gfx9_get_gs_info(key, pipeline, nir, infos, gs_info);
2736 }
2737
2738 if (nir[MESA_SHADER_FRAGMENT]) {
2739 if (!pipeline->shaders[MESA_SHADER_FRAGMENT]) {
2740 radv_start_feedback(stage_feedbacks[MESA_SHADER_FRAGMENT]);
2741
2742 bool aco = use_aco && radv_aco_supported_stage(MESA_SHADER_FRAGMENT, has_gs, has_ts);
2743 pipeline->shaders[MESA_SHADER_FRAGMENT] =
2744 radv_shader_variant_compile(device, modules[MESA_SHADER_FRAGMENT], &nir[MESA_SHADER_FRAGMENT], 1,
2745 pipeline->layout, keys + MESA_SHADER_FRAGMENT,
2746 infos + MESA_SHADER_FRAGMENT,
2747 keep_executable_info, aco,
2748 &binaries[MESA_SHADER_FRAGMENT]);
2749
2750 radv_stop_feedback(stage_feedbacks[MESA_SHADER_FRAGMENT], false);
2751 }
2752
2753 /* TODO: These are no longer used as keys we should refactor this */
2754 keys[MESA_SHADER_VERTEX].vs_common_out.export_prim_id =
2755 pipeline->shaders[MESA_SHADER_FRAGMENT]->info.ps.prim_id_input;
2756 keys[MESA_SHADER_VERTEX].vs_common_out.export_layer_id =
2757 pipeline->shaders[MESA_SHADER_FRAGMENT]->info.ps.layer_input;
2758 keys[MESA_SHADER_VERTEX].vs_common_out.export_clip_dists =
2759 !!pipeline->shaders[MESA_SHADER_FRAGMENT]->info.ps.num_input_clips_culls;
2760 keys[MESA_SHADER_TESS_EVAL].vs_common_out.export_prim_id =
2761 pipeline->shaders[MESA_SHADER_FRAGMENT]->info.ps.prim_id_input;
2762 keys[MESA_SHADER_TESS_EVAL].vs_common_out.export_layer_id =
2763 pipeline->shaders[MESA_SHADER_FRAGMENT]->info.ps.layer_input;
2764 keys[MESA_SHADER_TESS_EVAL].vs_common_out.export_clip_dists =
2765 !!pipeline->shaders[MESA_SHADER_FRAGMENT]->info.ps.num_input_clips_culls;
2766 }
2767
2768 if (device->physical_device->rad_info.chip_class >= GFX9 && modules[MESA_SHADER_TESS_CTRL]) {
2769 if (!pipeline->shaders[MESA_SHADER_TESS_CTRL]) {
2770 struct nir_shader *combined_nir[] = {nir[MESA_SHADER_VERTEX], nir[MESA_SHADER_TESS_CTRL]};
2771 struct radv_shader_variant_key key = keys[MESA_SHADER_TESS_CTRL];
2772 key.tcs.vs_key = keys[MESA_SHADER_VERTEX].vs;
2773
2774 radv_start_feedback(stage_feedbacks[MESA_SHADER_TESS_CTRL]);
2775
2776 pipeline->shaders[MESA_SHADER_TESS_CTRL] = radv_shader_variant_compile(device, modules[MESA_SHADER_TESS_CTRL], combined_nir, 2,
2777 pipeline->layout,
2778 &key, &infos[MESA_SHADER_TESS_CTRL], keep_executable_info,
2779 false, &binaries[MESA_SHADER_TESS_CTRL]);
2780
2781 radv_stop_feedback(stage_feedbacks[MESA_SHADER_TESS_CTRL], false);
2782 }
2783 modules[MESA_SHADER_VERTEX] = NULL;
2784 keys[MESA_SHADER_TESS_EVAL].tes.num_patches = pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.tcs.num_patches;
2785 keys[MESA_SHADER_TESS_EVAL].tes.tcs_num_outputs = util_last_bit64(pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.tcs.outputs_written);
2786 }
2787
2788 if (device->physical_device->rad_info.chip_class >= GFX9 && modules[MESA_SHADER_GEOMETRY]) {
2789 gl_shader_stage pre_stage = modules[MESA_SHADER_TESS_EVAL] ? MESA_SHADER_TESS_EVAL : MESA_SHADER_VERTEX;
2790 if (!pipeline->shaders[MESA_SHADER_GEOMETRY]) {
2791 struct nir_shader *combined_nir[] = {nir[pre_stage], nir[MESA_SHADER_GEOMETRY]};
2792
2793 radv_start_feedback(stage_feedbacks[MESA_SHADER_GEOMETRY]);
2794
2795 pipeline->shaders[MESA_SHADER_GEOMETRY] = radv_shader_variant_compile(device, modules[MESA_SHADER_GEOMETRY], combined_nir, 2,
2796 pipeline->layout,
2797 &keys[pre_stage], &infos[MESA_SHADER_GEOMETRY], keep_executable_info,
2798 false, &binaries[MESA_SHADER_GEOMETRY]);
2799
2800 radv_stop_feedback(stage_feedbacks[MESA_SHADER_GEOMETRY], false);
2801 }
2802 modules[pre_stage] = NULL;
2803 }
2804
2805 for (int i = 0; i < MESA_SHADER_STAGES; ++i) {
2806 if(modules[i] && !pipeline->shaders[i]) {
2807 if (i == MESA_SHADER_TESS_CTRL) {
2808 keys[MESA_SHADER_TESS_CTRL].tcs.num_inputs = util_last_bit64(pipeline->shaders[MESA_SHADER_VERTEX]->info.vs.ls_outputs_written);
2809 }
2810 if (i == MESA_SHADER_TESS_EVAL) {
2811 keys[MESA_SHADER_TESS_EVAL].tes.num_patches = pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.tcs.num_patches;
2812 keys[MESA_SHADER_TESS_EVAL].tes.tcs_num_outputs = util_last_bit64(pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.tcs.outputs_written);
2813 }
2814
2815 radv_start_feedback(stage_feedbacks[i]);
2816
2817 bool aco = use_aco && radv_aco_supported_stage(i, has_gs, has_ts);
2818 pipeline->shaders[i] = radv_shader_variant_compile(device, modules[i], &nir[i], 1,
2819 pipeline->layout,
2820 keys + i, infos + i,keep_executable_info,
2821 aco, &binaries[i]);
2822
2823 radv_stop_feedback(stage_feedbacks[i], false);
2824 }
2825 }
2826
2827 if(modules[MESA_SHADER_GEOMETRY]) {
2828 struct radv_shader_binary *gs_copy_binary = NULL;
2829 if (!pipeline->gs_copy_shader &&
2830 !radv_pipeline_has_ngg(pipeline)) {
2831 struct radv_shader_info info = {};
2832 struct radv_shader_variant_key key = {};
2833
2834 key.has_multiview_view_index =
2835 keys[MESA_SHADER_GEOMETRY].has_multiview_view_index;
2836
2837 radv_nir_shader_info_pass(nir[MESA_SHADER_GEOMETRY],
2838 pipeline->layout, &key,
2839 &info);
2840 info.wave_size = 64; /* Wave32 not supported. */
2841
2842 pipeline->gs_copy_shader = radv_create_gs_copy_shader(
2843 device, nir[MESA_SHADER_GEOMETRY], &info,
2844 &gs_copy_binary, keep_executable_info,
2845 keys[MESA_SHADER_GEOMETRY].has_multiview_view_index);
2846 }
2847
2848 if (!keep_executable_info && pipeline->gs_copy_shader) {
2849 struct radv_shader_binary *binaries[MESA_SHADER_STAGES] = {NULL};
2850 struct radv_shader_variant *variants[MESA_SHADER_STAGES] = {0};
2851
2852 binaries[MESA_SHADER_GEOMETRY] = gs_copy_binary;
2853 variants[MESA_SHADER_GEOMETRY] = pipeline->gs_copy_shader;
2854
2855 radv_pipeline_cache_insert_shaders(device, cache,
2856 gs_copy_hash,
2857 variants,
2858 binaries);
2859 }
2860 free(gs_copy_binary);
2861 }
2862
2863 if (!keep_executable_info) {
2864 radv_pipeline_cache_insert_shaders(device, cache, hash, pipeline->shaders,
2865 binaries);
2866 }
2867
2868 for (int i = 0; i < MESA_SHADER_STAGES; ++i) {
2869 free(binaries[i]);
2870 if (nir[i]) {
2871 ralloc_free(nir[i]);
2872
2873 if (radv_can_dump_shader_stats(device, modules[i]))
2874 radv_shader_dump_stats(device,
2875 pipeline->shaders[i],
2876 i, stderr);
2877 }
2878 }
2879
2880 if (fs_m.nir)
2881 ralloc_free(fs_m.nir);
2882
2883 radv_stop_feedback(pipeline_feedback, false);
2884 }
2885
2886 static uint32_t
2887 radv_pipeline_stage_to_user_data_0(struct radv_pipeline *pipeline,
2888 gl_shader_stage stage, enum chip_class chip_class)
2889 {
2890 bool has_gs = radv_pipeline_has_gs(pipeline);
2891 bool has_tess = radv_pipeline_has_tess(pipeline);
2892 bool has_ngg = radv_pipeline_has_ngg(pipeline);
2893
2894 switch (stage) {
2895 case MESA_SHADER_FRAGMENT:
2896 return R_00B030_SPI_SHADER_USER_DATA_PS_0;
2897 case MESA_SHADER_VERTEX:
2898 if (has_tess) {
2899 if (chip_class >= GFX10) {
2900 return R_00B430_SPI_SHADER_USER_DATA_HS_0;
2901 } else if (chip_class == GFX9) {
2902 return R_00B430_SPI_SHADER_USER_DATA_LS_0;
2903 } else {
2904 return R_00B530_SPI_SHADER_USER_DATA_LS_0;
2905 }
2906
2907 }
2908
2909 if (has_gs) {
2910 if (chip_class >= GFX10) {
2911 return R_00B230_SPI_SHADER_USER_DATA_GS_0;
2912 } else {
2913 return R_00B330_SPI_SHADER_USER_DATA_ES_0;
2914 }
2915 }
2916
2917 if (has_ngg)
2918 return R_00B230_SPI_SHADER_USER_DATA_GS_0;
2919
2920 return R_00B130_SPI_SHADER_USER_DATA_VS_0;
2921 case MESA_SHADER_GEOMETRY:
2922 return chip_class == GFX9 ? R_00B330_SPI_SHADER_USER_DATA_ES_0 :
2923 R_00B230_SPI_SHADER_USER_DATA_GS_0;
2924 case MESA_SHADER_COMPUTE:
2925 return R_00B900_COMPUTE_USER_DATA_0;
2926 case MESA_SHADER_TESS_CTRL:
2927 return chip_class == GFX9 ? R_00B430_SPI_SHADER_USER_DATA_LS_0 :
2928 R_00B430_SPI_SHADER_USER_DATA_HS_0;
2929 case MESA_SHADER_TESS_EVAL:
2930 if (has_gs) {
2931 return chip_class >= GFX10 ? R_00B230_SPI_SHADER_USER_DATA_GS_0 :
2932 R_00B330_SPI_SHADER_USER_DATA_ES_0;
2933 } else if (has_ngg) {
2934 return R_00B230_SPI_SHADER_USER_DATA_GS_0;
2935 } else {
2936 return R_00B130_SPI_SHADER_USER_DATA_VS_0;
2937 }
2938 default:
2939 unreachable("unknown shader");
2940 }
2941 }
2942
2943 struct radv_bin_size_entry {
2944 unsigned bpp;
2945 VkExtent2D extent;
2946 };
2947
2948 static VkExtent2D
2949 radv_gfx9_compute_bin_size(struct radv_pipeline *pipeline, const VkGraphicsPipelineCreateInfo *pCreateInfo)
2950 {
2951 static const struct radv_bin_size_entry color_size_table[][3][9] = {
2952 {
2953 /* One RB / SE */
2954 {
2955 /* One shader engine */
2956 { 0, {128, 128}},
2957 { 1, { 64, 128}},
2958 { 2, { 32, 128}},
2959 { 3, { 16, 128}},
2960 { 17, { 0, 0}},
2961 { UINT_MAX, { 0, 0}},
2962 },
2963 {
2964 /* Two shader engines */
2965 { 0, {128, 128}},
2966 { 2, { 64, 128}},
2967 { 3, { 32, 128}},
2968 { 5, { 16, 128}},
2969 { 17, { 0, 0}},
2970 { UINT_MAX, { 0, 0}},
2971 },
2972 {
2973 /* Four shader engines */
2974 { 0, {128, 128}},
2975 { 3, { 64, 128}},
2976 { 5, { 16, 128}},
2977 { 17, { 0, 0}},
2978 { UINT_MAX, { 0, 0}},
2979 },
2980 },
2981 {
2982 /* Two RB / SE */
2983 {
2984 /* One shader engine */
2985 { 0, {128, 128}},
2986 { 2, { 64, 128}},
2987 { 3, { 32, 128}},
2988 { 5, { 16, 128}},
2989 { 33, { 0, 0}},
2990 { UINT_MAX, { 0, 0}},
2991 },
2992 {
2993 /* Two shader engines */
2994 { 0, {128, 128}},
2995 { 3, { 64, 128}},
2996 { 5, { 32, 128}},
2997 { 9, { 16, 128}},
2998 { 33, { 0, 0}},
2999 { UINT_MAX, { 0, 0}},
3000 },
3001 {
3002 /* Four shader engines */
3003 { 0, {256, 256}},
3004 { 2, {128, 256}},
3005 { 3, {128, 128}},
3006 { 5, { 64, 128}},
3007 { 9, { 16, 128}},
3008 { 33, { 0, 0}},
3009 { UINT_MAX, { 0, 0}},
3010 },
3011 },
3012 {
3013 /* Four RB / SE */
3014 {
3015 /* One shader engine */
3016 { 0, {128, 256}},
3017 { 2, {128, 128}},
3018 { 3, { 64, 128}},
3019 { 5, { 32, 128}},
3020 { 9, { 16, 128}},
3021 { 33, { 0, 0}},
3022 { UINT_MAX, { 0, 0}},
3023 },
3024 {
3025 /* Two shader engines */
3026 { 0, {256, 256}},
3027 { 2, {128, 256}},
3028 { 3, {128, 128}},
3029 { 5, { 64, 128}},
3030 { 9, { 32, 128}},
3031 { 17, { 16, 128}},
3032 { 33, { 0, 0}},
3033 { UINT_MAX, { 0, 0}},
3034 },
3035 {
3036 /* Four shader engines */
3037 { 0, {256, 512}},
3038 { 2, {256, 256}},
3039 { 3, {128, 256}},
3040 { 5, {128, 128}},
3041 { 9, { 64, 128}},
3042 { 17, { 16, 128}},
3043 { 33, { 0, 0}},
3044 { UINT_MAX, { 0, 0}},
3045 },
3046 },
3047 };
3048 static const struct radv_bin_size_entry ds_size_table[][3][9] = {
3049 {
3050 // One RB / SE
3051 {
3052 // One shader engine
3053 { 0, {128, 256}},
3054 { 2, {128, 128}},
3055 { 4, { 64, 128}},
3056 { 7, { 32, 128}},
3057 { 13, { 16, 128}},
3058 { 49, { 0, 0}},
3059 { UINT_MAX, { 0, 0}},
3060 },
3061 {
3062 // Two shader engines
3063 { 0, {256, 256}},
3064 { 2, {128, 256}},
3065 { 4, {128, 128}},
3066 { 7, { 64, 128}},
3067 { 13, { 32, 128}},
3068 { 25, { 16, 128}},
3069 { 49, { 0, 0}},
3070 { UINT_MAX, { 0, 0}},
3071 },
3072 {
3073 // Four shader engines
3074 { 0, {256, 512}},
3075 { 2, {256, 256}},
3076 { 4, {128, 256}},
3077 { 7, {128, 128}},
3078 { 13, { 64, 128}},
3079 { 25, { 16, 128}},
3080 { 49, { 0, 0}},
3081 { UINT_MAX, { 0, 0}},
3082 },
3083 },
3084 {
3085 // Two RB / SE
3086 {
3087 // One shader engine
3088 { 0, {256, 256}},
3089 { 2, {128, 256}},
3090 { 4, {128, 128}},
3091 { 7, { 64, 128}},
3092 { 13, { 32, 128}},
3093 { 25, { 16, 128}},
3094 { 97, { 0, 0}},
3095 { UINT_MAX, { 0, 0}},
3096 },
3097 {
3098 // Two shader engines
3099 { 0, {256, 512}},
3100 { 2, {256, 256}},
3101 { 4, {128, 256}},
3102 { 7, {128, 128}},
3103 { 13, { 64, 128}},
3104 { 25, { 32, 128}},
3105 { 49, { 16, 128}},
3106 { 97, { 0, 0}},
3107 { UINT_MAX, { 0, 0}},
3108 },
3109 {
3110 // Four shader engines
3111 { 0, {512, 512}},
3112 { 2, {256, 512}},
3113 { 4, {256, 256}},
3114 { 7, {128, 256}},
3115 { 13, {128, 128}},
3116 { 25, { 64, 128}},
3117 { 49, { 16, 128}},
3118 { 97, { 0, 0}},
3119 { UINT_MAX, { 0, 0}},
3120 },
3121 },
3122 {
3123 // Four RB / SE
3124 {
3125 // One shader engine
3126 { 0, {256, 512}},
3127 { 2, {256, 256}},
3128 { 4, {128, 256}},
3129 { 7, {128, 128}},
3130 { 13, { 64, 128}},
3131 { 25, { 32, 128}},
3132 { 49, { 16, 128}},
3133 { UINT_MAX, { 0, 0}},
3134 },
3135 {
3136 // Two shader engines
3137 { 0, {512, 512}},
3138 { 2, {256, 512}},
3139 { 4, {256, 256}},
3140 { 7, {128, 256}},
3141 { 13, {128, 128}},
3142 { 25, { 64, 128}},
3143 { 49, { 32, 128}},
3144 { 97, { 16, 128}},
3145 { UINT_MAX, { 0, 0}},
3146 },
3147 {
3148 // Four shader engines
3149 { 0, {512, 512}},
3150 { 4, {256, 512}},
3151 { 7, {256, 256}},
3152 { 13, {128, 256}},
3153 { 25, {128, 128}},
3154 { 49, { 64, 128}},
3155 { 97, { 16, 128}},
3156 { UINT_MAX, { 0, 0}},
3157 },
3158 },
3159 };
3160
3161 RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass);
3162 struct radv_subpass *subpass = pass->subpasses + pCreateInfo->subpass;
3163 VkExtent2D extent = {512, 512};
3164
3165 unsigned log_num_rb_per_se =
3166 util_logbase2_ceil(pipeline->device->physical_device->rad_info.num_render_backends /
3167 pipeline->device->physical_device->rad_info.max_se);
3168 unsigned log_num_se = util_logbase2_ceil(pipeline->device->physical_device->rad_info.max_se);
3169
3170 unsigned total_samples = 1u << G_028BE0_MSAA_NUM_SAMPLES(pipeline->graphics.ms.pa_sc_aa_config);
3171 unsigned ps_iter_samples = 1u << G_028804_PS_ITER_SAMPLES(pipeline->graphics.ms.db_eqaa);
3172 unsigned effective_samples = total_samples;
3173 unsigned color_bytes_per_pixel = 0;
3174
3175 const VkPipelineColorBlendStateCreateInfo *vkblend = pCreateInfo->pColorBlendState;
3176 if (vkblend) {
3177 for (unsigned i = 0; i < subpass->color_count; i++) {
3178 if (!vkblend->pAttachments[i].colorWriteMask)
