vulkan_device.cpp 51 KB

1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991001011021031041051061071081091101111121131141151161171181191201211221231241251261271281291301311321331341351361371381391401411421431441451461471481491501511521531541551561571581591601611621631641651661671681691701711721731741751761771781791801811821831841851861871881891901911921931941951961971981992002012022032042052062072082092102112122132142152162172182192202212222232242252262272282292302312322332342352362372382392402412422432442452462472482492502512522532542552562572582592602612622632642652662672682692702712722732742752762772782792802812822832842852862872882892902912922932942952962972982993003013023033043053063073083093103113123133143153163173183193203213223233243253263273283293303313323333343353363373383393403413423433443453463473483493503513523533543553563573583593603613623633643653663673683693703713723733743753763773783793803813823833843853863873883893903913923933943953963973983994004014024034044054064074084094104114124134144154164174184194204214224234244254264274284294304314324334344354364374384394404414424434444454464474484494504514524534544554564574584594604614624634644654664674684694704714724734744754764774784794804814824834844854864874884894904914924934944954964974984995005015025035045055065075085095105115125135145155165175185195205215225235245255265275285295305315325335345355365375385395405415425435445455465475485495505515525535545555565575585595605615625635645655665675685695705715725735745755765775785795805815825835845855865875885895905915925935945955965975985996006016026036046056066076086096106116126136146156166176186196206216226236246256266276286296306316326336346356366376386396406416426436446456466476486496506516526536546556566576586596606616626636646656666676686696706716726736746756766776786796806816826836846856866876886896906916926936946956966976986997007017027037047057067077087097107117127137147157167177187197207217227237247257267277287297307317327337347357367377387397407417427437447457467477487497507517527537547557567577587597607617627637647657667677687697707717727737747757767777787797807817827837847857867877887897907917927937947957967977987998008018028038048058068078088098108118128138148158168178188198208218228238248258268278288298308318328338348358368378388398408418428438448458468478488498508518528538548558568578588598608618628638648658668678688698708718728738748758768778788798808818828838848858868878888898908918928938948958968978988999009019029039049059069079089099109119129139149159169179189199209219229239249259269279289299309319329339349359369379389399409419429439449459469479489499509519529539549559569579589599609619629639649659669679689699709719729739749759769779789799809819829839849859869879889899909919929939949959969979989991000100110021003100410051006100710081009101010111012101310141015101610171018101910201021102210231024102510261027102810291030103110321033103410351036103710381039104010411042104310441045104610471048104910501051105210531054105510561057105810591060106110621063106410651066106710681069107010711072107310741075107610771078107910801081108210831084108510861087108810891090109110921093109410951096109710981099110011011102110311041105110611071108110911101111111211131114111511161117111811191120112111221123112411251126112711281129113011311132113311341135113611371138113911401141114211431144114511461147114811491150115111521153115411551156115711581159116011611162116311641165116611671168116911701171117211731174117511761177117811791180118111821183118411851186118711881189119011911192119311941195119611971198119912001201120212031204120512061207120812091210121112121213121412151216121712181219122012211222122312241225122612271228122912301231123212331234123512361237123812391240124112421243124412451246
  1. // Copyright 2018 yuzu Emulator Project
  2. // Licensed under GPLv2 or any later version
  3. // Refer to the license.txt file included.
  4. #include <algorithm>
  5. #include <bitset>
  6. #include <chrono>
  7. #include <optional>
  8. #include <string_view>
  9. #include <thread>
  10. #include <unordered_set>
  11. #include <utility>
  12. #include <vector>
  13. #include "common/assert.h"
  14. #include "common/settings.h"
  15. #include "video_core/vulkan_common/nsight_aftermath_tracker.h"
  16. #include "video_core/vulkan_common/vulkan_device.h"
  17. #include "video_core/vulkan_common/vulkan_wrapper.h"
  18. namespace Vulkan {
  19. namespace {
  20. namespace Alternatives {
  21. constexpr std::array DEPTH24_UNORM_STENCIL8_UINT{
  22. VK_FORMAT_D32_SFLOAT_S8_UINT,
  23. VK_FORMAT_D16_UNORM_S8_UINT,
  24. VK_FORMAT_UNDEFINED,
  25. };
  26. constexpr std::array DEPTH16_UNORM_STENCIL8_UINT{
  27. VK_FORMAT_D24_UNORM_S8_UINT,
  28. VK_FORMAT_D32_SFLOAT_S8_UINT,
  29. VK_FORMAT_UNDEFINED,
  30. };
  31. } // namespace Alternatives
  32. enum class NvidiaArchitecture {
  33. AmpereOrNewer,
  34. Turing,
  35. VoltaOrOlder,
  36. };
  37. constexpr std::array REQUIRED_EXTENSIONS{
  38. VK_KHR_MAINTENANCE1_EXTENSION_NAME,
  39. VK_KHR_STORAGE_BUFFER_STORAGE_CLASS_EXTENSION_NAME,
  40. VK_KHR_SHADER_DRAW_PARAMETERS_EXTENSION_NAME,
  41. VK_KHR_16BIT_STORAGE_EXTENSION_NAME,
  42. VK_KHR_8BIT_STORAGE_EXTENSION_NAME,
  43. VK_KHR_DRIVER_PROPERTIES_EXTENSION_NAME,
  44. VK_KHR_DESCRIPTOR_UPDATE_TEMPLATE_EXTENSION_NAME,
  45. VK_KHR_TIMELINE_SEMAPHORE_EXTENSION_NAME,
  46. VK_KHR_SAMPLER_MIRROR_CLAMP_TO_EDGE_EXTENSION_NAME,
  47. VK_KHR_SHADER_FLOAT_CONTROLS_EXTENSION_NAME,
  48. VK_KHR_VARIABLE_POINTERS_EXTENSION_NAME,
  49. VK_EXT_VERTEX_ATTRIBUTE_DIVISOR_EXTENSION_NAME,
  50. VK_EXT_SHADER_SUBGROUP_BALLOT_EXTENSION_NAME,
  51. VK_EXT_SHADER_SUBGROUP_VOTE_EXTENSION_NAME,
  52. VK_EXT_ROBUSTNESS_2_EXTENSION_NAME,
  53. VK_EXT_HOST_QUERY_RESET_EXTENSION_NAME,
  54. VK_EXT_SHADER_DEMOTE_TO_HELPER_INVOCATION_EXTENSION_NAME,
  55. #ifdef _WIN32
  56. VK_KHR_EXTERNAL_MEMORY_WIN32_EXTENSION_NAME,
  57. #endif
  58. #ifdef __unix__
  59. VK_KHR_EXTERNAL_MEMORY_FD_EXTENSION_NAME,
  60. #endif
  61. };
  62. template <typename T>
  63. void SetNext(void**& next, T& data) {
  64. *next = &data;
  65. next = &data.pNext;
  66. }
  67. constexpr const VkFormat* GetFormatAlternatives(VkFormat format) {
  68. switch (format) {
  69. case VK_FORMAT_D24_UNORM_S8_UINT:
  70. return Alternatives::DEPTH24_UNORM_STENCIL8_UINT.data();
  71. case VK_FORMAT_D16_UNORM_S8_UINT:
  72. return Alternatives::DEPTH16_UNORM_STENCIL8_UINT.data();
  73. default:
  74. return nullptr;
  75. }
  76. }
  77. VkFormatFeatureFlags GetFormatFeatures(VkFormatProperties properties, FormatType format_type) {
  78. switch (format_type) {
  79. case FormatType::Linear:
  80. return properties.linearTilingFeatures;
  81. case FormatType::Optimal:
  82. return properties.optimalTilingFeatures;
  83. case FormatType::Buffer:
  84. return properties.bufferFeatures;
  85. default:
  86. return {};
  87. }
  88. }
  89. std::unordered_map<VkFormat, VkFormatProperties> GetFormatProperties(vk::PhysicalDevice physical) {
  90. static constexpr std::array formats{
  91. VK_FORMAT_A8B8G8R8_UNORM_PACK32,
  92. VK_FORMAT_A8B8G8R8_UINT_PACK32,
  93. VK_FORMAT_A8B8G8R8_SNORM_PACK32,
  94. VK_FORMAT_A8B8G8R8_SINT_PACK32,
  95. VK_FORMAT_A8B8G8R8_SRGB_PACK32,
  96. VK_FORMAT_B5G6R5_UNORM_PACK16,
  97. VK_FORMAT_A2B10G10R10_UNORM_PACK32,
  98. VK_FORMAT_A2B10G10R10_UINT_PACK32,
  99. VK_FORMAT_A1R5G5B5_UNORM_PACK16,
  100. VK_FORMAT_R32G32B32A32_SFLOAT,
  101. VK_FORMAT_R32G32B32A32_SINT,
  102. VK_FORMAT_R32G32B32A32_UINT,
  103. VK_FORMAT_R32G32_SFLOAT,
  104. VK_FORMAT_R32G32_SINT,
  105. VK_FORMAT_R32G32_UINT,
  106. VK_FORMAT_R16G16B16A16_SINT,
  107. VK_FORMAT_R16G16B16A16_UINT,
  108. VK_FORMAT_R16G16B16A16_SNORM,
  109. VK_FORMAT_R16G16B16A16_UNORM,
  110. VK_FORMAT_R16G16_UNORM,
  111. VK_FORMAT_R16G16_SNORM,
  112. VK_FORMAT_R16G16_SFLOAT,
  113. VK_FORMAT_R16G16_SINT,
  114. VK_FORMAT_R16_UNORM,
  115. VK_FORMAT_R16_SNORM,
  116. VK_FORMAT_R16_UINT,
  117. VK_FORMAT_R8G8B8A8_SRGB,
  118. VK_FORMAT_R8G8_UNORM,
  119. VK_FORMAT_R8G8_SNORM,
  120. VK_FORMAT_R8G8_SINT,
  121. VK_FORMAT_R8G8_UINT,
  122. VK_FORMAT_R8_UNORM,
  123. VK_FORMAT_R8_SNORM,
  124. VK_FORMAT_R8_SINT,
  125. VK_FORMAT_R8_UINT,
  126. VK_FORMAT_B10G11R11_UFLOAT_PACK32,
  127. VK_FORMAT_R32_SFLOAT,
  128. VK_FORMAT_R32_UINT,
  129. VK_FORMAT_R32_SINT,
  130. VK_FORMAT_R16_SFLOAT,
  131. VK_FORMAT_R16G16B16A16_SFLOAT,
  132. VK_FORMAT_B8G8R8A8_UNORM,
  133. VK_FORMAT_B8G8R8A8_SRGB,
  134. VK_FORMAT_R4G4B4A4_UNORM_PACK16,
  135. VK_FORMAT_D32_SFLOAT,
  136. VK_FORMAT_D16_UNORM,
  137. VK_FORMAT_D16_UNORM_S8_UINT,
  138. VK_FORMAT_D24_UNORM_S8_UINT,
  139. VK_FORMAT_D32_SFLOAT_S8_UINT,
  140. VK_FORMAT_BC1_RGBA_UNORM_BLOCK,
  141. VK_FORMAT_BC2_UNORM_BLOCK,
  142. VK_FORMAT_BC3_UNORM_BLOCK,
  143. VK_FORMAT_BC4_UNORM_BLOCK,
  144. VK_FORMAT_BC4_SNORM_BLOCK,
  145. VK_FORMAT_BC5_UNORM_BLOCK,
  146. VK_FORMAT_BC5_SNORM_BLOCK,
  147. VK_FORMAT_BC7_UNORM_BLOCK,
  148. VK_FORMAT_BC6H_UFLOAT_BLOCK,
  149. VK_FORMAT_BC6H_SFLOAT_BLOCK,
  150. VK_FORMAT_BC1_RGBA_SRGB_BLOCK,
  151. VK_FORMAT_BC2_SRGB_BLOCK,
  152. VK_FORMAT_BC3_SRGB_BLOCK,
  153. VK_FORMAT_BC7_SRGB_BLOCK,
  154. VK_FORMAT_ASTC_4x4_UNORM_BLOCK,
  155. VK_FORMAT_ASTC_4x4_SRGB_BLOCK,
  156. VK_FORMAT_ASTC_5x4_UNORM_BLOCK,
  157. VK_FORMAT_ASTC_5x4_SRGB_BLOCK,
  158. VK_FORMAT_ASTC_5x5_UNORM_BLOCK,
  159. VK_FORMAT_ASTC_5x5_SRGB_BLOCK,
  160. VK_FORMAT_ASTC_6x5_UNORM_BLOCK,
  161. VK_FORMAT_ASTC_6x5_SRGB_BLOCK,
  162. VK_FORMAT_ASTC_6x6_UNORM_BLOCK,
  163. VK_FORMAT_ASTC_6x6_SRGB_BLOCK,
  164. VK_FORMAT_ASTC_8x5_UNORM_BLOCK,
  165. VK_FORMAT_ASTC_8x5_SRGB_BLOCK,
  166. VK_FORMAT_ASTC_8x6_UNORM_BLOCK,
  167. VK_FORMAT_ASTC_8x6_SRGB_BLOCK,
  168. VK_FORMAT_ASTC_8x8_UNORM_BLOCK,
  169. VK_FORMAT_ASTC_8x8_SRGB_BLOCK,
  170. VK_FORMAT_ASTC_10x5_UNORM_BLOCK,
  171. VK_FORMAT_ASTC_10x5_SRGB_BLOCK,
  172. VK_FORMAT_ASTC_10x6_UNORM_BLOCK,
  173. VK_FORMAT_ASTC_10x6_SRGB_BLOCK,
  174. VK_FORMAT_ASTC_10x8_UNORM_BLOCK,
  175. VK_FORMAT_ASTC_10x8_SRGB_BLOCK,
  176. VK_FORMAT_ASTC_10x10_UNORM_BLOCK,
  177. VK_FORMAT_ASTC_10x10_SRGB_BLOCK,
  178. VK_FORMAT_ASTC_12x10_UNORM_BLOCK,
  179. VK_FORMAT_ASTC_12x10_SRGB_BLOCK,
  180. VK_FORMAT_ASTC_12x12_UNORM_BLOCK,
  181. VK_FORMAT_ASTC_12x12_SRGB_BLOCK,
  182. VK_FORMAT_ASTC_8x6_UNORM_BLOCK,
  183. VK_FORMAT_ASTC_8x6_SRGB_BLOCK,
  184. VK_FORMAT_ASTC_6x5_UNORM_BLOCK,
  185. VK_FORMAT_ASTC_6x5_SRGB_BLOCK,
  186. VK_FORMAT_E5B9G9R9_UFLOAT_PACK32,
  187. };
  188. std::unordered_map<VkFormat, VkFormatProperties> format_properties;
  189. for (const auto format : formats) {
  190. format_properties.emplace(format, physical.GetFormatProperties(format));
  191. }
  192. return format_properties;
  193. }
  194. std::vector<std::string> GetSupportedExtensions(vk::PhysicalDevice physical) {
  195. const std::vector extensions = physical.EnumerateDeviceExtensionProperties();
  196. std::vector<std::string> supported_extensions;
  197. supported_extensions.reserve(extensions.size());
  198. for (const auto& extension : extensions) {
  199. supported_extensions.emplace_back(extension.extensionName);
  200. }
  201. return supported_extensions;
  202. }
  203. NvidiaArchitecture GetNvidiaArchitecture(vk::PhysicalDevice physical,
  204. std::span<const std::string> exts) {
  205. if (std::ranges::find(exts, VK_KHR_FRAGMENT_SHADING_RATE_EXTENSION_NAME) != exts.end()) {
  206. VkPhysicalDeviceFragmentShadingRatePropertiesKHR shading_rate_props{};
  207. shading_rate_props.sType =
  208. VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FRAGMENT_SHADING_RATE_PROPERTIES_KHR;
  209. VkPhysicalDeviceProperties2KHR physical_properties{};
  210. physical_properties.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROPERTIES_2_KHR;
  211. physical_properties.pNext = &shading_rate_props;
  212. physical.GetProperties2KHR(physical_properties);
  213. if (shading_rate_props.primitiveFragmentShadingRateWithMultipleViewports) {
  214. // Only Ampere and newer support this feature
  215. return NvidiaArchitecture::AmpereOrNewer;
  216. }
  217. }
  218. if (std::ranges::find(exts, VK_NV_SHADING_RATE_IMAGE_EXTENSION_NAME) != exts.end()) {
  219. return NvidiaArchitecture::Turing;
  220. }
  221. return NvidiaArchitecture::VoltaOrOlder;
  222. }
  223. } // Anonymous namespace
  224. Device::Device(VkInstance instance_, vk::PhysicalDevice physical_, VkSurfaceKHR surface,
  225. const vk::InstanceDispatch& dld_)
  226. : instance{instance_}, dld{dld_}, physical{physical_}, properties{physical.GetProperties()},
  227. supported_extensions{GetSupportedExtensions(physical)},
  228. format_properties(GetFormatProperties(physical)) {
  229. CheckSuitability(surface != nullptr);
  230. SetupFamilies(surface);
  231. SetupFeatures();
  232. SetupProperties();
  233. const auto queue_cis = GetDeviceQueueCreateInfos();
  234. const std::vector extensions = LoadExtensions(surface != nullptr);
  235. VkPhysicalDeviceFeatures2 features2{
  236. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2,
  237. .pNext = nullptr,
  238. .features{
  239. .robustBufferAccess = true,
  240. .fullDrawIndexUint32 = false,
  241. .imageCubeArray = true,
  242. .independentBlend = true,
  243. .geometryShader = true,
  244. .tessellationShader = true,
  245. .sampleRateShading = true,
  246. .dualSrcBlend = true,
  247. .logicOp = false,
  248. .multiDrawIndirect = false,
  249. .drawIndirectFirstInstance = false,
  250. .depthClamp = true,
  251. .depthBiasClamp = true,
  252. .fillModeNonSolid = true,
  253. .depthBounds = is_depth_bounds_supported,
  254. .wideLines = true,
