diff options
Diffstat (limited to 'src')
38 files changed, 575 insertions, 127 deletions
diff --git a/src/common/settings.h b/src/common/settings.h index cfc1ab46f..a88ee045d 100644 --- a/src/common/settings.h +++ b/src/common/settings.h @@ -319,6 +319,7 @@ struct Values { // Renderer Setting<RendererBackend> renderer_backend{RendererBackend::OpenGL, "backend"}; BasicSetting<bool> renderer_debug{false, "debug"}; + BasicSetting<bool> renderer_shader_feedback{false, "shader_feedback"}; BasicSetting<bool> enable_nsight_aftermath{false, "nsight_aftermath"}; BasicSetting<bool> disable_shader_loop_safety_checks{false, "disable_shader_loop_safety_checks"}; diff --git a/src/common/uuid.cpp b/src/common/uuid.cpp index 26db03fba..18303a1e3 100644 --- a/src/common/uuid.cpp +++ b/src/common/uuid.cpp @@ -18,7 +18,7 @@ UUID UUID::Generate() { } std::string UUID::Format() const { - return fmt::format("0x{:016X}{:016X}", uuid[1], uuid[0]); + return fmt::format("{:016x}{:016x}", uuid[1], uuid[0]); } std::string UUID::FormatSwitch() const { diff --git a/src/core/hle/api_version.h b/src/core/hle/api_version.h index 5e10a7ad9..43d5670a9 100644 --- a/src/core/hle/api_version.h +++ b/src/core/hle/api_version.h @@ -12,9 +12,9 @@ namespace HLE::ApiVersion { // Horizon OS version constants. -constexpr u8 HOS_VERSION_MAJOR = 11; -constexpr u8 HOS_VERSION_MINOR = 0; -constexpr u8 HOS_VERSION_MICRO = 1; +constexpr u8 HOS_VERSION_MAJOR = 12; +constexpr u8 HOS_VERSION_MINOR = 1; +constexpr u8 HOS_VERSION_MICRO = 0; // NintendoSDK version constants. @@ -22,15 +22,15 @@ constexpr u8 SDK_REVISION_MAJOR = 1; constexpr u8 SDK_REVISION_MINOR = 0; constexpr char PLATFORM_STRING[] = "NX"; -constexpr char VERSION_HASH[] = "69103fcb2004dace877094c2f8c29e6113be5dbf"; -constexpr char DISPLAY_VERSION[] = "11.0.1"; -constexpr char DISPLAY_TITLE[] = "NintendoSDK Firmware for NX 11.0.1-1.0"; +constexpr char VERSION_HASH[] = "76b10c2dab7d3aa73fc162f8dff1655e6a21caf4"; +constexpr char DISPLAY_VERSION[] = "12.1.0"; +constexpr char DISPLAY_TITLE[] = "NintendoSDK Firmware for NX 12.1.0-1.0"; // Atmosphere version constants. constexpr u8 ATMOSPHERE_RELEASE_VERSION_MAJOR = 0; constexpr u8 ATMOSPHERE_RELEASE_VERSION_MINOR = 19; -constexpr u8 ATMOSPHERE_RELEASE_VERSION_MICRO = 4; +constexpr u8 ATMOSPHERE_RELEASE_VERSION_MICRO = 5; constexpr u32 GetTargetFirmware() { return u32{HOS_VERSION_MAJOR} << 24 | u32{HOS_VERSION_MINOR} << 16 | diff --git a/src/core/hle/service/acc/acc.cpp b/src/core/hle/service/acc/acc.cpp index 2e969f2a8..882fc1492 100644 --- a/src/core/hle/service/acc/acc.cpp +++ b/src/core/hle/service/acc/acc.cpp @@ -292,7 +292,7 @@ public: protected: void Get(Kernel::HLERequestContext& ctx) { - LOG_DEBUG(Service_ACC, "called user_id={}", user_id.Format()); + LOG_DEBUG(Service_ACC, "called user_id=0x{}", user_id.Format()); ProfileBase profile_base{}; ProfileData data{}; if (profile_manager.GetProfileBaseAndData(user_id, profile_base, data)) { @@ -301,7 +301,7 @@ protected: rb.Push(ResultSuccess); rb.PushRaw(profile_base); } else { - LOG_ERROR(Service_ACC, "Failed to get profile base and data for user={}", + LOG_ERROR(Service_ACC, "Failed to get profile base and data for user=0x{}", user_id.Format()); IPC::ResponseBuilder rb{ctx, 2}; rb.Push(ResultUnknown); // TODO(ogniK): Get actual error code @@ -309,14 +309,14 @@ protected: } void GetBase(Kernel::HLERequestContext& ctx) { - LOG_DEBUG(Service_ACC, "called user_id={}", user_id.Format()); + LOG_DEBUG(Service_ACC, "called user_id=0x{}", user_id.Format()); ProfileBase profile_base{}; if (profile_manager.GetProfileBase(user_id, profile_base)) { IPC::ResponseBuilder rb{ctx, 16}; rb.Push(ResultSuccess); rb.PushRaw(profile_base); } else { - LOG_ERROR(Service_ACC, "Failed to get profile base for user={}", user_id.Format()); + LOG_ERROR(Service_ACC, "Failed to get profile base for user=0x{}", user_id.Format()); IPC::ResponseBuilder rb{ctx, 2}; rb.Push(ResultUnknown); // TODO(ogniK): Get actual error code } @@ -372,7 +372,7 @@ protected: const auto user_data = ctx.ReadBuffer(); - LOG_DEBUG(Service_ACC, "called, username='{}', timestamp={:016X}, uuid={}", + LOG_DEBUG(Service_ACC, "called, username='{}', timestamp={:016X}, uuid=0x{}", Common::StringFromFixedZeroTerminatedBuffer( reinterpret_cast<const char*>(base.username.data()), base.username.size()), base.timestamp, base.user_uuid.Format()); @@ -405,7 +405,7 @@ protected: const auto user_data = ctx.ReadBuffer(); const auto image_data = ctx.ReadBuffer(1); - LOG_DEBUG(Service_ACC, "called, username='{}', timestamp={:016X}, uuid={}", + LOG_DEBUG(Service_ACC, "called, username='{}', timestamp={:016X}, uuid=0x{}", Common::StringFromFixedZeroTerminatedBuffer( reinterpret_cast<const char*>(base.username.data()), base.username.size()), base.timestamp, base.user_uuid.Format()); @@ -662,7 +662,7 @@ void Module::Interface::GetUserCount(Kernel::HLERequestContext& ctx) { void Module::Interface::GetUserExistence(Kernel::HLERequestContext& ctx) { IPC::RequestParser rp{ctx}; Common::UUID user_id = rp.PopRaw<Common::UUID>(); - LOG_DEBUG(Service_ACC, "called user_id={}", user_id.Format()); + LOG_DEBUG(Service_ACC, "called user_id=0x{}", user_id.Format()); IPC::ResponseBuilder rb{ctx, 3}; rb.Push(ResultSuccess); @@ -693,7 +693,7 @@ void Module::Interface::GetLastOpenedUser(Kernel::HLERequestContext& ctx) { void Module::Interface::GetProfile(Kernel::HLERequestContext& ctx) { IPC::RequestParser rp{ctx}; Common::UUID user_id = rp.PopRaw<Common::UUID>(); - LOG_DEBUG(Service_ACC, "called user_id={}", user_id.Format()); + LOG_DEBUG(Service_ACC, "called user_id=0x{}", user_id.Format()); IPC::ResponseBuilder rb{ctx, 2, 0, 1}; rb.Push(ResultSuccess); @@ -802,7 +802,7 @@ void Module::Interface::GetProfileEditor(Kernel::HLERequestContext& ctx) { IPC::RequestParser rp{ctx}; Common::UUID user_id = rp.PopRaw<Common::UUID>(); - LOG_DEBUG(Service_ACC, "called, user_id={}", user_id.Format()); + LOG_DEBUG(Service_ACC, "called, user_id=0x{}", user_id.Format()); IPC::ResponseBuilder rb{ctx, 2, 0, 1}; rb.Push(ResultSuccess); @@ -844,7 +844,7 @@ void Module::Interface::StoreSaveDataThumbnailApplication(Kernel::HLERequestCont IPC::RequestParser rp{ctx}; const auto uuid = rp.PopRaw<Common::UUID>(); - LOG_WARNING(Service_ACC, "(STUBBED) called, uuid={}", uuid.Format()); + LOG_WARNING(Service_ACC, "(STUBBED) called, uuid=0x{}", uuid.Format()); // TODO(ogniK): Check if application ID is zero on acc initialize. As we don't have a reliable // way of confirming things like the TID, we're going to assume a non zero value for the time @@ -858,7 +858,7 @@ void Module::Interface::StoreSaveDataThumbnailSystem(Kernel::HLERequestContext& const auto uuid = rp.PopRaw<Common::UUID>(); const auto tid = rp.Pop<u64_le>(); - LOG_WARNING(Service_ACC, "(STUBBED) called, uuid={}, tid={:016X}", uuid.Format(), tid); + LOG_WARNING(Service_ACC, "(STUBBED) called, uuid=0x{}, tid={:016X}", uuid.Format(), tid); StoreSaveDataThumbnail(ctx, uuid, tid); } diff --git a/src/core/hle/service/am/applets/applet_software_keyboard.cpp b/src/core/hle/service/am/applets/applet_software_keyboard.cpp index 7cae90609..673abb755 100644 --- a/src/core/hle/service/am/applets/applet_software_keyboard.cpp +++ b/src/core/hle/service/am/applets/applet_software_keyboard.cpp @@ -377,7 +377,7 @@ void SoftwareKeyboard::SubmitForTextCheck(std::u16string submitted_text) { if (swkbd_config_common.use_utf8) { std::string utf8_submitted_text = Common::UTF16ToUTF8(current_text); - const u64 buffer_size = sizeof(u64) + utf8_submitted_text.size(); + const u64 buffer_size = utf8_submitted_text.size(); LOG_DEBUG(Service_AM, "\nBuffer Size: {}\nUTF-8 Submitted Text: {}", buffer_size, utf8_submitted_text); @@ -386,7 +386,7 @@ void SoftwareKeyboard::SubmitForTextCheck(std::u16string submitted_text) { std::memcpy(out_data.data() + sizeof(u64), utf8_submitted_text.data(), utf8_submitted_text.size()); } else { - const u64 buffer_size = sizeof(u64) + current_text.size() * sizeof(char16_t); + const u64 buffer_size = current_text.size() * sizeof(char16_t); LOG_DEBUG(Service_AM, "\nBuffer Size: {}\nUTF-16 Submitted Text: {}", buffer_size, Common::UTF16ToUTF8(current_text)); diff --git a/src/core/hle/service/friend/friend.cpp b/src/core/hle/service/friend/friend.cpp index a3c939c0c..b58c152ce 100644 --- a/src/core/hle/service/friend/friend.cpp +++ b/src/core/hle/service/friend/friend.cpp @@ -158,7 +158,7 @@ private: const auto local_play = rp.Pop<bool>(); const auto uuid = rp.PopRaw<Common::UUID>(); - LOG_WARNING(Service_Friend, "(STUBBED) called local_play={} uuid={}", local_play, + LOG_WARNING(Service_Friend, "(STUBBED) called, local_play={}, uuid=0x{}", local_play, uuid.Format()); IPC::ResponseBuilder rb{ctx, 2}; @@ -171,7 +171,7 @@ private: const auto uuid = rp.PopRaw<Common::UUID>(); [[maybe_unused]] const auto filter = rp.PopRaw<SizedFriendFilter>(); const auto pid = rp.Pop<u64>(); - LOG_WARNING(Service_Friend, "(STUBBED) called, offset={}, uuid={}, pid={}", friend_offset, + LOG_WARNING(Service_Friend, "(STUBBED) called, offset={}, uuid=0x{}, pid={}", friend_offset, uuid.Format(), pid); IPC::ResponseBuilder rb{ctx, 3}; @@ -289,7 +289,7 @@ void Module::Interface::CreateNotificationService(Kernel::HLERequestContext& ctx IPC::RequestParser rp{ctx}; auto uuid = rp.PopRaw<Common::UUID>(); - LOG_DEBUG(Service_Friend, "called, uuid={}", uuid.Format()); + LOG_DEBUG(Service_Friend, "called, uuid=0x{}", uuid.Format()); IPC::ResponseBuilder rb{ctx, 2, 0, 1}; rb.Push(ResultSuccess); diff --git a/src/core/hle/service/ns/language.cpp b/src/core/hle/service/ns/language.cpp index 29c4a820c..7d9e4a20b 100644 --- a/src/core/hle/service/ns/language.cpp +++ b/src/core/hle/service/ns/language.cpp @@ -339,13 +339,16 @@ std::optional<ApplicationLanguage> ConvertToApplicationLanguage( case Set::LanguageCode::FR_CA: return ApplicationLanguage::CanadianFrench; case Set::LanguageCode::PT: + case Set::LanguageCode::PT_BR: return ApplicationLanguage::Portuguese; case Set::LanguageCode::RU: return ApplicationLanguage::Russian; case Set::LanguageCode::KO: return ApplicationLanguage::Korean; + case Set::LanguageCode::ZH_TW: case Set::LanguageCode::ZH_HANT: return ApplicationLanguage::TraditionalChinese; + case Set::LanguageCode::ZH_CN: case Set::LanguageCode::ZH_HANS: return ApplicationLanguage::SimplifiedChinese; default: diff --git a/src/core/hle/service/ns/ns_language.h b/src/core/hle/service/ns/ns_language.h deleted file mode 100644 index 59ac85a19..000000000 --- a/src/core/hle/service/ns/ns_language.h +++ /dev/null @@ -1,42 +0,0 @@ -// Copyright 2019 yuzu emulator team -// Licensed under GPLv2 or any later version -// Refer to the license.txt file included. - -#pragma once -#include <optional> -#include <string> -#include "common/common_types.h" -#include "core/hle/service/set/set.h" - -namespace Service::NS { -/// This is nn::ns::detail::ApplicationLanguage -enum class ApplicationLanguage : u8 { - AmericanEnglish = 0, - BritishEnglish, - Japanese, - French, - German, - LatinAmericanSpanish, - Spanish, - Italian, - Dutch, - CanadianFrench, - Portuguese, - Russian, - Korean, - TraditionalChinese, - SimplifiedChinese, - Count -}; -using ApplicationLanguagePriorityList = - const std::array<ApplicationLanguage, static_cast<std::size_t>(ApplicationLanguage::Count)>; - -constexpr u32 GetSupportedLanguageFlag(const ApplicationLanguage lang) { - return 1U << static_cast<u32>(lang); -} - -const ApplicationLanguagePriorityList* GetApplicationLanguagePriorityList(ApplicationLanguage lang); -std::optional<ApplicationLanguage> ConvertToApplicationLanguage( - Service::Set::LanguageCode language_code); -std::optional<Service::Set::LanguageCode> ConvertToLanguageCode(ApplicationLanguage lang); -} // namespace Service::NS
\ No newline at end of file diff --git a/src/core/hle/service/set/set.cpp b/src/core/hle/service/set/set.cpp index 522a604a5..f2e2e8306 100644 --- a/src/core/hle/service/set/set.cpp +++ b/src/core/hle/service/set/set.cpp @@ -12,7 +12,7 @@ namespace Service::Set { namespace { -constexpr std::array<LanguageCode, 17> available_language_codes = {{ +constexpr std::array<LanguageCode, 18> available_language_codes = {{ LanguageCode::JA, LanguageCode::EN_US, LanguageCode::FR, @@ -30,6 +30,7 @@ constexpr std::array<LanguageCode, 17> available_language_codes = {{ LanguageCode::ES_419, LanguageCode::ZH_HANS, LanguageCode::ZH_HANT, + LanguageCode::PT_BR, }}; enum class KeyboardLayout : u64 { @@ -50,7 +51,7 @@ enum class KeyboardLayout : u64 { ChineseTraditional = 14, }; -constexpr std::array<std::pair<LanguageCode, KeyboardLayout>, 17> language_to_layout{{ +constexpr std::array<std::pair<LanguageCode, KeyboardLayout>, 18> language_to_layout{{ {LanguageCode::JA, KeyboardLayout::Japanese}, {LanguageCode::EN_US, KeyboardLayout::EnglishUs}, {LanguageCode::FR, KeyboardLayout::French}, @@ -68,10 +69,11 @@ constexpr std::array<std::pair<LanguageCode, KeyboardLayout>, 17> language_to_la {LanguageCode::ES_419, KeyboardLayout::SpanishLatin}, {LanguageCode::ZH_HANS, KeyboardLayout::ChineseSimplified}, {LanguageCode::ZH_HANT, KeyboardLayout::ChineseTraditional}, + {LanguageCode::PT_BR, KeyboardLayout::Portuguese}, }}; -constexpr std::size_t pre4_0_0_max_entries = 15; -constexpr std::size_t post4_0_0_max_entries = 17; +constexpr std::size_t PRE_4_0_0_MAX_ENTRIES = 0xF; +constexpr std::size_t POST_4_0_0_MAX_ENTRIES = 0x40; constexpr ResultCode ERR_INVALID_LANGUAGE{ErrorModule::Settings, 625}; @@ -81,9 +83,9 @@ void PushResponseLanguageCode(Kernel::HLERequestContext& ctx, std::size_t num_la rb.Push(static_cast<u32>(num_language_codes)); } -void GetAvailableLanguageCodesImpl(Kernel::HLERequestContext& ctx, std::size_t max_size) { +void GetAvailableLanguageCodesImpl(Kernel::HLERequestContext& ctx, std::size_t max_entries) { const std::size_t requested_amount = ctx.GetWriteBufferSize() / sizeof(LanguageCode); - const std::size_t copy_amount = std::min(requested_amount, max_size); + const std::size_t copy_amount = std::min(requested_amount, max_entries); const std::size_t copy_size = copy_amount * sizeof(LanguageCode); ctx.WriteBuffer(available_language_codes.data(), copy_size); @@ -118,7 +120,7 @@ LanguageCode GetLanguageCodeFromIndex(std::size_t index) { void SET::GetAvailableLanguageCodes(Kernel::HLERequestContext& ctx) { LOG_DEBUG(Service_SET, "called"); - GetAvailableLanguageCodesImpl(ctx, pre4_0_0_max_entries); + GetAvailableLanguageCodesImpl(ctx, PRE_4_0_0_MAX_ENTRIES); } void SET::MakeLanguageCode(Kernel::HLERequestContext& ctx) { @@ -140,19 +142,19 @@ void SET::MakeLanguageCode(Kernel::HLERequestContext& ctx) { void SET::GetAvailableLanguageCodes2(Kernel::HLERequestContext& ctx) { LOG_DEBUG(Service_SET, "called"); - GetAvailableLanguageCodesImpl(ctx, post4_0_0_max_entries); + GetAvailableLanguageCodesImpl(ctx, POST_4_0_0_MAX_ENTRIES); } void SET::GetAvailableLanguageCodeCount(Kernel::HLERequestContext& ctx) { LOG_DEBUG(Service_SET, "called"); - PushResponseLanguageCode(ctx, pre4_0_0_max_entries); + PushResponseLanguageCode(ctx, PRE_4_0_0_MAX_ENTRIES); } void SET::GetAvailableLanguageCodeCount2(Kernel::HLERequestContext& ctx) { LOG_DEBUG(Service_SET, "called"); - PushResponseLanguageCode(ctx, post4_0_0_max_entries); + PushResponseLanguageCode(ctx, POST_4_0_0_MAX_ENTRIES); } void SET::GetQuestFlag(Kernel::HLERequestContext& ctx) { diff --git a/src/core/hle/service/set/set.h b/src/core/hle/service/set/set.h index d5bd7828d..acabebeaa 100644 --- a/src/core/hle/service/set/set.h +++ b/src/core/hle/service/set/set.h @@ -31,6 +31,7 @@ enum class LanguageCode : u64 { ES_419 = 0x00003931342D7365, ZH_HANS = 0x00736E61482D687A, ZH_HANT = 0x00746E61482D687A, + PT_BR = 0x00000052422D7470, }; LanguageCode GetLanguageCodeFromIndex(std::size_t idx); diff --git a/src/shader_recompiler/frontend/ir/value.h b/src/shader_recompiler/frontend/ir/value.h index 795194d41..334bb47aa 100644 --- a/src/shader_recompiler/frontend/ir/value.h +++ b/src/shader_recompiler/frontend/ir/value.h @@ -57,6 +57,7 @@ public: [[nodiscard]] IR::Inst* Inst() const; [[nodiscard]] IR::Inst* InstRecursive() const; + [[nodiscard]] IR::Inst* TryInstRecursive() const; [[nodiscard]] IR::Value Resolve() const; [[nodiscard]] IR::Reg Reg() const; [[nodiscard]] IR::Pred Pred() const; @@ -308,6 +309,13 @@ inline IR::Inst* Value::InstRecursive() const { return inst; } +inline IR::Inst* Value::TryInstRecursive() const { + if (IsIdentity()) { + return inst->Arg(0).TryInstRecursive(); + } + return type == Type::Opaque ? inst : nullptr; +} + inline IR::Value Value::Resolve() const { if (IsIdentity()) { return inst->Arg(0).Resolve(); diff --git a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp index 8dd6d6c2c..d089fdd12 100644 --- a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp +++ b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp @@ -3,6 +3,7 @@ // Refer to the license.txt file included. #include <algorithm> +#include <functional> #include <tuple> #include <type_traits> @@ -88,6 +89,26 @@ bool FoldWhenAllImmediates(IR::Inst& inst, Func&& func) { return true; } +/// Return true when all values in a range are equal +template <typename Range> +bool AreEqual(const Range& range) { + auto resolver{[](const auto& value) { return value.Resolve(); }}; + auto equal{[](const IR::Value& lhs, const IR::Value& rhs) { + if (lhs == rhs) { + return true; + } + // Not equal, but try to match if they read the same constant buffer + if (!lhs.IsImmediate() && !rhs.IsImmediate() && + lhs.Inst()->GetOpcode() == IR::Opcode::GetCbufU32 && + rhs.Inst()->GetOpcode() == IR::Opcode::GetCbufU32 && + lhs.Inst()->Arg(0) == rhs.Inst()->Arg(0) && lhs.Inst()->Arg(1) == rhs.Inst()->Arg(1)) { + return true; + } + return false; + }}; + return std::ranges::adjacent_find(range, std::not_fn(equal), resolver) == std::end(range); +} + void FoldGetRegister(IR::Inst& inst) { if (inst.Arg(0).Reg() == IR::Reg::RZ) { inst.ReplaceUsesWith(IR::Value{u32{0}}); @@ -100,6 +121,157 @@ void FoldGetPred(IR::Inst& inst) { } } +/// Replaces the XMAD pattern generated by an integer FMA +bool FoldXmadMultiplyAdd(IR::Block& block, IR::Inst& inst) { + /* + * We are looking for this specific pattern: + * %6 = BitFieldUExtract %op_b, #0, #16 + * %7 = BitFieldUExtract %op_a', #16, #16 + * %8 = IMul32 %6, %7 + * %10 = BitFieldUExtract %op_a', #0, #16 + * %11 = BitFieldInsert %8, %10, #16, #16 + * %15 = BitFieldUExtract %op_b, #0, #16 + * %16 = BitFieldUExtract %op_a, #0, #16 + * %17 = IMul32 %15, %16 + * %18 = IAdd32 %17, %op_c + * %22 = BitFieldUExtract %op_b, #16, #16 + * %23 = BitFieldUExtract %11, #16, #16 + * %24 = IMul32 %22, %23 + * %25 = ShiftLeftLogical32 %24, #16 + * %26 = ShiftLeftLogical32 %11, #16 + * %27 = IAdd32 %26, %18 + * %result = IAdd32 %25, %27 + * + * And replace it with: + * %temp = IMul32 %op_a, %op_b + * %result = IAdd32 %temp, %op_c + * + * This optimization has been proven safe by Nvidia's compiler logic being reversed. + * (If Nvidia generates this code from 'fma(a, b, c)', we can do the same in the reverse order.) + */ + const IR::Value zero{0u}; + const IR::Value sixteen{16u}; + IR::Inst* const _25{inst.Arg(0).TryInstRecursive()}; + IR::Inst* const _27{inst.Arg(1).TryInstRecursive()}; + if (!_25 || !_27) { + return false; + } + if (_27->GetOpcode() != IR::Opcode::IAdd32) { + return false; + } + if (_25->GetOpcode() != IR::Opcode::ShiftLeftLogical32 || _25->Arg(1) != sixteen) { + return false; + } + IR::Inst* const _24{_25->Arg(0).TryInstRecursive()}; + if (!_24 || _24->GetOpcode() != IR::Opcode::IMul32) { + return false; + } + IR::Inst* const _22{_24->Arg(0).TryInstRecursive()}; + IR::Inst* const _23{_24->Arg(1).TryInstRecursive()}; + if (!_22 || !_23) { + return false; + } + if (_22->GetOpcode() != IR::Opcode::BitFieldUExtract) { + return false; + } + if (_23->GetOpcode() != IR::Opcode::BitFieldUExtract) { + return false; + } + if (_22->Arg(1) != sixteen || _22->Arg(2) != sixteen) { + return false; + } + if (_23->Arg(1) != sixteen || _23->Arg(2) != sixteen) { + return false; + } + IR::Inst* const _11{_23->Arg(0).TryInstRecursive()}; + if (!_11 || _11->GetOpcode() != IR::Opcode::BitFieldInsert) { + return false; + } + if (_11->Arg(2) != sixteen || _11->Arg(3) != sixteen) { + return false; + } + IR::Inst* const _8{_11->Arg(0).TryInstRecursive()}; + IR::Inst* const _10{_11->Arg(1).TryInstRecursive()}; + if (!_8 || !_10) { + return false; + } + if (_8->GetOpcode() != IR::Opcode::IMul32) { + return false; + } + if (_10->GetOpcode() != IR::Opcode::BitFieldUExtract) { + return false; + } + IR::Inst* const _6{_8->Arg(0).TryInstRecursive()}; + IR::Inst* const _7{_8->Arg(1).TryInstRecursive()}; + if (!_6 || !_7) { + return false; + } + if (_6->GetOpcode() != IR::Opcode::BitFieldUExtract) { + return false; + } + if (_7->GetOpcode() != IR::Opcode::BitFieldUExtract) { + return false; + } + if (_6->Arg(1) != zero || _6->Arg(2) != sixteen) { + return false; + } + if (_7->Arg(1) != sixteen || _7->Arg(2) != sixteen) { + return false; + } + IR::Inst* const _26{_27->Arg(0).TryInstRecursive()}; + IR::Inst* const _18{_27->Arg(1).TryInstRecursive()}; + if (!_26 || !_18) { + return false; + } + if (_26->GetOpcode() != IR::Opcode::ShiftLeftLogical32 || _26->Arg(1) != sixteen) { + return false; + } + if (_26->Arg(0).InstRecursive() != _11) { + return false; + } + if (_18->GetOpcode() != IR::Opcode::IAdd32) { + return false; + } + IR::Inst* const _17{_18->Arg(0).TryInstRecursive()}; + if (!_17 || _17->GetOpcode() != IR::Opcode::IMul32) { + return false; + } + IR::Inst* const _15{_17->Arg(0).TryInstRecursive()}; + IR::Inst* const _16{_17->Arg(1).TryInstRecursive()}; + if (!_15 || !_16) { + return false; + } + if (_15->GetOpcode() != IR::Opcode::BitFieldUExtract) { + return false; + } + if (_16->GetOpcode() != IR::Opcode::BitFieldUExtract) { + return false; + } + if (_15->Arg(1) != zero || _16->Arg(1) != zero || _10->Arg(1) != zero) { + return false; + } + if (_15->Arg(2) != sixteen || _16->Arg(2) != sixteen || _10->Arg(2) != sixteen) { + return false; + } + const std::array<IR::Value, 3> op_as{ + _7->Arg(0).Resolve(), + _16->Arg(0).Resolve(), + _10->Arg(0).Resolve(), + }; + const std::array<IR::Value, 3> op_bs{ + _22->Arg(0).Resolve(), + _6->Arg(0).Resolve(), + _15->Arg(0).Resolve(), + }; + const IR::U32 op_c{_18->Arg(1)}; + if (!AreEqual(op_as) || !AreEqual(op_bs)) { + return false; + } + IR::IREmitter ir{block, IR::Block::InstructionList::s_iterator_to(inst)}; + inst.ReplaceUsesWith(ir.IAdd(ir.IMul(IR::U32{op_as[0]}, IR::U32{op_bs[1]}), op_c)); + return true; +} + /// Replaces the pattern generated by two XMAD multiplications bool FoldXmadMultiply(IR::Block& block, IR::Inst& inst) { /* @@ -116,33 +288,31 @@ bool FoldXmadMultiply(IR::Block& block, IR::Inst& inst) { * * This optimization has been proven safe by LLVM and MSVC. */ - const IR::Value lhs_arg{inst.Arg(0)}; - const IR::Value rhs_arg{inst.Arg(1)}; - if (lhs_arg.IsImmediate() || rhs_arg.IsImmediate()) { + IR::Inst* const lhs_shl{inst.Arg(0).TryInstRecursive()}; + IR::Inst* const rhs_mul{inst.Arg(1).TryInstRecursive()}; + if (!lhs_shl || !rhs_mul) { return false; } - IR::Inst* const lhs_shl{lhs_arg.InstRecursive()}; if (lhs_shl->GetOpcode() != IR::Opcode::ShiftLeftLogical32 || lhs_shl->Arg(1) != IR::Value{16U}) { return false; } - if (lhs_shl->Arg(0).IsImmediate()) { + IR::Inst* const lhs_mul{lhs_shl->Arg(0).TryInstRecursive()}; + if (!lhs_mul) { return false; } - IR::Inst* const lhs_mul{lhs_shl->Arg(0).InstRecursive()}; - IR::Inst* const rhs_mul{rhs_arg.InstRecursive()}; if (lhs_mul->GetOpcode() != IR::Opcode::IMul32 || rhs_mul->GetOpcode() != IR::Opcode::IMul32) { return false; } - if (lhs_mul->Arg(1).Resolve() != rhs_mul->Arg(1).Resolve()) { + const IR::U32 factor_b{lhs_mul->Arg(1)}; + if (factor_b.Resolve() != rhs_mul->Arg(1).Resolve()) { return false; } - const IR::U32 factor_b{lhs_mul->Arg(1)}; - if (lhs_mul->Arg(0).IsImmediate() || rhs_mul->Arg(0).IsImmediate()) { + IR::Inst* const lhs_bfe{lhs_mul->Arg(0).TryInstRecursive()}; + IR::Inst* const rhs_bfe{rhs_mul->Arg(0).TryInstRecursive()}; + if (!lhs_bfe || !rhs_bfe) { return