projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
vulkan/overlay: add queue present timing measurement
[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 case OVERLAY_PARAM_ENABLED_present_timing:
549 return "(us)";
550 case OVERLAY_PARAM_ENABLED_gpu_timing:
551 return "(ns)";
552 default:
553 return "";
554 }
555 }
556
557 static void snapshot_swapchain_frame(struct swapchain_data *data)
558 {
559 struct device_data *device_data = data->device;
560 struct instance_data *instance_data = device_data->instance;
561 uint32_t f_idx = data->n_frames % ARRAY_SIZE(data->frames_stats);
562 uint64_t now = os_time_get(); /* us */
563
564 if (data->last_present_time) {
565 data->frame_stats.stats[OVERLAY_PARAM_ENABLED_frame_timing] =
566 now - data->last_present_time;
567 }
568
569 memset(&data->frames_stats[f_idx], 0, sizeof(data->frames_stats[f_idx]));
570 for (int s = 0; s < OVERLAY_PARAM_ENABLED_MAX; s++) {
571 data->frames_stats[f_idx].stats[s] += device_data->frame_stats.stats[s] + data->frame_stats.stats[s];
572 data->accumulated_stats.stats[s] += device_data->frame_stats.stats[s] + data->frame_stats.stats[s];
573 }
574
575 if (data->last_fps_update) {
576 double elapsed = (double)(now - data->last_fps_update); /* us */
577 if (elapsed >= instance_data->params.fps_sampling_period) {
578 data->fps = 1000000.0f * data->n_frames_since_update / elapsed;
579 if (instance_data->params.output_file) {
580 if (!instance_data->first_line_printed) {
581 bool first_column = true;
582
583 instance_data->first_line_printed = true;
584
585 #define OVERLAY_PARAM_BOOL(name) \
586 if (instance_data->params.enabled[OVERLAY_PARAM_ENABLED_##name]) { \
587 fprintf(instance_data->params.output_file, \
588 "%s%s%s", first_column ? "" : ", ", #name, \
589 param_unit(OVERLAY_PARAM_ENABLED_##name)); \
590 first_column = false; \
591 }
592 #define OVERLAY_PARAM_CUSTOM(name)
593 OVERLAY_PARAMS
594 #undef OVERLAY_PARAM_BOOL
595 #undef OVERLAY_PARAM_CUSTOM
596 fprintf(instance_data->params.output_file, "\n");
597 }
598
599 for (int s = 0; s < OVERLAY_PARAM_ENABLED_MAX; s++) {
600 if (!instance_data->params.enabled[s])
601 continue;
602 if (s == OVERLAY_PARAM_ENABLED_fps) {
603 fprintf(instance_data->params.output_file,
604 "%s%.2f", s == 0 ? "" : ", ", data->fps);
605 } else {
606 fprintf(instance_data->params.output_file,
607 "%s%" PRIu64, s == 0 ? "" : ", ",
608 data->accumulated_stats.stats[s]);
609 }
610 }
611 fprintf(instance_data->params.output_file, "\n");
612 fflush(instance_data->params.output_file);
613 }
614
615 memset(&data->accumulated_stats, 0, sizeof(data->accumulated_stats));
616 data->n_frames_since_update = 0;
617 data->last_fps_update = now;
618 }
619 } else {
620 data->last_fps_update = now;
621 }
622
623 memset(&device_data->frame_stats, 0, sizeof(device_data->frame_stats));
624 memset(&data->frame_stats, 0, sizeof(device_data->frame_stats));
625
626 data->last_present_time = now;
627 data->n_frames++;
628 data->n_frames_since_update++;
629 }
630
631 static float get_time_stat(void *_data, int _idx)
632 {
633 struct swapchain_data *data = (struct swapchain_data *) _data;
634 if ((ARRAY_SIZE(data->frames_stats) - _idx) > data->n_frames)
635 return 0.0f;
636 int idx = ARRAY_SIZE(data->frames_stats) +
637 data->n_frames < ARRAY_SIZE(data->frames_stats) ?
638 _idx - data->n_frames :
639 _idx + data->n_frames;
640 idx %= ARRAY_SIZE(data->frames_stats);
641 /* Time stats are in us. */
642 return data->frames_stats[idx].stats[data->stat_selector] / data->time_dividor;
643 }
644
645 static float get_stat(void *_data, int _idx)
646 {
647 struct swapchain_data *data = (struct swapchain_data *) _data;
648 if ((ARRAY_SIZE(data->frames_stats) - _idx) > data->n_frames)
649 return 0.0f;
650 int idx = ARRAY_SIZE(data->frames_stats) +
651 data->n_frames < ARRAY_SIZE(data->frames_stats) ?
652 _idx - data->n_frames :
653 _idx + data->n_frames;
654 idx %= ARRAY_SIZE(data->frames_stats);
655 return data->frames_stats[idx].stats[data->stat_selector];
656 }
657
658 static void position_layer(struct swapchain_data *data)
659
660 {
661 struct device_data *device_data = data->device;
662 struct instance_data *instance_data = device_data->instance;
663 const float margin = 10.0f;
664
665 ImGui::SetNextWindowBgAlpha(0.5);
666 ImGui::SetNextWindowSize(data->window_size, ImGuiCond_Always);
667 switch (instance_data->params.position) {
668 case LAYER_POSITION_TOP_LEFT:
669 ImGui::SetNextWindowPos(ImVec2(margin, margin), ImGuiCond_Always);
670 break;
671 case LAYER_POSITION_TOP_RIGHT:
672 ImGui::SetNextWindowPos(ImVec2(data->width - data->window_size.x - margin, margin),
673 ImGuiCond_Always);
674 break;
675 case LAYER_POSITION_BOTTOM_LEFT:
676 ImGui::SetNextWindowPos(ImVec2(margin, data->height - data->window_size.y - margin),
677 ImGuiCond_Always);
678 break;
679 case LAYER_POSITION_BOTTOM_RIGHT:
680 ImGui::SetNextWindowPos(ImVec2(data->width - data->window_size.x - margin,
681 data->height - data->window_size.y - margin),
682 ImGuiCond_Always);
683 break;
684 }
685 }
686
687 static void compute_swapchain_display(struct swapchain_data *data)
688 {
689 struct device_data *device_data = data->device;
690 struct instance_data *instance_data = device_data->instance;
691
692 ImGui::SetCurrentContext(data->imgui_context);
693 ImGui::NewFrame();
694 position_layer(data);
695 ImGui::Begin("Mesa overlay");
696 ImGui::Text("Device: %s", device_data->properties.deviceName);
697
698 const char *format_name = vk_Format_to_str(data->format);
699 format_name = format_name ? (format_name + strlen("VK_FORMAT_")) : "unknown";
700 ImGui::Text("Swapchain format: %s", format_name);
701 ImGui::Text("Frames: %" PRIu64, data->n_frames);
702 if (instance_data->params.enabled[OVERLAY_PARAM_ENABLED_fps])
703 ImGui::Text("FPS: %.2f" , data->fps);
704
705 /* Recompute min/max */
706 for (uint32_t s = 0; s < OVERLAY_PARAM_ENABLED_MAX; s++) {
707 data->stats_min.stats[s] = UINT64_MAX;
708 data->stats_max.stats[s] = 0;
709 }
710 for (uint32_t f = 0; f < MIN2(data->n_frames, ARRAY_SIZE(data->frames_stats)); f++) {
711 for (uint32_t s = 0; s < OVERLAY_PARAM_ENABLED_MAX; s++) {
712 data->stats_min.stats[s] = MIN2(data->frames_stats[f].stats[s],
713 data->stats_min.stats[s]);
714 data->stats_max.stats[s] = MAX2(data->frames_stats[f].stats[s],
715 data->stats_max.stats[s]);
716 }
717 }
718 for (uint32_t s = 0; s < OVERLAY_PARAM_ENABLED_MAX; s++) {
719 assert(data->stats_min.stats[s] != UINT64_MAX);
720 }
721
722 for (uint32_t s = 0; s < OVERLAY_PARAM_ENABLED_MAX; s++) {
723 if (!instance_data->params.enabled[s] ||
724 s == OVERLAY_PARAM_ENABLED_fps ||
725 s == OVERLAY_PARAM_ENABLED_frame)
726 continue;
727
728 char hash[40];
729 snprintf(hash, sizeof(hash), "##%s", overlay_param_names[s]);
730 data->stat_selector = (enum overlay_param_enabled) s;
731 data->time_dividor = 1000.0f;
732 if (s == OVERLAY_PARAM_ENABLED_gpu_timing)
733 data->time_dividor = 1000000.0f;
734
735 if (s == OVERLAY_PARAM_ENABLED_frame_timing ||
736 s == OVERLAY_PARAM_ENABLED_acquire_timing ||
737 s == OVERLAY_PARAM_ENABLED_present_timing ||
738 s == OVERLAY_PARAM_ENABLED_gpu_timing) {
739 double min_time = data->stats_min.stats[s] / data->time_dividor;
740 double max_time = data->stats_max.stats[s] / data->time_dividor;
741 ImGui::PlotHistogram(hash, get_time_stat, data,
742 ARRAY_SIZE(data->frames_stats), 0,
743 NULL, min_time, max_time,
744 ImVec2(ImGui::GetContentRegionAvailWidth(), 30));
745 ImGui::Text("%s: %.3fms [%.3f, %.3f]", overlay_param_names[s],
746 get_time_stat(data, ARRAY_SIZE(data->frames_stats) - 1),
747 min_time, max_time);
748 } else {
749 ImGui::PlotHistogram(hash, get_stat, data,
750 ARRAY_SIZE(data->frames_stats), 0,
751 NULL,
752 data->stats_min.stats[s],
753 data->stats_max.stats[s],
754 ImVec2(ImGui::GetContentRegionAvailWidth(), 30));
755 ImGui::Text("%s: %.0f [%" PRIu64 ", %" PRIu64 "]", overlay_param_names[s],
756 get_stat(data, ARRAY_SIZE(data->frames_stats) - 1),
757 data->stats_min.stats[s], data->stats_max.stats[s]);
758 }
759 }
760 data->window_size = ImVec2(data->window_size.x, ImGui::GetCursorPosY() + 10.0f);
761 ImGui::End();
762 ImGui::EndFrame();
763 ImGui::Render();
764 }
765
766 static uint32_t vk_memory_type(struct device_data *data,
767 VkMemoryPropertyFlags properties,
768 uint32_t type_bits)
769 {
770 VkPhysicalDeviceMemoryProperties prop;
771 data->instance->vtable.GetPhysicalDeviceMemoryProperties(data->physical_device, &prop);
772 for (uint32_t i = 0; i < prop.memoryTypeCount; i++)
773 if ((prop.memoryTypes[i].propertyFlags & properties) == properties && type_bits & (1<<i))
774 return i;
775 return 0xFFFFFFFF; // Unable to find memoryType
776 }
777
778 static void ensure_swapchain_fonts(struct swapchain_data *data,
779 VkCommandBuffer command_buffer)
780 {
781 if (data->font_uploaded)
782 return;
783
784 data->font_uploaded = true;
785
786 struct device_data *device_data = data->device;
787 ImGuiIO& io = ImGui::GetIO();
788 unsigned char* pixels;
789 int width, height;
790 io.Fonts->GetTexDataAsRGBA32(&pixels, &width, &height);
791 size_t upload_size = width * height * 4 * sizeof(char);
792
793 /* Upload buffer */
794 VkBufferCreateInfo buffer_info = {};
795 buffer_info.sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO;
796 buffer_info.size = upload_size;
797 buffer_info.usage = VK_BUFFER_USAGE_TRANSFER_SRC_BIT;
798 buffer_info.sharingMode = VK_SHARING_MODE_EXCLUSIVE;
799 VK_CHECK(device_data->vtable.CreateBuffer(device_data->device, &buffer_info,
800 NULL, &data->upload_font_buffer));
801 VkMemoryRequirements upload_buffer_req;
802 device_data->vtable.GetBufferMemoryRequirements(device_data->device,
803 data->upload_font_buffer,
804 &upload_buffer_req);
805 VkMemoryAllocateInfo upload_alloc_info = {};
806 upload_alloc_info.sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO;
807 upload_alloc_info.allocationSize = upload_buffer_req.size;
808 upload_alloc_info.memoryTypeIndex = vk_memory_type(device_data,
809 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT,
810 upload_buffer_req.memoryTypeBits);
811 VK_CHECK(device_data->vtable.AllocateMemory(device_data->device,
812 &upload_alloc_info,
813 NULL,
814 &data->upload_font_buffer_mem));
815 VK_CHECK(device_data->vtable.BindBufferMemory(device_data->device,
816 data->upload_font_buffer,
817 data->upload_font_buffer_mem, 0));
818
