renderer_vulkan.cpp 15 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355
  1. // SPDX-FileCopyrightText: Copyright 2018 yuzu Emulator Project
  2. // SPDX-License-Identifier: GPL-2.0-or-later
  3. #include <algorithm>
  4. #include <array>
  5. #include <cstring>
  6. #include <memory>
  7. #include <optional>
  8. #include <string>
  9. #include <vector>
  10. #include <fmt/format.h>
  11. #include "common/logging/log.h"
  12. #include "common/scope_exit.h"
  13. #include "common/settings.h"
  14. #include "common/telemetry.h"
  15. #include "core/core_timing.h"
  16. #include "core/frontend/emu_window.h"
  17. #include "core/telemetry_session.h"
  18. #include "video_core/gpu.h"
  19. #include "video_core/renderer_vulkan/renderer_vulkan.h"
  20. #include "video_core/renderer_vulkan/vk_blit_screen.h"
  21. #include "video_core/renderer_vulkan/vk_rasterizer.h"
  22. #include "video_core/renderer_vulkan/vk_scheduler.h"
  23. #include "video_core/renderer_vulkan/vk_state_tracker.h"
  24. #include "video_core/renderer_vulkan/vk_swapchain.h"
  25. #include "video_core/vulkan_common/vulkan_debug_callback.h"
  26. #include "video_core/vulkan_common/vulkan_device.h"
  27. #include "video_core/vulkan_common/vulkan_instance.h"
  28. #include "video_core/vulkan_common/vulkan_library.h"
  29. #include "video_core/vulkan_common/vulkan_memory_allocator.h"
  30. #include "video_core/vulkan_common/vulkan_surface.h"
  31. #include "video_core/vulkan_common/vulkan_wrapper.h"
  32. namespace Vulkan {
  33. namespace {
  34. std::string GetReadableVersion(u32 version) {
  35. return fmt::format("{}.{}.{}", VK_VERSION_MAJOR(version), VK_VERSION_MINOR(version),
  36. VK_VERSION_PATCH(version));
  37. }
  38. std::string GetDriverVersion(const Device& device) {
  39. // Extracted from
  40. // https://github.com/SaschaWillems/vulkan.gpuinfo.org/blob/5dddea46ea1120b0df14eef8f15ff8e318e35462/functions.php#L308-L314
  41. const u32 version = device.GetDriverVersion();
  42. if (device.GetDriverID() == VK_DRIVER_ID_NVIDIA_PROPRIETARY) {
  43. const u32 major = (version >> 22) & 0x3ff;
  44. const u32 minor = (version >> 14) & 0x0ff;
  45. const u32 secondary = (version >> 6) & 0x0ff;
  46. const u32 tertiary = version & 0x003f;
  47. return fmt::format("{}.{}.{}.{}", major, minor, secondary, tertiary);
  48. }
  49. if (device.GetDriverID() == VK_DRIVER_ID_INTEL_PROPRIETARY_WINDOWS) {
  50. const u32 major = version >> 14;
  51. const u32 minor = version & 0x3fff;
  52. return fmt::format("{}.{}", major, minor);
  53. }
  54. return GetReadableVersion(version);
  55. }
  56. std::string BuildCommaSeparatedExtensions(std::vector<std::string> available_extensions) {
  57. std::sort(std::begin(available_extensions), std::end(available_extensions));
  58. static constexpr std::size_t AverageExtensionSize = 64;
  59. std::string separated_extensions;
  60. separated_extensions.reserve(available_extensions.size() * AverageExtensionSize);
  61. const auto end = std::end(available_extensions);
  62. for (auto extension = std::begin(available_extensions); extension != end; ++extension) {
  63. if (const bool is_last = extension + 1 == end; is_last) {
  64. separated_extensions += *extension;
  65. } else {
  66. separated_extensions += fmt::format("{},", *extension);
  67. }
  68. }
  69. return separated_extensions;
  70. }
  71. } // Anonymous namespace
  72. Device CreateDevice(const vk::Instance& instance, const vk::InstanceDispatch& dld,
  73. VkSurfaceKHR surface) {
  74. const std::vector<VkPhysicalDevice> devices = instance.EnumeratePhysicalDevices();
  75. const s32 device_index = Settings::values.vulkan_device.GetValue();
  76. if (device_index < 0 || device_index >= static_cast<s32>(devices.size())) {
  77. LOG_ERROR(Render_Vulkan, "Invalid device index {}!", device_index);
  78. throw vk::Exception(VK_ERROR_INITIALIZATION_FAILED);
  79. }