false; } - IR::Inst* const lhs_bfe{lhs_mul->Arg(0).InstRecursive()}; - IR::Inst* const rhs_bfe{rhs_mul->Arg(0).InstRecursive()}; if (lhs_bfe->GetOpcode() != IR::Opcode::BitFieldUExtract) { return false; } @@ -155,10 +325,10 @@ bool FoldXmadMultiply(IR::Block& block, IR::Inst& inst) { if (rhs_bfe->Arg(1) != IR::Value{0U} || rhs_bfe->Arg(2) != IR::Value{16U}) { return false; } - if (lhs_bfe->Arg(0).Resolve() != rhs_bfe->Arg(0).Resolve()) { + const IR::U32 factor_a{lhs_bfe->Arg(0)}; + if (factor_a.Resolve() != rhs_bfe->Arg(0).Resolve()) { return false; } - const IR::U32 factor_a{lhs_bfe->Arg(0)}; IR::IREmitter ir{block, IR::Block::InstructionList::s_iterator_to(inst)}; inst.ReplaceUsesWith(ir.IMul(factor_a, factor_b)); return true; @@ -181,6 +351,9 @@ void FoldAdd(IR::Block& block, IR::Inst& inst) { if (FoldXmadMultiply(block, inst)) { return; } + if (FoldXmadMultiplyAdd(block, inst)) { + return; + } } } @@ -476,6 +649,10 @@ void ConstantPropagation(IR::Block& block, IR::Inst& inst) { return FoldInverseFunc(inst, IR::Opcode::UnpackHalf2x16); case IR::Opcode::UnpackHalf2x16: return FoldInverseFunc(inst, IR::Opcode::PackHalf2x16); + case IR::Opcode::PackFloat2x16: + return FoldInverseFunc(inst, IR::Opcode::UnpackFloat2x16); + case IR::Opcode::UnpackFloat2x16: + return FoldInverseFunc(inst, IR::Opcode::PackFloat2x16); case IR::Opcode::SelectU1: case IR::Opcode::SelectU8: case IR::Opcode::SelectU16: diff --git a/src/video_core/CMakeLists.txt b/src/video_core/CMakeLists.txt index 007ecc13e..333f6f35f 100644 --- a/src/video_core/CMakeLists.txt +++ b/src/video_core/CMakeLists.txt @@ -106,6 +106,8 @@ add_library(video_core STATIC renderer_vulkan/maxwell_to_vk.cpp renderer_vulkan/maxwell_to_vk.h renderer_vulkan/pipeline_helper.h + renderer_vulkan/pipeline_statistics.cpp + renderer_vulkan/pipeline_statistics.h renderer_vulkan/renderer_vulkan.h renderer_vulkan/renderer_vulkan.cpp renderer_vulkan/vk_blit_screen.cpp diff --git a/src/video_core/renderer_vulkan/pipeline_statistics.cpp b/src/video_core/renderer_vulkan/pipeline_statistics.cpp new file mode 100644 index 000000000..bfec931a6 --- /dev/null +++ b/src/video_core/renderer_vulkan/pipeline_statistics.cpp @@ -0,0 +1,100 @@ +// Copyright 2021 yuzu Emulator Project +// Licensed under GPLv2 or any later version +// Refer to the license.txt file included. + +#include <string_view> + +#include <fmt/format.h> + +#include "common/common_types.h" +#include "common/logging/log.h" +#include "video_core/renderer_vulkan/pipeline_statistics.h" +#include "video_core/vulkan_common/vulkan_device.h" +#include "video_core/vulkan_common/vulkan_wrapper.h" + +namespace Vulkan { + +using namespace std::string_view_literals; + +static u64 GetUint64(const VkPipelineExecutableStatisticKHR& statistic) { + switch (statistic.format) { + case VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_INT64_KHR: + return static_cast<u64>(statistic.value.i64); + case VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR: + return statistic.value.u64; + case VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_FLOAT64_KHR: + return static_cast<u64>(statistic.value.f64); + default: + return 0; + } +} + +PipelineStatistics::PipelineStatistics(const Device& device_) : device{device_} {} + +void PipelineStatistics::Collect(VkPipeline pipeline) { + const auto& dev{device.GetLogical()}; + const std::vector properties{dev.GetPipelineExecutablePropertiesKHR(pipeline)}; + const u32 num_executables{static_cast<u32>(properties.size())}; + for (u32 executable = 0; executable < num_executables; ++executable) { + const auto statistics{dev.GetPipelineExecutableStatisticsKHR(pipeline, executable)}; + if (statistics.empty()) { + continue; + } + Stats stage_stats; + for (const auto& statistic : statistics) { + const char* const name{statistic.name}; + if (name == "Binary Size"sv || name == "Code size"sv || name == "Instruction Count"sv) { + stage_stats.code_size = GetUint64(statistic); + } else if (name == "Register Count"sv) { + stage_stats.register_count = GetUint64(statistic); + } else if (name == "SGPRs"sv || name == "numUsedSgprs"sv) { + stage_stats.sgpr_count = GetUint64(statistic); + } else if (name == "VGPRs"sv || name == "numUsedVgprs"sv) { + stage_stats.vgpr_count = GetUint64(statistic); + } else if (name == "Branches"sv) { + stage_stats.branches_count = GetUint64(statistic); + } else if (name == "Basic Block Count"sv) { + stage_stats.basic_block_count = GetUint64(statistic); + } + } + std::lock_guard lock{mutex}; + collected_stats.push_back(stage_stats); + } +} + +void PipelineStatistics::Report() const { + double num{}; + Stats total; + { + std::lock_guard lock{mutex}; + for (const Stats& stats : collected_stats) { + total.code_size += stats.code_size; + total.register_count += stats.register_count; + total.sgpr_count += stats.sgpr_count; + total.vgpr_count += stats.vgpr_count; + total.branches_count += stats.branches_count; + total.basic_block_count += stats.basic_block_count; + } + num = static_cast<double>(collected_stats.size()); + } + std::string report; + const auto add = [&](const char* fmt, u64 value) { + if (value > 0) { + report += fmt::format(fmt::runtime(fmt), static_cast<double>(value) / num); + } + }; + add("Code size: {:9.03f}\n", total.code_size); + add("Register count: {:9.03f}\n", total.register_count); + add("SGPRs: {:9.03f}\n", total.sgpr_count); + add("VGPRs: {:9.03f}\n", total.vgpr_count); + add("Branches count: {:9.03f}\n", total.branches_count); + add("Basic blocks: {:9.03f}\n", total.basic_block_count); + + LOG_INFO(Render_Vulkan, + "\nAverage pipeline statistics\n" + "==========================================\n" + "{}\n", + report); +} + +} // namespace Vulkan diff --git a/src/video_core/renderer_vulkan/pipeline_statistics.h b/src/video_core/renderer_vulkan/pipeline_statistics.h new file mode 100644 index 000000000..b61840107 --- /dev/null +++ b/src/video_core/renderer_vulkan/pipeline_statistics.h @@ -0,0 +1,40 @@ +// Copyright 2021 yuzu Emulator Project +// Licensed under GPLv2 or any later version +// Refer to the license.txt file included. + +#pragma once + +#include <mutex> +#include <vector> + +#include "common/common_types.h" +#include "video_core/vulkan_common/vulkan_wrapper.h" + +namespace Vulkan { + +class Device; + +class PipelineStatistics { +public: + explicit PipelineStatistics(const Device& device_); + + void Collect(VkPipeline pipeline); + + void Report() const; + +private: + struct Stats { + u64 code_size{}; + u64 register_count{}; + u64 sgpr_count{}; + u64 vgpr_count{}; + u64 branches_count{}; + u64 basic_block_count{}; + }; + + const Device& device; + mutable std::mutex mutex; + std::vector<Stats> collected_stats; +}; + +} // namespace Vulkan diff --git a/src/video_core/renderer_vulkan/vk_compute_pipeline.cpp b/src/video_core/renderer_vulkan/vk_compute_pipeline.cpp index 70b84c7a6..44faf626a 100644 --- a/src/video_core/renderer_vulkan/vk_compute_pipeline.cpp +++ b/src/video_core/renderer_vulkan/vk_compute_pipeline.cpp @@ -8,6 +8,7 @@ #include <boost/container/small_vector.hpp> #include "video_core/renderer_vulkan/pipeline_helper.h" +#include "video_core/renderer_vulkan/pipeline_statistics.h" #include "video_core/renderer_vulkan/vk_buffer_cache.h" #include "video_core/renderer_vulkan/vk_compute_pipeline.h" #include "video_core/renderer_vulkan/vk_descriptor_pool.h" @@ -26,6 +27,7 @@ using Tegra::Texture::TexturePair; ComputePipeline::ComputePipeline(const Device& device_, DescriptorPool& descriptor_pool, VKUpdateDescriptorQueue& update_descriptor_queue_, Common::ThreadWorker* thread_worker, + PipelineStatistics* pipeline_statistics, VideoCore::ShaderNotify* shader_notify, const Shader::Info& info_, vk::ShaderModule spv_module_) : device{device_}, update_descriptor_queue{update_descriptor_queue_}, info{info_}, @@ -36,7 +38,7 @@ ComputePipeline::ComputePipeline(const Device& device_, DescriptorPool& descript std::copy_n(info.constant_buffer_used_sizes.begin(), uniform_buffer_sizes.size(), uniform_buffer_sizes.begin()); - auto