projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
vulkan/overlay: fix crash on freeing NULL command buffer
[mesa.git] / src / vulkan / overlay-layer / overlay.cpp
1 /*
2 * Copyright © 2019 Intel Corporation
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 */
23
24 #include <string.h>
25 #include <stdlib.h>
26 #include <assert.h>
27
28 #include <vulkan/vulkan.h>
29 #include <vulkan/vk_layer.h>
30
31 #include "imgui.h"
32
33 #include "overlay_params.h"
34
35 #include "util/debug.h"
36 #include "util/hash_table.h"
37 #include "util/list.h"
38 #include "util/ralloc.h"
39 #include "util/os_time.h"
40 #include "util/simple_mtx.h"
41
42 #include "vk_enum_to_str.h"
43 #include "vk_util.h"
44
45 /* Mapped from VkInstace/VkPhysicalDevice */
46 struct instance_data {
47 struct vk_instance_dispatch_table vtable;
48 VkInstance instance;
49
50 struct overlay_params params;
51 bool pipeline_statistics_enabled;
52
53 bool first_line_printed;
54 };
55
56 struct frame_stat {
57 uint64_t stats[OVERLAY_PARAM_ENABLED_MAX];
58 };
59
60 /* Mapped from VkDevice */
61 struct queue_data;
62 struct device_data {
63 struct instance_data *instance;
64
65 PFN_vkSetDeviceLoaderData set_device_loader_data;
66
67 struct vk_device_dispatch_table vtable;
68 VkPhysicalDevice physical_device;
69 VkDevice device;
70
71 VkPhysicalDeviceProperties properties;
72
73 struct queue_data *graphic_queue;
74
75 struct queue_data **queues;
76 uint32_t n_queues;
77
78 /* For a single frame */
79 struct frame_stat frame_stats;
80 };
81
82 /* Mapped from VkCommandBuffer */
83 struct command_buffer_data {
84 struct device_data *device;
85
86 VkCommandBufferLevel level;
87
88 VkCommandBuffer cmd_buffer;
89 VkQueryPool pipeline_query_pool;
90 VkQueryPool timestamp_query_pool;
91 uint32_t query_index;
92
93 struct frame_stat stats;
94
95 struct list_head link; /* link into queue_data::running_command_buffer */
96 };
97
98 /* Mapped from VkQueue */
99 struct queue_data {
100 struct device_data *device;
101
102 VkQueue queue;
103 VkQueueFlags flags;
104 uint32_t family_index;
105 uint64_t timestamp_mask;
106
107 VkFence queries_fence;
108
109 struct list_head running_command_buffer;
110 };
111
112 struct overlay_draw {
113 struct list_head link;
114
115 VkCommandBuffer command_buffer;
116
117 VkSemaphore semaphore;
118 VkFence fence;
119
120 VkBuffer vertex_buffer;
121 VkDeviceMemory vertex_buffer_mem;
122 VkDeviceSize vertex_buffer_size;
123
124 VkBuffer index_buffer;
125 VkDeviceMemory index_buffer_mem;
126 VkDeviceSize index_buffer_size;
127 };
128
129 /* Mapped from VkSwapchainKHR */
130 struct swapchain_data {
131 struct device_data *device;
132
133 VkSwapchainKHR swapchain;
134 unsigned width, height;
135 VkFormat format;
136
137 uint32_t n_images;
138 VkImage *images;
139 VkImageView *image_views;
140 VkFramebuffer *framebuffers;
141
142 VkRenderPass render_pass;
143
144 VkDescriptorPool descriptor_pool;
145 VkDescriptorSetLayout descriptor_layout;
146 VkDescriptorSet descriptor_set;
147
148 VkSampler font_sampler;
149
150 VkPipelineLayout pipeline_layout;
151 VkPipeline pipeline;
152
153 VkCommandPool command_pool;
154
155 struct list_head draws; /* List of struct overlay_draw */
156
157 bool font_uploaded;
158 VkImage font_image;
159 VkImageView font_image_view;
160 VkDeviceMemory font_mem;
161 VkBuffer upload_font_buffer;
162 VkDeviceMemory upload_font_buffer_mem;
163
164 /**/
165 ImGuiContext* imgui_context;
166 ImVec2 window_size;
167
168 /**/
169 uint64_t n_frames;
170 uint64_t last_present_time;
171
172 unsigned n_frames_since_update;
173 uint64_t last_fps_update;
174 double fps;
175
176 enum overlay_param_enabled stat_selector;
177 double time_dividor;
178 struct frame_stat stats_min, stats_max;
179 struct frame_stat frames_stats[200];
180
181 /* Over a single frame */
182 struct frame_stat frame_stats;
183
184 /* Over fps_sampling_period */
185 struct frame_stat accumulated_stats;
186 };
187
188 static const VkQueryPipelineStatisticFlags overlay_query_flags =
189 VK_QUERY_PIPELINE_STATISTIC_INPUT_ASSEMBLY_VERTICES_BIT |
190 VK_QUERY_PIPELINE_STATISTIC_INPUT_ASSEMBLY_PRIMITIVES_BIT |
191 VK_QUERY_PIPELINE_STATISTIC_VERTEX_SHADER_INVOCATIONS_BIT |
192 VK_QUERY_PIPELINE_STATISTIC_GEOMETRY_SHADER_INVOCATIONS_BIT |
193 VK_QUERY_PIPELINE_STATISTIC_GEOMETRY_SHADER_PRIMITIVES_BIT |
194 VK_QUERY_PIPELINE_STATISTIC_CLIPPING_INVOCATIONS_BIT |
195 VK_QUERY_PIPELINE_STATISTIC_CLIPPING_PRIMITIVES_BIT |
196 VK_QUERY_PIPELINE_STATISTIC_FRAGMENT_SHADER_INVOCATIONS_BIT |
197 VK_QUERY_PIPELINE_STATISTIC_TESSELLATION_CONTROL_SHADER_PATCHES_BIT |
198 VK_QUERY_PIPELINE_STATISTIC_TESSELLATION_EVALUATION_SHADER_INVOCATIONS_BIT |
199 VK_QUERY_PIPELINE_STATISTIC_COMPUTE_SHADER_INVOCATIONS_BIT;
200 #define OVERLAY_QUERY_COUNT (11)
201
202 static struct hash_table_u64 *vk_object_to_data = NULL;
203 static simple_mtx_t vk_object_to_data_mutex = _SIMPLE_MTX_INITIALIZER_NP;
204
205 thread_local ImGuiContext* __MesaImGui;
206
207 static inline void ensure_vk_object_map(void)
208 {
209 if (!vk_object_to_data)
210 vk_object_to_data = _mesa_hash_table_u64_create(NULL);
211 }
212
213 #define HKEY(obj) ((uint64_t)(obj))
214 #define FIND_SWAPCHAIN_DATA(obj) ((struct swapchain_data *)find_object_data(HKEY(obj)))
215 #define FIND_CMD_BUFFER_DATA(obj) ((struct command_buffer_data *)find_object_data(HKEY(obj)))
216 #define FIND_DEVICE_DATA(obj) ((struct device_data *)find_object_data(HKEY(obj)))
217 #define FIND_QUEUE_DATA(obj) ((struct queue_data *)find_object_data(HKEY(obj)))
218 #define FIND_PHYSICAL_DEVICE_DATA(obj) ((struct instance_data *)find_object_data(HKEY(obj)))
219 #define FIND_INSTANCE_DATA(obj) ((struct instance_data *)find_object_data(HKEY(obj)))
220 static void *find_object_data(uint64_t obj)
221 {
222 simple_mtx_lock(&vk_object_to_data_mutex);
223 ensure_vk_object_map();
224 void *data = _mesa_hash_table_u64_search(vk_object_to_data, obj);
225 simple_mtx_unlock(&vk_object_to_data_mutex);
226 return data;
227 }
228
229 static void map_object(uint64_t obj, void *data)
230 {
231 simple_mtx_lock(&vk_object_to_data_mutex);
232 ensure_vk_object_map();
233 _mesa_hash_table_u64_insert(vk_object_to_data, obj, data);
234 simple_mtx_unlock(&vk_object_to_data_mutex);
235 }
236
237 static void unmap_object(uint64_t obj)
238 {
239 simple_mtx_lock(&vk_object_to_data_mutex);
240 _mesa_hash_table_u64_remove(vk_object_to_data, obj);
241 simple_mtx_unlock(&vk_object_to_data_mutex);
242 }
243
244 /**/
245
246 #define VK_CHECK(expr) \
247 do { \
248 VkResult __result = (expr); \
249 if (__result != VK_SUCCESS) { \
250 fprintf(stderr, "'%s' line %i failed with %s\n", \
251 #expr, __LINE__, vk_Result_to_str(__result)); \
252 } \
253 } while (0)
254
255 /**/
256
257 static VkLayerInstanceCreateInfo *get_instance_chain_info(const VkInstanceCreateInfo *pCreateInfo,
258 VkLayerFunction func)
259 {
260 vk_foreach_struct(item, pCreateInfo->pNext) {
261 if (item->sType == VK_STRUCTURE_TYPE_LOADER_INSTANCE_CREATE_INFO &&
262 ((VkLayerInstanceCreateInfo *) item)->function == func)
263 return (VkLayerInstanceCreateInfo *) item;
264 }
265 unreachable("instance chain info not found");
266 return NULL;
267 }
268
269 static VkLayerDeviceCreateInfo *get_device_chain_info(const VkDeviceCreateInfo *pCreateInfo,
270 VkLayerFunction func)
271 {
272 vk_foreach_struct(item, pCreateInfo->pNext) {
273 if (item->sType == VK_STRUCTURE_TYPE_LOADER_DEVICE_CREATE_INFO &&
274 ((VkLayerDeviceCreateInfo *) item)->function == func)
275 return (VkLayerDeviceCreateInfo *)item;
276 }
277 unreachable("device chain info not found");
278 return NULL;
279 }
280
281 static struct VkBaseOutStructure *
282 clone_chain(const struct VkBaseInStructure *chain)
283 {
284 struct VkBaseOutStructure *head = NULL, *tail = NULL;
285
286 vk_foreach_struct_const(item, chain) {
287 size_t item_size = vk_structure_type_size(item);
288 struct VkBaseOutStructure *new_item =
289 (struct VkBaseOutStructure *)malloc(item_size);;
290
291 memcpy(new_item, item, item_size);
292
293 if (!head)
294 head = new_item;
295 if (tail)
296 tail->pNext = new_item;
297 tail = new_item;
298 }
299
300 return head;
301 }
302
303 static void
304 free_chain(struct VkBaseOutStructure *chain)
305 {
306 while (chain) {
307 void *node = chain;
308 chain = chain->pNext;
309 free(node);
310 }
311 }
312
313 /**/
314
315 static struct instance_data *new_instance_data(VkInstance instance)
316 {
317 struct instance_data *data = rzalloc(NULL, struct instance_data);
318 data->instance = instance;
319 map_object(HKEY(data->instance), data);
320 return data;
321 }
322
323 static void destroy_instance_data(struct instance_data *data)
324 {
325 if (data->params.output_file)
326 fclose(data->params.output_file);
327 unmap_object(HKEY(data->instance));
328 ralloc_free(data);
329 }
330
331 static void instance_data_map_physical_devices(struct instance_data *instance_data,
332 bool map)
333 {
334 uint32_t physicalDeviceCount = 0;
335 instance_data->vtable.EnumeratePhysicalDevices(instance_data->instance,
336 &physicalDeviceCount,
337 NULL);
338
339 VkPhysicalDevice *physicalDevices = (VkPhysicalDevice *) malloc(sizeof(VkPhysicalDevice) * physicalDeviceCount);
340 instance_data->vtable.EnumeratePhysicalDevices(instance_data->instance,
341 &physicalDeviceCount,
342 physicalDevices);
343
344 for (uint32_t i = 0; i < physicalDeviceCount; i++) {
345 if (map)
346 map_object(HKEY(physicalDevices[i]), instance_data);
347 else
348 unmap_object(HKEY(physicalDevices[i]));
349 }
350
351 free(physicalDevices);
352 }
353
354 /**/
355 static struct device_data *new_device_data(VkDevice device, struct instance_data *instance)
356 {
357 struct device_data *data = rzalloc(NULL, struct device_data);
358 data->instance = instance;
359 data->device = device;
360 map_object(HKEY(data->device), data);
361 return data;
362 }
363
364 static struct queue_data *new_queue_data(VkQueue queue,
365 const VkQueueFamilyProperties *family_props,