819 /* Upload to Buffer */
820 char* map = NULL;
821 VK_CHECK(device_data->vtable.MapMemory(device_data->device,
822 data->upload_font_buffer_mem,
823 0, upload_size, 0, (void**)(&map)));
824 memcpy(map, pixels, upload_size);
825 VkMappedMemoryRange range[1] = {};
826 range[0].sType = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE;
827 range[0].memory = data->upload_font_buffer_mem;
828 range[0].size = upload_size;
829 VK_CHECK(device_data->vtable.FlushMappedMemoryRanges(device_data->device, 1, range));
830 device_data->vtable.UnmapMemory(device_data->device,
831 data->upload_font_buffer_mem);
832
833 /* Copy buffer to image */
834 VkImageMemoryBarrier copy_barrier[1] = {};
835 copy_barrier[0].sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER;
836 copy_barrier[0].dstAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT;
837 copy_barrier[0].oldLayout = VK_IMAGE_LAYOUT_UNDEFINED;
838 copy_barrier[0].newLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
839 copy_barrier[0].srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
840 copy_barrier[0].dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
841 copy_barrier[0].image = data->font_image;
842 copy_barrier[0].subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
843 copy_barrier[0].subresourceRange.levelCount = 1;
844 copy_barrier[0].subresourceRange.layerCount = 1;
845 device_data->vtable.CmdPipelineBarrier(command_buffer,
846 VK_PIPELINE_STAGE_HOST_BIT,
847 VK_PIPELINE_STAGE_TRANSFER_BIT,
848 0, 0, NULL, 0, NULL,
849 1, copy_barrier);
850
851 VkBufferImageCopy region = {};
852 region.imageSubresource.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
853 region.imageSubresource.layerCount = 1;
854 region.imageExtent.width = width;
855 region.imageExtent.height = height;
856 region.imageExtent.depth = 1;
857 device_data->vtable.CmdCopyBufferToImage(command_buffer,
858 data->upload_font_buffer,
859 data->font_image,
860 VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
861 1, &region);
862
863 VkImageMemoryBarrier use_barrier[1] = {};
864 use_barrier[0].sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER;
865 use_barrier[0].srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT;
866 use_barrier[0].dstAccessMask = VK_ACCESS_SHADER_READ_BIT;
867 use_barrier[0].oldLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
868 use_barrier[0].newLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
869 use_barrier[0].srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
870 use_barrier[0].dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
871 use_barrier[0].image = data->font_image;
872 use_barrier[0].subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
873 use_barrier[0].subresourceRange.levelCount = 1;
874 use_barrier[0].subresourceRange.layerCount = 1;
875 device_data->vtable.CmdPipelineBarrier(command_buffer,
876 VK_PIPELINE_STAGE_TRANSFER_BIT,
877 VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT,
878 0,
879 0, NULL,
880 0, NULL,
881 1, use_barrier);
882
883 /* Store our identifier */
884 io.Fonts->TexID = (ImTextureID)(intptr_t)data->font_image;
885 }
886
887 static void CreateOrResizeBuffer(struct device_data *data,
888 VkBuffer *buffer,
889 VkDeviceMemory *buffer_memory,
890 VkDeviceSize *buffer_size,
891 size_t new_size, VkBufferUsageFlagBits usage)
892 {
893 if (*buffer != VK_NULL_HANDLE)
894 data->vtable.DestroyBuffer(data->device, *buffer, NULL);
895 if (*buffer_memory)
896 data->vtable.FreeMemory(data->device, *buffer_memory, NULL);
897
898 VkBufferCreateInfo buffer_info = {};
899 buffer_info.sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO;
900 buffer_info.size = new_size;
901 buffer_info.usage = usage;
902 buffer_info.sharingMode = VK_SHARING_MODE_EXCLUSIVE;
903 VK_CHECK(data->vtable.CreateBuffer(data->device, &buffer_info, NULL, buffer));
904
905 VkMemoryRequirements req;
906 data->vtable.GetBufferMemoryRequirements(data->device, *buffer, &req);
907 VkMemoryAllocateInfo alloc_info = {};
908 alloc_info.sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO;
909 alloc_info.allocationSize = req.size;
910 alloc_info.memoryTypeIndex =
911 vk_memory_type(data, VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT, req.memoryTypeBits);
912 VK_CHECK(data->vtable.AllocateMemory(data->device, &alloc_info, NULL, buffer_memory));
913
914 VK_CHECK(data->vtable.BindBufferMemory(data->device, *buffer, *buffer_memory, 0));
915 *buffer_size = new_size;
916 }
917
918 static struct overlay_draw *render_swapchain_display(struct swapchain_data *data,
919 const VkSemaphore *wait_semaphores,
920 unsigned n_wait_semaphores,
921 unsigned image_index)
922 {
923 ImDrawData* draw_data = ImGui::GetDrawData();
924 if (draw_data->TotalVtxCount == 0)
925 return NULL;
926
927 struct device_data *device_data = data->device;
928 struct overlay_draw *draw = get_overlay_draw(data);
929
930 device_data->vtable.ResetCommandBuffer(draw->command_buffer, 0);
931
932 VkRenderPassBeginInfo render_pass_info = {};
933 render_pass_info.sType = VK_STRUCTURE_TYPE_RENDER_PASS_BEGIN_INFO;
934 render_pass_info.renderPass = data->render_pass;
935 render_pass_info.framebuffer = data->framebuffers[image_index];
936 render_pass_info.renderArea.extent.width = data->width;
937 render_pass_info.renderArea.extent.height = data->height;
938
939 VkCommandBufferBeginInfo buffer_begin_info = {};
940 buffer_begin_info.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO;
941
942 device_data->vtable.BeginCommandBuffer(draw->command_buffer, &buffer_begin_info);
943
944 ensure_swapchain_fonts(data, draw->command_buffer);
945
946 /* Bounce the image to display back to color attachment layout for
947 * rendering on top of it.
948 */
949 VkImageMemoryBarrier imb;
950 imb.sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER;
951 imb.pNext = nullptr;
952 imb.srcAccessMask = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT;
953 imb.dstAccessMask = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT;
954 imb.oldLayout = VK_IMAGE_LAYOUT_PRESENT_SRC_KHR;
955 imb.newLayout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL;
956 imb.image = data->images[image_index];
957 imb.subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
958 imb.subresourceRange.baseMipLevel = 0;
959 imb.subresourceRange.levelCount = 1;
960 imb.subresourceRange.baseArrayLayer = 0;
961 imb.subresourceRange.layerCount = 1;
962 imb.srcQueueFamilyIndex = device_data->graphic_queue->family_index;
963 imb.dstQueueFamilyIndex = device_data->graphic_queue->family_index;
964 device_data->vtable.CmdPipelineBarrier(draw->command_buffer,
965 VK_PIPELINE_STAGE_ALL_GRAPHICS_BIT,
966 VK_PIPELINE_STAGE_ALL_GRAPHICS_BIT,
967 0, /* dependency flags */
968 0, nullptr, /* memory barriers */
969 0, nullptr, /* buffer memory barriers */
970 1, &imb); /* image memory barriers */
971
972 device_data->vtable.CmdBeginRenderPass(draw->command_buffer, &render_pass_info,
973 VK_SUBPASS_CONTENTS_INLINE);
974
975 /* Create/Resize vertex & index buffers */
976 size_t vertex_size = draw_data->TotalVtxCount * sizeof(ImDrawVert);
977 size_t index_size = draw_data->TotalIdxCount * sizeof(ImDrawIdx);
978 if (draw->vertex_buffer_size < vertex_size) {
979 CreateOrResizeBuffer(device_data,
980 &draw->vertex_buffer,
981 &draw->vertex_buffer_mem,
982 &draw->vertex_buffer_size,
983 vertex_size, VK_BUFFER_USAGE_VERTEX_BUFFER_BIT);
984 }
985 if (draw->index_buffer_size < index_size) {
986 CreateOrResizeBuffer(device_data,
987 &draw->index_buffer,
988 &draw->index_buffer_mem,
989 &draw->index_buffer_size,
990 index_size, VK_BUFFER_USAGE_INDEX_BUFFER_BIT);
991 }
992
993 /* Upload vertex & index data */
994 ImDrawVert* vtx_dst = NULL;
995 ImDrawIdx* idx_dst = NULL;
996 VK_CHECK(device_data->vtable.MapMemory(device_data->device, draw->vertex_buffer_mem,
997 0, vertex_size, 0, (void**)(&vtx_dst)));
998 VK_CHECK(device_data->vtable.MapMemory(device_data->device, draw->index_buffer_mem,
999 0, index_size, 0, (void**)(&idx_dst)));
1000 for (int n = 0; n < draw_data->CmdListsCount; n++)
1001 {
1002 const ImDrawList* cmd_list = draw_data->CmdLists[n];
1003 memcpy(vtx_dst, cmd_list->VtxBuffer.Data, cmd_list->VtxBuffer.Size * sizeof(ImDrawVert));
1004 memcpy(idx_dst, cmd_list->IdxBuffer.Data, cmd_list->IdxBuffer.Size * sizeof(ImDrawIdx));
1005 vtx_dst += cmd_list->VtxBuffer.Size;
1006 idx_dst += cmd_list->IdxBuffer.Size;
1007 }
1008 VkMappedMemoryRange range[2] = {};
1009 range[0].sType = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE;
1010 range[0].memory = draw->vertex_buffer_mem;
1011 range[0].size = VK_WHOLE_SIZE;
1012 range[1].sType = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE;
1013 range[1].memory = draw->index_buffer_mem;
1014 range[1].size = VK_WHOLE_SIZE;
1015 VK_CHECK(device_data->vtable.FlushMappedMemoryRanges(device_data->device, 2, range));
1016 device_data->vtable.UnmapMemory(device_data->device, draw->vertex_buffer_mem);
1017 device_data->vtable.UnmapMemory(device_data->device, draw->index_buffer_mem);
1018
1019 /* Bind pipeline and descriptor sets */
1020 device_data->vtable.CmdBindPipeline(draw->command_buffer, VK_PIPELINE_BIND_POINT_GRAPHICS, data->pipeline);
1021 VkDescriptorSet desc_set[1] = { data->descriptor_set };
1022 device_data->vtable.CmdBindDescriptorSets(draw->command_buffer, VK_PIPELINE_BIND_POINT_GRAPHICS,
1023 data->pipeline_layout, 0, 1, desc_set, 0, NULL);
1024
1025 /* Bind vertex & index buffers */
1026 VkBuffer vertex_buffers[1] = { draw->vertex_buffer };
1027 VkDeviceSize vertex_offset[1] = { 0 };
1028 device_data->vtable.CmdBindVertexBuffers(draw->command_buffer, 0, 1, vertex_buffers, vertex_offset);
1029 device_data->vtable.CmdBindIndexBuffer(draw->command_buffer, draw->index_buffer, 0, VK_INDEX_TYPE_UINT16);
1030
1031 /* Setup viewport */
1032 VkViewport viewport;
1033 viewport.x = 0;
1034 viewport.y = 0;
1035 viewport.width = draw_data->DisplaySize.x;
1036 viewport.height = draw_data->DisplaySize.y;
1037 viewport.minDepth = 0.0f;
1038 viewport.maxDepth = 1.0f;
1039 device_data->vtable.CmdSetViewport(draw->command_buffer, 0, 1, &viewport);
1040
1041
1042 /* Setup scale and translation through push constants :
1043 *
1044 * Our visible imgui space lies from draw_data->DisplayPos (top left) to
1045 * draw_data->DisplayPos+data_data->DisplaySize (bottom right). DisplayMin
1046 * is typically (0,0) for single viewport apps.
