projects / mesa.git / blob
? search:


8d1343d2d8bb6ec81638d0913f312d451cb9186a
[mesa.git] / src / amd / vulkan / radv_descriptor_set.c
1 /*
2 * Copyright © 2016 Red Hat.
3 * Copyright © 2016 Bas Nieuwenhuizen
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
14 * Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
21 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
22 * IN THE SOFTWARE.
23 */
24 #include <assert.h>
25 #include <stdbool.h>
26 #include <string.h>
27 #include <unistd.h>
28 #include <fcntl.h>
29
30 #include "util/mesa-sha1.h"
31 #include "radv_private.h"
32 #include "sid.h"
33 #include "vk_format.h"
34 #include "vk_util.h"
35
36
37 static bool has_equal_immutable_samplers(const VkSampler *samplers, uint32_t count)
38 {
39 if (!samplers)
40 return false;
41 for(uint32_t i = 1; i < count; ++i) {
42 if (memcmp(radv_sampler_from_handle(samplers[0])->state,
43 radv_sampler_from_handle(samplers[i])->state, 16)) {
44 return false;
45 }
46 }
47 return true;
48 }
49
50 static int binding_compare(const void* av, const void *bv)
51 {
52 const VkDescriptorSetLayoutBinding *a = (const VkDescriptorSetLayoutBinding*)av;
53 const VkDescriptorSetLayoutBinding *b = (const VkDescriptorSetLayoutBinding*)bv;
54
55 return (a->binding < b->binding) ? -1 : (a->binding > b->binding) ? 1 : 0;
56 }
57
58 static VkDescriptorSetLayoutBinding *
59 create_sorted_bindings(const VkDescriptorSetLayoutBinding *bindings, unsigned count) {
60 VkDescriptorSetLayoutBinding *sorted_bindings = malloc(count * sizeof(VkDescriptorSetLayoutBinding));
61 if (!sorted_bindings)
62 return NULL;
63
64 memcpy(sorted_bindings, bindings, count * sizeof(VkDescriptorSetLayoutBinding));
65
66 qsort(sorted_bindings, count, sizeof(VkDescriptorSetLayoutBinding), binding_compare);
67
68 return sorted_bindings;
69 }
70
71 VkResult radv_CreateDescriptorSetLayout(
72 VkDevice _device,
73 const VkDescriptorSetLayoutCreateInfo* pCreateInfo,
74 const VkAllocationCallbacks* pAllocator,
75 VkDescriptorSetLayout* pSetLayout)
76 {
77 RADV_FROM_HANDLE(radv_device, device, _device);
78 struct radv_descriptor_set_layout *set_layout;
79
80 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO);
81 const VkDescriptorSetLayoutBindingFlagsCreateInfo *variable_flags =
82 vk_find_struct_const(pCreateInfo->pNext, DESCRIPTOR_SET_LAYOUT_BINDING_FLAGS_CREATE_INFO);
83
84 uint32_t max_binding = 0;
85 uint32_t immutable_sampler_count = 0;
86 uint32_t ycbcr_sampler_count = 0;
87 for (uint32_t j = 0; j < pCreateInfo->bindingCount; j++) {
88 max_binding = MAX2(max_binding, pCreateInfo->pBindings[j].binding);
89 if ((pCreateInfo->pBindings[j].descriptorType == VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER ||
90 pCreateInfo->pBindings[j].descriptorType == VK_DESCRIPTOR_TYPE_SAMPLER) &&
91 pCreateInfo->pBindings[j].pImmutableSamplers) {
92 immutable_sampler_count += pCreateInfo->pBindings[j].descriptorCount;
93
94 bool has_ycbcr_sampler = false;
95 for (unsigned i = 0; i < pCreateInfo->pBindings[j].descriptorCount; ++i) {
96 if (radv_sampler_from_handle(pCreateInfo->pBindings[j].pImmutableSamplers[i])->ycbcr_sampler)
97 has_ycbcr_sampler = true;
98 }
99
100 if (has_ycbcr_sampler)
101 ycbcr_sampler_count += pCreateInfo->pBindings[j].descriptorCount;
102 }
103 }
104
105 uint32_t samplers_offset = sizeof(struct radv_descriptor_set_layout) +
106 (max_binding + 1) * sizeof(set_layout->binding[0]);
107 size_t size = samplers_offset + immutable_sampler_count * 4 * sizeof(uint32_t);
108 if (ycbcr_sampler_count > 0) {
109 size += ycbcr_sampler_count * sizeof(struct radv_sampler_ycbcr_conversion) + (max_binding + 1) * sizeof(uint32_t);
110 }
111
112 set_layout = vk_zalloc2(&device->vk.alloc, pAllocator, size, 8,
113 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
114 if (!set_layout)
115 return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
116
117 vk_object_base_init(&device->vk, &set_layout->base,
118 VK_OBJECT_TYPE_DESCRIPTOR_SET_LAYOUT);
119
120 set_layout->flags = pCreateInfo->flags;
121 set_layout->layout_size = size;
122
123 /* We just allocate all the samplers at the end of the struct */
124 uint32_t *samplers = (uint32_t*)&set_layout->binding[max_binding + 1];
125 struct radv_sampler_ycbcr_conversion *ycbcr_samplers = NULL;
126 uint32_t *ycbcr_sampler_offsets = NULL;
127
128 if (ycbcr_sampler_count > 0) {
129 ycbcr_sampler_offsets = samplers + 4 * immutable_sampler_count;
130 set_layout->ycbcr_sampler_offsets_offset = (char*)ycbcr_sampler_offsets - (char*)set_layout;
131 ycbcr_samplers = (struct radv_sampler_ycbcr_conversion *)(ycbcr_sampler_offsets + max_binding + 1);
132 } else
133 set_layout->ycbcr_sampler_offsets_offset = 0;
134
135 VkDescriptorSetLayoutBinding *bindings = create_sorted_bindings(pCreateInfo->pBindings,
136 pCreateInfo->bindingCount);
137 if (!bindings) {
138 vk_object_base_finish(&set_layout->base);
139 vk_free2(&device->vk.alloc, pAllocator, set_layout);
140 return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
141 }
142
143 set_layout->binding_count = max_binding + 1;
144 set_layout->shader_stages = 0;
145 set_layout->dynamic_shader_stages = 0;
146 set_layout->has_immutable_samplers = false;
147 set_layout->size = 0;
148
149 memset(set_layout->binding, 0, size - sizeof(struct radv_descriptor_set_layout));
150
151 uint32_t buffer_count = 0;
152 uint32_t dynamic_offset_count = 0;
153
154 for (uint32_t j = 0; j < pCreateInfo->bindingCount; j++) {
155 const VkDescriptorSetLayoutBinding *binding = bindings + j;
156 uint32_t b = binding->binding;
157 uint32_t alignment = 0;
158 unsigned binding_buffer_count = 0;
159 uint32_t descriptor_count = binding->descriptorCount;
160 bool has_ycbcr_sampler = false;
161
162 /* main image + fmask */
163 uint32_t max_sampled_image_descriptors = 2;
164
165 if (binding->descriptorType == VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER &&
166 binding->pImmutableSamplers) {
167 for (unsigned i = 0; i < binding->descriptorCount; ++i) {
168 struct radv_sampler_ycbcr_conversion *conversion =
169 radv_sampler_from_handle(binding->pImmutableSamplers[i])->ycbcr_sampler;
170
171 if (conversion) {
172 has_ycbcr_sampler = true;
173 max_sampled_image_descriptors = MAX2(max_sampled_image_descriptors,
174 vk_format_get_plane_count(conversion->format));
175 }
176 }
177 }
178
179 switch (binding->descriptorType) {
180 case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
181 case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
182 assert(!(pCreateInfo->flags & VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR));
183 set_layout->binding[b].dynamic_offset_count = 1;
184 set_layout->dynamic_shader_stages |= binding->stageFlags;
185 set_layout->binding[b].size = 0;
186 binding_buffer_count = 1;
187 alignment = 1;
188 break;
189 case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
190 case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
191 case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
192 case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
193 set_layout->binding[b].size = 16;
194 binding_buffer_count = 1;
195 alignment = 16;
196 break;
197 case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE:
198 case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE:
199 case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT:
200 /* main descriptor + fmask descriptor */
201 set_layout->binding[b].size = 64;
202 binding_buffer_count = 1;
203 alignment = 32;
204 break;