  255. .largePoints = true,
  256. .alphaToOne = false,
  257. .multiViewport = true,
  258. .samplerAnisotropy = true,
  259. .textureCompressionETC2 = false,
  260. .textureCompressionASTC_LDR = is_optimal_astc_supported,
  261. .textureCompressionBC = false,
  262. .occlusionQueryPrecise = true,
  263. .pipelineStatisticsQuery = false,
  264. .vertexPipelineStoresAndAtomics = true,
  265. .fragmentStoresAndAtomics = true,
  266. .shaderTessellationAndGeometryPointSize = false,
  267. .shaderImageGatherExtended = true,
  268. .shaderStorageImageExtendedFormats = false,
  269. .shaderStorageImageMultisample = is_shader_storage_image_multisample,
  270. .shaderStorageImageReadWithoutFormat = is_formatless_image_load_supported,
  271. .shaderStorageImageWriteWithoutFormat = true,
  272. .shaderUniformBufferArrayDynamicIndexing = false,
  273. .shaderSampledImageArrayDynamicIndexing = false,
  274. .shaderStorageBufferArrayDynamicIndexing = false,
  275. .shaderStorageImageArrayDynamicIndexing = false,
  276. .shaderClipDistance = true,
  277. .shaderCullDistance = true,
  278. .shaderFloat64 = is_shader_float64_supported,
  279. .shaderInt64 = is_shader_int64_supported,
  280. .shaderInt16 = is_shader_int16_supported,
  281. .shaderResourceResidency = false,
  282. .shaderResourceMinLod = false,
  283. .sparseBinding = false,
  284. .sparseResidencyBuffer = false,
  285. .sparseResidencyImage2D = false,
  286. .sparseResidencyImage3D = false,
  287. .sparseResidency2Samples = false,
  288. .sparseResidency4Samples = false,
  289. .sparseResidency8Samples = false,
  290. .sparseResidency16Samples = false,
  291. .sparseResidencyAliased = false,
  292. .variableMultisampleRate = false,
  293. .inheritedQueries = false,
  294. },
  295. };
  296. const void* first_next = &features2;
  297. void** next = &features2.pNext;
  298. VkPhysicalDeviceTimelineSemaphoreFeaturesKHR timeline_semaphore{
  299. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_FEATURES_KHR,
  300. .pNext = nullptr,
  301. .timelineSemaphore = true,
  302. };
  303. SetNext(next, timeline_semaphore);
  304. VkPhysicalDevice16BitStorageFeaturesKHR bit16_storage{
  305. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES_KHR,
  306. .pNext = nullptr,
  307. .storageBuffer16BitAccess = true,
  308. .uniformAndStorageBuffer16BitAccess = true,
  309. .storagePushConstant16 = false,
  310. .storageInputOutput16 = false,
  311. };
  312. SetNext(next, bit16_storage);
  313. VkPhysicalDevice8BitStorageFeaturesKHR bit8_storage{
  314. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES_KHR,
  315. .pNext = nullptr,
  316. .storageBuffer8BitAccess = false,
  317. .uniformAndStorageBuffer8BitAccess = true,
  318. .storagePushConstant8 = false,
  319. };
  320. SetNext(next, bit8_storage);
  321. VkPhysicalDeviceRobustness2FeaturesEXT robustness2{
  322. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ROBUSTNESS_2_FEATURES_EXT,
  323. .pNext = nullptr,
  324. .robustBufferAccess2 = true,
  325. .robustImageAccess2 = true,
  326. .nullDescriptor = true,
  327. };
  328. SetNext(next, robustness2);
  329. VkPhysicalDeviceHostQueryResetFeaturesEXT host_query_reset{
  330. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_QUERY_RESET_FEATURES_EXT,
  331. .pNext = nullptr,
  332. .hostQueryReset = true,
  333. };
  334. SetNext(next, host_query_reset);
  335. VkPhysicalDeviceVariablePointerFeaturesKHR variable_pointers{
  336. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES_KHR,
  337. .pNext = nullptr,
  338. .variablePointersStorageBuffer = VK_TRUE,
  339. .variablePointers = VK_TRUE,
  340. };
  341. SetNext(next, variable_pointers);
  342. VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT demote{
  343. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DEMOTE_TO_HELPER_INVOCATION_FEATURES_EXT,
  344. .pNext = nullptr,
  345. .shaderDemoteToHelperInvocation = true,
  346. };
  347. SetNext(next, demote);
  348. VkPhysicalDeviceFloat16Int8FeaturesKHR float16_int8;
  349. if (is_int8_supported || is_float16_supported) {
  350. float16_int8 = {
  351. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR,
  352. .pNext = nullptr,
  353. .shaderFloat16 = is_float16_supported,
  354. .shaderInt8 = is_int8_supported,
  355. };
  356. SetNext(next, float16_int8);
  357. }
  358. if (!is_float16_supported) {
  359. LOG_INFO(Render_Vulkan, "Device doesn't support float16 natively");
  360. }
  361. if (!is_int8_supported) {
  362. LOG_INFO(Render_Vulkan, "Device doesn't support int8 natively");
  363. }
  364. if (!nv_viewport_swizzle) {
  365. LOG_INFO(Render_Vulkan, "Device doesn't support viewport swizzles");
  366. }
  367. if (!nv_viewport_array2) {
  368. LOG_INFO(Render_Vulkan, "Device doesn't support viewport masks");
  369. }
  370. if (!nv_geometry_shader_passthrough) {
  371. LOG_INFO(Render_Vulkan, "Device doesn't support passthrough geometry shaders");
  372. }
  373. VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR std430_layout;
  374. if (khr_uniform_buffer_standard_layout) {
  375. std430_layout = {
  376. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_UNIFORM_BUFFER_STANDARD_LAYOUT_FEATURES_KHR,
  377. .pNext = nullptr,
  378. .uniformBufferStandardLayout = true,
  379. };
  380. SetNext(next, std430_layout);
  381. } else {
  382. LOG_INFO(Render_Vulkan, "Device doesn't support packed UBOs");
  383. }
  384. VkPhysicalDeviceIndexTypeUint8FeaturesEXT index_type_uint8;
  385. if (ext_index_type_uint8) {
  386. index_type_uint8 = {
  387. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INDEX_TYPE_UINT8_FEATURES_EXT,
  388. .pNext = nullptr,
  389. .indexTypeUint8 = true,
  390. };
  391. SetNext(next, index_type_uint8);
  392. } else {
  393. LOG_INFO(Render_Vulkan, "Device doesn't support uint8 indexes");
  394. }
  395. VkPhysicalDeviceTransformFeedbackFeaturesEXT transform_feedback;
  396. if (ext_transform_feedback) {
  397. transform_feedback = {
  398. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT,
  399. .pNext = nullptr,
  400. .transformFeedback = true,
  401. .geometryStreams = true,
  402. };
  403. SetNext(next, transform_feedback);
  404. } else {
  405. LOG_INFO(Render_Vulkan, "Device doesn't support transform feedbacks");
  406. }
  407. VkPhysicalDeviceCustomBorderColorFeaturesEXT custom_border;
  408. if (ext_custom_border_color) {
  409. custom_border = {
  410. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CUSTOM_BORDER_COLOR_FEATURES_EXT,
  411. .pNext = nullptr,
  412. .customBorderColors = VK_TRUE,
  413. .customBorderColorWithoutFormat = VK_TRUE,
  414. };
  415. SetNext(next, custom_border);
  416. } else {
  417. LOG_INFO(Render_Vulkan, "Device doesn't support custom border colors");
  418. }
