projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
radv: set BYPASS_VTX_RATE_COMBINER_GFX103 on GFX 10.3
[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 "sid.h"
42 #include "ac_binary.h"
43 #include "ac_llvm_util.h"
44 #include "ac_nir_to_llvm.h"
45 #include "vk_format.h"
46 #include "util/debug.h"
47 #include "ac_exp_param.h"
48 #include "ac_shader_util.h"
49
50 struct radv_blend_state {
51 uint32_t blend_enable_4bit;
52 uint32_t need_src_alpha;
53
54 uint32_t cb_color_control;
55 uint32_t cb_target_mask;
56 uint32_t cb_target_enabled_4bit;
57 uint32_t sx_mrt_blend_opt[8];
58 uint32_t cb_blend_control[8];
59
60 uint32_t spi_shader_col_format;
61 uint32_t col_format_is_int8;
62 uint32_t col_format_is_int10;
63 uint32_t cb_shader_mask;
64 uint32_t db_alpha_to_mask;
65
66 uint32_t commutative_4bit;
67
68 bool single_cb_enable;
69 bool mrt0_is_dual_src;
70 };
71
72 struct radv_dsa_order_invariance {
73 /* Whether the final result in Z/S buffers is guaranteed to be
74 * invariant under changes to the order in which fragments arrive.
75 */
76 bool zs;
77
78 /* Whether the set of fragments that pass the combined Z/S test is
79 * guaranteed to be invariant under changes to the order in which
80 * fragments arrive.
81 */
82 bool pass_set;
83 };
84
85 static const VkPipelineMultisampleStateCreateInfo *
86 radv_pipeline_get_multisample_state(const VkGraphicsPipelineCreateInfo *pCreateInfo)
87 {
88 if (!pCreateInfo->pRasterizationState->rasterizerDiscardEnable)
89 return pCreateInfo->pMultisampleState;
90 return NULL;
91 }
92
93 static const VkPipelineTessellationStateCreateInfo *
94 radv_pipeline_get_tessellation_state(const VkGraphicsPipelineCreateInfo *pCreateInfo)
95 {
96 for (uint32_t i = 0; i < pCreateInfo->stageCount; i++) {
97 if (pCreateInfo->pStages[i].stage == VK_SHADER_STAGE_TESSELLATION_CONTROL_BIT ||
98 pCreateInfo->pStages[i].stage == VK_SHADER_STAGE_TESSELLATION_EVALUATION_BIT) {
99 return pCreateInfo->pTessellationState;
100 }
101 }
102 return NULL;
103 }
104
105 static const VkPipelineDepthStencilStateCreateInfo *
106 radv_pipeline_get_depth_stencil_state(const VkGraphicsPipelineCreateInfo *pCreateInfo)
107 {
108 RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass);
109 struct radv_subpass *subpass = pass->subpasses + pCreateInfo->subpass;
110
111 if (!pCreateInfo->pRasterizationState->rasterizerDiscardEnable &&
112 subpass->depth_stencil_attachment)
113 return pCreateInfo->pDepthStencilState;
114 return NULL;
115 }
116
117 static const VkPipelineColorBlendStateCreateInfo *
118 radv_pipeline_get_color_blend_state(const VkGraphicsPipelineCreateInfo *pCreateInfo)
119 {
120 RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass);
121 struct radv_subpass *subpass = pass->subpasses + pCreateInfo->subpass;
122
123 if (!pCreateInfo->pRasterizationState->rasterizerDiscardEnable &&
124 subpass->has_color_att)
125 return pCreateInfo->pColorBlendState;
126 return NULL;
127 }
128
129 bool radv_pipeline_has_ngg(const struct radv_pipeline *pipeline)
130 {
131 struct radv_shader_variant *variant = NULL;
132 if (pipeline->shaders[MESA_SHADER_GEOMETRY])
133 variant = pipeline->shaders[MESA_SHADER_GEOMETRY];
134 else if (pipeline->shaders[MESA_SHADER_TESS_EVAL])
135 variant = pipeline->shaders[MESA_SHADER_TESS_EVAL];
136 else if (pipeline->shaders[MESA_SHADER_VERTEX])
137 variant = pipeline->shaders[MESA_SHADER_VERTEX];
138 else
139 return false;
140 return variant->info.is_ngg;
141 }
142
143 bool radv_pipeline_has_ngg_passthrough(const struct radv_pipeline *pipeline)
144 {
145 assert(radv_pipeline_has_ngg(pipeline));
146
147 struct radv_shader_variant *variant = NULL;
148 if (pipeline->shaders[MESA_SHADER_GEOMETRY])
149 variant = pipeline->shaders[MESA_SHADER_GEOMETRY];
150 else if (pipeline->shaders[MESA_SHADER_TESS_EVAL])
151 variant = pipeline->shaders[MESA_SHADER_TESS_EVAL];
152 else if (pipeline->shaders[MESA_SHADER_VERTEX])
153 variant = pipeline->shaders[MESA_SHADER_VERTEX];
154 else
155 return false;
156 return variant->info.is_ngg_passthrough;
157 }
158
159 bool radv_pipeline_has_gs_copy_shader(const struct radv_pipeline *pipeline)
160 {
161 if (!radv_pipeline_has_gs(pipeline))
162 return false;
163
164 /* The GS copy shader is required if the pipeline has GS on GFX6-GFX9.
165 * On GFX10, it might be required in rare cases if it's not possible to
166 * enable NGG.
167 */
168 if (radv_pipeline_has_ngg(pipeline))
169 return false;
170
171 assert(pipeline->gs_copy_shader);
172 return true;
173 }
174
175 static void
176 radv_pipeline_destroy(struct radv_device *device,
177 struct radv_pipeline *pipeline,
178 const VkAllocationCallbacks* allocator)
179 {
180 for (unsigned i = 0; i < MESA_SHADER_STAGES; ++i)
181 if (pipeline->shaders[i])
182 radv_shader_variant_destroy(device, pipeline->shaders[i]);
183
184 if (pipeline->gs_copy_shader)
185 radv_shader_variant_destroy(device, pipeline->gs_copy_shader);
186
187 if(pipeline->cs.buf)
188 free(pipeline->cs.buf);
189
190 vk_object_base_finish(&pipeline->base);
191 vk_free2(&device->vk.alloc, allocator, pipeline);
192 }
193
194 void radv_DestroyPipeline(
195 VkDevice _device,
196 VkPipeline _pipeline,
197 const VkAllocationCallbacks* pAllocator)
198 {
199 RADV_FROM_HANDLE(radv_device, device, _device);
200 RADV_FROM_HANDLE(radv_pipeline, pipeline, _pipeline);
201
202 if (!_pipeline)
203 return;
204
205 radv_pipeline_destroy(device, pipeline, pAllocator);
206 }
207
208 static uint32_t get_hash_flags(struct radv_device *device)
209 {
210 uint32_t hash_flags = 0;
211
212 if (device->instance->debug_flags & RADV_DEBUG_NO_NGG)
213 hash_flags |= RADV_HASH_SHADER_NO_NGG;
214 if (device->physical_device->cs_wave_size == 32)
215 hash_flags |= RADV_HASH_SHADER_CS_WAVE32;
216 if (device->physical_device->ps_wave_size == 32)
217 hash_flags |= RADV_HASH_SHADER_PS_WAVE32;
218 if (device->physical_device->ge_wave_size == 32)
219 hash_flags |= RADV_HASH_SHADER_GE_WAVE32;
220 if (device->physical_device->use_llvm)
221 hash_flags |= RADV_HASH_SHADER_LLVM;
222 return hash_flags;
223 }
224
225 static void
226 radv_pipeline_init_scratch(struct radv_device *device,
227 struct radv_pipeline *pipeline)
228 {
229 unsigned scratch_bytes_per_wave = 0;
230 unsigned max_waves = 0;
231 unsigned min_waves = 1;
232
233 for (int i = 0; i < MESA_SHADER_STAGES; ++i) {
234 if (pipeline->shaders[i] &&
235 pipeline->shaders[i]->config.scratch_bytes_per_wave) {
236 unsigned max_stage_waves = device->scratch_waves;
237
238 scratch_bytes_per_wave = MAX2(scratch_bytes_per_wave,
239 pipeline->shaders[i]->config.scratch_bytes_per_wave);
240
241 max_stage_waves = MIN2(max_stage_waves,
242 4 * device->physical_device->rad_info.num_good_compute_units *
243 (256 / pipeline->shaders[i]->config.num_vgprs));
244 max_waves = MAX2(max_waves, max_stage_waves);
245 }
246 }
247
248 if (pipeline->shaders[MESA_SHADER_COMPUTE]) {
249 unsigned group_size = pipeline->shaders[MESA_SHADER_COMPUTE]->info.cs.block_size[0] *
250 pipeline->shaders[MESA_SHADER_COMPUTE]->info.cs.block_size[1] *
251 pipeline->shaders[MESA_SHADER_COMPUTE]->info.cs.block_size[2];
252 min_waves = MAX2(min_waves, round_up_u32(group_size, 64));
253 }
254
255 pipeline->scratch_bytes_per_wave = scratch_bytes_per_wave;
256 pipeline->max_waves = max_waves;
257 }
258
259 static uint32_t si_translate_blend_logic_op(VkLogicOp op)
260 {
261 switch (op) {
262 case VK_LOGIC_OP_CLEAR:
263 return V_028808_ROP3_CLEAR;
264 case VK_LOGIC_OP_AND:
265 return V_028808_ROP3_AND;
266 case VK_LOGIC_OP_AND_REVERSE:
267 return V_028808_ROP3_AND_REVERSE;
268 case VK_LOGIC_OP_COPY:
269 return V_028808_ROP3_COPY;
270 case VK_LOGIC_OP_AND_INVERTED:
271 return V_028808_ROP3_AND_INVERTED;
272 case VK_LOGIC_OP_NO_OP:
273 return V_028808_ROP3_NO_OP;
274 case VK_LOGIC_OP_XOR:
275 return V_028808_ROP3_XOR;
276 case VK_LOGIC_OP_OR:
277 return V_028808_ROP3_OR;
278 case VK_LOGIC_OP_NOR:
279 return V_028808_ROP3_NOR;
280 case VK_LOGIC_OP_EQUIVALENT:
281 return V_028808_ROP3_EQUIVALENT;
282 case VK_LOGIC_OP_INVERT:
283 return V_028808_ROP3_INVERT;
284 case VK_LOGIC_OP_OR_REVERSE:
285 return V_028808_ROP3_OR_REVERSE;
286 case VK_LOGIC_OP_COPY_INVERTED:
287 return V_028808_ROP3_COPY_INVERTED;
288 case VK_LOGIC_OP_OR_INVERTED:
289 return V_028808_ROP3_OR_INVERTED;
290 case VK_LOGIC_OP_NAND:
291 return V_028808_ROP3_NAND;
292 case VK_LOGIC_OP_SET:
293 return V_028808_ROP3_SET;
294 default:
295 unreachable("Unhandled logic op");
296 }
297 }
298
299
300 static uint32_t si_translate_blend_function(VkBlendOp op)
301 {
302 switch (op) {
303 case VK_BLEND_OP_ADD:
304 return V_028780_COMB_DST_PLUS_SRC;
305 case VK_BLEND_OP_SUBTRACT:
306 return V_028780_COMB_SRC_MINUS_DST;
307 case VK_BLEND_OP_REVERSE_SUBTRACT:
308 return V_028780_COMB_DST_MINUS_SRC;
309 case VK_BLEND_OP_MIN:
310 return V_028780_COMB_MIN_DST_SRC;
311 case VK_BLEND_OP_MAX:
312 return V_028780_COMB_MAX_DST_SRC;
313 default:
314 return 0;
315 }
316 }
317
318 static uint32_t si_translate_blend_factor(VkBlendFactor factor)
319 {
320 switch (factor) {
321 case VK_BLEND_FACTOR_ZERO:
322 return V_028780_BLEND_ZERO;
323 case VK_BLEND_FACTOR_ONE:
324 return V_028780_BLEND_ONE;
325 case VK_BLEND_FACTOR_SRC_COLOR:
326 return V_028780_BLEND_SRC_COLOR;
327 case VK_BLEND_FACTOR_ONE_MINUS_SRC_COLOR:
328 return V_028780_BLEND_ONE_MINUS_SRC_COLOR;
329 case VK_BLEND_FACTOR_DST_COLOR:
330 return V_028780_BLEND_DST_COLOR;
331 case VK_BLEND_FACTOR_ONE_MINUS_DST_COLOR:
332 return V_028780_BLEND_ONE_MINUS_DST_COLOR;
333 case VK_BLEND_FACTOR_SRC_ALPHA:
334 return V_028780_BLEND_SRC_ALPHA;
335 case VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA:
336 return V_028780_BLEND_ONE_MINUS_SRC_ALPHA;
337 case VK_BLEND_FACTOR_DST_ALPHA:
338 return V_028780_BLEND_DST_ALPHA;
339 case VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA:
340 return V_028780_BLEND_ONE_MINUS_DST_ALPHA;
341 case VK_BLEND_FACTOR_CONSTANT_COLOR:
342 return V_028780_BLEND_CONSTANT_COLOR;
343 case VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_COLOR:
344 return V_028780_BLEND_ONE_MINUS_CONSTANT_COLOR;
345 case VK_BLEND_FACTOR_CONSTANT_ALPHA:
346 return V_028780_BLEND_CONSTANT_ALPHA;
347 case VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_ALPHA:
348 return V_028780_BLEND_ONE_MINUS_CONSTANT_ALPHA;
349 case VK_BLEND_FACTOR_SRC_ALPHA_SATURATE:
350 return V_028780_BLEND_SRC_ALPHA_SATURATE;
351 case VK_BLEND_FACTOR_SRC1_COLOR:
352 return V_028780_BLEND_SRC1_COLOR;
353 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_COLOR:
354 return V_028780_BLEND_INV_SRC1_COLOR;
355 case VK_BLEND_FACTOR_SRC1_ALPHA:
356 return V_028780_BLEND_SRC1_ALPHA;
357 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA:
358 return V_028780_BLEND_INV_SRC1_ALPHA;
359 default:
360 return 0;
361 }
362 }
363
364 static uint32_t si_translate_blend_opt_function(VkBlendOp op)
365 {
366 switch (op) {
367 case VK_BLEND_OP_ADD:
368 return V_028760_OPT_COMB_ADD;
369 case VK_BLEND_OP_SUBTRACT:
370 return V_028760_OPT_COMB_SUBTRACT;
371 case VK_BLEND_OP_REVERSE_SUBTRACT:
372 return V_028760_OPT_COMB_REVSUBTRACT;
373 case VK_BLEND_OP_MIN:
374 return V_028760_OPT_COMB_MIN;
375 case VK_BLEND_OP_MAX:
376 return V_028760_OPT_COMB_MAX;
377 default:
378 return V_028760_OPT_COMB_BLEND_DISABLED;
379 }
380 }
381
382 static uint32_t si_translate_blend_opt_factor(VkBlendFactor factor, bool is_alpha)
383 {
384 switch (factor) {
385 case VK_BLEND_FACTOR_ZERO:
386 return V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_ALL;
387 case VK_BLEND_FACTOR_ONE:
388 return V_028760_BLEND_OPT_PRESERVE_ALL_IGNORE_NONE;
389 case VK_BLEND_FACTOR_SRC_COLOR:
390 return is_alpha ? V_028760_BLEND_OPT_PRESERVE_A1_IGNORE_A0
391 : V_028760_BLEND_OPT_PRESERVE_C1_IGNORE_C0;
392 case VK_BLEND_FACTOR_ONE_MINUS_SRC_COLOR:
393 return is_alpha ? V_028760_BLEND_OPT_PRESERVE_A0_IGNORE_A1
394 : V_028760_BLEND_OPT_PRESERVE_C0_IGNORE_C1;
395 case VK_BLEND_FACTOR_SRC_ALPHA:
396 return V_028760_BLEND_OPT_PRESERVE_A1_IGNORE_A0;
397 case VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA:
398 return V_028760_BLEND_OPT_PRESERVE_A0_IGNORE_A1;
399 case VK_BLEND_FACTOR_SRC_ALPHA_SATURATE:
400 return is_alpha ? V_028760_BLEND_OPT_PRESERVE_ALL_IGNORE_NONE
401 : V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_A0;
402 default:
403 return V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE;
404 }
405 }
406
407 /**
408 * Get rid of DST in the blend factors by commuting the operands:
409 * func(src * DST, dst * 0) ---> func(src * 0, dst * SRC)
410 */
411 static void si_blend_remove_dst(unsigned *func, unsigned *src_factor,
412 unsigned *dst_factor, unsigned expected_dst,
413 unsigned replacement_src)
414 {
415 if (*src_factor == expected_dst &&
416 *dst_factor == VK_BLEND_FACTOR_ZERO) {
417 *src_factor = VK_BLEND_FACTOR_ZERO;
418 *dst_factor = replacement_src;
419
420 /* Commuting the operands requires reversing subtractions. */
421 if (*func == VK_BLEND_OP_SUBTRACT)
422 *func = VK_BLEND_OP_REVERSE_SUBTRACT;
423 else if (*func == VK_BLEND_OP_REVERSE_SUBTRACT)
424 *func = VK_BLEND_OP_SUBTRACT;
425 }
426 }
427
428 static bool si_blend_factor_uses_dst(unsigned factor)
429 {
430 return factor == VK_BLEND_FACTOR_DST_COLOR ||
431 factor == VK_BLEND_FACTOR_DST_ALPHA ||
432 factor == VK_BLEND_FACTOR_SRC_ALPHA_SATURATE ||
433 factor == VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA ||
434 factor == VK_BLEND_FACTOR_ONE_MINUS_DST_COLOR;
435 }
436
437 static bool is_dual_src(VkBlendFactor factor)
438 {
439 switch (factor) {
440 case VK_BLEND_FACTOR_SRC1_COLOR:
441 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_COLOR:
442 case VK_BLEND_FACTOR_SRC1_ALPHA:
443 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA:
444 return true;
445 default:
446 return false;
447 }
448 }
449
450 static unsigned radv_choose_spi_color_format(VkFormat vk_format,
451 bool blend_enable,
452 bool blend_need_alpha)
453 {
454 const struct vk_format_description *desc = vk_format_description(vk_format);
455 struct ac_spi_color_formats formats = {};
456 unsigned format, ntype, swap;
457
458 format = radv_translate_colorformat(vk_format);
459 ntype = radv_translate_color_numformat(vk_format, desc,
460 vk_format_get_first_non_void_channel(vk_format));
461 swap = radv_translate_colorswap(vk_format, false);
462
463 ac_choose_spi_color_formats(format, swap, ntype, false, &formats);
464
465 if (blend_enable && blend_need_alpha)
466 return formats.blend_alpha;
467 else if(blend_need_alpha)
468 return formats.alpha;
469 else if(blend_enable)
470 return formats.blend;
471 else
472 return formats.normal;
473 }
474
475 static bool
476 format_is_int8(VkFormat format)
477 {
478 const struct vk_format_description *desc = vk_format_description(format);
479 int channel = vk_format_get_first_non_void_channel(format);
480
481 return channel >= 0 && desc->channel[channel].pure_integer &&
482 desc->channel[channel].size == 8;
483 }
484
485 static bool
486 format_is_int10(VkFormat format)
487 {
488 const struct vk_format_description *desc = vk_format_description(format);
489
490 if (desc->nr_channels != 4)
491 return false;
492 for (unsigned i = 0; i < 4; i++) {
493 if (desc->channel[i].pure_integer && desc->channel[i].size == 10)
494 return true;
495 }
496 return false;
497 }
498
499 static void
500 radv_pipeline_compute_spi_color_formats(struct radv_pipeline *pipeline,
501 const VkGraphicsPipelineCreateInfo *pCreateInfo,
502 struct radv_blend_state *blend)
503 {
504 RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass);
505 struct radv_subpass *subpass = pass->subpasses + pCreateInfo->subpass;
506 unsigned col_format = 0, is_int8 = 0, is_int10 = 0;
507 unsigned num_targets;
508
509 for (unsigned i = 0; i < (blend->single_cb_enable ? 1 : subpass->color_count); ++i) {
510 unsigned cf;
511
512 if (subpass->color_attachments[i].attachment == VK_ATTACHMENT_UNUSED ||
513 !(blend->cb_target_mask & (0xfu << (i * 4)))) {
514 cf = V_028714_SPI_SHADER_ZERO;
515 } else {
516 struct radv_render_pass_attachment *attachment = pass->attachments + subpass->color_attachments[i].attachment;
517 bool blend_enable =
518 blend->blend_enable_4bit & (0xfu << (i * 4));
519
520 cf = radv_choose_spi_color_format(attachment->format,
521 blend_enable,
522 blend->need_src_alpha & (1 << i));
523
524 if (format_is_int8(attachment->format))
525 is_int8 |= 1 << i;
526 if (format_is_int10(attachment->format))
527 is_int10 |= 1 << i;
528 }
529
530 col_format |= cf << (4 * i);
531 }
532
533 if (!(col_format & 0xf) && blend->need_src_alpha & (1 << 0)) {
534 /* When a subpass doesn't have any color attachments, write the
535 * alpha channel of MRT0 when alpha coverage is enabled because
536 * the depth attachment needs it.
537 */
538 col_format |= V_028714_SPI_SHADER_32_AR;
539 }
540
541 /* If the i-th target format is set, all previous target formats must
542 * be non-zero to avoid hangs.
543 */
544 num_targets = (util_last_bit(col_format) + 3) / 4;
545 for (unsigned i = 0; i < num_targets; i++) {
546 if (!(col_format & (0xf << (i * 4)))) {
547 col_format |= V_028714_SPI_SHADER_32_R << (i * 4);
548 }
549 }
550
551 /* The output for dual source blending should have the same format as
552 * the first output.
553 */
554 if (blend->mrt0_is_dual_src)
555 col_format |= (col_format & 0xf) << 4;
556
557 blend->spi_shader_col_format = col_format;
558 blend->col_format_is_int8 = is_int8;
559 blend->col_format_is_int10 = is_int10;
560 }
561
562 /*
563 * Ordered so that for each i,
564 * radv_format_meta_fs_key(radv_fs_key_format_exemplars[i]) == i.
565 */
566 const VkFormat radv_fs_key_format_exemplars[NUM_META_FS_KEYS] = {
567 VK_FORMAT_R32_SFLOAT,
568 VK_FORMAT_R32G32_SFLOAT,
569 VK_FORMAT_R8G8B8A8_UNORM,
570 VK_FORMAT_R16G16B16A16_UNORM,
571 VK_FORMAT_R16G16B16A16_SNORM,
572 VK_FORMAT_R16G16B16A16_UINT,
573 VK_FORMAT_R16G16B16A16_SINT,
574 VK_FORMAT_R32G32B32A32_SFLOAT,
575 VK_FORMAT_R8G8B8A8_UINT,
576 VK_FORMAT_R8G8B8A8_SINT,
577 VK_FORMAT_A2R10G10B10_UINT_PACK32,
578 VK_FORMAT_A2R10G10B10_SINT_PACK32,
579 };
580
581 unsigned radv_format_meta_fs_key(VkFormat format)
582 {
583 unsigned col_format = radv_choose_spi_color_format(format, false, false);
584
585 assert(col_format != V_028714_SPI_SHADER_32_AR);
586 if (col_format >= V_028714_SPI_SHADER_32_AR)
587 --col_format; /* Skip V_028714_SPI_SHADER_32_AR since there is no such VkFormat */
588
589 --col_format; /* Skip V_028714_SPI_SHADER_ZERO */
590 bool is_int8 = format_is_int8(format);
591 bool is_int10 = format_is_int10(format);
592
593 return col_format + (is_int8 ? 3 : is_int10 ? 5 : 0);
594 }
595
596 static void
597 radv_blend_check_commutativity(struct radv_blend_state *blend,
598 VkBlendOp op, VkBlendFactor src,
599 VkBlendFactor dst, unsigned chanmask)
600 {
601 /* Src factor is allowed when it does not depend on Dst. */
602 static const uint32_t src_allowed =
603 (1u << VK_BLEND_FACTOR_ONE) |
604 (1u << VK_BLEND_FACTOR_SRC_COLOR) |
605 (1u << VK_BLEND_FACTOR_SRC_ALPHA) |
606 (1u << VK_BLEND_FACTOR_SRC_ALPHA_SATURATE) |
607 (1u << VK_BLEND_FACTOR_CONSTANT_COLOR) |
608 (1u << VK_BLEND_FACTOR_CONSTANT_ALPHA) |
609 (1u << VK_BLEND_FACTOR_SRC1_COLOR) |
610 (1u << VK_BLEND_FACTOR_SRC1_ALPHA) |
611 (1u << VK_BLEND_FACTOR_ZERO) |
612 (1u << VK_BLEND_FACTOR_ONE_MINUS_SRC_COLOR) |
613 (1u << VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA) |
614 (1u << VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_COLOR) |
615 (1u << VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_ALPHA) |
616 (1u << VK_BLEND_FACTOR_ONE_MINUS_SRC1_COLOR) |
617 (1u << VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA);
618
619 if (dst == VK_BLEND_FACTOR_ONE &&
620 (src_allowed & (1u << src))) {
621 /* Addition is commutative, but floating point addition isn't
622 * associative: subtle changes can be introduced via different
623 * rounding. Be conservative, only enable for min and max.
624 */
625 if (op == VK_BLEND_OP_MAX || op == VK_BLEND_OP_MIN)
626 blend->commutative_4bit |= chanmask;
627 }
628 }
629
630 static struct radv_blend_state
631 radv_pipeline_init_blend_state(struct radv_pipeline *pipeline,
632 const VkGraphicsPipelineCreateInfo *pCreateInfo,
633 const struct radv_graphics_pipeline_create_info *extra)
634 {
635 const VkPipelineColorBlendStateCreateInfo *vkblend = radv_pipeline_get_color_blend_state(pCreateInfo);
636 const VkPipelineMultisampleStateCreateInfo *vkms = radv_pipeline_get_multisample_state(pCreateInfo);
637 struct radv_blend_state blend = {0};
638 unsigned mode = V_028808_CB_NORMAL;
639 int i;
640
641 if (extra && extra->custom_blend_mode) {
642 blend.single_cb_enable = true;
643 mode = extra->custom_blend_mode;
644 }
645
646 blend.cb_color_control = 0;
647 if (vkblend) {
648 if (vkblend->logicOpEnable)
649 blend.cb_color_control |= S_028808_ROP3(si_translate_blend_logic_op(vkblend->logicOp));
650 else
651 blend.cb_color_control |= S_028808_ROP3(V_028808_ROP3_COPY);
652 }
653
654 blend.db_alpha_to_mask = S_028B70_ALPHA_TO_MASK_OFFSET0(3) |
655 S_028B70_ALPHA_TO_MASK_OFFSET1(1) |
656 S_028B70_ALPHA_TO_MASK_OFFSET2(0) |
657 S_028B70_ALPHA_TO_MASK_OFFSET3(2) |
658 S_028B70_OFFSET_ROUND(1);
659
660 if (vkms && vkms->alphaToCoverageEnable) {
661 blend.db_alpha_to_mask |= S_028B70_ALPHA_TO_MASK_ENABLE(1);
662 blend.need_src_alpha |= 0x1;
663 }
664
665 blend.cb_target_mask = 0;
666 if (vkblend) {
667 for (i = 0; i < vkblend->attachmentCount; i++) {
668 const VkPipelineColorBlendAttachmentState *att = &vkblend->pAttachments[i];
669 unsigned blend_cntl = 0;
670 unsigned srcRGB_opt, dstRGB_opt, srcA_opt, dstA_opt;
671 VkBlendOp eqRGB = att->colorBlendOp;
672 VkBlendFactor srcRGB = att->srcColorBlendFactor;
673 VkBlendFactor dstRGB = att->dstColorBlendFactor;
674 VkBlendOp eqA = att->alphaBlendOp;
675 VkBlendFactor srcA = att->srcAlphaBlendFactor;
676 VkBlendFactor dstA = att->dstAlphaBlendFactor;
677
678 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);
679
680 if (!att->colorWriteMask)
681 continue;
682
683 blend.cb_target_mask |= (unsigned)att->colorWriteMask << (4 * i);
684 blend.cb_target_enabled_4bit |= 0xf << (4 * i);
685 if (!att->blendEnable) {
686 blend.cb_blend_control[i] = blend_cntl;
687 continue;
688 }
689
690 if (is_dual_src(srcRGB) || is_dual_src(dstRGB) || is_dual_src(srcA) || is_dual_src(dstA))
691 if (i == 0)
692 blend.mrt0_is_dual_src = true;
693
694 if (eqRGB == VK_BLEND_OP_MIN || eqRGB == VK_BLEND_OP_MAX) {
695 srcRGB = VK_BLEND_FACTOR_ONE;
696 dstRGB = VK_BLEND_FACTOR_ONE;
697 }
698 if (eqA == VK_BLEND_OP_MIN || eqA == VK_BLEND_OP_MAX) {
699 srcA = VK_BLEND_FACTOR_ONE;
700 dstA = VK_BLEND_FACTOR_ONE;
701 }
702
703 radv_blend_check_commutativity(&blend, eqRGB, srcRGB, dstRGB,
704 0x7 << (4 * i));
705 radv_blend_check_commutativity(&blend, eqA, srcA, dstA,
706 0x8 << (4 * i));
707
708 /* Blending optimizations for RB+.
709 * These transformations don't change the behavior.
710 *
711 * First, get rid of DST in the blend factors:
712 * func(src * DST, dst * 0) ---> func(src * 0, dst * SRC)
713 */
714 si_blend_remove_dst(&eqRGB, &srcRGB, &dstRGB,
715 VK_BLEND_FACTOR_DST_COLOR,
716 VK_BLEND_FACTOR_SRC_COLOR);
717
718 si_blend_remove_dst(&eqA, &srcA, &dstA,
719 VK_BLEND_FACTOR_DST_COLOR,
720 VK_BLEND_FACTOR_SRC_COLOR);
721
722 si_blend_remove_dst(&eqA, &srcA, &dstA,
723 VK_BLEND_FACTOR_DST_ALPHA,
724 VK_BLEND_FACTOR_SRC_ALPHA);
725
726 /* Look up the ideal settings from tables. */
727 srcRGB_opt = si_translate_blend_opt_factor(srcRGB, false);
728 dstRGB_opt = si_translate_blend_opt_factor(dstRGB, false);
729 srcA_opt = si_translate_blend_opt_factor(srcA, true);
730 dstA_opt = si_translate_blend_opt_factor(dstA, true);
731
732 /* Handle interdependencies. */
733 if (si_blend_factor_uses_dst(srcRGB))
734 dstRGB_opt = V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE;
735 if (si_blend_factor_uses_dst(srcA))
736 dstA_opt = V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE;
737
738 if (srcRGB == VK_BLEND_FACTOR_SRC_ALPHA_SATURATE &&
739 (dstRGB == VK_BLEND_FACTOR_ZERO ||
740 dstRGB == VK_BLEND_FACTOR_SRC_ALPHA ||
741 dstRGB == VK_BLEND_FACTOR_SRC_ALPHA_SATURATE))
742 dstRGB_opt = V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_A0;
743
744 /* Set the final value. */
745 blend.sx_mrt_blend_opt[i] =
746 S_028760_COLOR_SRC_OPT(srcRGB_opt) |
747 S_028760_COLOR_DST_OPT(dstRGB_opt) |
748 S_028760_COLOR_COMB_FCN(si_translate_blend_opt_function(eqRGB)) |
749 S_028760_ALPHA_SRC_OPT(srcA_opt) |
750 S_028760_ALPHA_DST_OPT(dstA_opt) |
751 S_028760_ALPHA_COMB_FCN(si_translate_blend_opt_function(eqA));
752 blend_cntl |= S_028780_ENABLE(1);
753
754 blend_cntl |= S_028780_COLOR_COMB_FCN(si_translate_blend_function(eqRGB));
755 blend_cntl |= S_028780_COLOR_SRCBLEND(si_translate_blend_factor(srcRGB));
756 blend_cntl |= S_028780_COLOR_DESTBLEND(si_translate_blend_factor(dstRGB));
757 if (srcA != srcRGB || dstA != dstRGB || eqA != eqRGB) {
758 blend_cntl |= S_028780_SEPARATE_ALPHA_BLEND(1);
759 blend_cntl |= S_028780_ALPHA_COMB_FCN(si_translate_blend_function(eqA));
760 blend_cntl |= S_028780_ALPHA_SRCBLEND(si_translate_blend_factor(srcA));
761 blend_cntl |= S_028780_ALPHA_DESTBLEND(si_translate_blend_factor(dstA));
762 }
763 blend.cb_blend_control[i] = blend_cntl;
764
765 blend.blend_enable_4bit |= 0xfu << (i * 4);
766
767 if (srcRGB == VK_BLEND_FACTOR_SRC_ALPHA ||
768 dstRGB == VK_BLEND_FACTOR_SRC_ALPHA ||
769 srcRGB == VK_BLEND_FACTOR_SRC_ALPHA_SATURATE ||
770 dstRGB == VK_BLEND_FACTOR_SRC_ALPHA_SATURATE ||
771 srcRGB == VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA ||
772 dstRGB == VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA)
773 blend.need_src_alpha |= 1 << i;
774 }
775 for (i = vkblend->attachmentCount; i < 8; i++) {
776 blend.cb_blend_control[i] = 0;
777 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);
778 }
779 }
780
781 if (pipeline->device->physical_device->rad_info.has_rbplus) {
782 /* Disable RB+ blend optimizations for dual source blending. */
783 if (blend.mrt0_is_dual_src) {
784 for (i = 0; i < 8; i++) {
785 blend.sx_mrt_blend_opt[i] =
786 S_028760_COLOR_COMB_FCN(V_028760_OPT_COMB_NONE) |
787 S_028760_ALPHA_COMB_FCN(V_028760_OPT_COMB_NONE);
788 }
789 }
790
791 /* RB+ doesn't work with dual source blending, logic op and
792 * RESOLVE.