205 case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
206 /* main descriptor + fmask descriptor + sampler */
207 set_layout->binding[b].size = 96;
208 binding_buffer_count = 1;
209 alignment = 32;
210 break;
211 case VK_DESCRIPTOR_TYPE_SAMPLER:
212 set_layout->binding[b].size = 16;
213 alignment = 16;
214 break;
215 case VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT:
216 alignment = 16;
217 set_layout->binding[b].size = descriptor_count;
218 descriptor_count = 1;
219 break;
220 default:
221 break;
222 }
223
224 set_layout->size = align(set_layout->size, alignment);
225 set_layout->binding[b].type = binding->descriptorType;
226 set_layout->binding[b].array_size = descriptor_count;
227 set_layout->binding[b].offset = set_layout->size;
228 set_layout->binding[b].buffer_offset = buffer_count;
229 set_layout->binding[b].dynamic_offset_offset = dynamic_offset_count;
230
231 if (variable_flags && binding->binding < variable_flags->bindingCount &&
232 (variable_flags->pBindingFlags[binding->binding] & VK_DESCRIPTOR_BINDING_VARIABLE_DESCRIPTOR_COUNT_BIT_EXT)) {
233 assert(!binding->pImmutableSamplers); /* Terribly ill defined how many samplers are valid */
234 assert(binding->binding == max_binding);
235
236 set_layout->has_variable_descriptors = true;
237 }
238
239 if ((binding->descriptorType == VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER ||
240 binding->descriptorType == VK_DESCRIPTOR_TYPE_SAMPLER) &&
241 binding->pImmutableSamplers) {
242 set_layout->binding[b].immutable_samplers_offset = samplers_offset;
243 set_layout->binding[b].immutable_samplers_equal =
244 has_equal_immutable_samplers(binding->pImmutableSamplers, binding->descriptorCount);
245 set_layout->has_immutable_samplers = true;
246
247
248 for (uint32_t i = 0; i < binding->descriptorCount; i++)
249 memcpy(samplers + 4 * i, &radv_sampler_from_handle(binding->pImmutableSamplers[i])->state, 16);
250
251 /* Don't reserve space for the samplers if they're not accessed. */
252 if (set_layout->binding[b].immutable_samplers_equal) {
253 if (binding->descriptorType == VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER &&
254 max_sampled_image_descriptors <= 2)
255 set_layout->binding[b].size -= 32;
256 else if (binding->descriptorType == VK_DESCRIPTOR_TYPE_SAMPLER)
257 set_layout->binding[b].size -= 16;
258 }
259 samplers += 4 * binding->descriptorCount;
260 samplers_offset += 4 * sizeof(uint32_t) * binding->descriptorCount;
261
262 if (has_ycbcr_sampler) {
263 ycbcr_sampler_offsets[b] = (const char*)ycbcr_samplers - (const char*)set_layout;
264 for (uint32_t i = 0; i < binding->descriptorCount; i++) {
265 if (radv_sampler_from_handle(binding->pImmutableSamplers[i])->ycbcr_sampler)
266 ycbcr_samplers[i] = *radv_sampler_from_handle(binding->pImmutableSamplers[i])->ycbcr_sampler;
267 else
268 ycbcr_samplers[i].format = VK_FORMAT_UNDEFINED;
269 }
270 ycbcr_samplers += binding->descriptorCount;
271 }
272 }
273
274 set_layout->size += descriptor_count * set_layout->binding[b].size;
275 buffer_count += descriptor_count * binding_buffer_count;
276 dynamic_offset_count += descriptor_count *
277 set_layout->binding[b].dynamic_offset_count;
278 set_layout->shader_stages |= binding->stageFlags;
279 }
280
281 free(bindings);
282
283 set_layout->buffer_count = buffer_count;
284 set_layout->dynamic_offset_count = dynamic_offset_count;
285
286 *pSetLayout = radv_descriptor_set_layout_to_handle(set_layout);
287
288 return VK_SUCCESS;
289 }
290
291 void radv_DestroyDescriptorSetLayout(
292 VkDevice _device,
293 VkDescriptorSetLayout _set_layout,
294 const VkAllocationCallbacks* pAllocator)
295 {
296 RADV_FROM_HANDLE(radv_device, device, _device);
297 RADV_FROM_HANDLE(radv_descriptor_set_layout, set_layout, _set_layout);
298
299 if (!set_layout)
300 return;
301
302 vk_object_base_finish(&set_layout->base);
303 vk_free2(&device->vk.alloc, pAllocator, set_layout);
304 }
305
306 void radv_GetDescriptorSetLayoutSupport(VkDevice device,
307 const VkDescriptorSetLayoutCreateInfo* pCreateInfo,
308 VkDescriptorSetLayoutSupport* pSupport)
309 {
310 VkDescriptorSetLayoutBinding *bindings = create_sorted_bindings(pCreateInfo->pBindings,
311 pCreateInfo->bindingCount);
312 if (!bindings) {
313 pSupport->supported = false;
314 return;
315 }
316
317 const VkDescriptorSetLayoutBindingFlagsCreateInfo *variable_flags =
318 vk_find_struct_const(pCreateInfo->pNext, DESCRIPTOR_SET_LAYOUT_BINDING_FLAGS_CREATE_INFO);
319 VkDescriptorSetVariableDescriptorCountLayoutSupport *variable_count =
320 vk_find_struct((void*)pCreateInfo->pNext, DESCRIPTOR_SET_VARIABLE_DESCRIPTOR_COUNT_LAYOUT_SUPPORT);
321 if (variable_count) {
322 variable_count->maxVariableDescriptorCount = 0;
323 }
324
325 bool supported = true;
326 uint64_t size = 0;
327 for (uint32_t i = 0; i < pCreateInfo->bindingCount; i++) {
328 const VkDescriptorSetLayoutBinding *binding = bindings + i;
329
330 uint64_t descriptor_size = 0;
331 uint64_t descriptor_alignment = 1;
332 uint32_t descriptor_count = binding->descriptorCount;
333 switch (binding->descriptorType) {
334 case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
335 case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
336 break;
337 case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
338 case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
339 case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
340 case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
341 descriptor_size = 16;
342 descriptor_alignment = 16;
343 break;
344 case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE:
345 case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE:
346 case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT:
347 descriptor_size = 64;
348 descriptor_alignment = 32;
349 break;
350 case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
351 if (!has_equal_immutable_samplers(binding->pImmutableSamplers, descriptor_count)) {
352 descriptor_size = 64;
353 } else {
354 descriptor_size = 96;
355 }
356 descriptor_alignment = 32;
357 break;
358 case VK_DESCRIPTOR_TYPE_SAMPLER:
359 if (!has_equal_immutable_samplers(binding->pImmutableSamplers, descriptor_count)) {
360 descriptor_size = 16;
361 descriptor_alignment = 16;
362 }
363 break;
364 case VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT:
365 descriptor_alignment = 16;
366 descriptor_size = descriptor_count;
367 descriptor_count = 1;
368 break;
369 default:
370 break;
371 }
372
373 if (size && !align_u64(size, descriptor_alignment)) {
374 supported = false;
375 }
376 size = align_u64(size, descriptor_alignment);
377
378 uint64_t max_count = INT32_MAX;
379 if (binding->descriptorType == VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT)
380 max_count = INT32_MAX - size;
381 else if (descriptor_size)
382 max_count = (INT32_MAX - size) / descriptor_size;
383
384 if (max_count < descriptor_count) {
385 supported = false;
386 }
387 if (variable_flags && binding->binding <variable_flags->bindingCount && variable_count &&
388 (variable_flags->pBindingFlags[binding->binding] & VK_DESCRIPTOR_BINDING_VARIABLE_DESCRIPTOR_COUNT_BIT_EXT)) {
389 variable_count->maxVariableDescriptorCount = MIN2(UINT32_MAX, max_count);
390 }
391 size += descriptor_count * descriptor_size;
392 }
393
394 free(bindings);
395
396 pSupport->supported = supported;
397 }
398
399 /*
400 * Pipeline layouts. These have nothing to do with the pipeline. They are
401 * just multiple descriptor set layouts pasted together.