366 uint32_t family_index,
367 struct device_data *device_data)
368 {
369 struct queue_data *data = rzalloc(device_data, struct queue_data);
370 data->device = device_data;
371 data->queue = queue;
372 data->flags = family_props->queueFlags;
373 data->timestamp_mask = (1ull << family_props->timestampValidBits) - 1;
374 data->family_index = family_index;
375 LIST_INITHEAD(&data->running_command_buffer);
376 map_object(HKEY(data->queue), data);
377
378 /* Fence synchronizing access to queries on that queue. */
379 VkFenceCreateInfo fence_info = {};
380 fence_info.sType = VK_STRUCTURE_TYPE_FENCE_CREATE_INFO;
381 fence_info.flags = VK_FENCE_CREATE_SIGNALED_BIT;
382 VK_CHECK(device_data->vtable.CreateFence(device_data->device,
383 &fence_info,
384 NULL,
385 &data->queries_fence));
386
387 if (data->flags & VK_QUEUE_GRAPHICS_BIT)
388 device_data->graphic_queue = data;
389
390 return data;
391 }
392
393 static void destroy_queue(struct queue_data *data)
394 {
395 struct device_data *device_data = data->device;
396 device_data->vtable.DestroyFence(device_data->device, data->queries_fence, NULL);
397 unmap_object(HKEY(data->queue));
398 ralloc_free(data);
399 }
400
401 static void device_map_queues(struct device_data *data,
402 const VkDeviceCreateInfo *pCreateInfo)
403 {
404 for (uint32_t i = 0; i < pCreateInfo->queueCreateInfoCount; i++)
405 data->n_queues += pCreateInfo->pQueueCreateInfos[i].queueCount;
406 data->queues = ralloc_array(data, struct queue_data *, data->n_queues);
407
408 struct instance_data *instance_data = data->instance;
409 uint32_t n_family_props;
410 instance_data->vtable.GetPhysicalDeviceQueueFamilyProperties(data->physical_device,
411 &n_family_props,
412 NULL);
413 VkQueueFamilyProperties *family_props =
414 (VkQueueFamilyProperties *)malloc(sizeof(VkQueueFamilyProperties) * n_family_props);
415 instance_data->vtable.GetPhysicalDeviceQueueFamilyProperties(data->physical_device,
416 &n_family_props,
417 family_props);
418
419 uint32_t queue_index = 0;
420 for (uint32_t i = 0; i < pCreateInfo->queueCreateInfoCount; i++) {
421 for (uint32_t j = 0; j < pCreateInfo->pQueueCreateInfos[i].queueCount; j++) {
422 VkQueue queue;
423 data->vtable.GetDeviceQueue(data->device,
424 pCreateInfo->pQueueCreateInfos[i].queueFamilyIndex,
425 j, &queue);
426
427 VK_CHECK(data->set_device_loader_data(data->device, queue));
428
429 data->queues[queue_index++] =
430 new_queue_data(queue, &family_props[pCreateInfo->pQueueCreateInfos[i].queueFamilyIndex],
431 pCreateInfo->pQueueCreateInfos[i].queueFamilyIndex, data);
432 }
433 }
434
435 free(family_props);
436 }
437
438 static void device_unmap_queues(struct device_data *data)
439 {
440 for (uint32_t i = 0; i < data->n_queues; i++)
441 destroy_queue(data->queues[i]);
442 }
443
444 static void destroy_device_data(struct device_data *data)
445 {
446 unmap_object(HKEY(data->device));
447 ralloc_free(data);
448 }
449
450 /**/
451 static struct command_buffer_data *new_command_buffer_data(VkCommandBuffer cmd_buffer,
452 VkCommandBufferLevel level,
453 VkQueryPool pipeline_query_pool,
454 VkQueryPool timestamp_query_pool,
455 uint32_t query_index,
456 struct device_data *device_data)
457 {
458 struct command_buffer_data *data = rzalloc(NULL, struct command_buffer_data);
459 data->device = device_data;
460 data->cmd_buffer = cmd_buffer;
461 data->level = level;
462 data->pipeline_query_pool = pipeline_query_pool;
463 data->timestamp_query_pool = timestamp_query_pool;
464 data->query_index = query_index;
465 list_inithead(&data->link);
466 map_object(HKEY(data->cmd_buffer), data);
467 return data;
468 }
469
470 static void destroy_command_buffer_data(struct command_buffer_data *data)
471 {
472 unmap_object(HKEY(data->cmd_buffer));
473 list_delinit(&data->link);
474 ralloc_free(data);
475 }
476
477 /**/
478 static struct swapchain_data *new_swapchain_data(VkSwapchainKHR swapchain,
479 struct device_data *device_data)
480 {
481 struct instance_data *instance_data = device_data->instance;
482 struct swapchain_data *data = rzalloc(NULL, struct swapchain_data);
483 data->device = device_data;
484 data->swapchain = swapchain;
485 data->window_size = ImVec2(instance_data->params.width, instance_data->params.height);
486 list_inithead(&data->draws);
487 map_object(HKEY(data->swapchain), data);
488 return data;
489 }
490
491 static void destroy_swapchain_data(struct swapchain_data *data)
492 {
493 unmap_object(HKEY(data->swapchain));
494 ralloc_free(data);
495 }
496
497 struct overlay_draw *get_overlay_draw(struct swapchain_data *data)
498 {
499 struct device_data *device_data = data->device;
500 struct overlay_draw *draw = list_empty(&data->draws) ?
501 NULL : list_first_entry(&data->draws, struct overlay_draw, link);
502
503 VkSemaphoreCreateInfo sem_info = {};
504 sem_info.sType = VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO;
505
506 if (draw && device_data->vtable.GetFenceStatus(device_data->device, draw->fence) == VK_SUCCESS) {
507 list_del(&draw->link);
508 VK_CHECK(device_data->vtable.ResetFences(device_data->device,
509 1, &draw->fence));
510 list_addtail(&draw->link, &data->draws);
511 return draw;
512 }
513
514 draw = rzalloc(data, struct overlay_draw);
515
516 VkCommandBufferAllocateInfo cmd_buffer_info = {};
517 cmd_buffer_info.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO;
518 cmd_buffer_info.commandPool = data->command_pool;
519 cmd_buffer_info.level = VK_COMMAND_BUFFER_LEVEL_PRIMARY;
520 cmd_buffer_info.commandBufferCount = 1;
521 VK_CHECK(device_data->vtable.AllocateCommandBuffers(device_data->device,
522 &cmd_buffer_info,
523 &draw->command_buffer));
524 VK_CHECK(device_data->set_device_loader_data(device_data->device,
525 draw->command_buffer));
526
527
528 VkFenceCreateInfo fence_info = {};
529 fence_info.sType = VK_STRUCTURE_TYPE_FENCE_CREATE_INFO;
530 VK_CHECK(device_data->vtable.CreateFence(device_data->device,
531 &fence_info,
532 NULL,
533 &draw->fence));
534
535 VK_CHECK(device_data->vtable.CreateSemaphore(device_data->device, &sem_info,
536 NULL, &draw->semaphore));
537
538 list_addtail(&draw->link, &data->draws);
539
540 return draw;
541 }
542
543 static const char *param_unit(enum overlay_param_enabled param)
544 {
545 switch (param) {
546 case OVERLAY_PARAM_ENABLED_frame_timing:
547 case OVERLAY_PARAM_ENABLED_acquire_timing:
548 return "(us)";
549 case OVERLAY_PARAM_ENABLED_gpu_timing:
550 return "(ns)";
551 default:
552 return "";
553 }
554 }
555
556 static void snapshot_swapchain_frame(struct swapchain_data *data)
557 {
558 struct device_data *device_data = data->device;
559 struct instance_data *instance_data = device_data->instance;
560 uint32_t f_idx = data->n_frames % ARRAY_SIZE(data->frames_stats);
561 uint64_t now = os_time_get(); /* us */
562
563 if (data->last_present_time) {
564 data->frame_stats.stats[OVERLAY_PARAM_ENABLED_frame_timing] =
565 now - data->last_present_time;
566 }
567
568 memset(&data->frames_stats[f_idx], 0, sizeof(data->frames_stats[f_idx]));
569 for (int s = 0; s < OVERLAY_PARAM_ENABLED_MAX; s++) {
570 data->frames_stats[f_idx].stats[s] += device_data->frame_stats.stats[s] + data->frame_stats.stats[s];
571 data->accumulated_stats.stats[s] += device_data->frame_stats.stats[s] + data->frame_stats.stats[s];
572 }
573
574 if (data->last_fps_update) {
575 double elapsed = (double)(now - data->last_fps_update); /* us */
576 if (elapsed >= instance_data->params.fps_sampling_period) {
577 data->fps = 1000000.0f * data->n_frames_since_update / elapsed;
578 if (instance_data->params.output_file) {
579 if (!instance_data->first_line_printed) {
580 bool first_column = true;
581
582 instance_data->first_line_printed = true;
583
584 #define OVERLAY_PARAM_BOOL(name) \
585 if (instance_data->params.enabled[OVERLAY_PARAM_ENABLED_##name]) { \
586 fprintf(instance_data->params.output_file, \
587 "%s%s%s", first_column ? "" : ", ", #name, \
588 param_unit(OVERLAY_PARAM_ENABLED_##name)); \
589 first_column = false; \
590 }
591 #define OVERLAY_PARAM_CUSTOM(name)
592 OVERLAY_PARAMS
593 #undef OVERLAY_PARAM_BOOL
594 #undef OVERLAY_PARAM_CUSTOM
595 fprintf(instance_data->params.output_file, "\n");
596 }
597
598 for (int s = 0; s < OVERLAY_PARAM_ENABLED_MAX; s++) {
599 if (!instance_data->params.enabled[s])
600 continue;
601 if (s == OVERLAY_PARAM_ENABLED_fps) {
602 fprintf(instance_data->params.output_file,
603 "%s%.2f", s == 0 ? "" : ", ", data->fps);
604 } else {
605 fprintf(instance_data->params.output_file,
606 "%s%" PRIu64, s == 0 ? "" : ", ",
607 data->accumulated_stats.stats[s]);
608 }
609 }
610 fprintf(instance_data->params.output_file, "\n");
611 fflush(instance_data->params.output_file);
612 }
613
614 memset(&data->accumulated_stats, 0, sizeof(data->accumulated_stats));
615 data->n_frames_since_update = 0;
616 data->last_fps_update = now;
617 }
618 } else {
619 data->last_fps_update = now;
620 }
621
622 memset(&device_data->frame_stats, 0, sizeof(device_data->frame_stats));
623 memset(&data->frame_stats, 0, sizeof(device_data->frame_stats));
624
625 data->last_present_time = now;
626 data->n_frames++;
627 data->n_frames_since_update++;
628 }
629
630 static float get_time_stat(void *_data, int _idx)
631 {
632 struct swapchain_data *data = (struct swapchain_data *) _data;
633 if ((ARRAY_SIZE(data->frames_stats) - _idx) > data->n_frames)
634 return 0.0f;
635 int idx = ARRAY_SIZE(data->frames_stats) +
636 data->n_frames < ARRAY_SIZE(data->frames_stats) ?
637 _idx - data->n_frames :
638 _idx + data->n_frames;
639 idx %= ARRAY_SIZE(data->frames_stats);
640 /* Time stats are in us. */
641 return data->frames_stats[idx].stats[data->stat_selector] / data->time_dividor;
642 }
643
644 static float get_stat(void *_data, int _idx)
645 {
646 struct swapchain_data *data = (struct swapchain_data *) _data;
647 if ((ARRAY_SIZE(data->frames_stats) - _idx) > data->n_frames)
648 return 0.0f;
649 int idx = ARRAY_SIZE(data->frames_stats) +
650 data->n_frames < ARRAY_SIZE(data->frames_stats) ?
