GPU: Implement predicated exit instructions in the shader programs.

GPU: Partially implemented the bra shader instruction

appveyor.yml

@@ -53,7 +53,7 @@ build_script:
           # https://www.appveyor.com/docs/build-phase
           msbuild msvc_build/yuzu.sln /maxcpucount /logger:"C:\Program Files\AppVeyor\BuildAgent\Appveyor.MSBuildLogger.dll"
         } else {
-          C:\msys64\usr\bin\bash.exe -lc 'mingw32-make -C mingw_build/ 2>&1'
+          C:\msys64\usr\bin\bash.exe -lc 'mingw32-make -j4 -C mingw_build/ 2>&1'
         }
 
 after_build:

src/core/hle/service/am/am.cpp

@@ -155,7 +155,7 @@ ISelfController::ISelfController(std::shared_ptr<NVFlinger::NVFlinger> nvflinger
     RegisterHandlers(functions);
 
     launchable_event =
-        Kernel::Event::Create(Kernel::ResetType::OneShot, "ISelfController:LaunchableEvent");
+        Kernel::Event::Create(Kernel::ResetType::Sticky, "ISelfController:LaunchableEvent");
 }
 
 void ISelfController::SetFocusHandlingMode(Kernel::HLERequestContext& ctx) {
@@ -436,13 +436,13 @@ public:
         static const FunctionInfo functions[] = {
             {0, &ILibraryAppletAccessor::GetAppletStateChangedEvent, "GetAppletStateChangedEvent"},
             {1, nullptr, "IsCompleted"},
-            {10, nullptr, "Start"},
+            {10, &ILibraryAppletAccessor::Start, "Start"},
             {20, nullptr, "RequestExit"},
             {25, nullptr, "Terminate"},
-            {30, nullptr, "GetResult"},
+            {30, &ILibraryAppletAccessor::GetResult, "GetResult"},
             {50, nullptr, "SetOutOfFocusApplicationSuspendingEnabled"},
             {100, &ILibraryAppletAccessor::PushInData, "PushInData"},
-            {101, nullptr, "PopOutData"},
+            {101, &ILibraryAppletAccessor::PopOutData, "PopOutData"},
             {102, nullptr, "PushExtraStorage"},
             {103, nullptr, "PushInteractiveInData"},
             {104, nullptr, "PopInteractiveOutData"},
@@ -470,6 +470,20 @@ private:
         NGLOG_WARNING(Service_AM, "(STUBBED) called");
     }
 
+    void GetResult(Kernel::HLERequestContext& ctx) {
+        IPC::ResponseBuilder rb{ctx, 2};
+        rb.Push(RESULT_SUCCESS);
+
+        NGLOG_WARNING(Service_AM, "(STUBBED) called");
+    }
+
+    void Start(Kernel::HLERequestContext& ctx) {
+        IPC::ResponseBuilder rb{ctx, 2};
+        rb.Push(RESULT_SUCCESS);
+
+        NGLOG_WARNING(Service_AM, "(STUBBED) called");
+    }
+
     void PushInData(Kernel::HLERequestContext& ctx) {
         IPC::RequestParser rp{ctx};
         storage_stack.push(rp.PopIpcInterface<AM::IStorage>());
@@ -480,6 +494,16 @@ private:
         NGLOG_DEBUG(Service_AM, "called");
     }
 
+    void PopOutData(Kernel::HLERequestContext& ctx) {
+        IPC::ResponseBuilder rb{ctx, 2, 0, 1};
+        rb.Push(RESULT_SUCCESS);
+        rb.PushIpcInterface<AM::IStorage>(std::move(storage_stack.top()));
+
+        storage_stack.pop();
+
+        NGLOG_DEBUG(Service_AM, "called");
+    }
+
     std::stack<std::shared_ptr<AM::IStorage>> storage_stack;
     Kernel::SharedPtr<Kernel::Event> state_changed_event;
 };