1047 */
1048 float scale[2];
1049 scale[0] = 2.0f / draw_data->DisplaySize.x;
1050 scale[1] = 2.0f / draw_data->DisplaySize.y;
1051 float translate[2];
1052 translate[0] = -1.0f - draw_data->DisplayPos.x * scale[0];
1053 translate[1] = -1.0f - draw_data->DisplayPos.y * scale[1];
1054 device_data->vtable.CmdPushConstants(draw->command_buffer, data->pipeline_layout,
1055 VK_SHADER_STAGE_VERTEX_BIT,
1056 sizeof(float) * 0, sizeof(float) * 2, scale);
1057 device_data->vtable.CmdPushConstants(draw->command_buffer, data->pipeline_layout,
1058 VK_SHADER_STAGE_VERTEX_BIT,
1059 sizeof(float) * 2, sizeof(float) * 2, translate);
1060
1061 // Render the command lists:
1062 int vtx_offset = 0;
1063 int idx_offset = 0;
1064 ImVec2 display_pos = draw_data->DisplayPos;
1065 for (int n = 0; n < draw_data->CmdListsCount; n++)
1066 {
1067 const ImDrawList* cmd_list = draw_data->CmdLists[n];
1068 for (int cmd_i = 0; cmd_i < cmd_list->CmdBuffer.Size; cmd_i++)
1069 {
1070 const ImDrawCmd* pcmd = &cmd_list->CmdBuffer[cmd_i];
1071 // Apply scissor/clipping rectangle
1072 // FIXME: We could clamp width/height based on clamped min/max values.
1073 VkRect2D scissor;
1074 scissor.offset.x = (int32_t)(pcmd->ClipRect.x - display_pos.x) > 0 ? (int32_t)(pcmd->ClipRect.x - display_pos.x) : 0;
1075 scissor.offset.y = (int32_t)(pcmd->ClipRect.y - display_pos.y) > 0 ? (int32_t)(pcmd->ClipRect.y - display_pos.y) : 0;
1076 scissor.extent.width = (uint32_t)(pcmd->ClipRect.z - pcmd->ClipRect.x);
1077 scissor.extent.height = (uint32_t)(pcmd->ClipRect.w - pcmd->ClipRect.y + 1); // FIXME: Why +1 here?
1078 device_data->vtable.CmdSetScissor(draw->command_buffer, 0, 1, &scissor);
1079
1080 // Draw
1081 device_data->vtable.CmdDrawIndexed(draw->command_buffer, pcmd->ElemCount, 1, idx_offset, vtx_offset, 0);
1082
1083 idx_offset += pcmd->ElemCount;
1084 }
1085 vtx_offset += cmd_list->VtxBuffer.Size;
1086 }
1087
1088 device_data->vtable.CmdEndRenderPass(draw->command_buffer);
1089 device_data->vtable.EndCommandBuffer(draw->command_buffer);
1090
1091 VkSubmitInfo submit_info = {};
1092 VkPipelineStageFlags stage_wait = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
1093 submit_info.sType = VK_STRUCTURE_TYPE_SUBMIT_INFO;
1094 submit_info.commandBufferCount = 1;
1095 submit_info.pCommandBuffers = &draw->command_buffer;
1096 submit_info.pWaitDstStageMask = &stage_wait;
1097 submit_info.waitSemaphoreCount = n_wait_semaphores;
1098 submit_info.pWaitSemaphores = wait_semaphores;
1099 submit_info.signalSemaphoreCount = 1;
1100 submit_info.pSignalSemaphores = &draw->semaphore;
1101
1102 device_data->vtable.QueueSubmit(device_data->graphic_queue->queue, 1, &submit_info, draw->fence);
1103
1104 return draw;
1105 }
1106
1107 static const uint32_t overlay_vert_spv[] = {
1108 #include "overlay.vert.spv.h"
1109 };
1110 static const uint32_t overlay_frag_spv[] = {
1111 #include "overlay.frag.spv.h"
1112 };
1113
1114 static void setup_swapchain_data_pipeline(struct swapchain_data *data)
1115 {
1116 struct device_data *device_data = data->device;
1117 VkShaderModule vert_module, frag_module;
1118
1119 /* Create shader modules */
1120 VkShaderModuleCreateInfo vert_info = {};
1121 vert_info.sType = VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO;
1122 vert_info.codeSize = sizeof(overlay_vert_spv);
1123 vert_info.pCode = overlay_vert_spv;
1124 VK_CHECK(device_data->vtable.CreateShaderModule(device_data->device,
1125 &vert_info, NULL, &vert_module));
1126 VkShaderModuleCreateInfo frag_info = {};
1127 frag_info.sType = VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO;
1128 frag_info.codeSize = sizeof(overlay_frag_spv);
1129 frag_info.pCode = (uint32_t*)overlay_frag_spv;
1130 VK_CHECK(device_data->vtable.CreateShaderModule(device_data->device,
1131 &frag_info, NULL, &frag_module));
1132
1133 /* Font sampler */
1134 VkSamplerCreateInfo sampler_info = {};
1135 sampler_info.sType = VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO;
1136 sampler_info.magFilter = VK_FILTER_LINEAR;
1137 sampler_info.minFilter = VK_FILTER_LINEAR;
1138 sampler_info.mipmapMode = VK_SAMPLER_MIPMAP_MODE_LINEAR;
1139 sampler_info.addressModeU = VK_SAMPLER_ADDRESS_MODE_REPEAT;
1140 sampler_info.addressModeV = VK_SAMPLER_ADDRESS_MODE_REPEAT;
1141 sampler_info.addressModeW = VK_SAMPLER_ADDRESS_MODE_REPEAT;
1142 sampler_info.minLod = -1000;
1143 sampler_info.maxLod = 1000;
1144 sampler_info.maxAnisotropy = 1.0f;
1145 VK_CHECK(device_data->vtable.CreateSampler(device_data->device, &sampler_info,
1146 NULL, &data->font_sampler));
1147
1148 /* Descriptor pool */
1149 VkDescriptorPoolSize sampler_pool_size = {};
1150 sampler_pool_size.type = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER;
1151 sampler_pool_size.descriptorCount = 1;
1152 VkDescriptorPoolCreateInfo desc_pool_info = {};
1153 desc_pool_info.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_POOL_CREATE_INFO;
1154 desc_pool_info.maxSets = 1;
1155 desc_pool_info.poolSizeCount = 1;
1156 desc_pool_info.pPoolSizes = &sampler_pool_size;
1157 VK_CHECK(device_data->vtable.CreateDescriptorPool(device_data->device,
1158 &desc_pool_info,
1159 NULL, &data->descriptor_pool));
1160
1161 /* Descriptor layout */
1162 VkSampler sampler[1] = { data->font_sampler };
1163 VkDescriptorSetLayoutBinding binding[1] = {};
1164 binding[0].descriptorType = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER;
1165 binding[0].descriptorCount = 1;
1166 binding[0].stageFlags = VK_SHADER_STAGE_FRAGMENT_BIT;
1167 binding[0].pImmutableSamplers = sampler;
1168 VkDescriptorSetLayoutCreateInfo set_layout_info = {};
1169 set_layout_info.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO;
1170 set_layout_info.bindingCount = 1;
1171 set_layout_info.pBindings = binding;
1172 VK_CHECK(device_data->vtable.CreateDescriptorSetLayout(device_data->device,
1173 &set_layout_info,
1174 NULL, &data->descriptor_layout));
1175
1176 /* Descriptor set */
1177 VkDescriptorSetAllocateInfo alloc_info = {};
1178 alloc_info.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_ALLOCATE_INFO;
1179 alloc_info.descriptorPool = data->descriptor_pool;
1180 alloc_info.descriptorSetCount = 1;
1181 alloc_info.pSetLayouts = &data->descriptor_layout;
1182 VK_CHECK(device_data->vtable.AllocateDescriptorSets(device_data->device,
1183 &alloc_info,
1184 &data->descriptor_set));
1185
1186 /* Constants: we are using 'vec2 offset' and 'vec2 scale' instead of a full
1187 * 3d projection matrix
1188 */
1189 VkPushConstantRange push_constants[1] = {};
1190 push_constants[0].stageFlags = VK_SHADER_STAGE_VERTEX_BIT;
1191 push_constants[0].offset = sizeof(float) * 0;
1192 push_constants[0].size = sizeof(float) * 4;
1193 VkPipelineLayoutCreateInfo layout_info = {};
1194 layout_info.sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO;
1195 layout_info.setLayoutCount = 1;
1196 layout_info.pSetLayouts = &data->descriptor_layout;
1197 layout_info.pushConstantRangeCount = 1;
1198 layout_info.pPushConstantRanges = push_constants;
1199 VK_CHECK(device_data->vtable.CreatePipelineLayout(device_data->device,
1200 &layout_info,
1201 NULL, &data->pipeline_layout));
1202
1203 VkPipelineShaderStageCreateInfo stage[2] = {};
1204 stage[0].sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO;
1205 stage[0].stage = VK_SHADER_STAGE_VERTEX_BIT;
1206 stage[0].module = vert_module;
1207 stage[0].pName = "main";
1208 stage[1].sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO;
1209 stage[1].stage = VK_SHADER_STAGE_FRAGMENT_BIT;
1210 stage[1].module = frag_module;
1211 stage[1].pName = "main";
1212
1213 VkVertexInputBindingDescription binding_desc[1] = {};
1214 binding_desc[0].stride = sizeof(ImDrawVert);
1215 binding_desc[0].inputRate = VK_VERTEX_INPUT_RATE_VERTEX;
1216
1217 VkVertexInputAttributeDescription attribute_desc[3] = {};
1218 attribute_desc[0].location = 0;
1219 attribute_desc[0].binding = binding_desc[0].binding;
1220 attribute_desc[0].format = VK_FORMAT_R32G32_SFLOAT;
1221 attribute_desc[0].offset = IM_OFFSETOF(ImDrawVert, pos);
1222 attribute_desc[1].location = 1;
1223 attribute_desc[1].binding = binding_desc[0].binding;
1224 attribute_desc[1].format = VK_FORMAT_R32G32_SFLOAT;
1225 attribute_desc[1].offset = IM_OFFSETOF(ImDrawVert, uv);
1226 attribute_desc[2].location = 2;
1227 attribute_desc[2].binding = binding_desc[0].binding;
1228 attribute_desc[2].format = VK_FORMAT_R8G8B8A8_UNORM;
1229 attribute_desc[2].offset = IM_OFFSETOF(ImDrawVert, col);
1230
1231 VkPipelineVertexInputStateCreateInfo vertex_info = {};
1232 vertex_info.sType = VK_STRUCTURE_TYPE_PIPELINE_VERTEX_INPUT_STATE_CREATE_INFO;
1233 vertex_info.vertexBindingDescriptionCount = 1;
1234 vertex_info.pVertexBindingDescriptions = binding_desc;
1235 vertex_info.vertexAttributeDescriptionCount = 3;
1236 vertex_info.pVertexAttributeDescriptions = attribute_desc;
1237
1238 VkPipelineInputAssemblyStateCreateInfo ia_info = {};