402 */
403
404 VkResult radv_CreatePipelineLayout(
405 VkDevice _device,
406 const VkPipelineLayoutCreateInfo* pCreateInfo,
407 const VkAllocationCallbacks* pAllocator,
408 VkPipelineLayout* pPipelineLayout)
409 {
410 RADV_FROM_HANDLE(radv_device, device, _device);
411 struct radv_pipeline_layout *layout;
412 struct mesa_sha1 ctx;
413
414 assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO);
415
416 layout = vk_alloc2(&device->vk.alloc, pAllocator, sizeof(*layout), 8,
417 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
418 if (layout == NULL)
419 return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
420
421 vk_object_base_init(&device->vk, &layout->base,
422 VK_OBJECT_TYPE_PIPELINE_LAYOUT);
423
424 layout->num_sets = pCreateInfo->setLayoutCount;
425
426 unsigned dynamic_offset_count = 0;
427 uint16_t dynamic_shader_stages = 0;
428
429
430 _mesa_sha1_init(&ctx);
431 for (uint32_t set = 0; set < pCreateInfo->setLayoutCount; set++) {
432 RADV_FROM_HANDLE(radv_descriptor_set_layout, set_layout,
433 pCreateInfo->pSetLayouts[set]);
434 layout->set[set].layout = set_layout;
435
436 layout->set[set].dynamic_offset_start = dynamic_offset_count;
437 for (uint32_t b = 0; b < set_layout->binding_count; b++) {
438 dynamic_offset_count += set_layout->binding[b].array_size * set_layout->binding[b].dynamic_offset_count;
439 dynamic_shader_stages |= set_layout->dynamic_shader_stages;
440 }
441 _mesa_sha1_update(&ctx, set_layout, set_layout->layout_size);
442 }
443
444 layout->dynamic_offset_count = dynamic_offset_count;
445 layout->dynamic_shader_stages = dynamic_shader_stages;
446 layout->push_constant_size = 0;
447
448 for (unsigned i = 0; i < pCreateInfo->pushConstantRangeCount; ++i) {
449 const VkPushConstantRange *range = pCreateInfo->pPushConstantRanges + i;
450 layout->push_constant_size = MAX2(layout->push_constant_size,
451 range->offset + range->size);
452 }
453
454 layout->push_constant_size = align(layout->push_constant_size, 16);
455 _mesa_sha1_update(&ctx, &layout->push_constant_size,
456 sizeof(layout->push_constant_size));
457 _mesa_sha1_final(&ctx, layout->sha1);
458 *pPipelineLayout = radv_pipeline_layout_to_handle(layout);
459
460 return VK_SUCCESS;
461 }
462
463 void radv_DestroyPipelineLayout(
464 VkDevice _device,
465 VkPipelineLayout _pipelineLayout,
466 const VkAllocationCallbacks* pAllocator)
467 {
468 RADV_FROM_HANDLE(radv_device, device, _device);
469 RADV_FROM_HANDLE(radv_pipeline_layout, pipeline_layout, _pipelineLayout);
470
471 if (!pipeline_layout)
472 return;
473
474 vk_object_base_finish(&pipeline_layout->base);
475 vk_free2(&device->vk.alloc, pAllocator, pipeline_layout);
476 }
477
478 #define EMPTY 1
479
480 static VkResult
481 radv_descriptor_set_create(struct radv_device *device,
482 struct radv_descriptor_pool *pool,
483 const struct radv_descriptor_set_layout *layout,
484 const uint32_t *variable_count,
485 struct radv_descriptor_set **out_set)
486 {
487 struct radv_descriptor_set *set;
488 uint32_t buffer_count = layout->buffer_count;
489 if (variable_count) {
490 unsigned stride = 1;
491 if (layout->binding[layout->binding_count - 1].type == VK_DESCRIPTOR_TYPE_SAMPLER ||
492 layout->binding[layout->binding_count - 1].type == VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT)
493 stride = 0;
494 buffer_count = layout->binding[layout->binding_count - 1].buffer_offset +
495 *variable_count * stride;
496 }
497 unsigned range_offset = sizeof(struct radv_descriptor_set) +
498 sizeof(struct radeon_winsys_bo *) * buffer_count;
499 unsigned mem_size = range_offset +
500 sizeof(struct radv_descriptor_range) * layout->dynamic_offset_count;
501
502 if (pool->host_memory_base) {
503 if (pool->host_memory_end - pool->host_memory_ptr < mem_size)
504 return vk_error(device->instance, VK_ERROR_OUT_OF_POOL_MEMORY);
505
506 set = (struct radv_descriptor_set*)pool->host_memory_ptr;
507 pool->host_memory_ptr += mem_size;
508 } else {
509 set = vk_alloc2(&device->vk.alloc, NULL, mem_size, 8,
510 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
511
512 if (!set)
513 return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
514 }
515
516 memset(set, 0, mem_size);
517
518 vk_object_base_init(&device->vk, &set->base,
519 VK_OBJECT_TYPE_DESCRIPTOR_SET);
520
521 if (layout->dynamic_offset_count) {
522 set->dynamic_descriptors = (struct radv_descriptor_range*)((uint8_t*)set + range_offset);
523 }
524
525 set->layout = layout;
526 set->buffer_count = buffer_count;
527 uint32_t layout_size = layout->size;
528 if (variable_count) {
529 assert(layout->has_variable_descriptors);
530 uint32_t stride = layout->binding[layout->binding_count - 1].size;
531 if (layout->binding[layout->binding_count - 1].type == VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT)
532 stride = 1;
533
534 layout_size = layout->binding[layout->binding_count - 1].offset +
535 *variable_count * stride;
536 }
537 layout_size = align_u32(layout_size, 32);
538 if (layout_size) {
539 set->size = layout_size;
540
541 if (!pool->host_memory_base && pool->entry_count == pool->max_entry_count) {
542 vk_free2(&device->vk.alloc, NULL, set);
543 return vk_error(device->instance, VK_ERROR_OUT_OF_POOL_MEMORY);
544 }
545
546 /* try to allocate linearly first, so that we don't spend
547 * time looking for gaps if the app only allocates &
548 * resets via the pool. */
549 if (pool->current_offset + layout_size <= pool->size) {
550 set->bo = pool->bo;
551 set->mapped_ptr = (uint32_t*)(pool->mapped_ptr + pool->current_offset);
552 set->va = radv_buffer_get_va(set->bo) + pool->current_offset;
553 if (!pool->host_memory_base) {
554 pool->entries[pool->entry_count].offset = pool->current_offset;
555 pool->entries[pool->entry_count].size = layout_size;
556 pool->entries[pool->entry_count].set = set;
557 pool->entry_count++;
558 }
559 pool->current_offset += layout_size;
560 } else if (!pool->host_memory_base) {
