diff options
| author | 2022-12-24 19:19:41 -0500 | |
|---|---|---|
| committer | 2023-01-01 16:43:58 -0500 | |
| commit | d09aa0182f18d1ac338ab47009b42fdeb67497a8 (patch) | |
| tree | 1b3ae726f7aaa465b30fba795cfca0ae88d2140b /src/video_core/macro | |
| parent | Rasterizer: Setup skeleton for Host Conditional rendering (diff) | |
| download | yuzu-d09aa0182f18d1ac338ab47009b42fdeb67497a8.tar.gz yuzu-d09aa0182f18d1ac338ab47009b42fdeb67497a8.tar.xz yuzu-d09aa0182f18d1ac338ab47009b42fdeb67497a8.zip | |
MacroHLE: Final cleanup and fixes.
Diffstat (limited to 'src/video_core/macro')
| -rw-r--r-- | src/video_core/macro/macro_hle.cpp | 98 |
1 files changed, 28 insertions, 70 deletions
diff --git a/src/video_core/macro/macro_hle.cpp b/src/video_core/macro/macro_hle.cpp index 294a338d2..3481fcd41 100644 --- a/src/video_core/macro/macro_hle.cpp +++ b/src/video_core/macro/macro_hle.cpp | |||
| @@ -47,21 +47,7 @@ public: | |||
| 47 | explicit HLEMacroImpl(Engines::Maxwell3D& maxwell3d_) : maxwell3d{maxwell3d_} {} | 47 | explicit HLEMacroImpl(Engines::Maxwell3D& maxwell3d_) : maxwell3d{maxwell3d_} {} |
| 48 | 48 | ||
| 49 | protected: | 49 | protected: |
| 50 | void advanceCheck() { | ||
| 51 | current_value = (current_value + 1) % fibonacci_post; | ||
| 52 | check_limit = current_value == 0; | ||
| 53 | if (check_limit) { | ||
| 54 | const u32 new_fibonacci = fibonacci_pre + fibonacci_post; | ||
| 55 | fibonacci_pre = fibonacci_post; | ||
| 56 | fibonacci_post = new_fibonacci; | ||
| 57 | } | ||
| 58 | } | ||
| 59 | |||
| 60 | Engines::Maxwell3D& maxwell3d; | 50 | Engines::Maxwell3D& maxwell3d; |
| 61 | u32 fibonacci_pre{89}; | ||
| 62 | u32 fibonacci_post{144}; | ||
| 63 | u32 current_value{fibonacci_post - 1}; | ||
| 64 | bool check_limit{}; | ||
| 65 | }; | 51 | }; |
| 66 | 52 | ||
| 67 | class HLE_771BB18C62444DA0 final : public HLEMacroImpl { | 53 | class HLE_771BB18C62444DA0 final : public HLEMacroImpl { |
| @@ -124,12 +110,13 @@ private: | |||
| 124 | maxwell3d.RefreshParameters(); | 110 | maxwell3d.RefreshParameters(); |
| 125 | const u32 instance_count = (maxwell3d.GetRegisterValue(0xD1B) & parameters[2]); | 111 | const u32 instance_count = (maxwell3d.GetRegisterValue(0xD1B) & parameters[2]); |
| 126 | 112 | ||
| 113 | auto topology = static_cast<Maxwell::Regs::PrimitiveTopology>(parameters[0]); | ||
| 127 | const u32 vertex_first = parameters[3]; | 114 | const u32 vertex_first = parameters[3]; |
| 128 | const u32 vertex_count = parameters[1]; | 115 | const u32 vertex_count = parameters[1]; |
| 129 | |||
| 130 | 116 | ||
| 131 | if (maxwell3d.AnyParametersDirty() && | 117 | if (!IsTopologySafe(topology) && |
| 132 | maxwell3d.GetMaxCurrentVertices() < vertex_first + vertex_count) { | 118 | static_cast<size_t>(maxwell3d.GetMaxCurrentVertices()) < |
| 119 | static_cast<size_t>(vertex_first) + static_cast<size_t>(vertex_count)) { | ||
| 133 | ASSERT_MSG(false, "Faulty draw!"); | 120 | ASSERT_MSG(false, "Faulty draw!"); |
| 134 | return; | 121 | return; |
| 135 | } | 122 | } |
| @@ -141,9 +128,8 @@ private: | |||
| 141 | maxwell3d.setHLEReplacementName(0, 0x640, Maxwell::HLEReplaceName::BaseInstance); | 128 | maxwell3d.setHLEReplacementName(0, 0x640, Maxwell::HLEReplaceName::BaseInstance); |