  419. VkPhysicalDeviceExtendedDynamicStateFeaturesEXT dynamic_state;
  420. if (ext_extended_dynamic_state) {
  421. dynamic_state = {
  422. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTENDED_DYNAMIC_STATE_FEATURES_EXT,
  423. .pNext = nullptr,
  424. .extendedDynamicState = VK_TRUE,
  425. };
  426. SetNext(next, dynamic_state);
  427. } else {
  428. LOG_INFO(Render_Vulkan, "Device doesn't support extended dynamic state");
  429. }
  430. VkPhysicalDeviceLineRasterizationFeaturesEXT line_raster;
  431. if (ext_line_rasterization) {
  432. line_raster = {
  433. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_FEATURES_EXT,
  434. .pNext = nullptr,
  435. .rectangularLines = VK_TRUE,
  436. .bresenhamLines = VK_FALSE,
  437. .smoothLines = VK_TRUE,
  438. .stippledRectangularLines = VK_FALSE,
  439. .stippledBresenhamLines = VK_FALSE,
  440. .stippledSmoothLines = VK_FALSE,
  441. };
  442. SetNext(next, line_raster);
  443. } else {
  444. LOG_INFO(Render_Vulkan, "Device doesn't support smooth lines");
  445. }
  446. if (!ext_conservative_rasterization) {
  447. LOG_INFO(Render_Vulkan, "Device doesn't support conservative rasterization");
  448. }
  449. VkPhysicalDeviceProvokingVertexFeaturesEXT provoking_vertex;
  450. if (ext_provoking_vertex) {
  451. provoking_vertex = {
  452. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROVOKING_VERTEX_FEATURES_EXT,
  453. .pNext = nullptr,
  454. .provokingVertexLast = VK_TRUE,
  455. .transformFeedbackPreservesProvokingVertex = VK_TRUE,
  456. };
  457. SetNext(next, provoking_vertex);
  458. } else {
  459. LOG_INFO(Render_Vulkan, "Device doesn't support provoking vertex last");
  460. }
  461. VkPhysicalDeviceVertexInputDynamicStateFeaturesEXT vertex_input_dynamic;
  462. if (ext_vertex_input_dynamic_state) {
  463. vertex_input_dynamic = {
  464. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_INPUT_DYNAMIC_STATE_FEATURES_EXT,
  465. .pNext = nullptr,
  466. .vertexInputDynamicState = VK_TRUE,
  467. };
  468. SetNext(next, vertex_input_dynamic);
  469. } else {
  470. LOG_INFO(Render_Vulkan, "Device doesn't support vertex input dynamic state");
  471. }
  472. VkPhysicalDeviceShaderAtomicInt64FeaturesKHR atomic_int64;
  473. if (ext_shader_atomic_int64) {
  474. atomic_int64 = {
  475. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_INT64_FEATURES_KHR,
  476. .pNext = nullptr,
  477. .shaderBufferInt64Atomics = VK_TRUE,
  478. .shaderSharedInt64Atomics = VK_TRUE,
  479. };
  480. SetNext(next, atomic_int64);
  481. }
  482. VkPhysicalDeviceWorkgroupMemoryExplicitLayoutFeaturesKHR workgroup_layout;
  483. if (khr_workgroup_memory_explicit_layout) {
  484. workgroup_layout = {
  485. .sType =
  486. VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_WORKGROUP_MEMORY_EXPLICIT_LAYOUT_FEATURES_KHR,
  487. .pNext = nullptr,
  488. .workgroupMemoryExplicitLayout = VK_TRUE,
  489. .workgroupMemoryExplicitLayoutScalarBlockLayout = VK_TRUE,
  490. .workgroupMemoryExplicitLayout8BitAccess = VK_TRUE,
  491. .workgroupMemoryExplicitLayout16BitAccess = VK_TRUE,
  492. };
  493. SetNext(next, workgroup_layout);
  494. }
  495. VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR executable_properties;
  496. if (khr_pipeline_executable_properties) {
  497. LOG_INFO(Render_Vulkan, "Enabling shader feedback, expect slower shader build times");
  498. executable_properties = {
  499. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR,
  500. .pNext = nullptr,
  501. .pipelineExecutableInfo = VK_TRUE,
  502. };
  503. SetNext(next, executable_properties);
  504. }
  505. if (!ext_depth_range_unrestricted) {
  506. LOG_INFO(Render_Vulkan, "Device doesn't support depth range unrestricted");
  507. }
  508. VkDeviceDiagnosticsConfigCreateInfoNV diagnostics_nv;
  509. if (Settings::values.enable_nsight_aftermath && nv_device_diagnostics_config) {
  510. nsight_aftermath_tracker = std::make_unique<NsightAftermathTracker>();
  511. diagnostics_nv = {
  512. .sType = VK_STRUCTURE_TYPE_DEVICE_DIAGNOSTICS_CONFIG_CREATE_INFO_NV,
  513. .pNext = &features2,
  514. .flags = VK_DEVICE_DIAGNOSTICS_CONFIG_ENABLE_SHADER_DEBUG_INFO_BIT_NV |
  515. VK_DEVICE_DIAGNOSTICS_CONFIG_ENABLE_RESOURCE_TRACKING_BIT_NV |
  516. VK_DEVICE_DIAGNOSTICS_CONFIG_ENABLE_AUTOMATIC_CHECKPOINTS_BIT_NV,
  517. };
  518. first_next = &diagnostics_nv;
  519. }
  520. logical = vk::Device::Create(physical, queue_cis, extensions, first_next, dld);
  521. CollectPhysicalMemoryInfo();
  522. CollectTelemetryParameters();
  523. CollectToolingInfo();
  524. if (driver_id == VK_DRIVER_ID_NVIDIA_PROPRIETARY_KHR) {
  525. const auto arch = GetNvidiaArchitecture(physical, supported_extensions);
  526. switch (arch) {
  527. case NvidiaArchitecture::AmpereOrNewer:
  528. LOG_WARNING(Render_Vulkan, "Blacklisting Ampere devices from float16 math");
  529. is_float16_supported = false;
  530. break;
  531. case NvidiaArchitecture::Turing:
  532. break;
  533. case NvidiaArchitecture::VoltaOrOlder:
  534. LOG_WARNING(Render_Vulkan, "Blacklisting Volta and older from VK_KHR_push_descriptor");
  535. khr_push_descriptor = false;
  536. break;
  537. }
  538. }
  539. if (ext_extended_dynamic_state && driver_id == VK_DRIVER_ID_MESA_RADV) {
  540. // Mask driver version variant
  541. const u32 version = (properties.driverVersion << 3) >> 3;
  542. if (version < VK_MAKE_API_VERSION(0, 21, 2, 0)) {
  543. LOG_WARNING(Render_Vulkan,
  544. "RADV versions older than 21.2 have broken VK_EXT_extended_dynamic_state");
  545. ext_extended_dynamic_state = false;
  546. }
  547. }
  548. sets_per_pool = 64;
  549. if (driver_id == VK_DRIVER_ID_AMD_PROPRIETARY || driver_id == VK_DRIVER_ID_AMD_OPEN_SOURCE) {
  550. // AMD drivers need a higher amount of Sets per Pool in certain circunstances like in XC2.
  551. sets_per_pool = 96;
  552. }
  553. const bool is_amd = driver_id == VK_DRIVER_ID_AMD_PROPRIETARY ||
  554. driver_id == VK_DRIVER_ID_MESA_RADV ||
  555. driver_id == VK_DRIVER_ID_AMD_OPEN_SOURCE;
  556. if (ext_sampler_filter_minmax && is_amd) {
  557. // Disable ext_sampler_filter_minmax on AMD GCN4 and lower as it is broken.
  558. if (!is_float16_supported) {
  559. LOG_WARNING(
  560. Render_Vulkan,
  561. "Blacklisting AMD GCN4 and lower for VK_EXT_SAMPLER_FILTER_MINMAX_EXTENSION_NAME");
  562. ext_sampler_filter_minmax = false;
  563. }
  564. }
  565. if (ext_vertex_input_dynamic_state && driver_id == VK_DRIVER_ID_INTEL_PROPRIETARY_WINDOWS) {
  566. LOG_WARNING(Render_Vulkan, "Blacklisting Intel for VK_EXT_vertex_input_dynamic_state");
  567. ext_vertex_input_dynamic_state = false;
  568. }
  569. if (is_float16_supported && driver_id == VK_DRIVER_ID_INTEL_PROPRIETARY_WINDOWS) {
  570. // Intel's compiler crashes when using fp16 on Astral Chain, disable it for the time being.