793 */
794 if (blend.mrt0_is_dual_src ||
795 (vkblend && vkblend->logicOpEnable) ||
796 mode == V_028808_CB_RESOLVE)
797 blend.cb_color_control |= S_028808_DISABLE_DUAL_QUAD(1);
798 }
799
800 if (blend.cb_target_mask)
801 blend.cb_color_control |= S_028808_MODE(mode);
802 else
803 blend.cb_color_control |= S_028808_MODE(V_028808_CB_DISABLE);
804
805 radv_pipeline_compute_spi_color_formats(pipeline, pCreateInfo, &blend);
806 return blend;
807 }
808
809 static uint32_t si_translate_fill(VkPolygonMode func)
810 {
811 switch(func) {
812 case VK_POLYGON_MODE_FILL:
813 return V_028814_X_DRAW_TRIANGLES;
814 case VK_POLYGON_MODE_LINE:
815 return V_028814_X_DRAW_LINES;
816 case VK_POLYGON_MODE_POINT:
817 return V_028814_X_DRAW_POINTS;
818 default:
819 assert(0);
820 return V_028814_X_DRAW_POINTS;
821 }
822 }
823
824 static uint8_t radv_pipeline_get_ps_iter_samples(const VkGraphicsPipelineCreateInfo *pCreateInfo)
825 {
826 const VkPipelineMultisampleStateCreateInfo *vkms = pCreateInfo->pMultisampleState;
827 RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass);
828 struct radv_subpass *subpass = &pass->subpasses[pCreateInfo->subpass];
829 uint32_t ps_iter_samples = 1;
830 uint32_t num_samples;
831
832 /* From the Vulkan 1.1.129 spec, 26.7. Sample Shading:
833 *
834 * "If the VK_AMD_mixed_attachment_samples extension is enabled and the
835 * subpass uses color attachments, totalSamples is the number of
836 * samples of the color attachments. Otherwise, totalSamples is the
837 * value of VkPipelineMultisampleStateCreateInfo::rasterizationSamples
838 * specified at pipeline creation time."
839 */
840 if (subpass->has_color_att) {
841 num_samples = subpass->color_sample_count;
842 } else {
843 num_samples = vkms->rasterizationSamples;
844 }
845
846 if (vkms->sampleShadingEnable) {
847 ps_iter_samples = ceilf(vkms->minSampleShading * num_samples);
848 ps_iter_samples = util_next_power_of_two(ps_iter_samples);
849 }
850 return ps_iter_samples;
851 }
852
853 static bool
854 radv_is_depth_write_enabled(const VkPipelineDepthStencilStateCreateInfo *pCreateInfo)
855 {
856 return pCreateInfo->depthTestEnable &&
857 pCreateInfo->depthWriteEnable &&
858 pCreateInfo->depthCompareOp != VK_COMPARE_OP_NEVER;
859 }
860
861 static bool
862 radv_writes_stencil(const VkStencilOpState *state)
863 {
864 return state->writeMask &&
865 (state->failOp != VK_STENCIL_OP_KEEP ||
866 state->passOp != VK_STENCIL_OP_KEEP ||
867 state->depthFailOp != VK_STENCIL_OP_KEEP);
868 }
869
870 static bool
871 radv_is_stencil_write_enabled(const VkPipelineDepthStencilStateCreateInfo *pCreateInfo)
872 {
873 return pCreateInfo->stencilTestEnable &&
874 (radv_writes_stencil(&pCreateInfo->front) ||
875 radv_writes_stencil(&pCreateInfo->back));
876 }
877
878 static bool
879 radv_is_ds_write_enabled(const VkPipelineDepthStencilStateCreateInfo *pCreateInfo)
880 {
881 return radv_is_depth_write_enabled(pCreateInfo) ||
882 radv_is_stencil_write_enabled(pCreateInfo);
883 }
884
885 static bool
886 radv_order_invariant_stencil_op(VkStencilOp op)
887 {
888 /* REPLACE is normally order invariant, except when the stencil
889 * reference value is written by the fragment shader. Tracking this
890 * interaction does not seem worth the effort, so be conservative.
891 */
892 return op != VK_STENCIL_OP_INCREMENT_AND_CLAMP &&
893 op != VK_STENCIL_OP_DECREMENT_AND_CLAMP &&
894 op != VK_STENCIL_OP_REPLACE;
895 }
896
897 static bool
898 radv_order_invariant_stencil_state(const VkStencilOpState *state)
899 {
900 /* Compute whether, assuming Z writes are disabled, this stencil state
901 * is order invariant in the sense that the set of passing fragments as
902 * well as the final stencil buffer result does not depend on the order
903 * of fragments.
904 */
905 return !state->writeMask ||
906 /* The following assumes that Z writes are disabled. */
907 (state->compareOp == VK_COMPARE_OP_ALWAYS &&
908 radv_order_invariant_stencil_op(state->passOp) &&
909 radv_order_invariant_stencil_op(state->depthFailOp)) ||
910 (state->compareOp == VK_COMPARE_OP_NEVER &&
911 radv_order_invariant_stencil_op(state->failOp));
912 }
913
914 static bool
915 radv_pipeline_has_dynamic_ds_states(const VkGraphicsPipelineCreateInfo *pCreateInfo)
916 {
917 VkDynamicState ds_states[] = {
918 VK_DYNAMIC_STATE_DEPTH_TEST_ENABLE_EXT,
919 VK_DYNAMIC_STATE_DEPTH_WRITE_ENABLE_EXT,
920 VK_DYNAMIC_STATE_DEPTH_COMPARE_OP_EXT,
921 VK_DYNAMIC_STATE_STENCIL_TEST_ENABLE_EXT,
922 VK_DYNAMIC_STATE_STENCIL_OP_EXT,
923 };
924
925 if (pCreateInfo->pDynamicState) {
926 uint32_t count = pCreateInfo->pDynamicState->dynamicStateCount;
927 for (uint32_t i = 0; i < count; i++) {
928 for (uint32_t j = 0; j < ARRAY_SIZE(ds_states); j++) {
929 if (pCreateInfo->pDynamicState->pDynamicStates[i] == ds_states[j])
930 return true;
931 }
932 }
933 }
934
935 return false;
936 }
937
938 static bool
939 radv_pipeline_out_of_order_rast(struct radv_pipeline *pipeline,
940 struct radv_blend_state *blend,
941 const VkGraphicsPipelineCreateInfo *pCreateInfo)
942 {
943 RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass);
944 struct radv_subpass *subpass = pass->subpasses + pCreateInfo->subpass;
945 const VkPipelineDepthStencilStateCreateInfo *vkds = radv_pipeline_get_depth_stencil_state(pCreateInfo);
946 const VkPipelineColorBlendStateCreateInfo *vkblend = radv_pipeline_get_color_blend_state(pCreateInfo);
947 unsigned colormask = blend->cb_target_enabled_4bit;
948
949 if (!pipeline->device->physical_device->out_of_order_rast_allowed)
950 return false;
951
952 /* Be conservative if a logic operation is enabled with color buffers. */
953 if (colormask && vkblend && vkblend->logicOpEnable)
954 return false;
955
956 /* Be conservative if an extended dynamic depth/stencil state is
957 * enabled because the driver can't update out-of-order rasterization
958 * dynamically.
959 */
960 if (radv_pipeline_has_dynamic_ds_states(pCreateInfo))
961 return false;
962
963 /* Default depth/stencil invariance when no attachment is bound. */
964 struct radv_dsa_order_invariance dsa_order_invariant = {
965 .zs = true, .pass_set = true
966 };
967
968 if (vkds) {
969 struct radv_render_pass_attachment *attachment =
970 pass->attachments + subpass->depth_stencil_attachment->attachment;
971 bool has_stencil = vk_format_is_stencil(attachment->format);
972 struct radv_dsa_order_invariance order_invariance[2];
973 struct radv_shader_variant *ps =
974 pipeline->shaders[MESA_SHADER_FRAGMENT];
975
976 /* Compute depth/stencil order invariance in order to know if
977 * it's safe to enable out-of-order.
978 */
979 bool zfunc_is_ordered =
980 vkds->depthCompareOp == VK_COMPARE_OP_NEVER ||
981 vkds->depthCompareOp == VK_COMPARE_OP_LESS ||
982 vkds->depthCompareOp == VK_COMPARE_OP_LESS_OR_EQUAL ||
983 vkds->depthCompareOp == VK_COMPARE_OP_GREATER ||
984 vkds->depthCompareOp == VK_COMPARE_OP_GREATER_OR_EQUAL;
985
986 bool nozwrite_and_order_invariant_stencil =
987 !radv_is_ds_write_enabled(vkds) ||
988 (!radv_is_depth_write_enabled(vkds) &&
989 radv_order_invariant_stencil_state(&vkds->front) &&
990 radv_order_invariant_stencil_state(&vkds->back));
991
992 order_invariance[1].zs =
993 nozwrite_and_order_invariant_stencil ||
994 (!radv_is_stencil_write_enabled(vkds) &&
995 zfunc_is_ordered);
996 order_invariance[0].zs =
997 !radv_is_depth_write_enabled(vkds) || zfunc_is_ordered;
998
999 order_invariance[1].pass_set =
1000 nozwrite_and_order_invariant_stencil ||
1001 (!radv_is_stencil_write_enabled(vkds) &&
1002 (vkds->depthCompareOp == VK_COMPARE_OP_ALWAYS ||
1003 vkds->depthCompareOp == VK_COMPARE_OP_NEVER));
1004 order_invariance[0].pass_set =
1005 !radv_is_depth_write_enabled(vkds) ||
1006 (vkds->depthCompareOp == VK_COMPARE_OP_ALWAYS ||
1007 vkds->depthCompareOp == VK_COMPARE_OP_NEVER);
1008
1009 dsa_order_invariant = order_invariance[has_stencil];
1010 if (!dsa_order_invariant.zs)
1011 return false;
1012
1013 /* The set of PS invocations is always order invariant,
1014 * except when early Z/S tests are requested.
1015 */
1016 if (ps &&
1017 ps->info.ps.writes_memory &&
1018 ps->info.ps.early_fragment_test &&
1019 !dsa_order_invariant.pass_set)
1020 return false;
1021
1022 /* Determine if out-of-order rasterization should be disabled
1023 * when occlusion queries are used.
1024 */
1025 pipeline->graphics.disable_out_of_order_rast_for_occlusion =
1026 !dsa_order_invariant.pass_set;
1027 }
1028
1029 /* No color buffers are enabled for writing. */
1030 if (!colormask)
1031 return true;
1032
1033 unsigned blendmask = colormask & blend->blend_enable_4bit;
1034
1035 if (blendmask) {
1036 /* Only commutative blending. */
1037 if (blendmask & ~blend->commutative_4bit)
1038 return false;
1039
1040 if (!dsa_order_invariant.pass_set)
1041 return false;
1042 }
1043
1044 if (colormask & ~blendmask)
1045 return false;
1046
1047 return true;
1048 }
1049
1050 static const VkConservativeRasterizationModeEXT
1051 radv_get_conservative_raster_mode(const VkPipelineRasterizationStateCreateInfo *pCreateInfo)
1052 {
1053 const VkPipelineRasterizationConservativeStateCreateInfoEXT *conservative_raster =
1054 vk_find_struct_const(pCreateInfo->pNext, PIPELINE_RASTERIZATION_CONSERVATIVE_STATE_CREATE_INFO_EXT);
1055
1056 if (!conservative_raster)
1057 return VK_CONSERVATIVE_RASTERIZATION_MODE_DISABLED_EXT;
1058 return conservative_raster->conservativeRasterizationMode;
1059 }
1060
1061 static void
1062 radv_pipeline_init_multisample_state(struct radv_pipeline *pipeline,
1063 struct radv_blend_state *blend,
1064 const VkGraphicsPipelineCreateInfo *pCreateInfo)
1065 {
1066 const VkPipelineMultisampleStateCreateInfo *vkms = radv_pipeline_get_multisample_state(pCreateInfo);
1067 struct radv_multisample_state *ms = &pipeline->graphics.ms;
1068 unsigned num_tile_pipes = pipeline->device->physical_device->rad_info.num_tile_pipes;
1069 const VkConservativeRasterizationModeEXT mode =
1070 radv_get_conservative_raster_mode(pCreateInfo->pRasterizationState);
1071 bool out_of_order_rast = false;
1072 int ps_iter_samples = 1;
1073 uint32_t mask = 0xffff;
1074
1075 if (vkms) {
1076 ms->num_samples = vkms->rasterizationSamples;
1077
1078 /* From the Vulkan 1.1.129 spec, 26.7. Sample Shading:
1079 *
1080 * "Sample shading is enabled for a graphics pipeline:
1081 *
1082 * - If the interface of the fragment shader entry point of the
1083 * graphics pipeline includes an input variable decorated
1084 * with SampleId or SamplePosition. In this case
1085 * minSampleShadingFactor takes the value 1.0.
1086 * - Else if the sampleShadingEnable member of the
1087 * VkPipelineMultisampleStateCreateInfo structure specified
1088 * when creating the graphics pipeline is set to VK_TRUE. In
1089 * this case minSampleShadingFactor takes the value of
1090 * VkPipelineMultisampleStateCreateInfo::minSampleShading.
1091 *
1092 * Otherwise, sample shading is considered disabled."
1093 */
1094 if (pipeline->shaders[MESA_SHADER_FRAGMENT]->info.ps.force_persample) {
1095 ps_iter_samples = ms->num_samples;
1096 } else {
1097 ps_iter_samples = radv_pipeline_get_ps_iter_samples(pCreateInfo);
1098 }
1099 } else {
1100 ms->num_samples = 1;
1101 }
1102
1103 const struct VkPipelineRasterizationStateRasterizationOrderAMD *raster_order =
1104 vk_find_struct_const(pCreateInfo->pRasterizationState->pNext, PIPELINE_RASTERIZATION_STATE_RASTERIZATION_ORDER_AMD);
1105 if (raster_order && raster_order->rasterizationOrder == VK_RASTERIZATION_ORDER_RELAXED_AMD) {
1106 /* Out-of-order rasterization is explicitly enabled by the
1107 * application.
1108 */
1109 out_of_order_rast = true;
1110 } else {
1111 /* Determine if the driver can enable out-of-order
1112 * rasterization internally.
1113 */
1114 out_of_order_rast =
1115 radv_pipeline_out_of_order_rast(pipeline, blend, pCreateInfo);
1116 }
1117
1118 ms->pa_sc_aa_config = 0;
1119 ms->db_eqaa = S_028804_HIGH_QUALITY_INTERSECTIONS(1) |
1120 S_028804_INCOHERENT_EQAA_READS(1) |
1121 S_028804_INTERPOLATE_COMP_Z(1) |
1122 S_028804_STATIC_ANCHOR_ASSOCIATIONS(1);
1123
1124 /* Adjust MSAA state if conservative rasterization is enabled. */
1125 if (mode != VK_CONSERVATIVE_RASTERIZATION_MODE_DISABLED_EXT) {
1126 ms->pa_sc_aa_config |= S_028BE0_AA_MASK_CENTROID_DTMN(1);
1127
1128 ms->db_eqaa |= S_028804_ENABLE_POSTZ_OVERRASTERIZATION(1) |
1129 S_028804_OVERRASTERIZATION_AMOUNT(4);
1130 }
1131
1132 ms->pa_sc_mode_cntl_1 =
1133 S_028A4C_WALK_FENCE_ENABLE(1) | //TODO linear dst fixes
1134 S_028A4C_WALK_FENCE_SIZE(num_tile_pipes == 2 ? 2 : 3) |
1135 S_028A4C_OUT_OF_ORDER_PRIMITIVE_ENABLE(out_of_order_rast) |
1136 S_028A4C_OUT_OF_ORDER_WATER_MARK(0x7) |
1137 /* always 1: */
1138 S_028A4C_WALK_ALIGN8_PRIM_FITS_ST(1) |
1139 S_028A4C_SUPERTILE_WALK_ORDER_ENABLE(1) |
1140 S_028A4C_TILE_WALK_ORDER_ENABLE(1) |
1141 S_028A4C_MULTI_SHADER_ENGINE_PRIM_DISCARD_ENABLE(1) |
1142 S_028A4C_FORCE_EOV_CNTDWN_ENABLE(1) |
1143 S_028A4C_FORCE_EOV_REZ_ENABLE(1);
1144 ms->pa_sc_mode_cntl_0 = S_028A48_ALTERNATE_RBS_PER_TILE(pipeline->device->physical_device->rad_info.chip_class >= GFX9) |
1145 S_028A48_VPORT_SCISSOR_ENABLE(1);
1146
1147 const VkPipelineRasterizationLineStateCreateInfoEXT *rast_line =
1148 vk_find_struct_const(pCreateInfo->pRasterizationState->pNext,
1149 PIPELINE_RASTERIZATION_LINE_STATE_CREATE_INFO_EXT);
1150 if (rast_line) {
1151 ms->pa_sc_mode_cntl_0 |= S_028A48_LINE_STIPPLE_ENABLE(rast_line->stippledLineEnable);
1152 if (rast_line->lineRasterizationMode == VK_LINE_RASTERIZATION_MODE_BRESENHAM_EXT) {
1153 /* From the Vulkan spec 1.1.129:
1154 *
1155 * "When VK_LINE_RASTERIZATION_MODE_BRESENHAM_EXT lines
1156 * are being rasterized, sample locations may all be
1157 * treated as being at the pixel center (this may
1158 * affect attribute and depth interpolation)."
1159 */
1160 ms->num_samples = 1;
1161 }
1162 }
1163
1164 if (ms->num_samples > 1) {
1165 RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass);
1166 struct radv_subpass *subpass = &pass->subpasses[pCreateInfo->subpass];
1167 uint32_t z_samples = subpass->depth_stencil_attachment ? subpass->depth_sample_count : ms->num_samples;
1168 unsigned log_samples = util_logbase2(ms->num_samples);
1169 unsigned log_z_samples = util_logbase2(z_samples);
1170 unsigned log_ps_iter_samples = util_logbase2(ps_iter_samples);
1171 ms->pa_sc_mode_cntl_0 |= S_028A48_MSAA_ENABLE(1);
1172 ms->db_eqaa |= S_028804_MAX_ANCHOR_SAMPLES(log_z_samples) |
1173 S_028804_PS_ITER_SAMPLES(log_ps_iter_samples) |
1174 S_028804_MASK_EXPORT_NUM_SAMPLES(log_samples) |
1175 S_028804_ALPHA_TO_MASK_NUM_SAMPLES(log_samples);
1176 ms->pa_sc_aa_config |= S_028BE0_MSAA_NUM_SAMPLES(log_samples) |
1177 S_028BE0_MAX_SAMPLE_DIST(radv_get_default_max_sample_dist(log_samples)) |
1178 S_028BE0_MSAA_EXPOSED_SAMPLES(log_samples) | /* CM_R_028BE0_PA_SC_AA_CONFIG */
1179 S_028BE0_COVERED_CENTROID_IS_CENTER_GFX103(pipeline->device->physical_device->rad_info.chip_class >= GFX10_3);
1180 ms->pa_sc_mode_cntl_1 |= S_028A4C_PS_ITER_SAMPLE(ps_iter_samples > 1);
1181 if (ps_iter_samples > 1)
1182 pipeline->graphics.spi_baryc_cntl |= S_0286E0_POS_FLOAT_LOCATION(2);
1183 }
1184
1185 if (vkms && vkms->pSampleMask) {
1186 mask = vkms->pSampleMask[0] & 0xffff;
1187 }
1188
1189 ms->pa_sc_aa_mask[0] = mask | (mask << 16);
1190 ms->pa_sc_aa_mask[1] = mask | (mask << 16);
1191 }
1192
1193 static bool
1194 radv_prim_can_use_guardband(enum VkPrimitiveTopology topology)
1195 {
1196 switch (topology) {
1197 case VK_PRIMITIVE_TOPOLOGY_POINT_LIST:
1198 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST:
1199 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP:
1200 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY:
1201 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY:
1202 return false;
1203 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST:
1204 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP:
1205 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN:
1206 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY:
1207 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY:
1208 case VK_PRIMITIVE_TOPOLOGY_PATCH_LIST:
1209 return true;
1210 default:
1211 unreachable("unhandled primitive type");
1212 }
1213 }
1214
1215 static uint32_t
1216 si_conv_gl_prim_to_gs_out(unsigned gl_prim)
1217 {
1218 switch (gl_prim) {
1219 case 0: /* GL_POINTS */
1220 return V_028A6C_OUTPRIM_TYPE_POINTLIST;
1221 case 1: /* GL_LINES */
1222 case 3: /* GL_LINE_STRIP */
1223 case 0xA: /* GL_LINE_STRIP_ADJACENCY_ARB */
1224 case 0x8E7A: /* GL_ISOLINES */
1225 return V_028A6C_OUTPRIM_TYPE_LINESTRIP;
1226
1227 case 4: /* GL_TRIANGLES */
1228 case 0xc: /* GL_TRIANGLES_ADJACENCY_ARB */
1229 case 5: /* GL_TRIANGLE_STRIP */
1230 case 7: /* GL_QUADS */
1231 return V_028A6C_OUTPRIM_TYPE_TRISTRIP;
1232 default:
1233 assert(0);
1234 return 0;
1235 }
1236 }
1237
1238 static uint32_t
1239 si_conv_prim_to_gs_out(enum VkPrimitiveTopology topology)
1240 {
1241 switch (topology) {
1242 case VK_PRIMITIVE_TOPOLOGY_POINT_LIST:
1243 case VK_PRIMITIVE_TOPOLOGY_PATCH_LIST:
1244 return V_028A6C_OUTPRIM_TYPE_POINTLIST;
1245 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST:
1246 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP:
1247 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY:
1248 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY:
1249 return V_028A6C_OUTPRIM_TYPE_LINESTRIP;
1250 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST:
1251 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP:
1252 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN:
1253 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY:
1254 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY:
1255 return V_028A6C_OUTPRIM_TYPE_TRISTRIP;
1256 default:
1257 assert(0);
1258 return 0;
1259 }
1260 }
1261
1262 static unsigned radv_dynamic_state_mask(VkDynamicState state)
1263 {
1264 switch(state) {
1265 case VK_DYNAMIC_STATE_VIEWPORT:
1266 case VK_DYNAMIC_STATE_VIEWPORT_WITH_COUNT_EXT:
1267 return RADV_DYNAMIC_VIEWPORT;
1268 case VK_DYNAMIC_STATE_SCISSOR:
1269 case VK_DYNAMIC_STATE_SCISSOR_WITH_COUNT_EXT:
1270 return RADV_DYNAMIC_SCISSOR;
1271 case VK_DYNAMIC_STATE_LINE_WIDTH:
1272 return RADV_DYNAMIC_LINE_WIDTH;
1273 case VK_DYNAMIC_STATE_DEPTH_BIAS:
1274 return RADV_DYNAMIC_DEPTH_BIAS;
1275 case VK_DYNAMIC_STATE_BLEND_CONSTANTS:
1276 return RADV_DYNAMIC_BLEND_CONSTANTS;
1277 case VK_DYNAMIC_STATE_DEPTH_BOUNDS:
1278 return RADV_DYNAMIC_DEPTH_BOUNDS;
1279 case VK_DYNAMIC_STATE_STENCIL_COMPARE_MASK:
1280 return RADV_DYNAMIC_STENCIL_COMPARE_MASK;
1281 case VK_DYNAMIC_STATE_STENCIL_WRITE_MASK:
1282 return RADV_DYNAMIC_STENCIL_WRITE_MASK;
1283 case VK_DYNAMIC_STATE_STENCIL_REFERENCE:
1284 return RADV_DYNAMIC_STENCIL_REFERENCE;
1285 case VK_DYNAMIC_STATE_DISCARD_RECTANGLE_EXT:
1286 return RADV_DYNAMIC_DISCARD_RECTANGLE;
1287 case VK_DYNAMIC_STATE_SAMPLE_LOCATIONS_EXT:
1288 return RADV_DYNAMIC_SAMPLE_LOCATIONS;
1289 case VK_DYNAMIC_STATE_LINE_STIPPLE_EXT:
1290 return RADV_DYNAMIC_LINE_STIPPLE;
1291 case VK_DYNAMIC_STATE_CULL_MODE_EXT:
1292 return RADV_DYNAMIC_CULL_MODE;
1293 case VK_DYNAMIC_STATE_FRONT_FACE_EXT:
1294 return RADV_DYNAMIC_FRONT_FACE;
1295 case VK_DYNAMIC_STATE_PRIMITIVE_TOPOLOGY_EXT:
1296 return RADV_DYNAMIC_PRIMITIVE_TOPOLOGY;
1297 case VK_DYNAMIC_STATE_DEPTH_TEST_ENABLE_EXT:
1298 return RADV_DYNAMIC_DEPTH_TEST_ENABLE;
1299 case VK_DYNAMIC_STATE_DEPTH_WRITE_ENABLE_EXT:
1300 return RADV_DYNAMIC_DEPTH_WRITE_ENABLE;
1301 case VK_DYNAMIC_STATE_DEPTH_COMPARE_OP_EXT:
1302 return RADV_DYNAMIC_DEPTH_COMPARE_OP;
1303 case VK_DYNAMIC_STATE_DEPTH_BOUNDS_TEST_ENABLE_EXT:
1304 return RADV_DYNAMIC_DEPTH_BOUNDS_TEST_ENABLE;
1305 case VK_DYNAMIC_STATE_STENCIL_TEST_ENABLE_EXT:
1306 return RADV_DYNAMIC_STENCIL_TEST_ENABLE;
1307 case VK_DYNAMIC_STATE_STENCIL_OP_EXT:
1308 return RADV_DYNAMIC_STENCIL_OP;
1309 case VK_DYNAMIC_STATE_VERTEX_INPUT_BINDING_STRIDE_EXT:
1310 return RADV_DYNAMIC_VERTEX_INPUT_BINDING_STRIDE;
1311 default:
1312 unreachable("Unhandled dynamic state");
1313 }
1314 }
1315
1316 static uint32_t radv_pipeline_needed_dynamic_state(const VkGraphicsPipelineCreateInfo *pCreateInfo)
1317 {
1318 uint32_t states = RADV_DYNAMIC_ALL;
1319
1320 /* If rasterization is disabled we do not care about any of the
1321 * dynamic states, since they are all rasterization related only,
1322 * except primitive topology and vertex binding stride.
1323 */
1324 if (pCreateInfo->pRasterizationState->rasterizerDiscardEnable)
1325 return RADV_DYNAMIC_PRIMITIVE_TOPOLOGY |
1326 RADV_DYNAMIC_VERTEX_INPUT_BINDING_STRIDE;
1327
1328 if (!pCreateInfo->pRasterizationState->depthBiasEnable)
1329 states &= ~RADV_DYNAMIC_DEPTH_BIAS;
1330
1331 if (!pCreateInfo->pDepthStencilState ||
1332 !pCreateInfo->pDepthStencilState->depthBoundsTestEnable)
1333 states &= ~RADV_DYNAMIC_DEPTH_BOUNDS;
1334
1335 if (!pCreateInfo->pDepthStencilState ||
1336 !pCreateInfo->pDepthStencilState->stencilTestEnable)
1337 states &= ~(RADV_DYNAMIC_STENCIL_COMPARE_MASK |
1338 RADV_DYNAMIC_STENCIL_WRITE_MASK |
1339 RADV_DYNAMIC_STENCIL_REFERENCE);
1340
1341 if (!vk_find_struct_const(pCreateInfo->pNext, PIPELINE_DISCARD_RECTANGLE_STATE_CREATE_INFO_EXT))
1342 states &= ~RADV_DYNAMIC_DISCARD_RECTANGLE;
1343
1344 if (!pCreateInfo->pMultisampleState ||
1345 !vk_find_struct_const(pCreateInfo->pMultisampleState->pNext,
1346 PIPELINE_SAMPLE_LOCATIONS_STATE_CREATE_INFO_EXT))
1347 states &= ~RADV_DYNAMIC_SAMPLE_LOCATIONS;
1348
1349 if (!pCreateInfo->pRasterizationState ||
1350 !vk_find_struct_const(pCreateInfo->pRasterizationState->pNext,
1351 PIPELINE_RASTERIZATION_LINE_STATE_CREATE_INFO_EXT))
1352 states &= ~RADV_DYNAMIC_LINE_STIPPLE;
1353
1354 /* TODO: blend constants & line width. */
1355
1356 return states;
1357 }
1358
1359 static struct radv_ia_multi_vgt_param_helpers
1360 radv_compute_ia_multi_vgt_param_helpers(struct radv_pipeline *pipeline)
1361 {
1362 struct radv_ia_multi_vgt_param_helpers ia_multi_vgt_param = {0};
1363 const struct radv_device *device = pipeline->device;
1364
1365 if (radv_pipeline_has_tess(pipeline))
1366 ia_multi_vgt_param.primgroup_size = pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.tcs.num_patches;
1367 else if (radv_pipeline_has_gs(pipeline))
1368 ia_multi_vgt_param.primgroup_size = 64;
1369 else
1370 ia_multi_vgt_param.primgroup_size = 128; /* recommended without a GS */
1371
1372 /* GS requirement. */
1373 ia_multi_vgt_param.partial_es_wave = false;
1374 if (radv_pipeline_has_gs(pipeline) && device->physical_device->rad_info.chip_class <= GFX8)
1375 if (SI_GS_PER_ES / ia_multi_vgt_param.primgroup_size >= pipeline->device->gs_table_depth - 3)
1376 ia_multi_vgt_param.partial_es_wave = true;
1377
1378 ia_multi_vgt_param.ia_switch_on_eoi = false;
1379 if (pipeline->shaders[MESA_SHADER_FRAGMENT]->info.ps.prim_id_input)
1380 ia_multi_vgt_param.ia_switch_on_eoi = true;
1381 if (radv_pipeline_has_gs(pipeline) &&
1382 pipeline->shaders[MESA_SHADER_GEOMETRY]->info.uses_prim_id)
1383 ia_multi_vgt_param.ia_switch_on_eoi = true;
1384 if (radv_pipeline_has_tess(pipeline)) {
1385 /* SWITCH_ON_EOI must be set if PrimID is used. */
1386 if (pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.uses_prim_id ||
1387 radv_get_shader(pipeline, MESA_SHADER_TESS_EVAL)->info.uses_prim_id)
1388 ia_multi_vgt_param.ia_switch_on_eoi = true;
1389 }
1390
1391 ia_multi_vgt_param.partial_vs_wave = false;
1392 if (radv_pipeline_has_tess(pipeline)) {
1393 /* Bug with tessellation and GS on Bonaire and older 2 SE chips. */
1394 if ((device->physical_device->rad_info.family == CHIP_TAHITI ||
1395 device->physical_device->rad_info.family == CHIP_PITCAIRN ||
1396 device->physical_device->rad_info.family == CHIP_BONAIRE) &&
1397 radv_pipeline_has_gs(pipeline))
1398 ia_multi_vgt_param.partial_vs_wave = true;
1399 /* Needed for 028B6C_DISTRIBUTION_MODE != 0 */
1400 if (device->physical_device->rad_info.has_distributed_tess) {
1401 if (radv_pipeline_has_gs(pipeline)) {
1402 if (device->physical_device->rad_info.chip_class <= GFX8)
1403 ia_multi_vgt_param.partial_es_wave = true;
1404 } else {
1405 ia_multi_vgt_param.partial_vs_wave = true;
1406 }
1407 }
1408 }
1409
1410 if (radv_pipeline_has_gs(pipeline)) {
1411 /* On these chips there is the possibility of a hang if the
1412 * pipeline uses a GS and partial_vs_wave is not set.
1413 *
1414 * This mostly does not hit 4-SE chips, as those typically set
1415 * ia_switch_on_eoi and then partial_vs_wave is set for pipelines
1416 * with GS due to another workaround.
1417 *
1418 * Reproducer: https://bugs.freedesktop.org/show_bug.cgi?id=109242
1419 */
1420 if (device->physical_device->rad_info.family == CHIP_TONGA ||
1421 device->physical_device->rad_info.family == CHIP_FIJI ||
1422 device->physical_device->rad_info.family == CHIP_POLARIS10 ||
1423 device->physical_device->rad_info.family == CHIP_POLARIS11 ||
1424 device->physical_device->rad_info.family == CHIP_POLARIS12 ||
1425 device->physical_device->rad_info.family == CHIP_VEGAM) {
1426 ia_multi_vgt_param.partial_vs_wave = true;
1427 }
1428 }
1429
1430 ia_multi_vgt_param.base =
1431 S_028AA8_PRIMGROUP_SIZE(ia_multi_vgt_param.primgroup_size - 1) |
1432 /* The following field was moved to VGT_SHADER_STAGES_EN in GFX9. */
1433 S_028AA8_MAX_PRIMGRP_IN_WAVE(device->physical_device->rad_info.chip_class == GFX8 ? 2 : 0) |
1434 S_030960_EN_INST_OPT_BASIC(device->physical_device->rad_info.chip_class >= GFX9) |
1435 S_030960_EN_INST_OPT_ADV(device->physical_device->rad_info.chip_class >= GFX9);
1436
1437 return ia_multi_vgt_param;
1438 }
1439
1440 static void
1441 radv_pipeline_init_input_assembly_state(struct radv_pipeline *pipeline,
1442 const VkGraphicsPipelineCreateInfo *pCreateInfo,
1443 const struct radv_graphics_pipeline_create_info *extra)
1444 {
1445 const VkPipelineInputAssemblyStateCreateInfo *ia_state = pCreateInfo->pInputAssemblyState;
1446 struct radv_shader_variant *tes = pipeline->shaders[MESA_SHADER_TESS_EVAL];
1447 struct radv_shader_variant *gs = pipeline->shaders[MESA_SHADER_GEOMETRY];
1448
1449 pipeline->graphics.prim_restart_enable = !!ia_state->primitiveRestartEnable;
1450 pipeline->graphics.can_use_guardband = radv_prim_can_use_guardband(ia_state->topology);
1451
1452 if (radv_pipeline_has_gs(pipeline)) {
1453 if (si_conv_gl_prim_to_gs_out(gs->info.gs.output_prim) == V_028A6C_OUTPRIM_TYPE_TRISTRIP)
1454 pipeline->graphics.can_use_guardband = true;
1455 } else if (radv_pipeline_has_tess(pipeline)) {
1456 if (!tes->info.tes.point_mode &&
1457 si_conv_gl_prim_to_gs_out(tes->info.tes.primitive_mode) == V_028A6C_OUTPRIM_TYPE_TRISTRIP)
1458 pipeline->graphics.can_use_guardband = true;
1459 }
1460
1461 if (extra && extra->use_rectlist) {
1462 pipeline->graphics.can_use_guardband = true;
1463 }
1464
1465 pipeline->graphics.ia_multi_vgt_param =
1466 radv_compute_ia_multi_vgt_param_helpers(pipeline);
1467 }
1468
1469 static void
1470 radv_pipeline_init_dynamic_state(struct radv_pipeline *pipeline,
1471 const VkGraphicsPipelineCreateInfo *pCreateInfo,
1472 const struct radv_graphics_pipeline_create_info *extra)
1473 {
1474 uint32_t needed_states = radv_pipeline_needed_dynamic_state(pCreateInfo);
1475 uint32_t states = needed_states;
1476 RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass);
1477 struct radv_subpass *subpass = &pass->subpasses[pCreateInfo->subpass];
1478
1479 pipeline->dynamic_state = default_dynamic_state;
1480 pipeline->graphics.needed_dynamic_state = needed_states;
1481
1482 if (pCreateInfo->pDynamicState) {
1483 /* Remove all of the states that are marked as dynamic */
1484 uint32_t count = pCreateInfo->pDynamicState->dynamicStateCount;
1485 for (uint32_t s = 0; s < count; s++)
1486 states &= ~radv_dynamic_state_mask(pCreateInfo->pDynamicState->pDynamicStates[s]);
1487 }
1488
1489 struct radv_dynamic_state *dynamic = &pipeline->dynamic_state;
1490
1491 if (needed_states & RADV_DYNAMIC_VIEWPORT) {
1492 assert(pCreateInfo->pViewportState);
1493
1494 dynamic->viewport.count = pCreateInfo->pViewportState->viewportCount;
1495 if (states & RADV_DYNAMIC_VIEWPORT) {
1496 typed_memcpy(dynamic->viewport.viewports,
1497 pCreateInfo->pViewportState->pViewports,
1498 pCreateInfo->pViewportState->viewportCount);
1499 }
1500 }
1501
1502 if (needed_states & RADV_DYNAMIC_SCISSOR) {
1503 dynamic->scissor.count = pCreateInfo->pViewportState->scissorCount;
1504 if (states & RADV_DYNAMIC_SCISSOR) {
1505 typed_memcpy(dynamic->scissor.scissors,
1506 pCreateInfo->pViewportState->pScissors,
1507 pCreateInfo->pViewportState->scissorCount);
1508 }
1509 }
1510
1511 if (states & RADV_DYNAMIC_LINE_WIDTH) {
1512 assert(pCreateInfo->pRasterizationState);
1513 dynamic->line_width = pCreateInfo->pRasterizationState->lineWidth;
1514 }
1515
1516 if (states & RADV_DYNAMIC_DEPTH_BIAS) {
1517 assert(pCreateInfo->pRasterizationState);
1518 dynamic->depth_bias.bias =
1519 pCreateInfo->pRasterizationState->depthBiasConstantFactor;
1520 dynamic->depth_bias.clamp =
1521 pCreateInfo->pRasterizationState->depthBiasClamp;
1522 dynamic->depth_bias.slope =
1523 pCreateInfo->pRasterizationState->depthBiasSlopeFactor;
1524 }
1525
1526 /* Section 9.2 of the Vulkan 1.0.15 spec says:
1527 *
1528 * pColorBlendState is [...] NULL if the pipeline has rasterization
1529 * disabled or if the subpass of the render pass the pipeline is
1530 * created against does not use any color attachments.