| 142 | } | 129 | } |
| 143 | 130 | ||
| 144 | maxwell3d.draw_manager->DrawArray( | 131 | maxwell3d.draw_manager->DrawArray(topology, vertex_first, vertex_count, base_instance, |
| 145 | static_cast<Tegra::Engines::Maxwell3D::Regs::PrimitiveTopology>(parameters[0]), | 132 | instance_count); |
| 146 | vertex_first, vertex_count, base_instance, instance_count); | ||
| 147 | 133 | ||
| 148 | if (extended) { | 134 | if (extended) { |
| 149 | maxwell3d.regs.global_base_instance_index = 0; | 135 | maxwell3d.regs.global_base_instance_index = 0; |
| @@ -166,13 +152,7 @@ public: | |||
| 166 | return; | 152 | return; |
| 167 | } | 153 | } |
| 168 | 154 | ||
| 169 | advanceCheck(); | ||
| 170 | if (check_limit) { | ||
| 171 | maxwell3d.RefreshParameters(); | ||
| 172 | minimum_limit = std::max(parameters[3], minimum_limit); | ||
| 173 | } | ||
| 174 | const u32 estimate = static_cast<u32>(maxwell3d.EstimateIndexBufferSize()); | 155 | const u32 estimate = static_cast<u32>(maxwell3d.EstimateIndexBufferSize()); |
| 175 | const u32 base_size = std::max<u32>(minimum_limit, estimate); | ||
| 176 | const u32 element_base = parameters[4]; | 156 | const u32 element_base = parameters[4]; |
| 177 | const u32 base_instance = parameters[5]; | 157 | const u32 base_instance = parameters[5]; |
| 178 | maxwell3d.regs.vertex_id_base = element_base; | 158 | maxwell3d.regs.vertex_id_base = element_base; |
| @@ -191,7 +171,7 @@ public: | |||
| 191 | params.max_draw_counts = 1; | 171 | params.max_draw_counts = 1; |
| 192 | params.stride = 0; | 172 | params.stride = 0; |
| 193 | maxwell3d.dirty.flags[VideoCommon::Dirty::IndexBuffer] = true; | 173 | maxwell3d.dirty.flags[VideoCommon::Dirty::IndexBuffer] = true; |
| 194 | maxwell3d.draw_manager->DrawIndexedIndirect(topology, 0, base_size); | 174 | maxwell3d.draw_manager->DrawIndexedIndirect(topology, 0, estimate); |
| 195 | maxwell3d.engine_state = Maxwell::EngineHint::None; | 175 | maxwell3d.engine_state = Maxwell::EngineHint::None; |
| 196 | maxwell3d.replace_table.clear(); | 176 | maxwell3d.replace_table.clear(); |
| 197 | maxwell3d.regs.vertex_id_base = 0x0; | 177 | maxwell3d.regs.vertex_id_base = 0x0; |
| @@ -223,8 +203,6 @@ private: | |||
| 223 | maxwell3d.engine_state = Maxwell::EngineHint::None; | 203 | maxwell3d.engine_state = Maxwell::EngineHint::None; |
| 224 | maxwell3d.replace_table.clear(); | 204 | maxwell3d.replace_table.clear(); |
| 225 | } | 205 | } |
| 226 | |||
| 227 | u32 minimum_limit{1 << 18}; | ||
| 228 | }; | 206 | }; |
| 229 | 207 | ||
| 230 | class HLE_MultiLayerClear final : public HLEMacroImpl { | 208 | class HLE_MultiLayerClear final : public HLEMacroImpl { |
| @@ -257,10 +235,6 @@ public: | |||
| 257 | return; | 235 | return; |
| 258 | } | 236 | } |
| 259 | 237 | ||
| 260 | advanceCheck(); | ||
| 261 | if (check_limit) { | ||
| 262 | maxwell3d.RefreshParameters(); | ||
| 263 | } | ||
| 264 | const u32 start_indirect = parameters[0]; | 238 | const u32 start_indirect = parameters[0]; |
| 265 | const u32 end_indirect = parameters[1]; | 239 | const u32 end_indirect = parameters[1]; |