src/core/hle/service/nvdrv/devices/nvhost_gpu.h

@@ -26,11 +26,19 @@ public:
 private:
     enum class IoctlCommand : u32_le {
         IocSetNVMAPfdCommand = 0x40044801,
+        IocAllocGPFIFOCommand = 0x40084805,
         IocSetClientDataCommand = 0x40084714,
         IocGetClientDataCommand = 0x80084715,
         IocZCullBind = 0xc010480b,
         IocSetErrorNotifierCommand = 0xC018480C,
         IocChannelSetPriorityCommand = 0x4004480D,
+        IocEnableCommand = 0x0000480E,
+        IocDisableCommand = 0x0000480F,
+        IocPreemptCommand = 0x00004810,
+        IocForceResetCommand = 0x00004811,
+        IocEventIdControlCommand = 0x40084812,
+        IocGetErrorNotificationCommand = 0xC0104817,
+        IocAllocGPFIFOExCommand = 0x40204818,
         IocAllocGPFIFOEx2Command = 0xC020481A,
         IocAllocObjCtxCommand = 0xC0104809,
         IocChannelGetWaitbaseCommand = 0xC0080003,
@@ -56,6 +64,12 @@ private:
     };
     static_assert(sizeof(IoctlChannelSetTimeout) == 4, "IoctlChannelSetTimeout is incorrect size");
 
+    struct IoctlAllocGPFIFO {
+        u32_le num_entries;
+        u32_le flags;
+    };
+    static_assert(sizeof(IoctlAllocGPFIFO) == 8, "IoctlAllocGPFIFO is incorrect size");
+
     struct IoctlClientData {
         u64_le data;
     };
@@ -76,12 +90,45 @@ private:
     };
     static_assert(sizeof(IoctlSetErrorNotifier) == 24, "IoctlSetErrorNotifier is incorrect size");
 
+    struct IoctlChannelSetPriority {
+        u32_le priority;
+    };
+    static_assert(sizeof(IoctlChannelSetPriority) == 4,
+                  "IoctlChannelSetPriority is incorrect size");
+
+    struct IoctlEventIdControl {
+        u32_le cmd; // 0=disable, 1=enable, 2=clear
+        u32_le id;
+    };
+    static_assert(sizeof(IoctlEventIdControl) == 8, "IoctlEventIdControl is incorrect size");
+
+    struct IoctlGetErrorNotification {
+        u64_le timestamp;
+        u32_le info32;
+        u16_le info16;
+        u16_le status; // always 0xFFFF
+    };
+    static_assert(sizeof(IoctlGetErrorNotification) == 16,
+                  "IoctlGetErrorNotification is incorrect size");
+
     struct IoctlFence {
         u32_le id;
         u32_le value;
     };
     static_assert(sizeof(IoctlFence) == 8, "IoctlFence is incorrect size");
 
+    struct IoctlAllocGpfifoEx {
+        u32_le num_entries;
+        u32_le flags;
+        u32_le unk0;
+        u32_le unk1;
+        u32_le unk2;
+        u32_le unk3;
+        u32_le unk4;
+        u32_le unk5;
+    };
+    static_assert(sizeof(IoctlAllocGpfifoEx) == 32, "IoctlAllocGpfifoEx is incorrect size");
+
     struct IoctlAllocGpfifoEx2 {
         u32_le num_entries;   // in
         u32_le flags;         // in

src/video_core/engines/maxwell_3d.h

@@ -354,10 +354,35 @@ public:
                     f32 scale_x;
                     f32 scale_y;
                     f32 scale_z;
-                    u32 translate_x;
-                    u32 translate_y;
-                    u32 translate_z;
+                    f32 translate_x;
+                    f32 translate_y;
+                    f32 translate_z;
                     INSERT_PADDING_WORDS(2);
+
+                    MathUtil::Rectangle<s32> GetRect() const {
+                        return {
+                            GetX(),               // left
+                            GetY() + GetHeight(), // top
+                            GetX() + GetWidth(),  // right
+                            GetY()                // bottom
+                        };
+                    };
+
+                    s32 GetX() const {
+                        return static_cast<s32>(std::max(0.0f, translate_x - std::fabs(scale_x)));
+                    }
+
+                    s32 GetY() const {
+                        return static_cast<s32>(std::max(0.0f, translate_y - std::fabs(scale_y)));
+                    }
+
+                    s32 GetWidth() const {
+                        return static_cast<s32>(translate_x + std::fabs(scale_x)) - GetX();
+                    }
+
+                    s32 GetHeight() const {
+                        return static_cast<s32>(translate_y + std::fabs(scale_y)) - GetY();
+                    }
                 } viewport_transform[NumViewports];
 