1531 */
1532 if (subpass->has_color_att && states & RADV_DYNAMIC_BLEND_CONSTANTS) {
1533 assert(pCreateInfo->pColorBlendState);
1534 typed_memcpy(dynamic->blend_constants,
1535 pCreateInfo->pColorBlendState->blendConstants, 4);
1536 }
1537
1538 if (states & RADV_DYNAMIC_CULL_MODE) {
1539 dynamic->cull_mode =
1540 pCreateInfo->pRasterizationState->cullMode;
1541 }
1542
1543 if (states & RADV_DYNAMIC_FRONT_FACE) {
1544 dynamic->front_face =
1545 pCreateInfo->pRasterizationState->frontFace;
1546 }
1547
1548 if (states & RADV_DYNAMIC_PRIMITIVE_TOPOLOGY) {
1549 dynamic->primitive_topology =
1550 si_translate_prim(pCreateInfo->pInputAssemblyState->topology);
1551 if (extra && extra->use_rectlist) {
1552 dynamic->primitive_topology = V_008958_DI_PT_RECTLIST;
1553 }
1554 }
1555
1556 /* If there is no depthstencil attachment, then don't read
1557 * pDepthStencilState. The Vulkan spec states that pDepthStencilState may
1558 * be NULL in this case. Even if pDepthStencilState is non-NULL, there is
1559 * no need to override the depthstencil defaults in
1560 * radv_pipeline::dynamic_state when there is no depthstencil attachment.
1561 *
1562 * Section 9.2 of the Vulkan 1.0.15 spec says:
1563 *
1564 * pDepthStencilState is [...] NULL if the pipeline has rasterization
1565 * disabled or if the subpass of the render pass the pipeline is created
1566 * against does not use a depth/stencil attachment.
1567 */
1568 if (needed_states && subpass->depth_stencil_attachment) {
1569 assert(pCreateInfo->pDepthStencilState);
1570
1571 if (states & RADV_DYNAMIC_DEPTH_BOUNDS) {
1572 dynamic->depth_bounds.min =
1573 pCreateInfo->pDepthStencilState->minDepthBounds;
1574 dynamic->depth_bounds.max =
1575 pCreateInfo->pDepthStencilState->maxDepthBounds;
1576 }
1577
1578 if (states & RADV_DYNAMIC_STENCIL_COMPARE_MASK) {
1579 dynamic->stencil_compare_mask.front =
1580 pCreateInfo->pDepthStencilState->front.compareMask;
1581 dynamic->stencil_compare_mask.back =
1582 pCreateInfo->pDepthStencilState->back.compareMask;
1583 }
1584
1585 if (states & RADV_DYNAMIC_STENCIL_WRITE_MASK) {
1586 dynamic->stencil_write_mask.front =
1587 pCreateInfo->pDepthStencilState->front.writeMask;
1588 dynamic->stencil_write_mask.back =
1589 pCreateInfo->pDepthStencilState->back.writeMask;
1590 }
1591
1592 if (states & RADV_DYNAMIC_STENCIL_REFERENCE) {
1593 dynamic->stencil_reference.front =
1594 pCreateInfo->pDepthStencilState->front.reference;
1595 dynamic->stencil_reference.back =
1596 pCreateInfo->pDepthStencilState->back.reference;
1597 }
1598
1599 if (states & RADV_DYNAMIC_DEPTH_TEST_ENABLE) {
1600 dynamic->depth_test_enable =
1601 pCreateInfo->pDepthStencilState->depthTestEnable;
1602 }
1603
1604 if (states & RADV_DYNAMIC_DEPTH_WRITE_ENABLE) {
1605 dynamic->depth_write_enable =
1606 pCreateInfo->pDepthStencilState->depthWriteEnable;
1607 }
1608
1609 if (states & RADV_DYNAMIC_DEPTH_COMPARE_OP) {
1610 dynamic->depth_compare_op =
1611 pCreateInfo->pDepthStencilState->depthCompareOp;
1612 }
1613
1614 if (states & RADV_DYNAMIC_DEPTH_BOUNDS_TEST_ENABLE) {
1615 dynamic->depth_bounds_test_enable =
1616 pCreateInfo->pDepthStencilState->depthBoundsTestEnable;
1617 }
1618
1619 if (states & RADV_DYNAMIC_STENCIL_TEST_ENABLE) {
1620 dynamic->stencil_test_enable =
1621 pCreateInfo->pDepthStencilState->stencilTestEnable;
1622 }
1623
1624 if (states & RADV_DYNAMIC_STENCIL_OP) {
1625 dynamic->stencil_op.front.compare_op =
1626 pCreateInfo->pDepthStencilState->front.compareOp;
1627 dynamic->stencil_op.front.fail_op =
1628 pCreateInfo->pDepthStencilState->front.failOp;
1629 dynamic->stencil_op.front.pass_op =
1630 pCreateInfo->pDepthStencilState->front.passOp;
1631 dynamic->stencil_op.front.depth_fail_op =
1632 pCreateInfo->pDepthStencilState->front.depthFailOp;
1633
1634 dynamic->stencil_op.back.compare_op =
1635 pCreateInfo->pDepthStencilState->back.compareOp;
1636 dynamic->stencil_op.back.fail_op =
1637 pCreateInfo->pDepthStencilState->back.failOp;
1638 dynamic->stencil_op.back.pass_op =
1639 pCreateInfo->pDepthStencilState->back.passOp;
1640 dynamic->stencil_op.back.depth_fail_op =
1641 pCreateInfo->pDepthStencilState->back.depthFailOp;
1642 }
1643 }
1644
1645 const VkPipelineDiscardRectangleStateCreateInfoEXT *discard_rectangle_info =
1646 vk_find_struct_const(pCreateInfo->pNext, PIPELINE_DISCARD_RECTANGLE_STATE_CREATE_INFO_EXT);
1647 if (needed_states & RADV_DYNAMIC_DISCARD_RECTANGLE) {
1648 dynamic->discard_rectangle.count = discard_rectangle_info->discardRectangleCount;
1649 if (states & RADV_DYNAMIC_DISCARD_RECTANGLE) {
1650 typed_memcpy(dynamic->discard_rectangle.rectangles,
1651 discard_rectangle_info->pDiscardRectangles,
1652 discard_rectangle_info->discardRectangleCount);
1653 }
1654 }
1655
1656 if (needed_states & RADV_DYNAMIC_SAMPLE_LOCATIONS) {
1657 const VkPipelineSampleLocationsStateCreateInfoEXT *sample_location_info =
1658 vk_find_struct_const(pCreateInfo->pMultisampleState->pNext,
1659 PIPELINE_SAMPLE_LOCATIONS_STATE_CREATE_INFO_EXT);
1660 /* If sampleLocationsEnable is VK_FALSE, the default sample
1661 * locations are used and the values specified in
1662 * sampleLocationsInfo are ignored.
1663 */
1664 if (sample_location_info->sampleLocationsEnable) {
1665 const VkSampleLocationsInfoEXT *pSampleLocationsInfo =
1666 &sample_location_info->sampleLocationsInfo;
1667
1668 assert(pSampleLocationsInfo->sampleLocationsCount <= MAX_SAMPLE_LOCATIONS);
1669
1670 dynamic->sample_location.per_pixel = pSampleLocationsInfo->sampleLocationsPerPixel;
1671 dynamic->sample_location.grid_size = pSampleLocationsInfo->sampleLocationGridSize;
1672 dynamic->sample_location.count = pSampleLocationsInfo->sampleLocationsCount;
1673 typed_memcpy(&dynamic->sample_location.locations[0],
1674 pSampleLocationsInfo->pSampleLocations,
1675 pSampleLocationsInfo->sampleLocationsCount);
1676 }
1677 }
1678
1679 const VkPipelineRasterizationLineStateCreateInfoEXT *rast_line_info =
1680 vk_find_struct_const(pCreateInfo->pRasterizationState->pNext,
1681 PIPELINE_RASTERIZATION_LINE_STATE_CREATE_INFO_EXT);
1682 if (needed_states & RADV_DYNAMIC_LINE_STIPPLE) {
1683 dynamic->line_stipple.factor = rast_line_info->lineStippleFactor;
1684 dynamic->line_stipple.pattern = rast_line_info->lineStipplePattern;
1685 }
1686
1687 if (!(states & RADV_DYNAMIC_VERTEX_INPUT_BINDING_STRIDE))
1688 pipeline->graphics.uses_dynamic_stride = true;
1689
1690 pipeline->dynamic_state.mask = states;
1691 }
1692
1693 static void
1694 radv_pipeline_init_raster_state(struct radv_pipeline *pipeline,
1695 const VkGraphicsPipelineCreateInfo *pCreateInfo)
1696 {
1697 const VkPipelineRasterizationStateCreateInfo *raster_info =
1698 pCreateInfo->pRasterizationState;
1699
1700 pipeline->graphics.pa_su_sc_mode_cntl =
1701 S_028814_FACE(raster_info->frontFace) |
1702 S_028814_CULL_FRONT(!!(raster_info->cullMode & VK_CULL_MODE_FRONT_BIT)) |
1703 S_028814_CULL_BACK(!!(raster_info->cullMode & VK_CULL_MODE_BACK_BIT)) |
1704 S_028814_POLY_MODE(raster_info->polygonMode != VK_POLYGON_MODE_FILL) |
1705 S_028814_POLYMODE_FRONT_PTYPE(si_translate_fill(raster_info->polygonMode)) |
1706 S_028814_POLYMODE_BACK_PTYPE(si_translate_fill(raster_info->polygonMode)) |
1707 S_028814_POLY_OFFSET_FRONT_ENABLE(raster_info->depthBiasEnable ? 1 : 0) |
1708 S_028814_POLY_OFFSET_BACK_ENABLE(raster_info->depthBiasEnable ? 1 : 0) |
1709 S_028814_POLY_OFFSET_PARA_ENABLE(raster_info->depthBiasEnable ? 1 : 0);
1710 }
1711
1712 static void
1713 radv_pipeline_init_depth_stencil_state(struct radv_pipeline *pipeline,
1714 const VkGraphicsPipelineCreateInfo *pCreateInfo)
1715 {
1716 const VkPipelineDepthStencilStateCreateInfo *ds_info
1717 = radv_pipeline_get_depth_stencil_state(pCreateInfo);
1718 RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass);
1719 struct radv_subpass *subpass = pass->subpasses + pCreateInfo->subpass;
1720 struct radv_render_pass_attachment *attachment = NULL;
1721 uint32_t db_depth_control = 0;
1722
1723 if (subpass->depth_stencil_attachment)
1724 attachment = pass->attachments + subpass->depth_stencil_attachment->attachment;
1725
1726 bool has_depth_attachment = attachment && vk_format_is_depth(attachment->format);
1727 bool has_stencil_attachment = attachment && vk_format_is_stencil(attachment->format);
1728
1729 if (ds_info) {
1730 if (has_depth_attachment) {
1731 db_depth_control = S_028800_Z_ENABLE(ds_info->depthTestEnable ? 1 : 0) |
1732 S_028800_Z_WRITE_ENABLE(ds_info->depthWriteEnable ? 1 : 0) |
1733 S_028800_ZFUNC(ds_info->depthCompareOp) |
1734 S_028800_DEPTH_BOUNDS_ENABLE(ds_info->depthBoundsTestEnable ? 1 : 0);
1735 }
1736
1737 if (has_stencil_attachment && ds_info->stencilTestEnable) {
1738 db_depth_control |= S_028800_STENCIL_ENABLE(1) | S_028800_BACKFACE_ENABLE(1);
1739 db_depth_control |= S_028800_STENCILFUNC(ds_info->front.compareOp);
1740 db_depth_control |= S_028800_STENCILFUNC_BF(ds_info->back.compareOp);
1741 }
1742 }
1743
1744 pipeline->graphics.db_depth_control = db_depth_control;
1745 }
1746
1747 static void
1748 gfx9_get_gs_info(const struct radv_pipeline_key *key,
1749 const struct radv_pipeline *pipeline,
1750 nir_shader **nir,
1751 struct radv_shader_info *infos,
1752 struct gfx9_gs_info *out)
1753 {
1754 struct radv_shader_info *gs_info = &infos[MESA_SHADER_GEOMETRY];
1755 struct radv_es_output_info *es_info;
1756 if (pipeline->device->physical_device->rad_info.chip_class >= GFX9)
1757 es_info = nir[MESA_SHADER_TESS_CTRL] ? &gs_info->tes.es_info : &gs_info->vs.es_info;
1758 else
1759 es_info = nir[MESA_SHADER_TESS_CTRL] ?
1760 &infos[MESA_SHADER_TESS_EVAL].tes.es_info :
1761 &infos[MESA_SHADER_VERTEX].vs.es_info;
1762
1763 unsigned gs_num_invocations = MAX2(gs_info->gs.invocations, 1);
1764 bool uses_adjacency;
1765 switch(key->topology) {
1766 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY:
1767 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY:
1768 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY:
1769 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY:
1770 uses_adjacency = true;
1771 break;
1772 default:
1773 uses_adjacency = false;
1774 break;
1775 }
1776
1777 /* All these are in dwords: */
1778 /* We can't allow using the whole LDS, because GS waves compete with
1779 * other shader stages for LDS space. */
1780 const unsigned max_lds_size = 8 * 1024;
1781 const unsigned esgs_itemsize = es_info->esgs_itemsize / 4;
1782 unsigned esgs_lds_size;
1783
1784 /* All these are per subgroup: */
1785 const unsigned max_out_prims = 32 * 1024;
1786 const unsigned max_es_verts = 255;
1787 const unsigned ideal_gs_prims = 64;
1788 unsigned max_gs_prims, gs_prims;
1789 unsigned min_es_verts, es_verts, worst_case_es_verts;
1790
1791 if (uses_adjacency || gs_num_invocations > 1)
1792 max_gs_prims = 127 / gs_num_invocations;
1793 else
1794 max_gs_prims = 255;
1795
1796 /* MAX_PRIMS_PER_SUBGROUP = gs_prims * max_vert_out * gs_invocations.
1797 * Make sure we don't go over the maximum value.
1798 */
1799 if (gs_info->gs.vertices_out > 0) {
1800 max_gs_prims = MIN2(max_gs_prims,
1801 max_out_prims /
1802 (gs_info->gs.vertices_out * gs_num_invocations));
1803 }
1804 assert(max_gs_prims > 0);
1805
1806 /* If the primitive has adjacency, halve the number of vertices
1807 * that will be reused in multiple primitives.
1808 */
1809 min_es_verts = gs_info->gs.vertices_in / (uses_adjacency ? 2 : 1);
1810
1811 gs_prims = MIN2(ideal_gs_prims, max_gs_prims);
1812 worst_case_es_verts = MIN2(min_es_verts * gs_prims, max_es_verts);
1813
1814 /* Compute ESGS LDS size based on the worst case number of ES vertices
1815 * needed to create the target number of GS prims per subgroup.
1816 */
1817 esgs_lds_size = esgs_itemsize * worst_case_es_verts;
1818
1819 /* If total LDS usage is too big, refactor partitions based on ratio
1820 * of ESGS item sizes.
1821 */
1822 if (esgs_lds_size > max_lds_size) {
1823 /* Our target GS Prims Per Subgroup was too large. Calculate
1824 * the maximum number of GS Prims Per Subgroup that will fit
1825 * into LDS, capped by the maximum that the hardware can support.
1826 */
1827 gs_prims = MIN2((max_lds_size / (esgs_itemsize * min_es_verts)),
1828 max_gs_prims);
1829 assert(gs_prims > 0);
1830 worst_case_es_verts = MIN2(min_es_verts * gs_prims,
1831 max_es_verts);
1832
1833 esgs_lds_size = esgs_itemsize * worst_case_es_verts;
1834 assert(esgs_lds_size <= max_lds_size);
1835 }
1836
1837 /* Now calculate remaining ESGS information. */
1838 if (esgs_lds_size)
1839 es_verts = MIN2(esgs_lds_size / esgs_itemsize, max_es_verts);
1840 else
1841 es_verts = max_es_verts;
1842
1843 /* Vertices for adjacency primitives are not always reused, so restore
1844 * it for ES_VERTS_PER_SUBGRP.
1845 */
1846 min_es_verts = gs_info->gs.vertices_in;
1847
1848 /* For normal primitives, the VGT only checks if they are past the ES
1849 * verts per subgroup after allocating a full GS primitive and if they
1850 * are, kick off a new subgroup. But if those additional ES verts are
1851 * unique (e.g. not reused) we need to make sure there is enough LDS
1852 * space to account for those ES verts beyond ES_VERTS_PER_SUBGRP.
1853 */
1854 es_verts -= min_es_verts - 1;
1855
1856 uint32_t es_verts_per_subgroup = es_verts;
1857 uint32_t gs_prims_per_subgroup = gs_prims;
1858 uint32_t gs_inst_prims_in_subgroup = gs_prims * gs_num_invocations;
1859 uint32_t max_prims_per_subgroup = gs_inst_prims_in_subgroup * gs_info->gs.vertices_out;
1860 out->lds_size = align(esgs_lds_size, 128) / 128;
1861 out->vgt_gs_onchip_cntl = S_028A44_ES_VERTS_PER_SUBGRP(es_verts_per_subgroup) |
1862 S_028A44_GS_PRIMS_PER_SUBGRP(gs_prims_per_subgroup) |
1863 S_028A44_GS_INST_PRIMS_IN_SUBGRP(gs_inst_prims_in_subgroup);
1864 out->vgt_gs_max_prims_per_subgroup = S_028A94_MAX_PRIMS_PER_SUBGROUP(max_prims_per_subgroup);
1865 out->vgt_esgs_ring_itemsize = esgs_itemsize;
1866 assert(max_prims_per_subgroup <= max_out_prims);
1867 }
1868
1869 static void clamp_gsprims_to_esverts(unsigned *max_gsprims, unsigned max_esverts,
1870 unsigned min_verts_per_prim, bool use_adjacency)
1871 {
1872 unsigned max_reuse = max_esverts - min_verts_per_prim;
1873 if (use_adjacency)
1874 max_reuse /= 2;
1875 *max_gsprims = MIN2(*max_gsprims, 1 + max_reuse);
1876 }
1877
1878 static unsigned
1879 radv_get_num_input_vertices(nir_shader **nir)
1880 {
1881 if (nir[MESA_SHADER_GEOMETRY]) {
1882 nir_shader *gs = nir[MESA_SHADER_GEOMETRY];
1883
1884 return gs->info.gs.vertices_in;
1885 }
1886
1887 if (nir[MESA_SHADER_TESS_CTRL]) {
1888 nir_shader *tes = nir[MESA_SHADER_TESS_EVAL];
1889
1890 if (tes->info.tess.point_mode)
1891 return 1;
1892 if (tes->info.tess.primitive_mode == GL_ISOLINES)
1893 return 2;
1894 return 3;
1895 }
1896
1897 return 3;
1898 }
1899
1900 static void
1901 gfx10_get_ngg_info(const struct radv_pipeline_key *key,
1902 struct radv_pipeline *pipeline,
1903 nir_shader **nir,
1904 struct radv_shader_info *infos,
1905 struct gfx10_ngg_info *ngg)
1906 {
1907 struct radv_shader_info *gs_info = &infos[MESA_SHADER_GEOMETRY];
1908 struct radv_es_output_info *es_info =
1909 nir[MESA_SHADER_TESS_CTRL] ? &gs_info->tes.es_info : &gs_info->vs.es_info;
1910 unsigned gs_type = nir[MESA_SHADER_GEOMETRY] ? MESA_SHADER_GEOMETRY : MESA_SHADER_VERTEX;
1911 unsigned max_verts_per_prim = radv_get_num_input_vertices(nir);
1912 unsigned min_verts_per_prim =
1913 gs_type == MESA_SHADER_GEOMETRY ? max_verts_per_prim : 1;
1914 unsigned gs_num_invocations = nir[MESA_SHADER_GEOMETRY] ? MAX2(gs_info->gs.invocations, 1) : 1;
1915 bool uses_adjacency;
1916 switch(key->topology) {
1917 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY:
1918 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY:
1919 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY:
1920 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY:
1921 uses_adjacency = true;
1922 break;
1923 default:
1924 uses_adjacency = false;
1925 break;
1926 }
1927
1928 /* All these are in dwords: */
1929 /* We can't allow using the whole LDS, because GS waves compete with
1930 * other shader stages for LDS space.
1931 *
1932 * TODO: We should really take the shader's internal LDS use into
1933 * account. The linker will fail if the size is greater than
1934 * 8K dwords.
1935 */
1936 const unsigned max_lds_size = 8 * 1024 - 768;
1937 const unsigned target_lds_size = max_lds_size;
1938 unsigned esvert_lds_size = 0;
1939 unsigned gsprim_lds_size = 0;
1940
1941 /* All these are per subgroup: */
1942 const unsigned min_esverts = pipeline->device->physical_device->rad_info.chip_class >= GFX10_3 ? 29 : 24;
1943 bool max_vert_out_per_gs_instance = false;
1944 unsigned max_esverts_base = 256;
1945 unsigned max_gsprims_base = 128; /* default prim group size clamp */
1946
1947 /* Hardware has the following non-natural restrictions on the value
1948 * of GE_CNTL.VERT_GRP_SIZE based on based on the primitive type of
1949 * the draw:
1950 * - at most 252 for any line input primitive type
1951 * - at most 251 for any quad input primitive type
1952 * - at most 251 for triangle strips with adjacency (this happens to
1953 * be the natural limit for triangle *lists* with adjacency)
1954 */
1955 max_esverts_base = MIN2(max_esverts_base, 251 + max_verts_per_prim - 1);
1956
1957 if (gs_type == MESA_SHADER_GEOMETRY) {
1958 unsigned max_out_verts_per_gsprim =
1959 gs_info->gs.vertices_out * gs_num_invocations;
1960
1961 if (max_out_verts_per_gsprim <= 256) {
1962 if (max_out_verts_per_gsprim) {
1963 max_gsprims_base = MIN2(max_gsprims_base,
1964 256 / max_out_verts_per_gsprim);
1965 }
1966 } else {
1967 /* Use special multi-cycling mode in which each GS
1968 * instance gets its own subgroup. Does not work with
1969 * tessellation. */
1970 max_vert_out_per_gs_instance = true;
1971 max_gsprims_base = 1;
1972 max_out_verts_per_gsprim = gs_info->gs.vertices_out;
1973 }
1974
1975 esvert_lds_size = es_info->esgs_itemsize / 4;
1976 gsprim_lds_size = (gs_info->gs.gsvs_vertex_size / 4 + 1) * max_out_verts_per_gsprim;
1977 } else {
1978 /* VS and TES. */
1979 /* LDS size for passing data from GS to ES. */
1980 struct radv_streamout_info *so_info = nir[MESA_SHADER_TESS_CTRL]
1981 ? &infos[MESA_SHADER_TESS_EVAL].so
1982 : &infos[MESA_SHADER_VERTEX].so;
1983
1984 if (so_info->num_outputs)
1985 esvert_lds_size = 4 * so_info->num_outputs + 1;
1986
1987 /* GS stores Primitive IDs (one DWORD) into LDS at the address
1988 * corresponding to the ES thread of the provoking vertex. All
1989 * ES threads load and export PrimitiveID for their thread.
1990 */
1991 if (!nir[MESA_SHADER_TESS_CTRL] &&
1992 infos[MESA_SHADER_VERTEX].vs.outinfo.export_prim_id)
1993 esvert_lds_size = MAX2(esvert_lds_size, 1);
1994 }
1995
1996 unsigned max_gsprims = max_gsprims_base;
1997 unsigned max_esverts = max_esverts_base;
1998
1999 if (esvert_lds_size)
2000 max_esverts = MIN2(max_esverts, target_lds_size / esvert_lds_size);
2001 if (gsprim_lds_size)
2002 max_gsprims = MIN2(max_gsprims, target_lds_size / gsprim_lds_size);
2003
2004 max_esverts = MIN2(max_esverts, max_gsprims * max_verts_per_prim);
2005 clamp_gsprims_to_esverts(&max_gsprims, max_esverts, min_verts_per_prim, uses_adjacency);
2006 assert(max_esverts >= max_verts_per_prim && max_gsprims >= 1);
2007
2008 if (esvert_lds_size || gsprim_lds_size) {
2009 /* Now that we have a rough proportionality between esverts
2010 * and gsprims based on the primitive type, scale both of them
2011 * down simultaneously based on required LDS space.
2012 *
2013 * We could be smarter about this if we knew how much vertex
2014 * reuse to expect.
2015 */
2016 unsigned lds_total = max_esverts * esvert_lds_size +
2017 max_gsprims * gsprim_lds_size;
2018 if (lds_total > target_lds_size) {
2019 max_esverts = max_esverts * target_lds_size / lds_total;
2020 max_gsprims = max_gsprims * target_lds_size / lds_total;
2021
2022 max_esverts = MIN2(max_esverts, max_gsprims * max_verts_per_prim);
2023 clamp_gsprims_to_esverts(&max_gsprims, max_esverts,
2024 min_verts_per_prim, uses_adjacency);
2025 assert(max_esverts >= max_verts_per_prim && max_gsprims >= 1);
2026 }
2027 }
2028
2029 /* Round up towards full wave sizes for better ALU utilization. */
2030 if (!max_vert_out_per_gs_instance) {
2031 unsigned orig_max_esverts;
2032 unsigned orig_max_gsprims;
2033 unsigned wavesize;
2034
2035 if (gs_type == MESA_SHADER_GEOMETRY) {
2036 wavesize = gs_info->wave_size;
2037 } else {
2038 wavesize = nir[MESA_SHADER_TESS_CTRL]
2039 ? infos[MESA_SHADER_TESS_EVAL].wave_size
2040 : infos[MESA_SHADER_VERTEX].wave_size;
2041 }
2042
2043 do {
2044 orig_max_esverts = max_esverts;
2045 orig_max_gsprims = max_gsprims;
2046
2047 max_esverts = align(max_esverts, wavesize);
2048 max_esverts = MIN2(max_esverts, max_esverts_base);
2049 if (esvert_lds_size)
2050 max_esverts = MIN2(max_esverts,
2051 (max_lds_size - max_gsprims * gsprim_lds_size) /
2052 esvert_lds_size);
2053 max_esverts = MIN2(max_esverts, max_gsprims * max_verts_per_prim);
2054
2055 max_gsprims = align(max_gsprims, wavesize);
2056 max_gsprims = MIN2(max_gsprims, max_gsprims_base);
2057 if (gsprim_lds_size)
2058 max_gsprims = MIN2(max_gsprims,
2059 (max_lds_size - max_esverts * esvert_lds_size) /
2060 gsprim_lds_size);
2061 clamp_gsprims_to_esverts(&max_gsprims, max_esverts,
2062 min_verts_per_prim, uses_adjacency);
2063 assert(max_esverts >= max_verts_per_prim && max_gsprims >= 1);
2064 } while (orig_max_esverts != max_esverts || orig_max_gsprims != max_gsprims);
2065 }
2066
2067 /* Hardware restriction: minimum value of max_esverts */
2068 max_esverts = MAX2(max_esverts, min_esverts - 1 + max_verts_per_prim);
2069
2070 unsigned max_out_vertices =
2071 max_vert_out_per_gs_instance ? gs_info->gs.vertices_out :
2072 gs_type == MESA_SHADER_GEOMETRY ?
2073 max_gsprims * gs_num_invocations * gs_info->gs.vertices_out :
2074 max_esverts;
2075 assert(max_out_vertices <= 256);
2076
2077 unsigned prim_amp_factor = 1;
2078 if (gs_type == MESA_SHADER_GEOMETRY) {
2079 /* Number of output primitives per GS input primitive after
2080 * GS instancing. */
2081 prim_amp_factor = gs_info->gs.vertices_out;
2082 }
2083
2084 /* The GE only checks against the maximum number of ES verts after
2085 * allocating a full GS primitive. So we need to ensure that whenever
2086 * this check passes, there is enough space for a full primitive without
2087 * vertex reuse.
2088 */
2089 ngg->hw_max_esverts = max_esverts - max_verts_per_prim + 1;
2090 ngg->max_gsprims = max_gsprims;
2091 ngg->max_out_verts = max_out_vertices;
2092 ngg->prim_amp_factor = prim_amp_factor;
2093 ngg->max_vert_out_per_gs_instance = max_vert_out_per_gs_instance;
2094 ngg->ngg_emit_size = max_gsprims * gsprim_lds_size;
2095 ngg->esgs_ring_size = 4 * max_esverts * esvert_lds_size;
2096
2097 if (gs_type == MESA_SHADER_GEOMETRY) {
2098 ngg->vgt_esgs_ring_itemsize = es_info->esgs_itemsize / 4;
2099 } else {
2100 ngg->vgt_esgs_ring_itemsize = 1;
2101 }
2102
2103 pipeline->graphics.esgs_ring_size = ngg->esgs_ring_size;
2104
2105 assert(ngg->hw_max_esverts >= min_esverts); /* HW limitation */
2106 }
2107
2108 static void
2109 radv_pipeline_init_gs_ring_state(struct radv_pipeline *pipeline,
2110 const struct gfx9_gs_info *gs)
2111 {
2112 struct radv_device *device = pipeline->device;
2113 unsigned num_se = device->physical_device->rad_info.max_se;
2114 unsigned wave_size = 64;
2115 unsigned max_gs_waves = 32 * num_se; /* max 32 per SE on GCN */
2116 /* On GFX6-GFX7, the value comes from VGT_GS_VERTEX_REUSE = 16.
2117 * On GFX8+, the value comes from VGT_VERTEX_REUSE_BLOCK_CNTL = 30 (+2).