3179 continue;
3180
3181 if (subpass->color_attachments[i].attachment == VK_ATTACHMENT_UNUSED)
3182 continue;
3183
3184 VkFormat format = pass->attachments[subpass->color_attachments[i].attachment].format;
3185 color_bytes_per_pixel += vk_format_get_blocksize(format);
3186 }
3187
3188 /* MSAA images typically don't use all samples all the time. */
3189 if (effective_samples >= 2 && ps_iter_samples <= 1)
3190 effective_samples = 2;
3191 color_bytes_per_pixel *= effective_samples;
3192 }
3193
3194 const struct radv_bin_size_entry *color_entry = color_size_table[log_num_rb_per_se][log_num_se];
3195 while(color_entry[1].bpp <= color_bytes_per_pixel)
3196 ++color_entry;
3197
3198 extent = color_entry->extent;
3199
3200 if (subpass->depth_stencil_attachment) {
3201 struct radv_render_pass_attachment *attachment = pass->attachments + subpass->depth_stencil_attachment->attachment;
3202
3203 /* Coefficients taken from AMDVLK */
3204 unsigned depth_coeff = vk_format_is_depth(attachment->format) ? 5 : 0;
3205 unsigned stencil_coeff = vk_format_is_stencil(attachment->format) ? 1 : 0;
3206 unsigned ds_bytes_per_pixel = 4 * (depth_coeff + stencil_coeff) * total_samples;
3207
3208 const struct radv_bin_size_entry *ds_entry = ds_size_table[log_num_rb_per_se][log_num_se];
3209 while(ds_entry[1].bpp <= ds_bytes_per_pixel)
3210 ++ds_entry;
3211
3212 if (ds_entry->extent.width * ds_entry->extent.height < extent.width * extent.height)
3213 extent = ds_entry->extent;
3214 }
3215
3216 return extent;
3217 }
3218
3219 static VkExtent2D
3220 radv_gfx10_compute_bin_size(struct radv_pipeline *pipeline, const VkGraphicsPipelineCreateInfo *pCreateInfo)
3221 {
3222 RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass);
3223 struct radv_subpass *subpass = pass->subpasses + pCreateInfo->subpass;
3224 VkExtent2D extent = {512, 512};
3225
3226 const unsigned db_tag_size = 64;
3227 const unsigned db_tag_count = 312;
3228 const unsigned color_tag_size = 1024;
3229 const unsigned color_tag_count = 31;
3230 const unsigned fmask_tag_size = 256;
3231 const unsigned fmask_tag_count = 44;
3232
3233 const unsigned rb_count = pipeline->device->physical_device->rad_info.num_render_backends;
3234 const unsigned pipe_count = MAX2(rb_count, pipeline->device->physical_device->rad_info.num_sdp_interfaces);
3235
3236 const unsigned db_tag_part = (db_tag_count * rb_count / pipe_count) * db_tag_size * pipe_count;
3237 const unsigned color_tag_part = (color_tag_count * rb_count / pipe_count) * color_tag_size * pipe_count;
3238 const unsigned fmask_tag_part = (fmask_tag_count * rb_count / pipe_count) * fmask_tag_size * pipe_count;
3239
3240 const unsigned total_samples = 1u << G_028BE0_MSAA_NUM_SAMPLES(pipeline->graphics.ms.pa_sc_aa_config);
3241 const unsigned samples_log = util_logbase2_ceil(total_samples);
3242
3243 unsigned color_bytes_per_pixel = 0;
3244 unsigned fmask_bytes_per_pixel = 0;
3245
3246 const VkPipelineColorBlendStateCreateInfo *vkblend = pCreateInfo->pColorBlendState;
3247 if (vkblend) {
3248 for (unsigned i = 0; i < subpass->color_count; i++) {
3249 if (!vkblend->pAttachments[i].colorWriteMask)
3250 continue;
3251
3252 if (subpass->color_attachments[i].attachment == VK_ATTACHMENT_UNUSED)
3253 continue;
3254
3255 VkFormat format = pass->attachments[subpass->color_attachments[i].attachment].format;
3256 color_bytes_per_pixel += vk_format_get_blocksize(format);
3257
3258 if (total_samples > 1) {
3259 assert(samples_log <= 3);
3260 const unsigned fmask_array[] = {0, 1, 1, 4};
3261 fmask_bytes_per_pixel += fmask_array[samples_log];
3262 }
3263 }
3264
3265 color_bytes_per_pixel *= total_samples;
3266 }
3267 color_bytes_per_pixel = MAX2(color_bytes_per_pixel, 1);
3268
3269 const unsigned color_pixel_count_log = util_logbase2(color_tag_part / color_bytes_per_pixel);
3270 extent.width = 1ull << ((color_pixel_count_log + 1) / 2);
3271 extent.height = 1ull << (color_pixel_count_log / 2);
3272
3273 if (fmask_bytes_per_pixel) {
3274 const unsigned fmask_pixel_count_log = util_logbase2(fmask_tag_part / fmask_bytes_per_pixel);
3275
3276 const VkExtent2D fmask_extent = (VkExtent2D){
3277 .width = 1ull << ((fmask_pixel_count_log + 1) / 2),
3278 .height = 1ull << (color_pixel_count_log / 2)
3279 };
3280
3281 if (fmask_extent.width * fmask_extent.height < extent.width * extent.height)
3282 extent = fmask_extent;
3283 }
3284
3285 if (subpass->depth_stencil_attachment) {
3286 struct radv_render_pass_attachment *attachment = pass->attachments + subpass->depth_stencil_attachment->attachment;
3287
3288 /* Coefficients taken from AMDVLK */
3289 unsigned depth_coeff = vk_format_is_depth(attachment->format) ? 5 : 0;
3290 unsigned stencil_coeff = vk_format_is_stencil(attachment->format) ? 1 : 0;
3291 unsigned db_bytes_per_pixel = (depth_coeff + stencil_coeff) * total_samples;
3292
3293 const unsigned db_pixel_count_log = util_logbase2(db_tag_part / db_bytes_per_pixel);
3294
3295 const VkExtent2D db_extent = (VkExtent2D){
3296 .width = 1ull << ((db_pixel_count_log + 1) / 2),
3297 .height = 1ull << (color_pixel_count_log / 2)
3298 };
3299
3300 if (db_extent.width * db_extent.height < extent.width * extent.height)
3301 extent = db_extent;
3302 }
3303
3304 extent.width = MAX2(extent.width, 128);
3305 extent.height = MAX2(extent.width, 64);
3306
3307 return extent;
3308 }
3309
3310 static void
3311 radv_pipeline_generate_disabled_binning_state(struct radeon_cmdbuf *ctx_cs,
3312 struct radv_pipeline *pipeline,
3313 const VkGraphicsPipelineCreateInfo *pCreateInfo)
3314 {
3315 uint32_t pa_sc_binner_cntl_0 =
3316 S_028C44_BINNING_MODE(V_028C44_DISABLE_BINNING_USE_LEGACY_SC) |
3317 S_028C44_DISABLE_START_OF_PRIM(1);
3318 uint32_t db_dfsm_control = S_028060_PUNCHOUT_MODE(V_028060_FORCE_OFF);
3319
3320 if (pipeline->device->physical_device->rad_info.chip_class >= GFX10) {
3321 RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass);
3322 struct radv_subpass *subpass = pass->subpasses + pCreateInfo->subpass;
3323 const VkPipelineColorBlendStateCreateInfo *vkblend = pCreateInfo->pColorBlendState;
3324 unsigned min_bytes_per_pixel = 0;
3325
3326 if (vkblend) {
3327 for (unsigned i = 0; i < subpass->color_count; i++) {
3328 if (!vkblend->pAttachments[i].colorWriteMask)
3329 continue;
3330
3331 if (subpass->color_attachments[i].attachment == VK_ATTACHMENT_UNUSED)
3332 continue;
3333
3334 VkFormat format = pass->attachments[subpass->color_attachments[i].attachment].format;
3335 unsigned bytes = vk_format_get_blocksize(format);
3336 if (!min_bytes_per_pixel || bytes < min_bytes_per_pixel)
3337 min_bytes_per_pixel = bytes;
3338 }
3339 }
3340
3341 pa_sc_binner_cntl_0 =
3342 S_028C44_BINNING_MODE(V_028C44_DISABLE_BINNING_USE_NEW_SC) |
3343 S_028C44_BIN_SIZE_X(0) |
3344 S_028C44_BIN_SIZE_Y(0) |
3345 S_028C44_BIN_SIZE_X_EXTEND(2) | /* 128 */
3346 S_028C44_BIN_SIZE_Y_EXTEND(min_bytes_per_pixel <= 4 ? 2 : 1) | /* 128 or 64 */
3347 S_028C44_DISABLE_START_OF_PRIM(1);
3348 }
3349
3350 pipeline->graphics.binning.pa_sc_binner_cntl_0 = pa_sc_binner_cntl_0;
3351 pipeline->graphics.binning.db_dfsm_control = db_dfsm_control;
3352 }
3353
3354 static void
3355 radv_pipeline_generate_binning_state(struct radeon_cmdbuf *ctx_cs,
3356 struct radv_pipeline *pipeline,
3357 const VkGraphicsPipelineCreateInfo *pCreateInfo,
3358 const struct radv_blend_state *blend)
3359 {
3360 if (pipeline->device->physical_device->rad_info.chip_class < GFX9)
3361 return;
3362
3363 VkExtent2D bin_size;
3364 if (pipeline->device->physical_device->rad_info.chip_class >= GFX10) {
3365 bin_size = radv_gfx10_compute_bin_size(pipeline, pCreateInfo);
3366 } else if (pipeline->device->physical_device->rad_info.chip_class == GFX9) {
3367 bin_size = radv_gfx9_compute_bin_size(pipeline, pCreateInfo);
3368 } else
3369 unreachable("Unhandled generation for binning bin size calculation");
3370
3371 if (pipeline->device->pbb_allowed && bin_size.width && bin_size.height) {
3372 unsigned context_states_per_bin; /* allowed range: [1, 6] */
3373 unsigned persistent_states_per_bin; /* allowed range: [1, 32] */
3374 unsigned fpovs_per_batch; /* allowed range: [0, 255], 0 = unlimited */
3375
3376 if (pipeline->device->physical_device->rad_info.has_dedicated_vram) {
3377 context_states_per_bin = 1;
3378 persistent_states_per_bin = 1;
3379 fpovs_per_batch = 63;
3380 } else {
3381 /* The context states are affected by the scissor bug. */
3382 context_states_per_bin = pipeline->device->physical_device->rad_info.has_gfx9_scissor_bug ? 1 : 6;
3383 /* 32 causes hangs for RAVEN. */
3384 persistent_states_per_bin = 16;
3385 fpovs_per_batch = 63;
3386 }
3387
3388 bool disable_start_of_prim = true;
3389 uint32_t db_dfsm_control = S_028060_PUNCHOUT_MODE(V_028060_FORCE_OFF);
3390
3391 const struct radv_shader_variant *ps = pipeline->shaders[MESA_SHADER_FRAGMENT];
3392
3393 if (pipeline->device->dfsm_allowed && ps &&
3394 !ps->info.ps.can_discard &&
3395 !ps->info.ps.writes_memory &&
3396 blend->cb_target_enabled_4bit) {
3397 db_dfsm_control = S_028060_PUNCHOUT_MODE(V_028060_AUTO);
3398 disable_start_of_prim = (blend->blend_enable_4bit & blend->cb_target_enabled_4bit) != 0;
3399 }
3400
3401 const uint32_t pa_sc_binner_cntl_0 =
3402 S_028C44_BINNING_MODE(V_028C44_BINNING_ALLOWED) |
3403 S_028C44_BIN_SIZE_X(bin_size.width == 16) |
3404 S_028C44_BIN_SIZE_Y(bin_size.height == 16) |
3405 S_028C44_BIN_SIZE_X_EXTEND(util_logbase2(MAX2(bin_size.width, 32)) - 5) |
3406 S_028C44_BIN_SIZE_Y_EXTEND(util_logbase2(MAX2(bin_size.height, 32)) - 5) |
3407 S_028C44_CONTEXT_STATES_PER_BIN(context_states_per_bin - 1) |
3408 S_028C44_PERSISTENT_STATES_PER_BIN(persistent_states_per_bin - 1) |
3409 S_028C44_DISABLE_START_OF_PRIM(disable_start_of_prim) |
3410 S_028C44_FPOVS_PER_BATCH(fpovs_per_batch) |
3411 S_028C44_OPTIMAL_BIN_SELECTION(1);
3412
3413 pipeline->graphics.binning.pa_sc_binner_cntl_0 = pa_sc_binner_cntl_0;
3414 pipeline->graphics.binning.db_dfsm_control = db_dfsm_control;
3415 } else
3416 radv_pipeline_generate_disabled_binning_state(ctx_cs, pipeline, pCreateInfo);
3417 }
3418
3419
3420 static void
3421 radv_pipeline_generate_depth_stencil_state(struct radeon_cmdbuf *ctx_cs,
3422 struct radv_pipeline *pipeline,
3423 const VkGraphicsPipelineCreateInfo *pCreateInfo,
3424 const struct radv_graphics_pipeline_create_info *extra)
3425 {
3426 const VkPipelineDepthStencilStateCreateInfo *vkds = pCreateInfo->pDepthStencilState;
3427 RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass);
3428 struct radv_subpass *subpass = pass->subpasses + pCreateInfo->subpass;
3429 struct radv_shader_variant *ps = pipeline->shaders[MESA_SHADER_FRAGMENT];
3430 struct radv_render_pass_attachment *attachment = NULL;
3431 uint32_t db_depth_control = 0, db_stencil_control = 0;
3432 uint32_t db_render_control = 0, db_render_override2 = 0;
3433 uint32_t db_render_override = 0;
3434
3435 if (subpass->depth_stencil_attachment)
3436 attachment = pass->attachments + subpass->depth_stencil_attachment->attachment;
3437
3438 bool has_depth_attachment = attachment && vk_format_is_depth(attachment->format);
3439 bool has_stencil_attachment = attachment && vk_format_is_stencil(attachment->format);
3440
3441 if (vkds && has_depth_attachment) {
3442 db_depth_control = S_028800_Z_ENABLE(vkds->depthTestEnable ? 1 : 0) |
3443 S_028800_Z_WRITE_ENABLE(vkds->depthWriteEnable ? 1 : 0) |
3444 S_028800_ZFUNC(vkds->depthCompareOp) |
3445 S_028800_DEPTH_BOUNDS_ENABLE(vkds->depthBoundsTestEnable ? 1 : 0);
3446
3447 /* from amdvlk: For 4xAA and 8xAA need to decompress on flush for better performance */
3448 db_render_override2 |= S_028010_DECOMPRESS_Z_ON_FLUSH(attachment->samples > 2);
3449 }
3450
3451 if (has_stencil_attachment && vkds && vkds->stencilTestEnable) {
3452 db_depth_control |= S_028800_STENCIL_ENABLE(1) | S_028800_BACKFACE_ENABLE(1);
3453 db_depth_control |= S_028800_STENCILFUNC(vkds->front.compareOp);
3454 db_stencil_control |= S_02842C_STENCILFAIL(si_translate_stencil_op(vkds->front.failOp));
3455 db_stencil_control |= S_02842C_STENCILZPASS(si_translate_stencil_op(vkds->front.passOp));
3456 db_stencil_control |= S_02842C_STENCILZFAIL(si_translate_stencil_op(vkds->front.depthFailOp));
3457
3458 db_depth_control |= S_028800_STENCILFUNC_BF(vkds->back.compareOp);
3459 db_stencil_control |= S_02842C_STENCILFAIL_BF(si_translate_stencil_op(vkds->back.failOp));
3460 db_stencil_control |= S_02842C_STENCILZPASS_BF(si_translate_stencil_op(vkds->back.passOp));
3461 db_stencil_control |= S_02842C_STENCILZFAIL_BF(si_translate_stencil_op(vkds->back.depthFailOp));
3462 }
3463
3464 if (attachment && extra) {
3465 db_render_control |= S_028000_DEPTH_CLEAR_ENABLE(extra->db_depth_clear);
3466 db_render_control |= S_028000_STENCIL_CLEAR_ENABLE(extra->db_stencil_clear);
3467
3468 db_render_control |= S_028000_RESUMMARIZE_ENABLE(extra->db_resummarize);
3469 db_render_control |= S_028000_DEPTH_COMPRESS_DISABLE(extra->db_flush_depth_inplace);
3470 db_render_control |= S_028000_STENCIL_COMPRESS_DISABLE(extra->db_flush_stencil_inplace);
3471 db_render_override2 |= S_028010_DISABLE_ZMASK_EXPCLEAR_OPTIMIZATION(extra->db_depth_disable_expclear);
3472 db_render_override2 |= S_028010_DISABLE_SMEM_EXPCLEAR_OPTIMIZATION(extra->db_stencil_disable_expclear);
3473 }
3474
3475 db_render_override |= S_02800C_FORCE_HIS_ENABLE0(V_02800C_FORCE_DISABLE) |
3476 S_02800C_FORCE_HIS_ENABLE1(V_02800C_FORCE_DISABLE);
3477
3478 if (!pCreateInfo->pRasterizationState->depthClampEnable &&
3479 ps->info.ps.writes_z) {
3480 /* From VK_EXT_depth_range_unrestricted spec:
3481 *
3482 * "The behavior described in Primitive Clipping still applies.