                 struct {
@@ -371,15 +396,6 @@ public:
                     };
                     float depth_range_near;
                     float depth_range_far;
-
-                    MathUtil::Rectangle<s32> GetRect() const {
-                        return {
-                            static_cast<s32>(x),          // left
-                            static_cast<s32>(y + height), // top
-                            static_cast<s32>(x + width),  // right
-                            static_cast<s32>(y)           // bottom
-                        };
-                    };
                 } viewport[NumViewports];
 
                 INSERT_PADDING_WORDS(0x1D);

src/video_core/engines/shader_bytecode.h

@@ -288,6 +288,19 @@ union Instruction {
         }
     } texs;
 
+    union {
+        BitField<20, 5, u64> target;
+        BitField<5, 1, u64> constant_buffer;
+
+        s32 GetBranchTarget() const {
+            // Sign extend the branch target offset
+            u32 mask = 1U << (5 - 1);
+            u32 value = static_cast<u32>(target);
+            // The branch offset is relative to the next instruction, so add 1 to it.
+            return static_cast<s32>((value ^ mask) - mask) + 1;
+        }
+    } bra;
+
     BitField<61, 1, u64> is_b_imm;
     BitField<60, 1, u64> is_b_gpr;
     BitField<59, 1, u64> is_c_gpr;
@@ -306,6 +319,7 @@ class OpCode {
 public:
     enum class Id {
         KIL,
+        BRA,
         LD_A,
         ST_A,
         TEX,
@@ -470,6 +484,7 @@ private:
         std::vector<Matcher> table = {
 #define INST(bitstring, op, type, name) Detail::GetMatcher(bitstring, op, type, name)
             INST("111000110011----", Id::KIL, Type::Flow, "KIL"),
+            INST("111000100100----", Id::BRA, Type::Flow, "BRA"),
             INST("1110111111011---", Id::LD_A, Type::Memory, "LD_A"),
             INST("1110111111110---", Id::ST_A, Type::Memory, "ST_A"),
             INST("1100000000111---", Id::TEX, Type::Memory, "TEX"),

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -298,7 +298,7 @@ void RasterizerOpenGL::DrawArrays() {
     const bool has_stencil = false;
     const bool using_color_fb = true;
     const bool using_depth_fb = false;
-    const MathUtil::Rectangle<s32> viewport_rect{regs.viewport[0].GetRect()};
+    const MathUtil::Rectangle<s32> viewport_rect{regs.viewport_transform[0].GetRect()};
 
     const bool write_color_fb =
         state.color_mask.red_enabled == GL_TRUE || state.color_mask.green_enabled == GL_TRUE ||
@@ -702,7 +702,7 @@ void RasterizerOpenGL::BindFramebufferSurfaces(const Surface& color_surface,
 
 void RasterizerOpenGL::SyncViewport(const MathUtil::Rectangle<u32>& surfaces_rect, u16 res_scale) {
     const auto& regs = Core::System().GetInstance().GPU().Maxwell3D().regs;
-    const MathUtil::Rectangle<s32> viewport_rect{regs.viewport[0].GetRect()};
+    const MathUtil::Rectangle<s32> viewport_rect{regs.viewport_transform[0].GetRect()};
 
     state.viewport.x = static_cast<GLint>(surfaces_rect.left) + viewport_rect.left * res_scale;
     state.viewport.y = static_cast<GLint>(surfaces_rect.bottom) + viewport_rect.bottom * res_scale;

src/video_core/renderer_opengl/gl_rasterizer_cache.cpp

@@ -933,7 +933,8 @@ Surface RasterizerCacheOpenGL::GetSurface(const SurfaceParams& params, ScaleMatc
     // Use GetSurfaceSubRect instead
     ASSERT(params.width == params.stride);
 
-    ASSERT(!params.is_tiled || (params.width % 8 == 0 && params.height % 8 == 0));
+    ASSERT(!params.is_tiled ||
+           (params.GetActualWidth() % 8 == 0 && params.GetActualHeight() % 8 == 0));
 