651 _idx - data->n_frames :
652 _idx + data->n_frames;
653 idx %= ARRAY_SIZE(data->frames_stats);
654 return data->frames_stats[idx].stats[data->stat_selector];
655 }
656
657 static void position_layer(struct swapchain_data *data)
658
659 {
660 struct device_data *device_data = data->device;
661 struct instance_data *instance_data = device_data->instance;
662 const float margin = 10.0f;
663
664 ImGui::SetNextWindowBgAlpha(0.5);
665 ImGui::SetNextWindowSize(data->window_size, ImGuiCond_Always);
666 switch (instance_data->params.position) {
667 case LAYER_POSITION_TOP_LEFT:
668 ImGui::SetNextWindowPos(ImVec2(margin, margin), ImGuiCond_Always);
669 break;
670 case LAYER_POSITION_TOP_RIGHT:
671 ImGui::SetNextWindowPos(ImVec2(data->width - data->window_size.x - margin, margin),
672 ImGuiCond_Always);
673 break;
674 case LAYER_POSITION_BOTTOM_LEFT:
675 ImGui::SetNextWindowPos(ImVec2(margin, data->height - data->window_size.y - margin),
676 ImGuiCond_Always);
677 break;
678 case LAYER_POSITION_BOTTOM_RIGHT:
679 ImGui::SetNextWindowPos(ImVec2(data->width - data->window_size.x - margin,
680 data->height - data->window_size.y - margin),
681 ImGuiCond_Always);
682 break;
683 }
684 }
685
686 static void compute_swapchain_display(struct swapchain_data *data)
687 {
688 struct device_data *device_data = data->device;
689 struct instance_data *instance_data = device_data->instance;
690
691 ImGui::SetCurrentContext(data->imgui_context);
692 ImGui::NewFrame();
693 position_layer(data);
694 ImGui::Begin("Mesa overlay");
695 ImGui::Text("Device: %s", device_data->properties.deviceName);
696
697 const char *format_name = vk_Format_to_str(data->format);
698 format_name = format_name ? (format_name + strlen("VK_FORMAT_")) : "unknown";
699 ImGui::Text("Swapchain format: %s", format_name);
700 ImGui::Text("Frames: %" PRIu64, data->n_frames);
701 if (instance_data->params.enabled[OVERLAY_PARAM_ENABLED_fps])
702 ImGui::Text("FPS: %.2f" , data->fps);
703
704 /* Recompute min/max */
705 for (uint32_t s = 0; s < OVERLAY_PARAM_ENABLED_MAX; s++) {
706 data->stats_min.stats[s] = UINT64_MAX;
707 data->stats_max.stats[s] = 0;
708 }
709 for (uint32_t f = 0; f < MIN2(data->n_frames, ARRAY_SIZE(data->frames_stats)); f++) {
710 for (uint32_t s = 0; s < OVERLAY_PARAM_ENABLED_MAX; s++) {
711 data->stats_min.stats[s] = MIN2(data->frames_stats[f].stats[s],
712 data->stats_min.stats[s]);
713 data->stats_max.stats[s] = MAX2(data->frames_stats[f].stats[s],
714 data->stats_max.stats[s]);
715 }
716 }
717 for (uint32_t s = 0; s < OVERLAY_PARAM_ENABLED_MAX; s++) {
718 assert(data->stats_min.stats[s] != UINT64_MAX);
719 }
720
721 for (uint32_t s = 0; s < OVERLAY_PARAM_ENABLED_MAX; s++) {
722 if (!instance_data->params.enabled[s] ||
723 s == OVERLAY_PARAM_ENABLED_fps ||
724 s == OVERLAY_PARAM_ENABLED_frame)
725 continue;
726
727 char hash[40];
728 snprintf(hash, sizeof(hash), "##%s", overlay_param_names[s]);
729 data->stat_selector = (enum overlay_param_enabled) s;
730 data->time_dividor = 1000.0f;
731 if (s == OVERLAY_PARAM_ENABLED_gpu_timing)
732 data->time_dividor = 1000000.0f;
733
734 if (s == OVERLAY_PARAM_ENABLED_frame_timing ||
735 s == OVERLAY_PARAM_ENABLED_acquire_timing ||
736 s == OVERLAY_PARAM_ENABLED_gpu_timing) {
737 double min_time = data->stats_min.stats[s] / data->time_dividor;
738 double max_time = data->stats_max.stats[s] / data->time_dividor;
739 ImGui::PlotHistogram(hash, get_time_stat, data,
740 ARRAY_SIZE(data->frames_stats), 0,
741 NULL, min_time, max_time,
742 ImVec2(ImGui::GetContentRegionAvailWidth(), 30));
743 ImGui::Text("%s: %.3fms [%.3f, %.3f]", overlay_param_names[s],
744 get_time_stat(data, ARRAY_SIZE(data->frames_stats) - 1),
745 min_time, max_time);
746 } else {
747 ImGui::PlotHistogram(hash, get_stat, data,
748 ARRAY_SIZE(data->frames_stats), 0,
749 NULL,
750 data->stats_min.stats[s],
751 data->stats_max.stats[s],
752 ImVec2(ImGui::GetContentRegionAvailWidth(), 30));
753 ImGui::Text("%s: %.0f [%" PRIu64 ", %" PRIu64 "]", overlay_param_names[s],
754 get_stat(data, ARRAY_SIZE(data->frames_stats) - 1),
755 data->stats_min.stats[s], data->stats_max.stats[s]);
756 }
757 }
758 data->window_size = ImVec2(data->window_size.x, ImGui::GetCursorPosY() + 10.0f);
759 ImGui::End();
760 ImGui::EndFrame();
761 ImGui::Render();
762 }
763
764 static uint32_t vk_memory_type(struct device_data *data,
765 VkMemoryPropertyFlags properties,
766 uint32_t type_bits)
767 {
768 VkPhysicalDeviceMemoryProperties prop;
769 data->instance->vtable.GetPhysicalDeviceMemoryProperties(data->physical_device, &prop);
770 for (uint32_t i = 0; i < prop.memoryTypeCount; i++)
771 if ((prop.memoryTypes[i].propertyFlags & properties) == properties && type_bits & (1<<i))
772 return i;
773 return 0xFFFFFFFF; // Unable to find memoryType
774 }
775
776 static void ensure_swapchain_fonts(struct swapchain_data *data,
777 VkCommandBuffer command_buffer)
778 {
779 if (data->font_uploaded)
780 return;
781
782 data->font_uploaded = true;
783
784 struct device_data *device_data = data->device;
785 ImGuiIO& io = ImGui::GetIO();
786 unsigned char* pixels;
787 int width, height;
788 io.Fonts->GetTexDataAsRGBA32(&pixels, &width, &height);
789 size_t upload_size = width * height * 4 * sizeof(char);
790
791 /* Upload buffer */
792 VkBufferCreateInfo buffer_info = {};
793 buffer_info.sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO;
794 buffer_info.size = upload_size;
795 buffer_info.usage = VK_BUFFER_USAGE_TRANSFER_SRC_BIT;
796 buffer_info.sharingMode = VK_SHARING_MODE_EXCLUSIVE;
797 VK_CHECK(device_data->vtable.CreateBuffer(device_data->device, &buffer_info,
798 NULL, &data->upload_font_buffer));
799 VkMemoryRequirements upload_buffer_req;
800 device_data->vtable.GetBufferMemoryRequirements(device_data->device,
801 data->upload_font_buffer,
802 &upload_buffer_req);
803 VkMemoryAllocateInfo upload_alloc_info = {};
804 upload_alloc_info.sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO;
805 upload_alloc_info.allocationSize = upload_buffer_req.size;
806 upload_alloc_info.memoryTypeIndex = vk_memory_type(device_data,
807 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT,
808 upload_buffer_req.memoryTypeBits);
809 VK_CHECK(device_data->vtable.AllocateMemory(device_data->device,
810 &upload_alloc_info,
811 NULL,
812 &data->upload_font_buffer_mem));
813 VK_CHECK(device_data->vtable.BindBufferMemory(device_data->device,
814 data->upload_font_buffer,
815 data->upload_font_buffer_mem, 0));
816
817 /* Upload to Buffer */
818 char* map = NULL;
819 VK_CHECK(device_data->vtable.MapMemory(device_data->device,
820 data->upload_font_buffer_mem,
821 0, upload_size, 0, (void**)(&map)));
822 memcpy(map, pixels, upload_size);
823 VkMappedMemoryRange range[1] = {};
824 range[0].sType = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE;
825 range[0].memory = data->upload_font_buffer_mem;
826 range[0].size = upload_size;
827 VK_CHECK(device_data->vtable.FlushMappedMemoryRanges(device_data->device, 1, range));
828 device_data->vtable.UnmapMemory(device_data->device,
829 data->upload_font_buffer_mem);
830
831 /* Copy buffer to image */
832 VkImageMemoryBarrier copy_barrier[1] = {};
833 copy_barrier[0].sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER;
834 copy_barrier[0].dstAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT;
835 copy_barrier[0].oldLayout = VK_IMAGE_LAYOUT_UNDEFINED;
836 copy_barrier[0].newLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
837 copy_barrier[0].srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
838 copy_barrier[0].dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
839 copy_barrier[0].image = data->font_image;
840 copy_barrier[0].subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
841 copy_barrier[0].subresourceRange.levelCount = 1;
842 copy_barrier[0].subresourceRange.layerCount = 1;
843 device_data->vtable.CmdPipelineBarrier(command_buffer,
844 VK_PIPELINE_STAGE_HOST_BIT,
845 VK_PIPELINE_STAGE_TRANSFER_BIT,
846 0, 0, NULL, 0, NULL,
847 1, copy_barrier);
848
849 VkBufferImageCopy region = {};
850 region.imageSubresource.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
851 region.imageSubresource.layerCount = 1;
852 region.imageExtent.width = width;
853 region.imageExtent.height = height;
854 region.imageExtent.depth = 1;
855 device_data->vtable.CmdCopyBufferToImage(command_buffer,
856 data->upload_font_buffer,
857 data->font_image,
858 VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
859 1, &region);
860
861 VkImageMemoryBarrier use_barrier[1] = {};
862 use_barrier[0].sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER;
863 use_barrier[0].srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT;
864 use_barrier[0].dstAccessMask = VK_ACCESS_SHADER_READ_BIT;
865 use_barrier[0].oldLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
866 use_barrier[0].newLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
867 use_barrier[0].srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
868 use_barrier[0].dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
869 use_barrier[0].image = data->font_image;
870 use_barrier[0].subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
871 use_barrier[0].subresourceRange.levelCount = 1;
872 use_barrier[0].subresourceRange.layerCount = 1;
873 device_data->vtable.CmdPipelineBarrier(command_buffer,
874 VK_PIPELINE_STAGE_TRANSFER_BIT,
875 VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT,
876 0,
877 0, NULL,
878 0, NULL,
879 1, use_barrier);
880
881 /* Store our identifier */
882 io.Fonts->TexID = (ImTextureID)(intptr_t)data->font_image;
883 }
884
885 static void CreateOrResizeBuffer(struct device_data *data,
886 VkBuffer *buffer,
887 VkDeviceMemory *buffer_memory,
888 VkDeviceSize *buffer_size,
889 size_t new_size, VkBufferUsageFlagBits usage)
890 {
891 if (*buffer != VK_NULL_HANDLE)
892 data->vtable.DestroyBuffer(data->device, *buffer, NULL);
893 if (*buffer_memory)
894 data->vtable.FreeMemory(data->device, *buffer_memory, NULL);
895
896 VkBufferCreateInfo buffer_info = {};
897 buffer_info.sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO;
898 buffer_info.size = new_size;
899 buffer_info.usage = usage;
900 buffer_info.sharingMode = VK_SHARING_MODE_EXCLUSIVE;
901 VK_CHECK(data->vtable.CreateBuffer(data->device, &buffer_info, NULL, buffer));
902
903 VkMemoryRequirements req;
904 data->vtable.GetBufferMemoryRequirements(data->device, *buffer, &req);
905 VkMemoryAllocateInfo alloc_info = {};
906 alloc_info.sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO;
907 alloc_info.allocationSize = req.size;
908 alloc_info.memoryTypeIndex =
909 vk_memory_type(data, VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT, req.memoryTypeBits);
910 VK_CHECK(data->vtable.AllocateMemory(data->device, &alloc_info, NULL, buffer_memory));
911
912 VK_CHECK(data->vtable.BindBufferMemory(data->device, *buffer, *buffer_memory, 0));
913 *buffer_size = new_size;
914 }
915
916 static struct overlay_draw *render_swapchain_display(struct swapchain_data *data,
917 const VkSemaphore *wait_semaphores,
918 unsigned n_wait_semaphores,
919 unsigned image_index)
920 {
921 ImDrawData* draw_data = ImGui::GetDrawData();
922 if (draw_data->TotalVtxCount == 0)
923 return NULL;
924
925 struct device_data *device_data = data->device;
926 struct overlay_draw *draw = get_overlay_draw(data);
927
928 device_data->vtable.ResetCommandBuffer(draw->command_buffer, 0);
929
930 VkRenderPassBeginInfo render_pass_info = {};
931 render_pass_info.sType = VK_STRUCTURE_TYPE_RENDER_PASS_BEGIN_INFO;
932 render_pass_info.renderPass = data->render_pass;
933 render_pass_info.framebuffer = data->framebuffers[image_index];
934 render_pass_info.renderArea.extent.width = data->width;
935 render_pass_info.renderArea.extent.height = data->height;
936
937 VkCommandBufferBeginInfo buffer_begin_info = {};
938 buffer_begin_info.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO;
939
940 device_data->vtable.BeginCommandBuffer(draw->command_buffer, &buffer_begin_info);
941
942 ensure_swapchain_fonts(data, draw->command_buffer);
943
944 /* Bounce the image to display back to color attachment layout for
945 * rendering on top of it.