2118 */
2119 unsigned gs_vertex_reuse =
2120 (device->physical_device->rad_info.chip_class >= GFX8 ? 32 : 16) * num_se;
2121 unsigned alignment = 256 * num_se;
2122 /* The maximum size is 63.999 MB per SE. */
2123 unsigned max_size = ((unsigned)(63.999 * 1024 * 1024) & ~255) * num_se;
2124 struct radv_shader_info *gs_info = &pipeline->shaders[MESA_SHADER_GEOMETRY]->info;
2125
2126 /* Calculate the minimum size. */
2127 unsigned min_esgs_ring_size = align(gs->vgt_esgs_ring_itemsize * 4 * gs_vertex_reuse *
2128 wave_size, alignment);
2129 /* These are recommended sizes, not minimum sizes. */
2130 unsigned esgs_ring_size = max_gs_waves * 2 * wave_size *
2131 gs->vgt_esgs_ring_itemsize * 4 * gs_info->gs.vertices_in;
2132 unsigned gsvs_ring_size = max_gs_waves * 2 * wave_size *
2133 gs_info->gs.max_gsvs_emit_size;
2134
2135 min_esgs_ring_size = align(min_esgs_ring_size, alignment);
2136 esgs_ring_size = align(esgs_ring_size, alignment);
2137 gsvs_ring_size = align(gsvs_ring_size, alignment);
2138
2139 if (pipeline->device->physical_device->rad_info.chip_class <= GFX8)
2140 pipeline->graphics.esgs_ring_size = CLAMP(esgs_ring_size, min_esgs_ring_size, max_size);
2141
2142 pipeline->graphics.gsvs_ring_size = MIN2(gsvs_ring_size, max_size);
2143 }
2144
2145 struct radv_shader_variant *
2146 radv_get_shader(const struct radv_pipeline *pipeline,
2147 gl_shader_stage stage)
2148 {
2149 if (stage == MESA_SHADER_VERTEX) {
2150 if (pipeline->shaders[MESA_SHADER_VERTEX])
2151 return pipeline->shaders[MESA_SHADER_VERTEX];
2152 if (pipeline->shaders[MESA_SHADER_TESS_CTRL])
2153 return pipeline->shaders[MESA_SHADER_TESS_CTRL];
2154 if (pipeline->shaders[MESA_SHADER_GEOMETRY])
2155 return pipeline->shaders[MESA_SHADER_GEOMETRY];
2156 } else if (stage == MESA_SHADER_TESS_EVAL) {
2157 if (!radv_pipeline_has_tess(pipeline))
2158 return NULL;
2159 if (pipeline->shaders[MESA_SHADER_TESS_EVAL])
2160 return pipeline->shaders[MESA_SHADER_TESS_EVAL];
2161 if (pipeline->shaders[MESA_SHADER_GEOMETRY])
2162 return pipeline->shaders[MESA_SHADER_GEOMETRY];
2163 }
2164 return pipeline->shaders[stage];
2165 }
2166
2167 static const struct radv_vs_output_info *get_vs_output_info(const struct radv_pipeline *pipeline)
2168 {
2169 if (radv_pipeline_has_gs(pipeline))
2170 if (radv_pipeline_has_ngg(pipeline))
2171 return &pipeline->shaders[MESA_SHADER_GEOMETRY]->info.vs.outinfo;
2172 else
2173 return &pipeline->gs_copy_shader->info.vs.outinfo;
2174 else if (radv_pipeline_has_tess(pipeline))
2175 return &pipeline->shaders[MESA_SHADER_TESS_EVAL]->info.tes.outinfo;
2176 else
2177 return &pipeline->shaders[MESA_SHADER_VERTEX]->info.vs.outinfo;
2178 }
2179
2180 static void
2181 radv_link_shaders(struct radv_pipeline *pipeline, nir_shader **shaders)
2182 {
2183 nir_shader* ordered_shaders[MESA_SHADER_STAGES];
2184 int shader_count = 0;
2185
2186 if(shaders[MESA_SHADER_FRAGMENT]) {
2187 ordered_shaders[shader_count++] = shaders[MESA_SHADER_FRAGMENT];
2188 }
2189 if(shaders[MESA_SHADER_GEOMETRY]) {
2190 ordered_shaders[shader_count++] = shaders[MESA_SHADER_GEOMETRY];
2191 }
2192 if(shaders[MESA_SHADER_TESS_EVAL]) {
2193 ordered_shaders[shader_count++] = shaders[MESA_SHADER_TESS_EVAL];
2194 }
2195 if(shaders[MESA_SHADER_TESS_CTRL]) {
2196 ordered_shaders[shader_count++] = shaders[MESA_SHADER_TESS_CTRL];
2197 }
2198 if(shaders[MESA_SHADER_VERTEX]) {
2199 ordered_shaders[shader_count++] = shaders[MESA_SHADER_VERTEX];
2200 }
2201
2202 if (shader_count > 1) {
2203 unsigned first = ordered_shaders[shader_count - 1]->info.stage;
2204 unsigned last = ordered_shaders[0]->info.stage;
2205
2206 if (ordered_shaders[0]->info.stage == MESA_SHADER_FRAGMENT &&
2207 ordered_shaders[1]->info.has_transform_feedback_varyings)
2208 nir_link_xfb_varyings(ordered_shaders[1], ordered_shaders[0]);
2209
2210 for (int i = 0; i < shader_count; ++i) {
2211 nir_variable_mode mask = 0;
2212
2213 if (ordered_shaders[i]->info.stage != first)
2214 mask = mask | nir_var_shader_in;
2215
2216 if (ordered_shaders[i]->info.stage != last)
2217 mask = mask | nir_var_shader_out;
2218
2219 nir_lower_io_to_scalar_early(ordered_shaders[i], mask);
2220 radv_optimize_nir(ordered_shaders[i], false, false);
2221 }
2222 }
2223
2224 for (int i = 1; i < shader_count; ++i) {
2225 nir_lower_io_arrays_to_elements(ordered_shaders[i],
2226 ordered_shaders[i - 1]);
2227
2228 if (nir_link_opt_varyings(ordered_shaders[i],
2229 ordered_shaders[i - 1]))
2230 radv_optimize_nir(ordered_shaders[i - 1], false, false);
2231
2232 nir_remove_dead_variables(ordered_shaders[i],
2233 nir_var_shader_out, NULL);
2234 nir_remove_dead_variables(ordered_shaders[i - 1],
2235 nir_var_shader_in, NULL);
2236
2237 bool progress = nir_remove_unused_varyings(ordered_shaders[i],
2238 ordered_shaders[i - 1]);
2239
2240 nir_compact_varyings(ordered_shaders[i],
2241 ordered_shaders[i - 1], true);
2242
2243 if (progress) {
2244 if (nir_lower_global_vars_to_local(ordered_shaders[i])) {
2245 ac_lower_indirect_derefs(ordered_shaders[i],
2246 pipeline->device->physical_device->rad_info.chip_class);
2247 }
2248 radv_optimize_nir(ordered_shaders[i], false, false);
2249
2250 if (nir_lower_global_vars_to_local(ordered_shaders[i - 1])) {
2251 ac_lower_indirect_derefs(ordered_shaders[i - 1],
2252 pipeline->device->physical_device->rad_info.chip_class);
2253 }
2254 radv_optimize_nir(ordered_shaders[i - 1], false, false);
2255 }
2256 }
2257 }
2258
2259 static void
2260 radv_set_linked_driver_locations(struct radv_pipeline *pipeline, nir_shader **shaders,
2261 struct radv_shader_info infos[MESA_SHADER_STAGES])
2262 {
2263 bool has_tess = shaders[MESA_SHADER_TESS_CTRL];
2264 bool has_gs = shaders[MESA_SHADER_GEOMETRY];
2265
2266 if (!has_tess && !has_gs)
2267 return;
2268
2269 unsigned vs_info_idx = MESA_SHADER_VERTEX;
2270 unsigned tes_info_idx = MESA_SHADER_TESS_EVAL;
2271
2272 if (pipeline->device->physical_device->rad_info.chip_class >= GFX9) {
2273 /* These are merged into the next stage */
2274 vs_info_idx = has_tess ? MESA_SHADER_TESS_CTRL : MESA_SHADER_GEOMETRY;
2275 tes_info_idx = has_gs ? MESA_SHADER_GEOMETRY : MESA_SHADER_TESS_EVAL;
2276 }
2277
2278 if (has_tess) {
2279 nir_linked_io_var_info vs2tcs =
2280 nir_assign_linked_io_var_locations(shaders[MESA_SHADER_VERTEX], shaders[MESA_SHADER_TESS_CTRL]);
2281 nir_linked_io_var_info tcs2tes =
2282 nir_assign_linked_io_var_locations(shaders[MESA_SHADER_TESS_CTRL], shaders[MESA_SHADER_TESS_EVAL]);
2283
2284 infos[vs_info_idx].vs.num_linked_outputs = vs2tcs.num_linked_io_vars;
2285 infos[MESA_SHADER_TESS_CTRL].tcs.num_linked_inputs = vs2tcs.num_linked_io_vars;
2286 infos[MESA_SHADER_TESS_CTRL].tcs.num_linked_outputs = tcs2tes.num_linked_io_vars;
2287 infos[MESA_SHADER_TESS_CTRL].tcs.num_linked_patch_outputs = tcs2tes.num_linked_patch_io_vars;
2288 infos[tes_info_idx].tes.num_linked_inputs = tcs2tes.num_linked_io_vars;
2289 infos[tes_info_idx].tes.num_linked_patch_inputs = tcs2tes.num_linked_patch_io_vars;
2290
2291 if (has_gs) {
2292 nir_linked_io_var_info tes2gs =
2293 nir_assign_linked_io_var_locations(shaders[MESA_SHADER_TESS_EVAL], shaders[MESA_SHADER_GEOMETRY]);
2294
2295 infos[tes_info_idx].tes.num_linked_outputs = tes2gs.num_linked_io_vars;
2296 infos[MESA_SHADER_GEOMETRY].gs.num_linked_inputs = tes2gs.num_linked_io_vars;
2297 }
2298 } else if (has_gs) {
2299 nir_linked_io_var_info vs2gs =
2300 nir_assign_linked_io_var_locations(shaders[MESA_SHADER_VERTEX], shaders[MESA_SHADER_GEOMETRY]);
2301
2302 infos[vs_info_idx].vs.num_linked_outputs = vs2gs.num_linked_io_vars;
2303 infos[MESA_SHADER_GEOMETRY].gs.num_linked_inputs = vs2gs.num_linked_io_vars;
2304 }
2305 }
2306
2307 static uint32_t
2308 radv_get_attrib_stride(const VkPipelineVertexInputStateCreateInfo *input_state,
2309 uint32_t attrib_binding)
2310 {
2311 for (uint32_t i = 0; i < input_state->vertexBindingDescriptionCount; i++) {
2312 const VkVertexInputBindingDescription *input_binding =
2313 &input_state->pVertexBindingDescriptions[i];
2314
2315 if (input_binding->binding == attrib_binding)
2316 return input_binding->stride;
2317 }
2318
2319 return 0;
2320 }
2321
2322 static struct radv_pipeline_key
2323 radv_generate_graphics_pipeline_key(struct radv_pipeline *pipeline,
2324 const VkGraphicsPipelineCreateInfo *pCreateInfo,
2325 const struct radv_blend_state *blend)
2326 {
2327 RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass);
2328 struct radv_subpass *subpass = pass->subpasses + pCreateInfo->subpass;
2329 const VkPipelineVertexInputStateCreateInfo *input_state =
2330 pCreateInfo->pVertexInputState;
2331 const VkPipelineVertexInputDivisorStateCreateInfoEXT *divisor_state =
2332 vk_find_struct_const(input_state->pNext, PIPELINE_VERTEX_INPUT_DIVISOR_STATE_CREATE_INFO_EXT);
2333
2334 struct radv_pipeline_key key;
2335 memset(&key, 0, sizeof(key));
2336
2337 if (pCreateInfo->flags & VK_PIPELINE_CREATE_DISABLE_OPTIMIZATION_BIT)
2338 key.optimisations_disabled = 1;
2339
2340 key.has_multiview_view_index = !!subpass->view_mask;
2341
2342 uint32_t binding_input_rate = 0;
2343 uint32_t instance_rate_divisors[MAX_VERTEX_ATTRIBS];
2344 for (unsigned i = 0; i < input_state->vertexBindingDescriptionCount; ++i) {
2345 if (input_state->pVertexBindingDescriptions[i].inputRate) {
2346 unsigned binding = input_state->pVertexBindingDescriptions[i].binding;
2347 binding_input_rate |= 1u << binding;
2348 instance_rate_divisors[binding] = 1;
2349 }
2350 }
2351 if (divisor_state) {
2352 for (unsigned i = 0; i < divisor_state->vertexBindingDivisorCount; ++i) {
2353 instance_rate_divisors[divisor_state->pVertexBindingDivisors[i].binding] =
2354 divisor_state->pVertexBindingDivisors[i].divisor;
2355 }
2356 }
2357
2358 for (unsigned i = 0; i < input_state->vertexAttributeDescriptionCount; ++i) {
2359 const VkVertexInputAttributeDescription *desc =
2360 &input_state->pVertexAttributeDescriptions[i];
2361 const struct vk_format_description *format_desc;
2362 unsigned location = desc->location;
2363 unsigned binding = desc->binding;
2364 unsigned num_format, data_format;
2365 int first_non_void;
2366
2367 if (binding_input_rate & (1u << binding)) {
2368 key.instance_rate_inputs |= 1u << location;
2369 key.instance_rate_divisors[location] = instance_rate_divisors[binding];
2370 }
2371
2372 format_desc = vk_format_description(desc->format);
2373 first_non_void = vk_format_get_first_non_void_channel(desc->format);
2374
2375 num_format = radv_translate_buffer_numformat(format_desc, first_non_void);
2376 data_format = radv_translate_buffer_dataformat(format_desc, first_non_void);
2377
2378 key.vertex_attribute_formats[location] = data_format | (num_format << 4);
2379 key.vertex_attribute_bindings[location] = desc->binding;
2380 key.vertex_attribute_offsets[location] = desc->offset;
2381 key.vertex_attribute_strides[location] = radv_get_attrib_stride(input_state, desc->binding);
2382
2383 if (pipeline->device->physical_device->rad_info.chip_class <= GFX8 &&
2384 pipeline->device->physical_device->rad_info.family != CHIP_STONEY) {
2385 VkFormat format = input_state->pVertexAttributeDescriptions[i].format;
2386 uint64_t adjust;
2387 switch(format) {
2388 case VK_FORMAT_A2R10G10B10_SNORM_PACK32:
2389 case VK_FORMAT_A2B10G10R10_SNORM_PACK32:
2390 adjust = RADV_ALPHA_ADJUST_SNORM;
2391 break;
2392 case VK_FORMAT_A2R10G10B10_SSCALED_PACK32:
2393 case VK_FORMAT_A2B10G10R10_SSCALED_PACK32:
2394 adjust = RADV_ALPHA_ADJUST_SSCALED;
2395 break;
2396 case VK_FORMAT_A2R10G10B10_SINT_PACK32:
2397 case VK_FORMAT_A2B10G10R10_SINT_PACK32:
2398 adjust = RADV_ALPHA_ADJUST_SINT;
2399 break;
2400 default:
2401 adjust = 0;
2402 break;
2403 }
2404 key.vertex_alpha_adjust |= adjust << (2 * location);
2405 }
2406
2407 switch (desc->format) {
2408 case VK_FORMAT_B8G8R8A8_UNORM:
2409 case VK_FORMAT_B8G8R8A8_SNORM:
2410 case VK_FORMAT_B8G8R8A8_USCALED:
2411 case VK_FORMAT_B8G8R8A8_SSCALED:
2412 case VK_FORMAT_B8G8R8A8_UINT:
2413 case VK_FORMAT_B8G8R8A8_SINT:
2414 case VK_FORMAT_B8G8R8A8_SRGB:
2415 case VK_FORMAT_A2R10G10B10_UNORM_PACK32:
2416 case VK_FORMAT_A2R10G10B10_SNORM_PACK32:
2417 case VK_FORMAT_A2R10G10B10_USCALED_PACK32:
2418 case VK_FORMAT_A2R10G10B10_SSCALED_PACK32:
2419 case VK_FORMAT_A2R10G10B10_UINT_PACK32:
2420 case VK_FORMAT_A2R10G10B10_SINT_PACK32:
2421 key.vertex_post_shuffle |= 1 << location;
2422 break;
2423 default:
2424 break;
2425 }
2426 }
2427
2428 const VkPipelineTessellationStateCreateInfo *tess =
2429 radv_pipeline_get_tessellation_state(pCreateInfo);
2430 if (tess)
2431 key.tess_input_vertices = tess->patchControlPoints;
2432
2433 const VkPipelineMultisampleStateCreateInfo *vkms =
2434 radv_pipeline_get_multisample_state(pCreateInfo);
2435 if (vkms && vkms->rasterizationSamples > 1) {
2436 uint32_t num_samples = vkms->rasterizationSamples;
2437 uint32_t ps_iter_samples = radv_pipeline_get_ps_iter_samples(pCreateInfo);
2438 key.num_samples = num_samples;
2439 key.log2_ps_iter_samples = util_logbase2(ps_iter_samples);
2440 }
2441
2442 key.col_format = blend->spi_shader_col_format;
2443 key.is_dual_src = blend->mrt0_is_dual_src;
2444 if (pipeline->device->physical_device->rad_info.chip_class < GFX8) {
2445 key.is_int8 = blend->col_format_is_int8;
2446 key.is_int10 = blend->col_format_is_int10;
2447 }
2448
2449 if (pipeline->device->physical_device->rad_info.chip_class >= GFX10)
2450 key.topology = pCreateInfo->pInputAssemblyState->topology;
2451
2452 return key;
2453 }
2454
2455 static bool
2456 radv_nir_stage_uses_xfb(const nir_shader *nir)
2457 {
2458 nir_xfb_info *xfb = nir_gather_xfb_info(nir, NULL);
2459 bool uses_xfb = !!xfb;
2460
2461 ralloc_free(xfb);
2462 return uses_xfb;
2463 }
2464
2465 static void
2466 radv_fill_shader_keys(struct radv_device *device,
2467 struct radv_shader_variant_key *keys,
2468 const struct radv_pipeline_key *key,
2469 nir_shader **nir)
2470 {
2471 keys[MESA_SHADER_VERTEX].vs.instance_rate_inputs = key->instance_rate_inputs;
2472 keys[MESA_SHADER_VERTEX].vs.alpha_adjust = key->vertex_alpha_adjust;
2473 keys[MESA_SHADER_VERTEX].vs.post_shuffle = key->vertex_post_shuffle;
2474 for (unsigned i = 0; i < MAX_VERTEX_ATTRIBS; ++i) {
2475 keys[MESA_SHADER_VERTEX].vs.instance_rate_divisors[i] = key->instance_rate_divisors[i];
2476 keys[MESA_SHADER_VERTEX].vs.vertex_attribute_formats[i] = key->vertex_attribute_formats[i];
2477 keys[MESA_SHADER_VERTEX].vs.vertex_attribute_bindings[i] = key->vertex_attribute_bindings[i];
2478 keys[MESA_SHADER_VERTEX].vs.vertex_attribute_offsets[i] = key->vertex_attribute_offsets[i];
2479 keys[MESA_SHADER_VERTEX].vs.vertex_attribute_strides[i] = key->vertex_attribute_strides[i];
2480 }
2481 keys[MESA_SHADER_VERTEX].vs.outprim = si_conv_prim_to_gs_out(key->topology);
2482
2483 if (nir[MESA_SHADER_TESS_CTRL]) {
2484 keys[MESA_SHADER_VERTEX].vs_common_out.as_ls = true;
2485 keys[MESA_SHADER_TESS_CTRL].tcs.num_inputs = 0;
2486 keys[MESA_SHADER_TESS_CTRL].tcs.input_vertices = key->tess_input_vertices;
2487 keys[MESA_SHADER_TESS_CTRL].tcs.primitive_mode = nir[MESA_SHADER_TESS_EVAL]->info.tess.primitive_mode;
2488
2489 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));
2490 }
2491
2492 if (nir[MESA_SHADER_GEOMETRY]) {
2493 if (nir[MESA_SHADER_TESS_CTRL])
2494 keys[MESA_SHADER_TESS_EVAL].vs_common_out.as_es = true;
2495 else
2496 keys[MESA_SHADER_VERTEX].vs_common_out.as_es = true;
2497 }
2498
2499 if (device->physical_device->use_ngg) {
2500 if (nir[MESA_SHADER_TESS_CTRL]) {
2501 keys[MESA_SHADER_TESS_EVAL].vs_common_out.as_ngg = true;
2502 } else {
2503 keys[MESA_SHADER_VERTEX].vs_common_out.as_ngg = true;
2504 }
2505
2506 if (nir[MESA_SHADER_TESS_CTRL] &&
2507 nir[MESA_SHADER_GEOMETRY] &&
2508 nir[MESA_SHADER_GEOMETRY]->info.gs.invocations *
2509 nir[MESA_SHADER_GEOMETRY]->info.gs.vertices_out > 256) {
2510 /* Fallback to the legacy path if tessellation is
2511 * enabled with extreme geometry because
2512 * EN_MAX_VERT_OUT_PER_GS_INSTANCE doesn't work and it
2513 * might hang.
2514 */
2515 keys[MESA_SHADER_TESS_EVAL].vs_common_out.as_ngg = false;
2516 }
2517
2518 if (!device->physical_device->use_ngg_gs) {
2519 if (nir[MESA_SHADER_GEOMETRY]) {
2520 if (nir[MESA_SHADER_TESS_CTRL])
2521 keys[MESA_SHADER_TESS_EVAL].vs_common_out.as_ngg = false;
2522 else
2523 keys[MESA_SHADER_VERTEX].vs_common_out.as_ngg = false;
2524 }
2525 }
2526
2527 gl_shader_stage last_xfb_stage = MESA_SHADER_VERTEX;
2528
2529 for (int i = MESA_SHADER_VERTEX; i <= MESA_SHADER_GEOMETRY; i++) {
2530 if (nir[i])
2531 last_xfb_stage = i;
2532 }
2533
2534 bool uses_xfb = nir[last_xfb_stage] &&
2535 radv_nir_stage_uses_xfb(nir[last_xfb_stage]);
2536
2537 if (!device->physical_device->use_ngg_streamout && uses_xfb) {
2538 if (nir[MESA_SHADER_TESS_CTRL])
2539 keys[MESA_SHADER_TESS_EVAL].vs_common_out.as_ngg = false;
2540 else
2541 keys[MESA_SHADER_VERTEX].vs_common_out.as_ngg = false;
2542 }
2543
2544 /* Determine if the pipeline is eligible for the NGG passthrough
2545 * mode. It can't be enabled for geometry shaders, for NGG
2546 * streamout or for vertex shaders that export the primitive ID
2547 * (this is checked later because we don't have the info here.)
2548 */
2549 if (!nir[MESA_SHADER_GEOMETRY] && !uses_xfb) {
2550 if (nir[MESA_SHADER_TESS_CTRL] &&
2551 keys[MESA_SHADER_TESS_EVAL].vs_common_out.as_ngg) {
2552 keys[MESA_SHADER_TESS_EVAL].vs_common_out.as_ngg_passthrough = true;
2553 } else if (nir[MESA_SHADER_VERTEX] &&
2554 keys[MESA_SHADER_VERTEX].vs_common_out.as_ngg) {
2555 keys[MESA_SHADER_VERTEX].vs_common_out.as_ngg_passthrough = true;
2556 }
2557 }
2558 }
2559
2560 for(int i = 0; i < MESA_SHADER_STAGES; ++i)
2561 keys[i].has_multiview_view_index = key->has_multiview_view_index;
2562
2563 keys[MESA_SHADER_FRAGMENT].fs.col_format = key->col_format;
2564 keys[MESA_SHADER_FRAGMENT].fs.is_int8 = key->is_int8;
2565 keys[MESA_SHADER_FRAGMENT].fs.is_int10 = key->is_int10;
2566 keys[MESA_SHADER_FRAGMENT].fs.log2_ps_iter_samples = key->log2_ps_iter_samples;
2567 keys[MESA_SHADER_FRAGMENT].fs.num_samples = key->num_samples;
2568 keys[MESA_SHADER_FRAGMENT].fs.is_dual_src = key->is_dual_src;
2569
2570 if (nir[MESA_SHADER_COMPUTE]) {
2571 keys[MESA_SHADER_COMPUTE].cs.subgroup_size = key->compute_subgroup_size;
2572 }
2573 }
2574
2575 static uint8_t
2576 radv_get_wave_size(struct radv_device *device,
2577 const VkPipelineShaderStageCreateInfo *pStage,
2578 gl_shader_stage stage,
2579 const struct radv_shader_variant_key *key)
2580 {
2581 if (stage == MESA_SHADER_GEOMETRY && !key->vs_common_out.as_ngg)
2582 return 64;
2583 else if (stage == MESA_SHADER_COMPUTE) {
2584 if (key->cs.subgroup_size) {
2585 /* Return the required subgroup size if specified. */
2586 return key->cs.subgroup_size;
2587 }
2588 return device->physical_device->cs_wave_size;
2589 }
2590 else if (stage == MESA_SHADER_FRAGMENT)
2591 return device->physical_device->ps_wave_size;
2592 else
2593 return device->physical_device->ge_wave_size;
2594 }
2595
2596 static uint8_t
2597 radv_get_ballot_bit_size(struct radv_device *device,
2598 const VkPipelineShaderStageCreateInfo *pStage,
2599 gl_shader_stage stage,
2600 const struct radv_shader_variant_key *key)
2601 {
2602 if (stage == MESA_SHADER_COMPUTE && key->cs.subgroup_size)
2603 return key->cs.subgroup_size;
2604 return 64;
2605 }
2606
2607 static void
2608 radv_fill_shader_info(struct radv_pipeline *pipeline,
2609 const VkPipelineShaderStageCreateInfo **pStages,
2610 struct radv_shader_variant_key *keys,
2611 struct radv_shader_info *infos,
2612 nir_shader **nir)
2613 {
2614 unsigned active_stages = 0;
2615 unsigned filled_stages = 0;
2616
2617 for (int i = 0; i < MESA_SHADER_STAGES; i++) {
2618 if (nir[i])
2619 active_stages |= (1 << i);
2620 }
2621
2622 if (nir[MESA_SHADER_FRAGMENT]) {
2623 radv_nir_shader_info_init(&infos[MESA_SHADER_FRAGMENT]);
2624 radv_nir_shader_info_pass(nir[MESA_SHADER_FRAGMENT],
2625 pipeline->layout,
2626 &keys[MESA_SHADER_FRAGMENT],
2627 &infos[MESA_SHADER_FRAGMENT],
2628 pipeline->device->physical_device->use_llvm);
2629
2630 /* TODO: These are no longer used as keys we should refactor this */
2631 keys[MESA_SHADER_VERTEX].vs_common_out.export_prim_id =
2632 infos[MESA_SHADER_FRAGMENT].ps.prim_id_input;
2633 keys[MESA_SHADER_VERTEX].vs_common_out.export_layer_id =
2634 infos[MESA_SHADER_FRAGMENT].ps.layer_input;
2635 keys[MESA_SHADER_VERTEX].vs_common_out.export_clip_dists =
2636 !!infos[MESA_SHADER_FRAGMENT].ps.num_input_clips_culls;
2637 keys[MESA_SHADER_VERTEX].vs_common_out.export_viewport_index =
2638 infos[MESA_SHADER_FRAGMENT].ps.viewport_index_input;
2639 keys[MESA_SHADER_TESS_EVAL].vs_common_out.export_prim_id =
2640 infos[MESA_SHADER_FRAGMENT].ps.prim_id_input;
2641 keys[MESA_SHADER_TESS_EVAL].vs_common_out.export_layer_id =
2642 infos[MESA_SHADER_FRAGMENT].ps.layer_input;
2643 keys[MESA_SHADER_TESS_EVAL].vs_common_out.export_clip_dists =
2644 !!infos[MESA_SHADER_FRAGMENT].ps.num_input_clips_culls;
2645 keys[MESA_SHADER_TESS_EVAL].vs_common_out.export_viewport_index =
2646 infos[MESA_SHADER_FRAGMENT].ps.viewport_index_input;
2647
2648 /* NGG passthrough mode can't be enabled for vertex shaders
2649 * that export the primitive ID.
2650 *
2651 * TODO: I should really refactor the keys logic.
2652 */
2653 if (nir[MESA_SHADER_VERTEX] &&
2654 keys[MESA_SHADER_VERTEX].vs_common_out.export_prim_id) {
2655 keys[MESA_SHADER_VERTEX].vs_common_out.as_ngg_passthrough = false;
2656 }
2657
2658 filled_stages |= (1 << MESA_SHADER_FRAGMENT);
2659 }
2660
2661 if (nir[MESA_SHADER_TESS_CTRL]) {
2662 infos[MESA_SHADER_TESS_CTRL].tcs.tes_inputs_read =
2663 nir[MESA_SHADER_TESS_EVAL]->info.inputs_read;
2664 infos[MESA_SHADER_TESS_CTRL].tcs.tes_patch_inputs_read =
2665 nir[MESA_SHADER_TESS_EVAL]->info.patch_inputs_read;
2666 }
2667
2668 if (pipeline->device->physical_device->rad_info.chip_class >= GFX9 &&
2669 nir[MESA_SHADER_TESS_CTRL]) {
2670 struct nir_shader *combined_nir[] = {nir[MESA_SHADER_VERTEX], nir[MESA_SHADER_TESS_CTRL]};
2671 struct radv_shader_variant_key key = keys[MESA_SHADER_TESS_CTRL];
2672 key.tcs.vs_key = keys[MESA_SHADER_VERTEX].vs;
2673
2674 radv_nir_shader_info_init(&infos[MESA_SHADER_TESS_CTRL]);
2675
2676 for (int i = 0; i < 2; i++) {
2677 radv_nir_shader_info_pass(combined_nir[i],
2678 pipeline->layout, &key,
2679 &infos[MESA_SHADER_TESS_CTRL],
2680 pipeline->device->physical_device->use_llvm);
2681 }
2682
2683 keys[MESA_SHADER_TESS_EVAL].tes.num_patches =
2684 infos[MESA_SHADER_TESS_CTRL].tcs.num_patches;
2685 keys[MESA_SHADER_TESS_EVAL].tes.tcs_num_outputs =
2686 util_last_bit64(infos[MESA_SHADER_TESS_CTRL].tcs.outputs_written);
2687
2688 filled_stages |= (1 << MESA_SHADER_VERTEX);
2689 filled_stages |= (1 << MESA_SHADER_TESS_CTRL);
2690 }
2691
2692 if (pipeline->device->physical_device->rad_info.chip_class >= GFX9 &&
2693 nir[MESA_SHADER_GEOMETRY]) {
2694 gl_shader_stage pre_stage = nir[MESA_SHADER_TESS_EVAL] ? MESA_SHADER_TESS_EVAL : MESA_SHADER_VERTEX;
2695 struct nir_shader *combined_nir[] = {nir[pre_stage], nir[MESA_SHADER_GEOMETRY]};
2696
2697 radv_nir_shader_info_init(&infos[MESA_SHADER_GEOMETRY]);
2698
2699 for (int i = 0; i < 2; i++) {
2700 radv_nir_shader_info_pass(combined_nir[i],
2701 pipeline->layout,
2702 &keys[pre_stage],
2703 &infos[MESA_SHADER_GEOMETRY],
2704 pipeline->device->physical_device->use_llvm);
2705 }
2706
2707 filled_stages |= (1 << pre_stage);
2708 filled_stages |= (1 << MESA_SHADER_GEOMETRY);
2709 }
2710
2711 active_stages ^= filled_stages;
2712 while (active_stages) {
2713 int i = u_bit_scan(&active_stages);
2714
2715 if (i == MESA_SHADER_TESS_CTRL) {
2716 keys[MESA_SHADER_TESS_CTRL].tcs.num_inputs =
2717 util_last_bit64(infos[MESA_SHADER_VERTEX].vs.ls_outputs_written);
2718 }
2719
2720 if (i == MESA_SHADER_TESS_EVAL) {
2721 keys[MESA_SHADER_TESS_EVAL].tes.num_patches =
2722 infos[MESA_SHADER_TESS_CTRL].tcs.num_patches;
2723 keys[MESA_SHADER_TESS_EVAL].tes.tcs_num_outputs =
2724 util_last_bit64(infos[MESA_SHADER_TESS_CTRL].tcs.outputs_written);
2725 }
2726
2727 radv_nir_shader_info_init(&infos[i]);
2728 radv_nir_shader_info_pass(nir[i], pipeline->layout,
2729 &keys[i], &infos[i], pipeline->device->physical_device->use_llvm);
2730 }
2731
2732 for (int i = 0; i < MESA_SHADER_STAGES; i++) {
2733 if (nir[i]) {
2734 infos[i].wave_size =
2735 radv_get_wave_size(pipeline->device, pStages[i],
2736 i, &keys[i]);
2737 infos[i].ballot_bit_size =
2738 radv_get_ballot_bit_size(pipeline->device,
2739 pStages[i], i,
2740 &keys[i]);
2741 }
2742 }
2743 }
2744
2745 static void
2746 merge_tess_info(struct shader_info *tes_info,
2747 const struct shader_info *tcs_info)
2748 {
2749 /* The Vulkan 1.0.38 spec, section 21.1 Tessellator says:
2750 *
2751 * "PointMode. Controls generation of points rather than triangles
2752 * or lines. This functionality defaults to disabled, and is
2753 * enabled if either shader stage includes the execution mode.
2754 *
2755 * and about Triangles, Quads, IsoLines, VertexOrderCw, VertexOrderCcw,
2756 * PointMode, SpacingEqual, SpacingFractionalEven, SpacingFractionalOdd,
2757 * and OutputVertices, it says:
2758 *
2759 * "One mode must be set in at least one of the tessellation
2760 * shader stages."
2761 *
2762 * So, the fields can be set in either the TCS or TES, but they must
2763 * agree if set in both. Our backend looks at TES, so bitwise-or in
2764 * the values from the TCS.