  80. const vk::PhysicalDevice physical_device(devices[device_index], dld);
  81. return Device(*instance, physical_device, surface, dld);
  82. }
  83. RendererVulkan::RendererVulkan(Core::TelemetrySession& telemetry_session_,
  84. Core::Frontend::EmuWindow& emu_window,
  85. Core::Memory::Memory& cpu_memory_, Tegra::GPU& gpu_,
  86. std::unique_ptr<Core::Frontend::GraphicsContext> context_) try
  87. : RendererBase(emu_window, std::move(context_)), telemetry_session(telemetry_session_),
  88. cpu_memory(cpu_memory_), gpu(gpu_), library(OpenLibrary()),
  89. instance(CreateInstance(library, dld, VK_API_VERSION_1_1, render_window.GetWindowInfo().type,
  90. Settings::values.renderer_debug.GetValue())),
  91. debug_callback(Settings::values.renderer_debug ? CreateDebugCallback(instance) : nullptr),
  92. surface(CreateSurface(instance, render_window)),
  93. device(CreateDevice(instance, dld, *surface)), memory_allocator(device, false),
  94. state_tracker(), scheduler(device, state_tracker),
  95. swapchain(*surface, device, scheduler, render_window.GetFramebufferLayout().width,
  96. render_window.GetFramebufferLayout().height, false),
  97. blit_screen(cpu_memory, render_window, device, memory_allocator, swapchain, scheduler,
  98. screen_info),
  99. rasterizer(render_window, gpu, cpu_memory, screen_info, device, memory_allocator,
  100. state_tracker, scheduler) {
  101. if (Settings::values.renderer_force_max_clock.GetValue() && device.ShouldBoostClocks()) {
  102. turbo_mode.emplace(instance, dld);
  103. }
  104. Report();
  105. } catch (const vk::Exception& exception) {
  106. LOG_ERROR(Render_Vulkan, "Vulkan initialization failed with error: {}", exception.what());
  107. throw std::runtime_error{fmt::format("Vulkan initialization error {}", exception.what())};
  108. }
  109. RendererVulkan::~RendererVulkan() {
  110. void(device.GetLogical().WaitIdle());
  111. }
  112. void RendererVulkan::SwapBuffers(const Tegra::FramebufferConfig* framebuffer) {
  113. if (!framebuffer) {
  114. return;
  115. }
  116. SCOPE_EXIT({ render_window.OnFrameDisplayed(); });
  117. if (!render_window.IsShown()) {
  118. return;
  119. }
  120. // Update screen info if the framebuffer size has changed.
  121. if (screen_info.width != framebuffer->width || screen_info.height != framebuffer->height) {
  122. screen_info.width = framebuffer->width;
  123. screen_info.height = framebuffer->height;
  124. }
  125. const VAddr framebuffer_addr = framebuffer->address + framebuffer->offset;
  126. const bool use_accelerated =
  127. rasterizer.AccelerateDisplay(*framebuffer, framebuffer_addr, framebuffer->stride);
  128. const bool is_srgb = use_accelerated && screen_info.is_srgb;
  129. RenderScreenshot(*framebuffer, use_accelerated);
  130. bool has_been_recreated = false;
  131. const auto recreate_swapchain = [&](u32 width, u32 height) {
  132. if (!has_been_recreated) {
  133. has_been_recreated = true;
  134. scheduler.Finish();
  135. }
  136. swapchain.Create(width, height, is_srgb);
  137. };
  138. const Layout::FramebufferLayout layout = render_window.GetFramebufferLayout();
  139. if (swapchain.NeedsRecreation(is_srgb) || swapchain.GetWidth() != layout.width ||
  140. swapchain.GetHeight() != layout.height) {
  141. recreate_swapchain(layout.width, layout.height);
  142. }
  143. bool is_outdated;
  144. do {
  145. swapchain.AcquireNextImage();
  146. is_outdated = swapchain.IsOutDated();
  147. if (is_outdated) {
  148. recreate_swapchain(layout.width, layout.height);
  149. }
  150. } while (is_outdated);
  151. if (has_been_recreated) {
  152. blit_screen.Recreate();
  153. }
  154. const VkSemaphore render_semaphore = blit_screen.DrawToSwapchain(*framebuffer, use_accelerated);
  155. const VkSemaphore present_semaphore = swapchain.CurrentPresentSemaphore();
  156. scheduler.Flush(render_semaphore, present_semaphore);