946 */
947 VkImageMemoryBarrier imb;
948 imb.sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER;
949 imb.pNext = nullptr;
950 imb.srcAccessMask = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT;
951 imb.dstAccessMask = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT;
952 imb.oldLayout = VK_IMAGE_LAYOUT_PRESENT_SRC_KHR;
953 imb.newLayout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL;
954 imb.image = data->images[image_index];
955 imb.subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
956 imb.subresourceRange.baseMipLevel = 0;
957 imb.subresourceRange.levelCount = 1;
958 imb.subresourceRange.baseArrayLayer = 0;
959 imb.subresourceRange.layerCount = 1;
960 imb.srcQueueFamilyIndex = device_data->graphic_queue->family_index;
961 imb.dstQueueFamilyIndex = device_data->graphic_queue->family_index;
962 device_data->vtable.CmdPipelineBarrier(draw->command_buffer,
963 VK_PIPELINE_STAGE_ALL_GRAPHICS_BIT,
964 VK_PIPELINE_STAGE_ALL_GRAPHICS_BIT,
965 0, /* dependency flags */
966 0, nullptr, /* memory barriers */
967 0, nullptr, /* buffer memory barriers */
968 1, &imb); /* image memory barriers */
969
970 device_data->vtable.CmdBeginRenderPass(draw->command_buffer, &render_pass_info,
971 VK_SUBPASS_CONTENTS_INLINE);
972
973 /* Create/Resize vertex & index buffers */
974 size_t vertex_size = draw_data->TotalVtxCount * sizeof(ImDrawVert);
975 size_t index_size = draw_data->TotalIdxCount * sizeof(ImDrawIdx);
976 if (draw->vertex_buffer_size < vertex_size) {
977 CreateOrResizeBuffer(device_data,
978 &draw->vertex_buffer,
979 &draw->vertex_buffer_mem,
980 &draw->vertex_buffer_size,
981 vertex_size, VK_BUFFER_USAGE_VERTEX_BUFFER_BIT);
982 }
983 if (draw->index_buffer_size < index_size) {
984 CreateOrResizeBuffer(device_data,
985 &draw->index_buffer,
986 &draw->index_buffer_mem,
987 &draw->index_buffer_size,
988 index_size, VK_BUFFER_USAGE_INDEX_BUFFER_BIT);
989 }
990
991 /* Upload vertex & index data */
992 ImDrawVert* vtx_dst = NULL;
993 ImDrawIdx* idx_dst = NULL;
994 VK_CHECK(device_data->vtable.MapMemory(device_data->device, draw->vertex_buffer_mem,
995 0, vertex_size, 0, (void**)(&vtx_dst)));
996 VK_CHECK(device_data->vtable.MapMemory(device_data->device, draw->index_buffer_mem,
997 0, index_size, 0, (void**)(&idx_dst)));
998 for (int n = 0; n < draw_data->CmdListsCount; n++)
999 {
1000 const ImDrawList* cmd_list = draw_data->CmdLists[n];
1001 memcpy(vtx_dst, cmd_list->VtxBuffer.Data, cmd_list->VtxBuffer.Size * sizeof(ImDrawVert));
1002 memcpy(idx_dst, cmd_list->IdxBuffer.Data, cmd_list->IdxBuffer.Size * sizeof(ImDrawIdx));
1003 vtx_dst += cmd_list->VtxBuffer.Size;
1004 idx_dst += cmd_list->IdxBuffer.Size;
1005 }
1006 VkMappedMemoryRange range[2] = {};
1007 range[0].sType = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE;
1008 range[0].memory = draw->vertex_buffer_mem;
1009 range[0].size = VK_WHOLE_SIZE;
1010 range[1].sType = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE;
1011 range[1].memory = draw->index_buffer_mem;
1012 range[1].size = VK_WHOLE_SIZE;
1013 VK_CHECK(device_data->vtable.FlushMappedMemoryRanges(device_data->device, 2, range));
1014 device_data->vtable.UnmapMemory(device_data->device, draw->vertex_buffer_mem);
1015 device_data->vtable.UnmapMemory(device_data->device, draw->index_buffer_mem);
1016
1017 /* Bind pipeline and descriptor sets */
1018 device_data->vtable.CmdBindPipeline(draw->command_buffer, VK_PIPELINE_BIND_POINT_GRAPHICS, data->pipeline);
1019 VkDescriptorSet desc_set[1] = { data->descriptor_set };
1020 device_data->vtable.CmdBindDescriptorSets(draw->command_buffer, VK_PIPELINE_BIND_POINT_GRAPHICS,
1021 data->pipeline_layout, 0, 1, desc_set, 0, NULL);
1022
1023 /* Bind vertex & index buffers */
1024 VkBuffer vertex_buffers[1] = { draw->vertex_buffer };
1025 VkDeviceSize vertex_offset[1] = { 0 };
1026 device_data->vtable.CmdBindVertexBuffers(draw->command_buffer, 0, 1, vertex_buffers, vertex_offset);
1027 device_data->vtable.CmdBindIndexBuffer(draw->command_buffer, draw->index_buffer, 0, VK_INDEX_TYPE_UINT16);
1028
1029 /* Setup viewport */
1030 VkViewport viewport;
1031 viewport.x = 0;
1032 viewport.y = 0;
1033 viewport.width = draw_data->DisplaySize.x;
1034 viewport.height = draw_data->DisplaySize.y;
1035 viewport.minDepth = 0.0f;
1036 viewport.maxDepth = 1.0f;
1037 device_data->vtable.CmdSetViewport(draw->command_buffer, 0, 1, &viewport);
1038
1039
1040 /* Setup scale and translation through push constants :
1041 *
1042 * Our visible imgui space lies from draw_data->DisplayPos (top left) to
1043 * draw_data->DisplayPos+data_data->DisplaySize (bottom right). DisplayMin
1044 * is typically (0,0) for single viewport apps.
1045 */
1046 float scale[2];
1047 scale[0] = 2.0f / draw_data->DisplaySize.x;
1048 scale[1] = 2.0f / draw_data->DisplaySize.y;
1049 float translate[2];
1050 translate[0] = -1.0f - draw_data->DisplayPos.x * scale[0];
1051 translate[1] = -1.0f - draw_data->DisplayPos.y * scale[1];
1052 device_data->vtable.CmdPushConstants(draw->command_buffer, data->pipeline_layout,
1053 VK_SHADER_STAGE_VERTEX_BIT,
1054 sizeof(float) * 0, sizeof(float) * 2, scale);
1055 device_data->vtable.CmdPushConstants(draw->command_buffer, data->pipeline_layout,
1056 VK_SHADER_STAGE_VERTEX_BIT,
1057 sizeof(float) * 2, sizeof(float) * 2, translate);
1058
1059 // Render the command lists:
1060 int vtx_offset = 0;
1061 int idx_offset = 0;
1062 ImVec2 display_pos = draw_data->DisplayPos;
1063 for (int n = 0; n < draw_data->CmdListsCount; n++)
1064 {
1065 const ImDrawList* cmd_list = draw_data->CmdLists[n];
1066 for (int cmd_i = 0; cmd_i < cmd_list->CmdBuffer.Size; cmd_i++)
1067 {
1068 const ImDrawCmd* pcmd = &cmd_list->CmdBuffer[cmd_i];
1069 // Apply scissor/clipping rectangle
1070 // FIXME: We could clamp width/height based on clamped min/max values.
1071 VkRect2D scissor;
1072 scissor.offset.x = (int32_t)(pcmd->ClipRect.x - display_pos.x) > 0 ? (int32_t)(pcmd->ClipRect.x - display_pos.x) : 0;
1073 scissor.offset.y = (int32_t)(pcmd->ClipRect.y - display_pos.y) > 0 ? (int32_t)(pcmd->ClipRect.y - display_pos.y) : 0;
1074 scissor.extent.width = (uint32_t)(pcmd->ClipRect.z - pcmd->ClipRect.x);
1075 scissor.extent.height = (uint32_t)(pcmd->ClipRect.w - pcmd->ClipRect.y + 1); // FIXME: Why +1 here?
1076 device_data->vtable.CmdSetScissor(draw->command_buffer, 0, 1, &scissor);
1077
1078 // Draw
1079 device_data->vtable.CmdDrawIndexed(draw->command_buffer, pcmd->ElemCount, 1, idx_offset, vtx_offset, 0);
1080
1081 idx_offset += pcmd->ElemCount;
1082 }
1083 vtx_offset += cmd_list->VtxBuffer.Size;
1084 }
1085
1086 device_data->vtable.CmdEndRenderPass(draw->command_buffer);
1087 device_data->vtable.EndCommandBuffer(draw->command_buffer);
1088
1089 VkSubmitInfo submit_info = {};
1090 VkPipelineStageFlags stage_wait = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
1091 submit_info.sType = VK_STRUCTURE_TYPE_SUBMIT_INFO;
1092 submit_info.commandBufferCount = 1;
1093 submit_info.pCommandBuffers = &draw->command_buffer;
1094 submit_info.pWaitDstStageMask = &stage_wait;
1095 submit_info.waitSemaphoreCount = n_wait_semaphores;
1096 submit_info.pWaitSemaphores = wait_semaphores;
1097 submit_info.signalSemaphoreCount = 1;
1098 submit_info.pSignalSemaphores = &draw->semaphore;
1099
1100 device_data->vtable.QueueSubmit(device_data->graphic_queue->queue, 1, &submit_info, draw->fence);
1101
1102 return draw;
1103 }
1104
1105 static const uint32_t overlay_vert_spv[] = {
1106 #include "overlay.vert.spv.h"
1107 };
1108 static const uint32_t overlay_frag_spv[] = {
1109 #include "overlay.frag.spv.h"
1110 };
1111
1112 static void setup_swapchain_data_pipeline(struct swapchain_data *data)
1113 {
1114 struct device_data *device_data = data->device;
1115 VkShaderModule vert_module, frag_module;
1116
1117 /* Create shader modules */
1118 VkShaderModuleCreateInfo vert_info = {};
1119 vert_info.sType = VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO;
1120 vert_info.codeSize = sizeof(overlay_vert_spv);
1121 vert_info.pCode = overlay_vert_spv;
1122 VK_CHECK(device_data->vtable.CreateShaderModule(device_data->device,
1123 &vert_info, NULL, &vert_module));
1124 VkShaderModuleCreateInfo frag_info = {};
1125 frag_info.sType = VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO;
1126 frag_info.codeSize = sizeof(overlay_frag_spv);
1127 frag_info.pCode = (uint32_t*)overlay_frag_spv;
1128 VK_CHECK(device_data->vtable.CreateShaderModule(device_data->device,
1129 &frag_info, NULL, &frag_module));
1130
1131 /* Font sampler */
1132 VkSamplerCreateInfo sampler_info = {};
1133 sampler_info.sType = VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO;
1134 sampler_info.magFilter = VK_FILTER_LINEAR;
1135 sampler_info.minFilter = VK_FILTER_LINEAR;
1136 sampler_info.mipmapMode = VK_SAMPLER_MIPMAP_MODE_LINEAR;
1137 sampler_info.addressModeU = VK_SAMPLER_ADDRESS_MODE_REPEAT;
1138 sampler_info.addressModeV = VK_SAMPLER_ADDRESS_MODE_REPEAT;
1139 sampler_info.addressModeW = VK_SAMPLER_ADDRESS_MODE_REPEAT;
1140 sampler_info.minLod = -1000;
1141 sampler_info.maxLod = 1000;
1142 sampler_info.maxAnisotropy = 1.0f;
1143 VK_CHECK(device_data->vtable.CreateSampler(device_data->device, &sampler_info,
1144 NULL, &data->font_sampler));
1145
1146 /* Descriptor pool */
1147 VkDescriptorPoolSize sampler_pool_size = {};
1148 sampler_pool_size.type = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER;
1149 sampler_pool_size.descriptorCount = 1;
1150 VkDescriptorPoolCreateInfo desc_pool_info = {};
1151 desc_pool_info.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_POOL_CREATE_INFO;
1152 desc_pool_info.maxSets = 1;
1153 desc_pool_info.poolSizeCount = 1;
1154 desc_pool_info.pPoolSizes = &sampler_pool_size;
1155 VK_CHECK(device_data->vtable.CreateDescriptorPool(device_data->device,
1156 &desc_pool_info,
1157 NULL, &data->descriptor_pool));
1158
1159 /* Descriptor layout */
1160 VkSampler sampler[1] = { data->font_sampler };
1161 VkDescriptorSetLayoutBinding binding[1] = {};
1162 binding[0].descriptorType = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER;
1163 binding[0].descriptorCount = 1;
1164 binding[0].stageFlags = VK_SHADER_STAGE_FRAGMENT_BIT;
1165 binding[0].pImmutableSamplers = sampler;
1166 VkDescriptorSetLayoutCreateInfo set_layout_info = {};
1167 set_layout_info.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO;
1168 set_layout_info.bindingCount = 1;
1169 set_layout_info.pBindings = binding;
1170 VK_CHECK(device_data->vtable.CreateDescriptorSetLayout(device_data->device,
1171 &set_layout_info,
1172 NULL, &data->descriptor_layout));
1173
1174 /* Descriptor set */
1175 VkDescriptorSetAllocateInfo alloc_info = {};
1176 alloc_info.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_ALLOCATE_INFO;
1177 alloc_info.descriptorPool = data->descriptor_pool;
1178 alloc_info.descriptorSetCount = 1;
1179 alloc_info.pSetLayouts = &data->descriptor_layout;
1180 VK_CHECK(device_data->vtable.AllocateDescriptorSets(device_data->device,