  571. LOG_WARNING(Render_Vulkan, "Blacklisting Intel proprietary from float16 math");
  572. is_float16_supported = false;
  573. }
  574. graphics_queue = logical.GetQueue(graphics_family);
  575. present_queue = logical.GetQueue(present_family);
  576. }
  577. Device::~Device() = default;
  578. VkFormat Device::GetSupportedFormat(VkFormat wanted_format, VkFormatFeatureFlags wanted_usage,
  579. FormatType format_type) const {
  580. if (IsFormatSupported(wanted_format, wanted_usage, format_type)) {
  581. return wanted_format;
  582. }
  583. // The wanted format is not supported by hardware, search for alternatives
  584. const VkFormat* alternatives = GetFormatAlternatives(wanted_format);
  585. if (alternatives == nullptr) {
  586. UNREACHABLE_MSG("Format={} with usage={} and type={} has no defined alternatives and host "
  587. "hardware does not support it",
  588. wanted_format, wanted_usage, format_type);
  589. return wanted_format;
  590. }
  591. std::size_t i = 0;
  592. for (VkFormat alternative = *alternatives; alternative; alternative = alternatives[++i]) {
  593. if (!IsFormatSupported(alternative, wanted_usage, format_type)) {
  594. continue;
  595. }
  596. LOG_WARNING(Render_Vulkan,
  597. "Emulating format={} with alternative format={} with usage={} and type={}",
  598. wanted_format, alternative, wanted_usage, format_type);
  599. return alternative;
  600. }
  601. // No alternatives found, panic
  602. UNREACHABLE_MSG("Format={} with usage={} and type={} is not supported by the host hardware and "
  603. "doesn't support any of the alternatives",
  604. wanted_format, wanted_usage, format_type);
  605. return wanted_format;
  606. }
  607. void Device::ReportLoss() const {
  608. LOG_CRITICAL(Render_Vulkan, "Device loss occured!");
  609. // Wait for the log to flush and for Nsight Aftermath to dump the results
  610. std::this_thread::sleep_for(std::chrono::seconds{15});
  611. }
  612. void Device::SaveShader(std::span<const u32> spirv) const {
  613. if (nsight_aftermath_tracker) {
  614. nsight_aftermath_tracker->SaveShader(spirv);
  615. }
  616. }
  617. bool Device::IsOptimalAstcSupported(const VkPhysicalDeviceFeatures& features) const {
  618. // Disable for now to avoid converting ASTC twice.
  619. static constexpr std::array astc_formats = {
  620. VK_FORMAT_ASTC_4x4_UNORM_BLOCK, VK_FORMAT_ASTC_4x4_SRGB_BLOCK,
  621. VK_FORMAT_ASTC_5x4_UNORM_BLOCK, VK_FORMAT_ASTC_5x4_SRGB_BLOCK,
  622. VK_FORMAT_ASTC_5x5_UNORM_BLOCK, VK_FORMAT_ASTC_5x5_SRGB_BLOCK,
  623. VK_FORMAT_ASTC_6x5_UNORM_BLOCK, VK_FORMAT_ASTC_6x5_SRGB_BLOCK,
  624. VK_FORMAT_ASTC_6x6_UNORM_BLOCK, VK_FORMAT_ASTC_6x6_SRGB_BLOCK,
  625. VK_FORMAT_ASTC_8x5_UNORM_BLOCK, VK_FORMAT_ASTC_8x5_SRGB_BLOCK,
  626. VK_FORMAT_ASTC_8x6_UNORM_BLOCK, VK_FORMAT_ASTC_8x6_SRGB_BLOCK,
  627. VK_FORMAT_ASTC_8x8_UNORM_BLOCK, VK_FORMAT_ASTC_8x8_SRGB_BLOCK,
  628. VK_FORMAT_ASTC_10x5_UNORM_BLOCK, VK_FORMAT_ASTC_10x5_SRGB_BLOCK,
  629. VK_FORMAT_ASTC_10x6_UNORM_BLOCK, VK_FORMAT_ASTC_10x6_SRGB_BLOCK,
  630. VK_FORMAT_ASTC_10x8_UNORM_BLOCK, VK_FORMAT_ASTC_10x8_SRGB_BLOCK,
  631. VK_FORMAT_ASTC_10x10_UNORM_BLOCK, VK_FORMAT_ASTC_10x10_SRGB_BLOCK,
  632. VK_FORMAT_ASTC_12x10_UNORM_BLOCK, VK_FORMAT_ASTC_12x10_SRGB_BLOCK,
  633. VK_FORMAT_ASTC_12x12_UNORM_BLOCK, VK_FORMAT_ASTC_12x12_SRGB_BLOCK,
  634. };
  635. if (!features.textureCompressionASTC_LDR) {
  636. return false;
  637. }
  638. const auto format_feature_usage{
  639. VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT | VK_FORMAT_FEATURE_BLIT_SRC_BIT |
  640. VK_FORMAT_FEATURE_BLIT_DST_BIT | VK_FORMAT_FEATURE_TRANSFER_SRC_BIT |
  641. VK_FORMAT_FEATURE_TRANSFER_DST_BIT};
  642. for (const auto format : astc_formats) {
  643. const auto physical_format_properties{physical.GetFormatProperties(format)};
  644. if ((physical_format_properties.optimalTilingFeatures & format_feature_usage) == 0) {
  645. return false;
  646. }
  647. }
  648. return true;
  649. }
  650. bool Device::TestDepthStencilBlits() const {
  651. static constexpr VkFormatFeatureFlags required_features =
  652. VK_FORMAT_FEATURE_BLIT_SRC_BIT | VK_FORMAT_FEATURE_BLIT_DST_BIT;
  653. const auto test_features = [](VkFormatProperties props) {
  654. return (props.optimalTilingFeatures & required_features) == required_features;
  655. };
  656. return test_features(format_properties.at(VK_FORMAT_D32_SFLOAT_S8_UINT)) &&
  657. test_features(format_properties.at(VK_FORMAT_D24_UNORM_S8_UINT));
  658. }
  659. bool Device::IsFormatSupported(VkFormat wanted_format, VkFormatFeatureFlags wanted_usage,
  660. FormatType format_type) const {
  661. const auto it = format_properties.find(wanted_format);
  662. if (it == format_properties.end()) {
  663. UNIMPLEMENTED_MSG("Unimplemented format query={}", wanted_format);
  664. return true;
  665. }
  666. const auto supported_usage = GetFormatFeatures(it->second, format_type);
  667. return (supported_usage & wanted_usage) == wanted_usage;
  668. }
  669. std::string Device::GetDriverName() const {
  670. switch (driver_id) {
  671. case VK_DRIVER_ID_AMD_PROPRIETARY:
  672. return "AMD";
  673. case VK_DRIVER_ID_AMD_OPEN_SOURCE:
  674. return "AMDVLK";
  675. case VK_DRIVER_ID_MESA_RADV:
  676. return "RADV";
  677. case VK_DRIVER_ID_NVIDIA_PROPRIETARY:
  678. return "NVIDIA";
  679. case VK_DRIVER_ID_INTEL_PROPRIETARY_WINDOWS:
  680. return "INTEL";
  681. case VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA:
  682. return "ANV";
  683. case VK_DRIVER_ID_MESA_LLVMPIPE:
  684. return "LAVAPIPE";
  685. default:
  686. return vendor_name;
  687. }
  688. }
  689. void Device::CheckSuitability(bool requires_swapchain) const {
  690. std::bitset<REQUIRED_EXTENSIONS.size()> available_extensions;
  691. bool has_swapchain = false;
  692. for (const VkExtensionProperties& property : physical.EnumerateDeviceExtensionProperties()) {
  693. const std::string_view name{property.extensionName};
  694. for (size_t i = 0; i < REQUIRED_EXTENSIONS.size(); ++i) {
  695. if (available_extensions[i]) {
  696. continue;
  697. }
  698. available_extensions[i] = name == REQUIRED_EXTENSIONS[i];
  699. }
  700. has_swapchain = has_swapchain || name == VK_KHR_SWAPCHAIN_EXTENSION_NAME;
  701. }
  702. for (size_t i = 0; i < REQUIRED_EXTENSIONS.size(); ++i) {
  703. if (available_extensions[i]) {
  704. continue;
  705. }
  706. LOG_ERROR(Render_Vulkan, "Missing required extension: {}", REQUIRED_EXTENSIONS[i]);
  707. throw vk::Exception(VK_ERROR_EXTENSION_NOT_PRESENT);
  708. }
  709. if (requires_swapchain && !has_swapchain) {
  710. LOG_ERROR(Render_Vulkan, "Missing required extension: VK_KHR_swapchain");
  711. throw vk::Exception(VK_ERROR_EXTENSION_NOT_PRESENT);
  712. }
  713. struct LimitTuple {
  714. u32 minimum;
  715. u32 value;
  716. const char* name;
  717. };
  718. const VkPhysicalDeviceLimits& limits{properties.limits};
  719. const std::array limits_report{
  720. LimitTuple{65536, limits.maxUniformBufferRange, "maxUniformBufferRange"},
  721. LimitTuple{16, limits.maxViewports, "maxViewports"},
  722. LimitTuple{8, limits.maxColorAttachments, "maxColorAttachments"},