1239 ia_info.sType = VK_STRUCTURE_TYPE_PIPELINE_INPUT_ASSEMBLY_STATE_CREATE_INFO;
1240 ia_info.topology = VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST;
1241
1242 VkPipelineViewportStateCreateInfo viewport_info = {};
1243 viewport_info.sType = VK_STRUCTURE_TYPE_PIPELINE_VIEWPORT_STATE_CREATE_INFO;
1244 viewport_info.viewportCount = 1;
1245 viewport_info.scissorCount = 1;
1246
1247 VkPipelineRasterizationStateCreateInfo raster_info = {};
1248 raster_info.sType = VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_STATE_CREATE_INFO;
1249 raster_info.polygonMode = VK_POLYGON_MODE_FILL;
1250 raster_info.cullMode = VK_CULL_MODE_NONE;
1251 raster_info.frontFace = VK_FRONT_FACE_COUNTER_CLOCKWISE;
1252 raster_info.lineWidth = 1.0f;
1253
1254 VkPipelineMultisampleStateCreateInfo ms_info = {};
1255 ms_info.sType = VK_STRUCTURE_TYPE_PIPELINE_MULTISAMPLE_STATE_CREATE_INFO;
1256 ms_info.rasterizationSamples = VK_SAMPLE_COUNT_1_BIT;
1257
1258 VkPipelineColorBlendAttachmentState color_attachment[1] = {};
1259 color_attachment[0].blendEnable = VK_TRUE;
1260 color_attachment[0].srcColorBlendFactor = VK_BLEND_FACTOR_SRC_ALPHA;
1261 color_attachment[0].dstColorBlendFactor = VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA;
1262 color_attachment[0].colorBlendOp = VK_BLEND_OP_ADD;
1263 color_attachment[0].srcAlphaBlendFactor = VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA;
1264 color_attachment[0].dstAlphaBlendFactor = VK_BLEND_FACTOR_ZERO;
1265 color_attachment[0].alphaBlendOp = VK_BLEND_OP_ADD;
1266 color_attachment[0].colorWriteMask = VK_COLOR_COMPONENT_R_BIT |
1267 VK_COLOR_COMPONENT_G_BIT | VK_COLOR_COMPONENT_B_BIT | VK_COLOR_COMPONENT_A_BIT;
1268
1269 VkPipelineDepthStencilStateCreateInfo depth_info = {};
1270 depth_info.sType = VK_STRUCTURE_TYPE_PIPELINE_DEPTH_STENCIL_STATE_CREATE_INFO;
1271
1272 VkPipelineColorBlendStateCreateInfo blend_info = {};
1273 blend_info.sType = VK_STRUCTURE_TYPE_PIPELINE_COLOR_BLEND_STATE_CREATE_INFO;
1274 blend_info.attachmentCount = 1;
1275 blend_info.pAttachments = color_attachment;
1276
1277 VkDynamicState dynamic_states[2] = { VK_DYNAMIC_STATE_VIEWPORT, VK_DYNAMIC_STATE_SCISSOR };
1278 VkPipelineDynamicStateCreateInfo dynamic_state = {};
1279 dynamic_state.sType = VK_STRUCTURE_TYPE_PIPELINE_DYNAMIC_STATE_CREATE_INFO;
1280 dynamic_state.dynamicStateCount = (uint32_t)IM_ARRAYSIZE(dynamic_states);
1281 dynamic_state.pDynamicStates = dynamic_states;
1282
1283 VkGraphicsPipelineCreateInfo info = {};
1284 info.sType = VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO;
1285 info.flags = 0;
1286 info.stageCount = 2;
1287 info.pStages = stage;
1288 info.pVertexInputState = &vertex_info;
1289 info.pInputAssemblyState = &ia_info;
1290 info.pViewportState = &viewport_info;
1291 info.pRasterizationState = &raster_info;
1292 info.pMultisampleState = &ms_info;
1293 info.pDepthStencilState = &depth_info;
1294 info.pColorBlendState = &blend_info;
1295 info.pDynamicState = &dynamic_state;
1296 info.layout = data->pipeline_layout;
1297 info.renderPass = data->render_pass;
1298 VK_CHECK(
1299 device_data->vtable.CreateGraphicsPipelines(device_data->device, VK_NULL_HANDLE,
1300 1, &info,
1301 NULL, &data->pipeline));
1302
1303 device_data->vtable.DestroyShaderModule(device_data->device, vert_module, NULL);
1304 device_data->vtable.DestroyShaderModule(device_data->device, frag_module, NULL);
1305
1306 ImGuiIO& io = ImGui::GetIO();
1307 unsigned char* pixels;
1308 int width, height;
1309 io.Fonts->GetTexDataAsRGBA32(&pixels, &width, &height);
1310
1311 /* Font image */
1312 VkImageCreateInfo image_info = {};
1313 image_info.sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO;
1314 image_info.imageType = VK_IMAGE_TYPE_2D;
1315 image_info.format = VK_FORMAT_R8G8B8A8_UNORM;
1316 image_info.extent.width = width;
1317 image_info.extent.height = height;
1318 image_info.extent.depth = 1;
1319 image_info.mipLevels = 1;
1320 image_info.arrayLayers = 1;
1321 image_info.samples = VK_SAMPLE_COUNT_1_BIT;
1322 image_info.tiling = VK_IMAGE_TILING_OPTIMAL;
1323 image_info.usage = VK_IMAGE_USAGE_SAMPLED_BIT | VK_IMAGE_USAGE_TRANSFER_DST_BIT;
1324 image_info.sharingMode = VK_SHARING_MODE_EXCLUSIVE;
1325 image_info.initialLayout = VK_IMAGE_LAYOUT_UNDEFINED;
1326 VK_CHECK(device_data->vtable.CreateImage(device_data->device, &image_info,
1327 NULL, &data->font_image));
1328 VkMemoryRequirements font_image_req;
1329 device_data->vtable.GetImageMemoryRequirements(device_data->device,
1330 data->font_image, &font_image_req);
1331 VkMemoryAllocateInfo image_alloc_info = {};
1332 image_alloc_info.sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO;
1333 image_alloc_info.allocationSize = font_image_req.size;
1334 image_alloc_info.memoryTypeIndex = vk_memory_type(device_data,
1335 VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT,
1336 font_image_req.memoryTypeBits);
1337 VK_CHECK(device_data->vtable.AllocateMemory(device_data->device, &image_alloc_info,
1338 NULL, &data->font_mem));
1339 VK_CHECK(device_data->vtable.BindImageMemory(device_data->device,
1340 data->font_image,
1341 data->font_mem, 0));
1342
1343 /* Font image view */
1344 VkImageViewCreateInfo view_info = {};
1345 view_info.sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO;
1346 view_info.image = data->font_image;
1347 view_info.viewType = VK_IMAGE_VIEW_TYPE_2D;
1348 view_info.format = VK_FORMAT_R8G8B8A8_UNORM;
1349 view_info.subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
1350 view_info.subresourceRange.levelCount = 1;
1351 view_info.subresourceRange.layerCount = 1;
1352 VK_CHECK(device_data->vtable.CreateImageView(device_data->device, &view_info,
1353 NULL, &data->font_image_view));
1354
1355 /* Descriptor set */
1356 VkDescriptorImageInfo desc_image[1] = {};
1357 desc_image[0].sampler = data->font_sampler;
1358 desc_image[0].imageView = data->font_image_view;
1359 desc_image[0].imageLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
1360 VkWriteDescriptorSet write_desc[1] = {};
1361 write_desc[0].sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET;
1362 write_desc[0].dstSet = data->descriptor_set;
1363 write_desc[0].descriptorCount = 1;
1364 write_desc[0].descriptorType = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER;
1365 write_desc[0].pImageInfo = desc_image;
1366 device_data->vtable.UpdateDescriptorSets(device_data->device, 1, write_desc, 0, NULL);
1367 }
1368
1369 static void setup_swapchain_data(struct swapchain_data *data,
1370 const VkSwapchainCreateInfoKHR *pCreateInfo)
1371 {
1372 data->width = pCreateInfo->imageExtent.width;
1373 data->height = pCreateInfo->imageExtent.height;
1374 data->format = pCreateInfo->imageFormat;
1375
1376 data->imgui_context = ImGui::CreateContext();
1377 ImGui::SetCurrentContext(data->imgui_context);
1378
1379 ImGui::GetIO().IniFilename = NULL;
1380 ImGui::GetIO().DisplaySize = ImVec2((float)data->width, (float)data->height);
1381
1382 struct device_data *device_data = data->device;
1383
1384 /* Render pass */
1385 VkAttachmentDescription attachment_desc = {};
1386 attachment_desc.format = pCreateInfo->imageFormat;
1387 attachment_desc.samples = VK_SAMPLE_COUNT_1_BIT;
1388 attachment_desc.loadOp = VK_ATTACHMENT_LOAD_OP_LOAD;
1389 attachment_desc.storeOp = VK_ATTACHMENT_STORE_OP_STORE;
1390 attachment_desc.stencilLoadOp = VK_ATTACHMENT_LOAD_OP_DONT_CARE;
1391 attachment_desc.stencilStoreOp = VK_ATTACHMENT_STORE_OP_DONT_CARE;
1392 attachment_desc.initialLayout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL;
1393 attachment_desc.finalLayout = VK_IMAGE_LAYOUT_PRESENT_SRC_KHR;
1394 VkAttachmentReference color_attachment = {};
1395 color_attachment.attachment = 0;
1396 color_attachment.layout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL;
1397 VkSubpassDescription subpass = {};
1398 subpass.pipelineBindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS;
1399 subpass.colorAttachmentCount = 1;
1400 subpass.pColorAttachments = &color_attachment;
1401 VkSubpassDependency dependency = {};
1402 dependency.srcSubpass = VK_SUBPASS_EXTERNAL;
1403 dependency.dstSubpass = 0;
1404 dependency.srcStageMask = VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT;
1405 dependency.dstStageMask = VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT;
1406 dependency.srcAccessMask = 0;
1407 dependency.dstAccessMask = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT;
1408 VkRenderPassCreateInfo render_pass_info = {};
1409 render_pass_info.sType = VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO;
1410 render_pass_info.attachmentCount = 1;
1411 render_pass_info.pAttachments = &attachment_desc;
1412 render_pass_info.subpassCount = 1;
1413 render_pass_info.pSubpasses = &subpass;
1414 render_pass_info.dependencyCount = 1;
1415 render_pass_info.pDependencies = &dependency;
1416 VK_CHECK(device_data->vtable.CreateRenderPass(device_data->device,
1417 &render_pass_info,
1418 NULL, &data->render_pass));