func{[this, &descriptor_pool, shader_notify] { + auto func{[this, &descriptor_pool, shader_notify, pipeline_statistics] { DescriptorLayoutBuilder builder{device}; builder.Add(info, VK_SHADER_STAGE_COMPUTE_BIT); @@ -50,10 +52,14 @@ ComputePipeline::ComputePipeline(const Device& device_, DescriptorPool& descript .pNext = nullptr, .requiredSubgroupSize = GuestWarpSize, }; + VkPipelineCreateFlags flags{}; + if (device.IsKhrPipelineEexecutablePropertiesEnabled()) { + flags |= VK_PIPELINE_CREATE_CAPTURE_STATISTICS_BIT_KHR; + } pipeline = device.GetLogical().CreateComputePipeline({ .sType = VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO, .pNext = nullptr, - .flags = 0, + .flags = flags, .stage{ .sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO, .pNext = device.IsExtSubgroupSizeControlSupported() ? &subgroup_size_ci : nullptr, @@ -67,6 +73,9 @@ ComputePipeline::ComputePipeline(const Device& device_, DescriptorPool& descript .basePipelineHandle = 0, .basePipelineIndex = 0, }); + if (pipeline_statistics) { + pipeline_statistics->Collect(*pipeline); + } std::lock_guard lock{build_mutex}; is_built = true; build_condvar.notify_one(); diff --git a/src/video_core/renderer_vulkan/vk_compute_pipeline.h b/src/video_core/renderer_vulkan/vk_compute_pipeline.h index 52fec04d3..8c4b0a301 100644 --- a/src/video_core/renderer_vulkan/vk_compute_pipeline.h +++ b/src/video_core/renderer_vulkan/vk_compute_pipeline.h @@ -25,6 +25,7 @@ class ShaderNotify; namespace Vulkan { class Device; +class PipelineStatistics; class VKScheduler; class ComputePipeline { @@ -32,6 +33,7 @@ public: explicit ComputePipeline(const Device& device, DescriptorPool& descriptor_pool, VKUpdateDescriptorQueue& update_descriptor_queue, Common::ThreadWorker* thread_worker, + PipelineStatistics* pipeline_statistics, VideoCore::ShaderNotify* shader_notify, const Shader::Info& info, vk::ShaderModule spv_module); diff --git a/src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp b/src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp index 18482e1d0..7c0f91007 100644 --- a/src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp +++ b/src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp @@ -11,6 +11,7 @@ #include "common/bit_field.h" #include "video_core/renderer_vulkan/maxwell_to_vk.h" #include "video_core/renderer_vulkan/pipeline_helper.h" +#include "video_core/renderer_vulkan/pipeline_statistics.h" #include "video_core/renderer_vulkan/vk_buffer_cache.h" #include "video_core/renderer_vulkan/vk_graphics_pipeline.h" #include "video_core/renderer_vulkan/vk_render_pass_cache.h" @@ -217,8 +218,8 @@ GraphicsPipeline::GraphicsPipeline( VKScheduler& scheduler_, BufferCache& buffer_cache_, TextureCache& texture_cache_, VideoCore::ShaderNotify* shader_notify, const Device& device_, DescriptorPool& descriptor_pool, VKUpdateDescriptorQueue& update_descriptor_queue_, Common::ThreadWorker* worker_thread, - RenderPassCache& render_pass_cache, const GraphicsPipelineCacheKey& key_, - std::array<vk::ShaderModule, NUM_STAGES> stages, + PipelineStatistics* pipeline_statistics, RenderPassCache& render_pass_cache, + const GraphicsPipelineCacheKey& key_, std::array<vk::ShaderModule, NUM_STAGES> stages, const std::array<const Shader::Info*, NUM_STAGES>& infos) : key{key_}, maxwell3d{maxwell3d_}, gpu_memory{gpu_memory_}, device{device_}, texture_cache{texture_cache_}, buffer_cache{buffer_cache_}, scheduler{scheduler_}, @@ -235,7 +236,7 @@ GraphicsPipeline::GraphicsPipeline( enabled_uniform_buffer_masks[stage] = info->constant_buffer_mask; std::ranges::copy(info->constant_buffer_used_sizes, uniform_buffer_sizes[stage].begin()); } - auto func{[this, shader_notify, &render_pass_cache, &descriptor_pool] { + auto func{[this, shader_notify, &render_pass_cache, &descriptor_pool, pipeline_statistics] { DescriptorLayoutBuilder builder{MakeBuilder(device, stage_infos)}; uses_push_descriptor = builder.CanUsePushDescriptor(); descriptor_set_layout = builder.CreateDescriptorSetLayout(uses_push_descriptor); @@ -250,6 +251,9 @@ GraphicsPipeline::GraphicsPipeline( const VkRenderPass render_pass{render_pass_cache.Get(MakeRenderPassKey(key.state))}; Validate(); MakePipeline(render_pass); + if (pipeline_statistics) { + pipeline_statistics->Collect(*pipeline); + } std::lock_guard lock{build_mutex}; is_built = true; @@ -782,10 +786,14 @@ void GraphicsPipeline::MakePipeline(VkRenderPass render_pass) { } */ } + VkPipelineCreateFlags flags{}; + if (device.IsKhrPipelineEexecutablePropertiesEnabled()) { + flags |= VK_PIPELINE_CREATE_CAPTURE_STATISTICS_BIT_KHR; + } pipeline = device.GetLogical().CreateGraphicsPipeline({ .sType = VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO, .pNext = nullptr, - .flags = 0, + .flags = flags, .stageCount = static_cast<u32>(shader_stages.size()), .pStages = shader_stages.data(), .pVertexInputState = &vertex_input_ci, diff --git a/src/video_core/renderer_vulkan/vk_graphics_pipeline.h b/src/video_core/renderer_vulkan/vk_graphics_pipeline.h index 2bd48d697..1c780e944 100644 --- a/src/video_core/renderer_vulkan/vk_graphics_pipeline.h +++ b/src/video_core/renderer_vulkan/vk_graphics_pipeline.h @@ -60,6 +60,7 @@ struct hash<Vulkan::GraphicsPipelineCacheKey> { namespace Vulkan { class Device; +class PipelineStatistics; class RenderPassCache; class VKScheduler; class VKUpdateDescriptorQueue; @@ -73,8 +74,9 @@ public: VKScheduler& scheduler, BufferCache& buffer_cache, TextureCache& texture_cache, VideoCore::ShaderNotify* shader_notify, const Device& device, DescriptorPool& descriptor_pool, VKUpdateDescriptorQueue& update_descriptor_queue, - Common::ThreadWorker* worker_thread, RenderPassCache& render_pass_cache, - const GraphicsPipelineCacheKey& key, std::array<vk::ShaderModule, NUM_STAGES> stages, + Common::ThreadWorker* worker_thread, PipelineStatistics* pipeline_statistics, + RenderPassCache& render_pass_cache, const GraphicsPipelineCacheKey& key, + std::array<vk::ShaderModule, NUM_STAGES> stages, const std::array<const Shader::Info*, NUM_STAGES>& infos); GraphicsPipeline& operator=(GraphicsPipeline&&) noexcept = delete; diff --git a/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp b/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp index 57b163247..a37ca1fdf 100644 --- a/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp +++ b/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp @@ -29,6 +29,7 @@ #include "video_core/renderer_vulkan/fixed_pipeline_state.h" #include "video_core/renderer_vulkan/maxwell_to_vk.h" #include "video_core/renderer_vulkan/pipeline_helper.h" +#include "video_core/renderer_vulkan/pipeline_statistics.h" #include "video_core/renderer_vulkan/vk_compute_pipeline.h" #include "video_core/renderer_vulkan/vk_descriptor_pool.h" #include "video_core/renderer_vulkan/vk_pipeline_cache.h" @@ -389,15 +390,19 @@ void PipelineCache::LoadDiskResources(u64 title_id, std::stop_token stop_loading size_t total{}; size_t built{}; bool has_loaded{}; + std::unique_ptr<PipelineStatistics> statistics; } state; + if (device.IsKhrPipelineEexecutablePropertiesEnabled()) { + state.statistics = std::make_unique<PipelineStatistics>(device); + } const auto load_compute{[&](std::ifstream& file, FileEnvironment env) { ComputePipelineCacheKey key; file.read(reinterpret_cast<char*>(&key), sizeof(key)); workers.QueueWork([this, key, env = std::move(env), &state, &callback]() mutable { ShaderPools pools; - auto pipeline{CreateComputePipeline(pools, key, env, false)}; + auto pipeline{CreateComputePipeline(pools, key, env, state.statistics.get(), false)}; std::lock_guard lock{state.mutex}; if (pipeline) { compute_cache.emplace(key, std::move(pipeline)); @@ -425,7 +430,8 @@ void PipelineCache::LoadDiskResources(u64 title_id, std::stop_token stop_loading for (auto& env : envs) { env_ptrs.push_back(&env); } - auto pipeline{CreateGraphicsPipeline(pools, key, MakeSpan(env_ptrs), false)}; + auto pipeline{CreateGraphicsPipeline(pools, key, MakeSpan(env_ptrs), + state.statistics.get(), false)}; std::lock_guard