3483 * If depth clamping is disabled the depth values are still
3484 * clipped to 0 ≤ zc ≤ wc before the viewport transform. If
3485 * depth clamping is enabled the above equation is ignored and
3486 * the depth values are instead clamped to the VkViewport
3487 * minDepth and maxDepth values, which in the case of this
3488 * extension can be outside of the 0.0 to 1.0 range."
3489 */
3490 db_render_override |= S_02800C_DISABLE_VIEWPORT_CLAMP(1);
3491 }
3492
3493 radeon_set_context_reg(ctx_cs, R_028800_DB_DEPTH_CONTROL, db_depth_control);
3494 radeon_set_context_reg(ctx_cs, R_02842C_DB_STENCIL_CONTROL, db_stencil_control);
3495
3496 radeon_set_context_reg(ctx_cs, R_028000_DB_RENDER_CONTROL, db_render_control);
3497 radeon_set_context_reg(ctx_cs, R_02800C_DB_RENDER_OVERRIDE, db_render_override);
3498 radeon_set_context_reg(ctx_cs, R_028010_DB_RENDER_OVERRIDE2, db_render_override2);
3499 }
3500
3501 static void
3502 radv_pipeline_generate_blend_state(struct radeon_cmdbuf *ctx_cs,
3503 struct radv_pipeline *pipeline,
3504 const struct radv_blend_state *blend)
3505 {
3506 radeon_set_context_reg_seq(ctx_cs, R_028780_CB_BLEND0_CONTROL, 8);
3507 radeon_emit_array(ctx_cs, blend->cb_blend_control,
3508 8);
3509 radeon_set_context_reg(ctx_cs, R_028808_CB_COLOR_CONTROL, blend->cb_color_control);
3510 radeon_set_context_reg(ctx_cs, R_028B70_DB_ALPHA_TO_MASK, blend->db_alpha_to_mask);
3511
3512 if (pipeline->device->physical_device->rad_info.has_rbplus) {
3513
3514 radeon_set_context_reg_seq(ctx_cs, R_028760_SX_MRT0_BLEND_OPT, 8);
3515 radeon_emit_array(ctx_cs, blend->sx_mrt_blend_opt, 8);
3516 }
3517
3518 radeon_set_context_reg(ctx_cs, R_028714_SPI_SHADER_COL_FORMAT, blend->spi_shader_col_format);
3519
3520 radeon_set_context_reg(ctx_cs, R_028238_CB_TARGET_MASK, blend->cb_target_mask);
3521 radeon_set_context_reg(ctx_cs, R_02823C_CB_SHADER_MASK, blend->cb_shader_mask);
3522
3523 pipeline->graphics.col_format = blend->spi_shader_col_format;
3524 pipeline->graphics.cb_target_mask = blend->cb_target_mask;
3525 }
3526
3527 static const VkConservativeRasterizationModeEXT
3528 radv_get_conservative_raster_mode(const VkPipelineRasterizationStateCreateInfo *pCreateInfo)
3529 {
3530 const VkPipelineRasterizationConservativeStateCreateInfoEXT *conservative_raster =
3531 vk_find_struct_const(pCreateInfo->pNext, PIPELINE_RASTERIZATION_CONSERVATIVE_STATE_CREATE_INFO_EXT);
3532
3533 if (!conservative_raster)
3534 return VK_CONSERVATIVE_RASTERIZATION_MODE_DISABLED_EXT;
3535 return conservative_raster->conservativeRasterizationMode;
3536 }
3537
3538 static void
3539 radv_pipeline_generate_raster_state(struct radeon_cmdbuf *ctx_cs,
3540 struct radv_pipeline *pipeline,
3541 const VkGraphicsPipelineCreateInfo *pCreateInfo)
3542 {
3543 const VkPipelineRasterizationStateCreateInfo *vkraster = pCreateInfo->pRasterizationState;
3544 const VkConservativeRasterizationModeEXT mode =
3545 radv_get_conservative_raster_mode(vkraster);
3546 uint32_t pa_sc_conservative_rast = S_028C4C_NULL_SQUAD_AA_MASK_ENABLE(1);
3547 bool depth_clip_disable = vkraster->depthClampEnable;
3548
3549 const VkPipelineRasterizationDepthClipStateCreateInfoEXT *depth_clip_state =
3550 vk_find_struct_const(vkraster->pNext, PIPELINE_RASTERIZATION_DEPTH_CLIP_STATE_CREATE_INFO_EXT);
3551 if (depth_clip_state) {
3552 depth_clip_disable = !depth_clip_state->depthClipEnable;
3553 }
3554
3555 radeon_set_context_reg(ctx_cs, R_028810_PA_CL_CLIP_CNTL,
3556 S_028810_DX_CLIP_SPACE_DEF(1) | // vulkan uses DX conventions.
3557 S_028810_ZCLIP_NEAR_DISABLE(depth_clip_disable ? 1 : 0) |
3558 S_028810_ZCLIP_FAR_DISABLE(depth_clip_disable ? 1 : 0) |
3559 S_028810_DX_RASTERIZATION_KILL(vkraster->rasterizerDiscardEnable ? 1 : 0) |
3560 S_028810_DX_LINEAR_ATTR_CLIP_ENA(1));
3561
3562 radeon_set_context_reg(ctx_cs, R_0286D4_SPI_INTERP_CONTROL_0,
3563 S_0286D4_FLAT_SHADE_ENA(1) |
3564 S_0286D4_PNT_SPRITE_ENA(1) |
3565 S_0286D4_PNT_SPRITE_OVRD_X(V_0286D4_SPI_PNT_SPRITE_SEL_S) |
3566 S_0286D4_PNT_SPRITE_OVRD_Y(V_0286D4_SPI_PNT_SPRITE_SEL_T) |
3567 S_0286D4_PNT_SPRITE_OVRD_Z(V_0286D4_SPI_PNT_SPRITE_SEL_0) |
3568 S_0286D4_PNT_SPRITE_OVRD_W(V_0286D4_SPI_PNT_SPRITE_SEL_1) |
3569 S_0286D4_PNT_SPRITE_TOP_1(0)); /* vulkan is top to bottom - 1.0 at bottom */
3570
3571 radeon_set_context_reg(ctx_cs, R_028BE4_PA_SU_VTX_CNTL,
3572 S_028BE4_PIX_CENTER(1) | // TODO verify
3573 S_028BE4_ROUND_MODE(V_028BE4_X_ROUND_TO_EVEN) |
3574 S_028BE4_QUANT_MODE(V_028BE4_X_16_8_FIXED_POINT_1_256TH));
3575
3576 radeon_set_context_reg(ctx_cs, R_028814_PA_SU_SC_MODE_CNTL,
3577 S_028814_FACE(vkraster->frontFace) |
3578 S_028814_CULL_FRONT(!!(vkraster->cullMode & VK_CULL_MODE_FRONT_BIT)) |
3579 S_028814_CULL_BACK(!!(vkraster->cullMode & VK_CULL_MODE_BACK_BIT)) |
3580 S_028814_POLY_MODE(vkraster->polygonMode != VK_POLYGON_MODE_FILL) |
3581 S_028814_POLYMODE_FRONT_PTYPE(si_translate_fill(vkraster->polygonMode)) |
3582 S_028814_POLYMODE_BACK_PTYPE(si_translate_fill(vkraster->polygonMode)) |
3583 S_028814_POLY_OFFSET_FRONT_ENABLE(vkraster->depthBiasEnable ? 1 : 0) |
3584 S_028814_POLY_OFFSET_BACK_ENABLE(vkraster->depthBiasEnable ? 1 : 0) |
3585 S_028814_POLY_OFFSET_PARA_ENABLE(vkraster->depthBiasEnable ? 1 : 0));
3586
3587 /* Conservative rasterization. */
3588 if (mode != VK_CONSERVATIVE_RASTERIZATION_MODE_DISABLED_EXT) {
3589 struct radv_multisample_state *ms = &pipeline->graphics.ms;
3590
3591 ms->pa_sc_aa_config |= S_028BE0_AA_MASK_CENTROID_DTMN(1);
3592 ms->db_eqaa |= S_028804_ENABLE_POSTZ_OVERRASTERIZATION(1) |
3593 S_028804_OVERRASTERIZATION_AMOUNT(4);
3594
3595 pa_sc_conservative_rast = S_028C4C_PREZ_AA_MASK_ENABLE(1) |
3596 S_028C4C_POSTZ_AA_MASK_ENABLE(1) |
3597 S_028C4C_CENTROID_SAMPLE_OVERRIDE(1);
3598
3599 if (mode == VK_CONSERVATIVE_RASTERIZATION_MODE_OVERESTIMATE_EXT) {
3600 pa_sc_conservative_rast |=
3601 S_028C4C_OVER_RAST_ENABLE(1) |
3602 S_028C4C_OVER_RAST_SAMPLE_SELECT(0) |
3603 S_028C4C_UNDER_RAST_ENABLE(0) |
3604 S_028C4C_UNDER_RAST_SAMPLE_SELECT(1) |
3605 S_028C4C_PBB_UNCERTAINTY_REGION_ENABLE(1);
3606 } else {
3607 assert(mode == VK_CONSERVATIVE_RASTERIZATION_MODE_UNDERESTIMATE_EXT);
3608 pa_sc_conservative_rast |=
3609 S_028C4C_OVER_RAST_ENABLE(0) |
3610 S_028C4C_OVER_RAST_SAMPLE_SELECT(1) |
3611 S_028C4C_UNDER_RAST_ENABLE(1) |
3612 S_028C4C_UNDER_RAST_SAMPLE_SELECT(0) |
3613 S_028C4C_PBB_UNCERTAINTY_REGION_ENABLE(0);
3614 }
3615 }
3616
3617 radeon_set_context_reg(ctx_cs, R_028C4C_PA_SC_CONSERVATIVE_RASTERIZATION_CNTL,
3618 pa_sc_conservative_rast);
3619 }
3620
3621
3622 static void
3623 radv_pipeline_generate_multisample_state(struct radeon_cmdbuf *ctx_cs,
3624 struct radv_pipeline *pipeline)
3625 {
3626 struct radv_multisample_state *ms = &pipeline->graphics.ms;
3627
3628 radeon_set_context_reg_seq(ctx_cs, R_028C38_PA_SC_AA_MASK_X0Y0_X1Y0, 2);
3629 radeon_emit(ctx_cs, ms->pa_sc_aa_mask[0]);
3630 radeon_emit(ctx_cs, ms->pa_sc_aa_mask[1]);
3631
3632 radeon_set_context_reg(ctx_cs, R_028804_DB_EQAA, ms->db_eqaa);
3633 radeon_set_context_reg(ctx_cs, R_028A4C_PA_SC_MODE_CNTL_1, ms->pa_sc_mode_cntl_1);
3634
3635 /* The exclusion bits can be set to improve rasterization efficiency
3636 * if no sample lies on the pixel boundary (-8 sample offset). It's
3637 * currently always TRUE because the driver doesn't support 16 samples.
3638 */
3639 bool exclusion = pipeline->device->physical_device->rad_info.chip_class >= GFX7;
3640 radeon_set_context_reg(ctx_cs, R_02882C_PA_SU_PRIM_FILTER_CNTL,
3641 S_02882C_XMAX_RIGHT_EXCLUSION(exclusion) |
3642 S_02882C_YMAX_BOTTOM_EXCLUSION(exclusion));
3643 }
3644
3645 static void
3646 radv_pipeline_generate_vgt_gs_mode(struct radeon_cmdbuf *ctx_cs,
3647 struct radv_pipeline *pipeline)
3648 {
3649 const struct radv_vs_output_info *outinfo = get_vs_output_info(pipeline);
3650 const struct radv_shader_variant *vs =
3651 pipeline->shaders[MESA_SHADER_TESS_EVAL] ?
3652 pipeline->shaders[MESA_SHADER_TESS_EVAL] :
3653 pipeline->shaders[MESA_SHADER_VERTEX];
3654 unsigned vgt_primitiveid_en = 0;
3655 uint32_t vgt_gs_mode = 0;
3656
3657 if (radv_pipeline_has_ngg(pipeline))
3658 return;
3659
3660 if (radv_pipeline_has_gs(pipeline)) {
3661 const struct radv_shader_variant *gs =
3662 pipeline->shaders[MESA_SHADER_GEOMETRY];
3663
3664 vgt_gs_mode = ac_vgt_gs_mode(gs->info.gs.vertices_out,
3665 pipeline->device->physical_device->rad_info.chip_class);
3666 } else if (outinfo->export_prim_id || vs->info.uses_prim_id) {
3667 vgt_gs_mode = S_028A40_MODE(V_028A40_GS_SCENARIO_A);
3668 vgt_primitiveid_en |= S_028A84_PRIMITIVEID_EN(1);
3669 }
3670
3671 radeon_set_context_reg(ctx_cs, R_028A84_VGT_PRIMITIVEID_EN, vgt_primitiveid_en);
3672 radeon_set_context_reg(ctx_cs, R_028A40_VGT_GS_MODE, vgt_gs_mode);
3673 }
3674
3675 static void
3676 radv_pipeline_generate_hw_vs(struct radeon_cmdbuf *ctx_cs,
3677 struct radeon_cmdbuf *cs,
3678 struct radv_pipeline *pipeline,
3679 struct radv_shader_variant *shader)
3680 {
3681 uint64_t va = radv_buffer_get_va(shader->bo) + shader->bo_offset;
3682
3683 radeon_set_sh_reg_seq(cs, R_00B120_SPI_SHADER_PGM_LO_VS, 4);
3684 radeon_emit(cs, va >> 8);
3685 radeon_emit(cs, S_00B124_MEM_BASE(va >> 40));
3686 radeon_emit(cs, shader->config.rsrc1);
3687 radeon_emit(cs, shader->config.rsrc2);
3688
3689 const struct radv_vs_output_info *outinfo = get_vs_output_info(pipeline);
3690 unsigned clip_dist_mask, cull_dist_mask, total_mask;
3691 clip_dist_mask = outinfo->clip_dist_mask;
3692 cull_dist_mask = outinfo->cull_dist_mask;
3693 total_mask = clip_dist_mask | cull_dist_mask;
3694 bool misc_vec_ena = outinfo->writes_pointsize ||
3695 outinfo->writes_layer ||
3696 outinfo->writes_viewport_index;
3697 unsigned spi_vs_out_config, nparams;
3698
3699 /* VS is required to export at least one param. */
3700 nparams = MAX2(outinfo->param_exports, 1);
3701 spi_vs_out_config = S_0286C4_VS_EXPORT_COUNT(nparams - 1);
3702
3703 if (pipeline->device->physical_device->rad_info.chip_class >= GFX10) {
3704 spi_vs_out_config |= S_0286C4_NO_PC_EXPORT(outinfo->param_exports == 0);
3705 }
3706
3707 radeon_set_context_reg(ctx_cs, R_0286C4_SPI_VS_OUT_CONFIG, spi_vs_out_config);
