collect_shader_info_pass.cpp 16 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469
  1. // Copyright 2021 yuzu Emulator Project
  2. // Licensed under GPLv2 or any later version
  3. // Refer to the license.txt file included.
  4. #include "shader_recompiler/frontend/ir/microinstruction.h"
  5. #include "shader_recompiler/frontend/ir/modifiers.h"
  6. #include "shader_recompiler/frontend/ir/program.h"
  7. #include "shader_recompiler/shader_info.h"
  8. namespace Shader::Optimization {
  9. namespace {
  10. void AddConstantBufferDescriptor(Info& info, u32 index, u32 count) {
  11. if (count != 1) {
  12. throw NotImplementedException("Constant buffer descriptor indexing");
  13. }
  14. if ((info.constant_buffer_mask & (1U << index)) != 0) {
  15. return;
  16. }
  17. info.constant_buffer_mask |= 1U << index;
  18. auto& cbufs{info.constant_buffer_descriptors};
  19. cbufs.insert(std::ranges::lower_bound(cbufs, index, {}, &ConstantBufferDescriptor::index),
  20. ConstantBufferDescriptor{
  21. .index{index},
  22. .count{1},
  23. });
  24. }
  25. void GetAttribute(Info& info, IR::Attribute attribute) {
  26. if (IR::IsGeneric(attribute)) {
  27. info.loads_generics.at(IR::GenericAttributeIndex(attribute)) = true;
  28. return;
  29. }
  30. switch (attribute) {
  31. case IR::Attribute::PositionX:
  32. case IR::Attribute::PositionY:
  33. case IR::Attribute::PositionZ:
  34. case IR::Attribute::PositionW:
  35. info.loads_position = true;
  36. break;
  37. case IR::Attribute::InstanceId:
  38. info.loads_instance_id = true;
  39. break;
  40. case IR::Attribute::VertexId:
  41. info.loads_vertex_id = true;
  42. break;
  43. case IR::Attribute::FrontFace:
  44. info.loads_front_face = true;
  45. break;
  46. default:
  47. throw NotImplementedException("Get attribute {}", attribute);
  48. }
  49. }
  50. void SetAttribute(Info& info, IR::Attribute attribute) {
  51. if (IR::IsGeneric(attribute)) {
  52. info.stores_generics.at(IR::GenericAttributeIndex(attribute)) = true;
  53. return;
  54. }
  55. switch (attribute) {
  56. case IR::Attribute::PositionX:
  57. case IR::Attribute::PositionY:
  58. case IR::Attribute::PositionZ:
  59. case IR::Attribute::PositionW:
  60. info.stores_position = true;
  61. break;
  62. default:
  63. throw NotImplementedException("Set attribute {}", attribute);
  64. }
  65. }
  66. void VisitUsages(Info& info, IR::Inst& inst) {
  67. switch (inst.Opcode()) {
  68. case IR::Opcode::CompositeConstructF16x2:
  69. case IR::Opcode::CompositeConstructF16x3:
  70. case IR::Opcode::CompositeConstructF16x4:
  71. case IR::Opcode::CompositeExtractF16x2:
  72. case IR::Opcode::CompositeExtractF16x3:
  73. case IR::Opcode::CompositeExtractF16x4:
  74. case IR::Opcode::CompositeInsertF16x2:
  75. case IR::Opcode::CompositeInsertF16x3:
  76. case IR::Opcode::CompositeInsertF16x4:
  77. case IR::Opcode::SelectF16:
  78. case IR::Opcode::BitCastU16F16:
  79. case IR::Opcode::BitCastF16U16:
  80. case IR::Opcode::PackFloat2x16:
  81. case IR::Opcode::UnpackFloat2x16:
  82. case IR::Opcode::ConvertS16F16:
  83. case IR::Opcode::ConvertS32F16:
  84. case IR::Opcode::ConvertS64F16:
  85. case IR::Opcode::ConvertU16F16:
  86. case IR::Opcode::ConvertU32F16:
  87. case IR::Opcode::ConvertU64F16:
  88. case IR::Opcode::ConvertF16S8:
  89. case IR::Opcode::ConvertF16S16:
  90. case IR::Opcode::ConvertF16S32:
  91. case IR::Opcode::ConvertF16S64:
  92. case IR::Opcode::ConvertF16U8:
  93. case IR::Opcode::ConvertF16U16:
  94. case IR::Opcode::ConvertF16U32:
  95. case IR::Opcode::ConvertF16U64:
  96. case IR::Opcode::FPAbs16:
  97. case IR::Opcode::FPAdd16:
  98. case IR::Opcode::FPCeil16:
  99. case IR::Opcode::FPFloor16:
  100. case IR::Opcode::FPFma16:
  101. case IR::Opcode::FPMul16:
  102. case IR::Opcode::FPNeg16:
  103. case IR::Opcode::FPRoundEven16:
  104. case IR::Opcode::FPSaturate16:
  105. case IR::Opcode::FPClamp16:
  106. case IR::Opcode::FPTrunc16:
  107. case IR::Opcode::FPOrdEqual16:
  108. case IR::Opcode::FPUnordEqual16:
  109. case IR::Opcode::FPOrdNotEqual16:
  110. case IR::Opcode::FPUnordNotEqual16:
  111. case IR::Opcode::FPOrdLessThan16:
  112. case IR::Opcode::FPUnordLessThan16:
  113. case IR::Opcode::FPOrdGreaterThan16:
  114. case IR::Opcode::FPUnordGreaterThan16:
  115. case IR::Opcode::FPOrdLessThanEqual16:
  116. case IR::Opcode::FPUnordLessThanEqual16:
  117. case IR::Opcode::FPOrdGreaterThanEqual16:
  118. case IR::Opcode::FPUnordGreaterThanEqual16:
  119. case IR::Opcode::FPIsNan16:
  120. info.uses_fp16 = true;
  121. break;
  122. case IR::Opcode::CompositeConstructF64x2:
  123. case IR::Opcode::CompositeConstructF64x3:
  124. case IR::Opcode::CompositeConstructF64x4:
  125. case IR::Opcode::CompositeExtractF64x2:
  126. case IR::Opcode::CompositeExtractF64x3:
  127. case IR::Opcode::CompositeExtractF64x4:
  128. case IR::Opcode::CompositeInsertF64x2:
  129. case IR::Opcode::CompositeInsertF64x3:
  130. case IR::Opcode::CompositeInsertF64x4:
  131. case IR::Opcode::SelectF64:
  132. case IR::Opcode::BitCastU64F64:
  133. case IR::Opcode::BitCastF64U64:
  134. case IR::Opcode::PackDouble2x32:
  135. case IR::Opcode::UnpackDouble2x32:
  136. case IR::Opcode::FPAbs64:
  137. case IR::Opcode::FPAdd64:
  138. case IR::Opcode::FPCeil64:
  139. case IR::Opcode::FPFloor64:
  140. case IR::Opcode::FPFma64:
  141. case IR::Opcode::FPMax64:
  142. case IR::Opcode::FPMin64:
  143. case IR::Opcode::FPMul64:
  144. case IR::Opcode::FPNeg64:
  145. case IR::Opcode::FPRecip64:
  146. case IR::Opcode::FPRecipSqrt64:
  147. case IR::Opcode::FPRoundEven64:
  148. case IR::Opcode::FPSaturate64:
  149. case IR::Opcode::FPClamp64:
  150. case IR::Opcode::FPTrunc64:
  151. case IR::Opcode::FPOrdEqual64:
  152. case IR::Opcode::FPUnordEqual64:
  153. case IR::Opcode::FPOrdNotEqual64:
  154. case IR::Opcode::FPUnordNotEqual64:
  155. case IR::Opcode::FPOrdLessThan64:
  156. case IR::Opcode::FPUnordLessThan64:
  157. case IR::Opcode::FPOrdGreaterThan64:
  158. case IR::Opcode::FPUnordGreaterThan64:
  159. case IR::Opcode::FPOrdLessThanEqual64:
  160. case IR::Opcode::FPUnordLessThanEqual64:
  161. case IR::Opcode::FPOrdGreaterThanEqual64:
  162. case IR::Opcode::FPUnordGreaterThanEqual64:
  163. case IR::Opcode::FPIsNan64:
  164. case IR::Opcode::ConvertS16F64:
  165. case IR::Opcode::ConvertS32F64:
  166. case IR::Opcode::ConvertS64F64:
  167. case IR::Opcode::ConvertU16F64:
  168. case IR::Opcode::ConvertU32F64:
  169. case IR::Opcode::ConvertU64F64:
  170. case IR::Opcode::ConvertF32F64:
  171. case IR::Opcode::ConvertF64F32:
  172. case IR::Opcode::ConvertF64S8:
  173. case IR::Opcode::ConvertF64S16:
  174. case IR::Opcode::ConvertF64S32:
  175. case IR::Opcode::ConvertF64S64:
  176. case IR::Opcode::ConvertF64U8:
  177. case IR::Opcode::ConvertF64U16:
  178. case IR::Opcode::ConvertF64U32:
  179. case IR::Opcode::ConvertF64U64:
  180. info.uses_fp64 = true;
  181. break;
  182. default:
  183. break;
  184. }
  185. switch (inst.Opcode()) {
  186. case IR::Opcode::GetCbufU8:
  187. case IR::Opcode::GetCbufS8:
  188. case IR::Opcode::UndefU8:
  189. case IR::Opcode::LoadGlobalU8:
  190. case IR::Opcode::LoadGlobalS8:
  191. case IR::Opcode::WriteGlobalU8:
  192. case IR::Opcode::WriteGlobalS8:
  193. case IR::Opcode::LoadStorageU8:
  194. case IR::Opcode::LoadStorageS8:
  195. case IR::Opcode::WriteStorageU8:
  196. case IR::Opcode::WriteStorageS8:
  197. case IR::Opcode::SelectU8:
  198. case IR::Opcode::ConvertF16S8:
  199. case IR::Opcode::ConvertF16U8:
  200. case IR::Opcode::ConvertF32S8:
  201. case IR::Opcode::ConvertF32U8:
  202. case IR::Opcode::ConvertF64S8:
  203. case IR::Opcode::ConvertF64U8:
  204. info.uses_int8 = true;
  205. break;
  206. default:
  207. break;
  208. }
  209. switch (inst.Opcode()) {
  210. case IR::Opcode::GetCbufU16:
  211. case IR::Opcode::GetCbufS16:
  212. case IR::Opcode::UndefU16:
  213. case IR::Opcode::LoadGlobalU16:
  214. case IR::Opcode::LoadGlobalS16:
  215. case IR::Opcode::WriteGlobalU16:
  216. case IR::Opcode::WriteGlobalS16:
  217. case IR::Opcode::LoadStorageU16:
  218. case IR::Opcode::LoadStorageS16:
  219. case IR::Opcode::WriteStorageU16:
  220. case IR::Opcode::WriteStorageS16:
  221. case IR::Opcode::SelectU16:
  222. case IR::Opcode::BitCastU16F16:
  223. case IR::Opcode::BitCastF16U16:
  224. case IR::Opcode::ConvertS16F16:
  225. case IR::Opcode::ConvertS16F32:
  226. case IR::Opcode::ConvertS16F64:
  227. case IR::Opcode::ConvertU16F16:
  228. case IR::Opcode::ConvertU16F32:
  229. case IR::Opcode::ConvertU16F64:
  230. case IR::Opcode::ConvertF16S16:
  231. case IR::Opcode::ConvertF16U16:
  232. case IR::Opcode::ConvertF32S16:
  233. case IR::Opcode::ConvertF32U16:
  234. case IR::Opcode::ConvertF64S16:
  235. case IR::Opcode::ConvertF64U16:
  236. info.uses_int16 = true;
  237. break;
  238. default:
  239. break;
  240. }
  241. switch (inst.Opcode()) {
  242. case IR::Opcode::GetCbufU64:
  243. case IR::Opcode::UndefU64:
  244. case IR::Opcode::LoadGlobalU8:
  245. case IR::Opcode::LoadGlobalS8:
  246. case IR::Opcode::LoadGlobalU16:
  247. case IR::Opcode::LoadGlobalS16:
  248. case IR::Opcode::LoadGlobal32:
  249. case IR::Opcode::LoadGlobal64:
  250. case IR::Opcode::LoadGlobal128:
  251. case IR::Opcode::WriteGlobalU8:
  252. case IR::Opcode::WriteGlobalS8:
  253. case IR::Opcode::WriteGlobalU16:
  254. case IR::Opcode::WriteGlobalS16:
  255. case IR::Opcode::WriteGlobal32:
  256. case IR::Opcode::WriteGlobal64:
  257. case IR::Opcode::WriteGlobal128:
  258. case IR::Opcode::SelectU64:
  259. case IR::Opcode::BitCastU64F64:
  260. case IR::Opcode::BitCastF64U64:
  261. case IR::Opcode::PackUint2x32:
  262. case IR::Opcode::UnpackUint2x32:
  263. case IR::Opcode::IAdd64:
  264. case IR::Opcode::ISub64:
  265. case IR::Opcode::INeg64:
  266. case IR::Opcode::ShiftLeftLogical64:
  267. case IR::Opcode::ShiftRightLogical64:
  268. case IR::Opcode::ShiftRightArithmetic64:
  269. case IR::Opcode::ConvertS64F16:
  270. case IR::Opcode::ConvertS64F32:
  271. case IR::Opcode::ConvertS64F64:
  272. case IR::Opcode::ConvertU64F16:
  273. case IR::Opcode::ConvertU64F32:
  274. case IR::Opcode::ConvertU64F64:
  275. case IR::Opcode::ConvertU64U32:
  276. case IR::Opcode::ConvertU32U64:
  277. case IR::Opcode::ConvertF16U64:
  278. case IR::Opcode::ConvertF32U64:
  279. case IR::Opcode::ConvertF64U64:
  280. info.uses_int64 = true;
  281. break;
  282. default:
  283. break;
  284. }
  285. switch (inst.Opcode()) {
  286. case IR::Opcode::DemoteToHelperInvocation:
  287. info.uses_demote_to_helper_invocation = true;
  288. break;
  289. case IR::Opcode::GetAttribute:
  290. GetAttribute(info, inst.Arg(0).Attribute());
  291. break;
  292. case IR::Opcode::SetAttribute:
  293. SetAttribute(info, inst.Arg(0).Attribute());
  294. break;
  295. case IR::Opcode::SetFragColor:
  296. info.stores_frag_color[inst.Arg(0).U32()] = true;
  297. break;
  298. case IR::Opcode::SetFragDepth:
  299. info.stores_frag_depth = true;
  300. break;
  301. case IR::Opcode::WorkgroupId:
  302. info.uses_workgroup_id = true;
  303. break;
  304. case IR::Opcode::LocalInvocationId:
  305. info.uses_local_invocation_id = true;
  306. break;
  307. case IR::Opcode::ShuffleIndex:
  308. case IR::Opcode::ShuffleUp:
  309. case IR::Opcode::ShuffleDown:
  310. case IR::Opcode::ShuffleButterfly:
  311. info.uses_subgroup_invocation_id = true;
  312. break;
  313. case IR::Opcode::GetCbufU8:
  314. case IR::Opcode::GetCbufS8:
  315. case IR::Opcode::GetCbufU16:
  316. case IR::Opcode::GetCbufS16:
  317. case IR::Opcode::GetCbufU32:
  318. case IR::Opcode::GetCbufF32:
  319. case IR::Opcode::GetCbufU64: {
  320. if (const IR::Value index{inst.Arg(0)}; index.IsImmediate()) {
  321. AddConstantBufferDescriptor(info, index.U32(), 1);
  322. } else {
  323. throw NotImplementedException("Constant buffer with non-immediate index");
  324. }
  325. switch (inst.Opcode()) {
  326. case IR::Opcode::GetCbufU8:
  327. case IR::Opcode::GetCbufS8:
  328. info.used_constant_buffer_types |= IR::Type::U8;
  329. break;
  330. case IR::Opcode::GetCbufU16:
  331. case IR::Opcode::GetCbufS16:
  332. info.used_constant_buffer_types |= IR::Type::U16;
  333. break;
  334. case IR::Opcode::GetCbufU32:
  335. info.used_constant_buffer_types |= IR::Type::U32;
  336. break;
  337. case IR::Opcode::GetCbufF32:
  338. info.used_constant_buffer_types |= IR::Type::F32;
  339. break;
  340. case IR::Opcode::GetCbufU64:
  341. info.used_constant_buffer_types |= IR::Type::U64;
  342. break;
  343. default:
  344. break;
  345. }
  346. break;
  347. }
  348. case IR::Opcode::BindlessImageSampleImplicitLod:
  349. case IR::Opcode::BindlessImageSampleExplicitLod:
  350. case IR::Opcode::BindlessImageSampleDrefImplicitLod:
  351. case IR::Opcode::BindlessImageSampleDrefExplicitLod:
  352. case IR::Opcode::BindlessImageGather:
  353. case IR::Opcode::BindlessImageGatherDref:
  354. case IR::Opcode::BoundImageSampleImplicitLod:
  355. case IR::Opcode::BoundImageSampleExplicitLod:
  356. case IR::Opcode::BoundImageSampleDrefImplicitLod:
  357. case IR::Opcode::BoundImageSampleDrefExplicitLod:
  358. case IR::Opcode::BoundImageGather:
  359. case IR::Opcode::BoundImageGatherDref:
  360. case IR::Opcode::ImageSampleImplicitLod:
  361. case IR::Opcode::ImageSampleExplicitLod:
  362. case IR::Opcode::ImageSampleDrefImplicitLod:
  363. case IR::Opcode::ImageSampleDrefExplicitLod:
  364. case IR::Opcode::ImageGather:
  365. case IR::Opcode::ImageGatherDref:
  366. case IR::Opcode::ImageQueryDimensions: {
  367. const TextureType type{inst.Flags<IR::TextureInstInfo>().type};
  368. info.uses_sampled_1d |= type == TextureType::Color1D || type == TextureType::ColorArray1D ||
  369. type == TextureType::Shadow1D || type == TextureType::ShadowArray1D;
  370. info.uses_sparse_residency |=
  371. inst.GetAssociatedPseudoOperation(IR::Opcode::GetSparseFromOp) != nullptr;
  372. break;
  373. }
  374. case IR::Opcode::VoteAll:
  375. case IR::Opcode::VoteAny:
  376. case IR::Opcode::VoteEqual:
  377. case IR::Opcode::SubgroupBallot:
  378. info.uses_subgroup_vote = true;
  379. break;
  380. default:
  381. break;
  382. }
  383. }
  384. void VisitFpModifiers(Info& info, IR::Inst& inst) {
  385. switch (inst.Opcode()) {
  386. case IR::Opcode::FPAdd16:
  387. case IR::Opcode::FPFma16:
  388. case IR::Opcode::FPMul16:
  389. case IR::Opcode::FPRoundEven16:
  390. case IR::Opcode::FPFloor16:
  391. case IR::Opcode::FPCeil16:
  392. case IR::Opcode::FPTrunc16: {
  393. const auto control{inst.Flags<IR::FpControl>()};
  394. switch (control.fmz_mode) {
  395. case IR::FmzMode::DontCare:
  396. break;
  397. case IR::FmzMode::FTZ:
  398. case IR::FmzMode::FMZ:
  399. info.uses_fp16_denorms_flush = true;
  400. break;
  401. case IR::FmzMode::None:
  402. info.uses_fp16_denorms_preserve = true;
  403. break;
  404. }
  405. break;
  406. }
  407. case IR::Opcode::FPAdd32:
  408. case IR::Opcode::FPFma32:
  409. case IR::Opcode::FPMul32:
  410. case IR::Opcode::FPRoundEven32:
  411. case IR::Opcode::FPFloor32:
  412. case IR::Opcode::FPCeil32:
  413. case IR::Opcode::FPTrunc32:
  414. case IR::Opcode::FPOrdEqual32:
  415. case IR::Opcode::FPUnordEqual32:
  416. case IR::Opcode::FPOrdNotEqual32:
  417. case IR::Opcode::FPUnordNotEqual32:
  418. case IR::Opcode::FPOrdLessThan32:
  419. case IR::Opcode::FPUnordLessThan32:
  420. case IR::Opcode::FPOrdGreaterThan32:
  421. case IR::Opcode::FPUnordGreaterThan32:
  422. case IR::Opcode::FPOrdLessThanEqual32:
  423. case IR::Opcode::FPUnordLessThanEqual32:
  424. case IR::Opcode::FPOrdGreaterThanEqual32:
  425. case IR::Opcode::FPUnordGreaterThanEqual32:
  426. case IR::Opcode::ConvertF16F32:
  427. case IR::Opcode::ConvertF64F32: {
  428. const auto control{inst.Flags<IR::FpControl>()};
  429. switch (control.fmz_mode) {
  430. case IR::FmzMode::DontCare:
  431. break;
  432. case IR::FmzMode::FTZ:
  433. case IR::FmzMode::FMZ:
  434. info.uses_fp32_denorms_flush = true;
  435. break;
  436. case IR::FmzMode::None:
  437. info.uses_fp32_denorms_preserve = true;
  438. break;
  439. }
  440. break;
  441. }
  442. default:
  443. break;
  444. }
  445. }
  446. void Visit(Info& info, IR::Inst& inst) {
  447. VisitUsages(info, inst);
  448. VisitFpModifiers(info, inst);
  449. }
  450. } // Anonymous namespace
  451. void CollectShaderInfoPass(IR::Program& program) {
  452. Info& info{program.info};
  453. for (IR::Block* const block : program.post_order_blocks) {
  454. for (IR::Inst& inst : block->Instructions()) {
  455. Visit(info, inst);
  456. }
  457. }
  458. }
  459. } // namespace Shader::Optimization