1181 &alloc_info,
1182 &data->descriptor_set));
1183
1184 /* Constants: we are using 'vec2 offset' and 'vec2 scale' instead of a full
1185 * 3d projection matrix
1186 */
1187 VkPushConstantRange push_constants[1] = {};
1188 push_constants[0].stageFlags = VK_SHADER_STAGE_VERTEX_BIT;
1189 push_constants[0].offset = sizeof(float) * 0;
1190 push_constants[0].size = sizeof(float) * 4;
1191 VkPipelineLayoutCreateInfo layout_info = {};
1192 layout_info.sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO;
1193 layout_info.setLayoutCount = 1;
1194 layout_info.pSetLayouts = &data->descriptor_layout;
1195 layout_info.pushConstantRangeCount = 1;
1196 layout_info.pPushConstantRanges = push_constants;
1197 VK_CHECK(device_data->vtable.CreatePipelineLayout(device_data->device,
1198 &layout_info,
1199 NULL, &data->pipeline_layout));
1200
1201 VkPipelineShaderStageCreateInfo stage[2] = {};
1202 stage[0].sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO;
1203 stage[0].stage = VK_SHADER_STAGE_VERTEX_BIT;
1204 stage[0].module = vert_module;
1205 stage[0].pName = "main";
1206 stage[1].sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO;
1207 stage[1].stage = VK_SHADER_STAGE_FRAGMENT_BIT;
1208 stage[1].module = frag_module;
1209 stage[1].pName = "main";
1210
1211 VkVertexInputBindingDescription binding_desc[1] = {};
1212 binding_desc[0].stride = sizeof(ImDrawVert);
1213 binding_desc[0].inputRate = VK_VERTEX_INPUT_RATE_VERTEX;
1214
1215 VkVertexInputAttributeDescription attribute_desc[3] = {};
1216 attribute_desc[0].location = 0;
1217 attribute_desc[0].binding = binding_desc[0].binding;
1218 attribute_desc[0].format = VK_FORMAT_R32G32_SFLOAT;
1219 attribute_desc[0].offset = IM_OFFSETOF(ImDrawVert, pos);
1220 attribute_desc[1].location = 1;
1221 attribute_desc[1].binding = binding_desc[0].binding;
1222 attribute_desc[1].format = VK_FORMAT_R32G32_SFLOAT;
1223 attribute_desc[1].offset = IM_OFFSETOF(ImDrawVert, uv);
1224 attribute_desc[2].location = 2;
1225 attribute_desc[2].binding = binding_desc[0].binding;
1226 attribute_desc[2].format = VK_FORMAT_R8G8B8A8_UNORM;
1227 attribute_desc[2].offset = IM_OFFSETOF(ImDrawVert, col);
1228
1229 VkPipelineVertexInputStateCreateInfo vertex_info = {};
1230 vertex_info.sType = VK_STRUCTURE_TYPE_PIPELINE_VERTEX_INPUT_STATE_CREATE_INFO;
1231 vertex_info.vertexBindingDescriptionCount = 1;
1232 vertex_info.pVertexBindingDescriptions = binding_desc;
1233 vertex_info.vertexAttributeDescriptionCount = 3;
1234 vertex_info.pVertexAttributeDescriptions = attribute_desc;
1235
1236 VkPipelineInputAssemblyStateCreateInfo ia_info = {};
1237 ia_info.sType = VK_STRUCTURE_TYPE_PIPELINE_INPUT_ASSEMBLY_STATE_CREATE_INFO;
1238 ia_info.topology = VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST;
1239
1240 VkPipelineViewportStateCreateInfo viewport_info = {};
1241 viewport_info.sType = VK_STRUCTURE_TYPE_PIPELINE_VIEWPORT_STATE_CREATE_INFO;
1242 viewport_info.viewportCount = 1;
1243 viewport_info.scissorCount = 1;
1244
1245 VkPipelineRasterizationStateCreateInfo raster_info = {};
1246 raster_info.sType = VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_STATE_CREATE_INFO;
1247 raster_info.polygonMode = VK_POLYGON_MODE_FILL;
1248 raster_info.cullMode = VK_CULL_MODE_NONE;
1249 raster_info.frontFace = VK_FRONT_FACE_COUNTER_CLOCKWISE;
1250 raster_info.lineWidth = 1.0f;
1251
1252 VkPipelineMultisampleStateCreateInfo ms_info = {};
1253 ms_info.sType = VK_STRUCTURE_TYPE_PIPELINE_MULTISAMPLE_STATE_CREATE_INFO;
1254 ms_info.rasterizationSamples = VK_SAMPLE_COUNT_1_BIT;
1255
1256 VkPipelineColorBlendAttachmentState color_attachment[1] = {};
1257 color_attachment[0].blendEnable = VK_TRUE;
1258 color_attachment[0].srcColorBlendFactor = VK_BLEND_FACTOR_SRC_ALPHA;
1259 color_attachment[0].dstColorBlendFactor = VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA;
1260 color_attachment[0].colorBlendOp = VK_BLEND_OP_ADD;
1261 color_attachment[0].srcAlphaBlendFactor = VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA;
1262 color_attachment[0].dstAlphaBlendFactor = VK_BLEND_FACTOR_ZERO;
1263 color_attachment[0].alphaBlendOp = VK_BLEND_OP_ADD;
1264 color_attachment[0].colorWriteMask = VK_COLOR_COMPONENT_R_BIT |
1265 VK_COLOR_COMPONENT_G_BIT | VK_COLOR_COMPONENT_B_BIT | VK_COLOR_COMPONENT_A_BIT;
1266
1267 VkPipelineDepthStencilStateCreateInfo depth_info = {};
1268 depth_info.sType = VK_STRUCTURE_TYPE_PIPELINE_DEPTH_STENCIL_STATE_CREATE_INFO;
1269
1270 VkPipelineColorBlendStateCreateInfo blend_info = {};
1271 blend_info.sType = VK_STRUCTURE_TYPE_PIPELINE_COLOR_BLEND_STATE_CREATE_INFO;
1272 blend_info.attachmentCount = 1;
1273 blend_info.pAttachments = color_attachment;
1274
1275 VkDynamicState dynamic_states[2] = { VK_DYNAMIC_STATE_VIEWPORT, VK_DYNAMIC_STATE_SCISSOR };
1276 VkPipelineDynamicStateCreateInfo dynamic_state = {};
1277 dynamic_state.sType = VK_STRUCTURE_TYPE_PIPELINE_DYNAMIC_STATE_CREATE_INFO;
1278 dynamic_state.dynamicStateCount = (uint32_t)IM_ARRAYSIZE(dynamic_states);
1279 dynamic_state.pDynamicStates = dynamic_states;
1280
1281 VkGraphicsPipelineCreateInfo info = {};
1282 info.sType = VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO;
1283 info.flags = 0;
1284 info.stageCount = 2;
1285 info.pStages = stage;
1286 info.pVertexInputState = &vertex_info;
1287 info.pInputAssemblyState = &ia_info;
1288 info.pViewportState = &viewport_info;
1289 info.pRasterizationState = &raster_info;
1290 info.pMultisampleState = &ms_info;
1291 info.pDepthStencilState = &depth_info;
1292 info.pColorBlendState = &blend_info;
1293 info.pDynamicState = &dynamic_state;
1294 info.layout = data->pipeline_layout;
1295 info.renderPass = data->render_pass;
1296 VK_CHECK(
1297 device_data->vtable.CreateGraphicsPipelines(device_data->device, VK_NULL_HANDLE,
1298 1, &info,
1299 NULL, &data->pipeline));
1300
1301 device_data->vtable.DestroyShaderModule(device_data->device, vert_module, NULL);
1302 device_data->vtable.DestroyShaderModule(device_data->device, frag_module, NULL);
1303
1304 ImGuiIO& io = ImGui::GetIO();
1305 unsigned char* pixels;
1306 int width, height;
1307 io.Fonts->GetTexDataAsRGBA32(&pixels, &width, &height);
1308
1309 /* Font image */
1310 VkImageCreateInfo image_info = {};
1311 image_info.sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO;
1312 image_info.imageType = VK_IMAGE_TYPE_2D;
1313 image_info.format = VK_FORMAT_R8G8B8A8_UNORM;
1314 image_info.extent.width = width;
1315 image_info.extent.height = height;
1316 image_info.extent.depth = 1;
1317 image_info.mipLevels = 1;
1318 image_info.arrayLayers = 1;
1319 image_info.samples = VK_SAMPLE_COUNT_1_BIT;
1320 image_info.tiling = VK_IMAGE_TILING_OPTIMAL;
1321 image_info.usage = VK_IMAGE_USAGE_SAMPLED_BIT | VK_IMAGE_USAGE_TRANSFER_DST_BIT;
1322 image_info.sharingMode = VK_SHARING_MODE_EXCLUSIVE;
1323 image_info.initialLayout = VK_IMAGE_LAYOUT_UNDEFINED;
1324 VK_CHECK(device_data->vtable.CreateImage(device_data->device, &image_info,
1325 NULL, &data->font_image));
1326 VkMemoryRequirements font_image_req;
1327 device_data->vtable.GetImageMemoryRequirements(device_data->device,
1328 data->font_image, &font_image_req);
1329 VkMemoryAllocateInfo image_alloc_info = {};
1330 image_alloc_info.sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO;
1331 image_alloc_info.allocationSize = font_image_req.size;
1332 image_alloc_info.memoryTypeIndex = vk_memory_type(device_data,
1333 VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT,
1334 font_image_req.memoryTypeBits);
1335 VK_CHECK(device_data->vtable.AllocateMemory(device_data->device, &image_alloc_info,
1336 NULL, &data->font_mem));
1337 VK_CHECK(device_data->vtable.BindImageMemory(device_data->device,
1338 data->font_image,
1339 data->font_mem, 0));
1340
1341 /* Font image view */
1342 VkImageViewCreateInfo view_info = {};
1343 view_info.sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO;
1344 view_info.image = data->font_image;
1345 view_info.viewType = VK_IMAGE_VIEW_TYPE_2D;
1346 view_info.format = VK_FORMAT_R8G8B8A8_UNORM;
1347 view_info.subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
1348 view_info.subresourceRange.levelCount = 1;
1349 view_info.subresourceRange.layerCount = 1;
1350 VK_CHECK(device_data->vtable.CreateImageView(device_data->device, &view_info,
1351 NULL, &data->font_image_view));
1352
1353 /* Descriptor set */
1354 VkDescriptorImageInfo desc_image[1] = {};
1355 desc_image[0].sampler = data->font_sampler;
1356 desc_image[0].imageView = data->font_image_view;
1357 desc_image[0].imageLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
1358 VkWriteDescriptorSet write_desc[1] = {};
1359 write_desc[0].sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET;
1360 write_desc[0].dstSet = data->descriptor_set;
1361 write_desc[0].descriptorCount = 1;
1362 write_desc[0].descriptorType = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER;
1363 write_desc[0].pImageInfo = desc_image;
1364 device_data->vtable.UpdateDescriptorSets(device_data->device, 1, write_desc, 0, NULL);
1365 }
1366
1367 static void setup_swapchain_data(struct swapchain_data *data,
1368 const VkSwapchainCreateInfoKHR *pCreateInfo)
1369 {
1370 data->width = pCreateInfo->imageExtent.width;
1371 data->height = pCreateInfo->imageExtent.height;
1372 data->format = pCreateInfo->imageFormat;
1373
1374 data->imgui_context = ImGui::CreateContext();
1375 ImGui::SetCurrentContext(data->imgui_context);
1376
1377 ImGui::GetIO().IniFilename = NULL;
1378 ImGui::GetIO().DisplaySize = ImVec2((float)data->width, (float)data->height);
1379
1380 struct device_data *device_data = data->device;
1381
1382 /* Render pass */
1383 VkAttachmentDescription attachment_desc = {};
1384 attachment_desc.format = pCreateInfo->imageFormat;
1385 attachment_desc.samples = VK_SAMPLE_COUNT_1_BIT;
1386 attachment_desc.loadOp = VK_ATTACHMENT_LOAD_OP_LOAD;
1387 attachment_desc.storeOp = VK_ATTACHMENT_STORE_OP_STORE;
1388 attachment_desc.stencilLoadOp = VK_ATTACHMENT_LOAD_OP_DONT_CARE;
1389 attachment_desc.stencilStoreOp = VK_ATTACHMENT_STORE_OP_DONT_CARE;
1390 attachment_desc.initialLayout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL;
1391 attachment_desc.finalLayout = VK_IMAGE_LAYOUT_PRESENT_SRC_KHR;
1392 VkAttachmentReference color_attachment = {};
1393 color_attachment.attachment = 0;
1394 color_attachment.layout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL;
1395 VkSubpassDescription subpass = {};
1396 subpass.pipelineBindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS;
1397 subpass.colorAttachmentCount = 1;
1398 subpass.pColorAttachments = &color_attachment;
1399 VkSubpassDependency dependency = {};
1400 dependency.srcSubpass = VK_SUBPASS_EXTERNAL;
1401 dependency.dstSubpass = 0;
1402 dependency.srcStageMask = VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT;
1403 dependency.dstStageMask = VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT;
