rescaling_pass.cpp 12 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304
  1. // Copyright 2021 yuzu Emulator Project
  2. // Licensed under GPLv2 or any later version
  3. // Refer to the license.txt file included.
  4. #include "common/alignment.h"
  5. #include "common/settings.h"
  6. #include "shader_recompiler/environment.h"
  7. #include "shader_recompiler/frontend/ir/ir_emitter.h"
  8. #include "shader_recompiler/frontend/ir/modifiers.h"
  9. #include "shader_recompiler/frontend/ir/program.h"
  10. #include "shader_recompiler/frontend/ir/value.h"
  11. #include "shader_recompiler/ir_opt/passes.h"
  12. #include "shader_recompiler/shader_info.h"
  13. namespace Shader::Optimization {
  14. namespace {
  15. void VisitMark(const IR::Program& program, const IR::Inst& inst) {
  16. const bool is_fragment_shader{program.stage == Stage::Fragment};
  17. if (!is_fragment_shader) {
  18. return;
  19. }
  20. switch (inst.GetOpcode()) {
  21. case IR::Opcode::ShuffleIndex:
  22. case IR::Opcode::ShuffleUp:
  23. case IR::Opcode::ShuffleDown:
  24. case IR::Opcode::ShuffleButterfly: {
  25. const IR::Value shfl_arg{inst.Arg(0)};
  26. if (shfl_arg.IsImmediate()) {
  27. break;
  28. }
  29. const IR::Inst* const arg_inst{shfl_arg.InstRecursive()};
  30. if (arg_inst->GetOpcode() != IR::Opcode::BitCastU32F32) {
  31. break;
  32. }
  33. const IR::Value bitcast_arg{arg_inst->Arg(0)};
  34. if (bitcast_arg.IsImmediate()) {
  35. break;
  36. }
  37. IR::Inst* const bitcast_inst{bitcast_arg.InstRecursive()};
  38. if (bitcast_inst->GetOpcode() == IR::Opcode::GetAttribute) {
  39. const IR::Attribute attr{bitcast_inst->Arg(0).Attribute()};
  40. switch (attr) {
  41. case IR::Attribute::PositionX:
  42. case IR::Attribute::PositionY:
  43. bitcast_inst->SetFlags<u32>(0xDEADBEEF);
  44. break;
  45. default:
  46. break;
  47. }
  48. }
  49. break;
  50. }
  51. default:
  52. break;
  53. }
  54. }
  55. void PatchFragCoord(IR::Block& block, IR::Inst& inst) {
  56. IR::IREmitter ir{block, IR::Block::InstructionList::s_iterator_to(inst)};
  57. const IR::F32 down_factor{ir.ResolutionDownFactor()};
  58. const IR::F32 frag_coord{ir.GetAttribute(inst.Arg(0).Attribute())};
  59. const IR::F32 downscaled_frag_coord{ir.FPMul(frag_coord, down_factor)};
  60. inst.ReplaceUsesWith(downscaled_frag_coord);
  61. }
  62. [[nodiscard]] IR::U32 Scale(IR::IREmitter& ir, const IR::U1& is_scaled, const IR::U32& value) {
  63. IR::U32 scaled_value{value};
  64. bool changed{};
  65. if (const u32 up_scale = Settings::values.resolution_info.up_scale; up_scale != 1) {
  66. scaled_value = ir.IMul(scaled_value, ir.Imm32(up_scale));
  67. changed = true;
  68. }
  69. if (const u32 down_shift = Settings::values.resolution_info.down_shift; down_shift != 0) {
  70. scaled_value = ir.ShiftRightArithmetic(scaled_value, ir.Imm32(down_shift));
  71. changed = true;
  72. }
  73. if (changed) {
  74. return IR::U32{ir.Select(is_scaled, scaled_value, value)};
  75. } else {
  76. return value;
  77. }
  78. }
  79. [[nodiscard]] IR::U32 SubScale(IR::IREmitter& ir, const IR::U1& is_scaled, const IR::U32& value,
  80. const IR::Attribute attrib) {
  81. const IR::F32 opt1{ir.Imm32(Settings::values.resolution_info.up_factor)};
  82. const IR::F32 base{ir.FPMul(ir.ConvertUToF(32, 32, value), opt1)};
  83. const IR::F32 frag_coord{ir.GetAttribute(attrib)};
  84. const IR::F32 opt2{ir.Imm32(Settings::values.resolution_info.down_factor)};
  85. const IR::F32 floor{ir.FPMul(opt1, ir.FPFloor(ir.FPMul(frag_coord, opt2)))};
  86. const IR::U32 deviation{
  87. ir.ConvertFToU(32, ir.FPAdd(base, ir.FPAdd(frag_coord, ir.FPNeg(floor))))};
  88. return IR::U32{ir.Select(is_scaled, deviation, value)};
  89. }
  90. [[nodiscard]] IR::U32 DownScale(IR::IREmitter& ir, const IR::U1& is_scaled, IR::U32 value) {
  91. IR::U32 scaled_value{value};
  92. bool changed{};
  93. if (const u32 down_shift = Settings::values.resolution_info.down_shift; down_shift != 0) {