1419
1420 setup_swapchain_data_pipeline(data);
1421
1422 VK_CHECK(device_data->vtable.GetSwapchainImagesKHR(device_data->device,
1423 data->swapchain,
1424 &data->n_images,
1425 NULL));
1426
1427 data->images = ralloc_array(data, VkImage, data->n_images);
1428 data->image_views = ralloc_array(data, VkImageView, data->n_images);
1429 data->framebuffers = ralloc_array(data, VkFramebuffer, data->n_images);
1430
1431 VK_CHECK(device_data->vtable.GetSwapchainImagesKHR(device_data->device,
1432 data->swapchain,
1433 &data->n_images,
1434 data->images));
1435
1436 /* Image views */
1437 VkImageViewCreateInfo view_info = {};
1438 view_info.sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO;
1439 view_info.viewType = VK_IMAGE_VIEW_TYPE_2D;
1440 view_info.format = pCreateInfo->imageFormat;
1441 view_info.components.r = VK_COMPONENT_SWIZZLE_R;
1442 view_info.components.g = VK_COMPONENT_SWIZZLE_G;
1443 view_info.components.b = VK_COMPONENT_SWIZZLE_B;
1444 view_info.components.a = VK_COMPONENT_SWIZZLE_A;
1445 view_info.subresourceRange = { VK_IMAGE_ASPECT_COLOR_BIT, 0, 1, 0, 1 };
1446 for (uint32_t i = 0; i < data->n_images; i++) {
1447 view_info.image = data->images[i];
1448 VK_CHECK(device_data->vtable.CreateImageView(device_data->device,
1449 &view_info, NULL,
1450 &data->image_views[i]));
1451 }
1452
1453 /* Framebuffers */
1454 VkImageView attachment[1];
1455 VkFramebufferCreateInfo fb_info = {};
1456 fb_info.sType = VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO;
1457 fb_info.renderPass = data->render_pass;
1458 fb_info.attachmentCount = 1;
1459 fb_info.pAttachments = attachment;
1460 fb_info.width = data->width;
1461 fb_info.height = data->height;
1462 fb_info.layers = 1;
1463 for (uint32_t i = 0; i < data->n_images; i++) {
1464 attachment[0] = data->image_views[i];
1465 VK_CHECK(device_data->vtable.CreateFramebuffer(device_data->device, &fb_info,
1466 NULL, &data->framebuffers[i]));
1467 }
1468
1469 /* Command buffer pool */
1470 VkCommandPoolCreateInfo cmd_buffer_pool_info = {};
1471 cmd_buffer_pool_info.sType = VK_STRUCTURE_TYPE_COMMAND_POOL_CREATE_INFO;
1472 cmd_buffer_pool_info.flags = VK_COMMAND_POOL_CREATE_RESET_COMMAND_BUFFER_BIT;
1473 cmd_buffer_pool_info.queueFamilyIndex = device_data->graphic_queue->family_index;
1474 VK_CHECK(device_data->vtable.CreateCommandPool(device_data->device,
1475 &cmd_buffer_pool_info,
1476 NULL, &data->command_pool));
1477 }
1478
1479 static void shutdown_swapchain_data(struct swapchain_data *data)
1480 {
1481 struct device_data *device_data = data->device;
1482
1483 list_for_each_entry_safe(struct overlay_draw, draw, &data->draws, link) {
1484 device_data->vtable.DestroySemaphore(device_data->device, draw->semaphore, NULL);
1485 device_data->vtable.DestroyFence(device_data->device, draw->fence, NULL);
1486 device_data->vtable.DestroyBuffer(device_data->device, draw->vertex_buffer, NULL);
1487 device_data->vtable.DestroyBuffer(device_data->device, draw->index_buffer, NULL);
1488 device_data->vtable.FreeMemory(device_data->device, draw->vertex_buffer_mem, NULL);
1489 device_data->vtable.FreeMemory(device_data->device, draw->index_buffer_mem, NULL);
1490 }
1491
1492 for (uint32_t i = 0; i < data->n_images; i++) {
1493 device_data->vtable.DestroyImageView(device_data->device, data->image_views[i], NULL);
1494 device_data->vtable.DestroyFramebuffer(device_data->device, data->framebuffers[i], NULL);
1495 }
1496
1497 device_data->vtable.DestroyRenderPass(device_data->device, data->render_pass, NULL);
1498
1499 device_data->vtable.DestroyCommandPool(device_data->device, data->command_pool, NULL);
1500
1501 device_data->vtable.DestroyPipeline(device_data->device, data->pipeline, NULL);
1502 device_data->vtable.DestroyPipelineLayout(device_data->device, data->pipeline_layout, NULL);
1503
1504 device_data->vtable.DestroyDescriptorPool(device_data->device,
1505 data->descriptor_pool, NULL);
1506 device_data->vtable.DestroyDescriptorSetLayout(device_data->device,
1507 data->descriptor_layout, NULL);
1508
1509 device_data->vtable.DestroySampler(device_data->device, data->font_sampler, NULL);
1510 device_data->vtable.DestroyImageView(device_data->device, data->font_image_view, NULL);
1511 device_data->vtable.DestroyImage(device_data->device, data->font_image, NULL);
1512 device_data->vtable.FreeMemory(device_data->device, data->font_mem, NULL);
1513
1514 device_data->vtable.DestroyBuffer(device_data->device, data->upload_font_buffer, NULL);
1515 device_data->vtable.FreeMemory(device_data->device, data->upload_font_buffer_mem, NULL);
1516
1517 ImGui::DestroyContext(data->imgui_context);
1518 }
1519
1520 static struct overlay_draw *before_present(struct swapchain_data *swapchain_data,
1521 const VkSemaphore *wait_semaphores,
1522 unsigned n_wait_semaphores,
1523 unsigned imageIndex)
1524 {
1525 struct instance_data *instance_data = swapchain_data->device->instance;
1526 struct overlay_draw *draw = NULL;
1527
1528 snapshot_swapchain_frame(swapchain_data);
1529
1530 if (!instance_data->params.no_display && swapchain_data->n_frames > 0) {
1531 compute_swapchain_display(swapchain_data);
1532 draw = render_swapchain_display(swapchain_data,
1533 wait_semaphores, n_wait_semaphores,
1534 imageIndex);
1535 }
1536
1537 return draw;
1538 }
1539
1540 static VkResult overlay_CreateSwapchainKHR(
1541 VkDevice device,
1542 const VkSwapchainCreateInfoKHR* pCreateInfo,
1543 const VkAllocationCallbacks* pAllocator,
1544 VkSwapchainKHR* pSwapchain)
1545 {
1546 struct device_data *device_data = FIND_DEVICE_DATA(device);
1547 VkResult result = device_data->vtable.CreateSwapchainKHR(device, pCreateInfo, pAllocator, pSwapchain);
1548 if (result != VK_SUCCESS) return result;
1549
1550 struct swapchain_data *swapchain_data = new_swapchain_data(*pSwapchain, device_data);
1551 setup_swapchain_data(swapchain_data, pCreateInfo);
1552 return result;
1553 }
1554
1555 static void overlay_DestroySwapchainKHR(
1556 VkDevice device,
1557 VkSwapchainKHR swapchain,
1558 const VkAllocationCallbacks* pAllocator)
1559 {
1560 struct swapchain_data *swapchain_data = FIND_SWAPCHAIN_DATA(swapchain);
1561
1562 shutdown_swapchain_data(swapchain_data);
1563 swapchain_data->device->vtable.DestroySwapchainKHR(device, swapchain, pAllocator);
1564 destroy_swapchain_data(swapchain_data);
1565 }
1566
1567 static VkResult overlay_QueuePresentKHR(
1568 VkQueue queue,
1569 const VkPresentInfoKHR* pPresentInfo)
1570 {
1571 struct queue_data *queue_data = FIND_QUEUE_DATA(queue);
1572 struct device_data *device_data = queue_data->device;
1573 struct instance_data *instance_data = device_data->instance;
1574 uint32_t query_results[OVERLAY_QUERY_COUNT];
1575
1576 device_data->frame_stats.stats[OVERLAY_PARAM_ENABLED_frame]++;
1577
1578 if (list_length(&queue_data->running_command_buffer) > 0) {
1579 /* Before getting the query results, make sure the operations have
1580 * completed.
1581 */
1582 VK_CHECK(device_data->vtable.ResetFences(device_data->device,
1583 1, &queue_data->queries_fence));
1584 VK_CHECK(device_data->vtable.QueueSubmit(queue, 0, NULL, queue_data->queries_fence));
1585 VK_CHECK(device_data->vtable.WaitForFences(device_data->device,
1586 1, &queue_data->queries_fence,
1587 VK_FALSE, UINT64_MAX));
1588
1589 /* Now get the results. */
1590 list_for_each_entry_safe(struct command_buffer_data, cmd_buffer_data,
1591 &queue_data->running_command_buffer, link) {
1592 list_delinit(&cmd_buffer_data->link);
1593
1594 if (cmd_buffer_data->pipeline_query_pool) {
1595 memset(query_results, 0, sizeof(query_results));
1596 VK_CHECK(device_data->vtable.GetQueryPoolResults(device_data->device,
1597 cmd_buffer_data->pipeline_query_pool,
1598 cmd_buffer_data->query_index, 1,
1599 sizeof(uint32_t) * OVERLAY_QUERY_COUNT,
1600 query_results, 0, VK_QUERY_RESULT_WAIT_BIT));
1601
1602 for (uint32_t i = OVERLAY_PARAM_ENABLED_vertices;
1603 i <= OVERLAY_PARAM_ENABLED_compute_invocations; i++) {
1604 device_data->frame_stats.stats[i] += query_results[i - OVERLAY_PARAM_ENABLED_vertices];
1605 }
1606 }
1607 if (cmd_buffer_data->timestamp_query_pool) {
1608 uint64_t gpu_timestamps[2] = { 0 };
1609 VK_CHECK(device_data->vtable.GetQueryPoolResults(device_data->device,
1610 cmd_buffer_data->timestamp_query_pool,
1611 cmd_buffer_data->query_index * 2, 2,
1612 2 * sizeof(uint64_t), gpu_timestamps, sizeof(uint64_t),
1613 VK_QUERY_RESULT_WAIT_BIT | VK_QUERY_RESULT_64_BIT));
1614
1615 gpu_timestamps[0] &= queue_data->timestamp_mask;
1616 gpu_timestamps[1] &= queue_data->timestamp_mask;
1617 device_data->frame_stats.stats[OVERLAY_PARAM_ENABLED_gpu_timing] +=
1618 (gpu_timestamps[1] - gpu_timestamps[0]) *
1619 device_data->properties.limits.timestampPeriod;
1620 }
1621 }
1622 }
1623
1624 /* Otherwise we need to add our overlay drawing semaphore to the list of
1625 * semaphores to wait on. If we don't do that the presented picture might
1626 * be have incomplete overlay drawings.