3708
3709 radeon_set_context_reg(ctx_cs, R_02870C_SPI_SHADER_POS_FORMAT,
3710 S_02870C_POS0_EXPORT_FORMAT(V_02870C_SPI_SHADER_4COMP) |
3711 S_02870C_POS1_EXPORT_FORMAT(outinfo->pos_exports > 1 ?
3712 V_02870C_SPI_SHADER_4COMP :
3713 V_02870C_SPI_SHADER_NONE) |
3714 S_02870C_POS2_EXPORT_FORMAT(outinfo->pos_exports > 2 ?
3715 V_02870C_SPI_SHADER_4COMP :
3716 V_02870C_SPI_SHADER_NONE) |
3717 S_02870C_POS3_EXPORT_FORMAT(outinfo->pos_exports > 3 ?
3718 V_02870C_SPI_SHADER_4COMP :
3719 V_02870C_SPI_SHADER_NONE));
3720
3721 radeon_set_context_reg(ctx_cs, R_028818_PA_CL_VTE_CNTL,
3722 S_028818_VTX_W0_FMT(1) |
3723 S_028818_VPORT_X_SCALE_ENA(1) | S_028818_VPORT_X_OFFSET_ENA(1) |
3724 S_028818_VPORT_Y_SCALE_ENA(1) | S_028818_VPORT_Y_OFFSET_ENA(1) |
3725 S_028818_VPORT_Z_SCALE_ENA(1) | S_028818_VPORT_Z_OFFSET_ENA(1));
3726
3727 radeon_set_context_reg(ctx_cs, R_02881C_PA_CL_VS_OUT_CNTL,
3728 S_02881C_USE_VTX_POINT_SIZE(outinfo->writes_pointsize) |
3729 S_02881C_USE_VTX_RENDER_TARGET_INDX(outinfo->writes_layer) |
3730 S_02881C_USE_VTX_VIEWPORT_INDX(outinfo->writes_viewport_index) |
3731 S_02881C_VS_OUT_MISC_VEC_ENA(misc_vec_ena) |
3732 S_02881C_VS_OUT_MISC_SIDE_BUS_ENA(misc_vec_ena) |
3733 S_02881C_VS_OUT_CCDIST0_VEC_ENA((total_mask & 0x0f) != 0) |
3734 S_02881C_VS_OUT_CCDIST1_VEC_ENA((total_mask & 0xf0) != 0) |
3735 cull_dist_mask << 8 |
3736 clip_dist_mask);
3737
3738 if (pipeline->device->physical_device->rad_info.chip_class <= GFX8)
3739 radeon_set_context_reg(ctx_cs, R_028AB4_VGT_REUSE_OFF,
3740 outinfo->writes_viewport_index);
3741 }
3742
3743 static void
3744 radv_pipeline_generate_hw_es(struct radeon_cmdbuf *cs,
3745 struct radv_pipeline *pipeline,
3746 struct radv_shader_variant *shader)
3747 {
3748 uint64_t va = radv_buffer_get_va(shader->bo) + shader->bo_offset;
3749
3750 radeon_set_sh_reg_seq(cs, R_00B320_SPI_SHADER_PGM_LO_ES, 4);
3751 radeon_emit(cs, va >> 8);
3752 radeon_emit(cs, S_00B324_MEM_BASE(va >> 40));
3753 radeon_emit(cs, shader->config.rsrc1);
3754 radeon_emit(cs, shader->config.rsrc2);
3755 }
3756
3757 static void
3758 radv_pipeline_generate_hw_ls(struct radeon_cmdbuf *cs,
3759 struct radv_pipeline *pipeline,
3760 struct radv_shader_variant *shader,
3761 const struct radv_tessellation_state *tess)
3762 {
3763 uint64_t va = radv_buffer_get_va(shader->bo) + shader->bo_offset;
3764 uint32_t rsrc2 = shader->config.rsrc2;
3765
3766 radeon_set_sh_reg_seq(cs, R_00B520_SPI_SHADER_PGM_LO_LS, 2);
3767 radeon_emit(cs, va >> 8);
3768 radeon_emit(cs, S_00B524_MEM_BASE(va >> 40));
3769
3770 rsrc2 |= S_00B52C_LDS_SIZE(tess->lds_size);
3771 if (pipeline->device->physical_device->rad_info.chip_class == GFX7 &&
3772 pipeline->device->physical_device->rad_info.family != CHIP_HAWAII)
3773 radeon_set_sh_reg(cs, R_00B52C_SPI_SHADER_PGM_RSRC2_LS, rsrc2);
3774
3775 radeon_set_sh_reg_seq(cs, R_00B528_SPI_SHADER_PGM_RSRC1_LS, 2);
3776 radeon_emit(cs, shader->config.rsrc1);
3777 radeon_emit(cs, rsrc2);
3778 }
3779
3780 static void
3781 radv_pipeline_generate_hw_ngg(struct radeon_cmdbuf *ctx_cs,
3782 struct radeon_cmdbuf *cs,
3783 struct radv_pipeline *pipeline,
3784 struct radv_shader_variant *shader)
3785 {
3786 uint64_t va = radv_buffer_get_va(shader->bo) + shader->bo_offset;
3787 gl_shader_stage es_type =
3788 radv_pipeline_has_tess(pipeline) ? MESA_SHADER_TESS_EVAL : MESA_SHADER_VERTEX;
3789 struct radv_shader_variant *es =
3790 es_type == MESA_SHADER_TESS_EVAL ? pipeline->shaders[MESA_SHADER_TESS_EVAL] : pipeline->shaders[MESA_SHADER_VERTEX];
3791 const struct gfx10_ngg_info *ngg_state = &shader->info.ngg_info;
3792
3793 radeon_set_sh_reg_seq(cs, R_00B320_SPI_SHADER_PGM_LO_ES, 2);
3794 radeon_emit(cs, va >> 8);
3795 radeon_emit(cs, S_00B324_MEM_BASE(va >> 40));
3796 radeon_set_sh_reg_seq(cs, R_00B228_SPI_SHADER_PGM_RSRC1_GS, 2);
3797 radeon_emit(cs, shader->config.rsrc1);
3798 radeon_emit(cs, shader->config.rsrc2);
3799
3800 const struct radv_vs_output_info *outinfo = get_vs_output_info(pipeline);
3801 unsigned clip_dist_mask, cull_dist_mask, total_mask;
3802 clip_dist_mask = outinfo->clip_dist_mask;
3803 cull_dist_mask = outinfo->cull_dist_mask;
3804 total_mask = clip_dist_mask | cull_dist_mask;
3805 bool misc_vec_ena = outinfo->writes_pointsize ||
3806 outinfo->writes_layer ||
3807 outinfo->writes_viewport_index;
3808 bool es_enable_prim_id = outinfo->export_prim_id ||
3809 (es && es->info.uses_prim_id);
3810 bool break_wave_at_eoi = false;
3811 unsigned ge_cntl;
3812 unsigned nparams;
3813
3814 if (es_type == MESA_SHADER_TESS_EVAL) {
3815 struct radv_shader_variant *gs =
3816 pipeline->shaders[MESA_SHADER_GEOMETRY];
3817
3818 if (es_enable_prim_id || (gs && gs->info.uses_prim_id))
3819 break_wave_at_eoi = true;
3820 }
3821
3822 nparams = MAX2(outinfo->param_exports, 1);
3823 radeon_set_context_reg(ctx_cs, R_0286C4_SPI_VS_OUT_CONFIG,
3824 S_0286C4_VS_EXPORT_COUNT(nparams - 1) |
3825 S_0286C4_NO_PC_EXPORT(outinfo->param_exports == 0));
3826
3827 radeon_set_context_reg(ctx_cs, R_028708_SPI_SHADER_IDX_FORMAT,
3828 S_028708_IDX0_EXPORT_FORMAT(V_028708_SPI_SHADER_1COMP));
3829 radeon_set_context_reg(ctx_cs, R_02870C_SPI_SHADER_POS_FORMAT,
3830 S_02870C_POS0_EXPORT_FORMAT(V_02870C_SPI_SHADER_4COMP) |
3831 S_02870C_POS1_EXPORT_FORMAT(outinfo->pos_exports > 1 ?
3832 V_02870C_SPI_SHADER_4COMP :
3833 V_02870C_SPI_SHADER_NONE) |
3834 S_02870C_POS2_EXPORT_FORMAT(outinfo->pos_exports > 2 ?
3835 V_02870C_SPI_SHADER_4COMP :
3836 V_02870C_SPI_SHADER_NONE) |
3837 S_02870C_POS3_EXPORT_FORMAT(outinfo->pos_exports > 3 ?
3838 V_02870C_SPI_SHADER_4COMP :
3839 V_02870C_SPI_SHADER_NONE));
3840
3841 radeon_set_context_reg(ctx_cs, R_028818_PA_CL_VTE_CNTL,
3842 S_028818_VTX_W0_FMT(1) |
3843 S_028818_VPORT_X_SCALE_ENA(1) | S_028818_VPORT_X_OFFSET_ENA(1) |
3844 S_028818_VPORT_Y_SCALE_ENA(1) | S_028818_VPORT_Y_OFFSET_ENA(1) |
3845 S_028818_VPORT_Z_SCALE_ENA(1) | S_028818_VPORT_Z_OFFSET_ENA(1));
3846 radeon_set_context_reg(ctx_cs, R_02881C_PA_CL_VS_OUT_CNTL,
3847 S_02881C_USE_VTX_POINT_SIZE(outinfo->writes_pointsize) |
3848 S_02881C_USE_VTX_RENDER_TARGET_INDX(outinfo->writes_layer) |
3849 S_02881C_USE_VTX_VIEWPORT_INDX(outinfo->writes_viewport_index) |
3850 S_02881C_VS_OUT_MISC_VEC_ENA(misc_vec_ena) |
3851 S_02881C_VS_OUT_MISC_SIDE_BUS_ENA(misc_vec_ena) |
3852 S_02881C_VS_OUT_CCDIST0_VEC_ENA((total_mask & 0x0f) != 0) |
3853 S_02881C_VS_OUT_CCDIST1_VEC_ENA((total_mask & 0xf0) != 0) |
3854 cull_dist_mask << 8 |
3855 clip_dist_mask);
3856
3857 radeon_set_context_reg(ctx_cs, R_028A84_VGT_PRIMITIVEID_EN,
3858 S_028A84_PRIMITIVEID_EN(es_enable_prim_id) |
3859 S_028A84_NGG_DISABLE_PROVOK_REUSE(es_enable_prim_id));
3860
3861 radeon_set_context_reg(ctx_cs, R_028AAC_VGT_ESGS_RING_ITEMSIZE,
3862 ngg_state->vgt_esgs_ring_itemsize);
3863
3864 /* NGG specific registers. */
3865 struct radv_shader_variant *gs = pipeline->shaders[MESA_SHADER_GEOMETRY];
3866 uint32_t gs_num_invocations = gs ? gs->info.gs.invocations : 1;
3867
3868 radeon_set_context_reg(ctx_cs, R_028A44_VGT_GS_ONCHIP_CNTL,
3869 S_028A44_ES_VERTS_PER_SUBGRP(ngg_state->hw_max_esverts) |
3870 S_028A44_GS_PRIMS_PER_SUBGRP(ngg_state->max_gsprims) |
3871 S_028A44_GS_INST_PRIMS_IN_SUBGRP(ngg_state->max_gsprims * gs_num_invocations));
3872 radeon_set_context_reg(ctx_cs, R_0287FC_GE_MAX_OUTPUT_PER_SUBGROUP,
3873 S_0287FC_MAX_VERTS_PER_SUBGROUP(ngg_state->max_out_verts));
3874 radeon_set_context_reg(ctx_cs, R_028B4C_GE_NGG_SUBGRP_CNTL,
3875 S_028B4C_PRIM_AMP_FACTOR(ngg_state->prim_amp_factor) |
3876 S_028B4C_THDS_PER_SUBGRP(0)); /* for fast launch */
3877 radeon_set_context_reg(ctx_cs, R_028B90_VGT_GS_INSTANCE_CNT,
3878 S_028B90_CNT(gs_num_invocations) |
3879 S_028B90_ENABLE(gs_num_invocations > 1) |
3880 S_028B90_EN_MAX_VERT_OUT_PER_GS_INSTANCE(ngg_state->max_vert_out_per_gs_instance));
3881
3882 /* User edge flags are set by the pos exports. If user edge flags are
3883 * not used, we must use hw-generated edge flags and pass them via
3884 * the prim export to prevent drawing lines on internal edges of
3885 * decomposed primitives (such as quads) with polygon mode = lines.
3886 *
3887 * TODO: We should combine hw-generated edge flags with user edge
3888 * flags in the shader.
3889 */
3890 radeon_set_context_reg(ctx_cs, R_028838_PA_CL_NGG_CNTL,
3891 S_028838_INDEX_BUF_EDGE_FLAG_ENA(!radv_pipeline_has_tess(pipeline) &&
3892 !radv_pipeline_has_gs(pipeline)));
3893
3894 ge_cntl = S_03096C_PRIM_GRP_SIZE(ngg_state->max_gsprims) |
3895 S_03096C_VERT_GRP_SIZE(ngg_state->hw_max_esverts) |
3896 S_03096C_BREAK_WAVE_AT_EOI(break_wave_at_eoi);
3897
3898 /* Bug workaround for a possible hang with non-tessellation cases.
3899 * Tessellation always sets GE_CNTL.VERT_GRP_SIZE = 0
3900 *
3901 * Requirement: GE_CNTL.VERT_GRP_SIZE = VGT_GS_ONCHIP_CNTL.ES_VERTS_PER_SUBGRP - 5
3902 */
3903 if ((pipeline->device->physical_device->rad_info.family == CHIP_NAVI10 ||
3904 pipeline->device->physical_device->rad_info.family == CHIP_NAVI12 ||
3905 pipeline->device->physical_device->rad_info.family == CHIP_NAVI14) &&
3906 !radv_pipeline_has_tess(pipeline) &&
3907 ngg_state->hw_max_esverts != 256) {
3908 ge_cntl &= C_03096C_VERT_GRP_SIZE;
3909
3910 if (ngg_state->hw_max_esverts > 5) {
3911 ge_cntl |= S_03096C_VERT_GRP_SIZE(ngg_state->hw_max_esverts - 5);
3912 }
3913 }
3914
3915 radeon_set_uconfig_reg(ctx_cs, R_03096C_GE_CNTL, ge_cntl);
3916 }
3917
3918 static void
3919 radv_pipeline_generate_hw_hs(struct radeon_cmdbuf *cs,
3920 struct radv_pipeline *pipeline,
3921 struct radv_shader_variant *shader,
3922 const struct radv_tessellation_state *tess)
3923 {
3924 uint64_t va = radv_buffer_get_va(shader->bo) + shader->bo_offset;
3925
3926 if (pipeline->device->physical_device->rad_info.chip_class >= GFX9) {
3927 unsigned hs_rsrc2 = shader->config.rsrc2;
3928
3929 if (pipeline->device->physical_device->rad_info.chip_class >= GFX10) {
3930 hs_rsrc2 |= S_00B42C_LDS_SIZE_GFX10(tess->lds_size);
3931 } else {
3932 hs_rsrc2 |= S_00B42C_LDS_SIZE_GFX9(tess->lds_size);
3933 }
3934
3935 if (pipeline->device->physical_device->rad_info.chip_class >= GFX10) {
3936 radeon_set_sh_reg_seq(cs, R_00B520_SPI_SHADER_PGM_LO_LS, 2);
3937 radeon_emit(cs, va >> 8);
3938 radeon_emit(cs, S_00B524_MEM_BASE(va >> 40));
3939 } else {
3940 radeon_set_sh_reg_seq(cs, R_00B410_SPI_SHADER_PGM_LO_LS, 2);
3941 radeon_emit(cs, va >> 8);
3942 radeon_emit(cs, S_00B414_MEM_BASE(va >> 40));
3943 }
3944
3945 radeon_set_sh_reg_seq(cs, R_00B428_SPI_SHADER_PGM_RSRC1_HS, 2);
3946 radeon_emit(cs, shader->config.rsrc1);
3947 radeon_emit(cs, hs_rsrc2);
3948 } else {
3949 radeon_set_sh_reg_seq(cs, R_00B420_SPI_SHADER_PGM_LO_HS, 4);
3950 radeon_emit(cs, va >> 8);
3951 radeon_emit(cs, S_00B424_MEM_BASE(va >> 40));
3952 radeon_emit(cs, shader->config.rsrc1);
3953 radeon_emit(cs, shader->config.rsrc2);
3954 }
3955 }
3956
3957 static void
3958 radv_pipeline_generate_vertex_shader(struct radeon_cmdbuf *ctx_cs,
3959 struct radeon_cmdbuf *cs,
3960 struct radv_pipeline *pipeline,
3961 const struct radv_tessellation_state *tess)
3962 {
3963 struct radv_shader_variant *vs;
3964
3965 /* Skip shaders merged into HS/GS */
3966 vs = pipeline->shaders[MESA_SHADER_VERTEX];
3967 if (!vs)
3968 return;
3969
3970 if (vs->info.vs.as_ls)
3971 radv_pipeline_generate_hw_ls(cs, pipeline, vs, tess);
3972 else if (vs->info.vs.as_es)
3973 radv_pipeline_generate_hw_es(cs, pipeline, vs);
3974 else if (vs->info.is_ngg)
3975 radv_pipeline_generate_hw_ngg(ctx_cs, cs, pipeline, vs);
3976 else
3977 radv_pipeline_generate_hw_vs(ctx_cs, cs, pipeline, vs);
3978 }
3979
3980 static void
3981 radv_pipeline_generate_tess_shaders(struct radeon_cmdbuf *ctx_cs,
3982 struct radeon_cmdbuf *cs,
3983 struct radv_pipeline *pipeline,
3984 const struct radv_tessellation_state *tess)
3985 {
3986 if (!radv_pipeline_has_tess(pipeline))
3987 return;
3988
3989 struct radv_shader_variant *tes, *tcs;
3990
3991 tcs = pipeline->shaders[MESA_SHADER_TESS_CTRL];
3992 tes = pipeline->shaders[MESA_SHADER_TESS_EVAL];
3993
3994 if (tes) {
3995 if (tes->info.is_ngg) {
3996 radv_pipeline_generate_hw_ngg(ctx_cs, cs, pipeline, tes);
3997 } else if (tes->info.tes.as_es)
3998 radv_pipeline_generate_hw_es(cs, pipeline, tes);
3999 else
4000 radv_pipeline_generate_hw_vs(ctx_cs, cs, pipeline, tes);
4001 }
4002
4003 radv_pipeline_generate_hw_hs(cs, pipeline, tcs, tess);
4004
4005 radeon_set_context_reg(ctx_cs, R_028B6C_VGT_TF_PARAM,
4006 tess->tf_param);
4007
4008 if (pipeline->device->physical_device->rad_info.chip_class >= GFX7)
4009 radeon_set_context_reg_idx(ctx_cs, R_028B58_VGT_LS_HS_CONFIG, 2,
4010 tess->ls_hs_config);
4011 else
4012 radeon_set_context_reg(ctx_cs, R_028B58_VGT_LS_HS_CONFIG,
4013 tess->ls_hs_config);
4014
4015 if (pipeline->device->physical_device->rad_info.chip_class >= GFX10 &&
4016 !radv_pipeline_has_gs(pipeline) && !radv_pipeline_has_ngg(pipeline)) {
4017 radeon_set_context_reg(ctx_cs, R_028A44_VGT_GS_ONCHIP_CNTL,
4018 S_028A44_ES_VERTS_PER_SUBGRP(250) |
4019 S_028A44_GS_PRIMS_PER_SUBGRP(126) |
4020 S_028A44_GS_INST_PRIMS_IN_SUBGRP(126));
4021 }
4022 }
4023
4024 static void
4025 radv_pipeline_generate_hw_gs(struct radeon_cmdbuf *ctx_cs,
4026 struct radeon_cmdbuf *cs,
4027 struct radv_pipeline *pipeline,
4028 struct radv_shader_variant *gs)
4029 {
4030 const struct gfx9_gs_info *gs_state = &gs->info.gs_ring_info;
4031 unsigned gs_max_out_vertices;
4032 uint8_t *num_components;
4033 uint8_t max_stream;
4034 unsigned offset;
4035 uint64_t va;
4036
4037 gs_max_out_vertices = gs->info.gs.vertices_out;
4038 max_stream = gs->info.gs.max_stream;
4039 num_components = gs->info.gs.num_stream_output_components;
4040
4041 offset = num_components[0] * gs_max_out_vertices;
4042
4043 radeon_set_context_reg_seq(ctx_cs, R_028A60_VGT_GSVS_RING_OFFSET_1, 3);
4044 radeon_emit(ctx_cs, offset);
4045 if (max_stream >= 1)
4046 offset += num_components[1] * gs_max_out_vertices;
4047 radeon_emit(ctx_cs, offset);
4048 if (max_stream >= 2)
4049 offset += num_components[2] * gs_max_out_vertices;
4050 radeon_emit(ctx_cs, offset);
4051 if (max_stream >= 3)
4052 offset += num_components[3] * gs_max_out_vertices;