  157. scheduler.WaitWorker();
  158. swapchain.Present(render_semaphore);
  159. gpu.RendererFrameEndNotify();
  160. rasterizer.TickFrame();
  161. }
  162. void RendererVulkan::Report() const {
  163. using namespace Common::Literals;
  164. const std::string vendor_name{device.GetVendorName()};
  165. const std::string model_name{device.GetModelName()};
  166. const std::string driver_version = GetDriverVersion(device);
  167. const std::string driver_name = fmt::format("{} {}", vendor_name, driver_version);
  168. const std::string api_version = GetReadableVersion(device.ApiVersion());
  169. const std::string extensions = BuildCommaSeparatedExtensions(device.GetAvailableExtensions());
  170. const auto available_vram = static_cast<f64>(device.GetDeviceLocalMemory()) / f64{1_GiB};
  171. LOG_INFO(Render_Vulkan, "Driver: {}", driver_name);
  172. LOG_INFO(Render_Vulkan, "Device: {}", model_name);
  173. LOG_INFO(Render_Vulkan, "Vulkan: {}", api_version);
  174. LOG_INFO(Render_Vulkan, "Available VRAM: {:.2f} GiB", available_vram);
  175. static constexpr auto field = Common::Telemetry::FieldType::UserSystem;
  176. telemetry_session.AddField(field, "GPU_Vendor", vendor_name);
  177. telemetry_session.AddField(field, "GPU_Model", model_name);
  178. telemetry_session.AddField(field, "GPU_Vulkan_Driver", driver_name);
  179. telemetry_session.AddField(field, "GPU_Vulkan_Version", api_version);
  180. telemetry_session.AddField(field, "GPU_Vulkan_Extensions", extensions);
  181. }
  182. void Vulkan::RendererVulkan::RenderScreenshot(const Tegra::FramebufferConfig& framebuffer,
  183. bool use_accelerated) {
  184. if (!renderer_settings.screenshot_requested) {
  185. return;
  186. }
  187. const Layout::FramebufferLayout layout{renderer_settings.screenshot_framebuffer_layout};
  188. vk::Image staging_image = device.GetLogical().CreateImage(VkImageCreateInfo{
  189. .sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO,
  190. .pNext = nullptr,
  191. .flags = VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT,
  192. .imageType = VK_IMAGE_TYPE_2D,
  193. .format = VK_FORMAT_B8G8R8A8_UNORM,
  194. .extent =
  195. {
  196. .width = layout.width,
  197. .height = layout.height,
  198. .depth = 1,
  199. },
  200. .mipLevels = 1,
  201. .arrayLayers = 1,
  202. .samples = VK_SAMPLE_COUNT_1_BIT,
  203. .tiling = VK_IMAGE_TILING_OPTIMAL,
  204. .usage = VK_IMAGE_USAGE_TRANSFER_SRC_BIT | VK_IMAGE_USAGE_TRANSFER_DST_BIT |
  205. VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT,
  206. .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
  207. .queueFamilyIndexCount = 0,
  208. .pQueueFamilyIndices = nullptr,
  209. .initialLayout = VK_IMAGE_LAYOUT_UNDEFINED,
  210. });
  211. const auto image_commit = memory_allocator.Commit(staging_image, MemoryUsage::DeviceLocal);
  212. const vk::ImageView dst_view = device.GetLogical().CreateImageView(VkImageViewCreateInfo{
  213. .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
  214. .pNext = nullptr,
  215. .flags = 0,
  216. .image = *staging_image,
  217. .viewType = VK_IMAGE_VIEW_TYPE_2D,
  218. .format = screen_info.is_srgb ? VK_FORMAT_B8G8R8A8_SRGB : VK_FORMAT_B8G8R8A8_UNORM,
  219. .components{
  220. .r = VK_COMPONENT_SWIZZLE_IDENTITY,
  221. .g = VK_COMPONENT_SWIZZLE_IDENTITY,
  222. .b = VK_COMPONENT_SWIZZLE_IDENTITY,
  223. .a = VK_COMPONENT_SWIZZLE_IDENTITY,
  224. },
  225. .subresourceRange{
  226. .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
  227. .baseMipLevel = 0,
  228. .levelCount = 1,
  229. .baseArrayLayer = 0,
  230. .layerCount = VK_REMAINING_ARRAY_LAYERS,
  231. },
  232. });
  233. const VkExtent2D render_area{.width = layout.width, .height = layout.height};
  234. const vk::Framebuffer screenshot_fb = blit_screen.CreateFramebuffer(*dst_view, render_area);
  235. // Since we're not rendering to the screen, ignore the render semaphore.