1404 dependency.srcAccessMask = 0;
1405 dependency.dstAccessMask = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT;
1406 VkRenderPassCreateInfo render_pass_info = {};
1407 render_pass_info.sType = VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO;
1408 render_pass_info.attachmentCount = 1;
1409 render_pass_info.pAttachments = &attachment_desc;
1410 render_pass_info.subpassCount = 1;
1411 render_pass_info.pSubpasses = &subpass;
1412 render_pass_info.dependencyCount = 1;
1413 render_pass_info.pDependencies = &dependency;
1414 VK_CHECK(device_data->vtable.CreateRenderPass(device_data->device,
1415 &render_pass_info,
1416 NULL, &data->render_pass));
1417
1418 setup_swapchain_data_pipeline(data);
1419
1420 VK_CHECK(device_data->vtable.GetSwapchainImagesKHR(device_data->device,
1421 data->swapchain,
1422 &data->n_images,
1423 NULL));
1424
1425 data->images = ralloc_array(data, VkImage, data->n_images);
1426 data->image_views = ralloc_array(data, VkImageView, data->n_images);
1427 data->framebuffers = ralloc_array(data, VkFramebuffer, data->n_images);
1428
1429 VK_CHECK(device_data->vtable.GetSwapchainImagesKHR(device_data->device,
1430 data->swapchain,
1431 &data->n_images,
1432 data->images));
1433
1434 /* Image views */
1435 VkImageViewCreateInfo view_info = {};
1436 view_info.sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO;
1437 view_info.viewType = VK_IMAGE_VIEW_TYPE_2D;
1438 view_info.format = pCreateInfo->imageFormat;
1439 view_info.components.r = VK_COMPONENT_SWIZZLE_R;
1440 view_info.components.g = VK_COMPONENT_SWIZZLE_G;
1441 view_info.components.b = VK_COMPONENT_SWIZZLE_B;
1442 view_info.components.a = VK_COMPONENT_SWIZZLE_A;
1443 view_info.subresourceRange = { VK_IMAGE_ASPECT_COLOR_BIT, 0, 1, 0, 1 };
1444 for (uint32_t i = 0; i < data->n_images; i++) {
1445 view_info.image = data->images[i];
1446 VK_CHECK(device_data->vtable.CreateImageView(device_data->device,
1447 &view_info, NULL,
1448 &data->image_views[i]));
1449 }
1450
1451 /* Framebuffers */
1452 VkImageView attachment[1];
1453 VkFramebufferCreateInfo fb_info = {};
1454 fb_info.sType = VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO;
1455 fb_info.renderPass = data->render_pass;
1456 fb_info.attachmentCount = 1;
1457 fb_info.pAttachments = attachment;
1458 fb_info.width = data->width;
1459 fb_info.height = data->height;
1460 fb_info.layers = 1;
1461 for (uint32_t i = 0; i < data->n_images; i++) {
1462 attachment[0] = data->image_views[i];
1463 VK_CHECK(device_data->vtable.CreateFramebuffer(device_data->device, &fb_info,
1464 NULL, &data->framebuffers[i]));
1465 }
1466
1467 /* Command buffer pool */
1468 VkCommandPoolCreateInfo cmd_buffer_pool_info = {};
1469 cmd_buffer_pool_info.sType = VK_STRUCTURE_TYPE_COMMAND_POOL_CREATE_INFO;
1470 cmd_buffer_pool_info.flags = VK_COMMAND_POOL_CREATE_RESET_COMMAND_BUFFER_BIT;
1471 cmd_buffer_pool_info.queueFamilyIndex = device_data->graphic_queue->family_index;
1472 VK_CHECK(device_data->vtable.CreateCommandPool(device_data->device,
1473 &cmd_buffer_pool_info,
1474 NULL, &data->command_pool));
1475 }
1476
1477 static void shutdown_swapchain_data(struct swapchain_data *data)
1478 {
1479 struct device_data *device_data = data->device;
1480
1481 list_for_each_entry_safe(struct overlay_draw, draw, &data->draws, link) {
1482 device_data->vtable.DestroySemaphore(device_data->device, draw->semaphore, NULL);
1483 device_data->vtable.DestroyFence(device_data->device, draw->fence, NULL);
1484 device_data->vtable.DestroyBuffer(device_data->device, draw->vertex_buffer, NULL);
1485 device_data->vtable.DestroyBuffer(device_data->device, draw->index_buffer, NULL);
1486 device_data->vtable.FreeMemory(device_data->device, draw->vertex_buffer_mem, NULL);
1487 device_data->vtable.FreeMemory(device_data->device, draw->index_buffer_mem, NULL);
1488 }
1489
1490 for (uint32_t i = 0; i < data->n_images; i++) {
1491 device_data->vtable.DestroyImageView(device_data->device, data->image_views[i], NULL);
1492 device_data->vtable.DestroyFramebuffer(device_data->device, data->framebuffers[i], NULL);
1493 }
1494
1495 device_data->vtable.DestroyRenderPass(device_data->device, data->render_pass, NULL);
1496
1497 device_data->vtable.DestroyCommandPool(device_data->device, data->command_pool, NULL);
1498
1499 device_data->vtable.DestroyPipeline(device_data->device, data->pipeline, NULL);
1500 device_data->vtable.DestroyPipelineLayout(device_data->device, data->pipeline_layout, NULL);
1501
1502 device_data->vtable.DestroyDescriptorPool(device_data->device,
1503 data->descriptor_pool, NULL);
1504 device_data->vtable.DestroyDescriptorSetLayout(device_data->device,
1505 data->descriptor_layout, NULL);
1506
1507 device_data->vtable.DestroySampler(device_data->device, data->font_sampler, NULL);
1508 device_data->vtable.DestroyImageView(device_data->device, data->font_image_view, NULL);
1509 device_data->vtable.DestroyImage(device_data->device, data->font_image, NULL);
1510 device_data->vtable.FreeMemory(device_data->device, data->font_mem, NULL);
1511
1512 device_data->vtable.DestroyBuffer(device_data->device, data->upload_font_buffer, NULL);
1513 device_data->vtable.FreeMemory(device_data->device, data->upload_font_buffer_mem, NULL);
1514
1515 ImGui::DestroyContext(data->imgui_context);
1516 }
1517
1518 static struct overlay_draw *before_present(struct swapchain_data *swapchain_data,
1519 const VkSemaphore *wait_semaphores,
1520 unsigned n_wait_semaphores,
1521 unsigned imageIndex)
1522 {
1523 struct instance_data *instance_data = swapchain_data->device->instance;
1524 struct overlay_draw *draw = NULL;
1525
1526 snapshot_swapchain_frame(swapchain_data);
1527
1528 if (!instance_data->params.no_display && swapchain_data->n_frames > 0) {
1529 compute_swapchain_display(swapchain_data);
1530 draw = render_swapchain_display(swapchain_data,
1531 wait_semaphores, n_wait_semaphores,
1532 imageIndex);
1533 }
1534
1535 return draw;
1536 }
1537
1538 static VkResult overlay_CreateSwapchainKHR(
1539 VkDevice device,
1540 const VkSwapchainCreateInfoKHR* pCreateInfo,
1541 const VkAllocationCallbacks* pAllocator,
1542 VkSwapchainKHR* pSwapchain)
1543 {
1544 struct device_data *device_data = FIND_DEVICE_DATA(device);
1545 VkResult result = device_data->vtable.CreateSwapchainKHR(device, pCreateInfo, pAllocator, pSwapchain);
1546 if (result != VK_SUCCESS) return result;
1547
1548 struct swapchain_data *swapchain_data = new_swapchain_data(*pSwapchain, device_data);
1549 setup_swapchain_data(swapchain_data, pCreateInfo);
1550 return result;
1551 }
1552
1553 static void overlay_DestroySwapchainKHR(
1554 VkDevice device,
1555 VkSwapchainKHR swapchain,
1556 const VkAllocationCallbacks* pAllocator)
1557 {
1558 struct swapchain_data *swapchain_data = FIND_SWAPCHAIN_DATA(swapchain);
1559
1560 shutdown_swapchain_data(swapchain_data);
1561 swapchain_data->device->vtable.DestroySwapchainKHR(device, swapchain, pAllocator);
1562 destroy_swapchain_data(swapchain_data);
1563 }
1564
1565 static VkResult overlay_QueuePresentKHR(
1566 VkQueue queue,
1567 const VkPresentInfoKHR* pPresentInfo)
1568 {
1569 struct queue_data *queue_data = FIND_QUEUE_DATA(queue);
1570 struct device_data *device_data = queue_data->device;
1571 struct instance_data *instance_data = device_data->instance;
1572 uint32_t query_results[OVERLAY_QUERY_COUNT];
1573
1574 device_data->frame_stats.stats[OVERLAY_PARAM_ENABLED_frame]++;
1575
1576 if (list_length(&queue_data->running_command_buffer) > 0) {
1577 /* Before getting the query results, make sure the operations have
1578 * completed.
1579 */
1580 VK_CHECK(device_data->vtable.ResetFences(device_data->device,
1581 1, &queue_data->queries_fence));
1582 VK_CHECK(device_data->vtable.QueueSubmit(queue, 0, NULL, queue_data->queries_fence));
1583 VK_CHECK(device_data->vtable.WaitForFences(device_data->device,
1584 1, &queue_data->queries_fence,
1585 VK_FALSE, UINT64_MAX));
1586
1587 /* Now get the results. */
1588 list_for_each_entry_safe(struct command_buffer_data, cmd_buffer_data,
1589 &queue_data->running_command_buffer, link) {
1590 list_delinit(&cmd_buffer_data->link);
1591
1592 if (cmd_buffer_data->pipeline_query_pool) {
1593 memset(query_results, 0, sizeof(query_results));
1594 VK_CHECK(device_data->vtable.GetQueryPoolResults(device_data->device,
1595 cmd_buffer_data->pipeline_query_pool,
1596 cmd_buffer_data->query_index, 1,
1597 sizeof(uint32_t) * OVERLAY_QUERY_COUNT,
1598 query_results, 0, VK_QUERY_RESULT_WAIT_BIT));
1599
1600 for (uint32_t i = OVERLAY_PARAM_ENABLED_vertices;
1601 i <= OVERLAY_PARAM_ENABLED_compute_invocations; i++) {
1602 device_data->frame_stats.stats[i] += query_results[i - OVERLAY_PARAM_ENABLED_vertices];
1603 }
1604 }
1605 if (cmd_buffer_data->timestamp_query_pool) {
1606 uint64_t gpu_timestamps[2] = { 0 };
1607 VK_CHECK(device_data->vtable.GetQueryPoolResults(device_data->device,
1608 cmd_buffer_data->timestamp_query_pool,
1609 cmd_buffer_data->query_index * 2, 2,
1610 2 * sizeof(uint64_t), gpu_timestamps, sizeof(uint64_t),
1611 VK_QUERY_RESULT_WAIT_BIT | VK_QUERY_RESULT_64_BIT));
1612
1613 gpu_timestamps[0] &= queue_data->timestamp_mask;
1614 gpu_timestamps[1] &= queue_data->timestamp_mask;
1615 device_data->frame_stats.stats[OVERLAY_PARAM_ENABLED_gpu_timing] +=
1616 (gpu_timestamps[1] - gpu_timestamps[0]) *
1617 device_data->properties.limits.timestampPeriod;
1618 }
1619 }
1620 }
1621
1622 /* Otherwise we need to add our overlay drawing semaphore to the list of
1623 * semaphores to wait on. If we don't do that the presented picture might
1624 * be have incomplete overlay drawings.
1625 */
1626 VkResult result = VK_SUCCESS;
1627 if (instance_data->params.no_display) {
1628 for (uint32_t i = 0; i < pPresentInfo->swapchainCount; i++) {
1629 VkSwapchainKHR swapchain = pPresentInfo->pSwapchains[i];
1630 struct swapchain_data *swapchain_data = FIND_SWAPCHAIN_DATA(swapchain);
1631
1632 before_present(swapchain_data,
1633 pPresentInfo->pWaitSemaphores,
1634 pPresentInfo->waitSemaphoreCount,
1635 pPresentInfo->pImageIndices[i]);
1636 }
1637 result = queue_data->device->vtable.QueuePresentKHR(queue, pPresentInfo);
1638 } else {
1639 for (uint32_t i = 0; i < pPresentInfo->swapchainCount; i++) {
1640 VkSwapchainKHR swapchain = pPresentInfo->pSwapchains[i];
1641 struct swapchain_data *swapchain_data = FIND_SWAPCHAIN_DATA(swapchain);
1642 VkPresentInfoKHR present_info = *pPresentInfo;
1643 present_info.swapchainCount = 1;
1644 present_info.pSwapchains = &swapchain;
1645
1646 uint32_t image_index = pPresentInfo->pImageIndices[i];
1647
1648 struct overlay_draw *draw = before_present(swapchain_data,
1649 pPresentInfo->pWaitSemaphores,
1650 pPresentInfo->waitSemaphoreCount,
1651 image_index);
1652
1653 /* Because the submission of the overlay draw waits on the semaphores
1654 * handed for present, we don't need to have this present operation
1655 * wait on them as well, we can just wait on the overlay submission
1656 * semaphore.