4053 radeon_set_context_reg(ctx_cs, R_028AB0_VGT_GSVS_RING_ITEMSIZE, offset);
4054
4055 radeon_set_context_reg_seq(ctx_cs, R_028B5C_VGT_GS_VERT_ITEMSIZE, 4);
4056 radeon_emit(ctx_cs, num_components[0]);
4057 radeon_emit(ctx_cs, (max_stream >= 1) ? num_components[1] : 0);
4058 radeon_emit(ctx_cs, (max_stream >= 2) ? num_components[2] : 0);
4059 radeon_emit(ctx_cs, (max_stream >= 3) ? num_components[3] : 0);
4060
4061 uint32_t gs_num_invocations = gs->info.gs.invocations;
4062 radeon_set_context_reg(ctx_cs, R_028B90_VGT_GS_INSTANCE_CNT,
4063 S_028B90_CNT(MIN2(gs_num_invocations, 127)) |
4064 S_028B90_ENABLE(gs_num_invocations > 0));
4065
4066 radeon_set_context_reg(ctx_cs, R_028AAC_VGT_ESGS_RING_ITEMSIZE,
4067 gs_state->vgt_esgs_ring_itemsize);
4068
4069 va = radv_buffer_get_va(gs->bo) + gs->bo_offset;
4070
4071 if (pipeline->device->physical_device->rad_info.chip_class >= GFX9) {
4072 if (pipeline->device->physical_device->rad_info.chip_class >= GFX10) {
4073 radeon_set_sh_reg_seq(cs, R_00B320_SPI_SHADER_PGM_LO_ES, 2);
4074 radeon_emit(cs, va >> 8);
4075 radeon_emit(cs, S_00B324_MEM_BASE(va >> 40));
4076 } else {
4077 radeon_set_sh_reg_seq(cs, R_00B210_SPI_SHADER_PGM_LO_ES, 2);
4078 radeon_emit(cs, va >> 8);
4079 radeon_emit(cs, S_00B214_MEM_BASE(va >> 40));
4080 }
4081
4082 radeon_set_sh_reg_seq(cs, R_00B228_SPI_SHADER_PGM_RSRC1_GS, 2);
4083 radeon_emit(cs, gs->config.rsrc1);
4084 radeon_emit(cs, gs->config.rsrc2 | S_00B22C_LDS_SIZE(gs_state->lds_size));
4085
4086 radeon_set_context_reg(ctx_cs, R_028A44_VGT_GS_ONCHIP_CNTL, gs_state->vgt_gs_onchip_cntl);
4087 radeon_set_context_reg(ctx_cs, R_028A94_VGT_GS_MAX_PRIMS_PER_SUBGROUP, gs_state->vgt_gs_max_prims_per_subgroup);
4088 } else {
4089 radeon_set_sh_reg_seq(cs, R_00B220_SPI_SHADER_PGM_LO_GS, 4);
4090 radeon_emit(cs, va >> 8);
4091 radeon_emit(cs, S_00B224_MEM_BASE(va >> 40));
4092 radeon_emit(cs, gs->config.rsrc1);
4093 radeon_emit(cs, gs->config.rsrc2);
4094 }
4095
4096 radv_pipeline_generate_hw_vs(ctx_cs, cs, pipeline, pipeline->gs_copy_shader);
4097 }
4098
4099 static void
4100 radv_pipeline_generate_geometry_shader(struct radeon_cmdbuf *ctx_cs,
4101 struct radeon_cmdbuf *cs,
4102 struct radv_pipeline *pipeline)
4103 {
4104 struct radv_shader_variant *gs;
4105
4106 gs = pipeline->shaders[MESA_SHADER_GEOMETRY];
4107 if (!gs)
4108 return;
4109
4110 if (gs->info.is_ngg)
4111 radv_pipeline_generate_hw_ngg(ctx_cs, cs, pipeline, gs);
4112 else
4113 radv_pipeline_generate_hw_gs(ctx_cs, cs, pipeline, gs);
4114
4115 radeon_set_context_reg(ctx_cs, R_028B38_VGT_GS_MAX_VERT_OUT,
4116 gs->info.gs.vertices_out);
4117 }
4118
4119 static uint32_t offset_to_ps_input(uint32_t offset, bool flat_shade, bool float16)
4120 {
4121 uint32_t ps_input_cntl;
4122 if (offset <= AC_EXP_PARAM_OFFSET_31) {
4123 ps_input_cntl = S_028644_OFFSET(offset);
4124 if (flat_shade)
4125 ps_input_cntl |= S_028644_FLAT_SHADE(1);
4126 if (float16) {
4127 ps_input_cntl |= S_028644_FP16_INTERP_MODE(1) |
4128 S_028644_ATTR0_VALID(1);
4129 }
4130 } else {
4131 /* The input is a DEFAULT_VAL constant. */
4132 assert(offset >= AC_EXP_PARAM_DEFAULT_VAL_0000 &&
4133 offset <= AC_EXP_PARAM_DEFAULT_VAL_1111);
4134 offset -= AC_EXP_PARAM_DEFAULT_VAL_0000;
4135 ps_input_cntl = S_028644_OFFSET(0x20) |
4136 S_028644_DEFAULT_VAL(offset);
4137 }
4138 return ps_input_cntl;
4139 }
4140
4141 static void
4142 radv_pipeline_generate_ps_inputs(struct radeon_cmdbuf *ctx_cs,
4143 struct radv_pipeline *pipeline)
4144 {
4145 struct radv_shader_variant *ps = pipeline->shaders[MESA_SHADER_FRAGMENT];
4146 const struct radv_vs_output_info *outinfo = get_vs_output_info(pipeline);
4147 uint32_t ps_input_cntl[32];
4148
4149 unsigned ps_offset = 0;
4150
4151 if (ps->info.ps.prim_id_input) {
4152 unsigned vs_offset = outinfo->vs_output_param_offset[VARYING_SLOT_PRIMITIVE_ID];
4153 if (vs_offset != AC_EXP_PARAM_UNDEFINED) {
4154 ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, true, false);
4155 ++ps_offset;
4156 }
4157 }
4158
4159 if (ps->info.ps.layer_input ||
4160 ps->info.needs_multiview_view_index) {
4161 unsigned vs_offset = outinfo->vs_output_param_offset[VARYING_SLOT_LAYER];
4162 if (vs_offset != AC_EXP_PARAM_UNDEFINED)
4163 ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, true, false);
4164 else
4165 ps_input_cntl[ps_offset] = offset_to_ps_input(AC_EXP_PARAM_DEFAULT_VAL_0000, true, false);
4166 ++ps_offset;
4167 }
4168
4169 if (ps->info.ps.has_pcoord) {
4170 unsigned val;
4171 val = S_028644_PT_SPRITE_TEX(1) | S_028644_OFFSET(0x20);
4172 ps_input_cntl[ps_offset] = val;
4173 ps_offset++;
4174 }
4175
4176 if (ps->info.ps.num_input_clips_culls) {
4177 unsigned vs_offset;
4178
4179 vs_offset = outinfo->vs_output_param_offset[VARYING_SLOT_CLIP_DIST0];
4180 if (vs_offset != AC_EXP_PARAM_UNDEFINED) {
4181 ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, false, false);
4182 ++ps_offset;
4183 }
4184
4185 vs_offset = outinfo->vs_output_param_offset[VARYING_SLOT_CLIP_DIST1];
4186 if (vs_offset != AC_EXP_PARAM_UNDEFINED &&
4187 ps->info.ps.num_input_clips_culls > 4) {
4188 ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, false, false);
4189 ++ps_offset;
4190 }
4191 }
4192
4193 for (unsigned i = 0; i < 32 && (1u << i) <= ps->info.ps.input_mask; ++i) {
4194 unsigned vs_offset;
4195 bool flat_shade;
4196 bool float16;
4197 if (!(ps->info.ps.input_mask & (1u << i)))
4198 continue;
4199
4200 vs_offset = outinfo->vs_output_param_offset[VARYING_SLOT_VAR0 + i];
4201 if (vs_offset == AC_EXP_PARAM_UNDEFINED) {
4202 ps_input_cntl[ps_offset] = S_028644_OFFSET(0x20);
4203 ++ps_offset;
4204 continue;
4205 }
4206
4207 flat_shade = !!(ps->info.ps.flat_shaded_mask & (1u << ps_offset));
4208 float16 = !!(ps->info.ps.float16_shaded_mask & (1u << ps_offset));
4209
4210 ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, flat_shade, float16);
4211 ++ps_offset;
4212 }
4213
4214 if (ps_offset) {
4215 radeon_set_context_reg_seq(ctx_cs, R_028644_SPI_PS_INPUT_CNTL_0, ps_offset);
4216 for (unsigned i = 0; i < ps_offset; i++) {
4217 radeon_emit(ctx_cs, ps_input_cntl[i]);
4218 }
4219 }
4220 }
4221
4222 static uint32_t
4223 radv_compute_db_shader_control(const struct radv_device *device,
4224 const struct radv_pipeline *pipeline,
4225 const struct radv_shader_variant *ps)
4226 {
4227 unsigned z_order;
4228 if (ps->info.ps.early_fragment_test || !ps->info.ps.writes_memory)
4229 z_order = V_02880C_EARLY_Z_THEN_LATE_Z;
4230 else
4231 z_order = V_02880C_LATE_Z;
4232
4233 bool disable_rbplus = device->physical_device->rad_info.has_rbplus &&
4234 !device->physical_device->rad_info.rbplus_allowed;
4235
4236 /* It shouldn't be needed to export gl_SampleMask when MSAA is disabled
4237 * but this appears to break Project Cars (DXVK). See
4238 * https://bugs.freedesktop.org/show_bug.cgi?id=109401
4239 */
4240 bool mask_export_enable = ps->info.ps.writes_sample_mask;
4241
4242 return S_02880C_Z_EXPORT_ENABLE(ps->info.ps.writes_z) |
4243 S_02880C_STENCIL_TEST_VAL_EXPORT_ENABLE(ps->info.ps.writes_stencil) |
4244 S_02880C_KILL_ENABLE(!!ps->info.ps.can_discard) |
4245 S_02880C_MASK_EXPORT_ENABLE(mask_export_enable) |
4246 S_02880C_Z_ORDER(z_order) |
4247 S_02880C_DEPTH_BEFORE_SHADER(ps->info.ps.early_fragment_test) |
4248 S_02880C_PRE_SHADER_DEPTH_COVERAGE_ENABLE(ps->info.ps.post_depth_coverage) |
4249 S_02880C_EXEC_ON_HIER_FAIL(ps->info.ps.writes_memory) |
4250 S_02880C_EXEC_ON_NOOP(ps->info.ps.writes_memory) |
4251 S_02880C_DUAL_QUAD_DISABLE(disable_rbplus);
4252 }
4253
4254 static void
4255 radv_pipeline_generate_fragment_shader(struct radeon_cmdbuf *ctx_cs,
4256 struct radeon_cmdbuf *cs,
4257 struct radv_pipeline *pipeline)
4258 {
4259 struct radv_shader_variant *ps;
4260 uint64_t va;
4261 assert (pipeline->shaders[MESA_SHADER_FRAGMENT]);
4262
4263 ps = pipeline->shaders[MESA_SHADER_FRAGMENT];
4264 va = radv_buffer_get_va(ps->bo) + ps->bo_offset;
4265
4266 radeon_set_sh_reg_seq(cs, R_00B020_SPI_SHADER_PGM_LO_PS, 4);
4267 radeon_emit(cs, va >> 8);
4268 radeon_emit(cs, S_00B024_MEM_BASE(va >> 40));
4269 radeon_emit(cs, ps->config.rsrc1);
4270 radeon_emit(cs, ps->config.rsrc2);
4271
4272 radeon_set_context_reg(ctx_cs, R_02880C_DB_SHADER_CONTROL,
4273 radv_compute_db_shader_control(pipeline->device,
4274 pipeline, ps));
4275
4276 radeon_set_context_reg(ctx_cs, R_0286CC_SPI_PS_INPUT_ENA,
4277 ps->config.spi_ps_input_ena);
4278
4279 radeon_set_context_reg(ctx_cs, R_0286D0_SPI_PS_INPUT_ADDR,
4280 ps->config.spi_ps_input_addr);
4281
4282 radeon_set_context_reg(ctx_cs, R_0286D8_SPI_PS_IN_CONTROL,
4283 S_0286D8_NUM_INTERP(ps->info.ps.num_interp) |
4284 S_0286D8_PS_W32_EN(ps->info.wave_size == 32));
4285
4286 radeon_set_context_reg(ctx_cs, R_0286E0_SPI_BARYC_CNTL, pipeline->graphics.spi_baryc_cntl);
4287
4288 radeon_set_context_reg(ctx_cs, R_028710_SPI_SHADER_Z_FORMAT,
4289 ac_get_spi_shader_z_format(ps->info.ps.writes_z,
4290 ps->info.ps.writes_stencil,
4291 ps->info.ps.writes_sample_mask));
4292
4293 if (pipeline->device->dfsm_allowed) {
4294 /* optimise this? */
4295 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
4296 radeon_emit(cs, EVENT_TYPE(V_028A90_FLUSH_DFSM) | EVENT_INDEX(0));
4297 }
4298 }
4299
4300 static void
4301 radv_pipeline_generate_vgt_vertex_reuse(struct radeon_cmdbuf *ctx_cs,
4302 struct radv_pipeline *pipeline)
4303 {
4304 if (pipeline->device->physical_device->rad_info.family < CHIP_POLARIS10 ||
4305 pipeline->device->physical_device->rad_info.chip_class >= GFX10)
4306 return;
4307
4308 unsigned vtx_reuse_depth = 30;
4309 if (radv_pipeline_has_tess(pipeline) &&
4310 radv_get_shader(pipeline, MESA_SHADER_TESS_EVAL)->info.tes.spacing == TESS_SPACING_FRACTIONAL_ODD) {
4311 vtx_reuse_depth = 14;
4312 }
4313 radeon_set_context_reg(ctx_cs, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL,
4314 S_028C58_VTX_REUSE_DEPTH(vtx_reuse_depth));
4315 }
4316
4317 static uint32_t
4318 radv_compute_vgt_shader_stages_en(const struct radv_pipeline *pipeline)
4319 {
4320 uint32_t stages = 0;
4321 if (radv_pipeline_has_tess(pipeline)) {
4322 stages |= S_028B54_LS_EN(V_028B54_LS_STAGE_ON) |
4323 S_028B54_HS_EN(1) | S_028B54_DYNAMIC_HS(1);
4324
4325 if (radv_pipeline_has_gs(pipeline))
4326 stages |= S_028B54_ES_EN(V_028B54_ES_STAGE_DS) |
4327 S_028B54_GS_EN(1);
4328 else if (radv_pipeline_has_ngg(pipeline))
4329 stages |= S_028B54_ES_EN(V_028B54_ES_STAGE_DS);
4330 else
4331 stages |= S_028B54_VS_EN(V_028B54_VS_STAGE_DS);
4332 } else if (radv_pipeline_has_gs(pipeline)) {
4333 stages |= S_028B54_ES_EN(V_028B54_ES_STAGE_REAL) |
4334 S_028B54_GS_EN(1);
4335 } else if (radv_pipeline_has_ngg(pipeline)) {
4336 stages |= S_028B54_ES_EN(V_028B54_ES_STAGE_REAL);
4337 }
4338
4339 if (radv_pipeline_has_ngg(pipeline)) {
4340 stages |= S_028B54_PRIMGEN_EN(1);
4341 if (pipeline->streamout_shader)
4342 stages |= S_028B54_NGG_WAVE_ID_EN(1);
4343 } else if (radv_pipeline_has_gs(pipeline)) {
4344 stages |= S_028B54_VS_EN(V_028B54_VS_STAGE_COPY_SHADER);
4345 }
4346
4347 if (pipeline->device->physical_device->rad_info.chip_class >= GFX9)
4348 stages |= S_028B54_MAX_PRIMGRP_IN_WAVE(2);
4349
4350 if (pipeline->device->physical_device->rad_info.chip_class >= GFX10) {
4351 uint8_t hs_size = 64, gs_size = 64, vs_size = 64;
4352
4353 if (radv_pipeline_has_tess(pipeline))
4354 hs_size = pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.wave_size;
4355
4356 if (pipeline->shaders[MESA_SHADER_GEOMETRY]) {
4357 vs_size = gs_size = pipeline->shaders[MESA_SHADER_GEOMETRY]->info.wave_size;
4358 if (pipeline->gs_copy_shader)
4359 vs_size = pipeline->gs_copy_shader->info.wave_size;
4360 } else if (pipeline->shaders[MESA_SHADER_TESS_EVAL])
4361 vs_size = pipeline->shaders[MESA_SHADER_TESS_EVAL]->info.wave_size;
4362 else if (pipeline->shaders[MESA_SHADER_VERTEX])
4363 vs_size = pipeline->shaders[MESA_SHADER_VERTEX]->info.wave_size;
4364
4365 if (radv_pipeline_has_ngg(pipeline))
4366 gs_size = vs_size;
4367
4368 /* legacy GS only supports Wave64 */
4369 stages |= S_028B54_HS_W32_EN(hs_size == 32 ? 1 : 0) |
4370 S_028B54_GS_W32_EN(gs_size == 32 ? 1 : 0) |
4371 S_028B54_VS_W32_EN(vs_size == 32 ? 1 : 0);
4372 }
4373
4374 return stages;
4375 }
4376
4377 static uint32_t
4378 radv_compute_cliprect_rule(const VkGraphicsPipelineCreateInfo *pCreateInfo)
4379 {
4380 const VkPipelineDiscardRectangleStateCreateInfoEXT *discard_rectangle_info =
4381 vk_find_struct_const(pCreateInfo->pNext, PIPELINE_DISCARD_RECTANGLE_STATE_CREATE_INFO_EXT);
4382
4383 if (!discard_rectangle_info)
4384 return 0xffff;
4385
4386 unsigned mask = 0;
4387
4388 for (unsigned i = 0; i < (1u << MAX_DISCARD_RECTANGLES); ++i) {
4389 /* Interpret i as a bitmask, and then set the bit in the mask if
4390 * that combination of rectangles in which the pixel is contained
4391 * should pass the cliprect test. */
4392 unsigned relevant_subset = i & ((1u << discard_rectangle_info->discardRectangleCount) - 1);
4393
4394 if (discard_rectangle_info->discardRectangleMode == VK_DISCARD_RECTANGLE_MODE_INCLUSIVE_EXT &&
4395 !relevant_subset)
4396 continue;
4397
4398 if (discard_rectangle_info->discardRectangleMode == VK_DISCARD_RECTANGLE_MODE_EXCLUSIVE_EXT &&
4399 relevant_subset)
4400 continue;
4401
4402 mask |= 1u << i;
4403 }
4404
4405 return mask;
4406 }
4407
4408 static void
4409 gfx10_pipeline_generate_ge_cntl(struct radeon_cmdbuf *ctx_cs,
4410 struct radv_pipeline *pipeline,
4411 const struct radv_tessellation_state *tess)
4412 {
4413 bool break_wave_at_eoi = false;
4414 unsigned primgroup_size;
4415 unsigned vertgroup_size;
4416
4417 if (radv_pipeline_has_tess(pipeline)) {
4418 primgroup_size = tess->num_patches; /* must be a multiple of NUM_PATCHES */
4419 vertgroup_size = 0;
4420 } else if (radv_pipeline_has_gs(pipeline)) {
4421 const struct gfx9_gs_info *gs_state =
4422 &pipeline->shaders[MESA_SHADER_GEOMETRY]->info.gs_ring_info;
4423 unsigned vgt_gs_onchip_cntl = gs_state->vgt_gs_onchip_cntl;
4424 primgroup_size = G_028A44_GS_PRIMS_PER_SUBGRP(vgt_gs_onchip_cntl);
4425 vertgroup_size = G_028A44_ES_VERTS_PER_SUBGRP(vgt_gs_onchip_cntl);
4426 } else {
4427 primgroup_size = 128; /* recommended without a GS and tess */
4428 vertgroup_size = 0;
4429 }