lock{state.mutex}; graphics_cache.emplace(key, std::move(pipeline)); @@ -445,6 +451,10 @@ void PipelineCache::LoadDiskResources(u64 title_id, std::stop_token stop_loading lock.unlock(); workers.WaitForRequests(); + + if (state.statistics) { + state.statistics->Report(); + } } GraphicsPipeline* PipelineCache::CurrentGraphicsPipelineSlowPath() { @@ -486,7 +496,8 @@ GraphicsPipeline* PipelineCache::BuiltPipeline(GraphicsPipeline* pipeline) const std::unique_ptr<GraphicsPipeline> PipelineCache::CreateGraphicsPipeline( ShaderPools& pools, const GraphicsPipelineCacheKey& key, - std::span<Shader::Environment* const> envs, bool build_in_parallel) try { + std::span<Shader::Environment* const> envs, PipelineStatistics* statistics, + bool build_in_parallel) try { LOG_INFO(Render_Vulkan, "0x{:016x}", key.Hash()); size_t env_index{0}; std::array<Shader::IR::Program, Maxwell::MaxShaderProgram> programs; @@ -540,7 +551,7 @@ std::unique_ptr<GraphicsPipeline> PipelineCache::CreateGraphicsPipeline( Common::ThreadWorker* const thread_worker{build_in_parallel ? &workers : nullptr}; return std::make_unique<GraphicsPipeline>( maxwell3d, gpu_memory, scheduler, buffer_cache, texture_cache, &shader_notify, device, - descriptor_pool, update_descriptor_queue, thread_worker, render_pass_cache, key, + descriptor_pool, update_descriptor_queue, thread_worker, statistics, render_pass_cache, key, std::move(modules), infos); } catch (const Shader::Exception& exception) { @@ -553,7 +564,8 @@ std::unique_ptr<GraphicsPipeline> PipelineCache::CreateGraphicsPipeline() { GetGraphicsEnvironments(environments, graphics_key.unique_hashes); main_pools.ReleaseContents(); - auto pipeline{CreateGraphicsPipeline(main_pools, graphics_key, environments.Span(), true)}; + auto pipeline{ + CreateGraphicsPipeline(main_pools, graphics_key, environments.Span(), nullptr, true)}; if (!pipeline || pipeline_cache_filename.empty()) { return pipeline; } @@ -578,7 +590,7 @@ std::unique_ptr<ComputePipeline> PipelineCache::CreateComputePipeline( env.SetCachedSize(shader->size_bytes); main_pools.ReleaseContents(); - auto pipeline{CreateComputePipeline(main_pools, key, env, true)}; + auto pipeline{CreateComputePipeline(main_pools, key, env, nullptr, true)}; if (!pipeline || pipeline_cache_filename.empty()) { return pipeline; } @@ -591,7 +603,7 @@ std::unique_ptr<ComputePipeline> PipelineCache::CreateComputePipeline( std::unique_ptr<ComputePipeline> PipelineCache::CreateComputePipeline( ShaderPools& pools, const ComputePipelineCacheKey& key, Shader::Environment& env, - bool build_in_parallel) try { + PipelineStatistics* statistics, bool build_in_parallel) try { LOG_INFO(Render_Vulkan, "0x{:016x}", key.Hash()); Shader::Maxwell::Flow::CFG cfg{env, pools.flow_block, env.StartAddress()}; @@ -605,8 +617,8 @@ std::unique_ptr<ComputePipeline> PipelineCache::CreateComputePipeline( } Common::ThreadWorker* const thread_worker{build_in_parallel ? &workers : nullptr}; return std::make_unique<ComputePipeline>(device, descriptor_pool, update_descriptor_queue, - thread_worker, &shader_notify, program.info, - std::move(spv_module)); + thread_worker, statistics, &shader_notify, + program.info, std::move(spv_module)); } catch (const Shader::Exception& exception) { LOG_ERROR(Render_Vulkan, "{}", exception.what()); diff --git a/src/video_core/renderer_vulkan/vk_pipeline_cache.h b/src/video_core/renderer_vulkan/vk_pipeline_cache.h index efe5a7ed8..4c135b5dd 100644 --- a/src/video_core/renderer_vulkan/vk_pipeline_cache.h +++ b/src/video_core/renderer_vulkan/vk_pipeline_cache.h @@ -80,8 +80,9 @@ struct hash<Vulkan::ComputePipelineCacheKey> { namespace Vulkan { class ComputePipeline; -class Device; class DescriptorPool; +class Device; +class PipelineStatistics; class RasterizerVulkan; class RenderPassCache; class VKScheduler; @@ -128,7 +129,8 @@ private: std::unique_ptr<GraphicsPipeline> CreateGraphicsPipeline( ShaderPools& pools, const GraphicsPipelineCacheKey& key, - std::span<Shader::Environment* const> envs, bool build_in_parallel); + std::span<Shader::Environment* const> envs, PipelineStatistics* statistics, + bool build_in_parallel); std::unique_ptr<ComputePipeline> CreateComputePipeline(const ComputePipelineCacheKey& key, const ShaderInfo* shader); @@ -136,6 +138,7 @@ private: std::unique_ptr<ComputePipeline> CreateComputePipeline(ShaderPools& pools, const ComputePipelineCacheKey& key, Shader::Environment& env, + PipelineStatistics* statistics, bool build_in_parallel); const Device& device; diff --git a/src/video_core/vulkan_common/vulkan_device.cpp b/src/video_core/vulkan_common/vulkan_device.cpp index 44afdc1cd..8e56a89e1 100644 --- a/src/video_core/vulkan_common/vulkan_device.cpp +++ b/src/video_core/vulkan_common/vulkan_device.cpp @@ -526,6 +526,17 @@ Device::Device(VkInstance instance_, vk::PhysicalDevice physical_, VkSurfaceKHR SetNext(next, workgroup_layout); } + VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR executable_properties; + if (khr_pipeline_executable_properties) { + LOG_INFO(Render_Vulkan, "Enabling shader feedback, expect slower shader build times"); + executable_properties = { + .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR, + .pNext = nullptr, + .pipelineExecutableInfo = VK_TRUE, + }; + SetNext(next, executable_properties); + } + if (!ext_depth_range_unrestricted) { LOG_INFO(Render_Vulkan, "Device doesn't support depth range unrestricted"); } @@ -824,6 +835,7 @@ std::vector<const char*> Device::LoadExtensions(bool requires_surface) { bool has_khr_shader_float16_int8{}; bool has_khr_workgroup_memory_explicit_layout{}; + bool has_khr_pipeline_executable_properties{}; bool has_ext_subgroup_size_control{}; bool has_ext_transform_feedback{}; bool has_ext_custom_border_color{}; @@ -878,6 +890,10 @@ std::vector<const char*> Device::LoadExtensions(bool requires_surface) { test(nv_device_diagnostics_config, VK_NV_DEVICE_DIAGNOSTICS_CONFIG_EXTENSION_NAME, true); } + if (Settings::values.renderer_shader_feedback) { + test(has_khr_pipeline_executable_properties, + VK_KHR_PIPELINE_EXECUTABLE_PROPERTIES_EXTENSION_NAME, false); + } } VkPhysicalDeviceFeatures2KHR features{}; features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2_KHR; @@ -1033,6 +1049,19 @@ std::vector<const char*> Device::LoadExtensions(bool requires_surface) { khr_workgroup_memory_explicit_layout = true; } } + if (has_khr_pipeline_executable_properties) { + VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR executable_properties; + executable_properties.sType = + VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR; + executable_properties.pNext = nullptr; + features.pNext = &executable_properties; + physical.GetFeatures2KHR(features); + + if (executable_properties.pipelineExecutableInfo) { + extensions.push_back(VK_KHR_PIPELINE_EXECUTABLE_PROPERTIES_EXTENSION_NAME); + khr_pipeline_executable_properties = true; + } + } if (khr_push_descriptor) { VkPhysicalDevicePushDescriptorPropertiesKHR push_descriptor; push_descriptor.