1657 */
1658 present_info.pWaitSemaphores = &draw->semaphore;
1659 present_info.waitSemaphoreCount = 1;
1660
1661 VkResult chain_result = queue_data->device->vtable.QueuePresentKHR(queue, &present_info);
1662 if (pPresentInfo->pResults)
1663 pPresentInfo->pResults[i] = chain_result;
1664 if (chain_result != VK_SUCCESS && result == VK_SUCCESS)
1665 result = chain_result;
1666 }
1667 }
1668 return result;
1669 }
1670
1671 static VkResult overlay_AcquireNextImageKHR(
1672 VkDevice device,
1673 VkSwapchainKHR swapchain,
1674 uint64_t timeout,
1675 VkSemaphore semaphore,
1676 VkFence fence,
1677 uint32_t* pImageIndex)
1678 {
1679 struct swapchain_data *swapchain_data = FIND_SWAPCHAIN_DATA(swapchain);
1680 struct device_data *device_data = swapchain_data->device;
1681
1682 uint64_t ts0 = os_time_get();
1683 VkResult result = device_data->vtable.AcquireNextImageKHR(device, swapchain, timeout,
1684 semaphore, fence, pImageIndex);
1685 uint64_t ts1 = os_time_get();
1686
1687 swapchain_data->frame_stats.stats[OVERLAY_PARAM_ENABLED_acquire_timing] += ts1 - ts0;
1688 swapchain_data->frame_stats.stats[OVERLAY_PARAM_ENABLED_acquire]++;
1689
1690 return result;
1691 }
1692
1693 static VkResult overlay_AcquireNextImage2KHR(
1694 VkDevice device,
1695 const VkAcquireNextImageInfoKHR* pAcquireInfo,
1696 uint32_t* pImageIndex)
1697 {
1698 struct swapchain_data *swapchain_data = FIND_SWAPCHAIN_DATA(pAcquireInfo->swapchain);
1699 struct device_data *device_data = swapchain_data->device;
1700
1701 uint64_t ts0 = os_time_get();
1702 VkResult result = device_data->vtable.AcquireNextImage2KHR(device, pAcquireInfo, pImageIndex);
1703 uint64_t ts1 = os_time_get();
1704
1705 swapchain_data->frame_stats.stats[OVERLAY_PARAM_ENABLED_acquire_timing] += ts1 - ts0;
1706 swapchain_data->frame_stats.stats[OVERLAY_PARAM_ENABLED_acquire]++;
1707
1708 return result;
1709 }
1710
1711 static void overlay_CmdDraw(
1712 VkCommandBuffer commandBuffer,
1713 uint32_t vertexCount,
1714 uint32_t instanceCount,
1715 uint32_t firstVertex,
1716 uint32_t firstInstance)
1717 {
1718 struct command_buffer_data *cmd_buffer_data = FIND_CMD_BUFFER_DATA(commandBuffer);
1719 cmd_buffer_data->stats.stats[OVERLAY_PARAM_ENABLED_draw]++;
1720 struct device_data *device_data = cmd_buffer_data->device;
1721 device_data->vtable.CmdDraw(commandBuffer, vertexCount, instanceCount,
1722 firstVertex, firstInstance);
1723 }
1724
1725 static void overlay_CmdDrawIndexed(
1726 VkCommandBuffer commandBuffer,
1727 uint32_t indexCount,
1728 uint32_t instanceCount,
1729 uint32_t firstIndex,
1730 int32_t vertexOffset,
1731 uint32_t firstInstance)
1732 {
1733 struct command_buffer_data *cmd_buffer_data = FIND_CMD_BUFFER_DATA(commandBuffer);
1734 cmd_buffer_data->stats.stats[OVERLAY_PARAM_ENABLED_draw_indexed]++;
1735 struct device_data *device_data = cmd_buffer_data->device;
1736 device_data->vtable.CmdDrawIndexed(commandBuffer, indexCount, instanceCount,
1737 firstIndex, vertexOffset, firstInstance);
1738 }
1739
1740 static void overlay_CmdDrawIndirect(
1741 VkCommandBuffer commandBuffer,
1742 VkBuffer buffer,
1743 VkDeviceSize offset,
1744 uint32_t drawCount,
1745 uint32_t stride)
1746 {
1747 struct command_buffer_data *cmd_buffer_data = FIND_CMD_BUFFER_DATA(commandBuffer);
1748 cmd_buffer_data->stats.stats[OVERLAY_PARAM_ENABLED_draw_indirect]++;
1749 struct device_data *device_data = cmd_buffer_data->device;
1750 device_data->vtable.CmdDrawIndirect(commandBuffer, buffer, offset, drawCount, stride);
1751 }
1752
1753 static void overlay_CmdDrawIndexedIndirect(
1754 VkCommandBuffer commandBuffer,
1755 VkBuffer buffer,
1756 VkDeviceSize offset,
1757 uint32_t drawCount,
1758 uint32_t stride)
1759 {
1760 struct command_buffer_data *cmd_buffer_data = FIND_CMD_BUFFER_DATA(commandBuffer);
1761 cmd_buffer_data->stats.stats[OVERLAY_PARAM_ENABLED_draw_indexed_indirect]++;
1762 struct device_data *device_data = cmd_buffer_data->device;
1763 device_data->vtable.CmdDrawIndexedIndirect(commandBuffer, buffer, offset, drawCount, stride);
1764 }
1765
1766 static void overlay_CmdDrawIndirectCountKHR(
1767 VkCommandBuffer commandBuffer,
1768 VkBuffer buffer,
1769 VkDeviceSize offset,
1770 VkBuffer countBuffer,
1771 VkDeviceSize countBufferOffset,
1772 uint32_t maxDrawCount,
1773 uint32_t stride)
1774 {
1775 struct command_buffer_data *cmd_buffer_data = FIND_CMD_BUFFER_DATA(commandBuffer);
1776 cmd_buffer_data->stats.stats[OVERLAY_PARAM_ENABLED_draw_indirect_count]++;
1777 struct device_data *device_data = cmd_buffer_data->device;
1778 device_data->vtable.CmdDrawIndirectCountKHR(commandBuffer, buffer, offset,
1779 countBuffer, countBufferOffset,
1780 maxDrawCount, stride);
1781 }
1782
1783 static void overlay_CmdDrawIndexedIndirectCountKHR(
1784 VkCommandBuffer commandBuffer,
1785 VkBuffer buffer,
1786 VkDeviceSize offset,
1787 VkBuffer countBuffer,
1788 VkDeviceSize countBufferOffset,
1789 uint32_t maxDrawCount,
1790 uint32_t stride)
1791 {
1792 struct command_buffer_data *cmd_buffer_data = FIND_CMD_BUFFER_DATA(commandBuffer);
1793 cmd_buffer_data->stats.stats[OVERLAY_PARAM_ENABLED_draw_indexed_indirect_count]++;
1794 struct device_data *device_data = cmd_buffer_data->device;
1795 device_data->vtable.CmdDrawIndexedIndirectCountKHR(commandBuffer, buffer, offset,
1796 countBuffer, countBufferOffset,
1797 maxDrawCount, stride);
1798 }
1799
1800 static void overlay_CmdDispatch(
1801 VkCommandBuffer commandBuffer,
1802 uint32_t groupCountX,
1803 uint32_t groupCountY,
1804 uint32_t groupCountZ)
1805 {
1806 struct command_buffer_data *cmd_buffer_data = FIND_CMD_BUFFER_DATA(commandBuffer);
1807 cmd_buffer_data->stats.stats[OVERLAY_PARAM_ENABLED_dispatch]++;
1808 struct device_data *device_data = cmd_buffer_data->device;
1809 device_data->vtable.CmdDispatch(commandBuffer, groupCountX, groupCountY, groupCountZ);
1810 }
1811
1812 static void overlay_CmdDispatchIndirect(
1813 VkCommandBuffer commandBuffer,
1814 VkBuffer buffer,
1815 VkDeviceSize offset)
1816 {
1817 struct command_buffer_data *cmd_buffer_data = FIND_CMD_BUFFER_DATA(commandBuffer);
1818 cmd_buffer_data->stats.stats[OVERLAY_PARAM_ENABLED_dispatch_indirect]++;
1819 struct device_data *device_data = cmd_buffer_data->device;
1820 device_data->vtable.CmdDispatchIndirect(commandBuffer, buffer, offset);
1821 }
1822
1823 static void overlay_CmdBindPipeline(
1824 VkCommandBuffer commandBuffer,
1825 VkPipelineBindPoint pipelineBindPoint,
1826 VkPipeline pipeline)
1827 {
1828 struct command_buffer_data *cmd_buffer_data = FIND_CMD_BUFFER_DATA(commandBuffer);
1829 switch (pipelineBindPoint) {
1830 case VK_PIPELINE_BIND_POINT_GRAPHICS: cmd_buffer_data->stats.stats[OVERLAY_PARAM_ENABLED_pipeline_graphics]++; break;
1831 case VK_PIPELINE_BIND_POINT_COMPUTE: cmd_buffer_data->stats.stats[OVERLAY_PARAM_ENABLED_pipeline_compute]++; break;
1832 case VK_PIPELINE_BIND_POINT_RAY_TRACING_NV: cmd_buffer_data->stats.stats[OVERLAY_PARAM_ENABLED_pipeline_raytracing]++; break;
1833 default: break;
1834 }
1835 struct device_data *device_data = cmd_buffer_data->device;
1836 device_data->vtable.CmdBindPipeline(commandBuffer, pipelineBindPoint, pipeline);
1837 }
1838
1839 static VkResult overlay_BeginCommandBuffer(
1840 VkCommandBuffer commandBuffer,
1841 const VkCommandBufferBeginInfo* pBeginInfo)
1842 {
1843 struct command_buffer_data *cmd_buffer_data = FIND_CMD_BUFFER_DATA(commandBuffer);
1844 struct device_data *device_data = cmd_buffer_data->device;
1845
1846 memset(&cmd_buffer_data->stats, 0, sizeof(cmd_buffer_data->stats));
1847
1848 /* We don't record any query in secondary command buffers, just make sure
1849 * we have the right inheritance.
1850 */
1851 if (cmd_buffer_data->level == VK_COMMAND_BUFFER_LEVEL_SECONDARY) {
1852 VkCommandBufferBeginInfo *begin_info = (VkCommandBufferBeginInfo *)
1853 clone_chain((const struct VkBaseInStructure *)pBeginInfo);
1854 VkCommandBufferInheritanceInfo *parent_inhe_info = (VkCommandBufferInheritanceInfo *)
1855 vk_find_struct(begin_info, COMMAND_BUFFER_INHERITANCE_INFO);
1856 VkCommandBufferInheritanceInfo inhe_info = {
1857 VK_STRUCTURE_TYPE_COMMAND_BUFFER_INHERITANCE_INFO,
1858 NULL,
1859 VK_NULL_HANDLE,
1860 0,
1861 VK_NULL_HANDLE,
1862 VK_FALSE,
1863 0,
1864 overlay_query_flags,
1865 };
1866
1867 if (parent_inhe_info)
1868 parent_inhe_info->pipelineStatistics = overlay_query_flags;
1869 else {
1870 inhe_info.pNext = begin_info->pNext;
1871 begin_info->pNext = &inhe_info;
1872 }
1873
1874 VkResult result = device_data->vtable.BeginCommandBuffer(commandBuffer, pBeginInfo);
1875
1876 if (!parent_inhe_info)
1877 begin_info->pNext = inhe_info.pNext;
1878
1879 free_chain((struct VkBaseOutStructure *)begin_info);
1880
1881 return result;
1882 }
1883
1884 /* Otherwise record a begin query as first command. */
1885 VkResult result = device_data->vtable.BeginCommandBuffer(commandBuffer, pBeginInfo);
1886
1887 if (result == VK_SUCCESS) {
1888 if (cmd_buffer_data->pipeline_query_pool) {
1889 device_data->vtable.CmdResetQueryPool(commandBuffer,
1890 cmd_buffer_data->pipeline_query_pool,
1891 cmd_buffer_data->query_index, 1);
1892 }
1893 if (cmd_buffer_data->timestamp_query_pool) {
1894 device_data->vtable.CmdResetQueryPool(commandBuffer,
1895 cmd_buffer_data->timestamp_query_pool,
1896 cmd_buffer_data->query_index * 2, 2);
1897 }
1898 if (cmd_buffer_data->pipeline_query_pool) {
1899 device_data->vtable.CmdBeginQuery(commandBuffer,
1900 cmd_buffer_data->pipeline_query_pool,
1901 cmd_buffer_data->query_index, 0);
1902 }
1903 if (cmd_buffer_data->timestamp_query_pool) {
1904 device_data->vtable.CmdWriteTimestamp(commandBuffer,
1905 VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT,
1906 cmd_buffer_data->timestamp_query_pool,
1907 cmd_buffer_data->query_index * 2);
1908 }
1909 }
1910
1911 return result;
1912 }
1913
1914 static VkResult overlay_EndCommandBuffer(
1915 VkCommandBuffer commandBuffer)
1916 {
1917 struct command_buffer_data *cmd_buffer_data = FIND_CMD_BUFFER_DATA(commandBuffer);
1918 struct device_data *device_data = cmd_buffer_data->device;
1919
1920 if (cmd_buffer_data->timestamp_query_pool) {
1921 device_data->vtable.CmdWriteTimestamp(commandBuffer,
1922 VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT,
1923 cmd_buffer_data->timestamp_query_pool,
1924 cmd_buffer_data->query_index * 2 + 1);
1925 }
1926 if (cmd_buffer_data->pipeline_query_pool) {
1927 device_data->vtable.CmdEndQuery(commandBuffer,
1928 cmd_buffer_data->pipeline_query_pool,
1929 cmd_buffer_data->query_index);
1930 }
1931
1932 return device_data->vtable.EndCommandBuffer(commandBuffer);
1933 }
1934
1935 static VkResult overlay_ResetCommandBuffer(
1936 VkCommandBuffer commandBuffer,
1937 VkCommandBufferResetFlags flags)
1938 {
1939 struct command_buffer_data *cmd_buffer_data = FIND_CMD_BUFFER_DATA(commandBuffer);
1940 struct device_data *device_data = cmd_buffer_data->device;
1941
1942 memset(&cmd_buffer_data->stats, 0, sizeof(cmd_buffer_data->stats));
1943
1944 return device_data->vtable.ResetCommandBuffer(commandBuffer, flags);
1945 }
1946
1947 static void overlay_CmdExecuteCommands(
1948 VkCommandBuffer commandBuffer,
1949 uint32_t commandBufferCount,
1950 const VkCommandBuffer* pCommandBuffers)
1951 {
1952 struct command_buffer_data *cmd_buffer_data = FIND_CMD_BUFFER_DATA(commandBuffer);
1953 struct device_data *device_data = cmd_buffer_data->device;
1954
1955 /* Add the stats of the executed command buffers to the primary one. */
1956 for (uint32_t c = 0; c < commandBufferCount; c++) {
1957 struct command_buffer_data *sec_cmd_buffer_data = FIND_CMD_BUFFER_DATA(pCommandBuffers[c]);
1958
1959 for (uint32_t s = 0; s < OVERLAY_PARAM_ENABLED_MAX; s++)
1960 cmd_buffer_data->stats.stats[s] += sec_cmd_buffer_data->stats.stats[s];
1961 }
1962
1963 device_data->vtable.CmdExecuteCommands(commandBuffer, commandBufferCount, pCommandBuffers);
1964 }
1965
1966 static VkResult overlay_AllocateCommandBuffers(
1967 VkDevice device,
1968 const VkCommandBufferAllocateInfo* pAllocateInfo,
1969 VkCommandBuffer* pCommandBuffers)
1970 {
1971 struct device_data *device_data = FIND_DEVICE_DATA(device);
1972 VkResult result =
1973 device_data->vtable.AllocateCommandBuffers(device, pAllocateInfo, pCommandBuffers);
1974 if (result != VK_SUCCESS)
1975 return result;
1976
1977 VkQueryPool pipeline_query_pool = VK_NULL_HANDLE;
1978 VkQueryPool timestamp_query_pool = VK_NULL_HANDLE;
1979 if (device_data->instance->pipeline_statistics_enabled &&
1980 pAllocateInfo->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY) {
1981 VkQueryPoolCreateInfo pool_info = {
1982 VK_STRUCTURE_TYPE_QUERY_POOL_CREATE_INFO,
1983 NULL,
1984 0,
1985 VK_QUERY_TYPE_PIPELINE_STATISTICS,
1986 pAllocateInfo->commandBufferCount,
1987 overlay_query_flags,
1988 };
1989 VK_CHECK(device_data->vtable.CreateQueryPool(device_data->device, &pool_info,
1990 NULL, &pipeline_query_pool));
1991 }
1992 if (device_data->instance->params.enabled[OVERLAY_PARAM_ENABLED_gpu_timing]) {
1993 VkQueryPoolCreateInfo pool_info = {
1994 VK_STRUCTURE_TYPE_QUERY_POOL_CREATE_INFO,
1995 NULL,
1996 0,
1997 VK_QUERY_TYPE_TIMESTAMP,
1998 pAllocateInfo->commandBufferCount * 2,
1999 0,
2000 };
2001 VK_CHECK(device_data->vtable.CreateQueryPool(device_data->device, &pool_info,
2002 NULL, &timestamp_query_pool));
2003 }
2004
2005 for (uint32_t i = 0; i < pAllocateInfo->commandBufferCount; i++) {
2006 new_command_buffer_data(pCommandBuffers[i], pAllocateInfo->level,
2007 pipeline_query_pool, timestamp_query_pool,
2008 i, device_data);
2009 }
2010
2011 if (pipeline_query_pool)
2012 map_object(HKEY(pipeline_query_pool), (void *)(uintptr_t) pAllocateInfo->commandBufferCount);
2013 if (timestamp_query_pool)
2014 map_object(HKEY(timestamp_query_pool), (void *)(uintptr_t) pAllocateInfo->commandBufferCount);
2015
2016 return result;
2017 }
2018
2019 static void overlay_FreeCommandBuffers(
2020 VkDevice device,
2021 VkCommandPool commandPool,
2022 uint32_t commandBufferCount,
2023 const VkCommandBuffer* pCommandBuffers)
2024 {
2025 struct device_data *device_data = FIND_DEVICE_DATA(device);
2026 for (uint32_t i = 0; i < commandBufferCount; i++) {
2027 struct command_buffer_data *cmd_buffer_data =
2028 FIND_CMD_BUFFER_DATA(pCommandBuffers[i]);
2029 /* It is legal to free a NULL command buffer*/
2030 if (!cmd_buffer_data)
2031 continue;
2032
2033 uint64_t count = (uintptr_t)find_object_data(HKEY(cmd_buffer_data->pipeline_query_pool));
2034 if (count == 1) {
2035 unmap_object(HKEY(cmd_buffer_data->pipeline_query_pool));
2036 device_data->vtable.DestroyQueryPool(device_data->device,
2037 cmd_buffer_data->pipeline_query_pool, NULL);
2038 } else if (count != 0) {
2039 map_object(HKEY(cmd_buffer_data->pipeline_query_pool), (void *)(uintptr_t)(count - 1));
2040 }
2041 count = (uintptr_t)find_object_data(HKEY(cmd_buffer_data->timestamp_query_pool));
2042 if (count == 1) {
2043 unmap_object(HKEY(cmd_buffer_data->timestamp_query_pool));
2044 device_data->vtable.DestroyQueryPool(device_data->device,
2045 cmd_buffer_data->timestamp_query_pool, NULL);
2046 } else if (count != 0) {
2047 map_object(HKEY(cmd_buffer_data->timestamp_query_pool), (void *)(uintptr_t)(count - 1));
2048 }
2049 destroy_command_buffer_data(cmd_buffer_data);
2050 }
2051
2052 device_data->vtable.FreeCommandBuffers(device, commandPool,
2053 commandBufferCount, pCommandBuffers);
2054 }
2055
2056 static VkResult overlay_QueueSubmit(
2057 VkQueue queue,
2058 uint32_t submitCount,
2059 const VkSubmitInfo* pSubmits,
2060 VkFence fence)
2061 {
2062 struct queue_data *queue_data = FIND_QUEUE_DATA(queue);
2063 struct device_data *device_data = queue_data->device;
2064
2065 device_data->frame_stats.stats[OVERLAY_PARAM_ENABLED_submit]++;
2066
2067 for (uint32_t s = 0; s < submitCount; s++) {
2068 for (uint32_t c = 0; c < pSubmits[s].commandBufferCount; c++) {
2069 struct command_buffer_data *cmd_buffer_data =
2070 FIND_CMD_BUFFER_DATA(pSubmits[s].pCommandBuffers[c]);
2071
2072 /* Merge the submitted command buffer stats into the device. */
2073 for (uint32_t st = 0; st < OVERLAY_PARAM_ENABLED_MAX; st++)
2074 device_data->frame_stats.stats[st] += cmd_buffer_data->stats.stats[st];
2075
2076 /* Attach the command buffer to the queue so we remember to read its
2077 * pipeline statistics & timestamps at QueuePresent().