1627 */
1628 VkResult result = VK_SUCCESS;
1629 if (instance_data->params.no_display) {
1630 for (uint32_t i = 0; i < pPresentInfo->swapchainCount; i++) {
1631 VkSwapchainKHR swapchain = pPresentInfo->pSwapchains[i];
1632 struct swapchain_data *swapchain_data = FIND_SWAPCHAIN_DATA(swapchain);
1633
1634 before_present(swapchain_data,
1635 pPresentInfo->pWaitSemaphores,
1636 pPresentInfo->waitSemaphoreCount,
1637 pPresentInfo->pImageIndices[i]);
1638
1639 VkPresentInfoKHR present_info = *pPresentInfo;
1640 present_info.swapchainCount = 1;
1641 present_info.pSwapchains = &swapchain;
1642
1643 uint64_t ts0 = os_time_get();
1644 result = queue_data->device->vtable.QueuePresentKHR(queue, &present_info);
1645 uint64_t ts1 = os_time_get();
1646 swapchain_data->frame_stats.stats[OVERLAY_PARAM_ENABLED_present_timing] += ts1 - ts0;
1647 }
1648 } else {
1649 for (uint32_t i = 0; i < pPresentInfo->swapchainCount; i++) {
1650 VkSwapchainKHR swapchain = pPresentInfo->pSwapchains[i];
1651 struct swapchain_data *swapchain_data = FIND_SWAPCHAIN_DATA(swapchain);
1652 VkPresentInfoKHR present_info = *pPresentInfo;
1653 present_info.swapchainCount = 1;
1654 present_info.pSwapchains = &swapchain;
1655
1656 uint32_t image_index = pPresentInfo->pImageIndices[i];
1657
1658 struct overlay_draw *draw = before_present(swapchain_data,
1659 pPresentInfo->pWaitSemaphores,
1660 pPresentInfo->waitSemaphoreCount,
1661 image_index);
1662
1663 /* Because the submission of the overlay draw waits on the semaphores
1664 * handed for present, we don't need to have this present operation
1665 * wait on them as well, we can just wait on the overlay submission
1666 * semaphore.
1667 */
1668 present_info.pWaitSemaphores = &draw->semaphore;
1669 present_info.waitSemaphoreCount = 1;
1670
1671 uint64_t ts0 = os_time_get();
1672 VkResult chain_result = queue_data->device->vtable.QueuePresentKHR(queue, &present_info);
1673 uint64_t ts1 = os_time_get();
1674 swapchain_data->frame_stats.stats[OVERLAY_PARAM_ENABLED_present_timing] += ts1 - ts0;
1675 if (pPresentInfo->pResults)
1676 pPresentInfo->pResults[i] = chain_result;
1677 if (chain_result != VK_SUCCESS && result == VK_SUCCESS)
1678 result = chain_result;
1679 }
1680 }
1681 return result;
1682 }
1683
1684 static VkResult overlay_AcquireNextImageKHR(
1685 VkDevice device,
1686 VkSwapchainKHR swapchain,
1687 uint64_t timeout,
1688 VkSemaphore semaphore,
1689 VkFence fence,
1690 uint32_t* pImageIndex)
1691 {
1692 struct swapchain_data *swapchain_data = FIND_SWAPCHAIN_DATA(swapchain);
1693 struct device_data *device_data = swapchain_data->device;
1694
1695 uint64_t ts0 = os_time_get();
1696 VkResult result = device_data->vtable.AcquireNextImageKHR(device, swapchain, timeout,
1697 semaphore, fence, pImageIndex);
1698 uint64_t ts1 = os_time_get();
1699
1700 swapchain_data->frame_stats.stats[OVERLAY_PARAM_ENABLED_acquire_timing] += ts1 - ts0;
1701 swapchain_data->frame_stats.stats[OVERLAY_PARAM_ENABLED_acquire]++;
1702
1703 return result;
1704 }
1705
1706 static VkResult overlay_AcquireNextImage2KHR(
1707 VkDevice device,
1708 const VkAcquireNextImageInfoKHR* pAcquireInfo,
1709 uint32_t* pImageIndex)
1710 {
1711 struct swapchain_data *swapchain_data = FIND_SWAPCHAIN_DATA(pAcquireInfo->swapchain);
1712 struct device_data *device_data = swapchain_data->device;
1713
1714 uint64_t ts0 = os_time_get();
1715 VkResult result = device_data->vtable.AcquireNextImage2KHR(device, pAcquireInfo, pImageIndex);
1716 uint64_t ts1 = os_time_get();
1717
1718 swapchain_data->frame_stats.stats[OVERLAY_PARAM_ENABLED_acquire_timing] += ts1 - ts0;
1719 swapchain_data->frame_stats.stats[OVERLAY_PARAM_ENABLED_acquire]++;
1720
1721 return result;
1722 }
1723
1724 static void overlay_CmdDraw(
1725 VkCommandBuffer commandBuffer,
1726 uint32_t vertexCount,
1727 uint32_t instanceCount,
1728 uint32_t firstVertex,
1729 uint32_t firstInstance)
1730 {
1731 struct command_buffer_data *cmd_buffer_data = FIND_CMD_BUFFER_DATA(commandBuffer);
1732 cmd_buffer_data->stats.stats[OVERLAY_PARAM_ENABLED_draw]++;
1733 struct device_data *device_data = cmd_buffer_data->device;
1734 device_data->vtable.CmdDraw(commandBuffer, vertexCount, instanceCount,
1735 firstVertex, firstInstance);
1736 }
1737
1738 static void overlay_CmdDrawIndexed(
1739 VkCommandBuffer commandBuffer,
1740 uint32_t indexCount,
1741 uint32_t instanceCount,
1742 uint32_t firstIndex,
1743 int32_t vertexOffset,
1744 uint32_t firstInstance)
1745 {
1746 struct command_buffer_data *cmd_buffer_data = FIND_CMD_BUFFER_DATA(commandBuffer);
1747 cmd_buffer_data->stats.stats[OVERLAY_PARAM_ENABLED_draw_indexed]++;
1748 struct device_data *device_data = cmd_buffer_data->device;
1749 device_data->vtable.CmdDrawIndexed(commandBuffer, indexCount, instanceCount,
1750 firstIndex, vertexOffset, firstInstance);
1751 }
1752
1753 static void overlay_CmdDrawIndirect(
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_indirect]++;
1762 struct device_data *device_data = cmd_buffer_data->device;
1763 device_data->vtable.CmdDrawIndirect(commandBuffer, buffer, offset, drawCount, stride);
1764 }
1765
1766 static void overlay_CmdDrawIndexedIndirect(
1767 VkCommandBuffer commandBuffer,
1768 VkBuffer buffer,
1769 VkDeviceSize offset,
1770 uint32_t drawCount,
1771 uint32_t stride)
1772 {
1773 struct command_buffer_data *cmd_buffer_data = FIND_CMD_BUFFER_DATA(commandBuffer);
1774 cmd_buffer_data->stats.stats[OVERLAY_PARAM_ENABLED_draw_indexed_indirect]++;
1775 struct device_data *device_data = cmd_buffer_data->device;
1776 device_data->vtable.CmdDrawIndexedIndirect(commandBuffer, buffer, offset, drawCount, stride);
1777 }
1778
1779 static void overlay_CmdDrawIndirectCountKHR(
1780 VkCommandBuffer commandBuffer,
1781 VkBuffer buffer,
1782 VkDeviceSize offset,
1783 VkBuffer countBuffer,
1784 VkDeviceSize countBufferOffset,
1785 uint32_t maxDrawCount,
1786 uint32_t stride)
1787 {
1788 struct command_buffer_data *cmd_buffer_data = FIND_CMD_BUFFER_DATA(commandBuffer);
1789 cmd_buffer_data->stats.stats[OVERLAY_PARAM_ENABLED_draw_indirect_count]++;
1790 struct device_data *device_data = cmd_buffer_data->device;
1791 device_data->vtable.CmdDrawIndirectCountKHR(commandBuffer, buffer, offset,
1792 countBuffer, countBufferOffset,
1793 maxDrawCount, stride);
1794 }
1795
1796 static void overlay_CmdDrawIndexedIndirectCountKHR(
1797 VkCommandBuffer commandBuffer,
1798 VkBuffer buffer,
1799 VkDeviceSize offset,
1800 VkBuffer countBuffer,
1801 VkDeviceSize countBufferOffset,
1802 uint32_t maxDrawCount,
1803 uint32_t stride)
1804 {
1805 struct command_buffer_data *cmd_buffer_data = FIND_CMD_BUFFER_DATA(commandBuffer);
1806 cmd_buffer_data->stats.stats[OVERLAY_PARAM_ENABLED_draw_indexed_indirect_count]++;
1807 struct device_data *device_data = cmd_buffer_data->device;
1808 device_data->vtable.CmdDrawIndexedIndirectCountKHR(commandBuffer, buffer, offset,
1809 countBuffer, countBufferOffset,
1810 maxDrawCount, stride);
1811 }
1812
1813 static void overlay_CmdDispatch(
1814 VkCommandBuffer commandBuffer,
1815 uint32_t groupCountX,
1816 uint32_t groupCountY,
1817 uint32_t groupCountZ)
1818 {
1819 struct command_buffer_data *cmd_buffer_data = FIND_CMD_BUFFER_DATA(commandBuffer);
1820 cmd_buffer_data->stats.stats[OVERLAY_PARAM_ENABLED_dispatch]++;
1821 struct device_data *device_data = cmd_buffer_data->device;
1822 device_data->vtable.CmdDispatch(commandBuffer, groupCountX, groupCountY, groupCountZ);
1823 }
1824
1825 static void overlay_CmdDispatchIndirect(
1826 VkCommandBuffer commandBuffer,
1827 VkBuffer buffer,
1828 VkDeviceSize offset)
1829 {
1830 struct command_buffer_data *cmd_buffer_data = FIND_CMD_BUFFER_DATA(commandBuffer);
1831 cmd_buffer_data->stats.stats[OVERLAY_PARAM_ENABLED_dispatch_indirect]++;
1832 struct device_data *device_data = cmd_buffer_data->device;
1833 device_data->vtable.CmdDispatchIndirect(commandBuffer, buffer, offset);
1834 }
1835
1836 static void overlay_CmdBindPipeline(
1837 VkCommandBuffer commandBuffer,
1838 VkPipelineBindPoint pipelineBindPoint,
1839 VkPipeline pipeline)
1840 {
1841 struct command_buffer_data *cmd_buffer_data = FIND_CMD_BUFFER_DATA(commandBuffer);
1842 switch (pipelineBindPoint) {
1843 case VK_PIPELINE_BIND_POINT_GRAPHICS: cmd_buffer_data->stats.stats[OVERLAY_PARAM_ENABLED_pipeline_graphics]++; break;
1844 case VK_PIPELINE_BIND_POINT_COMPUTE: cmd_buffer_data->stats.stats[OVERLAY_PARAM_ENABLED_pipeline_compute]++; break;
1845 case VK_PIPELINE_BIND_POINT_RAY_TRACING_NV: cmd_buffer_data->stats.stats[OVERLAY_PARAM_ENABLED_pipeline_raytracing]++; break;
1846 default: break;
1847 }
1848 struct device_data *device_data = cmd_buffer_data->device;
1849 device_data->vtable.CmdBindPipeline(commandBuffer, pipelineBindPoint, pipeline);
1850 }
1851
1852 static VkResult overlay_BeginCommandBuffer(
1853 VkCommandBuffer commandBuffer,
1854 const VkCommandBufferBeginInfo* pBeginInfo)
1855 {
1856 struct command_buffer_data *cmd_buffer_data = FIND_CMD_BUFFER_DATA(commandBuffer);
1857 struct device_data *device_data = cmd_buffer_data->device;
1858
1859 memset(&cmd_buffer_data->stats, 0, sizeof(cmd_buffer_data->stats));
1860
1861 /* We don't record any query in secondary command buffers, just make sure
1862 * we have the right inheritance.
