9 files changed, 234 insertions, 33 deletions

diff --git a/src/core/hle/service/nfp/nfp.cpp b/src/core/hle/service/nfp/nfp.cpp
index 2af4465de..2a9f84037 100644
--- a/src/core/hle/service/nfp/nfp.cpp
+++ b/src/core/hle/service/nfp/nfp.cpp
@@ -4,6 +4,8 @@
 
 #include "common/logging/log.h"
 #include "core/hle/ipc_helpers.h"
+#include "core/hle/kernel/event.h"
+#include "core/hle/service/hid/hid.h"
 #include "core/hle/service/nfp/nfp.h"
 #include "core/hle/service/nfp/nfp_user.h"
 
@@ -18,7 +20,7 @@ public:
         static const FunctionInfo functions[] = {
             {0, &IUser::Initialize, "Initialize"},
             {1, nullptr, "Finalize"},
-            {2, nullptr, "ListDevices"},
+            {2, &IUser::ListDevices, "ListDevices"},
             {3, nullptr, "StartDetection"},
             {4, nullptr, "StopDetection"},
             {5, nullptr, "Mount"},
@@ -33,24 +35,116 @@ public:
             {14, nullptr, "GetRegisterInfo"},
             {15, nullptr, "GetCommonInfo"},
             {16, nullptr, "GetModelInfo"},
-            {17, nullptr, "AttachActivateEvent"},
-            {18, nullptr, "AttachDeactivateEvent"},
-            {19, nullptr, "GetState"},
-            {20, nullptr, "GetDeviceState"},
-            {21, nullptr, "GetNpadId"},
+            {17, &IUser::AttachActivateEvent, "AttachActivateEvent"},
+            {18, &IUser::AttachDeactivateEvent, "AttachDeactivateEvent"},
+            {19, &IUser::GetState, "GetState"},
+            {20, &IUser::GetDeviceState, "GetDeviceState"},
+            {21, &IUser::GetNpadId, "GetNpadId"},
             {22, nullptr, "GetApplicationArea2"},
-            {23, nullptr, "AttachAvailabilityChangeEvent"},
+            {23, &IUser::AttachAvailabilityChangeEvent, "AttachAvailabilityChangeEvent"},
             {24, nullptr, "RecreateApplicationArea"},
         };
         RegisterHandlers(functions);
+
+        activate_event = Kernel::Event::Create(Kernel::ResetType::OneShot, "IUser:ActivateEvent");
+        deactivate_event =
+            Kernel::Event::Create(Kernel::ResetType::OneShot, "IUser:DeactivateEvent");
+        availability_change_event =
+            Kernel::Event::Create(Kernel::ResetType::OneShot, "IUser:AvailabilityChangeEvent");
     }
 