  236. void(blit_screen.Draw(framebuffer, *screenshot_fb, layout, render_area, use_accelerated));
  237. const auto buffer_size = static_cast<VkDeviceSize>(layout.width * layout.height * 4);
  238. const VkBufferCreateInfo dst_buffer_info{
  239. .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
  240. .pNext = nullptr,
  241. .flags = 0,
  242. .size = buffer_size,
  243. .usage = VK_BUFFER_USAGE_TRANSFER_DST_BIT,
  244. .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
  245. .queueFamilyIndexCount = 0,
  246. .pQueueFamilyIndices = nullptr,
  247. };
  248. const vk::Buffer dst_buffer = device.GetLogical().CreateBuffer(dst_buffer_info);
  249. MemoryCommit dst_buffer_memory = memory_allocator.Commit(dst_buffer, MemoryUsage::Download);
  250. scheduler.RequestOutsideRenderPassOperationContext();
  251. scheduler.Record([&](vk::CommandBuffer cmdbuf) {
  252. const VkImageMemoryBarrier read_barrier{
  253. .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
  254. .pNext = nullptr,
  255. .srcAccessMask = VK_ACCESS_MEMORY_WRITE_BIT,
  256. .dstAccessMask = VK_ACCESS_TRANSFER_READ_BIT,
  257. .oldLayout = VK_IMAGE_LAYOUT_PRESENT_SRC_KHR,
  258. .newLayout = VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
  259. .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
  260. .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
  261. .image = *staging_image,
  262. .subresourceRange{
  263. .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
  264. .baseMipLevel = 0,
  265. .levelCount = VK_REMAINING_MIP_LEVELS,
  266. .baseArrayLayer = 0,
  267. .layerCount = VK_REMAINING_ARRAY_LAYERS,
  268. },
  269. };
  270. const VkImageMemoryBarrier image_write_barrier{
  271. .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
  272. .pNext = nullptr,
  273. .srcAccessMask = 0,
  274. .dstAccessMask = VK_ACCESS_MEMORY_WRITE_BIT,
  275. .oldLayout = VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
  276. .newLayout = VK_IMAGE_LAYOUT_GENERAL,
  277. .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
  278. .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
  279. .image = *staging_image,
  280. .subresourceRange{
  281. .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
  282. .baseMipLevel = 0,
  283. .levelCount = VK_REMAINING_MIP_LEVELS,
  284. .baseArrayLayer = 0,
  285. .layerCount = VK_REMAINING_ARRAY_LAYERS,
  286. },
  287. };
  288. static constexpr VkMemoryBarrier memory_write_barrier{
  289. .sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER,
  290. .pNext = nullptr,
  291. .srcAccessMask = VK_ACCESS_MEMORY_WRITE_BIT,
  292. .dstAccessMask = VK_ACCESS_MEMORY_READ_BIT | VK_ACCESS_MEMORY_WRITE_BIT,
  293. };
  294. const VkBufferImageCopy copy{
  295. .bufferOffset = 0,
  296. .bufferRowLength = 0,
  297. .bufferImageHeight = 0,
  298. .imageSubresource{
  299. .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
  300. .mipLevel = 0,
  301. .baseArrayLayer = 0,
  302. .layerCount = 1,
  303. },
  304. .imageOffset{.x = 0, .y = 0, .z = 0},
  305. .imageExtent{
  306. .width = layout.width,
  307. .height = layout.height,
  308. .depth = 1,
  309. },
  310. };
  311. cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT,
  312. 0, read_barrier);
  313. cmdbuf.CopyImageToBuffer(*staging_image, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, *dst_buffer,
  314. copy);
  315. cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
  316. 0, memory_write_barrier, nullptr, image_write_barrier);
  317. });
  318. // Ensure the copy is fully completed before saving the screenshot
  319. scheduler.Finish();
  320. // Copy backing image data to the QImage screenshot buffer
  321. const auto dst_memory_map = dst_buffer_memory.Map();
  322. std::memcpy(renderer_settings.screenshot_bits, dst_memory_map.data(), dst_memory_map.size());
  323. renderer_settings.screenshot_complete_callback(false);
  324. renderer_settings.screenshot_requested = false;
  325. }
  326. } // namespace Vulkan