1863 */
1864 if (cmd_buffer_data->level == VK_COMMAND_BUFFER_LEVEL_SECONDARY) {
1865 VkCommandBufferBeginInfo *begin_info = (VkCommandBufferBeginInfo *)
1866 clone_chain((const struct VkBaseInStructure *)pBeginInfo);
1867 VkCommandBufferInheritanceInfo *parent_inhe_info = (VkCommandBufferInheritanceInfo *)
1868 vk_find_struct(begin_info, COMMAND_BUFFER_INHERITANCE_INFO);
1869 VkCommandBufferInheritanceInfo inhe_info = {
1870 VK_STRUCTURE_TYPE_COMMAND_BUFFER_INHERITANCE_INFO,
1871 NULL,
1872 VK_NULL_HANDLE,
1873 0,
1874 VK_NULL_HANDLE,
1875 VK_FALSE,
1876 0,
1877 overlay_query_flags,
1878 };
1879
1880 if (parent_inhe_info)
1881 parent_inhe_info->pipelineStatistics = overlay_query_flags;
1882 else {
1883 inhe_info.pNext = begin_info->pNext;
1884 begin_info->pNext = &inhe_info;
1885 }
1886
1887 VkResult result = device_data->vtable.BeginCommandBuffer(commandBuffer, pBeginInfo);
1888
1889 if (!parent_inhe_info)
1890 begin_info->pNext = inhe_info.pNext;
1891
1892 free_chain((struct VkBaseOutStructure *)begin_info);
1893
1894 return result;
1895 }
1896
1897 /* Otherwise record a begin query as first command. */
1898 VkResult result = device_data->vtable.BeginCommandBuffer(commandBuffer, pBeginInfo);
1899
1900 if (result == VK_SUCCESS) {
1901 if (cmd_buffer_data->pipeline_query_pool) {
1902 device_data->vtable.CmdResetQueryPool(commandBuffer,
1903 cmd_buffer_data->pipeline_query_pool,
1904 cmd_buffer_data->query_index, 1);
1905 }
1906 if (cmd_buffer_data->timestamp_query_pool) {
1907 device_data->vtable.CmdResetQueryPool(commandBuffer,
1908 cmd_buffer_data->timestamp_query_pool,
1909 cmd_buffer_data->query_index * 2, 2);
1910 }
1911 if (cmd_buffer_data->pipeline_query_pool) {
1912 device_data->vtable.CmdBeginQuery(commandBuffer,
1913 cmd_buffer_data->pipeline_query_pool,
1914 cmd_buffer_data->query_index, 0);
1915 }
1916 if (cmd_buffer_data->timestamp_query_pool) {
1917 device_data->vtable.CmdWriteTimestamp(commandBuffer,
1918 VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT,
1919 cmd_buffer_data->timestamp_query_pool,
1920 cmd_buffer_data->query_index * 2);
1921 }
1922 }
1923
1924 return result;
1925 }
1926
1927 static VkResult overlay_EndCommandBuffer(
1928 VkCommandBuffer commandBuffer)
1929 {
1930 struct command_buffer_data *cmd_buffer_data = FIND_CMD_BUFFER_DATA(commandBuffer);
1931 struct device_data *device_data = cmd_buffer_data->device;
1932
1933 if (cmd_buffer_data->timestamp_query_pool) {
1934 device_data->vtable.CmdWriteTimestamp(commandBuffer,
1935 VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT,
1936 cmd_buffer_data->timestamp_query_pool,
1937 cmd_buffer_data->query_index * 2 + 1);
1938 }
1939 if (cmd_buffer_data->pipeline_query_pool) {
1940 device_data->vtable.CmdEndQuery(commandBuffer,
1941 cmd_buffer_data->pipeline_query_pool,
1942 cmd_buffer_data->query_index);
1943 }
1944
1945 return device_data->vtable.EndCommandBuffer(commandBuffer);
1946 }
1947
1948 static VkResult overlay_ResetCommandBuffer(
1949 VkCommandBuffer commandBuffer,
1950 VkCommandBufferResetFlags flags)
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 memset(&cmd_buffer_data->stats, 0, sizeof(cmd_buffer_data->stats));
1956
1957 return device_data->vtable.ResetCommandBuffer(commandBuffer, flags);
1958 }
1959
1960 static void overlay_CmdExecuteCommands(
1961 VkCommandBuffer commandBuffer,
1962 uint32_t commandBufferCount,
1963 const VkCommandBuffer* pCommandBuffers)
1964 {
1965 struct command_buffer_data *cmd_buffer_data = FIND_CMD_BUFFER_DATA(commandBuffer);
1966 struct device_data *device_data = cmd_buffer_data->device;
1967
1968 /* Add the stats of the executed command buffers to the primary one. */
1969 for (uint32_t c = 0; c < commandBufferCount; c++) {
1970 struct command_buffer_data *sec_cmd_buffer_data = FIND_CMD_BUFFER_DATA(pCommandBuffers[c]);
1971
1972 for (uint32_t s = 0; s < OVERLAY_PARAM_ENABLED_MAX; s++)
1973 cmd_buffer_data->stats.stats[s] += sec_cmd_buffer_data->stats.stats[s];
1974 }
1975
1976 device_data->vtable.CmdExecuteCommands(commandBuffer, commandBufferCount, pCommandBuffers);
1977 }
1978
1979 static VkResult overlay_AllocateCommandBuffers(
1980 VkDevice device,
1981 const VkCommandBufferAllocateInfo* pAllocateInfo,
1982 VkCommandBuffer* pCommandBuffers)
1983 {
1984 struct device_data *device_data = FIND_DEVICE_DATA(device);
1985 VkResult result =
1986 device_data->vtable.AllocateCommandBuffers(device, pAllocateInfo, pCommandBuffers);
1987 if (result != VK_SUCCESS)
1988 return result;
1989
1990 VkQueryPool pipeline_query_pool = VK_NULL_HANDLE;
1991 VkQueryPool timestamp_query_pool = VK_NULL_HANDLE;
1992 if (device_data->instance->pipeline_statistics_enabled &&
1993 pAllocateInfo->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY) {
1994 VkQueryPoolCreateInfo pool_info = {
1995 VK_STRUCTURE_TYPE_QUERY_POOL_CREATE_INFO,
1996 NULL,
1997 0,
1998 VK_QUERY_TYPE_PIPELINE_STATISTICS,
1999 pAllocateInfo->commandBufferCount,
2000 overlay_query_flags,
2001 };
2002 VK_CHECK(device_data->vtable.CreateQueryPool(device_data->device, &pool_info,
2003 NULL, &pipeline_query_pool));
2004 }
2005 if (device_data->instance->params.enabled[OVERLAY_PARAM_ENABLED_gpu_timing]) {
2006 VkQueryPoolCreateInfo pool_info = {
2007 VK_STRUCTURE_TYPE_QUERY_POOL_CREATE_INFO,
2008 NULL,
2009 0,
2010 VK_QUERY_TYPE_TIMESTAMP,
2011 pAllocateInfo->commandBufferCount * 2,
2012 0,
2013 };
2014 VK_CHECK(device_data->vtable.CreateQueryPool(device_data->device, &pool_info,
2015 NULL, &timestamp_query_pool));
2016 }
2017
2018 for (uint32_t i = 0; i < pAllocateInfo->commandBufferCount; i++) {
2019 new_command_buffer_data(pCommandBuffers[i], pAllocateInfo->level,
2020 pipeline_query_pool, timestamp_query_pool,
2021 i, device_data);
2022 }
2023
2024 if (pipeline_query_pool)
2025 map_object(HKEY(pipeline_query_pool), (void *)(uintptr_t) pAllocateInfo->commandBufferCount);
2026 if (timestamp_query_pool)
2027 map_object(HKEY(timestamp_query_pool), (void *)(uintptr_t) pAllocateInfo->commandBufferCount);
2028
2029 return result;
2030 }
2031
2032 static void overlay_FreeCommandBuffers(
2033 VkDevice device,
2034 VkCommandPool commandPool,
2035 uint32_t commandBufferCount,
2036 const VkCommandBuffer* pCommandBuffers)
2037 {
2038 struct device_data *device_data = FIND_DEVICE_DATA(device);
2039 for (uint32_t i = 0; i < commandBufferCount; i++) {
2040 struct command_buffer_data *cmd_buffer_data =
2041 FIND_CMD_BUFFER_DATA(pCommandBuffers[i]);
2042 /* It is legal to free a NULL command buffer*/
2043 if (!cmd_buffer_data)
2044 continue;
2045
2046 uint64_t count = (uintptr_t)find_object_data(HKEY(cmd_buffer_data->pipeline_query_pool));
2047 if (count == 1) {
2048 unmap_object(HKEY(cmd_buffer_data->pipeline_query_pool));
2049 device_data->vtable.DestroyQueryPool(device_data->device,
2050 cmd_buffer_data->pipeline_query_pool, NULL);
2051 } else if (count != 0) {
2052 map_object(HKEY(cmd_buffer_data->pipeline_query_pool), (void *)(uintptr_t)(count - 1));
2053 }
2054 count = (uintptr_t)find_object_data(HKEY(cmd_buffer_data->timestamp_query_pool));
2055 if (count == 1) {
2056 unmap_object(HKEY(cmd_buffer_data->timestamp_query_pool));
2057 device_data->vtable.DestroyQueryPool(device_data->device,
2058 cmd_buffer_data->timestamp_query_pool, NULL);
2059 } else if (count != 0) {
2060 map_object(HKEY(cmd_buffer_data->timestamp_query_pool), (void *)(uintptr_t)(count - 1));
2061 }
2062 destroy_command_buffer_data(cmd_buffer_data);
2063 }
2064
2065 device_data->vtable.FreeCommandBuffers(device, commandPool,
2066 commandBufferCount, pCommandBuffers);
2067 }
2068
2069 static VkResult overlay_QueueSubmit(
2070 VkQueue queue,
2071 uint32_t submitCount,
2072 const VkSubmitInfo* pSubmits,
2073 VkFence fence)
2074 {
2075 struct queue_data *queue_data = FIND_QUEUE_DATA(queue);
2076 struct device_data *device_data = queue_data->device;
2077
2078 device_data->frame_stats.stats[OVERLAY_PARAM_ENABLED_submit]++;
2079
2080 for (uint32_t s = 0; s < submitCount; s++) {
2081 for (uint32_t c = 0; c < pSubmits[s].commandBufferCount; c++) {
2082 struct command_buffer_data *cmd_buffer_data =
2083 FIND_CMD_BUFFER_DATA(pSubmits[s].pCommandBuffers[c]);
2084
2085 /* Merge the submitted command buffer stats into the device. */
2086 for (uint32_t st = 0; st < OVERLAY_PARAM_ENABLED_MAX; st++)
2087 device_data->frame_stats.stats[st] += cmd_buffer_data->stats.stats[st];
2088
2089 /* Attach the command buffer to the queue so we remember to read its
2090 * pipeline statistics & timestamps at QueuePresent().