4430
4431 if (radv_pipeline_has_tess(pipeline)) {
4432 if (pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.uses_prim_id ||
4433 radv_get_shader(pipeline, MESA_SHADER_TESS_EVAL)->info.uses_prim_id)
4434 break_wave_at_eoi = true;
4435 }
4436
4437 radeon_set_uconfig_reg(ctx_cs, R_03096C_GE_CNTL,
4438 S_03096C_PRIM_GRP_SIZE(primgroup_size) |
4439 S_03096C_VERT_GRP_SIZE(vertgroup_size) |
4440 S_03096C_PACKET_TO_ONE_PA(0) /* line stipple */ |
4441 S_03096C_BREAK_WAVE_AT_EOI(break_wave_at_eoi));
4442 }
4443
4444 static void
4445 radv_pipeline_generate_pm4(struct radv_pipeline *pipeline,
4446 const VkGraphicsPipelineCreateInfo *pCreateInfo,
4447 const struct radv_graphics_pipeline_create_info *extra,
4448 const struct radv_blend_state *blend,
4449 const struct radv_tessellation_state *tess,
4450 unsigned prim, unsigned gs_out)
4451 {
4452 struct radeon_cmdbuf *ctx_cs = &pipeline->ctx_cs;
4453 struct radeon_cmdbuf *cs = &pipeline->cs;
4454
4455 cs->max_dw = 64;
4456 ctx_cs->max_dw = 256;
4457 cs->buf = malloc(4 * (cs->max_dw + ctx_cs->max_dw));
4458 ctx_cs->buf = cs->buf + cs->max_dw;
4459
4460 radv_pipeline_generate_depth_stencil_state(ctx_cs, pipeline, pCreateInfo, extra);
4461 radv_pipeline_generate_blend_state(ctx_cs, pipeline, blend);
4462 radv_pipeline_generate_raster_state(ctx_cs, pipeline, pCreateInfo);
4463 radv_pipeline_generate_multisample_state(ctx_cs, pipeline);
4464 radv_pipeline_generate_vgt_gs_mode(ctx_cs, pipeline);
4465 radv_pipeline_generate_vertex_shader(ctx_cs, cs, pipeline, tess);
4466 radv_pipeline_generate_tess_shaders(ctx_cs, cs, pipeline, tess);
4467 radv_pipeline_generate_geometry_shader(ctx_cs, cs, pipeline);
4468 radv_pipeline_generate_fragment_shader(ctx_cs, cs, pipeline);
4469 radv_pipeline_generate_ps_inputs(ctx_cs, pipeline);
4470 radv_pipeline_generate_vgt_vertex_reuse(ctx_cs, pipeline);
4471 radv_pipeline_generate_binning_state(ctx_cs, pipeline, pCreateInfo, blend);
4472
4473 if (pipeline->device->physical_device->rad_info.chip_class >= GFX10 && !radv_pipeline_has_ngg(pipeline))
4474 gfx10_pipeline_generate_ge_cntl(ctx_cs, pipeline, tess);
4475
4476 radeon_set_context_reg(ctx_cs, R_0286E8_SPI_TMPRING_SIZE,
4477 S_0286E8_WAVES(pipeline->max_waves) |
4478 S_0286E8_WAVESIZE(pipeline->scratch_bytes_per_wave >> 10));
4479
4480 radeon_set_context_reg(ctx_cs, R_028B54_VGT_SHADER_STAGES_EN, radv_compute_vgt_shader_stages_en(pipeline));
4481
4482 if (pipeline->device->physical_device->rad_info.chip_class >= GFX7) {
4483 radeon_set_uconfig_reg_idx(pipeline->device->physical_device,
4484 cs, R_030908_VGT_PRIMITIVE_TYPE, 1, prim);
4485 } else {
4486 radeon_set_config_reg(cs, R_008958_VGT_PRIMITIVE_TYPE, prim);
4487 }
4488 radeon_set_context_reg(ctx_cs, R_028A6C_VGT_GS_OUT_PRIM_TYPE, gs_out);
4489
4490 radeon_set_context_reg(ctx_cs, R_02820C_PA_SC_CLIPRECT_RULE, radv_compute_cliprect_rule(pCreateInfo));
4491
4492 pipeline->ctx_cs_hash = _mesa_hash_data(ctx_cs->buf, ctx_cs->cdw * 4);
4493
4494 assert(ctx_cs->cdw <= ctx_cs->max_dw);
4495 assert(cs->cdw <= cs->max_dw);
4496 }
4497
4498 static struct radv_ia_multi_vgt_param_helpers
4499 radv_compute_ia_multi_vgt_param_helpers(struct radv_pipeline *pipeline,
4500 const struct radv_tessellation_state *tess,
4501 uint32_t prim)
4502 {
4503 struct radv_ia_multi_vgt_param_helpers ia_multi_vgt_param = {0};
4504 const struct radv_device *device = pipeline->device;
4505
4506 if (radv_pipeline_has_tess(pipeline))
4507 ia_multi_vgt_param.primgroup_size = tess->num_patches;
4508 else if (radv_pipeline_has_gs(pipeline))
4509 ia_multi_vgt_param.primgroup_size = 64;
4510 else
4511 ia_multi_vgt_param.primgroup_size = 128; /* recommended without a GS */
4512
4513 /* GS requirement. */
4514 ia_multi_vgt_param.partial_es_wave = false;
4515 if (radv_pipeline_has_gs(pipeline) && device->physical_device->rad_info.chip_class <= GFX8)
4516 if (SI_GS_PER_ES / ia_multi_vgt_param.primgroup_size >= pipeline->device->gs_table_depth - 3)
4517 ia_multi_vgt_param.partial_es_wave = true;
4518
4519 ia_multi_vgt_param.wd_switch_on_eop = false;
4520 if (device->physical_device->rad_info.chip_class >= GFX7) {
4521 /* WD_SWITCH_ON_EOP has no effect on GPUs with less than
4522 * 4 shader engines. Set 1 to pass the assertion below.
4523 * The other cases are hardware requirements. */
4524 if (device->physical_device->rad_info.max_se < 4 ||
4525 prim == V_008958_DI_PT_POLYGON ||
4526 prim == V_008958_DI_PT_LINELOOP ||
4527 prim == V_008958_DI_PT_TRIFAN ||
4528 prim == V_008958_DI_PT_TRISTRIP_ADJ ||
4529 (pipeline->graphics.prim_restart_enable &&
4530 (device->physical_device->rad_info.family < CHIP_POLARIS10 ||
4531 (prim != V_008958_DI_PT_POINTLIST &&
4532 prim != V_008958_DI_PT_LINESTRIP))))
4533 ia_multi_vgt_param.wd_switch_on_eop = true;
4534 }
4535
4536 ia_multi_vgt_param.ia_switch_on_eoi = false;
4537 if (pipeline->shaders[MESA_SHADER_FRAGMENT]->info.ps.prim_id_input)
4538 ia_multi_vgt_param.ia_switch_on_eoi = true;
4539 if (radv_pipeline_has_gs(pipeline) &&
4540 pipeline->shaders[MESA_SHADER_GEOMETRY]->info.uses_prim_id)
4541 ia_multi_vgt_param.ia_switch_on_eoi = true;
4542 if (radv_pipeline_has_tess(pipeline)) {
4543 /* SWITCH_ON_EOI must be set if PrimID is used. */
4544 if (pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.uses_prim_id ||
4545 radv_get_shader(pipeline, MESA_SHADER_TESS_EVAL)->info.uses_prim_id)
4546 ia_multi_vgt_param.ia_switch_on_eoi = true;
4547 }
4548
4549 ia_multi_vgt_param.partial_vs_wave = false;
4550 if (radv_pipeline_has_tess(pipeline)) {
4551 /* Bug with tessellation and GS on Bonaire and older 2 SE chips. */
4552 if ((device->physical_device->rad_info.family == CHIP_TAHITI ||
4553 device->physical_device->rad_info.family == CHIP_PITCAIRN ||
4554 device->physical_device->rad_info.family == CHIP_BONAIRE) &&
4555 radv_pipeline_has_gs(pipeline))
4556 ia_multi_vgt_param.partial_vs_wave = true;
4557 /* Needed for 028B6C_DISTRIBUTION_MODE != 0 */
4558 if (device->physical_device->rad_info.has_distributed_tess) {
4559 if (radv_pipeline_has_gs(pipeline)) {
4560 if (device->physical_device->rad_info.chip_class <= GFX8)
4561 ia_multi_vgt_param.partial_es_wave = true;
4562 } else {
4563 ia_multi_vgt_param.partial_vs_wave = true;
4564 }
4565 }
4566 }
4567
4568 /* Workaround for a VGT hang when strip primitive types are used with
4569 * primitive restart.
4570 */
4571 if (pipeline->graphics.prim_restart_enable &&
4572 (prim == V_008958_DI_PT_LINESTRIP ||
4573 prim == V_008958_DI_PT_TRISTRIP ||
4574 prim == V_008958_DI_PT_LINESTRIP_ADJ ||
4575 prim == V_008958_DI_PT_TRISTRIP_ADJ)) {
4576 ia_multi_vgt_param.partial_vs_wave = true;
4577 }
4578
4579 if (radv_pipeline_has_gs(pipeline)) {
4580 /* On these chips there is the possibility of a hang if the
4581 * pipeline uses a GS and partial_vs_wave is not set.
4582 *
4583 * This mostly does not hit 4-SE chips, as those typically set
4584 * ia_switch_on_eoi and then partial_vs_wave is set for pipelines
4585 * with GS due to another workaround.
4586 *
4587 * Reproducer: https://bugs.freedesktop.org/show_bug.cgi?id=109242
4588 */
4589 if (device->physical_device->rad_info.family == CHIP_TONGA ||
4590 device->physical_device->rad_info.family == CHIP_FIJI ||
4591 device->physical_device->rad_info.family == CHIP_POLARIS10 ||
4592 device->physical_device->rad_info.family == CHIP_POLARIS11 ||
4593 device->physical_device->rad_info.family == CHIP_POLARIS12 ||
4594 device->physical_device->rad_info.family == CHIP_VEGAM) {
4595 ia_multi_vgt_param.partial_vs_wave = true;
4596 }
4597 }
4598
4599 ia_multi_vgt_param.base =
4600 S_028AA8_PRIMGROUP_SIZE(ia_multi_vgt_param.primgroup_size - 1) |
4601 /* The following field was moved to VGT_SHADER_STAGES_EN in GFX9. */
4602 S_028AA8_MAX_PRIMGRP_IN_WAVE(device->physical_device->rad_info.chip_class == GFX8 ? 2 : 0) |
4603 S_030960_EN_INST_OPT_BASIC(device->physical_device->rad_info.chip_class >= GFX9) |
4604 S_030960_EN_INST_OPT_ADV(device->physical_device->rad_info.chip_class >= GFX9);
4605
4606 return ia_multi_vgt_param;
4607 }
4608
4609
4610 static void
4611 radv_compute_vertex_input_state(struct radv_pipeline *pipeline,
4612 const VkGraphicsPipelineCreateInfo *pCreateInfo)
4613 {
4614 const VkPipelineVertexInputStateCreateInfo *vi_info =
4615 pCreateInfo->pVertexInputState;
4616 struct radv_vertex_elements_info *velems = &pipeline->vertex_elements;
4617
4618 for (uint32_t i = 0; i < vi_info->vertexAttributeDescriptionCount; i++) {
4619 const VkVertexInputAttributeDescription *desc =
4620 &vi_info->pVertexAttributeDescriptions[i];
4621 unsigned loc = desc->location;
4622 const struct vk_format_description *format_desc;
4623
4624 format_desc = vk_format_description(desc->format);
4625
4626 velems->format_size[loc] = format_desc->block.bits / 8;
4627 }
4628
4629 for (uint32_t i = 0; i < vi_info->vertexBindingDescriptionCount; i++) {
4630 const VkVertexInputBindingDescription *desc =
4631 &vi_info->pVertexBindingDescriptions[i];
4632
4633 pipeline->binding_stride[desc->binding] = desc->stride;
4634 pipeline->num_vertex_bindings =
4635 MAX2(pipeline->num_vertex_bindings, desc->binding + 1);
4636 }
4637 }
4638
4639 static struct radv_shader_variant *
4640 radv_pipeline_get_streamout_shader(struct radv_pipeline *pipeline)
4641 {
4642 int i;
4643
4644 for (i = MESA_SHADER_GEOMETRY; i >= MESA_SHADER_VERTEX; i--) {
4645 struct radv_shader_variant *shader =
4646 radv_get_shader(pipeline, i);
4647
4648 if (shader && shader->info.so.num_outputs > 0)
4649 return shader;
4650 }
4651
4652 return NULL;
4653 }
4654
4655 static VkResult
4656 radv_secure_compile(struct radv_pipeline *pipeline,
4657 struct radv_device *device,
4658 const struct radv_pipeline_key *key,
4659 const VkPipelineShaderStageCreateInfo **pStages,
4660 const VkPipelineCreateFlags flags,
4661 unsigned num_stages)
4662 {
4663 uint8_t allowed_pipeline_hashes[2][20];
4664 radv_hash_shaders(allowed_pipeline_hashes[0], pStages,
4665 pipeline->layout, key, get_hash_flags(device));
4666
4667 /* Generate the GC copy hash */
4668 memcpy(allowed_pipeline_hashes[1], allowed_pipeline_hashes[0], 20);
4669 allowed_pipeline_hashes[1][0] ^= 1;
4670
4671 uint8_t allowed_hashes[2][20];
4672 for (unsigned i = 0; i < 2; ++i) {
4673 disk_cache_compute_key(device->physical_device->disk_cache,
4674 allowed_pipeline_hashes[i], 20,
4675 allowed_hashes[i]);
4676 }
4677
4678 /* Do an early exit if all cache entries are already there. */
4679 bool may_need_copy_shader = pStages[MESA_SHADER_GEOMETRY];
4680 void *main_entry = disk_cache_get(device->physical_device->disk_cache, allowed_hashes[0], NULL);
4681 void *copy_entry = NULL;
4682 if (may_need_copy_shader)
4683 copy_entry = disk_cache_get(device->physical_device->disk_cache, allowed_hashes[1], NULL);
4684
4685 bool has_all_cache_entries = main_entry && (!may_need_copy_shader || copy_entry);
4686 free(main_entry);
4687 free(copy_entry);
4688
4689 if(has_all_cache_entries)
4690 return VK_SUCCESS;
4691
4692 unsigned process = 0;
4693 uint8_t sc_threads = device->instance->num_sc_threads;
4694 while (true) {
4695 mtx_lock(&device->sc_state->secure_compile_mutex);
4696 if (device->sc_state->secure_compile_thread_counter < sc_threads) {
4697 device->sc_state->secure_compile_thread_counter++;
4698 for (unsigned i = 0; i < sc_threads; i++) {
4699 if (!device->sc_state->secure_compile_processes[i].in_use) {
4700 device->sc_state->secure_compile_processes[i].in_use = true;
4701 process = i;
4702 break;
4703 }
4704 }
4705 mtx_unlock(&device->sc_state->secure_compile_mutex);
4706 break;
4707 }
4708 mtx_unlock(&device->sc_state->secure_compile_mutex);
4709 }
4710
4711 int fd_secure_input = device->sc_state->secure_compile_processes[process].fd_secure_input;
4712 int fd_secure_output = device->sc_state->secure_compile_processes[process].fd_secure_output;
4713
4714 /* Write pipeline / shader module out to secure process via pipe */
4715 enum radv_secure_compile_type sc_type = RADV_SC_TYPE_COMPILE_PIPELINE;
4716 write(fd_secure_input, &sc_type, sizeof(sc_type));
4717
4718 /* Write pipeline layout out to secure process */
4719 struct radv_pipeline_layout *layout = pipeline->layout;
4720 write(fd_secure_input, layout, sizeof(struct radv_pipeline_layout));
4721 write(fd_secure_input, &layout->num_sets, sizeof(uint32_t));
4722 for (uint32_t set = 0; set < layout->num_sets; set++) {
4723 write(fd_secure_input, &layout->set[set].layout->layout_size, sizeof(uint32_t));
4724 write(fd_secure_input, layout->set[set].layout, layout->set[set].layout->layout_size);
4725 }
4726
4727 /* Write pipeline key out to secure process */
4728 write(fd_secure_input, key, sizeof(struct radv_pipeline_key));
4729
4730 /* Write pipeline create flags out to secure process */
4731 write(fd_secure_input, &flags, sizeof(VkPipelineCreateFlags));
4732
4733 /* Write stage and shader information out to secure process */
4734 write(fd_secure_input, &num_stages, sizeof(uint32_t));
4735 for (uint32_t i = 0; i < MESA_SHADER_STAGES; i++) {
4736 if (!pStages[i])
4737 continue;
4738
4739 /* Write stage out to secure process */
4740 gl_shader_stage stage = ffs(pStages[i]->stage) - 1;
4741 write(fd_secure_input, &stage, sizeof(gl_shader_stage));
4742
4743 /* Write entry point name out to secure process */
4744 size_t name_size = strlen(pStages[i]->pName) + 1;
4745 write(fd_secure_input, &name_size, sizeof(size_t));
4746 write(fd_secure_input, pStages[i]->pName, name_size);
4747
4748 /* Write shader module out to secure process */
4749 struct radv_shader_module *module = radv_shader_module_from_handle(pStages[i]->module);
4750 assert(!module->nir);
4751 size_t module_size = sizeof(struct radv_shader_module) + module->size;
4752 write(fd_secure_input, &module_size, sizeof(size_t));
4753 write(fd_secure_input, module, module_size);
4754
4755 /* Write specialization info out to secure process */
4756 const VkSpecializationInfo *specInfo = pStages[i]->pSpecializationInfo;
4757 bool has_spec_info = specInfo ? true : false;
4758 write(fd_secure_input, &has_spec_info, sizeof(bool));
4759 if (specInfo) {
4760 write(fd_secure_input, &specInfo->dataSize, sizeof(size_t));