| 266 | if (start_indirect >= end_indirect) { | 240 | if (start_indirect >= end_indirect) { |
| @@ -274,20 +248,7 @@ public: | |||
| 274 | const u32 indirect_words = 5 + padding; | 248 | const u32 indirect_words = 5 + padding; |
| 275 | const u32 stride = indirect_words * sizeof(u32); | 249 | const u32 stride = indirect_words * sizeof(u32); |
| 276 | const std::size_t draw_count = end_indirect - start_indirect; | 250 | const std::size_t draw_count = end_indirect - start_indirect; |
| 277 | u32 lowest_first = std::numeric_limits<u32>::max(); | ||
| 278 | u32 highest_limit = std::numeric_limits<u32>::min(); | ||
| 279 | for (std::size_t index = 0; index < draw_count; index++) { | ||
| 280 | const std::size_t base = index * indirect_words + 5; | ||
| 281 | const u32 count = parameters[base]; | ||
| 282 | const u32 first_index = parameters[base + 2]; | ||
| 283 | lowest_first = std::min(lowest_first, first_index); | ||
| 284 | highest_limit = std::max(highest_limit, first_index + count); | ||
| 285 | } | ||
| 286 | if (check_limit) { | ||
| 287 | minimum_limit = std::max(highest_limit, minimum_limit); | ||
| 288 | } | ||
| 289 | const u32 estimate = static_cast<u32>(maxwell3d.EstimateIndexBufferSize()); | 251 | const u32 estimate = static_cast<u32>(maxwell3d.EstimateIndexBufferSize()); |
| 290 | const u32 base_size = std::max(minimum_limit, estimate); | ||
| 291 | maxwell3d.dirty.flags[VideoCommon::Dirty::IndexBuffer] = true; | 252 | maxwell3d.dirty.flags[VideoCommon::Dirty::IndexBuffer] = true; |
| 292 | auto& params = maxwell3d.draw_manager->GetIndirectParams(); | 253 | auto& params = maxwell3d.draw_manager->GetIndirectParams(); |
| 293 | params.is_indexed = true; | 254 | params.is_indexed = true; |
| @@ -301,7 +262,7 @@ public: | |||
| 301 | maxwell3d.engine_state = Maxwell::EngineHint::OnHLEMacro; | 262 | maxwell3d.engine_state = Maxwell::EngineHint::OnHLEMacro; |
| 302 | maxwell3d.setHLEReplacementName(0, 0x640, Maxwell::HLEReplaceName::BaseVertex); | 263 | maxwell3d.setHLEReplacementName(0, 0x640, Maxwell::HLEReplaceName::BaseVertex); |
| 303 | maxwell3d.setHLEReplacementName(0, 0x644, Maxwell::HLEReplaceName::BaseInstance); | 264 | maxwell3d.setHLEReplacementName(0, 0x644, Maxwell::HLEReplaceName::BaseInstance); |
| 304 | maxwell3d.draw_manager->DrawIndexedIndirect(topology, 0, base_size); | 265 | maxwell3d.draw_manager->DrawIndexedIndirect(topology, 0, estimate); |
| 305 | maxwell3d.engine_state = Maxwell::EngineHint::None; | 266 | maxwell3d.engine_state = Maxwell::EngineHint::None; |
| 306 | maxwell3d.replace_table.clear(); | 267 | maxwell3d.replace_table.clear(); |
| 307 | } | 268 | } |
| @@ -323,7 +284,6 @@ private: | |||
| 323 | return; | 284 | return; |
| 324 | } | 285 | } |
| 325 | const auto topology = static_cast<Maxwell::Regs::PrimitiveTopology>(parameters[2]); | 286 | const auto topology = static_cast<Maxwell::Regs::PrimitiveTopology>(parameters[2]); |
| 326 | maxwell3d.regs.draw.topology.Assign(topology); | ||
| 327 | const u32 padding = parameters[3]; | 287 | const u32 padding = parameters[3]; |
| 328 | const std::size_t max_draws = parameters[4]; | 288 | const std::size_t max_draws = parameters[4]; |
| 329 | 289 | ||
| @@ -345,8 +305,6 @@ private: | |||