561 uint64_t offset = 0;
562 int index;
563
564 for (index = 0; index < pool->entry_count; ++index) {
565 if (pool->entries[index].offset - offset >= layout_size)
566 break;
567 offset = pool->entries[index].offset + pool->entries[index].size;
568 }
569
570 if (pool->size - offset < layout_size) {
571 vk_free2(&device->vk.alloc, NULL, set);
572 return vk_error(device->instance, VK_ERROR_OUT_OF_POOL_MEMORY);
573 }
574 set->bo = pool->bo;
575 set->mapped_ptr = (uint32_t*)(pool->mapped_ptr + offset);
576 set->va = radv_buffer_get_va(set->bo) + offset;
577 memmove(&pool->entries[index + 1], &pool->entries[index],
578 sizeof(pool->entries[0]) * (pool->entry_count - index));
579 pool->entries[index].offset = offset;
580 pool->entries[index].size = layout_size;
581 pool->entries[index].set = set;
582 pool->entry_count++;
583 } else
584 return vk_error(device->instance, VK_ERROR_OUT_OF_POOL_MEMORY);
585 }
586
587 if (layout->has_immutable_samplers) {
588 for (unsigned i = 0; i < layout->binding_count; ++i) {
589 if (!layout->binding[i].immutable_samplers_offset ||
590 layout->binding[i].immutable_samplers_equal)
591 continue;
592
593 unsigned offset = layout->binding[i].offset / 4;
594 if (layout->binding[i].type == VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER)
595 offset += radv_combined_image_descriptor_sampler_offset(layout->binding + i) / 4;
596
597 const uint32_t *samplers = (const uint32_t*)((const char*)layout + layout->binding[i].immutable_samplers_offset);
598 for (unsigned j = 0; j < layout->binding[i].array_size; ++j) {
599 memcpy(set->mapped_ptr + offset, samplers + 4 * j, 16);
600 offset += layout->binding[i].size / 4;
601 }
602
603 }
604 }
605 *out_set = set;
606 return VK_SUCCESS;
607 }
608
609 static void
610 radv_descriptor_set_destroy(struct radv_device *device,
611 struct radv_descriptor_pool *pool,
612 struct radv_descriptor_set *set,
613 bool free_bo)
614 {
615 assert(!pool->host_memory_base);
616
617 if (free_bo && set->size && !pool->host_memory_base) {
618 uint32_t offset = (uint8_t*)set->mapped_ptr - pool->mapped_ptr;
619 for (int i = 0; i < pool->entry_count; ++i) {
620 if (pool->entries[i].offset == offset) {
621 memmove(&pool->entries[i], &pool->entries[i+1],
622 sizeof(pool->entries[i]) * (pool->entry_count - i - 1));
623 --pool->entry_count;
624 break;
625 }
626 }
627 }
628 vk_object_base_finish(&set->base);
629 vk_free2(&device->vk.alloc, NULL, set);
630 }
631
632 VkResult radv_CreateDescriptorPool(
633 VkDevice _device,
634 const VkDescriptorPoolCreateInfo* pCreateInfo,
635 const VkAllocationCallbacks* pAllocator,
636 VkDescriptorPool* pDescriptorPool)
637 {
638 RADV_FROM_HANDLE(radv_device, device, _device);
639 struct radv_descriptor_pool *pool;
640 uint64_t size = sizeof(struct radv_descriptor_pool);
641 uint64_t bo_size = 0, bo_count = 0, range_count = 0;
642
643 vk_foreach_struct(ext, pCreateInfo->pNext) {
644 switch (ext->sType) {
645 case VK_STRUCTURE_TYPE_DESCRIPTOR_POOL_INLINE_UNIFORM_BLOCK_CREATE_INFO_EXT: {
646 const struct VkDescriptorPoolInlineUniformBlockCreateInfoEXT *info =
647 (const struct VkDescriptorPoolInlineUniformBlockCreateInfoEXT*)ext;
648 /* the sizes are 4 aligned, and we need to align to at
649 * most 32, which needs at most 28 bytes extra per
650 * binding. */
651 bo_size += 28llu * info->maxInlineUniformBlockBindings;
652 break;
653 }
654 default:
655 break;
656 }
657 }
658
659 for (unsigned i = 0; i < pCreateInfo->poolSizeCount; ++i) {
660 if (pCreateInfo->pPoolSizes[i].type != VK_DESCRIPTOR_TYPE_SAMPLER)
661 bo_count += pCreateInfo->pPoolSizes[i].descriptorCount;
662
663 switch(pCreateInfo->pPoolSizes[i].type) {
664 case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
665 case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
666 range_count += pCreateInfo->pPoolSizes[i].descriptorCount;
667 break;
668 case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
669 case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
670 case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
671 case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
672 case VK_DESCRIPTOR_TYPE_SAMPLER:
673 /* 32 as we may need to align for images */
674 bo_size += 32 * pCreateInfo->pPoolSizes[i].descriptorCount;
675 break;
676 case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE:
677 case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE:
678 case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT:
679 bo_size += 64 * pCreateInfo->pPoolSizes[i].descriptorCount;
680 break;
681 case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
682 bo_size += 96 * pCreateInfo->pPoolSizes[i].descriptorCount;
683 break;
684 case VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT:
685 bo_size += pCreateInfo->pPoolSizes[i].descriptorCount;
686 break;
687 default:
688 break;
689 }
690 }
691
692 if (!(pCreateInfo->flags & VK_DESCRIPTOR_POOL_CREATE_FREE_DESCRIPTOR_SET_BIT)) {
693 uint64_t host_size = pCreateInfo->maxSets * sizeof(struct radv_descriptor_set);
694 host_size += sizeof(struct radeon_winsys_bo*) * bo_count;
695 host_size += sizeof(struct radv_descriptor_range) * range_count;
696 size += host_size;
697 } else {
698 size += sizeof(struct radv_descriptor_pool_entry) * pCreateInfo->maxSets;
699 }
700
701 pool = vk_alloc2(&device->vk.alloc, pAllocator, size, 8,
702 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
703 if (!pool)
704 return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
705
706 memset(pool, 0, sizeof(*pool));
707
708 vk_object_base_init(&device->vk, &pool->base,
709 VK_OBJECT_TYPE_DESCRIPTOR_POOL);
710
711 if (!(pCreateInfo->flags & VK_DESCRIPTOR_POOL_CREATE_FREE_DESCRIPTOR_SET_BIT)) {
712 pool->host_memory_base = (uint8_t*)pool + sizeof(struct radv_descriptor_pool);
713 pool->host_memory_ptr = pool->host_memory_base;
714 pool->host_memory_end = (uint8_t*)pool + size;
715 }
716