     // Check for an exact match in existing surfaces
     Surface surface =

src/video_core/renderer_opengl/gl_shader_decompiler.cpp

@@ -88,6 +88,20 @@ private:
         return *subroutines.insert(std::move(subroutine)).first;
     }
 
+    /// Merges exit method of two parallel branches.
+    static ExitMethod ParallelExit(ExitMethod a, ExitMethod b) {
+        if (a == ExitMethod::Undetermined) {
+            return b;
+        }
+        if (b == ExitMethod::Undetermined) {
+            return a;
+        }
+        if (a == b) {
+            return a;
+        }
+        return ExitMethod::Conditional;
+    }
+
     /// Scans a range of code for labels and determines the exit method.
     ExitMethod Scan(u32 begin, u32 end, std::set<u32>& labels) {
         auto [iter, inserted] =
@@ -97,10 +111,27 @@ private:
             return exit_method;
 
         for (u32 offset = begin; offset != end && offset != PROGRAM_END; ++offset) {
-            if (const auto opcode = OpCode::Decode({program_code[offset]})) {
+            const Instruction instr = {program_code[offset]};
+            if (const auto opcode = OpCode::Decode(instr)) {
                 switch (opcode->GetId()) {
                 case OpCode::Id::EXIT: {
-                    return exit_method = ExitMethod::AlwaysEnd;
+                    // The EXIT instruction can be predicated, which means that the shader can
+                    // conditionally end on this instruction. We have to consider the case where the
+                    // condition is not met and check the exit method of that other basic block.
+                    using Tegra::Shader::Pred;
+                    if (instr.pred.pred_index == static_cast<u64>(Pred::UnusedIndex)) {
+                        return exit_method = ExitMethod::AlwaysEnd;
+                    } else {
+                        ExitMethod not_met = Scan(offset + 1, end, labels);
+                        return exit_method = ParallelExit(ExitMethod::AlwaysEnd, not_met);
+                    }
+                }
+                case OpCode::Id::BRA: {
+                    u32 target = offset + instr.bra.GetBranchTarget();
+                    labels.insert(target);
+                    ExitMethod no_jmp = Scan(offset + 1, end, labels);
+                    ExitMethod jmp = Scan(target, end, labels);
+                    return exit_method = ParallelExit(no_jmp, jmp);
                 }
                 }
             }
@@ -1062,9 +1093,6 @@ private:
         default: {
             switch (opcode->GetId()) {
             case OpCode::Id::EXIT: {
-                ASSERT_MSG(instr.pred.pred_index == static_cast<u64>(Pred::UnusedIndex),
-                           "Predicated exits not implemented");
-
                 // Final color output is currently hardcoded to GPR0-3 for fragment shaders
                 if (stage == Maxwell3D::Regs::ShaderStage::Fragment) {
                     shader.AddLine("color.r = " + regs.GetRegisterAsFloat(0) + ';');
@@ -1074,13 +1102,25 @@ private:
                 }
 
                 shader.AddLine("return true;");
-                offset = PROGRAM_END - 1;
+                if (instr.pred.pred_index == static_cast<u64>(Pred::UnusedIndex)) {
+                    // If this is an unconditional exit then just end processing here, otherwise we
+                    // have to account for the possibility of the condition not being met, so
+                    // continue processing the next instruction.
+                    offset = PROGRAM_END - 1;
+                }
                 break;
             }
             case OpCode::Id::KIL: {
                 shader.AddLine("discard;");
                 break;
             }
+            case OpCode::Id::BRA: {
+                ASSERT_MSG(instr.bra.constant_buffer == 0,
+                           "BRA with constant buffers are not implemented");
+                u32 target = offset + instr.bra.GetBranchTarget();
+                shader.AddLine("{ jmp_to = " + std::to_string(target) + "u; break; }");
+                break;
+            }
             case OpCode::Id::IPA: {
                 const auto& attribute = instr.attribute.fmt28;
                 regs.SetRegisterToInputAttibute(instr.gpr0, attribute.element, attribute.index);