  723. LimitTuple{8, limits.maxClipDistances, "maxClipDistances"},
  724. };
  725. for (const auto& tuple : limits_report) {
  726. if (tuple.value < tuple.minimum) {
  727. LOG_ERROR(Render_Vulkan, "{} has to be {} or greater but it is {}", tuple.name,
  728. tuple.minimum, tuple.value);
  729. throw vk::Exception(VK_ERROR_FEATURE_NOT_PRESENT);
  730. }
  731. }
  732. VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT demote{};
  733. demote.sType =
  734. VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DEMOTE_TO_HELPER_INVOCATION_FEATURES_EXT;
  735. demote.pNext = nullptr;
  736. VkPhysicalDeviceVariablePointerFeaturesKHR variable_pointers{};
  737. variable_pointers.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES_KHR;
  738. variable_pointers.pNext = &demote;
  739. VkPhysicalDeviceRobustness2FeaturesEXT robustness2{};
  740. robustness2.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ROBUSTNESS_2_FEATURES_EXT;
  741. robustness2.pNext = &variable_pointers;
  742. VkPhysicalDeviceFeatures2KHR features2{};
  743. features2.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2;
  744. features2.pNext = &robustness2;
  745. physical.GetFeatures2KHR(features2);
  746. const VkPhysicalDeviceFeatures& features{features2.features};
  747. const std::array feature_report{
  748. std::make_pair(features.robustBufferAccess, "robustBufferAccess"),
  749. std::make_pair(features.vertexPipelineStoresAndAtomics, "vertexPipelineStoresAndAtomics"),
  750. std::make_pair(features.imageCubeArray, "imageCubeArray"),
  751. std::make_pair(features.independentBlend, "independentBlend"),
  752. std::make_pair(features.depthClamp, "depthClamp"),
  753. std::make_pair(features.samplerAnisotropy, "samplerAnisotropy"),
  754. std::make_pair(features.largePoints, "largePoints"),
  755. std::make_pair(features.multiViewport, "multiViewport"),
  756. std::make_pair(features.depthBiasClamp, "depthBiasClamp"),
  757. std::make_pair(features.fillModeNonSolid, "fillModeNonSolid"),
  758. std::make_pair(features.wideLines, "wideLines"),
  759. std::make_pair(features.geometryShader, "geometryShader"),
  760. std::make_pair(features.tessellationShader, "tessellationShader"),
  761. std::make_pair(features.sampleRateShading, "sampleRateShading"),
  762. std::make_pair(features.dualSrcBlend, "dualSrcBlend"),
  763. std::make_pair(features.occlusionQueryPrecise, "occlusionQueryPrecise"),
  764. std::make_pair(features.fragmentStoresAndAtomics, "fragmentStoresAndAtomics"),
  765. std::make_pair(features.shaderImageGatherExtended, "shaderImageGatherExtended"),
  766. std::make_pair(features.shaderStorageImageWriteWithoutFormat,
  767. "shaderStorageImageWriteWithoutFormat"),
  768. std::make_pair(features.shaderClipDistance, "shaderClipDistance"),
  769. std::make_pair(features.shaderCullDistance, "shaderCullDistance"),
  770. std::make_pair(demote.shaderDemoteToHelperInvocation, "shaderDemoteToHelperInvocation"),
  771. std::make_pair(variable_pointers.variablePointers, "variablePointers"),
  772. std::make_pair(variable_pointers.variablePointersStorageBuffer,
  773. "variablePointersStorageBuffer"),
  774. std::make_pair(robustness2.robustBufferAccess2, "robustBufferAccess2"),
  775. std::make_pair(robustness2.robustImageAccess2, "robustImageAccess2"),
  776. std::make_pair(robustness2.nullDescriptor, "nullDescriptor"),
  777. };
  778. for (const auto& [is_supported, name] : feature_report) {
  779. if (is_supported) {
  780. continue;
  781. }
  782. LOG_ERROR(Render_Vulkan, "Missing required feature: {}", name);
  783. throw vk::Exception(VK_ERROR_FEATURE_NOT_PRESENT);
  784. }
  785. }
  786. std::vector<const char*> Device::LoadExtensions(bool requires_surface) {
  787. std::vector<const char*> extensions;
  788. extensions.reserve(8 + REQUIRED_EXTENSIONS.size());
  789. extensions.insert(extensions.begin(), REQUIRED_EXTENSIONS.begin(), REQUIRED_EXTENSIONS.end());
  790. if (requires_surface) {
  791. extensions.push_back(VK_KHR_SWAPCHAIN_EXTENSION_NAME);
  792. }
  793. bool has_khr_shader_float16_int8{};
  794. bool has_khr_workgroup_memory_explicit_layout{};
  795. bool has_khr_pipeline_executable_properties{};
  796. bool has_khr_image_format_list{};
  797. bool has_khr_swapchain_mutable_format{};
  798. bool has_ext_subgroup_size_control{};
  799. bool has_ext_transform_feedback{};
  800. bool has_ext_custom_border_color{};
  801. bool has_ext_extended_dynamic_state{};
  802. bool has_ext_shader_atomic_int64{};
  803. bool has_ext_provoking_vertex{};
  804. bool has_ext_vertex_input_dynamic_state{};
  805. bool has_ext_line_rasterization{};
  806. for (const std::string& extension : supported_extensions) {
  807. const auto test = [&](std::optional<std::reference_wrapper<bool>> status, const char* name,
  808. bool push) {
  809. if (extension != name) {
  810. return;
  811. }
  812. if (push) {
  813. extensions.push_back(name);
  814. }
  815. if (status) {
  816. status->get() = true;
  817. }
  818. };
  819. test(nv_viewport_swizzle, VK_NV_VIEWPORT_SWIZZLE_EXTENSION_NAME, true);
  820. test(nv_viewport_array2, VK_NV_VIEWPORT_ARRAY2_EXTENSION_NAME, true);
  821. test(nv_geometry_shader_passthrough, VK_NV_GEOMETRY_SHADER_PASSTHROUGH_EXTENSION_NAME,
  822. true);
  823. test(khr_uniform_buffer_standard_layout,
  824. VK_KHR_UNIFORM_BUFFER_STANDARD_LAYOUT_EXTENSION_NAME, true);
  825. test(khr_spirv_1_4, VK_KHR_SPIRV_1_4_EXTENSION_NAME, true);
  826. test(khr_push_descriptor, VK_KHR_PUSH_DESCRIPTOR_EXTENSION_NAME, true);
  827. test(has_khr_shader_float16_int8, VK_KHR_SHADER_FLOAT16_INT8_EXTENSION_NAME, false);
  828. test(ext_depth_range_unrestricted, VK_EXT_DEPTH_RANGE_UNRESTRICTED_EXTENSION_NAME, true);
  829. test(ext_index_type_uint8, VK_EXT_INDEX_TYPE_UINT8_EXTENSION_NAME, true);
  830. test(ext_sampler_filter_minmax, VK_EXT_SAMPLER_FILTER_MINMAX_EXTENSION_NAME, true);
  831. test(ext_shader_viewport_index_layer, VK_EXT_SHADER_VIEWPORT_INDEX_LAYER_EXTENSION_NAME,
  832. true);
  833. test(ext_tooling_info, VK_EXT_TOOLING_INFO_EXTENSION_NAME, true);
  834. test(ext_shader_stencil_export, VK_EXT_SHADER_STENCIL_EXPORT_EXTENSION_NAME, true);
  835. test(ext_conservative_rasterization, VK_EXT_CONSERVATIVE_RASTERIZATION_EXTENSION_NAME,
  836. true);
  837. test(has_ext_transform_feedback, VK_EXT_TRANSFORM_FEEDBACK_EXTENSION_NAME, false);
  838. test(has_ext_custom_border_color, VK_EXT_CUSTOM_BORDER_COLOR_EXTENSION_NAME, false);
  839. test(has_ext_extended_dynamic_state, VK_EXT_EXTENDED_DYNAMIC_STATE_EXTENSION_NAME, false);
  840. test(has_ext_subgroup_size_control, VK_EXT_SUBGROUP_SIZE_CONTROL_EXTENSION_NAME, false);
  841. test(has_ext_provoking_vertex, VK_EXT_PROVOKING_VERTEX_EXTENSION_NAME, false);
  842. test(has_ext_vertex_input_dynamic_state, VK_EXT_VERTEX_INPUT_DYNAMIC_STATE_EXTENSION_NAME,
  843. false);
  844. test(has_ext_shader_atomic_int64, VK_KHR_SHADER_ATOMIC_INT64_EXTENSION_NAME, false);
  845. test(has_khr_workgroup_memory_explicit_layout,
  846. VK_KHR_WORKGROUP_MEMORY_EXPLICIT_LAYOUT_EXTENSION_NAME, false);
  847. test(has_khr_image_format_list, VK_KHR_IMAGE_FORMAT_LIST_EXTENSION_NAME, false);