2765 */
2766 assert(tcs_info->tess.tcs_vertices_out == 0 ||
2767 tes_info->tess.tcs_vertices_out == 0 ||
2768 tcs_info->tess.tcs_vertices_out == tes_info->tess.tcs_vertices_out);
2769 tes_info->tess.tcs_vertices_out |= tcs_info->tess.tcs_vertices_out;
2770
2771 assert(tcs_info->tess.spacing == TESS_SPACING_UNSPECIFIED ||
2772 tes_info->tess.spacing == TESS_SPACING_UNSPECIFIED ||
2773 tcs_info->tess.spacing == tes_info->tess.spacing);
2774 tes_info->tess.spacing |= tcs_info->tess.spacing;
2775
2776 assert(tcs_info->tess.primitive_mode == 0 ||
2777 tes_info->tess.primitive_mode == 0 ||
2778 tcs_info->tess.primitive_mode == tes_info->tess.primitive_mode);
2779 tes_info->tess.primitive_mode |= tcs_info->tess.primitive_mode;
2780 tes_info->tess.ccw |= tcs_info->tess.ccw;
2781 tes_info->tess.point_mode |= tcs_info->tess.point_mode;
2782 }
2783
2784 static
2785 void radv_init_feedback(const VkPipelineCreationFeedbackCreateInfoEXT *ext)
2786 {
2787 if (!ext)
2788 return;
2789
2790 if (ext->pPipelineCreationFeedback) {
2791 ext->pPipelineCreationFeedback->flags = 0;
2792 ext->pPipelineCreationFeedback->duration = 0;
2793 }
2794
2795 for (unsigned i = 0; i < ext->pipelineStageCreationFeedbackCount; ++i) {
2796 ext->pPipelineStageCreationFeedbacks[i].flags = 0;
2797 ext->pPipelineStageCreationFeedbacks[i].duration = 0;
2798 }
2799 }
2800
2801 static
2802 void radv_start_feedback(VkPipelineCreationFeedbackEXT *feedback)
2803 {
2804 if (!feedback)
2805 return;
2806
2807 feedback->duration -= radv_get_current_time();
2808 feedback ->flags = VK_PIPELINE_CREATION_FEEDBACK_VALID_BIT_EXT;
2809 }
2810
2811 static
2812 void radv_stop_feedback(VkPipelineCreationFeedbackEXT *feedback, bool cache_hit)
2813 {
2814 if (!feedback)
2815 return;
2816
2817 feedback->duration += radv_get_current_time();
2818 feedback ->flags = VK_PIPELINE_CREATION_FEEDBACK_VALID_BIT_EXT |
2819 (cache_hit ? VK_PIPELINE_CREATION_FEEDBACK_APPLICATION_PIPELINE_CACHE_HIT_BIT_EXT : 0);
2820 }
2821
2822 VkResult radv_create_shaders(struct radv_pipeline *pipeline,
2823 struct radv_device *device,
2824 struct radv_pipeline_cache *cache,
2825 const struct radv_pipeline_key *key,
2826 const VkPipelineShaderStageCreateInfo **pStages,
2827 const VkPipelineCreateFlags flags,
2828 VkPipelineCreationFeedbackEXT *pipeline_feedback,
2829 VkPipelineCreationFeedbackEXT **stage_feedbacks)
2830 {
2831 struct radv_shader_module fs_m = {0};
2832 struct radv_shader_module *modules[MESA_SHADER_STAGES] = { 0, };
2833 nir_shader *nir[MESA_SHADER_STAGES] = {0};
2834 struct radv_shader_binary *binaries[MESA_SHADER_STAGES] = {NULL};
2835 struct radv_shader_variant_key keys[MESA_SHADER_STAGES] = {{{{{0}}}}};
2836 struct radv_shader_info infos[MESA_SHADER_STAGES] = {0};
2837 unsigned char hash[20], gs_copy_hash[20];
2838 bool keep_executable_info = (flags & VK_PIPELINE_CREATE_CAPTURE_INTERNAL_REPRESENTATIONS_BIT_KHR) || device->keep_shader_info;
2839 bool keep_statistic_info = (flags & VK_PIPELINE_CREATE_CAPTURE_STATISTICS_BIT_KHR) ||
2840 (device->instance->debug_flags & RADV_DEBUG_DUMP_SHADER_STATS) ||
2841 device->keep_shader_info;
2842
2843 radv_start_feedback(pipeline_feedback);
2844
2845 for (unsigned i = 0; i < MESA_SHADER_STAGES; ++i) {
2846 if (pStages[i]) {
2847 modules[i] = radv_shader_module_from_handle(pStages[i]->module);
2848 if (modules[i]->nir)
2849 _mesa_sha1_compute(modules[i]->nir->info.name,
2850 strlen(modules[i]->nir->info.name),
2851 modules[i]->sha1);
2852
2853 pipeline->active_stages |= mesa_to_vk_shader_stage(i);
2854 }
2855 }
2856
2857 radv_hash_shaders(hash, pStages, pipeline->layout, key, get_hash_flags(device));
2858 memcpy(gs_copy_hash, hash, 20);
2859 gs_copy_hash[0] ^= 1;
2860
2861 bool found_in_application_cache = true;
2862 if (modules[MESA_SHADER_GEOMETRY] && !keep_executable_info && !keep_statistic_info) {
2863 struct radv_shader_variant *variants[MESA_SHADER_STAGES] = {0};
2864 radv_create_shader_variants_from_pipeline_cache(device, cache, gs_copy_hash, variants,
2865 &found_in_application_cache);
2866 pipeline->gs_copy_shader = variants[MESA_SHADER_GEOMETRY];
2867 }
2868
2869 if (!keep_executable_info && !keep_statistic_info &&
2870 radv_create_shader_variants_from_pipeline_cache(device, cache, hash, pipeline->shaders,
2871 &found_in_application_cache) &&
2872 (!modules[MESA_SHADER_GEOMETRY] || pipeline->gs_copy_shader)) {
2873 radv_stop_feedback(pipeline_feedback, found_in_application_cache);
2874 return VK_SUCCESS;
2875 }
2876
2877 if (flags & VK_PIPELINE_CREATE_FAIL_ON_PIPELINE_COMPILE_REQUIRED_BIT_EXT) {
2878 radv_stop_feedback(pipeline_feedback, found_in_application_cache);
2879 return VK_PIPELINE_COMPILE_REQUIRED_EXT;
2880 }
2881
2882 if (!modules[MESA_SHADER_FRAGMENT] && !modules[MESA_SHADER_COMPUTE]) {
2883 nir_builder fs_b;
2884 nir_builder_init_simple_shader(&fs_b, NULL, MESA_SHADER_FRAGMENT, NULL);
2885 fs_b.shader->info.name = ralloc_strdup(fs_b.shader, "noop_fs");
2886 fs_m.nir = fs_b.shader;
2887 modules[MESA_SHADER_FRAGMENT] = &fs_m;
2888 }
2889
2890 for (unsigned i = 0; i < MESA_SHADER_STAGES; ++i) {
2891 const VkPipelineShaderStageCreateInfo *stage = pStages[i];
2892 unsigned subgroup_size = 64, ballot_bit_size = 64;
2893
2894 if (!modules[i])
2895 continue;
2896
2897 radv_start_feedback(stage_feedbacks[i]);
2898
2899 if (key->compute_subgroup_size) {
2900 /* Only compute shaders currently support requiring a
2901 * specific subgroup size.
2902 */
2903 assert(i == MESA_SHADER_COMPUTE);
2904 subgroup_size = key->compute_subgroup_size;
2905 ballot_bit_size = key->compute_subgroup_size;
2906 }
2907
2908 nir[i] = radv_shader_compile_to_nir(device, modules[i],
2909 stage ? stage->pName : "main", i,
2910 stage ? stage->pSpecializationInfo : NULL,
2911 flags, pipeline->layout,
2912 subgroup_size, ballot_bit_size);
2913
2914 /* We don't want to alter meta shaders IR directly so clone it
2915 * first.
2916 */
2917 if (nir[i]->info.name) {
2918 nir[i] = nir_shader_clone(NULL, nir[i]);
2919 }
2920
2921 radv_stop_feedback(stage_feedbacks[i], false);
2922 }
2923
2924 if (nir[MESA_SHADER_TESS_CTRL]) {
2925 nir_lower_patch_vertices(nir[MESA_SHADER_TESS_EVAL], nir[MESA_SHADER_TESS_CTRL]->info.tess.tcs_vertices_out, NULL);
2926 merge_tess_info(&nir[MESA_SHADER_TESS_EVAL]->info, &nir[MESA_SHADER_TESS_CTRL]->info);
2927 }
2928
2929 if (!(flags & VK_PIPELINE_CREATE_DISABLE_OPTIMIZATION_BIT))
2930 radv_link_shaders(pipeline, nir);
2931
2932 radv_set_linked_driver_locations(pipeline, nir, infos);
2933
2934 for (int i = 0; i < MESA_SHADER_STAGES; ++i) {
2935 if (nir[i]) {
2936 /* do this again since information such as outputs_read can be out-of-date */
2937 nir_shader_gather_info(nir[i], nir_shader_get_entrypoint(nir[i]));
2938
2939 if (device->physical_device->use_llvm) {
2940 NIR_PASS_V(nir[i], nir_lower_bool_to_int32);
2941 } else {
2942 NIR_PASS_V(nir[i], nir_lower_non_uniform_access,
2943 nir_lower_non_uniform_ubo_access |
2944 nir_lower_non_uniform_ssbo_access |
2945 nir_lower_non_uniform_texture_access |
2946 nir_lower_non_uniform_image_access);
2947 }
2948 NIR_PASS_V(nir[i], nir_lower_memory_model);
2949 }
2950 }
2951
2952 if (nir[MESA_SHADER_FRAGMENT])
2953 radv_lower_fs_io(nir[MESA_SHADER_FRAGMENT]);
2954
2955 for (int i = 0; i < MESA_SHADER_STAGES; ++i) {
2956 if (radv_can_dump_shader(device, modules[i], false))
2957 nir_print_shader(nir[i], stderr);
2958 }
2959
2960 radv_fill_shader_keys(device, keys, key, nir);
2961
2962 radv_fill_shader_info(pipeline, pStages, keys, infos, nir);
2963
2964 if ((nir[MESA_SHADER_VERTEX] &&
2965 keys[MESA_SHADER_VERTEX].vs_common_out.as_ngg) ||
2966 (nir[MESA_SHADER_TESS_EVAL] &&
2967 keys[MESA_SHADER_TESS_EVAL].vs_common_out.as_ngg)) {
2968 struct gfx10_ngg_info *ngg_info;
2969
2970 if (nir[MESA_SHADER_GEOMETRY])
2971 ngg_info = &infos[MESA_SHADER_GEOMETRY].ngg_info;
2972 else if (nir[MESA_SHADER_TESS_CTRL])
2973 ngg_info = &infos[MESA_SHADER_TESS_EVAL].ngg_info;
2974 else
2975 ngg_info = &infos[MESA_SHADER_VERTEX].ngg_info;
2976
2977 gfx10_get_ngg_info(key, pipeline, nir, infos, ngg_info);
2978 } else if (nir[MESA_SHADER_GEOMETRY]) {
2979 struct gfx9_gs_info *gs_info =
2980 &infos[MESA_SHADER_GEOMETRY].gs_ring_info;
2981
2982 gfx9_get_gs_info(key, pipeline, nir, infos, gs_info);
2983 }
2984
2985 if(modules[MESA_SHADER_GEOMETRY]) {
2986 struct radv_shader_binary *gs_copy_binary = NULL;
2987 if (!pipeline->gs_copy_shader &&
2988 !radv_pipeline_has_ngg(pipeline)) {
2989 struct radv_shader_info info = {};
2990 struct radv_shader_variant_key key = {};
2991
2992 key.has_multiview_view_index =
2993 keys[MESA_SHADER_GEOMETRY].has_multiview_view_index;
2994
2995 radv_nir_shader_info_pass(nir[MESA_SHADER_GEOMETRY],
2996 pipeline->layout, &key,
2997 &info, pipeline->device->physical_device->use_llvm);
2998 info.wave_size = 64; /* Wave32 not supported. */
2999 info.ballot_bit_size = 64;
3000
3001 pipeline->gs_copy_shader = radv_create_gs_copy_shader(
3002 device, nir[MESA_SHADER_GEOMETRY], &info,
3003 &gs_copy_binary, keep_executable_info, keep_statistic_info,
3004 keys[MESA_SHADER_GEOMETRY].has_multiview_view_index);
3005 }
3006
3007 if (!keep_executable_info && !keep_statistic_info && pipeline->gs_copy_shader) {
3008 struct radv_shader_binary *binaries[MESA_SHADER_STAGES] = {NULL};
3009 struct radv_shader_variant *variants[MESA_SHADER_STAGES] = {0};
3010
3011 binaries[MESA_SHADER_GEOMETRY] = gs_copy_binary;
3012 variants[MESA_SHADER_GEOMETRY] = pipeline->gs_copy_shader;
3013
3014 radv_pipeline_cache_insert_shaders(device, cache,
3015 gs_copy_hash,
3016 variants,
3017 binaries);
3018 }
3019 free(gs_copy_binary);
3020 }
3021
3022 if (nir[MESA_SHADER_FRAGMENT]) {
3023 if (!pipeline->shaders[MESA_SHADER_FRAGMENT]) {
3024 radv_start_feedback(stage_feedbacks[MESA_SHADER_FRAGMENT]);
3025
3026 pipeline->shaders[MESA_SHADER_FRAGMENT] =
3027 radv_shader_variant_compile(device, modules[MESA_SHADER_FRAGMENT], &nir[MESA_SHADER_FRAGMENT], 1,
3028 pipeline->layout, keys + MESA_SHADER_FRAGMENT,
3029 infos + MESA_SHADER_FRAGMENT,
3030 keep_executable_info, keep_statistic_info,
3031 &binaries[MESA_SHADER_FRAGMENT]);
3032
3033 radv_stop_feedback(stage_feedbacks[MESA_SHADER_FRAGMENT], false);
3034 }
3035 }
3036
3037 if (device->physical_device->rad_info.chip_class >= GFX9 && modules[MESA_SHADER_TESS_CTRL]) {
3038 if (!pipeline->shaders[MESA_SHADER_TESS_CTRL]) {
3039 struct nir_shader *combined_nir[] = {nir[MESA_SHADER_VERTEX], nir[MESA_SHADER_TESS_CTRL]};
3040 struct radv_shader_variant_key key = keys[MESA_SHADER_TESS_CTRL];
3041 key.tcs.vs_key = keys[MESA_SHADER_VERTEX].vs;
3042
3043 radv_start_feedback(stage_feedbacks[MESA_SHADER_TESS_CTRL]);
3044
3045 pipeline->shaders[MESA_SHADER_TESS_CTRL] = radv_shader_variant_compile(device, modules[MESA_SHADER_TESS_CTRL], combined_nir, 2,
3046 pipeline->layout,
3047 &key, &infos[MESA_SHADER_TESS_CTRL], keep_executable_info,
3048 keep_statistic_info, &binaries[MESA_SHADER_TESS_CTRL]);
3049
3050 radv_stop_feedback(stage_feedbacks[MESA_SHADER_TESS_CTRL], false);
3051 }
3052 modules[MESA_SHADER_VERTEX] = NULL;
3053 keys[MESA_SHADER_TESS_EVAL].tes.num_patches = pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.tcs.num_patches;
3054 keys[MESA_SHADER_TESS_EVAL].tes.tcs_num_outputs = util_last_bit64(pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.tcs.outputs_written);
3055 }
3056
3057 if (device->physical_device->rad_info.chip_class >= GFX9 && modules[MESA_SHADER_GEOMETRY]) {
3058 gl_shader_stage pre_stage = modules[MESA_SHADER_TESS_EVAL] ? MESA_SHADER_TESS_EVAL : MESA_SHADER_VERTEX;
3059 if (!pipeline->shaders[MESA_SHADER_GEOMETRY]) {
3060 struct nir_shader *combined_nir[] = {nir[pre_stage], nir[MESA_SHADER_GEOMETRY]};
3061
3062 radv_start_feedback(stage_feedbacks[MESA_SHADER_GEOMETRY]);
3063
3064 pipeline->shaders[MESA_SHADER_GEOMETRY] = radv_shader_variant_compile(device, modules[MESA_SHADER_GEOMETRY], combined_nir, 2,
3065 pipeline->layout,
3066 &keys[pre_stage], &infos[MESA_SHADER_GEOMETRY], keep_executable_info,
3067 keep_statistic_info, &binaries[MESA_SHADER_GEOMETRY]);
3068
3069 radv_stop_feedback(stage_feedbacks[MESA_SHADER_GEOMETRY], false);
3070 }
3071 modules[pre_stage] = NULL;
3072 }
3073
3074 for (int i = 0; i < MESA_SHADER_STAGES; ++i) {
3075 if(modules[i] && !pipeline->shaders[i]) {
3076 if (i == MESA_SHADER_TESS_CTRL) {
3077 keys[MESA_SHADER_TESS_CTRL].tcs.num_inputs = util_last_bit64(pipeline->shaders[MESA_SHADER_VERTEX]->info.vs.ls_outputs_written);
3078 }
3079 if (i == MESA_SHADER_TESS_EVAL) {
3080 keys[MESA_SHADER_TESS_EVAL].tes.num_patches = pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.tcs.num_patches;
3081 keys[MESA_SHADER_TESS_EVAL].tes.tcs_num_outputs = util_last_bit64(pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.tcs.outputs_written);
3082 }
3083
3084 radv_start_feedback(stage_feedbacks[i]);
3085
3086 pipeline->shaders[i] = radv_shader_variant_compile(device, modules[i], &nir[i], 1,
3087 pipeline->layout,
3088 keys + i, infos + i, keep_executable_info,
3089 keep_statistic_info, &binaries[i]);
3090
3091 radv_stop_feedback(stage_feedbacks[i], false);
3092 }
3093 }
3094
3095 if (!keep_executable_info && !keep_statistic_info) {
3096 radv_pipeline_cache_insert_shaders(device, cache, hash, pipeline->shaders,
3097 binaries);
3098 }
3099
3100 for (int i = 0; i < MESA_SHADER_STAGES; ++i) {
3101 free(binaries[i]);
3102 if (nir[i]) {
3103 ralloc_free(nir[i]);
3104
3105 if (radv_can_dump_shader_stats(device, modules[i]))
3106 radv_shader_dump_stats(device,
3107 pipeline->shaders[i],
3108 i, stderr);
3109 }
3110 }
3111
3112 if (fs_m.nir)
3113 ralloc_free(fs_m.nir);
3114
3115 radv_stop_feedback(pipeline_feedback, false);
3116 return VK_SUCCESS;
3117 }
3118
3119 static uint32_t
3120 radv_pipeline_stage_to_user_data_0(struct radv_pipeline *pipeline,
3121 gl_shader_stage stage, enum chip_class chip_class)
3122 {
3123 bool has_gs = radv_pipeline_has_gs(pipeline);
3124 bool has_tess = radv_pipeline_has_tess(pipeline);
3125 bool has_ngg = radv_pipeline_has_ngg(pipeline);
3126
3127 switch (stage) {
3128 case MESA_SHADER_FRAGMENT:
3129 return R_00B030_SPI_SHADER_USER_DATA_PS_0;
3130 case MESA_SHADER_VERTEX:
3131 if (has_tess) {
3132 if (chip_class >= GFX10) {
3133 return R_00B430_SPI_SHADER_USER_DATA_HS_0;
3134 } else if (chip_class == GFX9) {
3135 return R_00B430_SPI_SHADER_USER_DATA_LS_0;
3136 } else {
3137 return R_00B530_SPI_SHADER_USER_DATA_LS_0;
3138 }
3139
3140 }
3141
3142 if (has_gs) {
3143 if (chip_class >= GFX10) {
3144 return R_00B230_SPI_SHADER_USER_DATA_GS_0;
3145 } else {
3146 return R_00B330_SPI_SHADER_USER_DATA_ES_0;
3147 }
3148 }
3149
3150 if (has_ngg)
3151 return R_00B230_SPI_SHADER_USER_DATA_GS_0;
3152
3153 return R_00B130_SPI_SHADER_USER_DATA_VS_0;
3154 case MESA_SHADER_GEOMETRY:
3155 return chip_class == GFX9 ? R_00B330_SPI_SHADER_USER_DATA_ES_0 :
3156 R_00B230_SPI_SHADER_USER_DATA_GS_0;
3157 case MESA_SHADER_COMPUTE:
3158 return R_00B900_COMPUTE_USER_DATA_0;
3159 case MESA_SHADER_TESS_CTRL:
3160 return chip_class == GFX9 ? R_00B430_SPI_SHADER_USER_DATA_LS_0 :
3161 R_00B430_SPI_SHADER_USER_DATA_HS_0;
3162 case MESA_SHADER_TESS_EVAL:
3163 if (has_gs) {
3164 return chip_class >= GFX10 ? R_00B230_SPI_SHADER_USER_DATA_GS_0 :
3165 R_00B330_SPI_SHADER_USER_DATA_ES_0;
3166 } else if (has_ngg) {
3167 return R_00B230_SPI_SHADER_USER_DATA_GS_0;
3168 } else {
3169 return R_00B130_SPI_SHADER_USER_DATA_VS_0;
3170 }
3171 default:
3172 unreachable("unknown shader");
3173 }
3174 }
3175
3176 struct radv_bin_size_entry {
3177 unsigned bpp;
3178 VkExtent2D extent;
3179 };
3180
3181 static VkExtent2D
3182 radv_gfx9_compute_bin_size(struct radv_pipeline *pipeline, const VkGraphicsPipelineCreateInfo *pCreateInfo)
3183 {
3184 static const struct radv_bin_size_entry color_size_table[][3][9] = {
3185 {
3186 /* One RB / SE */
3187 {
3188 /* One shader engine */
3189 { 0, {128, 128}},
3190 { 1, { 64, 128}},
3191 { 2, { 32, 128}},
3192 { 3, { 16, 128}},
3193 { 17, { 0, 0}},
3194 { UINT_MAX, { 0, 0}},
3195 },
3196 {
3197 /* Two shader engines */
3198 { 0, {128, 128}},
3199 { 2, { 64, 128}},
3200 { 3, { 32, 128}},
3201 { 5, { 16, 128}},
3202 { 17, { 0, 0}},
3203 { UINT_MAX, { 0, 0}},
3204 },
3205 {
3206 /* Four shader engines */
3207 { 0, {128, 128}},
3208 { 3, { 64, 128}},
3209 { 5, { 16, 128}},
3210 { 17, { 0, 0}},
3211 { UINT_MAX, { 0, 0}},
3212 },
3213 },
3214 {
3215 /* Two RB / SE */
3216 {
3217 /* One shader engine */
3218 { 0, {128, 128}},
3219 { 2, { 64, 128}},
3220 { 3, { 32, 128}},
3221 { 5, { 16, 128}},
3222 { 33, { 0, 0}},
3223 { UINT_MAX, { 0, 0}},
3224 },
3225 {
3226 /* Two shader engines */
3227 { 0, {128, 128}},
3228 { 3, { 64, 128}},
3229 { 5, { 32, 128}},
3230 { 9, { 16, 128}},
3231 { 33, { 0, 0}},
3232 { UINT_MAX, { 0, 0}},
3233 },
3234 {
3235 /* Four shader engines */
3236 { 0, {256, 256}},
3237 { 2, {128, 256}},
3238 { 3, {128, 128}},
3239 { 5, { 64, 128}},
3240 { 9, { 16, 128}},
3241 { 33, { 0, 0}},
3242 { UINT_MAX, { 0, 0}},
3243 },
3244 },
3245 {
3246 /* Four RB / SE */
3247 {
3248 /* One shader engine */
3249 { 0, {128, 256}},
3250 { 2, {128, 128}},
3251 { 3, { 64, 128}},
3252 { 5, { 32, 128}},
3253 { 9, { 16, 128}},
3254 { 33, { 0, 0}},
3255 { UINT_MAX, { 0, 0}},
3256 },
3257 {
3258 /* Two shader engines */
3259 { 0, {256, 256}},
3260 { 2, {128, 256}},
3261 { 3, {128, 128}},
3262 { 5, { 64, 128}},
3263 { 9, { 32, 128}},
3264 { 17, { 16, 128}},
3265 { 33, { 0, 0}},
3266 { UINT_MAX, { 0, 0}},
3267 },
3268 {
3269 /* Four shader engines */
3270 { 0, {256, 512}},
3271 { 2, {256, 256}},
3272 { 3, {128, 256}},
3273 { 5, {128, 128}},
3274 { 9, { 64, 128}},
3275 { 17, { 16, 128}},
3276 { 33, { 0, 0}},
3277 { UINT_MAX, { 0, 0}},
3278 },
3279 },
3280 };
3281 static const struct radv_bin_size_entry ds_size_table[][3][9] = {
3282 {
3283 // One RB / SE
3284 {
3285 // One shader engine
3286 { 0, {128, 256}},
3287 { 2, {128, 128}},
3288 { 4, { 64, 128}},
3289 { 7, { 32, 128}},
3290 { 13, { 16, 128}},
3291 { 49, { 0, 0}},
3292 { UINT_MAX, { 0, 0}},
3293 },
3294 {
3295 // Two shader engines
3296 { 0, {256, 256}},
3297 { 2, {128, 256}},
3298 { 4, {128, 128}},
3299 { 7, { 64, 128}},
3300 { 13, { 32, 128}},
3301 { 25, { 16, 128}},
3302 { 49, { 0, 0}},
3303 { UINT_MAX, { 0, 0}},
3304 },
3305 {
3306 // Four shader engines
3307 { 0, {256, 512}},
3308 { 2, {256, 256}},
3309 { 4, {128, 256}},
3310 { 7, {128, 128}},
3311 { 13, { 64, 128}},
3312 { 25, { 16, 128}},
3313 { 49, { 0, 0}},
3314 { UINT_MAX, { 0, 0}},
3315 },
3316 },
3317 {
3318 // Two RB / SE
3319 {
3320 // One shader engine
3321 { 0, {256, 256}},
3322 { 2, {128, 256}},
3323 { 4, {128, 128}},
3324 { 7, { 64, 128}},
3325 { 13, { 32, 128}},
3326 { 25, { 16, 128}},
3327 { 97, { 0, 0}},
3328 { UINT_MAX, { 0, 0}},
3329 },
3330 {
3331 // Two shader engines
3332 { 0, {256, 512}},
3333 { 2, {256, 256}},
3334 { 4, {128, 256}},
3335 { 7, {128, 128}},
3336 { 13, { 64, 128}},
3337 { 25, { 32, 128}},
3338 { 49, { 16, 128}},
3339 { 97, { 0, 0}},
3340 { UINT_MAX, { 0, 0}},
3341 },
3342 {
3343 // Four shader engines
3344 { 0, {512, 512}},
3345 { 2, {256, 512}},
3346 { 4, {256, 256}},
3347 { 7, {128, 256}},
3348 { 13, {128, 128}},
3349 { 25, { 64, 128}},
3350 { 49, { 16, 128}},
3351 { 97, { 0, 0}},
3352 { UINT_MAX, { 0, 0}},
3353 },
3354 },
3355 {
3356 // Four RB / SE
3357 {
3358 // One shader engine
3359 { 0, {256, 512}},
3360 { 2, {256, 256}},
3361 { 4, {128, 256}},
3362 { 7, {128, 128}},
3363 { 13, { 64, 128}},
3364 { 25, { 32, 128}},
3365 { 49, { 16, 128}},
3366 { UINT_MAX, { 0, 0}},
3367 },
3368 {
3369 // Two shader engines
3370 { 0, {512, 512}},
3371 { 2, {256, 512}},
3372 { 4, {256, 256}},
3373 { 7, {128, 256}},
3374 { 13, {128, 128}},
3375 { 25, { 64, 128}},
3376 { 49, { 32, 128}},
3377 { 97, { 16, 128}},
3378 { UINT_MAX, { 0, 0}},
3379 },
3380 {
3381 // Four shader engines
3382 { 0, {512, 512}},
3383 { 4, {256, 512}},
3384 { 7, {256, 256}},
3385 { 13, {128, 256}},
3386 { 25, {128, 128}},
3387 { 49, { 64, 128}},
3388 { 97, { 16, 128}},
3389 { UINT_MAX, { 0, 0}},
3390 },
3391 },
3392 };
3393
3394 RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass);
3395 struct radv_subpass *subpass = pass->subpasses + pCreateInfo->subpass;
3396 VkExtent2D extent = {512, 512};
3397
3398 unsigned log_num_rb_per_se =
3399 util_logbase2_ceil(pipeline->device->physical_device->rad_info.num_render_backends /
3400 pipeline->device->physical_device->rad_info.max_se);
3401 unsigned log_num_se = util_logbase2_ceil(pipeline->device->physical_device->rad_info.max_se);
3402
3403 unsigned total_samples = 1u << G_028BE0_MSAA_NUM_SAMPLES(pipeline->graphics.ms.pa_sc_aa_config);
3404 unsigned ps_iter_samples = 1u << G_028804_PS_ITER_SAMPLES(pipeline->graphics.ms.db_eqaa);
3405 unsigned effective_samples = total_samples;
3406 unsigned color_bytes_per_pixel = 0;
3407
3408 const VkPipelineColorBlendStateCreateInfo *vkblend =
3409 radv_pipeline_get_color_blend_state(pCreateInfo);
3410 if (vkblend) {
3411 for (unsigned i = 0; i < subpass->color_count; i++) {
3412 if (!vkblend->pAttachments[i].colorWriteMask)
3413 continue;
3414
3415 if (subpass->color_attachments[i].attachment == VK_ATTACHMENT_UNUSED)
3416 continue;
3417
3418 VkFormat format = pass->attachments[subpass->color_attachments[i].attachment].format;
3419 color_bytes_per_pixel += vk_format_get_blocksize(format);
3420 }
3421
3422 /* MSAA images typically don't use all samples all the time. */
3423 if (effective_samples >= 2 && ps_iter_samples <= 1)
3424 effective_samples = 2;
3425 color_bytes_per_pixel *= effective_samples;
3426 }
3427
3428 const struct radv_bin_size_entry *color_entry = color_size_table[log_num_rb_per_se][log_num_se];
3429 while(color_entry[1].bpp <= color_bytes_per_pixel)
3430 ++color_entry;
3431
3432 extent = color_entry->extent;
3433
3434 if (subpass->depth_stencil_attachment) {
3435 struct radv_render_pass_attachment *attachment = pass->attachments + subpass->depth_stencil_attachment->attachment;
3436
3437 /* Coefficients taken from AMDVLK */
3438 unsigned depth_coeff = vk_format_is_depth(attachment->format) ? 5 : 0;
3439 unsigned stencil_coeff = vk_format_is_stencil(attachment->format) ? 1 : 0;
3440 unsigned ds_bytes_per_pixel = 4 * (depth_coeff + stencil_coeff) * total_samples;
3441
3442 const struct radv_bin_size_entry *ds_entry = ds_size_table[log_num_rb_per_se][log_num_se];
3443 while(ds_entry[1].bpp <= ds_bytes_per_pixel)
3444 ++ds_entry;
3445
3446 if (ds_entry->extent.width * ds_entry->extent.height < extent.width * extent.height)
3447 extent = ds_entry->extent;
3448 }
3449
3450 return extent;
3451 }
3452
3453 static VkExtent2D
3454 radv_gfx10_compute_bin_size(struct radv_pipeline *pipeline, const VkGraphicsPipelineCreateInfo *pCreateInfo)
3455 {
3456 RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass);
3457 struct radv_subpass *subpass = pass->subpasses + pCreateInfo->subpass;
3458 VkExtent2D extent = {512, 512};
3459
3460 const unsigned db_tag_size = 64;
3461 const unsigned db_tag_count = 312;
3462 const unsigned color_tag_size = 1024;
3463 const unsigned color_tag_count = 31;
3464 const unsigned fmask_tag_size = 256;
3465 const unsigned fmask_tag_count = 44;
3466
3467 const unsigned rb_count = pipeline->device->physical_device->rad_info.num_render_backends;
3468 const unsigned pipe_count = MAX2(rb_count, pipeline->device->physical_device->rad_info.num_sdp_interfaces);
3469
3470 const unsigned db_tag_part = (db_tag_count * rb_count / pipe_count) * db_tag_size * pipe_count;
3471 const unsigned color_tag_part = (color_tag_count * rb_count / pipe_count) * color_tag_size * pipe_count;
3472 const unsigned fmask_tag_part = (fmask_tag_count * rb_count / pipe_count) * fmask_tag_size * pipe_count;
3473
3474 const unsigned total_samples = 1u << G_028BE0_MSAA_NUM_SAMPLES(pipeline->graphics.ms.pa_sc_aa_config);
3475 const unsigned samples_log = util_logbase2_ceil(total_samples);
3476
3477 unsigned color_bytes_per_pixel = 0;
3478 unsigned fmask_bytes_per_pixel = 0;
3479
3480 const VkPipelineColorBlendStateCreateInfo *vkblend =
3481 radv_pipeline_get_color_blend_state(pCreateInfo);
3482 if (vkblend) {
3483 for (unsigned i = 0; i < subpass->color_count; i++) {
3484 if (!vkblend->pAttachments[i].colorWriteMask)
3485 continue;
3486
3487 if (subpass->color_attachments[i].attachment == VK_ATTACHMENT_UNUSED)
3488 continue;
3489
3490 VkFormat format = pass->attachments[subpass->color_attachments[i].attachment].format;
3491 color_bytes_per_pixel += vk_format_get_blocksize(format);
3492
3493 if (total_samples > 1) {
3494 assert(samples_log <= 3);
3495 const unsigned fmask_array[] = {0, 1, 1, 4};
3496 fmask_bytes_per_pixel += fmask_array[samples_log];
3497 }
3498 }
3499
3500 color_bytes_per_pixel *= total_samples;
3501 }
3502 color_bytes_per_pixel = MAX2(color_bytes_per_pixel, 1);
3503
3504 const unsigned color_pixel_count_log = util_logbase2(color_tag_part / color_bytes_per_pixel);
3505 extent.width = 1ull << ((color_pixel_count_log + 1) / 2);
3506 extent.height = 1ull << (color_pixel_count_log / 2);
3507
3508 if (fmask_bytes_per_pixel) {
3509 const unsigned fmask_pixel_count_log = util_logbase2(fmask_tag_part / fmask_bytes_per_pixel);
3510
3511 const VkExtent2D fmask_extent = (VkExtent2D){
3512 .width = 1ull << ((fmask_pixel_count_log + 1) / 2),
3513 .height = 1ull << (color_pixel_count_log / 2)
3514 };
3515
3516 if (fmask_extent.width * fmask_extent.height < extent.width * extent.height)
3517 extent = fmask_extent;
3518 }
3519
3520 if (subpass->depth_stencil_attachment) {
3521 struct radv_render_pass_attachment *attachment = pass->attachments + subpass->depth_stencil_attachment->attachment;
3522
3523 /* Coefficients taken from AMDVLK */
3524 unsigned depth_coeff = vk_format_is_depth(attachment->format) ? 5 : 0;
3525 unsigned stencil_coeff = vk_format_is_stencil(attachment->format) ? 1 : 0;
3526 unsigned db_bytes_per_pixel = (depth_coeff + stencil_coeff) * total_samples;
3527
3528 const unsigned db_pixel_count_log = util_logbase2(db_tag_part / db_bytes_per_pixel);
3529
3530 const VkExtent2D db_extent = (VkExtent2D){
3531 .width = 1ull << ((db_pixel_count_log + 1) / 2),
3532 .height = 1ull << (color_pixel_count_log / 2)
3533 };
3534
3535 if (db_extent.width * db_extent.height < extent.width * extent.height)
3536 extent = db_extent;
3537 }
3538
3539 extent.width = MAX2(extent.width, 128);
3540 extent.height = MAX2(extent.width, 64);
3541
3542 return extent;
3543 }
3544
3545 static void
3546 radv_pipeline_init_disabled_binning_state(struct radv_pipeline *pipeline,
3547 const VkGraphicsPipelineCreateInfo *pCreateInfo)
3548 {
3549 uint32_t pa_sc_binner_cntl_0 =
3550 S_028C44_BINNING_MODE(V_028C44_DISABLE_BINNING_USE_LEGACY_SC) |
3551 S_028C44_DISABLE_START_OF_PRIM(1);
3552 uint32_t db_dfsm_control = S_028060_PUNCHOUT_MODE(V_028060_FORCE_OFF);
3553
3554 if (pipeline->device->physical_device->rad_info.chip_class >= GFX10) {
3555 RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass);
3556 struct radv_subpass *subpass = pass->subpasses + pCreateInfo->subpass;
3557 const VkPipelineColorBlendStateCreateInfo *vkblend =
3558 radv_pipeline_get_color_blend_state(pCreateInfo);
3559 unsigned min_bytes_per_pixel = 0;
3560
3561 if (vkblend) {
3562 for (unsigned i = 0; i < subpass->color_count; i++) {
3563 if (!vkblend->pAttachments[i].colorWriteMask)
3564 continue;
3565
3566 if (subpass->color_attachments[i].attachment == VK_ATTACHMENT_UNUSED)
3567 continue;
3568
3569 VkFormat format = pass->attachments[subpass->color_attachments[i].attachment].format;
3570 unsigned bytes = vk_format_get_blocksize(format);
3571 if (!min_bytes_per_pixel || bytes < min_bytes_per_pixel)
3572 min_bytes_per_pixel = bytes;
3573 }
3574 }
3575
3576 pa_sc_binner_cntl_0 =
3577 S_028C44_BINNING_MODE(V_028C44_DISABLE_BINNING_USE_NEW_SC) |
3578 S_028C44_BIN_SIZE_X(0) |
3579 S_028C44_BIN_SIZE_Y(0) |
3580 S_028C44_BIN_SIZE_X_EXTEND(2) | /* 128 */
3581 S_028C44_BIN_SIZE_Y_EXTEND(min_bytes_per_pixel <= 4 ? 2 : 1) | /* 128 or 64 */
3582 S_028C44_DISABLE_START_OF_PRIM(1);
3583 }
3584
3585 pipeline->graphics.binning.pa_sc_binner_cntl_0 = pa_sc_binner_cntl_0;
3586 pipeline->graphics.binning.db_dfsm_control = db_dfsm_control;
3587 }
3588
3589 struct radv_binning_settings
3590 radv_get_binning_settings(const struct radv_physical_device *pdev)
3591 {
3592 struct radv_binning_settings settings;
3593 if (pdev->rad_info.has_dedicated_vram) {
3594 if (pdev->rad_info.num_render_backends > 4) {
3595 settings.context_states_per_bin = 1;
3596 settings.persistent_states_per_bin = 1;
3597 } else {
3598 settings.context_states_per_bin = 3;
3599 settings.persistent_states_per_bin = 8;
3600 }
3601 settings.fpovs_per_batch = 63;
3602 } else {
3603 /* The context states are affected by the scissor bug. */
3604 settings.context_states_per_bin = 6;
3605 /* 32 causes hangs for RAVEN. */
3606 settings.persistent_states_per_bin = 16;
3607 settings.fpovs_per_batch = 63;
3608 }
3609
3610 if (pdev->rad_info.has_gfx9_scissor_bug)
3611 settings.context_states_per_bin = 1;
3612
3613 return settings;
3614 }
3615
3616 static void
3617 radv_pipeline_init_binning_state(struct radv_pipeline *pipeline,
3618 const VkGraphicsPipelineCreateInfo *pCreateInfo,
3619 const struct radv_blend_state *blend)
3620 {
3621 if (pipeline->device->physical_device->rad_info.chip_class < GFX9)
3622 return;
3623
3624 VkExtent2D bin_size;
3625 if (pipeline->device->physical_device->rad_info.chip_class >= GFX10) {
3626 bin_size = radv_gfx10_compute_bin_size(pipeline, pCreateInfo);
3627 } else if (pipeline->device->physical_device->rad_info.chip_class == GFX9) {
3628 bin_size = radv_gfx9_compute_bin_size(pipeline, pCreateInfo);
3629 } else
3630 unreachable("Unhandled generation for binning bin size calculation");
3631
3632 if (pipeline->device->pbb_allowed && bin_size.width && bin_size.height) {
3633 struct radv_binning_settings settings =
3634 radv_get_binning_settings(pipeline->device->physical_device);
3635
3636 bool disable_start_of_prim = true;
3637 uint32_t db_dfsm_control = S_028060_PUNCHOUT_MODE(V_028060_FORCE_OFF);
3638
3639 const struct radv_shader_variant *ps = pipeline->shaders[MESA_SHADER_FRAGMENT];
3640
3641 if (pipeline->device->dfsm_allowed && ps &&
3642 !ps->info.ps.can_discard &&
3643 !ps->info.ps.writes_memory &&
3644 blend->cb_target_enabled_4bit) {
3645 db_dfsm_control = S_028060_PUNCHOUT_MODE(V_028060_AUTO);
3646 disable_start_of_prim = (blend->blend_enable_4bit & blend->cb_target_enabled_4bit) != 0;
3647 }
3648
3649 const uint32_t pa_sc_binner_cntl_0 =
3650 S_028C44_BINNING_MODE(V_028C44_BINNING_ALLOWED) |
3651 S_028C44_BIN_SIZE_X(bin_size.width == 16) |
3652 S_028C44_BIN_SIZE_Y(bin_size.height == 16) |
3653 S_028C44_BIN_SIZE_X_EXTEND(util_logbase2(MAX2(bin_size.width, 32)) - 5) |
3654 S_028C44_BIN_SIZE_Y_EXTEND(util_logbase2(MAX2(bin_size.height, 32)) - 5) |
3655 S_028C44_CONTEXT_STATES_PER_BIN(settings.context_states_per_bin - 1) |
3656 S_028C44_PERSISTENT_STATES_PER_BIN(settings.persistent_states_per_bin - 1) |
3657 S_028C44_DISABLE_START_OF_PRIM(disable_start_of_prim) |
3658 S_028C44_FPOVS_PER_BATCH(settings.fpovs_per_batch) |
3659 S_028C44_OPTIMAL_BIN_SELECTION(1);
3660
3661 pipeline->graphics.binning.pa_sc_binner_cntl_0 = pa_sc_binner_cntl_0;
3662 pipeline->graphics.binning.db_dfsm_control = db_dfsm_control;
3663 } else
3664 radv_pipeline_init_disabled_binning_state(pipeline, pCreateInfo);
3665 }
3666
3667
3668 static void
3669 radv_pipeline_generate_depth_stencil_state(struct radeon_cmdbuf *ctx_cs,
3670 const struct radv_pipeline *pipeline,
3671 const VkGraphicsPipelineCreateInfo *pCreateInfo,
3672 const struct radv_graphics_pipeline_create_info *extra)
3673 {
3674 const VkPipelineDepthStencilStateCreateInfo *vkds = radv_pipeline_get_depth_stencil_state(pCreateInfo);
3675 RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass);
3676 struct radv_subpass *subpass = pass->subpasses + pCreateInfo->subpass;
3677 struct radv_shader_variant *ps = pipeline->shaders[MESA_SHADER_FRAGMENT];
3678 struct radv_render_pass_attachment *attachment = NULL;
3679 uint32_t db_render_control = 0, db_render_override2 = 0;
3680 uint32_t db_render_override = 0;
3681
3682 if (subpass->depth_stencil_attachment)
3683 attachment = pass->attachments + subpass->depth_stencil_attachment->attachment;
3684
3685 bool has_depth_attachment = attachment && vk_format_is_depth(attachment->format);
3686
3687 if (vkds && has_depth_attachment) {
3688 /* from amdvlk: For 4xAA and 8xAA need to decompress on flush for better performance */
3689 db_render_override2 |= S_028010_DECOMPRESS_Z_ON_FLUSH(attachment->samples > 2);
3690
3691 if (pipeline->device->physical_device->rad_info.chip_class >= GFX10_3)
3692 db_render_override2 |= S_028010_CENTROID_COMPUTATION_MODE_GFX103(2);
3693 }
3694
3695 if (attachment && extra) {
3696 db_render_control |= S_028000_DEPTH_CLEAR_ENABLE(extra->db_depth_clear);
3697 db_render_control |= S_028000_STENCIL_CLEAR_ENABLE(extra->db_stencil_clear);
3698
3699 db_render_control |= S_028000_RESUMMARIZE_ENABLE(extra->resummarize_enable);
3700 db_render_control |= S_028000_DEPTH_COMPRESS_DISABLE(extra->depth_compress_disable);
3701 db_render_control |= S_028000_STENCIL_COMPRESS_DISABLE(extra->stencil_compress_disable);
3702 db_render_override2 |= S_028010_DISABLE_ZMASK_EXPCLEAR_OPTIMIZATION(extra->db_depth_disable_expclear);
3703 db_render_override2 |= S_028010_DISABLE_SMEM_EXPCLEAR_OPTIMIZATION(extra->db_stencil_disable_expclear);
3704 }
3705
3706 db_render_override |= S_02800C_FORCE_HIS_ENABLE0(V_02800C_FORCE_DISABLE) |
3707 S_02800C_FORCE_HIS_ENABLE1(V_02800C_FORCE_DISABLE);
3708
3709 if (!pCreateInfo->pRasterizationState->depthClampEnable &&
3710 ps->info.ps.writes_z) {
3711 /* From VK_EXT_depth_range_unrestricted spec:
3712 *
3713 * "The behavior described in Primitive Clipping still applies.