717 if (bo_size) {
718 pool->bo = device->ws->buffer_create(device->ws, bo_size, 32,
719 RADEON_DOMAIN_VRAM,
720 RADEON_FLAG_NO_INTERPROCESS_SHARING |
721 RADEON_FLAG_READ_ONLY |
722 RADEON_FLAG_32BIT,
723 RADV_BO_PRIORITY_DESCRIPTOR);
724 pool->mapped_ptr = (uint8_t*)device->ws->buffer_map(pool->bo);
725 }
726 pool->size = bo_size;
727 pool->max_entry_count = pCreateInfo->maxSets;
728
729 *pDescriptorPool = radv_descriptor_pool_to_handle(pool);
730 return VK_SUCCESS;
731 }
732
733 void radv_DestroyDescriptorPool(
734 VkDevice _device,
735 VkDescriptorPool _pool,
736 const VkAllocationCallbacks* pAllocator)
737 {
738 RADV_FROM_HANDLE(radv_device, device, _device);
739 RADV_FROM_HANDLE(radv_descriptor_pool, pool, _pool);
740
741 if (!pool)
742 return;
743
744 if (!pool->host_memory_base) {
745 for(int i = 0; i < pool->entry_count; ++i) {
746 radv_descriptor_set_destroy(device, pool, pool->entries[i].set, false);
747 }
748 }
749
750 if (pool->bo)
751 device->ws->buffer_destroy(pool->bo);
752
753 vk_object_base_finish(&pool->base);
754 vk_free2(&device->vk.alloc, pAllocator, pool);
755 }
756
757 VkResult radv_ResetDescriptorPool(
758 VkDevice _device,
759 VkDescriptorPool descriptorPool,
760 VkDescriptorPoolResetFlags flags)
761 {
762 RADV_FROM_HANDLE(radv_device, device, _device);
763 RADV_FROM_HANDLE(radv_descriptor_pool, pool, descriptorPool);
764
765 if (!pool->host_memory_base) {
766 for(int i = 0; i < pool->entry_count; ++i) {
767 radv_descriptor_set_destroy(device, pool, pool->entries[i].set, false);
768 }
769 pool->entry_count = 0;
770 }
771
772 pool->current_offset = 0;
773 pool->host_memory_ptr = pool->host_memory_base;
774
775 return VK_SUCCESS;
776 }
777
778 VkResult radv_AllocateDescriptorSets(
779 VkDevice _device,
780 const VkDescriptorSetAllocateInfo* pAllocateInfo,
781 VkDescriptorSet* pDescriptorSets)
782 {
783 RADV_FROM_HANDLE(radv_device, device, _device);
784 RADV_FROM_HANDLE(radv_descriptor_pool, pool, pAllocateInfo->descriptorPool);
785
786 VkResult result = VK_SUCCESS;
787 uint32_t i;
788 struct radv_descriptor_set *set = NULL;
789
790 const VkDescriptorSetVariableDescriptorCountAllocateInfo *variable_counts =
791 vk_find_struct_const(pAllocateInfo->pNext, DESCRIPTOR_SET_VARIABLE_DESCRIPTOR_COUNT_ALLOCATE_INFO);
792 const uint32_t zero = 0;
793
794 /* allocate a set of buffers for each shader to contain descriptors */
795 for (i = 0; i < pAllocateInfo->descriptorSetCount; i++) {
796 RADV_FROM_HANDLE(radv_descriptor_set_layout, layout,
797 pAllocateInfo->pSetLayouts[i]);
798
799 const uint32_t *variable_count = NULL;
800 if (variable_counts) {
801 if (i < variable_counts->descriptorSetCount)
802 variable_count = variable_counts->pDescriptorCounts + i;
803 else
804 variable_count = &zero;
805 }
806
807 assert(!(layout->flags & VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR));
808
809 result = radv_descriptor_set_create(device, pool, layout, variable_count, &set);
810 if (result != VK_SUCCESS)
811 break;
812
813 pDescriptorSets[i] = radv_descriptor_set_to_handle(set);
814 }
815
816 if (result != VK_SUCCESS) {
817 radv_FreeDescriptorSets(_device, pAllocateInfo->descriptorPool,
818 i, pDescriptorSets);
819 for (i = 0; i < pAllocateInfo->descriptorSetCount; i++) {
820 pDescriptorSets[i] = VK_NULL_HANDLE;
821 }
822 }
823 return result;
824 }
825
826 VkResult radv_FreeDescriptorSets(
827 VkDevice _device,
828 VkDescriptorPool descriptorPool,
829 uint32_t count,
830 const VkDescriptorSet* pDescriptorSets)
831 {
832 RADV_FROM_HANDLE(radv_device, device, _device);
833 RADV_FROM_HANDLE(radv_descriptor_pool, pool, descriptorPool);
834
835 for (uint32_t i = 0; i < count; i++) {
836 RADV_FROM_HANDLE(radv_descriptor_set, set, pDescriptorSets[i]);
837
838 if (set && !pool->host_memory_base)
839 radv_descriptor_set_destroy(device, pool, set, true);
840 }
841 return VK_SUCCESS;
842 }
843
844 static void write_texel_buffer_descriptor(struct radv_device *device,
845 struct radv_cmd_buffer *cmd_buffer,
846 unsigned *dst,
847 struct radeon_winsys_bo **buffer_list,
848 const VkBufferView _buffer_view)
849 {
850 RADV_FROM_HANDLE(radv_buffer_view, buffer_view, _buffer_view);
851
852 if (!buffer_view) {
853 memset(dst, 0, 4 * 4);
854 return;
855 }
856
857 memcpy(dst, buffer_view->state, 4 * 4);
858
859 if (cmd_buffer)
860 radv_cs_add_buffer(device->ws, cmd_buffer->cs, buffer_view->bo);
861 else
862 *buffer_list = buffer_view->bo;
863 }
864
865 static void write_buffer_descriptor(struct radv_device *device,
866 struct radv_cmd_buffer *cmd_buffer,
867 unsigned *dst,
868 struct radeon_winsys_bo **buffer_list,
869 const VkDescriptorBufferInfo *buffer_info)
870 {
871 RADV_FROM_HANDLE(radv_buffer, buffer, buffer_info->buffer);
872
873 if (!buffer) {
874 memset(dst, 0, 4 * 4);
875 return;
876 }
877
878 uint64_t va = radv_buffer_get_va(buffer->bo);
879 uint32_t range = buffer_info->range;
880
881 if (buffer_info->range == VK_WHOLE_SIZE)
882 range = buffer->size - buffer_info->offset;
883
884 /* robustBufferAccess is relaxed enough to allow this (in combination
885 * with the alignment/size we return from vkGetBufferMemoryRequirements)
886 * and this allows the shader compiler to create more efficient 8/16-bit
887 * buffer accesses. */
888 range = align(range, 4);
889
890 va += buffer_info->offset + buffer->offset;
891 dst[0] = va;
892 dst[1] = S_008F04_BASE_ADDRESS_HI(va >> 32);
893 dst[2] = range;
894 dst[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) |
895 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
896 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) |
897 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W);
898
899 if (device->physical_device->rad_info.chip_class >= GFX10) {
900 dst[3] |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT) |
901 S_008F0C_OOB_SELECT(V_008F0C_OOB_SELECT_RAW) |
902 S_008F0C_RESOURCE_LEVEL(1);
903 } else {