  94. scaled_value = ir.ShiftLeftLogical(scaled_value, ir.Imm32(down_shift));
  95. changed = true;
  96. }
  97. if (const u32 up_scale = Settings::values.resolution_info.up_scale; up_scale != 1) {
  98. scaled_value = ir.IDiv(scaled_value, ir.Imm32(up_scale));
  99. changed = true;
  100. }
  101. if (changed) {
  102. return IR::U32{ir.Select(is_scaled, scaled_value, value)};
  103. } else {
  104. return value;
  105. }
  106. }
  107. void PatchImageQueryDimensions(IR::Block& block, IR::Inst& inst) {
  108. const auto it{IR::Block::InstructionList::s_iterator_to(inst)};
  109. IR::IREmitter ir{block, IR::Block::InstructionList::s_iterator_to(inst)};
  110. const auto info{inst.Flags<IR::TextureInstInfo>()};
  111. const IR::U1 is_scaled{ir.IsTextureScaled(ir.Imm32(info.descriptor_index))};
  112. switch (info.type) {
  113. case TextureType::Color2D:
  114. case TextureType::ColorArray2D: {
  115. const IR::Value new_inst{&*block.PrependNewInst(it, inst)};
  116. const IR::U32 width{DownScale(ir, is_scaled, IR::U32{ir.CompositeExtract(new_inst, 0)})};
  117. const IR::U32 height{DownScale(ir, is_scaled, IR::U32{ir.CompositeExtract(new_inst, 1)})};
  118. const IR::Value replacement{ir.CompositeConstruct(
  119. width, height, ir.CompositeExtract(new_inst, 2), ir.CompositeExtract(new_inst, 3))};
  120. inst.ReplaceUsesWith(replacement);
  121. break;
  122. }
  123. case TextureType::Color1D:
  124. case TextureType::ColorArray1D:
  125. case TextureType::Color3D:
  126. case TextureType::ColorCube:
  127. case TextureType::ColorArrayCube:
  128. case TextureType::Buffer:
  129. // Nothing to patch here
  130. break;
  131. }
  132. }
  133. void ScaleIntegerCoord(IR::IREmitter& ir, IR::Inst& inst, const IR::U1& is_scaled) {
  134. const auto info{inst.Flags<IR::TextureInstInfo>()};
  135. const IR::Value coord{inst.Arg(1)};
  136. switch (info.type) {
  137. case TextureType::Color2D: {
  138. const IR::U32 x{Scale(ir, is_scaled, IR::U32{ir.CompositeExtract(coord, 0)})};
  139. const IR::U32 y{Scale(ir, is_scaled, IR::U32{ir.CompositeExtract(coord, 1)})};
  140. inst.SetArg(1, ir.CompositeConstruct(x, y));
  141. break;
  142. }
  143. case TextureType::ColorArray2D: {
  144. const IR::U32 x{Scale(ir, is_scaled, IR::U32{ir.CompositeExtract(coord, 0)})};
  145. const IR::U32 y{Scale(ir, is_scaled, IR::U32{ir.CompositeExtract(coord, 1)})};
  146. const IR::U32 z{ir.CompositeExtract(coord, 2)};
  147. inst.SetArg(1, ir.CompositeConstruct(x, y, z));
  148. break;
  149. }
  150. case TextureType::Color1D:
  151. case TextureType::ColorArray1D:
  152. case TextureType::Color3D:
  153. case TextureType::ColorCube:
  154. case TextureType::ColorArrayCube:
  155. case TextureType::Buffer:
  156. // Nothing to patch here
  157. break;
  158. }
  159. }
  160. void SubScaleImageFetch(IR::Block& block, IR::Inst& inst) {
  161. IR::IREmitter ir{block, IR::Block::InstructionList::s_iterator_to(inst)};
  162. const auto info{inst.Flags<IR::TextureInstInfo>()};
  163. const IR::U1 is_scaled{ir.IsTextureScaled(ir.Imm32(info.descriptor_index))};
  164. const IR::Value coord{inst.Arg(1)};
  165. switch (info.type) {
  166. case TextureType::Color2D: {
  167. const IR::U32 x{SubScale(ir, is_scaled, IR::U32{ir.CompositeExtract(coord, 0)},
  168. IR::Attribute::PositionX)};
  169. const IR::U32 y{SubScale(ir, is_scaled, IR::U32{ir.CompositeExtract(coord, 1)},
  170. IR::Attribute::PositionY)};
  171. inst.SetArg(1, ir.CompositeConstruct(x, y));
  172. break;
  173. }
  174. case TextureType::ColorArray2D: {
  175. const IR::U32 x{SubScale(ir, is_scaled, IR::U32{ir.CompositeExtract(coord, 0)},
  176. IR::Attribute::PositionX)};
  177. const IR::U32 y{SubScale(ir, is_scaled, IR::U32{ir.CompositeExtract(coord, 1)},
  178. IR::Attribute::PositionY)};
  179. const IR::U32 z{ir.CompositeExtract(coord, 2)};
  180. inst.SetArg(1, ir.CompositeConstruct(x, y, z));
  181. break;
  182. }
  183. case TextureType::Color1D:
  184. case TextureType::ColorArray1D:
  185. case TextureType::Color3D:
  186. case TextureType::ColorCube:
  187. case TextureType::ColorArrayCube:
  188. case TextureType::Buffer:
  189. // Nothing to patch here
  190. break;
  191. }
  192. }
  193. void SubScaleImageRead(IR::Block& block, IR::Inst& inst) {
  194. IR::IREmitter ir{block, IR::Block::InstructionList::s_iterator_to(inst)};
  195. const auto info{inst.Flags<IR::TextureInstInfo>()};
  196. const IR::U1 is_scaled{ir.IsImageScaled(ir.Imm32(info.descriptor_index))};
  197. const IR::Value coord{inst.Arg(1)};
  198. switch (info.type) {
  199. case TextureType::Color2D: {
  200. const IR::U32 x{SubScale(ir, is_scaled, IR::U32{ir.CompositeExtract(coord, 0)},
  201. IR::Attribute::PositionX)};
  202. const IR::U32 y{SubScale(ir, is_scaled, IR::U32{ir.CompositeExtract(coord, 1)},
  203. IR::Attribute::PositionY)};
  204. inst.SetArg(1, ir.CompositeConstruct(x, y));
  205. break;
  206. }
  207. case TextureType::ColorArray2D: {
  208. const IR::U32 x{SubScale(ir, is_scaled, IR::U32{ir.CompositeExtract(coord, 0)},
  209. IR::Attribute::PositionX)};
  210. const IR::U32 y{SubScale(ir, is_scaled, IR::U32{ir.CompositeExtract(coord, 1)},
  211. IR::Attribute::PositionY)};
  212. const IR::U32 z{ir.CompositeExtract(coord, 2)};
  213. inst.SetArg(1, ir.CompositeConstruct(x, y, z));
  214. break;
  215. }
  216. case TextureType::Color1D:
  217. case TextureType::ColorArray1D:
  218. case TextureType::Color3D:
  219. case TextureType::ColorCube:
  220. case TextureType::ColorArrayCube:
  221. case TextureType::Buffer:
  222. // Nothing to patch here
  223. break;
  224. }
  225. }
  226. void PatchImageFetch(IR::Block& block, IR::Inst& inst) {
  227. IR::IREmitter ir{block, IR::Block::InstructionList::s_iterator_to(inst)};
  228. const auto info{inst.Flags<IR::TextureInstInfo>()};
  229. const IR::U1 is_scaled{ir.IsTextureScaled(ir.Imm32(info.descriptor_index))};
  230. ScaleIntegerCoord(ir, inst, is_scaled);
  231. }
  232. void PatchImageRead(IR::Block& block, IR::Inst& inst) {
  233. IR::IREmitter ir{block, IR::Block::InstructionList::s_iterator_to(inst)};
  234. const auto info{inst.Flags<IR::TextureInstInfo>()};
  235. const IR::U1 is_scaled{ir.IsImageScaled(ir.Imm32(info.descriptor_index))};
  236. ScaleIntegerCoord(ir, inst, is_scaled);
  237. }
  238. void Visit(const IR::Program& program, IR::Block& block, IR::Inst& inst) {
  239. const bool is_fragment_shader{program.stage == Stage::Fragment};
  240. const bool is_compute_shader{program.stage == Stage::Compute};
  241. switch (inst.GetOpcode()) {
  242. case IR::Opcode::GetAttribute: {
  243. const IR::Attribute attr{inst.Arg(0).Attribute()};
  244. switch (attr) {
  245. case IR::Attribute::PositionX:
  246. case IR::Attribute::PositionY:
  247. if (is_fragment_shader && inst.Flags<u32>() != 0xDEADBEEF) {
  248. PatchFragCoord(block, inst);
  249. }
  250. break;
  251. default:
  252. break;
  253. }
  254. break;
  255. }
  256. case IR::Opcode::ImageQueryDimensions:
  257. PatchImageQueryDimensions(block, inst);
  258. break;
  259. case IR::Opcode::ImageFetch:
  260. if (is_fragment_shader) {
  261. SubScaleImageFetch(block, inst);
  262. } else if (is_compute_shader) {
  263. PatchImageFetch(block, inst);
  264. }
  265. break;
  266. case IR::Opcode::ImageRead:
  267. if (is_fragment_shader) {
  268. SubScaleImageRead(block, inst);
  269. } else if (is_compute_shader) {
  270. PatchImageRead(block, inst);
  271. }
  272. break;
  273. default:
  274. break;
  275. }
  276. }
  277. } // Anonymous namespace
  278. void RescalingPass(IR::Program& program) {
  279. for (IR::Block* const block : program.post_order_blocks) {
  280. for (IR::Inst& inst : block->Instructions()) {
  281. VisitMark(program, inst);
  282. }
  283. }
  284. for (IR::Block* const block : program.post_order_blocks) {
  285. for (IR::Inst& inst : block->Instructions()) {
  286. Visit(program, *block, inst);
  287. }
  288. }
  289. }
  290. } // namespace Shader::Optimization