src/yuzu/main.cpp

@@ -335,6 +335,24 @@ void GMainWindow::OnDisplayTitleBars(bool show) {
     }
 }
 
+bool GMainWindow::SupportsRequiredGLExtensions() {
+    QStringList unsupported_ext;
+
+    if (!GLAD_GL_ARB_program_interface_query)
+        unsupported_ext.append("ARB_program_interface_query");
+    if (!GLAD_GL_ARB_separate_shader_objects)
+        unsupported_ext.append("ARB_separate_shader_objects");
+    if (!GLAD_GL_ARB_shader_storage_buffer_object)
+        unsupported_ext.append("ARB_shader_storage_buffer_object");
+    if (!GLAD_GL_ARB_vertex_attrib_binding)
+        unsupported_ext.append("ARB_vertex_attrib_binding");
+
+    for (const QString& ext : unsupported_ext)
+        NGLOG_CRITICAL(Frontend, "Unsupported GL extension: {}", ext.toStdString());
+
+    return unsupported_ext.empty();
+}
+
 bool GMainWindow::LoadROM(const QString& filename) {
     // Shutdown previous session if the emu thread is still active...
     if (emu_thread != nullptr)
@@ -350,6 +368,14 @@ bool GMainWindow::LoadROM(const QString& filename) {
         return false;
     }
 
+    if (!SupportsRequiredGLExtensions()) {
+        QMessageBox::critical(
+            this, tr("Error while initializing OpenGL Core!"),
+            tr("Your GPU may not support one or more required OpenGL extensions. Please "
+               "ensure you have the latest graphics driver. See the log for more details."));
+        return false;
+    }
+
     Core::System& system{Core::System::GetInstance()};
 
     system.SetGPUDebugContext(debug_context);

src/yuzu/main.h

@@ -79,6 +79,7 @@ private:
     void ConnectWidgetEvents();
     void ConnectMenuEvents();
 
+    bool SupportsRequiredGLExtensions();
     bool LoadROM(const QString& filename);
     void BootGame(const QString& filename);
     void ShutdownGame();

src/yuzu_cmd/emu_window/emu_window_sdl2.cpp

@@ -78,6 +78,24 @@ void EmuWindow_SDL2::Fullscreen() {
     SDL_MaximizeWindow(render_window);
 }
 
+bool EmuWindow_SDL2::SupportsRequiredGLExtensions() {
+    std::vector<std::string> unsupported_ext;
+
+    if (!GLAD_GL_ARB_program_interface_query)
+        unsupported_ext.push_back("ARB_program_interface_query");
+    if (!GLAD_GL_ARB_separate_shader_objects)
+        unsupported_ext.push_back("ARB_separate_shader_objects");
+    if (!GLAD_GL_ARB_shader_storage_buffer_object)
+        unsupported_ext.push_back("ARB_shader_storage_buffer_object");
+    if (!GLAD_GL_ARB_vertex_attrib_binding)
+        unsupported_ext.push_back("ARB_vertex_attrib_binding");
+
+    for (const std::string& ext : unsupported_ext)
+        NGLOG_CRITICAL(Frontend, "Unsupported GL extension: {}", ext);
+
+    return unsupported_ext.empty();
+}
+
 EmuWindow_SDL2::EmuWindow_SDL2(bool fullscreen) {
     InputCommon::Init();
 
@@ -128,6 +146,11 @@ EmuWindow_SDL2::EmuWindow_SDL2(bool fullscreen) {
         exit(1);
     }
 
+    if (!SupportsRequiredGLExtensions()) {
+        NGLOG_CRITICAL(Frontend, "GPU does not support all required OpenGL extensions! Exiting...");
+        exit(1);
+    }
+
     OnResize();
     OnMinimalClientAreaChangeRequest(GetActiveConfig().min_client_area_size);
     SDL_PumpEvents();

src/yuzu_cmd/emu_window/emu_window_sdl2.h

@@ -46,6 +46,9 @@ private:
     /// Called when user passes the fullscreen parameter flag
     void Fullscreen();
 
+    /// Whether the GPU and driver supports the OpenGL extension required
+    bool SupportsRequiredGLExtensions();
+
     /// Called when a configuration change affects the minimal size of the window
     void OnMinimalClientAreaChangeRequest(
         const std::pair<unsigned, unsigned>& minimal_size) override;