904 dst[3] |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
905 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32);
906 }
907
908 if (cmd_buffer)
909 radv_cs_add_buffer(device->ws, cmd_buffer->cs, buffer->bo);
910 else
911 *buffer_list = buffer->bo;
912 }
913
914 static void write_block_descriptor(struct radv_device *device,
915 struct radv_cmd_buffer *cmd_buffer,
916 void *dst,
917 const VkWriteDescriptorSet *writeset)
918 {
919 const VkWriteDescriptorSetInlineUniformBlockEXT *inline_ub =
920 vk_find_struct_const(writeset->pNext, WRITE_DESCRIPTOR_SET_INLINE_UNIFORM_BLOCK_EXT);
921
922 memcpy(dst, inline_ub->pData, inline_ub->dataSize);
923 }
924
925 static void write_dynamic_buffer_descriptor(struct radv_device *device,
926 struct radv_descriptor_range *range,
927 struct radeon_winsys_bo **buffer_list,
928 const VkDescriptorBufferInfo *buffer_info)
929 {
930 RADV_FROM_HANDLE(radv_buffer, buffer, buffer_info->buffer);
931 uint64_t va;
932 unsigned size;
933
934 if (!buffer)
935 return;
936
937 va = radv_buffer_get_va(buffer->bo);
938 size = buffer_info->range;
939
940 if (buffer_info->range == VK_WHOLE_SIZE)
941 size = buffer->size - buffer_info->offset;
942
943 /* robustBufferAccess is relaxed enough to allow this (in combination
944 * with the alignment/size we return from vkGetBufferMemoryRequirements)
945 * and this allows the shader compiler to create more efficient 8/16-bit
946 * buffer accesses. */
947 size = align(size, 4);
948
949 va += buffer_info->offset + buffer->offset;
950 range->va = va;
951 range->size = size;
952
953 *buffer_list = buffer->bo;
954 }
955
956 static void
957 write_image_descriptor(struct radv_device *device,
958 struct radv_cmd_buffer *cmd_buffer,
959 unsigned size, unsigned *dst,
960 struct radeon_winsys_bo **buffer_list,
961 VkDescriptorType descriptor_type,
962 const VkDescriptorImageInfo *image_info)
963 {
964 RADV_FROM_HANDLE(radv_image_view, iview, image_info->imageView);
965 union radv_descriptor *descriptor;
966
967 if (!iview) {
968 memset(dst, 0, size);
969 return;
970 }
971
972 if (descriptor_type == VK_DESCRIPTOR_TYPE_STORAGE_IMAGE) {
973 descriptor = &iview->storage_descriptor;
974 } else {
975 descriptor = &iview->descriptor;
976 }
977
978 memcpy(dst, descriptor, size);
979
980 if (cmd_buffer)
981 radv_cs_add_buffer(device->ws, cmd_buffer->cs, iview->bo);
982 else
983 *buffer_list = iview->bo;
984 }
985
986 static void
987 write_combined_image_sampler_descriptor(struct radv_device *device,
988 struct radv_cmd_buffer *cmd_buffer,
989 unsigned sampler_offset,
990 unsigned *dst,
991 struct radeon_winsys_bo **buffer_list,
992 VkDescriptorType descriptor_type,
993 const VkDescriptorImageInfo *image_info,
994 bool has_sampler)
995 {
996 RADV_FROM_HANDLE(radv_sampler, sampler, image_info->sampler);
997
998 write_image_descriptor(device, cmd_buffer, sampler_offset, dst, buffer_list,
999 descriptor_type, image_info);
1000 /* copy over sampler state */
1001 if (has_sampler) {
1002 memcpy(dst + sampler_offset / sizeof(*dst), sampler->state, 16);
1003 }
1004 }
1005
1006 static void
1007 write_sampler_descriptor(struct radv_device *device,
1008 unsigned *dst,
1009 const VkDescriptorImageInfo *image_info)
1010 {
1011 RADV_FROM_HANDLE(radv_sampler, sampler, image_info->sampler);
1012
1013 memcpy(dst, sampler->state, 16);
1014 }
1015
1016 void radv_update_descriptor_sets(
1017 struct radv_device* device,
1018 struct radv_cmd_buffer* cmd_buffer,
1019 VkDescriptorSet dstSetOverride,
1020 uint32_t descriptorWriteCount,
1021 const VkWriteDescriptorSet* pDescriptorWrites,
1022 uint32_t descriptorCopyCount,
1023 const VkCopyDescriptorSet* pDescriptorCopies)
1024 {
1025 uint32_t i, j;
1026 for (i = 0; i < descriptorWriteCount; i++) {
1027 const VkWriteDescriptorSet *writeset = &pDescriptorWrites[i];
1028 RADV_FROM_HANDLE(radv_descriptor_set, set,
1029 dstSetOverride ? dstSetOverride : writeset->dstSet);
1030 const struct radv_descriptor_set_binding_layout *binding_layout =
1031 set->layout->binding + writeset->dstBinding;
1032 uint32_t *ptr = set->mapped_ptr;
1033 struct radeon_winsys_bo **buffer_list = set->descriptors;
1034 /* Immutable samplers are not copied into push descriptors when they are
1035 * allocated, so if we are writing push descriptors we have to copy the
1036 * immutable samplers into them now.
1037 */
1038 const bool copy_immutable_samplers = cmd_buffer &&
1039 binding_layout->immutable_samplers_offset && !binding_layout->immutable_samplers_equal;
1040 const uint32_t *samplers = radv_immutable_samplers(set->layout, binding_layout);
1041
1042 ptr += binding_layout->offset / 4;
1043
1044 if (writeset->descriptorType == VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT) {
1045 write_block_descriptor(device, cmd_buffer, (uint8_t*)ptr + writeset->dstArrayElement, writeset);
1046 continue;
1047 }
1048
1049 ptr += binding_layout->size * writeset->dstArrayElement / 4;
1050 buffer_list += binding_layout->buffer_offset;
1051 buffer_list += writeset->dstArrayElement;
1052 for (j = 0; j < writeset->descriptorCount; ++j) {
1053 switch(writeset->descriptorType) {
1054 case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
1055 case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC: {
1056 unsigned idx = writeset->dstArrayElement + j;
1057 idx += binding_layout->dynamic_offset_offset;
1058 assert(!(set->layout->flags & VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR));
1059 write_dynamic_buffer_descriptor(device, set->dynamic_descriptors + idx,
1060 buffer_list, writeset->pBufferInfo + j);
1061 break;
1062 }
1063 case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
1064 case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
1065 write_buffer_descriptor(device, cmd_buffer, ptr, buffer_list,
1066 writeset->pBufferInfo + j);
1067 break;
1068 case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