3714 * If depth clamping is disabled the depth values are still
3715 * clipped to 0 ≤ zc ≤ wc before the viewport transform. If
3716 * depth clamping is enabled the above equation is ignored and
3717 * the depth values are instead clamped to the VkViewport
3718 * minDepth and maxDepth values, which in the case of this
3719 * extension can be outside of the 0.0 to 1.0 range."
3720 */
3721 db_render_override |= S_02800C_DISABLE_VIEWPORT_CLAMP(1);
3722 }
3723
3724 radeon_set_context_reg(ctx_cs, R_028000_DB_RENDER_CONTROL, db_render_control);
3725 radeon_set_context_reg(ctx_cs, R_02800C_DB_RENDER_OVERRIDE, db_render_override);
3726 radeon_set_context_reg(ctx_cs, R_028010_DB_RENDER_OVERRIDE2, db_render_override2);
3727 }
3728
3729 static void
3730 radv_pipeline_generate_blend_state(struct radeon_cmdbuf *ctx_cs,
3731 const struct radv_pipeline *pipeline,
3732 const struct radv_blend_state *blend)
3733 {
3734 radeon_set_context_reg_seq(ctx_cs, R_028780_CB_BLEND0_CONTROL, 8);
3735 radeon_emit_array(ctx_cs, blend->cb_blend_control,
3736 8);
3737 radeon_set_context_reg(ctx_cs, R_028808_CB_COLOR_CONTROL, blend->cb_color_control);
3738 radeon_set_context_reg(ctx_cs, R_028B70_DB_ALPHA_TO_MASK, blend->db_alpha_to_mask);
3739
3740 if (pipeline->device->physical_device->rad_info.has_rbplus) {
3741
3742 radeon_set_context_reg_seq(ctx_cs, R_028760_SX_MRT0_BLEND_OPT, 8);
3743 radeon_emit_array(ctx_cs, blend->sx_mrt_blend_opt, 8);
3744 }
3745
3746 radeon_set_context_reg(ctx_cs, R_028714_SPI_SHADER_COL_FORMAT, blend->spi_shader_col_format);
3747
3748 radeon_set_context_reg(ctx_cs, R_028238_CB_TARGET_MASK, blend->cb_target_mask);
3749 radeon_set_context_reg(ctx_cs, R_02823C_CB_SHADER_MASK, blend->cb_shader_mask);
3750 }
3751
3752 static void
3753 radv_pipeline_generate_raster_state(struct radeon_cmdbuf *ctx_cs,
3754 const struct radv_pipeline *pipeline,
3755 const VkGraphicsPipelineCreateInfo *pCreateInfo)
3756 {
3757 const VkPipelineRasterizationStateCreateInfo *vkraster = pCreateInfo->pRasterizationState;
3758 const VkConservativeRasterizationModeEXT mode =
3759 radv_get_conservative_raster_mode(vkraster);
3760 uint32_t pa_sc_conservative_rast = S_028C4C_NULL_SQUAD_AA_MASK_ENABLE(1);
3761 bool depth_clip_disable = vkraster->depthClampEnable;
3762
3763 const VkPipelineRasterizationDepthClipStateCreateInfoEXT *depth_clip_state =
3764 vk_find_struct_const(vkraster->pNext, PIPELINE_RASTERIZATION_DEPTH_CLIP_STATE_CREATE_INFO_EXT);
3765 if (depth_clip_state) {
3766 depth_clip_disable = !depth_clip_state->depthClipEnable;
3767 }
3768
3769 radeon_set_context_reg(ctx_cs, R_028810_PA_CL_CLIP_CNTL,
3770 S_028810_DX_CLIP_SPACE_DEF(1) | // vulkan uses DX conventions.
3771 S_028810_ZCLIP_NEAR_DISABLE(depth_clip_disable ? 1 : 0) |
3772 S_028810_ZCLIP_FAR_DISABLE(depth_clip_disable ? 1 : 0) |
3773 S_028810_DX_RASTERIZATION_KILL(vkraster->rasterizerDiscardEnable ? 1 : 0) |
3774 S_028810_DX_LINEAR_ATTR_CLIP_ENA(1));
3775
3776 radeon_set_context_reg(ctx_cs, R_028BDC_PA_SC_LINE_CNTL,
3777 S_028BDC_DX10_DIAMOND_TEST_ENA(1));
3778
3779 /* Conservative rasterization. */
3780 if (mode != VK_CONSERVATIVE_RASTERIZATION_MODE_DISABLED_EXT) {
3781 pa_sc_conservative_rast = S_028C4C_PREZ_AA_MASK_ENABLE(1) |
3782 S_028C4C_POSTZ_AA_MASK_ENABLE(1) |
3783 S_028C4C_CENTROID_SAMPLE_OVERRIDE(1);
3784
3785 if (mode == VK_CONSERVATIVE_RASTERIZATION_MODE_OVERESTIMATE_EXT) {
3786 pa_sc_conservative_rast |=
3787 S_028C4C_OVER_RAST_ENABLE(1) |
3788 S_028C4C_OVER_RAST_SAMPLE_SELECT(0) |
3789 S_028C4C_UNDER_RAST_ENABLE(0) |
3790 S_028C4C_UNDER_RAST_SAMPLE_SELECT(1) |
3791 S_028C4C_PBB_UNCERTAINTY_REGION_ENABLE(1);
3792 } else {
3793 assert(mode == VK_CONSERVATIVE_RASTERIZATION_MODE_UNDERESTIMATE_EXT);
3794 pa_sc_conservative_rast |=
3795 S_028C4C_OVER_RAST_ENABLE(0) |
3796 S_028C4C_OVER_RAST_SAMPLE_SELECT(1) |
3797 S_028C4C_UNDER_RAST_ENABLE(1) |
3798 S_028C4C_UNDER_RAST_SAMPLE_SELECT(0) |
3799 S_028C4C_PBB_UNCERTAINTY_REGION_ENABLE(0);
3800 }
3801 }
3802
3803 radeon_set_context_reg(ctx_cs, R_028C4C_PA_SC_CONSERVATIVE_RASTERIZATION_CNTL,
3804 pa_sc_conservative_rast);
3805 }
3806
3807
3808 static void
3809 radv_pipeline_generate_multisample_state(struct radeon_cmdbuf *ctx_cs,
3810 const struct radv_pipeline *pipeline)
3811 {
3812 const struct radv_multisample_state *ms = &pipeline->graphics.ms;
3813
3814 radeon_set_context_reg_seq(ctx_cs, R_028C38_PA_SC_AA_MASK_X0Y0_X1Y0, 2);
3815 radeon_emit(ctx_cs, ms->pa_sc_aa_mask[0]);
3816 radeon_emit(ctx_cs, ms->pa_sc_aa_mask[1]);
3817
3818 radeon_set_context_reg(ctx_cs, R_028804_DB_EQAA, ms->db_eqaa);
3819 radeon_set_context_reg(ctx_cs, R_028A48_PA_SC_MODE_CNTL_0, ms->pa_sc_mode_cntl_0);
3820 radeon_set_context_reg(ctx_cs, R_028A4C_PA_SC_MODE_CNTL_1, ms->pa_sc_mode_cntl_1);
3821 radeon_set_context_reg(ctx_cs, R_028BE0_PA_SC_AA_CONFIG, ms->pa_sc_aa_config);
3822
3823 /* The exclusion bits can be set to improve rasterization efficiency
3824 * if no sample lies on the pixel boundary (-8 sample offset). It's
3825 * currently always TRUE because the driver doesn't support 16 samples.
3826 */
3827 bool exclusion = pipeline->device->physical_device->rad_info.chip_class >= GFX7;
3828 radeon_set_context_reg(ctx_cs, R_02882C_PA_SU_PRIM_FILTER_CNTL,
3829 S_02882C_XMAX_RIGHT_EXCLUSION(exclusion) |
3830 S_02882C_YMAX_BOTTOM_EXCLUSION(exclusion));
3831
3832 /* GFX9: Flush DFSM when the AA mode changes. */
3833 if (pipeline->device->dfsm_allowed) {
3834 radeon_emit(ctx_cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
3835 radeon_emit(ctx_cs, EVENT_TYPE(V_028A90_FLUSH_DFSM) | EVENT_INDEX(0));
3836 }
3837 }
3838
3839 static void
3840 radv_pipeline_generate_vgt_gs_mode(struct radeon_cmdbuf *ctx_cs,
3841 const struct radv_pipeline *pipeline)
3842 {
3843 const struct radv_vs_output_info *outinfo = get_vs_output_info(pipeline);
3844 const struct radv_shader_variant *vs =
3845 pipeline->shaders[MESA_SHADER_TESS_EVAL] ?
3846 pipeline->shaders[MESA_SHADER_TESS_EVAL] :
3847 pipeline->shaders[MESA_SHADER_VERTEX];
3848 unsigned vgt_primitiveid_en = 0;
3849 uint32_t vgt_gs_mode = 0;
3850
3851 if (radv_pipeline_has_ngg(pipeline))
3852 return;
3853
3854 if (radv_pipeline_has_gs(pipeline)) {
3855 const struct radv_shader_variant *gs =
3856 pipeline->shaders[MESA_SHADER_GEOMETRY];
3857
3858 vgt_gs_mode = ac_vgt_gs_mode(gs->info.gs.vertices_out,
3859 pipeline->device->physical_device->rad_info.chip_class);
3860 } else if (outinfo->export_prim_id || vs->info.uses_prim_id) {
3861 vgt_gs_mode = S_028A40_MODE(V_028A40_GS_SCENARIO_A);
3862 vgt_primitiveid_en |= S_028A84_PRIMITIVEID_EN(1);
3863 }
3864
3865 radeon_set_context_reg(ctx_cs, R_028A84_VGT_PRIMITIVEID_EN, vgt_primitiveid_en);
3866 radeon_set_context_reg(ctx_cs, R_028A40_VGT_GS_MODE, vgt_gs_mode);
3867 }
3868
3869 static void
3870 radv_pipeline_generate_hw_vs(struct radeon_cmdbuf *ctx_cs,
3871 struct radeon_cmdbuf *cs,
3872 const struct radv_pipeline *pipeline,
3873 const struct radv_shader_variant *shader)
3874 {
3875 uint64_t va = radv_buffer_get_va(shader->bo) + shader->bo_offset;
3876
3877 radeon_set_sh_reg_seq(cs, R_00B120_SPI_SHADER_PGM_LO_VS, 4);
3878 radeon_emit(cs, va >> 8);
3879 radeon_emit(cs, S_00B124_MEM_BASE(va >> 40));
3880 radeon_emit(cs, shader->config.rsrc1);
3881 radeon_emit(cs, shader->config.rsrc2);
3882
3883 const struct radv_vs_output_info *outinfo = get_vs_output_info(pipeline);
3884 unsigned clip_dist_mask, cull_dist_mask, total_mask;
3885 clip_dist_mask = outinfo->clip_dist_mask;
3886 cull_dist_mask = outinfo->cull_dist_mask;
3887 total_mask = clip_dist_mask | cull_dist_mask;
3888 bool misc_vec_ena = outinfo->writes_pointsize ||
3889 outinfo->writes_layer ||
3890 outinfo->writes_viewport_index;
3891 unsigned spi_vs_out_config, nparams;
3892
3893 /* VS is required to export at least one param. */
3894 nparams = MAX2(outinfo->param_exports, 1);
3895 spi_vs_out_config = S_0286C4_VS_EXPORT_COUNT(nparams - 1);
3896
3897 if (pipeline->device->physical_device->rad_info.chip_class >= GFX10) {
3898 spi_vs_out_config |= S_0286C4_NO_PC_EXPORT(outinfo->param_exports == 0);
3899 }
3900
3901 radeon_set_context_reg(ctx_cs, R_0286C4_SPI_VS_OUT_CONFIG, spi_vs_out_config);
3902
3903 radeon_set_context_reg(ctx_cs, R_02870C_SPI_SHADER_POS_FORMAT,
3904 S_02870C_POS0_EXPORT_FORMAT(V_02870C_SPI_SHADER_4COMP) |
3905 S_02870C_POS1_EXPORT_FORMAT(outinfo->pos_exports > 1 ?
3906 V_02870C_SPI_SHADER_4COMP :
3907 V_02870C_SPI_SHADER_NONE) |
3908 S_02870C_POS2_EXPORT_FORMAT(outinfo->pos_exports > 2 ?
3909 V_02870C_SPI_SHADER_4COMP :
3910 V_02870C_SPI_SHADER_NONE) |
3911 S_02870C_POS3_EXPORT_FORMAT(outinfo->pos_exports > 3 ?
3912 V_02870C_SPI_SHADER_4COMP :
3913 V_02870C_SPI_SHADER_NONE));
3914
3915 radeon_set_context_reg(ctx_cs, R_02881C_PA_CL_VS_OUT_CNTL,
3916 S_02881C_USE_VTX_POINT_SIZE(outinfo->writes_pointsize) |
3917 S_02881C_USE_VTX_RENDER_TARGET_INDX(outinfo->writes_layer) |
3918 S_02881C_USE_VTX_VIEWPORT_INDX(outinfo->writes_viewport_index) |
3919 S_02881C_VS_OUT_MISC_VEC_ENA(misc_vec_ena) |
3920 S_02881C_VS_OUT_MISC_SIDE_BUS_ENA(misc_vec_ena) |
3921 S_02881C_VS_OUT_CCDIST0_VEC_ENA((total_mask & 0x0f) != 0) |
3922 S_02881C_VS_OUT_CCDIST1_VEC_ENA((total_mask & 0xf0) != 0) |
3923 S_02881C_BYPASS_PRIM_RATE_COMBINER_GFX103(pipeline->device->physical_device->rad_info.chip_class >= GFX10_3) |
3924 S_02881C_BYPASS_VTX_RATE_COMBINER_GFX103(pipeline->device->physical_device->rad_info.chip_class >= GFX10_3) |
3925 cull_dist_mask << 8 |
3926 clip_dist_mask);
3927
3928 if (pipeline->device->physical_device->rad_info.chip_class <= GFX8)
3929 radeon_set_context_reg(ctx_cs, R_028AB4_VGT_REUSE_OFF,
3930 outinfo->writes_viewport_index);
3931 }
3932
3933 static void
3934 radv_pipeline_generate_hw_es(struct radeon_cmdbuf *cs,
3935 const struct radv_pipeline *pipeline,
3936 const struct radv_shader_variant *shader)
3937 {
3938 uint64_t va = radv_buffer_get_va(shader->bo) + shader->bo_offset;
3939
3940 radeon_set_sh_reg_seq(cs, R_00B320_SPI_SHADER_PGM_LO_ES, 4);
3941 radeon_emit(cs, va >> 8);
3942 radeon_emit(cs, S_00B324_MEM_BASE(va >> 40));
3943 radeon_emit(cs, shader->config.rsrc1);
3944 radeon_emit(cs, shader->config.rsrc2);
3945 }
3946
3947 static void
3948 radv_pipeline_generate_hw_ls(struct radeon_cmdbuf *cs,
3949 const struct radv_pipeline *pipeline,
3950 const struct radv_shader_variant *shader)
3951 {
3952 unsigned num_lds_blocks = pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.tcs.num_lds_blocks;
3953 uint64_t va = radv_buffer_get_va(shader->bo) + shader->bo_offset;
3954 uint32_t rsrc2 = shader->config.rsrc2;
3955
3956 radeon_set_sh_reg_seq(cs, R_00B520_SPI_SHADER_PGM_LO_LS, 2);
3957 radeon_emit(cs, va >> 8);
3958 radeon_emit(cs, S_00B524_MEM_BASE(va >> 40));
3959
3960 rsrc2 |= S_00B52C_LDS_SIZE(num_lds_blocks);
3961 if (pipeline->device->physical_device->rad_info.chip_class == GFX7 &&
3962 pipeline->device->physical_device->rad_info.family != CHIP_HAWAII)
3963 radeon_set_sh_reg(cs, R_00B52C_SPI_SHADER_PGM_RSRC2_LS, rsrc2);
3964
3965 radeon_set_sh_reg_seq(cs, R_00B528_SPI_SHADER_PGM_RSRC1_LS, 2);
3966 radeon_emit(cs, shader->config.rsrc1);
3967 radeon_emit(cs, rsrc2);
3968 }
3969
3970 static void
3971 radv_pipeline_generate_hw_ngg(struct radeon_cmdbuf *ctx_cs,
3972 struct radeon_cmdbuf *cs,
3973 const struct radv_pipeline *pipeline,
3974 const struct radv_shader_variant *shader)
3975 {
3976 uint64_t va = radv_buffer_get_va(shader->bo) + shader->bo_offset;
3977 gl_shader_stage es_type =
3978 radv_pipeline_has_tess(pipeline) ? MESA_SHADER_TESS_EVAL : MESA_SHADER_VERTEX;
3979 struct radv_shader_variant *es =
3980 es_type == MESA_SHADER_TESS_EVAL ? pipeline->shaders[MESA_SHADER_TESS_EVAL] : pipeline->shaders[MESA_SHADER_VERTEX];
3981 const struct gfx10_ngg_info *ngg_state = &shader->info.ngg_info;
3982
3983 radeon_set_sh_reg_seq(cs, R_00B320_SPI_SHADER_PGM_LO_ES, 2);
3984 radeon_emit(cs, va >> 8);
3985 radeon_emit(cs, S_00B324_MEM_BASE(va >> 40));
3986 radeon_set_sh_reg_seq(cs, R_00B228_SPI_SHADER_PGM_RSRC1_GS, 2);
3987 radeon_emit(cs, shader->config.rsrc1);
3988 radeon_emit(cs, shader->config.rsrc2);
3989
3990 const struct radv_vs_output_info *outinfo = get_vs_output_info(pipeline);
3991 unsigned clip_dist_mask, cull_dist_mask, total_mask;
3992 clip_dist_mask = outinfo->clip_dist_mask;
3993 cull_dist_mask = outinfo->cull_dist_mask;
3994 total_mask = clip_dist_mask | cull_dist_mask;
3995 bool misc_vec_ena = outinfo->writes_pointsize ||
3996 outinfo->writes_layer ||
3997 outinfo->writes_viewport_index;
3998 bool es_enable_prim_id = outinfo->export_prim_id ||
3999 (es && es->info.uses_prim_id);
4000 bool break_wave_at_eoi = false;
4001 unsigned ge_cntl;
4002 unsigned nparams;
4003
4004 if (es_type == MESA_SHADER_TESS_EVAL) {
4005 struct radv_shader_variant *gs =
4006 pipeline->shaders[MESA_SHADER_GEOMETRY];
4007
4008 if (es_enable_prim_id || (gs && gs->info.uses_prim_id))
4009 break_wave_at_eoi = true;
4010 }
4011
4012 nparams = MAX2(outinfo->param_exports, 1);
4013 radeon_set_context_reg(ctx_cs, R_0286C4_SPI_VS_OUT_CONFIG,
4014 S_0286C4_VS_EXPORT_COUNT(nparams - 1) |
4015 S_0286C4_NO_PC_EXPORT(outinfo->param_exports == 0));
4016
4017 radeon_set_context_reg(ctx_cs, R_028708_SPI_SHADER_IDX_FORMAT,
4018 S_028708_IDX0_EXPORT_FORMAT(V_028708_SPI_SHADER_1COMP));
4019 radeon_set_context_reg(ctx_cs, R_02870C_SPI_SHADER_POS_FORMAT,
4020 S_02870C_POS0_EXPORT_FORMAT(V_02870C_SPI_SHADER_4COMP) |
4021 S_02870C_POS1_EXPORT_FORMAT(outinfo->pos_exports > 1 ?
4022 V_02870C_SPI_SHADER_4COMP :
4023 V_02870C_SPI_SHADER_NONE) |
4024 S_02870C_POS2_EXPORT_FORMAT(outinfo->pos_exports > 2 ?
4025 V_02870C_SPI_SHADER_4COMP :
4026 V_02870C_SPI_SHADER_NONE) |
4027 S_02870C_POS3_EXPORT_FORMAT(outinfo->pos_exports > 3 ?
4028 V_02870C_SPI_SHADER_4COMP :
4029 V_02870C_SPI_SHADER_NONE));
4030
4031 radeon_set_context_reg(ctx_cs, R_02881C_PA_CL_VS_OUT_CNTL,
4032 S_02881C_USE_VTX_POINT_SIZE(outinfo->writes_pointsize) |
4033 S_02881C_USE_VTX_RENDER_TARGET_INDX(outinfo->writes_layer) |
4034 S_02881C_USE_VTX_VIEWPORT_INDX(outinfo->writes_viewport_index) |
4035 S_02881C_VS_OUT_MISC_VEC_ENA(misc_vec_ena) |
4036 S_02881C_VS_OUT_MISC_SIDE_BUS_ENA(misc_vec_ena) |
4037 S_02881C_VS_OUT_CCDIST0_VEC_ENA((total_mask & 0x0f) != 0) |
4038 S_02881C_VS_OUT_CCDIST1_VEC_ENA((total_mask & 0xf0) != 0) |
4039 S_02881C_BYPASS_PRIM_RATE_COMBINER_GFX103(pipeline->device->physical_device->rad_info.chip_class >= GFX10_3) |
4040 S_02881C_BYPASS_VTX_RATE_COMBINER_GFX103(pipeline->device->physical_device->rad_info.chip_class >= GFX10_3) |
4041 cull_dist_mask << 8 |
4042 clip_dist_mask);
4043
4044 radeon_set_context_reg(ctx_cs, R_028A84_VGT_PRIMITIVEID_EN,
4045 S_028A84_PRIMITIVEID_EN(es_enable_prim_id) |
4046 S_028A84_NGG_DISABLE_PROVOK_REUSE(outinfo->export_prim_id));
4047
4048 radeon_set_context_reg(ctx_cs, R_028AAC_VGT_ESGS_RING_ITEMSIZE,
4049 ngg_state->vgt_esgs_ring_itemsize);
4050
4051 /* NGG specific registers. */
4052 struct radv_shader_variant *gs = pipeline->shaders[MESA_SHADER_GEOMETRY];
4053 uint32_t gs_num_invocations = gs ? gs->info.gs.invocations : 1;
4054
4055 radeon_set_context_reg(ctx_cs, R_028A44_VGT_GS_ONCHIP_CNTL,
4056 S_028A44_ES_VERTS_PER_SUBGRP(ngg_state->hw_max_esverts) |
4057 S_028A44_GS_PRIMS_PER_SUBGRP(ngg_state->max_gsprims) |
4058 S_028A44_GS_INST_PRIMS_IN_SUBGRP(ngg_state->max_gsprims * gs_num_invocations));
4059 radeon_set_context_reg(ctx_cs, R_0287FC_GE_MAX_OUTPUT_PER_SUBGROUP,
4060 S_0287FC_MAX_VERTS_PER_SUBGROUP(ngg_state->max_out_verts));
4061 radeon_set_context_reg(ctx_cs, R_028B4C_GE_NGG_SUBGRP_CNTL,
4062 S_028B4C_PRIM_AMP_FACTOR(ngg_state->prim_amp_factor) |
4063 S_028B4C_THDS_PER_SUBGRP(0)); /* for fast launch */
4064 radeon_set_context_reg(ctx_cs, R_028B90_VGT_GS_INSTANCE_CNT,
4065 S_028B90_CNT(gs_num_invocations) |
4066 S_028B90_ENABLE(gs_num_invocations > 1) |
4067 S_028B90_EN_MAX_VERT_OUT_PER_GS_INSTANCE(ngg_state->max_vert_out_per_gs_instance));
4068
4069 /* User edge flags are set by the pos exports. If user edge flags are
4070 * not used, we must use hw-generated edge flags and pass them via
4071 * the prim export to prevent drawing lines on internal edges of
4072 * decomposed primitives (such as quads) with polygon mode = lines.
4073 *
4074 * TODO: We should combine hw-generated edge flags with user edge
4075 * flags in the shader.
4076 */
4077 radeon_set_context_reg(ctx_cs, R_028838_PA_CL_NGG_CNTL,
4078 S_028838_INDEX_BUF_EDGE_FLAG_ENA(!radv_pipeline_has_tess(pipeline) &&
4079 !radv_pipeline_has_gs(pipeline)) |
4080 /* Reuse for NGG. */
4081 S_028838_VERTEX_REUSE_DEPTH_GFX103(pipeline->device->physical_device->rad_info.chip_class >= GFX10_3 ? 30 : 0));
4082
4083 ge_cntl = S_03096C_PRIM_GRP_SIZE(ngg_state->max_gsprims) |
4084 S_03096C_VERT_GRP_SIZE(256) | /* 256 = disable vertex grouping */
4085 S_03096C_BREAK_WAVE_AT_EOI(break_wave_at_eoi);
4086
4087 /* Bug workaround for a possible hang with non-tessellation cases.