| 345 | base_vertex, base_instance, parameters[base + 1]); | 305 | base_vertex, base_instance, parameters[base + 1]); |
| 346 | } | 306 | } |
| 347 | } | 307 | } |
| 348 | |||
| 349 | u32 minimum_limit{1 << 12}; | ||
| 350 | }; | 308 | }; |
| 351 | 309 | ||
| 352 | class HLE_C713C83D8F63CCF3 final : public HLEMacroImpl { | 310 | class HLE_C713C83D8F63CCF3 final : public HLEMacroImpl { |
| @@ -431,53 +389,53 @@ public: | |||
| 431 | HLEMacro::HLEMacro(Engines::Maxwell3D& maxwell3d_) : maxwell3d{maxwell3d_} { | 389 | HLEMacro::HLEMacro(Engines::Maxwell3D& maxwell3d_) : maxwell3d{maxwell3d_} { |
| 432 | builders.emplace(0x771BB18C62444DA0ULL, | 390 | builders.emplace(0x771BB18C62444DA0ULL, |
| 433 | std::function<std::unique_ptr<CachedMacro>(Engines::Maxwell3D&)>( | 391 | std::function<std::unique_ptr<CachedMacro>(Engines::Maxwell3D&)>( |
| 434 | [](Engines::Maxwell3D& maxwell3d) -> std::unique_ptr<CachedMacro> { | 392 | [](Engines::Maxwell3D& maxwell3d__) -> std::unique_ptr<CachedMacro> { |
| 435 | return std::make_unique<HLE_771BB18C62444DA0>(maxwell3d); | 393 | return std::make_unique<HLE_771BB18C62444DA0>(maxwell3d__); |
| 436 | })); | 394 | })); |
| 437 | builders.emplace(0x0D61FC9FAAC9FCADULL, | 395 | builders.emplace(0x0D61FC9FAAC9FCADULL, |
| 438 | std::function<std::unique_ptr<CachedMacro>(Engines::Maxwell3D&)>( | 396 | std::function<std::unique_ptr<CachedMacro>(Engines::Maxwell3D&)>( |
| 439 | [](Engines::Maxwell3D& maxwell3d) -> std::unique_ptr<CachedMacro> { | 397 | [](Engines::Maxwell3D& maxwell3d__) -> std::unique_ptr<CachedMacro> { |
| 440 | return std::make_unique<HLE_DrawArraysIndirect>(maxwell3d); | 398 | return std::make_unique<HLE_DrawArraysIndirect>(maxwell3d__); |
| 441 | })); | 399 | })); |
| 442 | builders.emplace(0x8A4D173EB99A8603ULL, | 400 | builders.emplace(0x8A4D173EB99A8603ULL, |
| 443 | std::function<std::unique_ptr<CachedMacro>(Engines::Maxwell3D&)>( | 401 | std::function<std::unique_ptr<CachedMacro>(Engines::Maxwell3D&)>( |
| 444 | [](Engines::Maxwell3D& maxwell3d) -> std::unique_ptr<CachedMacro> { | 402 | [](Engines::Maxwell3D& maxwell3d__) -> std::unique_ptr<CachedMacro> { |
| 445 | return std::make_unique<HLE_DrawArraysIndirect>(maxwell3d, true); | 403 | return std::make_unique<HLE_DrawArraysIndirect>(maxwell3d__, true); |
| 446 | })); | 404 | })); |
| 447 | builders.emplace(0x0217920100488FF7ULL, | 405 | builders.emplace(0x0217920100488FF7ULL, |
| 448 | std::function<std::unique_ptr<CachedMacro>(Engines::Maxwell3D&)>( | 406 | std::function<std::unique_ptr<CachedMacro>(Engines::Maxwell3D&)>( |
| 449 | [](Engines::Maxwell3D& maxwell3d) -> std::unique_ptr<CachedMacro> { | 407 | [](Engines::Maxwell3D& maxwell3d__) -> std::unique_ptr<CachedMacro> { |
| 450 | return std::make_unique<HLE_DrawIndexedIndirect>(maxwell3d); | 408 | return std::make_unique<HLE_DrawIndexedIndirect>(maxwell3d__); |
| 451 | })); | 409 | })); |
| 452 | builders.emplace(0x3F5E74B9C9A50164ULL, | 410 | builders.emplace(0x3F5E74B9C9A50164ULL, |
| 453 | std::function<std::unique_ptr<CachedMacro>(Engines::Maxwell3D&)>( | 411 | std::function<std::unique_ptr<CachedMacro>(Engines::Maxwell3D&)>( |