4761 write(fd_secure_input, specInfo->pData, specInfo->dataSize);
4762
4763 write(fd_secure_input, &specInfo->mapEntryCount, sizeof(uint32_t));
4764 for (uint32_t j = 0; j < specInfo->mapEntryCount; j++)
4765 write(fd_secure_input, &specInfo->pMapEntries[j], sizeof(VkSpecializationMapEntry));
4766 }
4767 }
4768
4769 /* Read the data returned from the secure process */
4770 while (sc_type != RADV_SC_TYPE_COMPILE_PIPELINE_FINISHED) {
4771 if (!radv_sc_read(fd_secure_output, &sc_type, sizeof(sc_type), true))
4772 return VK_ERROR_DEVICE_LOST;
4773
4774 if (sc_type == RADV_SC_TYPE_WRITE_DISK_CACHE) {
4775 assert(device->physical_device->disk_cache);
4776
4777 uint8_t disk_sha1[20];
4778 if (!radv_sc_read(fd_secure_output, disk_sha1, sizeof(uint8_t) * 20, true))
4779 return VK_ERROR_DEVICE_LOST;
4780
4781 if (memcmp(disk_sha1, allowed_hashes[0], 20) &&
4782 memcmp(disk_sha1, allowed_hashes[1], 20))
4783 return VK_ERROR_DEVICE_LOST;
4784
4785 uint32_t entry_size;
4786 if (!radv_sc_read(fd_secure_output, &entry_size, sizeof(uint32_t), true))
4787 return VK_ERROR_DEVICE_LOST;
4788
4789 struct cache_entry *entry = malloc(entry_size);
4790 if (!radv_sc_read(fd_secure_output, entry, entry_size, true))
4791 return VK_ERROR_DEVICE_LOST;
4792
4793 disk_cache_put(device->physical_device->disk_cache,
4794 disk_sha1, entry, entry_size,
4795 NULL);
4796
4797 free(entry);
4798 } else if (sc_type == RADV_SC_TYPE_READ_DISK_CACHE) {
4799 uint8_t disk_sha1[20];
4800 if (!radv_sc_read(fd_secure_output, disk_sha1, sizeof(uint8_t) * 20, true))
4801 return VK_ERROR_DEVICE_LOST;
4802
4803 if (memcmp(disk_sha1, allowed_hashes[0], 20) &&
4804 memcmp(disk_sha1, allowed_hashes[1], 20))
4805 return VK_ERROR_DEVICE_LOST;
4806
4807 size_t size;
4808 struct cache_entry *entry = (struct cache_entry *)
4809 disk_cache_get(device->physical_device->disk_cache,
4810 disk_sha1, &size);
4811
4812 uint8_t found = entry ? 1 : 0;
4813 write(fd_secure_input, &found, sizeof(uint8_t));
4814
4815 if (found) {
4816 write(fd_secure_input, &size, sizeof(size_t));
4817 write(fd_secure_input, entry, size);
4818 }
4819
4820 free(entry);
4821 }
4822 }
4823
4824 mtx_lock(&device->sc_state->secure_compile_mutex);
4825 device->sc_state->secure_compile_thread_counter--;
4826 device->sc_state->secure_compile_processes[process].in_use = false;
4827 mtx_unlock(&device->sc_state->secure_compile_mutex);
4828
4829 return VK_SUCCESS;
4830 }
4831
4832 static VkResult
4833 radv_pipeline_init(struct radv_pipeline *pipeline,
4834 struct radv_device *device,
4835 struct radv_pipeline_cache *cache,
4836 const VkGraphicsPipelineCreateInfo *pCreateInfo,
4837 const struct radv_graphics_pipeline_create_info *extra)
4838 {
4839 VkResult result;
4840 bool has_view_index = false;
4841
4842 RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass);
4843 struct radv_subpass *subpass = pass->subpasses + pCreateInfo->subpass;
4844 if (subpass->view_mask)
4845 has_view_index = true;
4846
4847 pipeline->device = device;
4848 pipeline->layout = radv_pipeline_layout_from_handle(pCreateInfo->layout);
4849 assert(pipeline->layout);
4850
4851 struct radv_blend_state blend = radv_pipeline_init_blend_state(pipeline, pCreateInfo, extra);
4852
4853 const VkPipelineCreationFeedbackCreateInfoEXT *creation_feedback =
4854 vk_find_struct_const(pCreateInfo->pNext, PIPELINE_CREATION_FEEDBACK_CREATE_INFO_EXT);
4855 radv_init_feedback(creation_feedback);
4856
4857 VkPipelineCreationFeedbackEXT *pipeline_feedback = creation_feedback ? creation_feedback->pPipelineCreationFeedback : NULL;
4858
4859 const VkPipelineShaderStageCreateInfo *pStages[MESA_SHADER_STAGES] = { 0, };
4860 VkPipelineCreationFeedbackEXT *stage_feedbacks[MESA_SHADER_STAGES] = { 0 };
4861 for (uint32_t i = 0; i < pCreateInfo->stageCount; i++) {
4862 gl_shader_stage stage = ffs(pCreateInfo->pStages[i].stage) - 1;
4863 pStages[stage] = &pCreateInfo->pStages[i];
4864 if(creation_feedback)
4865 stage_feedbacks[stage] = &creation_feedback->pPipelineStageCreationFeedbacks[i];
4866 }
4867
4868 struct radv_pipeline_key key = radv_generate_graphics_pipeline_key(pipeline, pCreateInfo, &blend, has_view_index);
4869 if (radv_device_use_secure_compile(device->instance)) {
4870 return radv_secure_compile(pipeline, device, &key, pStages, pCreateInfo->flags, pCreateInfo->stageCount);
4871 } else {
4872 radv_create_shaders(pipeline, device, cache, &key, pStages, pCreateInfo->flags, pipeline_feedback, stage_feedbacks);
4873 }
4874
4875 pipeline->graphics.spi_baryc_cntl = S_0286E0_FRONT_FACE_ALL_BITS(1);
4876 radv_pipeline_init_multisample_state(pipeline, &blend, pCreateInfo);
4877 uint32_t gs_out;
4878 uint32_t prim = si_translate_prim(pCreateInfo->pInputAssemblyState->topology);
4879
4880 pipeline->graphics.can_use_guardband = radv_prim_can_use_guardband(pCreateInfo->pInputAssemblyState->topology);
4881
4882 if (radv_pipeline_has_gs(pipeline)) {
4883 gs_out = si_conv_gl_prim_to_gs_out(pipeline->shaders[MESA_SHADER_GEOMETRY]->info.gs.output_prim);
4884 pipeline->graphics.can_use_guardband = gs_out == V_028A6C_OUTPRIM_TYPE_TRISTRIP;
4885 } else if (radv_pipeline_has_tess(pipeline)) {
4886 if (pipeline->shaders[MESA_SHADER_TESS_EVAL]->info.tes.point_mode)
4887 gs_out = V_028A6C_OUTPRIM_TYPE_POINTLIST;
4888 else
4889 gs_out = si_conv_gl_prim_to_gs_out(pipeline->shaders[MESA_SHADER_TESS_EVAL]->info.tes.primitive_mode);
4890 pipeline->graphics.can_use_guardband = gs_out == V_028A6C_OUTPRIM_TYPE_TRISTRIP;
4891 } else {
4892 gs_out = si_conv_prim_to_gs_out(pCreateInfo->pInputAssemblyState->topology);
4893 }
4894 if (extra && extra->use_rectlist) {
4895 prim = V_008958_DI_PT_RECTLIST;
4896 gs_out = V_028A6C_OUTPRIM_TYPE_TRISTRIP;
4897 pipeline->graphics.can_use_guardband = true;
4898 if (radv_pipeline_has_ngg(pipeline))
4899 gs_out = V_028A6C_VGT_OUT_RECT_V0;
4900 }
4901 pipeline->graphics.prim_restart_enable = !!pCreateInfo->pInputAssemblyState->primitiveRestartEnable;
4902 /* prim vertex count will need TESS changes */
4903 pipeline->graphics.prim_vertex_count = prim_size_table[prim];
4904
4905 radv_pipeline_init_dynamic_state(pipeline, pCreateInfo);
4906
4907 /* Ensure that some export memory is always allocated, for two reasons:
4908 *
4909 * 1) Correctness: The hardware ignores the EXEC mask if no export
4910 * memory is allocated, so KILL and alpha test do not work correctly
4911 * without this.
4912 * 2) Performance: Every shader needs at least a NULL export, even when
4913 * it writes no color/depth output. The NULL export instruction
4914 * stalls without this setting.
4915 *
4916 * Don't add this to CB_SHADER_MASK.
4917 *
4918 * GFX10 supports pixel shaders without exports by setting both the
4919 * color and Z formats to SPI_SHADER_ZERO. The hw will skip export
4920 * instructions if any are present.
4921 */
4922 struct radv_shader_variant *ps = pipeline->shaders[MESA_SHADER_FRAGMENT];
4923 if ((pipeline->device->physical_device->rad_info.chip_class <= GFX9 ||
4924 ps->info.ps.can_discard) &&
4925 !blend.spi_shader_col_format) {
4926 if (!ps->info.ps.writes_z &&
4927 !ps->info.ps.writes_stencil &&
4928 !ps->info.ps.writes_sample_mask)
4929 blend.spi_shader_col_format = V_028714_SPI_SHADER_32_R;
4930 }
4931
4932 for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
4933 if (pipeline->shaders[i]) {
4934 pipeline->need_indirect_descriptor_sets |= pipeline->shaders[i]->info.need_indirect_descriptor_sets;
4935 }
4936 }
4937
4938 if (radv_pipeline_has_gs(pipeline) && !radv_pipeline_has_ngg(pipeline)) {
4939 struct radv_shader_variant *gs =
4940 pipeline->shaders[MESA_SHADER_GEOMETRY];
4941
4942 calculate_gs_ring_sizes(pipeline, &gs->info.gs_ring_info);
4943 }
4944
4945 struct radv_tessellation_state tess = {0};
4946 if (radv_pipeline_has_tess(pipeline)) {
4947 if (prim == V_008958_DI_PT_PATCH) {
4948 pipeline->graphics.prim_vertex_count.min = pCreateInfo->pTessellationState->patchControlPoints;
4949 pipeline->graphics.prim_vertex_count.incr = 1;
4950 }
4951 tess = calculate_tess_state(pipeline, pCreateInfo);
4952 }
4953
4954 pipeline->graphics.ia_multi_vgt_param = radv_compute_ia_multi_vgt_param_helpers(pipeline, &tess, prim);
4955
4956 radv_compute_vertex_input_state(pipeline, pCreateInfo);
4957
4958 for (uint32_t i = 0; i < MESA_SHADER_STAGES; i++)
4959 pipeline->user_data_0[i] = radv_pipeline_stage_to_user_data_0(pipeline, i, device->physical_device->rad_info.chip_class);
4960
4961 struct radv_userdata_info *loc = radv_lookup_user_sgpr(pipeline, MESA_SHADER_VERTEX,
4962 AC_UD_VS_BASE_VERTEX_START_INSTANCE);
4963 if (loc->sgpr_idx != -1) {
4964 pipeline->graphics.vtx_base_sgpr = pipeline->user_data_0[MESA_SHADER_VERTEX];
4965 pipeline->graphics.vtx_base_sgpr += loc->sgpr_idx * 4;
4966 if (radv_get_shader(pipeline, MESA_SHADER_VERTEX)->info.vs.needs_draw_id)
4967 pipeline->graphics.vtx_emit_num = 3;
4968 else
4969 pipeline->graphics.vtx_emit_num = 2;
4970 }
4971
4972 /* Find the last vertex shader stage that eventually uses streamout. */
4973 pipeline->streamout_shader = radv_pipeline_get_streamout_shader(pipeline);
4974
4975 result = radv_pipeline_scratch_init(device, pipeline);
4976 radv_pipeline_generate_pm4(pipeline, pCreateInfo, extra, &blend, &tess, prim, gs_out);
4977
4978 return result;
4979 }
4980
4981 VkResult
4982 radv_graphics_pipeline_create(
4983 VkDevice _device,
4984 VkPipelineCache _cache,
4985 const VkGraphicsPipelineCreateInfo *pCreateInfo,
4986 const struct radv_graphics_pipeline_create_info *extra,
4987 const VkAllocationCallbacks *pAllocator,
4988 VkPipeline *pPipeline)
4989 {
4990 RADV_FROM_HANDLE(radv_device, device, _device);
4991 RADV_FROM_HANDLE(radv_pipeline_cache, cache, _cache);
4992 struct radv_pipeline *pipeline;
4993 VkResult result;
4994
4995 pipeline = vk_zalloc2(&device->alloc, pAllocator, sizeof(*pipeline), 8,
4996 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
4997 if (pipeline == NULL)
4998 return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
4999
5000 result = radv_pipeline_init(pipeline, device, cache,
5001 pCreateInfo, extra);
5002 if (result != VK_SUCCESS) {
5003 radv_pipeline_destroy(device, pipeline, pAllocator);
5004 return result;
5005 }
5006
5007 *pPipeline = radv_pipeline_to_handle(pipeline);
5008
5009 return VK_SUCCESS;
5010 }
5011
5012 VkResult radv_CreateGraphicsPipelines(
5013 VkDevice _device,
5014 VkPipelineCache pipelineCache,
5015 uint32_t count,
5016 const VkGraphicsPipelineCreateInfo* pCreateInfos,
5017 const VkAllocationCallbacks* pAllocator,
5018 VkPipeline* pPipelines)
5019 {
5020 VkResult result = VK_SUCCESS;
5021 unsigned i = 0;
5022
5023 for (; i < count; i++) {
5024 VkResult r;
5025 r = radv_graphics_pipeline_create(_device,
5026 pipelineCache,
5027 &pCreateInfos[i],
5028 NULL, pAllocator, &pPipelines[i]);
5029 if (r != VK_SUCCESS) {
5030 result = r;
5031 pPipelines[i] = VK_NULL_HANDLE;
5032 }
5033 }
5034
5035 return result;
5036 }
5037
5038
5039 static void
5040 radv_compute_generate_pm4(struct radv_pipeline *pipeline)
5041 {
5042 struct radv_shader_variant *compute_shader;
5043 struct radv_device *device = pipeline->device;
5044 unsigned threads_per_threadgroup;
5045 unsigned threadgroups_per_cu = 1;
5046 unsigned waves_per_threadgroup;
5047 unsigned max_waves_per_sh = 0;
5048 uint64_t va;
5049
5050 pipeline->cs.max_dw = device->physical_device->rad_info.chip_class >= GFX10 ? 22 : 20;
5051 pipeline->cs.buf = malloc(pipeline->cs.max_dw * 4);
5052
5053 compute_shader = pipeline->shaders[MESA_SHADER_COMPUTE];
5054 va = radv_buffer_get_va(compute_shader->bo) + compute_shader->bo_offset;
5055
5056 radeon_set_sh_reg_seq(&pipeline->cs, R_00B830_COMPUTE_PGM_LO, 2);
5057 radeon_emit(&pipeline->cs, va >> 8);
5058 radeon_emit(&pipeline->cs, S_00B834_DATA(va >> 40));
5059
5060 radeon_set_sh_reg_seq(&pipeline->cs, R_00B848_COMPUTE_PGM_RSRC1, 2);
5061 radeon_emit(&pipeline->cs, compute_shader->config.rsrc1);
5062 radeon_emit(&pipeline->cs, compute_shader->config.rsrc2);
5063 if (device->physical_device->rad_info.chip_class >= GFX10) {
5064 radeon_set_sh_reg(&pipeline->cs, R_00B8A0_COMPUTE_PGM_RSRC3, compute_shader->config.rsrc3);
5065 }
5066
5067 radeon_set_sh_reg(&pipeline->cs, R_00B860_COMPUTE_TMPRING_SIZE,
5068 S_00B860_WAVES(pipeline->max_waves) |
5069 S_00B860_WAVESIZE(pipeline->scratch_bytes_per_wave >> 10));
5070
5071 /* Calculate best compute resource limits. */
5072 threads_per_threadgroup = compute_shader->info.cs.block_size[0] *
5073 compute_shader->info.cs.block_size[1] *
5074 compute_shader->info.cs.block_size[2];
5075 waves_per_threadgroup = DIV_ROUND_UP(threads_per_threadgroup,
5076 compute_shader->info.wave_size);
5077
5078 if (device->physical_device->rad_info.chip_class >= GFX10 &&
5079 waves_per_threadgroup == 1)
5080 threadgroups_per_cu = 2;
5081
5082 radeon_set_sh_reg(&pipeline->cs, R_00B854_COMPUTE_RESOURCE_LIMITS,
5083 ac_get_compute_resource_limits(&device->physical_device->rad_info,
5084 waves_per_threadgroup,
5085 max_waves_per_sh,
5086 threadgroups_per_cu));
5087
5088 radeon_set_sh_reg_seq(&pipeline->cs, R_00B81C_COMPUTE_NUM_THREAD_X, 3);
5089 radeon_emit(&pipeline->cs,
5090 S_00B81C_NUM_THREAD_FULL(compute_shader->info.cs.block_size[0]));
5091 radeon_emit(&pipeline->cs,
5092 S_00B81C_NUM_THREAD_FULL(compute_shader->info.cs.block_size[1]));
5093 radeon_emit(&pipeline->cs,
5094 S_00B81C_NUM_THREAD_FULL(compute_shader->info.cs.block_size[2]));
5095
5096 assert(pipeline->cs.cdw <= pipeline->cs.max_dw);
5097 }
5098
5099 static struct radv_pipeline_key
5100 radv_generate_compute_pipeline_key(struct radv_pipeline *pipeline,
5101 const VkComputePipelineCreateInfo *pCreateInfo)
5102 {
5103 struct radv_pipeline_key key;
5104 memset(&key, 0, sizeof(key));
5105
5106 if (pCreateInfo->flags & VK_PIPELINE_CREATE_DISABLE_OPTIMIZATION_BIT)
5107 key.optimisations_disabled = 1;
5108
5109 return key;
5110 }
5111
5112 static VkResult radv_compute_pipeline_create(
5113 VkDevice _device,
5114 VkPipelineCache _cache,
5115 const VkComputePipelineCreateInfo* pCreateInfo,
5116 const VkAllocationCallbacks* pAllocator,
5117 VkPipeline* pPipeline)