1069 case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
1070 write_texel_buffer_descriptor(device, cmd_buffer, ptr, buffer_list,
1071 writeset->pTexelBufferView[j]);
1072 break;
1073 case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE:
1074 case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE:
1075 case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT:
1076 write_image_descriptor(device, cmd_buffer, 64, ptr, buffer_list,
1077 writeset->descriptorType,
1078 writeset->pImageInfo + j);
1079 break;
1080 case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER: {
1081 unsigned sampler_offset = radv_combined_image_descriptor_sampler_offset(binding_layout);
1082 write_combined_image_sampler_descriptor(device, cmd_buffer, sampler_offset,
1083 ptr, buffer_list,
1084 writeset->descriptorType,
1085 writeset->pImageInfo + j,
1086 !binding_layout->immutable_samplers_offset);
1087 if (copy_immutable_samplers) {
1088 const unsigned idx = writeset->dstArrayElement + j;
1089 memcpy((char*)ptr + sampler_offset, samplers + 4 * idx, 16);
1090 }
1091 break;
1092 }
1093 case VK_DESCRIPTOR_TYPE_SAMPLER:
1094 if (!binding_layout->immutable_samplers_offset) {
1095 write_sampler_descriptor(device, ptr,
1096 writeset->pImageInfo + j);
1097 } else if (copy_immutable_samplers) {
1098 unsigned idx = writeset->dstArrayElement + j;
1099 memcpy(ptr, samplers + 4 * idx, 16);
1100 }
1101 break;
1102 default:
1103 break;
1104 }
1105 ptr += binding_layout->size / 4;
1106 ++buffer_list;
1107 }
1108
1109 }
1110
1111 for (i = 0; i < descriptorCopyCount; i++) {
1112 const VkCopyDescriptorSet *copyset = &pDescriptorCopies[i];
1113 RADV_FROM_HANDLE(radv_descriptor_set, src_set,
1114 copyset->srcSet);
1115 RADV_FROM_HANDLE(radv_descriptor_set, dst_set,
1116 copyset->dstSet);
1117 const struct radv_descriptor_set_binding_layout *src_binding_layout =
1118 src_set->layout->binding + copyset->srcBinding;
1119 const struct radv_descriptor_set_binding_layout *dst_binding_layout =
1120 dst_set->layout->binding + copyset->dstBinding;
1121 uint32_t *src_ptr = src_set->mapped_ptr;
1122 uint32_t *dst_ptr = dst_set->mapped_ptr;
1123 struct radeon_winsys_bo **src_buffer_list = src_set->descriptors;
1124 struct radeon_winsys_bo **dst_buffer_list = dst_set->descriptors;
1125
1126 src_ptr += src_binding_layout->offset / 4;
1127 dst_ptr += dst_binding_layout->offset / 4;
1128
1129 if (src_binding_layout->type == VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT) {
1130 src_ptr += copyset->srcArrayElement / 4;
1131 dst_ptr += copyset->dstArrayElement / 4;
1132
1133 memcpy(dst_ptr, src_ptr, copyset->descriptorCount);
1134 continue;
1135 }
1136
1137 src_ptr += src_binding_layout->size * copyset->srcArrayElement / 4;
1138 dst_ptr += dst_binding_layout->size * copyset->dstArrayElement / 4;
1139
1140 src_buffer_list += src_binding_layout->buffer_offset;
1141 src_buffer_list += copyset->srcArrayElement;
1142
1143 dst_buffer_list += dst_binding_layout->buffer_offset;
1144 dst_buffer_list += copyset->dstArrayElement;
1145
1146 for (j = 0; j < copyset->descriptorCount; ++j) {
1147 switch (src_binding_layout->type) {
1148 case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
1149 case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC: {
1150 unsigned src_idx = copyset->srcArrayElement + j;
1151 unsigned dst_idx = copyset->dstArrayElement + j;
1152 struct radv_descriptor_range *src_range, *dst_range;
1153 src_idx += src_binding_layout->dynamic_offset_offset;
1154 dst_idx += dst_binding_layout->dynamic_offset_offset;
1155
1156 src_range = src_set->dynamic_descriptors + src_idx;
1157 dst_range = dst_set->dynamic_descriptors + dst_idx;
1158 *dst_range = *src_range;
1159 break;
1160 }
1161 default:
1162 memcpy(dst_ptr, src_ptr, src_binding_layout->size);
1163 }
1164 src_ptr += src_binding_layout->size / 4;
1165 dst_ptr += dst_binding_layout->size / 4;
1166
1167 if (src_binding_layout->type != VK_DESCRIPTOR_TYPE_SAMPLER) {
1168 /* Sampler descriptors don't have a buffer list. */
1169 dst_buffer_list[j] = src_buffer_list[j];
1170 }
1171 }
1172 }
1173 }
1174
1175 void radv_UpdateDescriptorSets(
1176 VkDevice _device,
1177 uint32_t descriptorWriteCount,
1178 const VkWriteDescriptorSet* pDescriptorWrites,
1179 uint32_t descriptorCopyCount,
1180 const VkCopyDescriptorSet* pDescriptorCopies)
1181 {
1182 RADV_FROM_HANDLE(radv_device, device, _device);
1183
1184 radv_update_descriptor_sets(device, NULL, VK_NULL_HANDLE, descriptorWriteCount, pDescriptorWrites,
1185 descriptorCopyCount, pDescriptorCopies);
1186 }
1187
1188 VkResult radv_CreateDescriptorUpdateTemplate(VkDevice _device,
1189 const VkDescriptorUpdateTemplateCreateInfo *pCreateInfo,
1190 const VkAllocationCallbacks *pAllocator,
1191 VkDescriptorUpdateTemplate *pDescriptorUpdateTemplate)
1192 {
1193 RADV_FROM_HANDLE(radv_device, device, _device);
1194 RADV_FROM_HANDLE(radv_descriptor_set_layout, set_layout, pCreateInfo->descriptorSetLayout);
1195 const uint32_t entry_count = pCreateInfo->descriptorUpdateEntryCount;
1196 const size_t size = sizeof(struct radv_descriptor_update_template) +
1197 sizeof(struct radv_descriptor_update_template_entry) * entry_count;
1198 struct radv_descriptor_update_template *templ;
1199 uint32_t i;
1200
1201 templ = vk_alloc2(&device->vk.alloc, pAllocator, size, 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1202 if (!templ)
1203 return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
1204
1205 vk_object_base_init(&device->vk, &templ->base,
1206 VK_OBJECT_TYPE_DESCRIPTOR_UPDATE_TEMPLATE);
1207
1208 templ->entry_count = entry_count;
1209
1210 if (pCreateInfo->templateType == VK_DESCRIPTOR_UPDATE_TEMPLATE_TYPE_PUSH_DESCRIPTORS_KHR) {
1211 RADV_FROM_HANDLE(radv_pipeline_layout, pipeline_layout, pCreateInfo->pipelineLayout);
1212
1213 /* descriptorSetLayout should be ignored for push descriptors
1214 * and instead it refers to pipelineLayout and set.