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR; diff --git a/src/video_core/vulkan_common/vulkan_device.h b/src/video_core/vulkan_common/vulkan_device.h index df394e384..c19f40746 100644 --- a/src/video_core/vulkan_common/vulkan_device.h +++ b/src/video_core/vulkan_common/vulkan_device.h @@ -214,6 +214,11 @@ public: return khr_push_descriptor; } + /// Returns true if VK_KHR_pipeline_executable_properties is enabled. + bool IsKhrPipelineEexecutablePropertiesEnabled() const { + return khr_pipeline_executable_properties; + } + /// Returns true if the device supports VK_KHR_workgroup_memory_explicit_layout. bool IsKhrWorkgroupMemoryExplicitLayoutSupported() const { return khr_workgroup_memory_explicit_layout; @@ -378,6 +383,7 @@ private: bool khr_spirv_1_4{}; ///< Support for VK_KHR_spirv_1_4. bool khr_workgroup_memory_explicit_layout{}; ///< Support for explicit workgroup layouts. bool khr_push_descriptor{}; ///< Support for VK_KHR_push_descritor. + bool khr_pipeline_executable_properties{}; ///< Support for executable properties. bool ext_index_type_uint8{}; ///< Support for VK_EXT_index_type_uint8. bool ext_sampler_filter_minmax{}; ///< Support for VK_EXT_sampler_filter_minmax. bool ext_depth_range_unrestricted{}; ///< Support for VK_EXT_depth_range_unrestricted. diff --git a/src/video_core/vulkan_common/vulkan_wrapper.cpp b/src/video_core/vulkan_common/vulkan_wrapper.cpp index 70898004a..a9faa4807 100644 --- a/src/video_core/vulkan_common/vulkan_wrapper.cpp +++ b/src/video_core/vulkan_common/vulkan_wrapper.cpp @@ -181,6 +181,8 @@ void Load(VkDevice device, DeviceDispatch& dld) noexcept { X(vkGetMemoryWin32HandleKHR); #endif X(vkGetQueryPoolResults); + X(vkGetPipelineExecutablePropertiesKHR); + X(vkGetPipelineExecutableStatisticsKHR); X(vkGetSemaphoreCounterValueKHR); X(vkMapMemory); X(vkQueueSubmit); @@ -809,6 +811,42 @@ VkMemoryRequirements Device::GetImageMemoryRequirements(VkImage image) const noe return requirements; } +std::vector<VkPipelineExecutablePropertiesKHR> Device::GetPipelineExecutablePropertiesKHR( + VkPipeline pipeline) const { + const VkPipelineInfoKHR info{ + .sType = VK_STRUCTURE_TYPE_PIPELINE_INFO_KHR, + .pNext = nullptr, + .pipeline = pipeline, + }; + u32 num{}; + dld->vkGetPipelineExecutablePropertiesKHR(handle, &info, &num, nullptr); + std::vector<VkPipelineExecutablePropertiesKHR> properties(num); + for (auto& property : properties) { + property.sType = VK_STRUCTURE_TYPE_PIPELINE_EXECUTABLE_PROPERTIES_KHR; + } + Check(dld->vkGetPipelineExecutablePropertiesKHR(handle, &info, &num, properties.data())); + return properties; +} + +std::vector<VkPipelineExecutableStatisticKHR> Device::GetPipelineExecutableStatisticsKHR( + VkPipeline pipeline, u32 executable_index) const { + const VkPipelineExecutableInfoKHR executable_info{ + .sType = VK_STRUCTURE_TYPE_PIPELINE_EXECUTABLE_INFO_KHR, + .pNext = nullptr, + .pipeline = pipeline, + .executableIndex = executable_index, + }; + u32 num{}; + dld->vkGetPipelineExecutableStatisticsKHR(handle, &executable_info, &num, nullptr); + std::vector<VkPipelineExecutableStatisticKHR> statistics(num); + for (auto& statistic : statistics) { + statistic.sType = VK_STRUCTURE_TYPE_PIPELINE_EXECUTABLE_STATISTIC_KHR; + } + Check(dld->vkGetPipelineExecutableStatisticsKHR(handle, &executable_info, &num, + statistics.data())); + return statistics; +} + void Device::UpdateDescriptorSets(Span<VkWriteDescriptorSet> writes, Span<VkCopyDescriptorSet> copies) const noexcept { dld->vkUpdateDescriptorSets(handle, writes.size(), writes.data(), copies.size(), copies.data()); diff --git a/src/video_core/vulkan_common/vulkan_wrapper.h b/src/video_core/vulkan_common/vulkan_wrapper.h index d76bb4324..b7ae01c6c 100644 --- a/src/video_core/vulkan_common/vulkan_wrapper.h +++ b/src/video_core/vulkan_common/vulkan_wrapper.h @@ -295,6 +295,8 @@ struct DeviceDispatch : InstanceDispatch { #ifdef _WIN32 PFN_vkGetMemoryWin32HandleKHR vkGetMemoryWin32HandleKHR{}; #endif + PFN_vkGetPipelineExecutablePropertiesKHR vkGetPipelineExecutablePropertiesKHR{}; + PFN_vkGetPipelineExecutableStatisticsKHR vkGetPipelineExecutableStatisticsKHR{}; PFN_vkGetQueryPoolResults vkGetQueryPoolResults{}; PFN_vkGetSemaphoreCounterValueKHR vkGetSemaphoreCounterValueKHR{}; PFN_vkMapMemory vkMapMemory{}; @@ -879,6 +881,12 @@ public: VkMemoryRequirements GetImageMemoryRequirements(VkImage image) const noexcept; + std::vector<VkPipelineExecutablePropertiesKHR> GetPipelineExecutablePropertiesKHR( + VkPipeline pipeline) const; + + std::vector<VkPipelineExecutableStatisticKHR> GetPipelineExecutableStatisticsKHR( + VkPipeline pipeline, u32 executable_index) const; + void UpdateDescriptorSets(Span<VkWriteDescriptorSet> writes, Span<VkCopyDescriptorSet> copies) const noexcept; diff --git a/src/yuzu/configuration/config.cpp b/src/yuzu/configuration/config.cpp index ecd5dfac1..f3b8787f5 100644 --- a/src/yuzu/configuration/config.cpp +++ b/src/yuzu/configuration/config.cpp @@ -825,6 +825,7 @@ void Config::ReadRendererValues() { if (global) { ReadBasicSetting(Settings::values.fps_cap); ReadBasicSetting(Settings::values.renderer_debug); + ReadBasicSetting(Settings::values.renderer_shader_feedback); ReadBasicSetting(Settings::values.enable_nsight_aftermath); ReadBasicSetting(Settings::values.disable_shader_loop_safety_checks); } @@ -1363,6 +1364,7 @@ void Config::SaveRendererValues() { if (global) { WriteBasicSetting(Settings::values.fps_cap); WriteBasicSetting(Settings::values.renderer_debug); + WriteBasicSetting(Settings::values.renderer_shader_feedback); WriteBasicSetting(Settings::values.enable_nsight_aftermath); WriteBasicSetting(Settings::values.disable_shader_loop_safety_checks); } diff --git a/src/yuzu/configuration/configure_debug.cpp b/src/yuzu/configuration/configure_debug.cpp index f7e29dbd7..c0b240c1e 100644 --- a/src/yuzu/configuration/configure_debug.cpp +++ b/src/yuzu/configuration/configure_debug.cpp @@ -43,6 +43,8 @@ void ConfigureDebug::SetConfiguration() { ui->use_auto_stub->setChecked(Settings::values.use_auto_stub.GetValue()); ui->enable_graphics_debugging->setEnabled(runtime_lock); ui->enable_graphics_debugging->setChecked(Settings::values.renderer_debug.GetValue()); + ui->enable_shader_feedback->setEnabled(runtime_lock); + ui->enable_shader_feedback->setChecked(Settings::values.renderer_shader_feedback.GetValue()); ui->enable_cpu_debugging->setEnabled(runtime_lock); ui->enable_cpu_debugging->setChecked(Settings::values.cpu_debug_mode.GetValue()); ui->enable_nsight_aftermath->setEnabled(runtime_lock); @@ -65,6 +67,7 @@ void ConfigureDebug::ApplyConfiguration() { Settings::values.use_debug_asserts = ui->use_debug_asserts->isChecked(); Settings::values.use_auto_stub = ui->use_auto_stub->isChecked(); Settings::values.renderer_debug = ui->enable_graphics_debugging->isChecked(); + Settings::values.renderer_shader_feedback = ui->enable_shader_feedback->isChecked(); Settings::values.cpu_debug_mode = ui->enable_cpu_debugging->isChecked(); Settings::values.enable_nsight_aftermath = ui->enable_nsight_aftermath->isChecked(); Settings::values.disable_shader_loop_safety_checks = diff --git a/src/yuzu/configuration/configure_debug.ui b/src/yuzu/configuration/configure_debug.ui index c8baf2921..3fe9ff7de 100644 --- a/src/yuzu/configuration/configure_debug.ui +++ b/src/yuzu/configuration/configure_debug.ui @@ -111,8 +111,8 @@ <property name="title"> <string>Graphics</string> </property> - <layout class="QVBoxLayout" name="verticalLayout_6"> - <item> + <layout class="QGridLayout" name="gridLayout_3"> + <item row="0" column="0"> <widget class="QCheckBox" name="enable_graphics_debugging"> <property name="enabled"> <bool>true</bool> @@ -125,7 +125,7 @@ </property> </widget> </item> - <item> + <item row="2" column="0"> <widget class="QCheckBox" name="enable_nsight_aftermath"> <property name="toolTip"> <string>When checked, it enables Nsight Aftermath crash dumps</string> @@ -135,7 +135,7 @@ </property> </widget> </item> - <item> + <item row="0" column="1"> <widget class="QCheckBox" name="disable_macro_jit"> <property name="enabled"> <bool>true</bool> @@ -148,7 +148,17 @@ </property> </widget> </item> - <item> + <item row="1" column="0"> + <widget class="QCheckBox" name="enable_shader_feedback"> + <property name="toolTip"> + <string>When checked, yuzu will log statistics about the compiled pipeline cache</string> + </property> + <property name="text"> + <string>Enable Shader Feedback</string> + </property> + </widget> + </item> + <item row="1" column="1"> <widget class="QCheckBox" name="disable_loop_safety_checks"> <property name="toolTip"> <string>When checked, it executes shaders without loop logic changes</string> @@ -276,11 +286,14 @@ <tabstop>open_log_button</tabstop> <tabstop>homebrew_args_edit</tabstop> <tabstop>enable_graphics_debugging</tabstop> + <tabstop>enable_shader_feedback</tabstop> <tabstop>enable_nsight_aftermath</tabstop> <tabstop>disable_macro_jit</tabstop> <tabstop>disable_loop_safety_checks</tabstop> + <tabstop>fs_access_log</tabstop> <tabstop>reporting_services</tabstop> <tabstop>quest_flag</tabstop> + <tabstop>enable_cpu_debugging</tabstop> <tabstop>use_debug_asserts</tabstop> <tabstop>use_auto_stub</tabstop> </tabstops> diff --git a/src/yuzu/configuration/configure_system.ui b/src/yuzu/configuration/configure_system.ui index 53b95658b..27f552f59 100644 --- a/src/yuzu/configuration/configure_system.ui +++ b/src/yuzu/configuration/configure_system.ui @@ -401,6 +401,11 @@ <string>Traditional Chinese (正體中文)</string> </property> </item> + <item> + <property name="text"> + <string>Brazilian Portuguese (português do Brasil)</string> + </property> + </item> </widget> </item> <item row="5" column="0"> diff --git a/src/yuzu_cmd/config.cpp b/src/yuzu_cmd/config.cpp index 5af1ee6a8..064ecaafa 100644 --- a/src/yuzu_cmd/config.cpp +++ b/src/yuzu_cmd/config.cpp @@ -444,6 +444,7 @@ void Config::ReadValues() { // Renderer ReadSetting("Renderer", Settings::values.renderer_backend); ReadSetting("Renderer", Settings::values.renderer_debug); + ReadSetting("Renderer", Settings::values.renderer_shader_feedback); ReadSetting("Renderer", Settings::values.enable_nsight_aftermath); ReadSetting("Renderer", Settings::values.disable_shader_loop_safety_checks); ReadSetting("Renderer", Settings::values.vulkan_device); diff --git a/src/yuzu_cmd/default_ini.h b/src/yuzu_cmd/default_ini.h index e646e2d2f..e02eceb99 100644 --- a/src/yuzu_cmd/default_ini.h +++ b/src/yuzu_cmd/default_ini.h @@ -221,6 +221,10 @@ backend = # 0 (default): Disabled, 1: Enabled debug = +# Enable shader feedback. +# 0 (default): Disabled, 1: Enabled +renderer_shader_feedback = + # Enable Nsight Aftermath crash dumps # 0 (default): Disabled, 1: Enabled nsight_aftermath = @@ -363,7 +367,7 @@ custom_rtc = # Sets the systems language index # 0: Japanese, 1: English (default), 2: French, 3: German, 4: Italian, 5: Spanish, 6: Chinese, # 7: Korean, 8: Dutch, 9: Portuguese, 10: Russian, 11: Taiwanese, 12: British English, 13: Canadian French, -# 14: Latin American Spanish, 15: Simplified Chinese, 16: Traditional Chinese +# 14: Latin American Spanish, 15: Simplified Chinese, 16: Traditional Chinese, 17: Brazilian Portuguese language_index = # The system region that yuzu will use during emulation diff --git a/src/yuzu_cmd/emu_window/emu_window_sdl2.cpp b/src/yuzu_cmd/emu_window/emu_window_sdl2.cpp index ea3e0ada4..f643a4b0b 100644 --- a/src/yuzu_cmd/emu_window/emu_window_sdl2.cpp +++ b/src/yuzu_cmd/emu_window/emu_window_sdl2.cpp @@ -16,8 +16,8 @@ #include "yuzu_cmd/emu_window/emu_window_sdl2.h" #include "yuzu_cmd/yuzu_icon.h" -EmuWindow_SDL2::EmuWindow_SDL2(InputCommon::InputSubsystem* input_subsystem_) - : input_subsystem{input_subsystem_} { +EmuWindow_SDL2::EmuWindow_SDL2(InputCommon::InputSubsystem* input_subsystem_, Core::System& system_) + : input_subsystem{input_subsystem_}, system{system_} { if (SDL_Init(SDL_INIT_VIDEO | SDL_INIT_JOYSTICK) < 0) { LOG_CRITICAL(Frontend, "Failed to initialize SDL2! Exiting..."); exit(1); @@ -218,7 +218,7 @@ void EmuWindow_SDL2::WaitEvent() { const u32 current_time = SDL_GetTicks(); if (current_time > last_time + 2000) { - const auto results = Core::System::GetInstance().GetAndResetPerfStats(); + const auto results = system.GetAndResetPerfStats(); const auto title = fmt::format("yuzu {} | {}-{} | FPS: {:.0f} ({:.0f}%)", Common::g_build_fullname, Common::g_scm_branch, Common::g_scm_desc, results.average_game_fps, diff --git a/src/yuzu_cmd/emu_window/emu_window_sdl2.h b/src/yuzu_cmd/emu_window/emu_window_sdl2.h index 1b9ab5b93..aa0d52ae4 100644 --- a/src/yuzu_cmd/emu_window/emu_window_sdl2.h +++ b/src/yuzu_cmd/emu_window/emu_window_sdl2.h @@ -24,7 +24,7 @@ enum class MouseButton; class EmuWindow_SDL2 : public Core::Frontend::EmuWindow { public: - explicit EmuWindow_SDL2(InputCommon::InputSubsystem* input_subsystem); + explicit EmuWindow_SDL2(InputCommon::InputSubsystem* input_subsystem, Core::System& system_); ~EmuWindow_SDL2(); /// Whether the window is still open, and a close request hasn't yet been sent @@ -87,4 +87,7 @@ protected: /// Input subsystem to use with this window. InputCommon::InputSubsystem* input_subsystem; + + /// yuzu core instance + Core::System& system; }; diff --git a/src/yuzu_cmd/emu_window/emu_window_sdl2_gl.cpp b/src/yuzu_cmd/emu_window/emu_window_sdl2_gl.cpp index eadb41790..5b98c255b 100644 --- a/src/yuzu_cmd/emu_window/emu_window_sdl2_gl.cpp +++ b/src/yuzu_cmd/emu_window/emu_window_sdl2_gl.cpp @@ -76,8 +76,9 @@ bool EmuWindow_SDL2_GL::SupportsRequiredGLExtensions() { return unsupported_ext.empty(); } -EmuWindow_SDL2_GL::EmuWindow_SDL2_GL(InputCommon::InputSubsystem* input_subsystem, bool fullscreen) - : EmuWindow_SDL2{input_subsystem} { +EmuWindow_SDL2_GL::EmuWindow_SDL2_GL(InputCommon::InputSubsystem* input_subsystem, + Core::System& system_, bool fullscreen) + : EmuWindow_SDL2{input_subsystem, system_} { SDL_GL_SetAttribute(SDL_GL_CONTEXT_MAJOR_VERSION, 4); SDL_GL_SetAttribute(SDL_GL_CONTEXT_MINOR_VERSION, 6); SDL_GL_SetAttribute(SDL_GL_CONTEXT_PROFILE_MASK, SDL_GL_CONTEXT_PROFILE_COMPATIBILITY); diff --git a/src/yuzu_cmd/emu_window/emu_window_sdl2_gl.h b/src/yuzu_cmd/emu_window/emu_window_sdl2_gl.h index 9e694d985..d7f2c83d8 100644 --- a/src/yuzu_cmd/emu_window/emu_window_sdl2_gl.h +++ b/src/yuzu_cmd/emu_window/emu_window_sdl2_gl.h @@ -8,13 +8,18 @@ #include "core/frontend/emu_window.h" #include "yuzu_cmd/emu_window/emu_window_sdl2.h" +namespace Core { +class System; +} + namespace InputCommon { class InputSubsystem; } class EmuWindow_SDL2_GL final : public EmuWindow_SDL2 { public: - explicit EmuWindow_SDL2_GL(InputCommon::InputSubsystem* input_subsystem, bool fullscreen); + explicit EmuWindow_SDL2_GL(InputCommon::InputSubsystem* input_subsystem, Core::System& system_, + bool fullscreen); ~EmuWindow_SDL2_GL(); std::unique_ptr<Core::Frontend::GraphicsContext> CreateSharedContext() const override; diff --git a/src/yuzu_cmd/emu_window/emu_window_sdl2_vk.cpp b/src/yuzu_cmd/emu_window/emu_window_sdl2_vk.cpp index d1473dbab..cdda375d8 100644 --- a/src/yuzu_cmd/emu_window/emu_window_sdl2_vk.cpp +++ b/src/yuzu_cmd/emu_window/emu_window_sdl2_vk.cpp @@ -24,8 +24,9 @@ #include <SDL.h> #include <SDL_syswm.h> -EmuWindow_SDL2_VK::EmuWindow_SDL2_VK(InputCommon::InputSubsystem* input_subsystem, bool fullscreen) - : EmuWindow_SDL2{input_subsystem} { +EmuWindow_SDL2_VK::EmuWindow_SDL2_VK(InputCommon::InputSubsystem* input_subsystem, + Core::System& system_, bool fullscreen) + : EmuWindow_SDL2{input_subsystem, system_} { const std::string window_title = fmt::format("yuzu {} | {}-{} (Vulkan)", Common::g_build_name, Common::g_scm_branch, Common::g_scm_desc); render_window = diff --git a/src/yuzu_cmd/emu_window/emu_window_sdl2_vk.h b/src/yuzu_cmd/emu_window/emu_window_sdl2_vk.h index de53844f0..3ea521b2a 100644 --- a/src/yuzu_cmd/emu_window/emu_window_sdl2_vk.h +++ b/src/yuzu_cmd/emu_window/emu_window_sdl2_vk.h @@ -19,7 +19,8 @@ class InputSubsystem; class EmuWindow_SDL2_VK final : public EmuWindow_SDL2 { public: - explicit EmuWindow_SDL2_VK(InputCommon::InputSubsystem* input_subsystem, bool fullscreen); + explicit EmuWindow_SDL2_VK(InputCommon::InputSubsystem* input_subsystem, Core::System& system, + bool fullscreen); ~EmuWindow_SDL2_VK() override; std::unique_ptr<Core::Frontend::GraphicsContext> CreateSharedContext() const override; diff --git a/src/yuzu_cmd/yuzu.cpp b/src/yuzu_cmd/yuzu.cpp index 35ce23696..c10093820 100644 --- a/src/yuzu_cmd/yuzu.cpp +++ b/src/yuzu_cmd/yuzu.cpp @@ -172,10 +172,10 @@ int main(int argc, char** argv) { std::unique_ptr<EmuWindow_SDL2> emu_window; switch (Settings::values.renderer_backend.GetValue()) { case Settings::RendererBackend::OpenGL: - emu_window = std::make_unique<EmuWindow_SDL2_GL>(&input_subsystem, fullscreen); + emu_window = std::make_unique<EmuWindow_SDL2_GL>(&input_subsystem, system, fullscreen); break; case Settings::RendererBackend::Vulkan: - emu_window = std::make_unique<EmuWindow_SDL2_VK>(&input_subsystem, fullscreen); + emu_window = std::make_unique<EmuWindow_SDL2_VK>(&input_subsystem, system, fullscreen); break; } |