4088 * Tessellation always sets GE_CNTL.VERT_GRP_SIZE = 0
4089 *
4090 * Requirement: GE_CNTL.VERT_GRP_SIZE = VGT_GS_ONCHIP_CNTL.ES_VERTS_PER_SUBGRP - 5
4091 */
4092 if (pipeline->device->physical_device->rad_info.chip_class == GFX10 &&
4093 !radv_pipeline_has_tess(pipeline) &&
4094 ngg_state->hw_max_esverts != 256) {
4095 ge_cntl &= C_03096C_VERT_GRP_SIZE;
4096
4097 if (ngg_state->hw_max_esverts > 5) {
4098 ge_cntl |= S_03096C_VERT_GRP_SIZE(ngg_state->hw_max_esverts - 5);
4099 }
4100 }
4101
4102 radeon_set_uconfig_reg(ctx_cs, R_03096C_GE_CNTL, ge_cntl);
4103 }
4104
4105 static void
4106 radv_pipeline_generate_hw_hs(struct radeon_cmdbuf *cs,
4107 const struct radv_pipeline *pipeline,
4108 const struct radv_shader_variant *shader)
4109 {
4110 uint64_t va = radv_buffer_get_va(shader->bo) + shader->bo_offset;
4111
4112 if (pipeline->device->physical_device->rad_info.chip_class >= GFX9) {
4113 if (pipeline->device->physical_device->rad_info.chip_class >= GFX10) {
4114 radeon_set_sh_reg_seq(cs, R_00B520_SPI_SHADER_PGM_LO_LS, 2);
4115 radeon_emit(cs, va >> 8);
4116 radeon_emit(cs, S_00B524_MEM_BASE(va >> 40));
4117 } else {
4118 radeon_set_sh_reg_seq(cs, R_00B410_SPI_SHADER_PGM_LO_LS, 2);
4119 radeon_emit(cs, va >> 8);
4120 radeon_emit(cs, S_00B414_MEM_BASE(va >> 40));
4121 }
4122
4123 radeon_set_sh_reg_seq(cs, R_00B428_SPI_SHADER_PGM_RSRC1_HS, 2);
4124 radeon_emit(cs, shader->config.rsrc1);
4125 radeon_emit(cs, shader->config.rsrc2);
4126 } else {
4127 radeon_set_sh_reg_seq(cs, R_00B420_SPI_SHADER_PGM_LO_HS, 4);
4128 radeon_emit(cs, va >> 8);
4129 radeon_emit(cs, S_00B424_MEM_BASE(va >> 40));
4130 radeon_emit(cs, shader->config.rsrc1);
4131 radeon_emit(cs, shader->config.rsrc2);
4132 }
4133 }
4134
4135 static void
4136 radv_pipeline_generate_vertex_shader(struct radeon_cmdbuf *ctx_cs,
4137 struct radeon_cmdbuf *cs,
4138 const struct radv_pipeline *pipeline)
4139 {
4140 struct radv_shader_variant *vs;
4141
4142 /* Skip shaders merged into HS/GS */
4143 vs = pipeline->shaders[MESA_SHADER_VERTEX];
4144 if (!vs)
4145 return;
4146
4147 if (vs->info.vs.as_ls)
4148 radv_pipeline_generate_hw_ls(cs, pipeline, vs);
4149 else if (vs->info.vs.as_es)
4150 radv_pipeline_generate_hw_es(cs, pipeline, vs);
4151 else if (vs->info.is_ngg)
4152 radv_pipeline_generate_hw_ngg(ctx_cs, cs, pipeline, vs);
4153 else
4154 radv_pipeline_generate_hw_vs(ctx_cs, cs, pipeline, vs);
4155 }
4156
4157 static void
4158 radv_pipeline_generate_tess_shaders(struct radeon_cmdbuf *ctx_cs,
4159 struct radeon_cmdbuf *cs,
4160 const struct radv_pipeline *pipeline)
4161 {
4162 struct radv_shader_variant *tes, *tcs;
4163
4164 tcs = pipeline->shaders[MESA_SHADER_TESS_CTRL];
4165 tes = pipeline->shaders[MESA_SHADER_TESS_EVAL];
4166
4167 if (tes) {
4168 if (tes->info.is_ngg) {
4169 radv_pipeline_generate_hw_ngg(ctx_cs, cs, pipeline, tes);
4170 } else if (tes->info.tes.as_es)
4171 radv_pipeline_generate_hw_es(cs, pipeline, tes);
4172 else
4173 radv_pipeline_generate_hw_vs(ctx_cs, cs, pipeline, tes);
4174 }
4175
4176 radv_pipeline_generate_hw_hs(cs, pipeline, tcs);
4177
4178 if (pipeline->device->physical_device->rad_info.chip_class >= GFX10 &&
4179 !radv_pipeline_has_gs(pipeline) && !radv_pipeline_has_ngg(pipeline)) {
4180 radeon_set_context_reg(ctx_cs, R_028A44_VGT_GS_ONCHIP_CNTL,
4181 S_028A44_ES_VERTS_PER_SUBGRP(250) |
4182 S_028A44_GS_PRIMS_PER_SUBGRP(126) |
4183 S_028A44_GS_INST_PRIMS_IN_SUBGRP(126));
4184 }
4185 }
4186
4187 static void
4188 radv_pipeline_generate_tess_state(struct radeon_cmdbuf *ctx_cs,
4189 const struct radv_pipeline *pipeline,
4190 const VkGraphicsPipelineCreateInfo *pCreateInfo)
4191 {
4192 struct radv_shader_variant *tes = radv_get_shader(pipeline, MESA_SHADER_TESS_EVAL);
4193 unsigned type = 0, partitioning = 0, topology = 0, distribution_mode = 0;
4194 unsigned num_tcs_input_cp, num_tcs_output_cp, num_patches;
4195 unsigned ls_hs_config;
4196
4197 num_tcs_input_cp = pCreateInfo->pTessellationState->patchControlPoints;
4198 num_tcs_output_cp = pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.tcs.tcs_vertices_out; //TCS VERTICES OUT
4199 num_patches = pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.tcs.num_patches;
4200
4201 ls_hs_config = S_028B58_NUM_PATCHES(num_patches) |
4202 S_028B58_HS_NUM_INPUT_CP(num_tcs_input_cp) |
4203 S_028B58_HS_NUM_OUTPUT_CP(num_tcs_output_cp);
4204
4205 if (pipeline->device->physical_device->rad_info.chip_class >= GFX7) {
4206 radeon_set_context_reg_idx(ctx_cs, R_028B58_VGT_LS_HS_CONFIG,
4207 2, ls_hs_config);
4208 } else {
4209 radeon_set_context_reg(ctx_cs, R_028B58_VGT_LS_HS_CONFIG,
4210 ls_hs_config);
4211 }
4212
4213 switch (tes->info.tes.primitive_mode) {
4214 case GL_TRIANGLES:
4215 type = V_028B6C_TESS_TRIANGLE;
4216 break;
4217 case GL_QUADS:
4218 type = V_028B6C_TESS_QUAD;
4219 break;
4220 case GL_ISOLINES:
4221 type = V_028B6C_TESS_ISOLINE;
4222 break;
4223 }
4224
4225 switch (tes->info.tes.spacing) {
4226 case TESS_SPACING_EQUAL:
4227 partitioning = V_028B6C_PART_INTEGER;
4228 break;
4229 case TESS_SPACING_FRACTIONAL_ODD:
4230 partitioning = V_028B6C_PART_FRAC_ODD;
4231 break;
4232 case TESS_SPACING_FRACTIONAL_EVEN:
4233 partitioning = V_028B6C_PART_FRAC_EVEN;
4234 break;
4235 default:
4236 break;
4237 }
4238
4239 bool ccw = tes->info.tes.ccw;
4240 const VkPipelineTessellationDomainOriginStateCreateInfo *domain_origin_state =
4241 vk_find_struct_const(pCreateInfo->pTessellationState,
4242 PIPELINE_TESSELLATION_DOMAIN_ORIGIN_STATE_CREATE_INFO);
4243
4244 if (domain_origin_state && domain_origin_state->domainOrigin != VK_TESSELLATION_DOMAIN_ORIGIN_UPPER_LEFT)
4245 ccw = !ccw;
4246
4247 if (tes->info.tes.point_mode)
4248 topology = V_028B6C_OUTPUT_POINT;
4249 else if (tes->info.tes.primitive_mode == GL_ISOLINES)
4250 topology = V_028B6C_OUTPUT_LINE;
4251 else if (ccw)
4252 topology = V_028B6C_OUTPUT_TRIANGLE_CCW;
4253 else
4254 topology = V_028B6C_OUTPUT_TRIANGLE_CW;
4255
4256 if (pipeline->device->physical_device->rad_info.has_distributed_tess) {
4257 if (pipeline->device->physical_device->rad_info.family == CHIP_FIJI ||
4258 pipeline->device->physical_device->rad_info.family >= CHIP_POLARIS10)
4259 distribution_mode = V_028B6C_DISTRIBUTION_MODE_TRAPEZOIDS;
4260 else
4261 distribution_mode = V_028B6C_DISTRIBUTION_MODE_DONUTS;
4262 } else
4263 distribution_mode = V_028B6C_DISTRIBUTION_MODE_NO_DIST;
4264
4265 radeon_set_context_reg(ctx_cs, R_028B6C_VGT_TF_PARAM,
4266 S_028B6C_TYPE(type) |
4267 S_028B6C_PARTITIONING(partitioning) |
4268 S_028B6C_TOPOLOGY(topology) |
4269 S_028B6C_DISTRIBUTION_MODE(distribution_mode));
4270 }
4271
4272 static void
4273 radv_pipeline_generate_hw_gs(struct radeon_cmdbuf *ctx_cs,
4274 struct radeon_cmdbuf *cs,
4275 const struct radv_pipeline *pipeline,
4276 const struct radv_shader_variant *gs)
4277 {
4278 const struct gfx9_gs_info *gs_state = &gs->info.gs_ring_info;
4279 unsigned gs_max_out_vertices;
4280 const uint8_t *num_components;
4281 uint8_t max_stream;
4282 unsigned offset;
4283 uint64_t va;
4284
4285 gs_max_out_vertices = gs->info.gs.vertices_out;
4286 max_stream = gs->info.gs.max_stream;
4287 num_components = gs->info.gs.num_stream_output_components;
4288
4289 offset = num_components[0] * gs_max_out_vertices;
4290
4291 radeon_set_context_reg_seq(ctx_cs, R_028A60_VGT_GSVS_RING_OFFSET_1, 3);
4292 radeon_emit(ctx_cs, offset);
4293 if (max_stream >= 1)
4294 offset += num_components[1] * gs_max_out_vertices;
4295 radeon_emit(ctx_cs, offset);
4296 if (max_stream >= 2)
4297 offset += num_components[2] * gs_max_out_vertices;
4298 radeon_emit(ctx_cs, offset);
4299 if (max_stream >= 3)
4300 offset += num_components[3] * gs_max_out_vertices;
4301 radeon_set_context_reg(ctx_cs, R_028AB0_VGT_GSVS_RING_ITEMSIZE, offset);
4302
4303 radeon_set_context_reg_seq(ctx_cs, R_028B5C_VGT_GS_VERT_ITEMSIZE, 4);
4304 radeon_emit(ctx_cs, num_components[0]);
4305 radeon_emit(ctx_cs, (max_stream >= 1) ? num_components[1] : 0);
4306 radeon_emit(ctx_cs, (max_stream >= 2) ? num_components[2] : 0);
4307 radeon_emit(ctx_cs, (max_stream >= 3) ? num_components[3] : 0);
4308
4309 uint32_t gs_num_invocations = gs->info.gs.invocations;
4310 radeon_set_context_reg(ctx_cs, R_028B90_VGT_GS_INSTANCE_CNT,
4311 S_028B90_CNT(MIN2(gs_num_invocations, 127)) |
4312 S_028B90_ENABLE(gs_num_invocations > 0));
4313
4314 radeon_set_context_reg(ctx_cs, R_028AAC_VGT_ESGS_RING_ITEMSIZE,
4315 gs_state->vgt_esgs_ring_itemsize);
4316
4317 va = radv_buffer_get_va(gs->bo) + gs->bo_offset;
4318
4319 if (pipeline->device->physical_device->rad_info.chip_class >= GFX9) {
4320 if (pipeline->device->physical_device->rad_info.chip_class >= GFX10) {
4321 radeon_set_sh_reg_seq(cs, R_00B320_SPI_SHADER_PGM_LO_ES, 2);
4322 radeon_emit(cs, va >> 8);
4323 radeon_emit(cs, S_00B324_MEM_BASE(va >> 40));
4324 } else {
4325 radeon_set_sh_reg_seq(cs, R_00B210_SPI_SHADER_PGM_LO_ES, 2);
4326 radeon_emit(cs, va >> 8);
4327 radeon_emit(cs, S_00B214_MEM_BASE(va >> 40));
4328 }
4329
4330 radeon_set_sh_reg_seq(cs, R_00B228_SPI_SHADER_PGM_RSRC1_GS, 2);
4331 radeon_emit(cs, gs->config.rsrc1);
4332 radeon_emit(cs, gs->config.rsrc2 | S_00B22C_LDS_SIZE(gs_state->lds_size));
4333
4334 radeon_set_context_reg(ctx_cs, R_028A44_VGT_GS_ONCHIP_CNTL, gs_state->vgt_gs_onchip_cntl);
4335 radeon_set_context_reg(ctx_cs, R_028A94_VGT_GS_MAX_PRIMS_PER_SUBGROUP, gs_state->vgt_gs_max_prims_per_subgroup);
4336 } else {
4337 radeon_set_sh_reg_seq(cs, R_00B220_SPI_SHADER_PGM_LO_GS, 4);
4338 radeon_emit(cs, va >> 8);
4339 radeon_emit(cs, S_00B224_MEM_BASE(va >> 40));
4340 radeon_emit(cs, gs->config.rsrc1);
4341 radeon_emit(cs, gs->config.rsrc2);
4342 }
4343
4344 radv_pipeline_generate_hw_vs(ctx_cs, cs, pipeline, pipeline->gs_copy_shader);
4345 }
4346
4347 static void
4348 radv_pipeline_generate_geometry_shader(struct radeon_cmdbuf *ctx_cs,
4349 struct radeon_cmdbuf *cs,
4350 const struct radv_pipeline *pipeline)
4351 {
4352 struct radv_shader_variant *gs;
4353
4354 gs = pipeline->shaders[MESA_SHADER_GEOMETRY];
4355 if (!gs)
4356 return;
4357
4358 if (gs->info.is_ngg)
4359 radv_pipeline_generate_hw_ngg(ctx_cs, cs, pipeline, gs);
4360 else
4361 radv_pipeline_generate_hw_gs(ctx_cs, cs, pipeline, gs);
4362
4363 radeon_set_context_reg(ctx_cs, R_028B38_VGT_GS_MAX_VERT_OUT,
4364 gs->info.gs.vertices_out);
4365 }
4366
4367 static uint32_t offset_to_ps_input(uint32_t offset, bool flat_shade,
4368 bool explicit, bool float16)
4369 {
4370 uint32_t ps_input_cntl;
4371 if (offset <= AC_EXP_PARAM_OFFSET_31) {
4372 ps_input_cntl = S_028644_OFFSET(offset);
4373 if (flat_shade || explicit)
4374 ps_input_cntl |= S_028644_FLAT_SHADE(1);
4375 if (explicit) {
4376 /* Force parameter cache to be read in passthrough
4377 * mode.
4378 */
4379 ps_input_cntl |= S_028644_OFFSET(1 << 5);
4380 }
4381 if (float16) {
4382 ps_input_cntl |= S_028644_FP16_INTERP_MODE(1) |
4383 S_028644_ATTR0_VALID(1);
4384 }
4385 } else {
4386 /* The input is a DEFAULT_VAL constant. */
4387 assert(offset >= AC_EXP_PARAM_DEFAULT_VAL_0000 &&
4388 offset <= AC_EXP_PARAM_DEFAULT_VAL_1111);
4389 offset -= AC_EXP_PARAM_DEFAULT_VAL_0000;
4390 ps_input_cntl = S_028644_OFFSET(0x20) |
4391 S_028644_DEFAULT_VAL(offset);
4392 }
4393 return ps_input_cntl;
4394 }
4395
4396 static void
4397 radv_pipeline_generate_ps_inputs(struct radeon_cmdbuf *ctx_cs,
4398 const struct radv_pipeline *pipeline)
4399 {
4400 struct radv_shader_variant *ps = pipeline->shaders[MESA_SHADER_FRAGMENT];
4401 const struct radv_vs_output_info *outinfo = get_vs_output_info(pipeline);
4402 uint32_t ps_input_cntl[32];
4403
4404 unsigned ps_offset = 0;
4405
4406 if (ps->info.ps.prim_id_input) {
4407 unsigned vs_offset = outinfo->vs_output_param_offset[VARYING_SLOT_PRIMITIVE_ID];
4408 if (vs_offset != AC_EXP_PARAM_UNDEFINED) {
4409 ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, true, false, false);
4410 ++ps_offset;
4411 }
4412 }
4413
4414 if (ps->info.ps.layer_input ||
4415 ps->info.needs_multiview_view_index) {
4416 unsigned vs_offset = outinfo->vs_output_param_offset[VARYING_SLOT_LAYER];
4417 if (vs_offset != AC_EXP_PARAM_UNDEFINED)
4418 ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, true, false, false);
4419 else
4420 ps_input_cntl[ps_offset] = offset_to_ps_input(AC_EXP_PARAM_DEFAULT_VAL_0000, true, false, false);
4421 ++ps_offset;
4422 }
4423
4424 if (ps->info.ps.viewport_index_input) {
4425 unsigned vs_offset = outinfo->vs_output_param_offset[VARYING_SLOT_VIEWPORT];
4426 if (vs_offset != AC_EXP_PARAM_UNDEFINED)
4427 ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, true, false, false);
4428 else
4429 ps_input_cntl[ps_offset] = offset_to_ps_input(AC_EXP_PARAM_DEFAULT_VAL_0000, true, false, false);
4430 ++ps_offset;
4431 }
4432
4433 if (ps->info.ps.has_pcoord) {
4434 unsigned val;
4435 val = S_028644_PT_SPRITE_TEX(1) | S_028644_OFFSET(0x20);
4436 ps_input_cntl[ps_offset] = val;
4437 ps_offset++;
4438 }
4439
4440 if (ps->info.ps.num_input_clips_culls) {
4441 unsigned vs_offset;
4442
4443 vs_offset = outinfo->vs_output_param_offset[VARYING_SLOT_CLIP_DIST0];
4444 if (vs_offset != AC_EXP_PARAM_UNDEFINED) {
4445 ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, false, false, false);
4446 ++ps_offset;
4447 }
4448
4449 vs_offset = outinfo->vs_output_param_offset[VARYING_SLOT_CLIP_DIST1];
4450 if (vs_offset != AC_EXP_PARAM_UNDEFINED &&
4451 ps->info.ps.num_input_clips_culls > 4) {
4452 ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, false, false, false);
4453 ++ps_offset;
4454 }
4455 }
4456
4457 for (unsigned i = 0; i < 32 && (1u << i) <= ps->info.ps.input_mask; ++i) {
4458 unsigned vs_offset;
4459 bool flat_shade;
4460 bool explicit;
4461 bool float16;
4462 if (!(ps->info.ps.input_mask & (1u << i)))
4463 continue;
4464
4465 vs_offset = outinfo->vs_output_param_offset[VARYING_SLOT_VAR0 + i];
4466 if (vs_offset == AC_EXP_PARAM_UNDEFINED) {
4467 ps_input_cntl[ps_offset] = S_028644_OFFSET(0x20);
4468 ++ps_offset;
4469 continue;
4470 }
4471
4472 flat_shade = !!(ps->info.ps.flat_shaded_mask & (1u << ps_offset));
4473 explicit = !!(ps->info.ps.explicit_shaded_mask & (1u << ps_offset));
4474 float16 = !!(ps->info.ps.float16_shaded_mask & (1u << ps_offset));
4475
4476 ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, flat_shade, explicit, float16);
4477 ++ps_offset;
4478 }
4479
4480 if (ps_offset) {
4481 radeon_set_context_reg_seq(ctx_cs, R_028644_SPI_PS_INPUT_CNTL_0, ps_offset);
4482 for (unsigned i = 0; i < ps_offset; i++) {
4483 radeon_emit(ctx_cs, ps_input_cntl[i]);
4484 }
4485 }
4486 }
4487
4488 static uint32_t
4489 radv_compute_db_shader_control(const struct radv_device *device,
4490 const struct radv_pipeline *pipeline,
4491 const struct radv_shader_variant *ps)
4492 {
4493 unsigned conservative_z_export = V_02880C_EXPORT_ANY_Z;
4494 unsigned z_order;
4495 if (ps->info.ps.early_fragment_test || !ps->info.ps.writes_memory)
4496 z_order = V_02880C_EARLY_Z_THEN_LATE_Z;
4497 else
4498 z_order = V_02880C_LATE_Z;
4499
4500 if (ps->info.ps.depth_layout == FRAG_DEPTH_LAYOUT_GREATER)
4501 conservative_z_export = V_02880C_EXPORT_GREATER_THAN_Z;
4502 else if (ps->info.ps.depth_layout == FRAG_DEPTH_LAYOUT_LESS)
4503 conservative_z_export = V_02880C_EXPORT_LESS_THAN_Z;
4504
4505 bool disable_rbplus = device->physical_device->rad_info.has_rbplus &&
4506 !device->physical_device->rad_info.rbplus_allowed;
4507
4508 /* It shouldn't be needed to export gl_SampleMask when MSAA is disabled
4509 * but this appears to break Project Cars (DXVK). See
4510 * https://bugs.freedesktop.org/show_bug.cgi?id=109401
4511 */
4512 bool mask_export_enable = ps->info.ps.writes_sample_mask;
4513
4514 return S_02880C_Z_EXPORT_ENABLE(ps->info.ps.writes_z) |
4515 S_02880C_STENCIL_TEST_VAL_EXPORT_ENABLE(ps->info.ps.writes_stencil) |
4516 S_02880C_KILL_ENABLE(!!ps->info.ps.can_discard) |
4517 S_02880C_MASK_EXPORT_ENABLE(mask_export_enable) |
4518 S_02880C_CONSERVATIVE_Z_EXPORT(conservative_z_export) |
4519 S_02880C_Z_ORDER(z_order) |
4520 S_02880C_DEPTH_BEFORE_SHADER(ps->info.ps.early_fragment_test) |
4521 S_02880C_PRE_SHADER_DEPTH_COVERAGE_ENABLE(ps->info.ps.post_depth_coverage) |
4522 S_02880C_EXEC_ON_HIER_FAIL(ps->info.ps.writes_memory) |
4523 S_02880C_EXEC_ON_NOOP(ps->info.ps.writes_memory) |
4524 S_02880C_DUAL_QUAD_DISABLE(disable_rbplus);
4525 }
4526
4527 static void
4528 radv_pipeline_generate_fragment_shader(struct radeon_cmdbuf *ctx_cs,
4529 struct radeon_cmdbuf *cs,
4530 struct radv_pipeline *pipeline)
4531 {
4532 struct radv_shader_variant *ps;
4533 uint64_t va;
4534 assert (pipeline->shaders[MESA_SHADER_FRAGMENT]);
4535
4536 ps = pipeline->shaders[MESA_SHADER_FRAGMENT];
4537 va = radv_buffer_get_va(ps->bo) + ps->bo_offset;
4538
4539 radeon_set_sh_reg_seq(cs, R_00B020_SPI_SHADER_PGM_LO_PS, 4);
4540 radeon_emit(cs, va >> 8);
4541 radeon_emit(cs, S_00B024_MEM_BASE(va >> 40));
4542 radeon_emit(cs, ps->config.rsrc1);
4543 radeon_emit(cs, ps->config.rsrc2);
4544
4545 radeon_set_context_reg(ctx_cs, R_02880C_DB_SHADER_CONTROL,
4546 radv_compute_db_shader_control(pipeline->device,
4547 pipeline, ps));
4548
4549 radeon_set_context_reg(ctx_cs, R_0286CC_SPI_PS_INPUT_ENA,
4550 ps->config.spi_ps_input_ena);
4551
4552 radeon_set_context_reg(ctx_cs, R_0286D0_SPI_PS_INPUT_ADDR,
4553 ps->config.spi_ps_input_addr);
4554
4555 radeon_set_context_reg(ctx_cs, R_0286D8_SPI_PS_IN_CONTROL,
4556 S_0286D8_NUM_INTERP(ps->info.ps.num_interp) |
4557 S_0286D8_PS_W32_EN(ps->info.wave_size == 32));
4558
4559 radeon_set_context_reg(ctx_cs, R_0286E0_SPI_BARYC_CNTL, pipeline->graphics.spi_baryc_cntl);
4560
4561 radeon_set_context_reg(ctx_cs, R_028710_SPI_SHADER_Z_FORMAT,
4562 ac_get_spi_shader_z_format(ps->info.ps.writes_z,
4563 ps->info.ps.writes_stencil,
4564 ps->info.ps.writes_sample_mask));
4565
4566 if (pipeline->device->dfsm_allowed) {
4567 /* optimise this? */
4568 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
4569 radeon_emit(cs, EVENT_TYPE(V_028A90_FLUSH_DFSM) | EVENT_INDEX(0));
4570 }
4571 }
4572
4573 static void
4574 radv_pipeline_generate_vgt_vertex_reuse(struct radeon_cmdbuf *ctx_cs,
4575 const struct radv_pipeline *pipeline)
4576 {
4577 if (pipeline->device->physical_device->rad_info.family < CHIP_POLARIS10 ||
4578 pipeline->device->physical_device->rad_info.chip_class >= GFX10)
4579 return;
4580
4581 unsigned vtx_reuse_depth = 30;
4582 if (radv_pipeline_has_tess(pipeline) &&
4583 radv_get_shader(pipeline, MESA_SHADER_TESS_EVAL)->info.tes.spacing == TESS_SPACING_FRACTIONAL_ODD) {
4584 vtx_reuse_depth = 14;
4585 }
4586 radeon_set_context_reg(ctx_cs, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL,
4587 S_028C58_VTX_REUSE_DEPTH(vtx_reuse_depth));
4588 }
4589
4590 static void
4591 radv_pipeline_generate_vgt_shader_config(struct radeon_cmdbuf *ctx_cs,
4592 const struct radv_pipeline *pipeline)
4593 {
4594 uint32_t stages = 0;
4595 if (radv_pipeline_has_tess(pipeline)) {
4596 stages |= S_028B54_LS_EN(V_028B54_LS_STAGE_ON) |
4597 S_028B54_HS_EN(1) | S_028B54_DYNAMIC_HS(1);
4598
4599 if (radv_pipeline_has_gs(pipeline))
4600 stages |= S_028B54_ES_EN(V_028B54_ES_STAGE_DS) |
4601 S_028B54_GS_EN(1);
4602 else if (radv_pipeline_has_ngg(pipeline))
4603 stages |= S_028B54_ES_EN(V_028B54_ES_STAGE_DS);
4604 else
4605 stages |= S_028B54_VS_EN(V_028B54_VS_STAGE_DS);
4606 } else if (radv_pipeline_has_gs(pipeline)) {
4607 stages |= S_028B54_ES_EN(V_028B54_ES_STAGE_REAL) |
4608 S_028B54_GS_EN(1);
4609 } else if (radv_pipeline_has_ngg(pipeline)) {
4610 stages |= S_028B54_ES_EN(V_028B54_ES_STAGE_REAL);
4611 }
4612
4613 if (radv_pipeline_has_ngg(pipeline)) {
4614 stages |= S_028B54_PRIMGEN_EN(1);
4615 if (pipeline->streamout_shader)
4616 stages |= S_028B54_NGG_WAVE_ID_EN(1);
4617 if (radv_pipeline_has_ngg_passthrough(pipeline))
4618 stages |= S_028B54_PRIMGEN_PASSTHRU_EN(1);
4619 } else if (radv_pipeline_has_gs(pipeline)) {
4620 stages |= S_028B54_VS_EN(V_028B54_VS_STAGE_COPY_SHADER);
4621 }
4622
4623 if (pipeline->device->physical_device->rad_info.chip_class >= GFX9)
4624 stages |= S_028B54_MAX_PRIMGRP_IN_WAVE(2);
4625
4626 if (pipeline->device->physical_device->rad_info.chip_class >= GFX10) {
4627 uint8_t hs_size = 64, gs_size = 64, vs_size = 64;
4628
4629 if (radv_pipeline_has_tess(pipeline))
4630 hs_size = pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.wave_size;
4631
4632 if (pipeline->shaders[MESA_SHADER_GEOMETRY]) {
4633 vs_size = gs_size = pipeline->shaders[MESA_SHADER_GEOMETRY]->info.wave_size;
4634 if (pipeline->gs_copy_shader)
4635 vs_size = pipeline->gs_copy_shader->info.wave_size;
4636 } else if (pipeline->shaders[MESA_SHADER_TESS_EVAL])
4637 vs_size = pipeline->shaders[MESA_SHADER_TESS_EVAL]->info.wave_size;
4638 else if (pipeline->shaders[MESA_SHADER_VERTEX])
4639 vs_size = pipeline->shaders[MESA_SHADER_VERTEX]->info.wave_size;
4640
4641 if (radv_pipeline_has_ngg(pipeline))
4642 gs_size = vs_size;
4643
4644 /* legacy GS only supports Wave64 */
4645 stages |= S_028B54_HS_W32_EN(hs_size == 32 ? 1 : 0) |
4646 S_028B54_GS_W32_EN(gs_size == 32 ? 1 : 0) |
4647 S_028B54_VS_W32_EN(vs_size == 32 ? 1 : 0);
4648 }
4649
4650 radeon_set_context_reg(ctx_cs, R_028B54_VGT_SHADER_STAGES_EN, stages);
4651 }
4652
4653 static void
4654 radv_pipeline_generate_cliprect_rule(struct radeon_cmdbuf *ctx_cs,
4655 const VkGraphicsPipelineCreateInfo *pCreateInfo)
4656 {
4657 const VkPipelineDiscardRectangleStateCreateInfoEXT *discard_rectangle_info =
4658 vk_find_struct_const(pCreateInfo->pNext, PIPELINE_DISCARD_RECTANGLE_STATE_CREATE_INFO_EXT);
4659 uint32_t cliprect_rule = 0;
4660
4661 if (!discard_rectangle_info) {
4662 cliprect_rule = 0xffff;
4663 } else {
4664 for (unsigned i = 0; i < (1u << MAX_DISCARD_RECTANGLES); ++i) {
4665 /* Interpret i as a bitmask, and then set the bit in
4666 * the mask if that combination of rectangles in which
4667 * the pixel is contained should pass the cliprect
4668 * test.