2091 */
2092 if (!cmd_buffer_data->pipeline_query_pool &&
2093 !cmd_buffer_data->timestamp_query_pool)
2094 continue;
2095
2096 if (list_empty(&cmd_buffer_data->link)) {
2097 list_addtail(&cmd_buffer_data->link,
2098 &queue_data->running_command_buffer);
2099 } else {
2100 fprintf(stderr, "Command buffer submitted multiple times before present.\n"
2101 "This could lead to invalid data.\n");
2102 }
2103 }
2104 }
2105
2106 return device_data->vtable.QueueSubmit(queue, submitCount, pSubmits, fence);
2107 }
2108
2109 static VkResult overlay_CreateDevice(
2110 VkPhysicalDevice physicalDevice,
2111 const VkDeviceCreateInfo* pCreateInfo,
2112 const VkAllocationCallbacks* pAllocator,
2113 VkDevice* pDevice)
2114 {
2115 struct instance_data *instance_data = FIND_PHYSICAL_DEVICE_DATA(physicalDevice);
2116 VkLayerDeviceCreateInfo *chain_info =
2117 get_device_chain_info(pCreateInfo, VK_LAYER_LINK_INFO);
2118
2119 assert(chain_info->u.pLayerInfo);
2120 PFN_vkGetInstanceProcAddr fpGetInstanceProcAddr = chain_info->u.pLayerInfo->pfnNextGetInstanceProcAddr;
2121 PFN_vkGetDeviceProcAddr fpGetDeviceProcAddr = chain_info->u.pLayerInfo->pfnNextGetDeviceProcAddr;
2122 PFN_vkCreateDevice fpCreateDevice = (PFN_vkCreateDevice)fpGetInstanceProcAddr(NULL, "vkCreateDevice");
2123 if (fpCreateDevice == NULL) {
2124 return VK_ERROR_INITIALIZATION_FAILED;
2125 }
2126
2127 // Advance the link info for the next element on the chain
2128 chain_info->u.pLayerInfo = chain_info->u.pLayerInfo->pNext;
2129
2130 VkPhysicalDeviceFeatures device_features = {};
2131 VkDeviceCreateInfo device_info = *pCreateInfo;
2132
2133 if (pCreateInfo->pEnabledFeatures)
2134 device_features = *(pCreateInfo->pEnabledFeatures);
2135 if (instance_data->pipeline_statistics_enabled) {
2136 device_features.inheritedQueries = true;
2137 device_features.pipelineStatisticsQuery = true;
2138 }
2139 device_info.pEnabledFeatures = &device_features;
2140
2141
2142 VkResult result = fpCreateDevice(physicalDevice, &device_info, pAllocator, pDevice);
2143 if (result != VK_SUCCESS) return result;
2144
2145 struct device_data *device_data = new_device_data(*pDevice, instance_data);
2146 device_data->physical_device = physicalDevice;
2147 vk_load_device_commands(*pDevice, fpGetDeviceProcAddr, &device_data->vtable);
2148
2149 instance_data->vtable.GetPhysicalDeviceProperties(device_data->physical_device,
2150 &device_data->properties);
2151
2152 VkLayerDeviceCreateInfo *load_data_info =
2153 get_device_chain_info(pCreateInfo, VK_LOADER_DATA_CALLBACK);
2154 device_data->set_device_loader_data = load_data_info->u.pfnSetDeviceLoaderData;
2155
2156 device_map_queues(device_data, pCreateInfo);
2157
2158 return result;
2159 }
2160
2161 static void overlay_DestroyDevice(
2162 VkDevice device,
2163 const VkAllocationCallbacks* pAllocator)
2164 {
2165 struct device_data *device_data = FIND_DEVICE_DATA(device);
2166 device_unmap_queues(device_data);
2167 device_data->vtable.DestroyDevice(device, pAllocator);
2168 destroy_device_data(device_data);
2169 }
2170
2171 static VkResult overlay_CreateInstance(
2172 const VkInstanceCreateInfo* pCreateInfo,
2173 const VkAllocationCallbacks* pAllocator,
2174 VkInstance* pInstance)
2175 {
2176 VkLayerInstanceCreateInfo *chain_info =
2177 get_instance_chain_info(pCreateInfo, VK_LAYER_LINK_INFO);
2178
2179 assert(chain_info->u.pLayerInfo);
2180 PFN_vkGetInstanceProcAddr fpGetInstanceProcAddr =
2181 chain_info->u.pLayerInfo->pfnNextGetInstanceProcAddr;
2182 PFN_vkCreateInstance fpCreateInstance =
2183 (PFN_vkCreateInstance)fpGetInstanceProcAddr(NULL, "vkCreateInstance");
2184 if (fpCreateInstance == NULL) {
2185 return VK_ERROR_INITIALIZATION_FAILED;
2186 }
2187
2188 // Advance the link info for the next element on the chain
2189 chain_info->u.pLayerInfo = chain_info->u.pLayerInfo->pNext;
2190
2191 VkResult result = fpCreateInstance(pCreateInfo, pAllocator, pInstance);
2192 if (result != VK_SUCCESS) return result;
2193
2194 struct instance_data *instance_data = new_instance_data(*pInstance);
2195 vk_load_instance_commands(instance_data->instance,
2196 fpGetInstanceProcAddr,
2197 &instance_data->vtable);
2198 instance_data_map_physical_devices(instance_data, true);
2199
2200 parse_overlay_env(&instance_data->params, getenv("VK_LAYER_MESA_OVERLAY_CONFIG"));
2201
2202 for (int i = OVERLAY_PARAM_ENABLED_vertices;
2203 i <= OVERLAY_PARAM_ENABLED_compute_invocations; i++) {
2204 if (instance_data->params.enabled[i]) {
2205 instance_data->pipeline_statistics_enabled = true;
2206 break;
2207 }
2208 }
2209
2210 return result;
2211 }
2212
2213 static void overlay_DestroyInstance(
2214 VkInstance instance,
2215 const VkAllocationCallbacks* pAllocator)
2216 {
2217 struct instance_data *instance_data = FIND_INSTANCE_DATA(instance);
2218 instance_data_map_physical_devices(instance_data, false);
2219 instance_data->vtable.DestroyInstance(instance, pAllocator);
2220 destroy_instance_data(instance_data);
2221 }
2222
2223 static const struct {
2224 const char *name;
2225 void *ptr;
2226 } name_to_funcptr_map[] = {
2227 { "vkGetDeviceProcAddr", (void *) vkGetDeviceProcAddr },
2228 #define ADD_HOOK(fn) { "vk" # fn, (void *) overlay_ ## fn }
2229 ADD_HOOK(AllocateCommandBuffers),
2230 ADD_HOOK(FreeCommandBuffers),
2231 ADD_HOOK(ResetCommandBuffer),
2232 ADD_HOOK(BeginCommandBuffer),
2233 ADD_HOOK(EndCommandBuffer),
2234 ADD_HOOK(CmdExecuteCommands),
2235
2236 ADD_HOOK(CmdDraw),
2237 ADD_HOOK(CmdDrawIndexed),
2238 ADD_HOOK(CmdDrawIndirect),
2239 ADD_HOOK(CmdDrawIndexedIndirect),
2240 ADD_HOOK(CmdDispatch),
2241 ADD_HOOK(CmdDispatchIndirect),
2242 ADD_HOOK(CmdDrawIndirectCountKHR),
2243 ADD_HOOK(CmdDrawIndexedIndirectCountKHR),
2244
2245 ADD_HOOK(CmdBindPipeline),
2246
2247 ADD_HOOK(CreateSwapchainKHR),
2248 ADD_HOOK(QueuePresentKHR),
2249 ADD_HOOK(DestroySwapchainKHR),
2250 ADD_HOOK(AcquireNextImageKHR),
2251 ADD_HOOK(AcquireNextImage2KHR),
2252
2253 ADD_HOOK(QueueSubmit),
2254
2255 ADD_HOOK(CreateDevice),
2256 ADD_HOOK(DestroyDevice),
2257
2258 ADD_HOOK(CreateInstance),
2259 ADD_HOOK(DestroyInstance),
2260 #undef ADD_HOOK
2261 };
2262
2263 static void *find_ptr(const char *name)
2264 {
2265 for (uint32_t i = 0; i < ARRAY_SIZE(name_to_funcptr_map); i++) {
2266 if (strcmp(name, name_to_funcptr_map[i].name) == 0)
2267 return name_to_funcptr_map[i].ptr;
2268 }
2269
2270 return NULL;
2271 }
2272
2273 VK_LAYER_EXPORT VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vkGetDeviceProcAddr(VkDevice dev,
2274 const char *funcName)
2275 {
2276 void *ptr = find_ptr(funcName);
2277 if (ptr) return reinterpret_cast<PFN_vkVoidFunction>(ptr);
2278
2279 if (dev == NULL) return NULL;
2280
2281 struct device_data *device_data = FIND_DEVICE_DATA(dev);
2282 if (device_data->vtable.GetDeviceProcAddr == NULL) return NULL;
2283 return device_data->vtable.GetDeviceProcAddr(dev, funcName);
2284 }
2285
2286 VK_LAYER_EXPORT VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL vkGetInstanceProcAddr(VkInstance instance,
2287 const char *funcName)
2288 {
2289 void *ptr = find_ptr(funcName);
2290 if (ptr) return reinterpret_cast<PFN_vkVoidFunction>(ptr);
2291
2292 if (instance == NULL) return NULL;
2293
2294 struct instance_data *instance_data = FIND_INSTANCE_DATA(instance);
2295 if (instance_data->vtable.GetInstanceProcAddr == NULL) return NULL;
2296 return instance_data->vtable.GetInstanceProcAddr(instance, funcName);
2297 }