| 454 | [](Engines::Maxwell3D& maxwell3d) -> std::unique_ptr<CachedMacro> { | 412 | [](Engines::Maxwell3D& maxwell3d__) -> std::unique_ptr<CachedMacro> { |
| 455 | return std::make_unique<HLE_MultiDrawIndexedIndirectCount>(maxwell3d); | 413 | return std::make_unique<HLE_MultiDrawIndexedIndirectCount>(maxwell3d__); |
| 456 | })); | 414 | })); |
| 457 | builders.emplace(0xEAD26C3E2109B06BULL, | 415 | builders.emplace(0xEAD26C3E2109B06BULL, |
| 458 | std::function<std::unique_ptr<CachedMacro>(Engines::Maxwell3D&)>( | 416 | std::function<std::unique_ptr<CachedMacro>(Engines::Maxwell3D&)>( |
| 459 | [](Engines::Maxwell3D& maxwell3d) -> std::unique_ptr<CachedMacro> { | 417 | [](Engines::Maxwell3D& maxwell3d__) -> std::unique_ptr<CachedMacro> { |
| 460 | return std::make_unique<HLE_MultiLayerClear>(maxwell3d); | 418 | return std::make_unique<HLE_MultiLayerClear>(maxwell3d__); |
| 461 | })); | 419 | })); |
| 462 | builders.emplace(0xC713C83D8F63CCF3ULL, | 420 | builders.emplace(0xC713C83D8F63CCF3ULL, |
| 463 | std::function<std::unique_ptr<CachedMacro>(Engines::Maxwell3D&)>( | 421 | std::function<std::unique_ptr<CachedMacro>(Engines::Maxwell3D&)>( |
| 464 | [](Engines::Maxwell3D& maxwell3d) -> std::unique_ptr<CachedMacro> { | 422 | [](Engines::Maxwell3D& maxwell3d__) -> std::unique_ptr<CachedMacro> { |
| 465 | return std::make_unique<HLE_C713C83D8F63CCF3>(maxwell3d); | 423 | return std::make_unique<HLE_C713C83D8F63CCF3>(maxwell3d__); |
| 466 | })); | 424 | })); |
| 467 | builders.emplace(0xD7333D26E0A93EDEULL, | 425 | builders.emplace(0xD7333D26E0A93EDEULL, |
| 468 | std::function<std::unique_ptr<CachedMacro>(Engines::Maxwell3D&)>( | 426 | std::function<std::unique_ptr<CachedMacro>(Engines::Maxwell3D&)>( |
| 469 | [](Engines::Maxwell3D& maxwell3d) -> std::unique_ptr<CachedMacro> { | 427 | [](Engines::Maxwell3D& maxwell3d__) -> std::unique_ptr<CachedMacro> { |
| 470 | return std::make_unique<HLE_D7333D26E0A93EDE>(maxwell3d); | 428 | return std::make_unique<HLE_D7333D26E0A93EDE>(maxwell3d__); |
| 471 | })); | 429 | })); |
| 472 | builders.emplace(0xEB29B2A09AA06D38ULL, | 430 | builders.emplace(0xEB29B2A09AA06D38ULL, |
| 473 | std::function<std::unique_ptr<CachedMacro>(Engines::Maxwell3D&)>( | 431 | std::function<std::unique_ptr<CachedMacro>(Engines::Maxwell3D&)>( |
| 474 | [](Engines::Maxwell3D& maxwell3d) -> std::unique_ptr<CachedMacro> { | 432 | [](Engines::Maxwell3D& maxwell3d__) -> std::unique_ptr<CachedMacro> { |
| 475 | return std::make_unique<HLE_BindShader>(maxwell3d); | 433 | return std::make_unique<HLE_BindShader>(maxwell3d__); |
| 476 | })); | 434 | })); |
| 477 | builders.emplace(0xDB1341DBEB4C8AF7ULL, | 435 | builders.emplace(0xDB1341DBEB4C8AF7ULL, |
| 478 | std::function<std::unique_ptr<CachedMacro>(Engines::Maxwell3D&)>( | 436 | std::function<std::unique_ptr<CachedMacro>(Engines::Maxwell3D&)>( |
| 479 | [](Engines::Maxwell3D& maxwell3d) -> std::unique_ptr<CachedMacro> { | 437 | [](Engines::Maxwell3D& maxwell3d__) -> std::unique_ptr<CachedMacro> { |
| 480 | return std::make_unique<HLE_SetRasterBoundingBox>(maxwell3d); | 438 | return std::make_unique<HLE_SetRasterBoundingBox>(maxwell3d__); |
| 481 | })); | 439 | })); |
| 482 | } | 440 | } |
| 483 | 441 | ||