  848. test(has_khr_swapchain_mutable_format, VK_KHR_SWAPCHAIN_MUTABLE_FORMAT_EXTENSION_NAME,
  849. false);
  850. test(has_ext_line_rasterization, VK_EXT_LINE_RASTERIZATION_EXTENSION_NAME, false);
  851. if (Settings::values.enable_nsight_aftermath) {
  852. test(nv_device_diagnostics_config, VK_NV_DEVICE_DIAGNOSTICS_CONFIG_EXTENSION_NAME,
  853. true);
  854. }
  855. if (Settings::values.renderer_shader_feedback) {
  856. test(has_khr_pipeline_executable_properties,
  857. VK_KHR_PIPELINE_EXECUTABLE_PROPERTIES_EXTENSION_NAME, false);
  858. }
  859. }
  860. VkPhysicalDeviceFeatures2KHR features{};
  861. features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2_KHR;
  862. VkPhysicalDeviceProperties2KHR physical_properties;
  863. physical_properties.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROPERTIES_2_KHR;
  864. if (has_khr_shader_float16_int8) {
  865. VkPhysicalDeviceFloat16Int8FeaturesKHR float16_int8_features;
  866. float16_int8_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR;
  867. float16_int8_features.pNext = nullptr;
  868. features.pNext = &float16_int8_features;
  869. physical.GetFeatures2KHR(features);
  870. is_float16_supported = float16_int8_features.shaderFloat16;
  871. is_int8_supported = float16_int8_features.shaderInt8;
  872. extensions.push_back(VK_KHR_SHADER_FLOAT16_INT8_EXTENSION_NAME);
  873. }
  874. if (has_ext_subgroup_size_control) {
  875. VkPhysicalDeviceSubgroupSizeControlFeaturesEXT subgroup_features;
  876. subgroup_features.sType =
  877. VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_FEATURES_EXT;
  878. subgroup_features.pNext = nullptr;
  879. features.pNext = &subgroup_features;
  880. physical.GetFeatures2KHR(features);
  881. VkPhysicalDeviceSubgroupSizeControlPropertiesEXT subgroup_properties;
  882. subgroup_properties.sType =
  883. VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_PROPERTIES_EXT;
  884. subgroup_properties.pNext = nullptr;
  885. physical_properties.pNext = &subgroup_properties;
  886. physical.GetProperties2KHR(physical_properties);
  887. is_warp_potentially_bigger = subgroup_properties.maxSubgroupSize > GuestWarpSize;
  888. if (subgroup_features.subgroupSizeControl &&
  889. subgroup_properties.minSubgroupSize <= GuestWarpSize &&
  890. subgroup_properties.maxSubgroupSize >= GuestWarpSize) {
  891. extensions.push_back(VK_EXT_SUBGROUP_SIZE_CONTROL_EXTENSION_NAME);
  892. guest_warp_stages = subgroup_properties.requiredSubgroupSizeStages;
  893. ext_subgroup_size_control = true;
  894. }
  895. } else {
  896. is_warp_potentially_bigger = true;
  897. }
  898. if (has_ext_provoking_vertex) {
  899. VkPhysicalDeviceProvokingVertexFeaturesEXT provoking_vertex;
  900. provoking_vertex.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROVOKING_VERTEX_FEATURES_EXT;
  901. provoking_vertex.pNext = nullptr;
  902. features.pNext = &provoking_vertex;
  903. physical.GetFeatures2KHR(features);
  904. if (provoking_vertex.provokingVertexLast &&
  905. provoking_vertex.transformFeedbackPreservesProvokingVertex) {
  906. extensions.push_back(VK_EXT_PROVOKING_VERTEX_EXTENSION_NAME);
  907. ext_provoking_vertex = true;
  908. }
  909. }
  910. if (has_ext_vertex_input_dynamic_state) {
  911. VkPhysicalDeviceVertexInputDynamicStateFeaturesEXT vertex_input;
  912. vertex_input.sType =
  913. VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_INPUT_DYNAMIC_STATE_FEATURES_EXT;
  914. vertex_input.pNext = nullptr;
  915. features.pNext = &vertex_input;
  916. physical.GetFeatures2KHR(features);
  917. if (vertex_input.vertexInputDynamicState) {
  918. extensions.push_back(VK_EXT_VERTEX_INPUT_DYNAMIC_STATE_EXTENSION_NAME);
  919. ext_vertex_input_dynamic_state = true;
  920. }
  921. }
  922. if (has_ext_shader_atomic_int64) {
  923. VkPhysicalDeviceShaderAtomicInt64Features atomic_int64;
  924. atomic_int64.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTENDED_DYNAMIC_STATE_FEATURES_EXT;
  925. atomic_int64.pNext = nullptr;
  926. features.pNext = &atomic_int64;
  927. physical.GetFeatures2KHR(features);
  928. if (atomic_int64.shaderBufferInt64Atomics && atomic_int64.shaderSharedInt64Atomics) {
  929. extensions.push_back(VK_KHR_SHADER_ATOMIC_INT64_EXTENSION_NAME);
  930. ext_shader_atomic_int64 = true;
  931. }
  932. }
  933. if (has_ext_transform_feedback) {
  934. VkPhysicalDeviceTransformFeedbackFeaturesEXT tfb_features;
  935. tfb_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT;
  936. tfb_features.pNext = nullptr;
  937. features.pNext = &tfb_features;
  938. physical.GetFeatures2KHR(features);
  939. VkPhysicalDeviceTransformFeedbackPropertiesEXT tfb_properties;
  940. tfb_properties.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT;
  941. tfb_properties.pNext = nullptr;
  942. physical_properties.pNext = &tfb_properties;
  943. physical.GetProperties2KHR(physical_properties);
  944. if (tfb_features.transformFeedback && tfb_features.geometryStreams &&
  945. tfb_properties.maxTransformFeedbackStreams >= 4 &&
  946. tfb_properties.maxTransformFeedbackBuffers && tfb_properties.transformFeedbackQueries &&
  947. tfb_properties.transformFeedbackDraw) {
  948. extensions.push_back(VK_EXT_TRANSFORM_FEEDBACK_EXTENSION_NAME);
  949. ext_transform_feedback = true;
  950. }
  951. }
  952. if (has_ext_custom_border_color) {
  953. VkPhysicalDeviceCustomBorderColorFeaturesEXT border_features;
  954. border_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CUSTOM_BORDER_COLOR_FEATURES_EXT;
  955. border_features.pNext = nullptr;
  956. features.pNext = &border_features;
  957. physical.GetFeatures2KHR(features);
  958. if (border_features.customBorderColors && border_features.customBorderColorWithoutFormat) {
  959. extensions.push_back(VK_EXT_CUSTOM_BORDER_COLOR_EXTENSION_NAME);
  960. ext_custom_border_color = true;
  961. }
  962. }
  963. if (has_ext_extended_dynamic_state) {
  964. VkPhysicalDeviceExtendedDynamicStateFeaturesEXT extended_dynamic_state;
  965. extended_dynamic_state.sType =
  966. VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTENDED_DYNAMIC_STATE_FEATURES_EXT;
  967. extended_dynamic_state.pNext = nullptr;
  968. features.pNext = &extended_dynamic_state;
  969. physical.GetFeatures2KHR(features);
  970. if (extended_dynamic_state.extendedDynamicState) {
  971. extensions.push_back(VK_EXT_EXTENDED_DYNAMIC_STATE_EXTENSION_NAME);
  972. ext_extended_dynamic_state = true;
  973. }
  974. }
  975. if (has_ext_line_rasterization) {
  976. VkPhysicalDeviceLineRasterizationFeaturesEXT line_raster;
  977. line_raster.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_FEATURES_EXT;
  978. line_raster.pNext = nullptr;
  979. features.pNext = &line_raster;
  980. physical.GetFeatures2KHR(features);
  981. if (line_raster.rectangularLines && line_raster.smoothLines) {
  982. extensions.push_back(VK_EXT_LINE_RASTERIZATION_EXTENSION_NAME);
  983. ext_line_rasterization = true;
  984. }
  985. }
  986. if (has_khr_workgroup_memory_explicit_layout) {
  987. VkPhysicalDeviceWorkgroupMemoryExplicitLayoutFeaturesKHR layout;
  988. layout.sType =
  989. VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_WORKGROUP_MEMORY_EXPLICIT_LAYOUT_FEATURES_KHR;