5118 {
5119 RADV_FROM_HANDLE(radv_device, device, _device);
5120 RADV_FROM_HANDLE(radv_pipeline_cache, cache, _cache);
5121 const VkPipelineShaderStageCreateInfo *pStages[MESA_SHADER_STAGES] = { 0, };
5122 VkPipelineCreationFeedbackEXT *stage_feedbacks[MESA_SHADER_STAGES] = { 0 };
5123 struct radv_pipeline *pipeline;
5124 VkResult result;
5125
5126 pipeline = vk_zalloc2(&device->alloc, pAllocator, sizeof(*pipeline), 8,
5127 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
5128 if (pipeline == NULL)
5129 return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
5130
5131 pipeline->device = device;
5132 pipeline->layout = radv_pipeline_layout_from_handle(pCreateInfo->layout);
5133 assert(pipeline->layout);
5134
5135 const VkPipelineCreationFeedbackCreateInfoEXT *creation_feedback =
5136 vk_find_struct_const(pCreateInfo->pNext, PIPELINE_CREATION_FEEDBACK_CREATE_INFO_EXT);
5137 radv_init_feedback(creation_feedback);
5138
5139 VkPipelineCreationFeedbackEXT *pipeline_feedback = creation_feedback ? creation_feedback->pPipelineCreationFeedback : NULL;
5140 if (creation_feedback)
5141 stage_feedbacks[MESA_SHADER_COMPUTE] = &creation_feedback->pPipelineStageCreationFeedbacks[0];
5142
5143 pStages[MESA_SHADER_COMPUTE] = &pCreateInfo->stage;
5144
5145 struct radv_pipeline_key key =
5146 radv_generate_compute_pipeline_key(pipeline, pCreateInfo);
5147
5148 if (radv_device_use_secure_compile(device->instance)) {
5149 result = radv_secure_compile(pipeline, device, &key, pStages, pCreateInfo->flags, 1);
5150 *pPipeline = radv_pipeline_to_handle(pipeline);
5151
5152 return result;
5153 } else {
5154 radv_create_shaders(pipeline, device, cache, &key, pStages, pCreateInfo->flags, pipeline_feedback, stage_feedbacks);
5155 }
5156
5157 pipeline->user_data_0[MESA_SHADER_COMPUTE] = radv_pipeline_stage_to_user_data_0(pipeline, MESA_SHADER_COMPUTE, device->physical_device->rad_info.chip_class);
5158 pipeline->need_indirect_descriptor_sets |= pipeline->shaders[MESA_SHADER_COMPUTE]->info.need_indirect_descriptor_sets;
5159 result = radv_pipeline_scratch_init(device, pipeline);
5160 if (result != VK_SUCCESS) {
5161 radv_pipeline_destroy(device, pipeline, pAllocator);
5162 return result;
5163 }
5164
5165 radv_compute_generate_pm4(pipeline);
5166
5167 *pPipeline = radv_pipeline_to_handle(pipeline);
5168
5169 return VK_SUCCESS;
5170 }
5171
5172 VkResult radv_CreateComputePipelines(
5173 VkDevice _device,
5174 VkPipelineCache pipelineCache,
5175 uint32_t count,
5176 const VkComputePipelineCreateInfo* pCreateInfos,
5177 const VkAllocationCallbacks* pAllocator,
5178 VkPipeline* pPipelines)
5179 {
5180 VkResult result = VK_SUCCESS;
5181
5182 unsigned i = 0;
5183 for (; i < count; i++) {
5184 VkResult r;
5185 r = radv_compute_pipeline_create(_device, pipelineCache,
5186 &pCreateInfos[i],
5187 pAllocator, &pPipelines[i]);
5188 if (r != VK_SUCCESS) {
5189 result = r;
5190 pPipelines[i] = VK_NULL_HANDLE;
5191 }
5192 }
5193
5194 return result;
5195 }
5196
5197
5198 static uint32_t radv_get_executable_count(const struct radv_pipeline *pipeline)
5199 {
5200 uint32_t ret = 0;
5201 for (int i = 0; i < MESA_SHADER_STAGES; ++i) {
5202 if (!pipeline->shaders[i])
5203 continue;
5204
5205 if (i == MESA_SHADER_GEOMETRY &&
5206 !radv_pipeline_has_ngg(pipeline)) {
5207 ret += 2u;
5208 } else {
5209 ret += 1u;
5210 }
5211
5212 }
5213 return ret;
5214 }
5215
5216 static struct radv_shader_variant *
5217 radv_get_shader_from_executable_index(const struct radv_pipeline *pipeline, int index, gl_shader_stage *stage)
5218 {
5219 for (int i = 0; i < MESA_SHADER_STAGES; ++i) {
5220 if (!pipeline->shaders[i])
5221 continue;
5222 if (!index) {
5223 *stage = i;
5224 return pipeline->shaders[i];
5225 }
5226
5227 --index;
5228
5229 if (i == MESA_SHADER_GEOMETRY &&
5230 !radv_pipeline_has_ngg(pipeline)) {
5231 if (!index) {
5232 *stage = i;
5233 return pipeline->gs_copy_shader;
5234 }
5235 --index;
5236 }
5237 }
5238
5239 *stage = -1;
5240 return NULL;
5241 }
5242
5243 /* Basically strlcpy (which does not exist on linux) specialized for
5244 * descriptions. */
5245 static void desc_copy(char *desc, const char *src) {
5246 int len = strlen(src);
5247 assert(len < VK_MAX_DESCRIPTION_SIZE);
5248 memcpy(desc, src, len);
5249 memset(desc + len, 0, VK_MAX_DESCRIPTION_SIZE - len);
5250 }
5251
5252 VkResult radv_GetPipelineExecutablePropertiesKHR(
5253 VkDevice _device,
5254 const VkPipelineInfoKHR* pPipelineInfo,
5255 uint32_t* pExecutableCount,
5256 VkPipelineExecutablePropertiesKHR* pProperties)
5257 {
5258 RADV_FROM_HANDLE(radv_pipeline, pipeline, pPipelineInfo->pipeline);
5259 const uint32_t total_count = radv_get_executable_count(pipeline);
5260
5261 if (!pProperties) {
5262 *pExecutableCount = total_count;
5263 return VK_SUCCESS;
5264 }
5265
5266 const uint32_t count = MIN2(total_count, *pExecutableCount);
5267 for (unsigned i = 0, executable_idx = 0;
5268 i < MESA_SHADER_STAGES && executable_idx < count; ++i) {
5269 if (!pipeline->shaders[i])
5270 continue;
5271 pProperties[executable_idx].stages = mesa_to_vk_shader_stage(i);
5272 const char *name = NULL;
5273 const char *description = NULL;
5274 switch(i) {
5275 case MESA_SHADER_VERTEX:
5276 name = "Vertex Shader";
5277 description = "Vulkan Vertex Shader";
5278 break;
5279 case MESA_SHADER_TESS_CTRL:
5280 if (!pipeline->shaders[MESA_SHADER_VERTEX]) {
5281 pProperties[executable_idx].stages |= VK_SHADER_STAGE_VERTEX_BIT;
5282 name = "Vertex + Tessellation Control Shaders";
5283 description = "Combined Vulkan Vertex and Tessellation Control Shaders";
5284 } else {
5285 name = "Tessellation Control Shader";
5286 description = "Vulkan Tessellation Control Shader";
5287 }
5288 break;
5289 case MESA_SHADER_TESS_EVAL:
5290 name = "Tessellation Evaluation Shader";
5291 description = "Vulkan Tessellation Evaluation Shader";
5292 break;
5293 case MESA_SHADER_GEOMETRY:
5294 if (radv_pipeline_has_tess(pipeline) && !pipeline->shaders[MESA_SHADER_TESS_EVAL]) {
5295 pProperties[executable_idx].stages |= VK_SHADER_STAGE_TESSELLATION_EVALUATION_BIT;
5296 name = "Tessellation Evaluation + Geometry Shaders";
5297 description = "Combined Vulkan Tessellation Evaluation and Geometry Shaders";
5298 } else if (!radv_pipeline_has_tess(pipeline) && !pipeline->shaders[MESA_SHADER_VERTEX]) {
5299 pProperties[executable_idx].stages |= VK_SHADER_STAGE_VERTEX_BIT;
5300 name = "Vertex + Geometry Shader";
5301 description = "Combined Vulkan Vertex and Geometry Shaders";
5302 } else {
5303 name = "Geometry Shader";
5304 description = "Vulkan Geometry Shader";
5305 }
5306 break;
5307 case MESA_SHADER_FRAGMENT:
5308 name = "Fragment Shader";
5309 description = "Vulkan Fragment Shader";
5310 break;
5311 case MESA_SHADER_COMPUTE:
5312 name = "Compute Shader";
5313 description = "Vulkan Compute Shader";
5314 break;
5315 }
5316
5317 desc_copy(pProperties[executable_idx].name, name);
5318 desc_copy(pProperties[executable_idx].description, description);
5319
5320 ++executable_idx;
5321 if (i == MESA_SHADER_GEOMETRY &&
5322 !radv_pipeline_has_ngg(pipeline)) {
5323 assert(pipeline->gs_copy_shader);
5324 if (executable_idx >= count)
5325 break;
5326
5327 pProperties[executable_idx].stages = VK_SHADER_STAGE_GEOMETRY_BIT;
5328 desc_copy(pProperties[executable_idx].name, "GS Copy Shader");
5329 desc_copy(pProperties[executable_idx].description,
5330 "Extra shader stage that loads the GS output ringbuffer into the rasterizer");
5331
5332 ++executable_idx;
5333 }
5334 }
5335
5336 for (unsigned i = 0; i < count; ++i)
5337 pProperties[i].subgroupSize = 64;
5338
5339 VkResult result = *pExecutableCount < total_count ? VK_INCOMPLETE : VK_SUCCESS;
5340 *pExecutableCount = count;
5341 return result;
5342 }
5343
5344 VkResult radv_GetPipelineExecutableStatisticsKHR(
5345 VkDevice _device,
5346 const VkPipelineExecutableInfoKHR* pExecutableInfo,
5347 uint32_t* pStatisticCount,
5348 VkPipelineExecutableStatisticKHR* pStatistics)
5349 {
5350 RADV_FROM_HANDLE(radv_device, device, _device);
5351 RADV_FROM_HANDLE(radv_pipeline, pipeline, pExecutableInfo->pipeline);
5352 gl_shader_stage stage;
5353 struct radv_shader_variant *shader = radv_get_shader_from_executable_index(pipeline, pExecutableInfo->executableIndex, &stage);
5354
5355 enum chip_class chip_class = device->physical_device->rad_info.chip_class;
5356 unsigned lds_increment = chip_class >= GFX7 ? 512 : 256;
5357 unsigned max_waves = radv_get_max_waves(device, shader, stage);
5358
5359 VkPipelineExecutableStatisticKHR *s = pStatistics;
5360 VkPipelineExecutableStatisticKHR *end = s + (pStatistics ? *pStatisticCount : 0);
5361 VkResult result = VK_SUCCESS;
5362
5363 if (s < end) {
5364 desc_copy(s->name, "SGPRs");
5365 desc_copy(s->description, "Number of SGPR registers allocated per subgroup");
5366 s->format = VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR;
5367 s->value.u64 = shader->config.num_sgprs;
5368 }
5369 ++s;
5370
5371 if (s < end) {
5372 desc_copy(s->name, "VGPRs");
5373 desc_copy(s->description, "Number of VGPR registers allocated per subgroup");
5374 s->format = VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR;
5375 s->value.u64 = shader->config.num_vgprs;
5376 }
5377 ++s;
5378
5379 if (s < end) {
5380 desc_copy(s->name, "Spilled SGPRs");
5381 desc_copy(s->description, "Number of SGPR registers spilled per subgroup");
5382 s->format = VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR;
5383 s->value.u64 = shader->config.spilled_sgprs;
5384 }
5385 ++s;
5386
5387 if (s < end) {
5388 desc_copy(s->name, "Spilled VGPRs");
5389 desc_copy(s->description, "Number of VGPR registers spilled per subgroup");
5390 s->format = VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR;
5391 s->value.u64 = shader->config.spilled_vgprs;
5392 }
5393 ++s;
5394
5395 if (s < end) {
5396 desc_copy(s->name, "PrivMem VGPRs");
5397 desc_copy(s->description, "Number of VGPRs stored in private memory per subgroup");
5398 s->format = VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR;
5399 s->value.u64 = shader->info.private_mem_vgprs;
5400 }
5401 ++s;
5402
5403 if (s < end) {
5404 desc_copy(s->name, "Code size");
5405 desc_copy(s->description, "Code size in bytes");
5406 s->format = VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR;
5407 s->value.u64 = shader->exec_size;
5408 }
5409 ++s;
5410
5411 if (s < end) {
5412 desc_copy(s->name, "LDS size");
5413 desc_copy(s->description, "LDS size in bytes per workgroup");
5414 s->format = VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR;
5415 s->value.u64 = shader->config.lds_size * lds_increment;
5416 }
5417 ++s;
5418
5419 if (s < end) {
5420 desc_copy(s->name, "Scratch size");
5421 desc_copy(s->description, "Private memory in bytes per subgroup");
5422 s->format = VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR;
5423 s->value.u64 = shader->config.scratch_bytes_per_wave;
5424 }
5425 ++s;
5426
5427 if (s < end) {
5428 desc_copy(s->name, "Subgroups per SIMD");
5429 desc_copy(s->description, "The maximum number of subgroups in flight on a SIMD unit");
5430 s->format = VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR;
5431 s->value.u64 = max_waves;
5432 }
5433 ++s;
5434
5435 if (!pStatistics)
5436 *pStatisticCount = s - pStatistics;
5437 else if (s > end) {
5438 *pStatisticCount = end - pStatistics;
5439 result = VK_INCOMPLETE;
5440 } else {
5441 *pStatisticCount = s - pStatistics;
5442 }
5443
5444 return result;
5445 }
5446
5447 static VkResult radv_copy_representation(void *data, size_t *data_size, const char *src)
5448 {
5449 size_t total_size = strlen(src) + 1;
5450
5451 if (!data) {
5452 *data_size = total_size;
5453 return VK_SUCCESS;
5454 }
5455
5456 size_t size = MIN2(total_size, *data_size);
5457
5458 memcpy(data, src, size);
5459 if (size)
5460 *((char*)data + size - 1) = 0;
5461 return size < total_size ? VK_INCOMPLETE : VK_SUCCESS;
5462 }
5463
5464 VkResult radv_GetPipelineExecutableInternalRepresentationsKHR(
5465 VkDevice device,
5466 const VkPipelineExecutableInfoKHR* pExecutableInfo,
5467 uint32_t* pInternalRepresentationCount,
5468 VkPipelineExecutableInternalRepresentationKHR* pInternalRepresentations)
5469 {
5470 RADV_FROM_HANDLE(radv_pipeline, pipeline, pExecutableInfo->pipeline);
5471 gl_shader_stage stage;
5472 struct radv_shader_variant *shader = radv_get_shader_from_executable_index(pipeline, pExecutableInfo->executableIndex, &stage);
5473
5474 VkPipelineExecutableInternalRepresentationKHR *p = pInternalRepresentations;
5475 VkPipelineExecutableInternalRepresentationKHR *end = p + (pInternalRepresentations ? *pInternalRepresentationCount : 0);
5476 VkResult result = VK_SUCCESS;
5477 /* optimized NIR */
5478 if (p < end) {
5479 p->isText = true;
5480 desc_copy(p->name, "NIR Shader(s)");
5481 desc_copy(p->description, "The optimized NIR shader(s)");
5482 if (radv_copy_representation(p->pData, &p->dataSize, shader->nir_string) != VK_SUCCESS)
5483 result = VK_INCOMPLETE;
5484 }
5485 ++p;
5486
5487 /* backend IR */
5488 if (p < end) {
5489 p->isText = true;
5490 if (shader->aco_used) {
5491 desc_copy(p->name, "ACO IR");
5492 desc_copy(p->description, "The ACO IR after some optimizations");
5493 } else {
5494 desc_copy(p->name, "LLVM IR");
5495 desc_copy(p->description, "The LLVM IR after some optimizations");
5496 }
5497 if (radv_copy_representation(p->pData, &p->dataSize, shader->ir_string) != VK_SUCCESS)
5498 result = VK_INCOMPLETE;
5499 }
5500 ++p;
5501
5502 /* Disassembler */
5503 if (p < end) {
5504 p->isText = true;
5505 desc_copy(p->name, "Assembly");
5506 desc_copy(p->description, "Final Assembly");
5507 if (radv_copy_representation(p->pData, &p->dataSize, shader->disasm_string) != VK_SUCCESS)
5508 result = VK_INCOMPLETE;
5509 }
5510 ++p;
5511
5512 if (!pInternalRepresentations)
5513 *pInternalRepresentationCount = p - pInternalRepresentations;
5514 else if(p > end) {
5515 result = VK_INCOMPLETE;
5516 *pInternalRepresentationCount = end - pInternalRepresentations;
5517 } else {
5518 *pInternalRepresentationCount = p - pInternalRepresentations;
5519 }
5520
5521 return result;
5522 }