2078 */
2079 if (!cmd_buffer_data->pipeline_query_pool &&
2080 !cmd_buffer_data->timestamp_query_pool)
2081 continue;
2082
2083 if (list_empty(&cmd_buffer_data->link)) {
2084 list_addtail(&cmd_buffer_data->link,
2085 &queue_data->running_command_buffer);
2086 } else {
2087 fprintf(stderr, "Command buffer submitted multiple times before present.\n"
2088 "This could lead to invalid data.\n");
2089 }
2090 }
2091 }
2092
2093 return device_data->vtable.QueueSubmit(queue, submitCount, pSubmits, fence);
2094 }
2095
2096 static VkResult overlay_CreateDevice(
2097 VkPhysicalDevice physicalDevice,
2098 const VkDeviceCreateInfo* pCreateInfo,
2099 const VkAllocationCallbacks* pAllocator,
2100 VkDevice* pDevice)
2101 {
2102 struct instance_data *instance_data = FIND_PHYSICAL_DEVICE_DATA(physicalDevice);
2103 VkLayerDeviceCreateInfo *chain_info =
2104 get_device_chain_info(pCreateInfo, VK_LAYER_LINK_INFO);
2105
2106 assert(chain_info->u.pLayerInfo);
2107 PFN_vkGetInstanceProcAddr fpGetInstanceProcAddr = chain_info->u.pLayerInfo->pfnNextGetInstanceProcAddr;
2108 PFN_vkGetDeviceProcAddr fpGetDeviceProcAddr = chain_info->u.pLayerInfo->pfnNextGetDeviceProcAddr;
2109 PFN_vkCreateDevice fpCreateDevice = (PFN_vkCreateDevice)fpGetInstanceProcAddr(NULL, "vkCreateDevice");
2110 if (fpCreateDevice == NULL) {
2111 return VK_ERROR_INITIALIZATION_FAILED;
2112 }
2113
2114 // Advance the link info for the next element on the chain
2115 chain_info->u.pLayerInfo = chain_info->u.pLayerInfo->pNext;
2116
2117 VkPhysicalDeviceFeatures device_features = {};
2118 VkDeviceCreateInfo device_info = *pCreateInfo;
2119
2120 if (pCreateInfo->pEnabledFeatures)
2121 device_features = *(pCreateInfo->pEnabledFeatures);
2122 if (instance_data->pipeline_statistics_enabled) {
2123 device_features.inheritedQueries = true;
2124 device_features.pipelineStatisticsQuery = true;
2125 }
2126 device_info.pEnabledFeatures = &device_features;
2127
2128
2129 VkResult result = fpCreateDevice(physicalDevice, &device_info, pAllocator, pDevice);
2130 if (result != VK_SUCCESS) return result;
2131
2132 struct device_data *device_data = new_device_data(*pDevice, instance_data);
2133 device_data->physical_device = physicalDevice;
2134 vk_load_device_commands(*pDevice, fpGetDeviceProcAddr, &device_data->vtable);
2135
2136 instance_data->vtable.GetPhysicalDeviceProperties(device_data->physical_device,
2137 &device_data->properties);
2138
2139 VkLayerDeviceCreateInfo *load_data_info =
2140 get_device_chain_info(pCreateInfo, VK_LOADER_DATA_CALLBACK);
2141 device_data->set_device_loader_data = load_data_info->u.pfnSetDeviceLoaderData;
2142
2143 device_map_queues(device_data, pCreateInfo);
2144
2145 return result;
2146 }
2147
2148 static void overlay_DestroyDevice(
2149 VkDevice device,
2150 const VkAllocationCallbacks* pAllocator)
2151 {
2152 struct device_data *device_data = FIND_DEVICE_DATA(device);
2153 device_unmap_queues(device_data);
2154 device_data->vtable.DestroyDevice(device, pAllocator);
2155 destroy_device_data(device_data);
2156 }
2157
2158 static VkResult overlay_CreateInstance(
2159 const VkInstanceCreateInfo* pCreateInfo,
2160 const VkAllocationCallbacks* pAllocator,
2161 VkInstance* pInstance)
2162 {
2163 VkLayerInstanceCreateInfo *chain_info =
2164 get_instance_chain_info(pCreateInfo, VK_LAYER_LINK_INFO);
2165
2166 assert(chain_info->u.pLayerInfo);
2167 PFN_vkGetInstanceProcAddr fpGetInstanceProcAddr =
2168 chain_info->u.pLayerInfo->pfnNextGetInstanceProcAddr;
2169 PFN_vkCreateInstance fpCreateInstance =
2170 (PFN_vkCreateInstance)fpGetInstanceProcAddr(NULL, "vkCreateInstance");
2171 if (fpCreateInstance == NULL) {
2172 return VK_ERROR_INITIALIZATION_FAILED;
2173 }
2174
2175 // Advance the link info for the next element on the chain
2176 chain_info->u.pLayerInfo = chain_info->u.pLayerInfo->pNext;
2177
2178 VkResult result = fpCreateInstance(pCreateInfo, pAllocator, pInstance);
2179 if (result != VK_SUCCESS) return result;
2180
2181 struct instance_data *instance_data = new_instance_data(*pInstance);
2182 vk_load_instance_commands(instance_data->instance,
2183 fpGetInstanceProcAddr,
2184 &instance_data->vtable);
2185 instance_data_map_physical_devices(instance_data, true);
2186
2187 parse_overlay_env(&instance_data->params, getenv("VK_LAYER_MESA_OVERLAY_CONFIG"));
2188
2189 for (int i = OVERLAY_PARAM_ENABLED_vertices;
2190 i <= OVERLAY_PARAM_ENABLED_compute_invocations; i++) {
2191 if (instance_data->params.enabled[i]) {
2192 instance_data->pipeline_statistics_enabled = true;
2193 break;
2194 }
2195 }
2196
2197 return result;
2198 }
2199
2200 static void overlay_DestroyInstance(
2201 VkInstance instance,
2202 const VkAllocationCallbacks* pAllocator)
2203 {
2204 struct instance_data *instance_data = FIND_INSTANCE_DATA(instance);
2205 instance_data_map_physical_devices(instance_data, false);
2206 instance_data->vtable.DestroyInstance(instance, pAllocator);
2207 destroy_instance_data(instance_data);
2208 }
2209
2210 static const struct {
2211 const char *name;
2212 void *ptr;
2213 } name_to_funcptr_map[] = {
2214 { "vkGetDeviceProcAddr", (void *) vkGetDeviceProcAddr },
2215 #define ADD_HOOK(fn) { "vk" # fn, (void *) overlay_ ## fn }
2216 ADD_HOOK(AllocateCommandBuffers),
2217 ADD_HOOK(FreeCommandBuffers),
2218 ADD_HOOK(ResetCommandBuffer),
2219 ADD_HOOK(BeginCommandBuffer),
2220 ADD_HOOK(EndCommandBuffer),
2221 ADD_HOOK(CmdExecuteCommands),
2222
2223 ADD_HOOK(CmdDraw),
2224 ADD_HOOK(CmdDrawIndexed),
2225 ADD_HOOK(CmdDrawIndirect),
2226 ADD_HOOK(CmdDrawIndexedIndirect),
2227 ADD_HOOK(CmdDispatch),
2228 ADD_HOOK(CmdDispatchIndirect),
2229 ADD_HOOK(CmdDrawIndirectCountKHR),
2230 ADD_HOOK(CmdDrawIndexedIndirectCountKHR),
2231
2232 ADD_HOOK(CmdBindPipeline),
2233
2234 ADD_HOOK(CreateSwapchainKHR),
2235 ADD_HOOK(QueuePresentKHR),
2236 ADD_HOOK(DestroySwapchainKHR),
2237 ADD_HOOK(AcquireNextImageKHR),
2238 ADD_HOOK(AcquireNextImage2KHR),
2239
2240 ADD_HOOK(QueueSubmit),
2241
2242 ADD_HOOK(CreateDevice),
2243 ADD_HOOK(DestroyDevice),
2244
2245 ADD_HOOK(CreateInstance),
2246 ADD_HOOK(DestroyInstance),
2247 #undef ADD_HOOK
2248 };
2249
2250 static void *find_ptr(const char *name)
2251 {
2252 for (uint32_t i = 0; i < ARRAY_SIZE(name_to_funcptr_map); i++) {
2253 if (strcmp(name, name_to_funcptr_map[i].name) == 0)
2254 return name_to_funcptr_map[i].ptr;
2255 }
2256
2257 return NULL;
2258 }
2259
2260 VK_LAYER_EXPORT VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vkGetDeviceProcAddr(VkDevice dev,
2261 const char *funcName)
2262 {
2263 void *ptr = find_ptr(funcName);
2264 if (ptr) return reinterpret_cast<PFN_vkVoidFunction>(ptr);
2265
2266 if (dev == NULL) return NULL;
2267
2268 struct device_data *device_data = FIND_DEVICE_DATA(dev);
2269 if (device_data->vtable.GetDeviceProcAddr == NULL) return NULL;
2270 return device_data->vtable.GetDeviceProcAddr(dev, funcName);
2271 }
2272
2273 VK_LAYER_EXPORT VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vkGetInstanceProcAddr(VkInstance instance,
2274 const char *funcName)
2275 {
2276 void *ptr = find_ptr(funcName);
2277 if (ptr) return reinterpret_cast<PFN_vkVoidFunction>(ptr);
2278
2279 if (instance == NULL) return NULL;
2280
2281 struct instance_data *instance_data = FIND_INSTANCE_DATA(instance);
2282 if (instance_data->vtable.GetInstanceProcAddr == NULL) return NULL;
2283 return instance_data->vtable.GetInstanceProcAddr(instance, funcName);
2284 }