  990. layout.pNext = nullptr;
  991. features.pNext = &layout;
  992. physical.GetFeatures2KHR(features);
  993. if (layout.workgroupMemoryExplicitLayout &&
  994. layout.workgroupMemoryExplicitLayout8BitAccess &&
  995. layout.workgroupMemoryExplicitLayout16BitAccess &&
  996. layout.workgroupMemoryExplicitLayoutScalarBlockLayout) {
  997. extensions.push_back(VK_KHR_WORKGROUP_MEMORY_EXPLICIT_LAYOUT_EXTENSION_NAME);
  998. khr_workgroup_memory_explicit_layout = true;
  999. }
  1000. }
  1001. if (has_khr_pipeline_executable_properties) {
  1002. VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR executable_properties;
  1003. executable_properties.sType =
  1004. VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR;
  1005. executable_properties.pNext = nullptr;
  1006. features.pNext = &executable_properties;
  1007. physical.GetFeatures2KHR(features);
  1008. if (executable_properties.pipelineExecutableInfo) {
  1009. extensions.push_back(VK_KHR_PIPELINE_EXECUTABLE_PROPERTIES_EXTENSION_NAME);
  1010. khr_pipeline_executable_properties = true;
  1011. }
  1012. }
  1013. if (has_khr_image_format_list && has_khr_swapchain_mutable_format) {
  1014. extensions.push_back(VK_KHR_IMAGE_FORMAT_LIST_EXTENSION_NAME);
  1015. extensions.push_back(VK_KHR_SWAPCHAIN_MUTABLE_FORMAT_EXTENSION_NAME);
  1016. khr_swapchain_mutable_format = true;
  1017. }
  1018. if (khr_push_descriptor) {
  1019. VkPhysicalDevicePushDescriptorPropertiesKHR push_descriptor;
  1020. push_descriptor.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR;
  1021. push_descriptor.pNext = nullptr;
  1022. physical_properties.pNext = &push_descriptor;
  1023. physical.GetProperties2KHR(physical_properties);
  1024. max_push_descriptors = push_descriptor.maxPushDescriptors;
  1025. }
  1026. return extensions;
  1027. }
  1028. void Device::SetupFamilies(VkSurfaceKHR surface) {
  1029. const std::vector queue_family_properties = physical.GetQueueFamilyProperties();
  1030. std::optional<u32> graphics;
  1031. std::optional<u32> present;
  1032. for (u32 index = 0; index < static_cast<u32>(queue_family_properties.size()); ++index) {
  1033. if (graphics && (present || !surface)) {
  1034. break;
  1035. }
  1036. const VkQueueFamilyProperties& queue_family = queue_family_properties[index];
  1037. if (queue_family.queueCount == 0) {
  1038. continue;
  1039. }
  1040. if (queue_family.queueFlags & VK_QUEUE_GRAPHICS_BIT) {
  1041. graphics = index;
  1042. }
  1043. if (surface && physical.GetSurfaceSupportKHR(index, surface)) {
  1044. present = index;
  1045. }
  1046. }
  1047. if (!graphics) {
  1048. LOG_ERROR(Render_Vulkan, "Device lacks a graphics queue");
  1049. throw vk::Exception(VK_ERROR_FEATURE_NOT_PRESENT);
  1050. }
  1051. if (surface && !present) {
  1052. LOG_ERROR(Render_Vulkan, "Device lacks a present queue");
  1053. throw vk::Exception(VK_ERROR_FEATURE_NOT_PRESENT);
  1054. }
  1055. graphics_family = *graphics;
  1056. present_family = *present;
  1057. }
  1058. void Device::SetupFeatures() {
  1059. const VkPhysicalDeviceFeatures features{physical.GetFeatures()};
  1060. is_depth_bounds_supported = features.depthBounds;
  1061. is_formatless_image_load_supported = features.shaderStorageImageReadWithoutFormat;
  1062. is_shader_float64_supported = features.shaderFloat64;
  1063. is_shader_int64_supported = features.shaderInt64;
  1064. is_shader_int16_supported = features.shaderInt16;
  1065. is_shader_storage_image_multisample = features.shaderStorageImageMultisample;
  1066. is_blit_depth_stencil_supported = TestDepthStencilBlits();
  1067. is_optimal_astc_supported = IsOptimalAstcSupported(features);
  1068. }
  1069. void Device::SetupProperties() {
  1070. float_controls.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT_CONTROLS_PROPERTIES_KHR;
  1071. VkPhysicalDeviceProperties2KHR properties2{};
  1072. properties2.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROPERTIES_2_KHR;
  1073. properties2.pNext = &float_controls;
  1074. physical.GetProperties2KHR(properties2);
  1075. }
  1076. void Device::CollectTelemetryParameters() {
  1077. VkPhysicalDeviceDriverPropertiesKHR driver{
  1078. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR,
  1079. .pNext = nullptr,
  1080. .driverID = {},
  1081. .driverName = {},
  1082. .driverInfo = {},
  1083. .conformanceVersion = {},
  1084. };
  1085. VkPhysicalDeviceProperties2KHR device_properties{
  1086. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROPERTIES_2_KHR,
  1087. .pNext = &driver,
  1088. .properties = {},
  1089. };
  1090. physical.GetProperties2KHR(device_properties);
  1091. driver_id = driver.driverID;
  1092. vendor_name = driver.driverName;
  1093. }
  1094. void Device::CollectPhysicalMemoryInfo() {
  1095. const auto mem_properties = physical.GetMemoryProperties();
  1096. const size_t num_properties = mem_properties.memoryHeapCount;
  1097. device_access_memory = 0;
  1098. for (size_t element = 0; element < num_properties; ++element) {
  1099. if ((mem_properties.memoryHeaps[element].flags & VK_MEMORY_HEAP_DEVICE_LOCAL_BIT) != 0) {
  1100. device_access_memory += mem_properties.memoryHeaps[element].size;
  1101. }
  1102. }
  1103. }
  1104. void Device::CollectToolingInfo() {
  1105. if (!ext_tooling_info) {
  1106. return;
  1107. }
  1108. const auto vkGetPhysicalDeviceToolPropertiesEXT =
  1109. reinterpret_cast<PFN_vkGetPhysicalDeviceToolPropertiesEXT>(
  1110. dld.vkGetInstanceProcAddr(instance, "vkGetPhysicalDeviceToolPropertiesEXT"));
  1111. if (!vkGetPhysicalDeviceToolPropertiesEXT) {
  1112. return;
  1113. }
  1114. u32 tool_count = 0;
  1115. if (vkGetPhysicalDeviceToolPropertiesEXT(physical, &tool_count, nullptr) != VK_SUCCESS) {
  1116. return;
  1117. }
  1118. std::vector<VkPhysicalDeviceToolPropertiesEXT> tools(tool_count);
  1119. if (vkGetPhysicalDeviceToolPropertiesEXT(physical, &tool_count, tools.data()) != VK_SUCCESS) {
  1120. return;
  1121. }
  1122. for (const VkPhysicalDeviceToolPropertiesEXT& tool : tools) {
  1123. const std::string_view name = tool.name;
  1124. LOG_INFO(Render_Vulkan, "{}", name);
  1125. has_renderdoc = has_renderdoc || name == "RenderDoc";
  1126. has_nsight_graphics = has_nsight_graphics || name == "NVIDIA Nsight Graphics";
  1127. }
  1128. }
  1129. std::vector<VkDeviceQueueCreateInfo> Device::GetDeviceQueueCreateInfos() const {
  1130. static constexpr float QUEUE_PRIORITY = 1.0f;
  1131. std::unordered_set<u32> unique_queue_families{graphics_family, present_family};
  1132. std::vector<VkDeviceQueueCreateInfo> queue_cis;
  1133. queue_cis.reserve(unique_queue_families.size());
  1134. for (const u32 queue_family : unique_queue_families) {
  1135. auto& ci = queue_cis.emplace_back(VkDeviceQueueCreateInfo{
  1136. .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO,
  1137. .pNext = nullptr,
  1138. .flags = 0,
  1139. .queueFamilyIndex = queue_family,
  1140. .queueCount = 1,
  1141. .pQueuePriorities = nullptr,
  1142. });
  1143. ci.pQueuePriorities = &QUEUE_PRIORITY;
  1144. }
  1145. return queue_cis;
  1146. }
  1147. } // namespace Vulkan