4669 */
4670 unsigned relevant_subset = i & ((1u << discard_rectangle_info->discardRectangleCount) - 1);
4671
4672 if (discard_rectangle_info->discardRectangleMode == VK_DISCARD_RECTANGLE_MODE_INCLUSIVE_EXT &&
4673 !relevant_subset)
4674 continue;
4675
4676 if (discard_rectangle_info->discardRectangleMode == VK_DISCARD_RECTANGLE_MODE_EXCLUSIVE_EXT &&
4677 relevant_subset)
4678 continue;
4679
4680 cliprect_rule |= 1u << i;
4681 }
4682 }
4683
4684 radeon_set_context_reg(ctx_cs, R_02820C_PA_SC_CLIPRECT_RULE, cliprect_rule);
4685 }
4686
4687 static void
4688 gfx10_pipeline_generate_ge_cntl(struct radeon_cmdbuf *ctx_cs,
4689 struct radv_pipeline *pipeline)
4690 {
4691 bool break_wave_at_eoi = false;
4692 unsigned primgroup_size;
4693 unsigned vertgroup_size = 256; /* 256 = disable vertex grouping */
4694
4695 if (radv_pipeline_has_tess(pipeline)) {
4696 primgroup_size = pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.tcs.num_patches;
4697 } else if (radv_pipeline_has_gs(pipeline)) {
4698 const struct gfx9_gs_info *gs_state =
4699 &pipeline->shaders[MESA_SHADER_GEOMETRY]->info.gs_ring_info;
4700 unsigned vgt_gs_onchip_cntl = gs_state->vgt_gs_onchip_cntl;
4701 primgroup_size = G_028A44_GS_PRIMS_PER_SUBGRP(vgt_gs_onchip_cntl);
4702 } else {
4703 primgroup_size = 128; /* recommended without a GS and tess */
4704 }
4705
4706 if (radv_pipeline_has_tess(pipeline)) {
4707 if (pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.uses_prim_id ||
4708 radv_get_shader(pipeline, MESA_SHADER_TESS_EVAL)->info.uses_prim_id)
4709 break_wave_at_eoi = true;
4710 }
4711
4712 radeon_set_uconfig_reg(ctx_cs, R_03096C_GE_CNTL,
4713 S_03096C_PRIM_GRP_SIZE(primgroup_size) |
4714 S_03096C_VERT_GRP_SIZE(vertgroup_size) |
4715 S_03096C_PACKET_TO_ONE_PA(0) /* line stipple */ |
4716 S_03096C_BREAK_WAVE_AT_EOI(break_wave_at_eoi));
4717 }
4718
4719 static void
4720 radv_pipeline_generate_vgt_gs_out(struct radeon_cmdbuf *ctx_cs,
4721 const struct radv_pipeline *pipeline,
4722 const VkGraphicsPipelineCreateInfo *pCreateInfo,
4723 const struct radv_graphics_pipeline_create_info *extra)
4724 {
4725 uint32_t gs_out;
4726
4727 if (radv_pipeline_has_gs(pipeline)) {
4728 gs_out = si_conv_gl_prim_to_gs_out(pipeline->shaders[MESA_SHADER_GEOMETRY]->info.gs.output_prim);
4729 } else if (radv_pipeline_has_tess(pipeline)) {
4730 if (pipeline->shaders[MESA_SHADER_TESS_EVAL]->info.tes.point_mode) {
4731 gs_out = V_028A6C_OUTPRIM_TYPE_POINTLIST;
4732 } else {
4733 gs_out = si_conv_gl_prim_to_gs_out(pipeline->shaders[MESA_SHADER_TESS_EVAL]->info.tes.primitive_mode);
4734 }
4735 } else {
4736 gs_out = si_conv_prim_to_gs_out(pCreateInfo->pInputAssemblyState->topology);
4737 }
4738
4739 if (extra && extra->use_rectlist) {
4740 gs_out = V_028A6C_OUTPRIM_TYPE_TRISTRIP;
4741 if (radv_pipeline_has_ngg(pipeline))
4742 gs_out = V_028A6C_VGT_OUT_RECT_V0;
4743 }
4744
4745 radeon_set_context_reg(ctx_cs, R_028A6C_VGT_GS_OUT_PRIM_TYPE, gs_out);
4746 }
4747
4748 static void
4749 radv_pipeline_generate_pm4(struct radv_pipeline *pipeline,
4750 const VkGraphicsPipelineCreateInfo *pCreateInfo,
4751 const struct radv_graphics_pipeline_create_info *extra,
4752 const struct radv_blend_state *blend)
4753 {
4754 struct radeon_cmdbuf *ctx_cs = &pipeline->ctx_cs;
4755 struct radeon_cmdbuf *cs = &pipeline->cs;
4756
4757 cs->max_dw = 64;
4758 ctx_cs->max_dw = 256;
4759 cs->buf = malloc(4 * (cs->max_dw + ctx_cs->max_dw));
4760 ctx_cs->buf = cs->buf + cs->max_dw;
4761
4762 radv_pipeline_generate_depth_stencil_state(ctx_cs, pipeline, pCreateInfo, extra);
4763 radv_pipeline_generate_blend_state(ctx_cs, pipeline, blend);
4764 radv_pipeline_generate_raster_state(ctx_cs, pipeline, pCreateInfo);
4765 radv_pipeline_generate_multisample_state(ctx_cs, pipeline);
4766 radv_pipeline_generate_vgt_gs_mode(ctx_cs, pipeline);
4767 radv_pipeline_generate_vertex_shader(ctx_cs, cs, pipeline);
4768
4769 if (radv_pipeline_has_tess(pipeline)) {
4770 radv_pipeline_generate_tess_shaders(ctx_cs, cs, pipeline);
4771 radv_pipeline_generate_tess_state(ctx_cs, pipeline, pCreateInfo);
4772 }
4773
4774 radv_pipeline_generate_geometry_shader(ctx_cs, cs, pipeline);
4775 radv_pipeline_generate_fragment_shader(ctx_cs, cs, pipeline);
4776 radv_pipeline_generate_ps_inputs(ctx_cs, pipeline);
4777 radv_pipeline_generate_vgt_vertex_reuse(ctx_cs, pipeline);
4778 radv_pipeline_generate_vgt_shader_config(ctx_cs, pipeline);
4779 radv_pipeline_generate_cliprect_rule(ctx_cs, pCreateInfo);
4780 radv_pipeline_generate_vgt_gs_out(ctx_cs, pipeline, pCreateInfo, extra);
4781
4782 if (pipeline->device->physical_device->rad_info.chip_class >= GFX10 && !radv_pipeline_has_ngg(pipeline))
4783 gfx10_pipeline_generate_ge_cntl(ctx_cs, pipeline);
4784
4785 pipeline->ctx_cs_hash = _mesa_hash_data(ctx_cs->buf, ctx_cs->cdw * 4);
4786
4787 assert(ctx_cs->cdw <= ctx_cs->max_dw);
4788 assert(cs->cdw <= cs->max_dw);
4789 }
4790
4791 static void
4792 radv_pipeline_init_vertex_input_state(struct radv_pipeline *pipeline,
4793 const VkGraphicsPipelineCreateInfo *pCreateInfo)
4794 {
4795 const VkPipelineVertexInputStateCreateInfo *vi_info =
4796 pCreateInfo->pVertexInputState;
4797
4798 for (uint32_t i = 0; i < vi_info->vertexBindingDescriptionCount; i++) {
4799 const VkVertexInputBindingDescription *desc =
4800 &vi_info->pVertexBindingDescriptions[i];
4801
4802 pipeline->binding_stride[desc->binding] = desc->stride;
4803 pipeline->num_vertex_bindings =
4804 MAX2(pipeline->num_vertex_bindings, desc->binding + 1);
4805 }
4806 }
4807
4808 static struct radv_shader_variant *
4809 radv_pipeline_get_streamout_shader(struct radv_pipeline *pipeline)
4810 {
4811 int i;
4812
4813 for (i = MESA_SHADER_GEOMETRY; i >= MESA_SHADER_VERTEX; i--) {
4814 struct radv_shader_variant *shader =
4815 radv_get_shader(pipeline, i);
4816
4817 if (shader && shader->info.so.num_outputs > 0)
4818 return shader;
4819 }
4820
4821 return NULL;
4822 }
4823
4824 static void
4825 radv_pipeline_init_shader_stages_state(struct radv_pipeline *pipeline)
4826 {
4827 struct radv_device *device = pipeline->device;
4828
4829 for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
4830 pipeline->user_data_0[i] =
4831 radv_pipeline_stage_to_user_data_0(pipeline, i,
4832 device->physical_device->rad_info.chip_class);
4833
4834 if (pipeline->shaders[i]) {
4835 pipeline->need_indirect_descriptor_sets |= pipeline->shaders[i]->info.need_indirect_descriptor_sets;
4836 }
4837 }
4838
4839 struct radv_userdata_info *loc = radv_lookup_user_sgpr(pipeline, MESA_SHADER_VERTEX,
4840 AC_UD_VS_BASE_VERTEX_START_INSTANCE);
4841 if (loc->sgpr_idx != -1) {
4842 pipeline->graphics.vtx_base_sgpr = pipeline->user_data_0[MESA_SHADER_VERTEX];
4843 pipeline->graphics.vtx_base_sgpr += loc->sgpr_idx * 4;
4844 if (radv_get_shader(pipeline, MESA_SHADER_VERTEX)->info.vs.needs_draw_id)
4845 pipeline->graphics.vtx_emit_num = 3;
4846 else
4847 pipeline->graphics.vtx_emit_num = 2;
4848 }
4849 }
4850
4851 static VkResult
4852 radv_pipeline_init(struct radv_pipeline *pipeline,
4853 struct radv_device *device,
4854 struct radv_pipeline_cache *cache,
4855 const VkGraphicsPipelineCreateInfo *pCreateInfo,
4856 const struct radv_graphics_pipeline_create_info *extra)
4857 {
4858 VkResult result;
4859
4860 pipeline->device = device;
4861 pipeline->layout = radv_pipeline_layout_from_handle(pCreateInfo->layout);
4862 assert(pipeline->layout);
4863
4864 struct radv_blend_state blend = radv_pipeline_init_blend_state(pipeline, pCreateInfo, extra);
4865
4866 const VkPipelineCreationFeedbackCreateInfoEXT *creation_feedback =
4867 vk_find_struct_const(pCreateInfo->pNext, PIPELINE_CREATION_FEEDBACK_CREATE_INFO_EXT);
4868 radv_init_feedback(creation_feedback);
4869
4870 VkPipelineCreationFeedbackEXT *pipeline_feedback = creation_feedback ? creation_feedback->pPipelineCreationFeedback : NULL;
4871
4872 const VkPipelineShaderStageCreateInfo *pStages[MESA_SHADER_STAGES] = { 0, };
4873 VkPipelineCreationFeedbackEXT *stage_feedbacks[MESA_SHADER_STAGES] = { 0 };
4874 for (uint32_t i = 0; i < pCreateInfo->stageCount; i++) {
4875 gl_shader_stage stage = ffs(pCreateInfo->pStages[i].stage) - 1;
4876 pStages[stage] = &pCreateInfo->pStages[i];
4877 if(creation_feedback)
4878 stage_feedbacks[stage] = &creation_feedback->pPipelineStageCreationFeedbacks[i];
4879 }
4880
4881 struct radv_pipeline_key key = radv_generate_graphics_pipeline_key(pipeline, pCreateInfo, &blend);
4882
4883 result = radv_create_shaders(pipeline, device, cache, &key, pStages,
4884 pCreateInfo->flags, pipeline_feedback,
4885 stage_feedbacks);
4886 if (result != VK_SUCCESS)
4887 return result;
4888
4889 pipeline->graphics.spi_baryc_cntl = S_0286E0_FRONT_FACE_ALL_BITS(1);
4890 radv_pipeline_init_multisample_state(pipeline, &blend, pCreateInfo);
4891 radv_pipeline_init_input_assembly_state(pipeline, pCreateInfo, extra);
4892 radv_pipeline_init_dynamic_state(pipeline, pCreateInfo, extra);
4893 radv_pipeline_init_raster_state(pipeline, pCreateInfo);
4894 radv_pipeline_init_depth_stencil_state(pipeline, pCreateInfo);
4895
4896 /* Ensure that some export memory is always allocated, for two reasons:
4897 *
4898 * 1) Correctness: The hardware ignores the EXEC mask if no export
4899 * memory is allocated, so KILL and alpha test do not work correctly
4900 * without this.
4901 * 2) Performance: Every shader needs at least a NULL export, even when
4902 * it writes no color/depth output. The NULL export instruction
4903 * stalls without this setting.
4904 *
4905 * Don't add this to CB_SHADER_MASK.
4906 *
4907 * GFX10 supports pixel shaders without exports by setting both the
4908 * color and Z formats to SPI_SHADER_ZERO. The hw will skip export
4909 * instructions if any are present.
4910 */
4911 struct radv_shader_variant *ps = pipeline->shaders[MESA_SHADER_FRAGMENT];
4912 if ((pipeline->device->physical_device->rad_info.chip_class <= GFX9 ||
4913 ps->info.ps.can_discard) &&
4914 !blend.spi_shader_col_format) {
4915 if (!ps->info.ps.writes_z &&
4916 !ps->info.ps.writes_stencil &&
4917 !ps->info.ps.writes_sample_mask)
4918 blend.spi_shader_col_format = V_028714_SPI_SHADER_32_R;
4919 }
4920
4921 blend.cb_shader_mask = ps->info.ps.cb_shader_mask;
4922
4923 if (extra &&
4924 (extra->custom_blend_mode == V_028808_CB_ELIMINATE_FAST_CLEAR ||
4925 extra->custom_blend_mode == V_028808_CB_FMASK_DECOMPRESS ||
4926 extra->custom_blend_mode == V_028808_CB_DCC_DECOMPRESS ||
4927 extra->custom_blend_mode == V_028808_CB_RESOLVE)) {
4928 /* According to the CB spec states, CB_SHADER_MASK should be
4929 * set to enable writes to all four channels of MRT0.
4930 */
4931 blend.cb_shader_mask = 0xf;
4932 }
4933
4934 pipeline->graphics.col_format = blend.spi_shader_col_format;
4935 pipeline->graphics.cb_target_mask = blend.cb_target_mask;
4936
4937 if (radv_pipeline_has_gs(pipeline) && !radv_pipeline_has_ngg(pipeline)) {
4938 struct radv_shader_variant *gs =
4939 pipeline->shaders[MESA_SHADER_GEOMETRY];
4940
4941 radv_pipeline_init_gs_ring_state(pipeline, &gs->info.gs_ring_info);
4942 }
4943
4944 if (radv_pipeline_has_tess(pipeline)) {
4945 pipeline->graphics.tess_patch_control_points =
4946 pCreateInfo->pTessellationState->patchControlPoints;
4947 }
4948
4949 radv_pipeline_init_vertex_input_state(pipeline, pCreateInfo);
4950 radv_pipeline_init_binning_state(pipeline, pCreateInfo, &blend);
4951 radv_pipeline_init_shader_stages_state(pipeline);
4952 radv_pipeline_init_scratch(device, pipeline);
4953
4954 /* Find the last vertex shader stage that eventually uses streamout. */
4955 pipeline->streamout_shader = radv_pipeline_get_streamout_shader(pipeline);
4956
4957 radv_pipeline_generate_pm4(pipeline, pCreateInfo, extra, &blend);
4958
4959 return result;
4960 }
4961
4962 VkResult
4963 radv_graphics_pipeline_create(
4964 VkDevice _device,
4965 VkPipelineCache _cache,
4966 const VkGraphicsPipelineCreateInfo *pCreateInfo,
4967 const struct radv_graphics_pipeline_create_info *extra,
4968 const VkAllocationCallbacks *pAllocator,
4969 VkPipeline *pPipeline)
4970 {
4971 RADV_FROM_HANDLE(radv_device, device, _device);
4972 RADV_FROM_HANDLE(radv_pipeline_cache, cache, _cache);
4973 struct radv_pipeline *pipeline;
4974 VkResult result;
4975
4976 pipeline = vk_zalloc2(&device->vk.alloc, pAllocator, sizeof(*pipeline), 8,
4977 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
4978 if (pipeline == NULL)
4979 return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
4980
4981 vk_object_base_init(&device->vk, &pipeline->base,
4982 VK_OBJECT_TYPE_PIPELINE);
4983
4984 result = radv_pipeline_init(pipeline, device, cache,
4985 pCreateInfo, extra);
4986 if (result != VK_SUCCESS) {
4987 radv_pipeline_destroy(device, pipeline, pAllocator);
4988 return result;
4989 }
4990
4991 *pPipeline = radv_pipeline_to_handle(pipeline);
4992
4993 return VK_SUCCESS;
4994 }
4995
4996 VkResult radv_CreateGraphicsPipelines(
4997 VkDevice _device,
4998 VkPipelineCache pipelineCache,
4999 uint32_t count,
5000 const VkGraphicsPipelineCreateInfo* pCreateInfos,
5001 const VkAllocationCallbacks* pAllocator,
5002 VkPipeline* pPipelines)
5003 {
5004 VkResult result = VK_SUCCESS;
5005 unsigned i = 0;
5006
5007 for (; i < count; i++) {
5008 VkResult r;
5009 r = radv_graphics_pipeline_create(_device,
5010 pipelineCache,
5011 &pCreateInfos[i],
5012 NULL, pAllocator, &pPipelines[i]);
5013 if (r != VK_SUCCESS) {
5014 result = r;
5015 pPipelines[i] = VK_NULL_HANDLE;
5016
5017 if (pCreateInfos[i].flags & VK_PIPELINE_CREATE_EARLY_RETURN_ON_FAILURE_BIT_EXT)
5018 break;
5019 }
5020 }
5021
5022 for (; i < count; ++i)
5023 pPipelines[i] = VK_NULL_HANDLE;
5024
5025 return result;
5026 }
5027
5028 static void
5029 radv_pipeline_generate_hw_cs(struct radeon_cmdbuf *cs,
5030 const struct radv_pipeline *pipeline)
5031 {
5032 struct radv_shader_variant *shader = pipeline->shaders[MESA_SHADER_COMPUTE];
5033 uint64_t va = radv_buffer_get_va(shader->bo) + shader->bo_offset;
5034 struct radv_device *device = pipeline->device;
5035
5036 radeon_set_sh_reg_seq(cs, R_00B830_COMPUTE_PGM_LO, 2);
5037 radeon_emit(cs, va >> 8);
5038 radeon_emit(cs, S_00B834_DATA(va >> 40));
5039
5040 radeon_set_sh_reg_seq(cs, R_00B848_COMPUTE_PGM_RSRC1, 2);
5041 radeon_emit(cs, shader->config.rsrc1);
5042 radeon_emit(cs, shader->config.rsrc2);
5043 if (device->physical_device->rad_info.chip_class >= GFX10) {
5044 radeon_set_sh_reg(cs, R_00B8A0_COMPUTE_PGM_RSRC3, shader->config.rsrc3);
5045 }
5046 }
5047
5048 static void
5049 radv_pipeline_generate_compute_state(struct radeon_cmdbuf *cs,
5050 const struct radv_pipeline *pipeline)
5051 {
5052 struct radv_shader_variant *shader = pipeline->shaders[MESA_SHADER_COMPUTE];
5053 struct radv_device *device = pipeline->device;
5054 unsigned threads_per_threadgroup;
5055 unsigned threadgroups_per_cu = 1;
5056 unsigned waves_per_threadgroup;
5057 unsigned max_waves_per_sh = 0;
5058
5059 /* Calculate best compute resource limits. */
5060 threads_per_threadgroup = shader->info.cs.block_size[0] *
5061 shader->info.cs.block_size[1] *
5062 shader->info.cs.block_size[2];
5063 waves_per_threadgroup = DIV_ROUND_UP(threads_per_threadgroup,
5064 shader->info.wave_size);
5065
5066 if (device->physical_device->rad_info.chip_class >= GFX10 &&
5067 waves_per_threadgroup == 1)
5068 threadgroups_per_cu = 2;
5069
5070 radeon_set_sh_reg(cs, R_00B854_COMPUTE_RESOURCE_LIMITS,
5071 ac_get_compute_resource_limits(&device->physical_device->rad_info,
5072 waves_per_threadgroup,
5073 max_waves_per_sh,
5074 threadgroups_per_cu));
5075
5076 radeon_set_sh_reg_seq(cs, R_00B81C_COMPUTE_NUM_THREAD_X, 3);
5077 radeon_emit(cs, S_00B81C_NUM_THREAD_FULL(shader->info.cs.block_size[0]));
5078 radeon_emit(cs, S_00B81C_NUM_THREAD_FULL(shader->info.cs.block_size[1]));
5079 radeon_emit(cs, S_00B81C_NUM_THREAD_FULL(shader->info.cs.block_size[2]));
5080 }
5081
5082 static void
5083 radv_compute_generate_pm4(struct radv_pipeline *pipeline)
5084 {
5085 struct radv_device *device = pipeline->device;
5086 struct radeon_cmdbuf *cs = &pipeline->cs;
5087
5088 cs->max_dw = device->physical_device->rad_info.chip_class >= GFX10 ? 19 : 16;
5089 cs->buf = malloc(cs->max_dw * 4);
5090
5091 radv_pipeline_generate_hw_cs(cs, pipeline);
5092 radv_pipeline_generate_compute_state(cs, pipeline);
5093
5094 assert(pipeline->cs.cdw <= pipeline->cs.max_dw);
5095 }
5096
5097 static struct radv_pipeline_key
5098 radv_generate_compute_pipeline_key(struct radv_pipeline *pipeline,
5099 const VkComputePipelineCreateInfo *pCreateInfo)
5100 {
5101 const VkPipelineShaderStageCreateInfo *stage = &pCreateInfo->stage;
5102 struct radv_pipeline_key key;
5103 memset(&key, 0, sizeof(key));
5104
5105 if (pCreateInfo->flags & VK_PIPELINE_CREATE_DISABLE_OPTIMIZATION_BIT)
5106 key.optimisations_disabled = 1;
5107
5108 const VkPipelineShaderStageRequiredSubgroupSizeCreateInfoEXT *subgroup_size =
5109 vk_find_struct_const(stage->pNext,
5110 PIPELINE_SHADER_STAGE_REQUIRED_SUBGROUP_SIZE_CREATE_INFO_EXT);
5111
5112 if (subgroup_size) {
5113 assert(subgroup_size->requiredSubgroupSize == 32 ||
5114 subgroup_size->requiredSubgroupSize == 64);
5115 key.compute_subgroup_size = subgroup_size->requiredSubgroupSize;
5116 }
5117
5118 return key;
5119 }
5120
5121 static VkResult radv_compute_pipeline_create(
5122 VkDevice _device,
5123 VkPipelineCache _cache,
5124 const VkComputePipelineCreateInfo* pCreateInfo,
5125 const VkAllocationCallbacks* pAllocator,
5126 VkPipeline* pPipeline)
5127 {
5128 RADV_FROM_HANDLE(radv_device, device, _device);
5129 RADV_FROM_HANDLE(radv_pipeline_cache, cache, _cache);
5130 const VkPipelineShaderStageCreateInfo *pStages[MESA_SHADER_STAGES] = { 0, };
5131 VkPipelineCreationFeedbackEXT *stage_feedbacks[MESA_SHADER_STAGES] = { 0 };
5132 struct radv_pipeline *pipeline;
5133 VkResult result;
5134
5135 pipeline = vk_zalloc2(&device->vk.alloc, pAllocator, sizeof(*pipeline), 8,
5136 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
5137 if (pipeline == NULL)
5138 return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
5139
5140 vk_object_base_init(&device->vk, &pipeline->base,
5141 VK_OBJECT_TYPE_PIPELINE);
5142
5143 pipeline->device = device;
5144 pipeline->layout = radv_pipeline_layout_from_handle(pCreateInfo->layout);
5145 assert(pipeline->layout);
5146
5147 const VkPipelineCreationFeedbackCreateInfoEXT *creation_feedback =
5148 vk_find_struct_const(pCreateInfo->pNext, PIPELINE_CREATION_FEEDBACK_CREATE_INFO_EXT);
5149 radv_init_feedback(creation_feedback);
5150
5151 VkPipelineCreationFeedbackEXT *pipeline_feedback = creation_feedback ? creation_feedback->pPipelineCreationFeedback : NULL;
5152 if (creation_feedback)
5153 stage_feedbacks[MESA_SHADER_COMPUTE] = &creation_feedback->pPipelineStageCreationFeedbacks[0];
5154
5155 pStages[MESA_SHADER_COMPUTE] = &pCreateInfo->stage;
5156
5157 struct radv_pipeline_key key =
5158 radv_generate_compute_pipeline_key(pipeline, pCreateInfo);
5159
5160 result = radv_create_shaders(pipeline, device, cache, &key, pStages,
5161 pCreateInfo->flags, pipeline_feedback,
5162 stage_feedbacks);
5163 if (result != VK_SUCCESS) {
5164 radv_pipeline_destroy(device, pipeline, pAllocator);
5165 return result;
5166 }
5167
5168 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);
5169 pipeline->need_indirect_descriptor_sets |= pipeline->shaders[MESA_SHADER_COMPUTE]->info.need_indirect_descriptor_sets;
5170 radv_pipeline_init_scratch(device, pipeline);
5171
5172 radv_compute_generate_pm4(pipeline);
5173
5174 *pPipeline = radv_pipeline_to_handle(pipeline);
5175
5176 return VK_SUCCESS;
5177 }
5178
5179 VkResult radv_CreateComputePipelines(
5180 VkDevice _device,
5181 VkPipelineCache pipelineCache,
5182 uint32_t count,
5183 const VkComputePipelineCreateInfo* pCreateInfos,
5184 const VkAllocationCallbacks* pAllocator,
5185 VkPipeline* pPipelines)
5186 {
5187 VkResult result = VK_SUCCESS;
5188
5189 unsigned i = 0;
5190 for (; i < count; i++) {
5191 VkResult r;
5192 r = radv_compute_pipeline_create(_device, pipelineCache,
5193 &pCreateInfos[i],
5194 pAllocator, &pPipelines[i]);
5195 if (r != VK_SUCCESS) {
5196 result = r;
5197 pPipelines[i] = VK_NULL_HANDLE;
5198
5199 if (pCreateInfos[i].flags & VK_PIPELINE_CREATE_EARLY_RETURN_ON_FAILURE_BIT_EXT)
5200 break;
5201 }
5202 }
5203
5204 for (; i < count; ++i)
5205 pPipelines[i] = VK_NULL_HANDLE;
5206
5207 return result;
5208 }
5209
5210
5211 static uint32_t radv_get_executable_count(const struct radv_pipeline *pipeline)
5212 {
5213 uint32_t ret = 0;
5214 for (int i = 0; i < MESA_SHADER_STAGES; ++i) {
5215 if (!pipeline->shaders[i])
5216 continue;
5217
5218 if (i == MESA_SHADER_GEOMETRY &&
5219 !radv_pipeline_has_ngg(pipeline)) {
5220 ret += 2u;
5221 } else {
5222 ret += 1u;
5223 }
5224
5225 }
5226 return ret;
5227 }
5228
5229 static struct radv_shader_variant *
5230 radv_get_shader_from_executable_index(const struct radv_pipeline *pipeline, int index, gl_shader_stage *stage)
5231 {
5232 for (int i = 0; i < MESA_SHADER_STAGES; ++i) {
5233 if (!pipeline->shaders[i])
5234 continue;
5235 if (!index) {
5236 *stage = i;
5237 return pipeline->shaders[i];
5238 }
5239
5240 --index;
5241
5242 if (i == MESA_SHADER_GEOMETRY &&
5243 !radv_pipeline_has_ngg(pipeline)) {
5244 if (!index) {
5245 *stage = i;
5246 return pipeline->gs_copy_shader;
5247 }
5248 --index;
5249 }
5250 }
5251
5252 *stage = -1;
5253 return NULL;
5254 }
5255
5256 /* Basically strlcpy (which does not exist on linux) specialized for
5257 * descriptions. */
5258 static void desc_copy(char *desc, const char *src) {
5259 int len = strlen(src);
5260 assert(len < VK_MAX_DESCRIPTION_SIZE);
5261 memcpy(desc, src, len);
5262 memset(desc + len, 0, VK_MAX_DESCRIPTION_SIZE - len);
5263 }
5264
5265 VkResult radv_GetPipelineExecutablePropertiesKHR(
5266 VkDevice _device,
5267 const VkPipelineInfoKHR* pPipelineInfo,
5268 uint32_t* pExecutableCount,
5269 VkPipelineExecutablePropertiesKHR* pProperties)
5270 {
5271 RADV_FROM_HANDLE(radv_pipeline, pipeline, pPipelineInfo->pipeline);
5272 const uint32_t total_count = radv_get_executable_count(pipeline);
5273
5274 if (!pProperties) {
5275 *pExecutableCount = total_count;
5276 return VK_SUCCESS;
5277 }
5278
5279 const uint32_t count = MIN2(total_count, *pExecutableCount);
5280 for (unsigned i = 0, executable_idx = 0;
5281 i < MESA_SHADER_STAGES && executable_idx < count; ++i) {
5282 if (!pipeline->shaders[i])
5283 continue;
5284 pProperties[executable_idx].stages = mesa_to_vk_shader_stage(i);
5285 const char *name = NULL;
5286 const char *description = NULL;
5287 switch(i) {
5288 case MESA_SHADER_VERTEX:
5289 name = "Vertex Shader";
5290 description = "Vulkan Vertex Shader";
5291 break;
5292 case MESA_SHADER_TESS_CTRL:
5293 if (!pipeline->shaders[MESA_SHADER_VERTEX]) {
5294 pProperties[executable_idx].stages |= VK_SHADER_STAGE_VERTEX_BIT;
5295 name = "Vertex + Tessellation Control Shaders";
5296 description = "Combined Vulkan Vertex and Tessellation Control Shaders";
5297 } else {
5298 name = "Tessellation Control Shader";
5299 description = "Vulkan Tessellation Control Shader";
5300 }
5301 break;
5302 case MESA_SHADER_TESS_EVAL:
5303 name = "Tessellation Evaluation Shader";
5304 description = "Vulkan Tessellation Evaluation Shader";
5305 break;
5306 case MESA_SHADER_GEOMETRY:
5307 if (radv_pipeline_has_tess(pipeline) && !pipeline->shaders[MESA_SHADER_TESS_EVAL]) {
5308 pProperties[executable_idx].stages |= VK_SHADER_STAGE_TESSELLATION_EVALUATION_BIT;
5309 name = "Tessellation Evaluation + Geometry Shaders";
5310 description = "Combined Vulkan Tessellation Evaluation and Geometry Shaders";
5311 } else if (!radv_pipeline_has_tess(pipeline) && !pipeline->shaders[MESA_SHADER_VERTEX]) {
5312 pProperties[executable_idx].stages |= VK_SHADER_STAGE_VERTEX_BIT;
5313 name = "Vertex + Geometry Shader";
5314 description = "Combined Vulkan Vertex and Geometry Shaders";
5315 } else {
5316 name = "Geometry Shader";
5317 description = "Vulkan Geometry Shader";
5318 }
5319 break;
5320 case MESA_SHADER_FRAGMENT:
5321 name = "Fragment Shader";
5322 description = "Vulkan Fragment Shader";
5323 break;
5324 case MESA_SHADER_COMPUTE:
5325 name = "Compute Shader";
5326 description = "Vulkan Compute Shader";
5327 break;
5328 }
5329
5330 pProperties[executable_idx].subgroupSize = pipeline->shaders[i]->info.wave_size;
5331 desc_copy(pProperties[executable_idx].name, name);
5332 desc_copy(pProperties[executable_idx].description, description);
5333
5334 ++executable_idx;
5335 if (i == MESA_SHADER_GEOMETRY &&
5336 !radv_pipeline_has_ngg(pipeline)) {
5337 assert(pipeline->gs_copy_shader);
5338 if (executable_idx >= count)
5339 break;
5340
5341 pProperties[executable_idx].stages = VK_SHADER_STAGE_GEOMETRY_BIT;
5342 pProperties[executable_idx].subgroupSize = 64;
5343 desc_copy(pProperties[executable_idx].name, "GS Copy Shader");
5344 desc_copy(pProperties[executable_idx].description,
5345 "Extra shader stage that loads the GS output ringbuffer into the rasterizer");
5346
5347 ++executable_idx;
5348 }
5349 }
5350
5351 VkResult result = *pExecutableCount < total_count ? VK_INCOMPLETE : VK_SUCCESS;
5352 *pExecutableCount = count;
5353 return result;
5354 }
5355
5356 VkResult radv_GetPipelineExecutableStatisticsKHR(
5357 VkDevice _device,
5358 const VkPipelineExecutableInfoKHR* pExecutableInfo,
5359 uint32_t* pStatisticCount,
5360 VkPipelineExecutableStatisticKHR* pStatistics)
5361 {
5362 RADV_FROM_HANDLE(radv_device, device, _device);
5363 RADV_FROM_HANDLE(radv_pipeline, pipeline, pExecutableInfo->pipeline);
5364 gl_shader_stage stage;
5365 struct radv_shader_variant *shader = radv_get_shader_from_executable_index(pipeline, pExecutableInfo->executableIndex, &stage);
5366
5367 enum chip_class chip_class = device->physical_device->rad_info.chip_class;
5368 unsigned lds_increment = chip_class >= GFX7 ? 512 : 256;
5369 unsigned max_waves = radv_get_max_waves(device, shader, stage);
5370
5371 VkPipelineExecutableStatisticKHR *s = pStatistics;
5372 VkPipelineExecutableStatisticKHR *end = s + (pStatistics ? *pStatisticCount : 0);
5373 VkResult result = VK_SUCCESS;
5374
5375 if (s < end) {
5376 desc_copy(s->name, "SGPRs");
5377 desc_copy(s->description, "Number of SGPR registers allocated per subgroup");
5378 s->format = VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR;
5379 s->value.u64 = shader->config.num_sgprs;
5380 }
5381 ++s;
5382
5383 if (s < end) {
5384 desc_copy(s->name, "VGPRs");
5385 desc_copy(s->description, "Number of VGPR registers allocated per subgroup");
5386 s->format = VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR;
5387 s->value.u64 = shader->config.num_vgprs;
5388 }
5389 ++s;
5390
5391 if (s < end) {
5392 desc_copy(s->name, "Spilled SGPRs");
5393 desc_copy(s->description, "Number of SGPR registers spilled per subgroup");
5394 s->format = VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR;
5395 s->value.u64 = shader->config.spilled_sgprs;
5396 }
5397 ++s;
5398
5399 if (s < end) {
5400 desc_copy(s->name, "Spilled VGPRs");
5401 desc_copy(s->description, "Number of VGPR registers spilled per subgroup");
5402 s->format = VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR;
5403 s->value.u64 = shader->config.spilled_vgprs;
5404 }
5405 ++s;
5406
5407 if (s < end) {
5408 desc_copy(s->name, "PrivMem VGPRs");
5409 desc_copy(s->description, "Number of VGPRs stored in private memory per subgroup");
5410 s->format = VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR;
5411 s->value.u64 = shader->info.private_mem_vgprs;
5412 }
5413 ++s;
5414
5415 if (s < end) {
5416 desc_copy(s->name, "Code size");
5417 desc_copy(s->description, "Code size in bytes");
5418 s->format = VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR;
5419 s->value.u64 = shader->exec_size;
5420 }
5421 ++s;
5422
5423 if (s < end) {
5424 desc_copy(s->name, "LDS size");
5425 desc_copy(s->description, "LDS size in bytes per workgroup");
5426 s->format = VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR;
5427 s->value.u64 = shader->config.lds_size * lds_increment;
5428 }
5429 ++s;
5430
5431 if (s < end) {
5432 desc_copy(s->name, "Scratch size");
5433 desc_copy(s->description, "Private memory in bytes per subgroup");
5434 s->format = VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR;
5435 s->value.u64 = shader->config.scratch_bytes_per_wave;
5436 }
5437 ++s;
5438
5439 if (s < end) {
5440 desc_copy(s->name, "Subgroups per SIMD");
5441 desc_copy(s->description, "The maximum number of subgroups in flight on a SIMD unit");
5442 s->format = VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR;
5443 s->value.u64 = max_waves;
5444 }
5445 ++s;
5446
5447 if (shader->statistics) {
5448 for (unsigned i = 0; i < shader->statistics->count; i++) {
5449 struct radv_compiler_statistic_info *info = &shader->statistics->infos[i];
5450 uint32_t value = shader->statistics->values[i];
5451 if (s < end) {
5452 desc_copy(s->name, info->name);
5453 desc_copy(s->description, info->desc);
5454 s->format = VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR;
5455 s->value.u64 = value;
5456 }
5457 ++s;
5458 }
5459 }
5460
5461 if (!pStatistics)
5462 *pStatisticCount = s - pStatistics;
5463 else if (s > end) {
5464 *pStatisticCount = end - pStatistics;
5465 result = VK_INCOMPLETE;
5466 } else {
5467 *pStatisticCount = s - pStatistics;
5468 }
5469
5470 return result;
5471 }
5472
5473 static VkResult radv_copy_representation(void *data, size_t *data_size, const char *src)
5474 {
5475 size_t total_size = strlen(src) + 1;
5476
5477 if (!data) {
5478 *data_size = total_size;
5479 return VK_SUCCESS;
5480 }
5481
5482 size_t size = MIN2(total_size, *data_size);
5483
5484 memcpy(data, src, size);
5485 if (size)
5486 *((char*)data + size - 1) = 0;
5487 return size < total_size ? VK_INCOMPLETE : VK_SUCCESS;
5488 }
5489
5490 VkResult radv_GetPipelineExecutableInternalRepresentationsKHR(
5491 VkDevice device,
5492 const VkPipelineExecutableInfoKHR* pExecutableInfo,
5493 uint32_t* pInternalRepresentationCount,
5494 VkPipelineExecutableInternalRepresentationKHR* pInternalRepresentations)
5495 {
5496 RADV_FROM_HANDLE(radv_pipeline, pipeline, pExecutableInfo->pipeline);
5497 gl_shader_stage stage;
5498 struct radv_shader_variant *shader = radv_get_shader_from_executable_index(pipeline, pExecutableInfo->executableIndex, &stage);
5499
5500 VkPipelineExecutableInternalRepresentationKHR *p = pInternalRepresentations;
5501 VkPipelineExecutableInternalRepresentationKHR *end = p + (pInternalRepresentations ? *pInternalRepresentationCount : 0);
5502 VkResult result = VK_SUCCESS;
5503 /* optimized NIR */
5504 if (p < end) {
5505 p->isText = true;
5506 desc_copy(p->name, "NIR Shader(s)");
5507 desc_copy(p->description, "The optimized NIR shader(s)");
5508 if (radv_copy_representation(p->pData, &p->dataSize, shader->nir_string) != VK_SUCCESS)
5509 result = VK_INCOMPLETE;
5510 }
5511 ++p;
5512
5513 /* backend IR */
5514 if (p < end) {
5515 p->isText = true;
5516 if (pipeline->device->physical_device->use_llvm) {
5517 desc_copy(p->name, "LLVM IR");
5518 desc_copy(p->description, "The LLVM IR after some optimizations");
5519 } else {
5520 desc_copy(p->name, "ACO IR");
5521 desc_copy(p->description, "The ACO IR after some optimizations");
5522 }
5523 if (radv_copy_representation(p->pData, &p->dataSize, shader->ir_string) != VK_SUCCESS)
5524 result = VK_INCOMPLETE;
5525 }
5526 ++p;
5527
5528 /* Disassembler */
5529 if (p < end) {
5530 p->isText = true;
5531 desc_copy(p->name, "Assembly");
5532 desc_copy(p->description, "Final Assembly");
5533 if (radv_copy_representation(p->pData, &p->dataSize, shader->disasm_string) != VK_SUCCESS)
5534 result = VK_INCOMPLETE;
5535 }
5536 ++p;
5537
5538 if (!pInternalRepresentations)
5539 *pInternalRepresentationCount = p - pInternalRepresentations;
5540 else if(p > end) {
5541 result = VK_INCOMPLETE;
5542 *pInternalRepresentationCount = end - pInternalRepresentations;
5543 } else {
5544 *pInternalRepresentationCount = p - pInternalRepresentations;
5545 }
5546
5547 return result;
5548 }