1215 */
1216 assert(pCreateInfo->set < MAX_SETS);
1217 set_layout = pipeline_layout->set[pCreateInfo->set].layout;
1218
1219 templ->bind_point = pCreateInfo->pipelineBindPoint;
1220 }
1221
1222 for (i = 0; i < entry_count; i++) {
1223 const VkDescriptorUpdateTemplateEntry *entry = &pCreateInfo->pDescriptorUpdateEntries[i];
1224 const struct radv_descriptor_set_binding_layout *binding_layout =
1225 set_layout->binding + entry->dstBinding;
1226 const uint32_t buffer_offset = binding_layout->buffer_offset + entry->dstArrayElement;
1227 const uint32_t *immutable_samplers = NULL;
1228 uint32_t dst_offset;
1229 uint32_t dst_stride;
1230
1231 /* dst_offset is an offset into dynamic_descriptors when the descriptor
1232 is dynamic, and an offset into mapped_ptr otherwise */
1233 switch (entry->descriptorType) {
1234 case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
1235 case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
1236 assert(pCreateInfo->templateType == VK_DESCRIPTOR_UPDATE_TEMPLATE_TYPE_DESCRIPTOR_SET);
1237 dst_offset = binding_layout->dynamic_offset_offset + entry->dstArrayElement;
1238 dst_stride = 0; /* Not used */
1239 break;
1240 default:
1241 switch (entry->descriptorType) {
1242 case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
1243 case VK_DESCRIPTOR_TYPE_SAMPLER:
1244 /* Immutable samplers are copied into push descriptors when they are pushed */
1245 if (pCreateInfo->templateType == VK_DESCRIPTOR_UPDATE_TEMPLATE_TYPE_PUSH_DESCRIPTORS_KHR &&
1246 binding_layout->immutable_samplers_offset && !binding_layout->immutable_samplers_equal) {
1247 immutable_samplers = radv_immutable_samplers(set_layout, binding_layout) + entry->dstArrayElement * 4;
1248 }
1249 break;
1250 default:
1251 break;
1252 }
1253 dst_offset = binding_layout->offset / 4;
1254 if (entry->descriptorType == VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT)
1255 dst_offset += entry->dstArrayElement / 4;
1256 else
1257 dst_offset += binding_layout->size * entry->dstArrayElement / 4;
1258
1259 dst_stride = binding_layout->size / 4;
1260 break;
1261 }
1262
1263 templ->entry[i] = (struct radv_descriptor_update_template_entry) {
1264 .descriptor_type = entry->descriptorType,
1265 .descriptor_count = entry->descriptorCount,
1266 .src_offset = entry->offset,
1267 .src_stride = entry->stride,
1268 .dst_offset = dst_offset,
1269 .dst_stride = dst_stride,
1270 .buffer_offset = buffer_offset,
1271 .has_sampler = !binding_layout->immutable_samplers_offset,
1272 .sampler_offset = radv_combined_image_descriptor_sampler_offset(binding_layout),
1273 .immutable_samplers = immutable_samplers
1274 };
1275 }
1276
1277 *pDescriptorUpdateTemplate = radv_descriptor_update_template_to_handle(templ);
1278 return VK_SUCCESS;
1279 }
1280
1281 void radv_DestroyDescriptorUpdateTemplate(VkDevice _device,
1282 VkDescriptorUpdateTemplate descriptorUpdateTemplate,
1283 const VkAllocationCallbacks *pAllocator)
1284 {
1285 RADV_FROM_HANDLE(radv_device, device, _device);
1286 RADV_FROM_HANDLE(radv_descriptor_update_template, templ, descriptorUpdateTemplate);
1287
1288 if (!templ)
1289 return;
1290
1291 vk_object_base_finish(&templ->base);
1292 vk_free2(&device->vk.alloc, pAllocator, templ);
1293 }
1294
1295 void radv_update_descriptor_set_with_template(struct radv_device *device,
1296 struct radv_cmd_buffer *cmd_buffer,
1297 struct radv_descriptor_set *set,
1298 VkDescriptorUpdateTemplate descriptorUpdateTemplate,
1299 const void *pData)
1300 {
1301 RADV_FROM_HANDLE(radv_descriptor_update_template, templ, descriptorUpdateTemplate);
1302 uint32_t i;
1303
1304 for (i = 0; i < templ->entry_count; ++i) {
1305 struct radeon_winsys_bo **buffer_list = set->descriptors + templ->entry[i].buffer_offset;
1306 uint32_t *pDst = set->mapped_ptr + templ->entry[i].dst_offset;
1307 const uint8_t *pSrc = ((const uint8_t *) pData) + templ->entry[i].src_offset;
1308 uint32_t j;
1309
1310 if (templ->entry[i].descriptor_type == VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT) {
1311 memcpy((uint8_t*)pDst, pSrc, templ->entry[i].descriptor_count);
1312 continue;
1313 }
1314
1315 for (j = 0; j < templ->entry[i].descriptor_count; ++j) {
1316 switch (templ->entry[i].descriptor_type) {
1317 case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
1318 case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC: {
1319 const unsigned idx = templ->entry[i].dst_offset + j;
1320 assert(!(set->layout->flags & VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR));
1321 write_dynamic_buffer_descriptor(device, set->dynamic_descriptors + idx,
1322 buffer_list, (struct VkDescriptorBufferInfo *) pSrc);
1323 break;
1324 }
1325 case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
1326 case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
1327 write_buffer_descriptor(device, cmd_buffer, pDst, buffer_list,
1328 (struct VkDescriptorBufferInfo *) pSrc);
1329 break;
1330 case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
1331 case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
1332 write_texel_buffer_descriptor(device, cmd_buffer, pDst, buffer_list,
1333 *(VkBufferView *) pSrc);
1334 break;
1335 case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE:
1336 case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE:
1337 case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT:
1338 write_image_descriptor(device, cmd_buffer, 64, pDst, buffer_list,
1339 templ->entry[i].descriptor_type,
1340 (struct VkDescriptorImageInfo *) pSrc);
1341 break;
1342 case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
1343 write_combined_image_sampler_descriptor(device, cmd_buffer, templ->entry[i].sampler_offset,
1344 pDst, buffer_list, templ->entry[i].descriptor_type,
1345 (struct VkDescriptorImageInfo *) pSrc,
1346 templ->entry[i].has_sampler);
1347 if (templ->entry[i].immutable_samplers) {
1348 memcpy((char*)pDst + templ->entry[i].sampler_offset, templ->entry[i].immutable_samplers + 4 * j, 16);
1349 }
1350 break;
1351 case VK_DESCRIPTOR_TYPE_SAMPLER:
1352 if (templ->entry[i].has_sampler)
1353 write_sampler_descriptor(device, pDst,
1354 (struct VkDescriptorImageInfo *) pSrc);
1355 else if (templ->entry[i].immutable_samplers)
1356 memcpy(pDst, templ->entry[i].immutable_samplers + 4 * j, 16);
1357 break;
1358 default:
1359 break;
1360 }
1361 pSrc += templ->entry[i].src_stride;
1362 pDst += templ->entry[i].dst_stride;
1363 ++buffer_list;
1364 }
1365 }
1366 }
1367
1368 void radv_UpdateDescriptorSetWithTemplate(VkDevice _device,
1369 VkDescriptorSet descriptorSet,
1370 VkDescriptorUpdateTemplate descriptorUpdateTemplate,
1371 const void *pData)
1372 {
1373 RADV_FROM_HANDLE(radv_device, device, _device);
1374 RADV_FROM_HANDLE(radv_descriptor_set, set, descriptorSet);
1375
1376 radv_update_descriptor_set_with_template(device, NULL, set, descriptorUpdateTemplate, pData);
1377 }
1378
1379
1380 VkResult radv_CreateSamplerYcbcrConversion(VkDevice _device,
1381 const VkSamplerYcbcrConversionCreateInfo* pCreateInfo,
1382 const VkAllocationCallbacks* pAllocator,
1383 VkSamplerYcbcrConversion* pYcbcrConversion)
1384 {
1385 RADV_FROM_HANDLE(radv_device, device, _device);
1386 struct radv_sampler_ycbcr_conversion *conversion = NULL;
1387
1388 conversion = vk_zalloc2(&device->vk.alloc, pAllocator, sizeof(*conversion), 8,
1389 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
1390
1391 if (conversion == NULL)
1392 return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
1393
1394 vk_object_base_init(&device->vk, &conversion->base,
1395 VK_OBJECT_TYPE_SAMPLER_YCBCR_CONVERSION);
1396
1397 conversion->format = pCreateInfo->format;
1398 conversion->ycbcr_model = pCreateInfo->ycbcrModel;
1399 conversion->ycbcr_range = pCreateInfo->ycbcrRange;
1400 conversion->components = pCreateInfo->components;
1401 conversion->chroma_offsets[0] = pCreateInfo->xChromaOffset;
1402 conversion->chroma_offsets[1] = pCreateInfo->yChromaOffset;
1403 conversion->chroma_filter = pCreateInfo->chromaFilter;
1404
1405 *pYcbcrConversion = radv_sampler_ycbcr_conversion_to_handle(conversion);
1406 return VK_SUCCESS;
1407 }
1408
1409
1410 void radv_DestroySamplerYcbcrConversion(VkDevice _device,
1411 VkSamplerYcbcrConversion ycbcrConversion,
1412 const VkAllocationCallbacks* pAllocator)
1413 {
1414 RADV_FROM_HANDLE(radv_device, device, _device);
1415 RADV_FROM_HANDLE(radv_sampler_ycbcr_conversion, ycbcr_conversion, ycbcrConversion);
1416
1417 if (!ycbcr_conversion)
1418 return;
1419
1420 vk_object_base_finish(&ycbcr_conversion->base);
1421 vk_free2(&device->vk.alloc, pAllocator, ycbcr_conversion);
1422 }