 private:
+    enum class State : u32 {
+        NonInitialized = 0,
+        Initialized = 1,
+    };
+
+    enum class DeviceState : u32 {
+        Initialized = 0,
+    };
+
     void Initialize(Kernel::HLERequestContext& ctx) {
         NGLOG_WARNING(Service_NFP, "(STUBBED) called");
+
+        state = State::Initialized;
+
         IPC::ResponseBuilder rb{ctx, 2};
         rb.Push(RESULT_SUCCESS);
     }
+
+    void ListDevices(Kernel::HLERequestContext& ctx) {
+        IPC::RequestParser rp{ctx};
+        const u32 array_size = rp.Pop<u32>();
+
+        ctx.WriteBuffer(&device_handle, sizeof(device_handle));
+
+        NGLOG_WARNING(Service_NFP, "(STUBBED) called, array_size={}", array_size);
+
+        IPC::ResponseBuilder rb{ctx, 3};
+        rb.Push(RESULT_SUCCESS);
+        rb.Push<u32>(0);
+    }
+
+    void AttachActivateEvent(Kernel::HLERequestContext& ctx) {
+        IPC::RequestParser rp{ctx};
+        const u64 dev_handle = rp.Pop<u64>();
+        NGLOG_WARNING(Service_NFP, "(STUBBED) called, dev_handle=0x{:X}", dev_handle);
+
+        IPC::ResponseBuilder rb{ctx, 2, 1};
+        rb.Push(RESULT_SUCCESS);
+        rb.PushCopyObjects(activate_event);
+    }
+
+    void AttachDeactivateEvent(Kernel::HLERequestContext& ctx) {
+        IPC::RequestParser rp{ctx};
+        const u64 dev_handle = rp.Pop<u64>();
+        NGLOG_WARNING(Service_NFP, "(STUBBED) called, dev_handle=0x{:X}", dev_handle);
+
+        IPC::ResponseBuilder rb{ctx, 2, 1};
+        rb.Push(RESULT_SUCCESS);
+        rb.PushCopyObjects(deactivate_event);
+    }
+
+    void GetState(Kernel::HLERequestContext& ctx) {
+        NGLOG_WARNING(Service_NFP, "(STUBBED) called");
+        IPC::ResponseBuilder rb{ctx, 3};
+        rb.Push(RESULT_SUCCESS);
+        rb.Push<u32>(static_cast<u32>(state));
+    }
+
+    void GetDeviceState(Kernel::HLERequestContext& ctx) {
+        NGLOG_WARNING(Service_NFP, "(STUBBED) called");
+        IPC::ResponseBuilder rb{ctx, 3};
+        rb.Push(RESULT_SUCCESS);
+        rb.Push<u32>(static_cast<u32>(device_state));
+    }
+
+    void GetNpadId(Kernel::HLERequestContext& ctx) {
+        IPC::RequestParser rp{ctx};
+        const u64 dev_handle = rp.Pop<u64>();
+        NGLOG_WARNING(Service_NFP, "(STUBBED) called, dev_handle=0x{:X}", dev_handle);
+        IPC::ResponseBuilder rb{ctx, 3};
+        rb.Push(RESULT_SUCCESS);
+        rb.Push<u32>(npad_id);
+    }
+
+    void AttachAvailabilityChangeEvent(Kernel::HLERequestContext& ctx) {
+        IPC::RequestParser rp{ctx};
+        const u64 dev_handle = rp.Pop<u64>();
+        NGLOG_WARNING(Service_NFP, "(STUBBED) called, dev_handle=0x{:X}", dev_handle);
+
+        IPC::ResponseBuilder rb{ctx, 2, 1};
+        rb.Push(RESULT_SUCCESS);
+        rb.PushCopyObjects(availability_change_event);
+    }
+
+    const u64 device_handle{0xDEAD};
+    const HID::ControllerID npad_id{HID::Controller_Player1};
+    State state{State::NonInitialized};
+    DeviceState device_state{DeviceState::Initialized};
+    Kernel::SharedPtr<Kernel::Event> activate_event;
+    Kernel::SharedPtr<Kernel::Event> deactivate_event;
+    Kernel::SharedPtr<Kernel::Event> availability_change_event;
 };
 
 void Module::Interface::CreateUserInterface(Kernel::HLERequestContext& ctx) {
diff --git a/src/video_core/engines/shader_bytecode.h b/src/video_core/engines/shader_bytecode.h
index af18c2d81..2cda1e63e 100644
--- a/src/video_core/engines/shader_bytecode.h
+++ b/src/video_core/engines/shader_bytecode.h
@@ -168,13 +168,22 @@ enum class SubOp : u64 {
     Min = 0x8,
 };
 
-enum class FloatRoundingOp : u64 {
+enum class F2iRoundingOp : u64 {
     None = 0,
     Floor = 1,
     Ceil = 2,
     Trunc = 3,
 };
 
+enum class F2fRoundingOp : u64 {
+    None = 0,
+    Pass = 3,
+    Round = 8,
+    Floor = 9,
+    Ceil = 10,
+    Trunc = 11,
+};
+
 enum class UniformType : u64 {
     UnsignedByte = 0,
     SignedByte = 1,
@@ -314,11 +323,11 @@ union Instruction {
         BitField<50, 1, u64> saturate_a;
 
         union {
-            BitField<39, 2, FloatRoundingOp> rounding;
+            BitField<39, 2, F2iRoundingOp> rounding;
         } f2i;
 
         union {
-            BitField<39, 4, u64> rounding;
+            BitField<39, 4, F2fRoundingOp> rounding;
         } f2f;
     } conversion;
 
@@ -390,6 +399,9 @@ class OpCode {
 public:
     enum class Id {
         KIL,
+        BFE_C,
+        BFE_R,
+        BFE_IMM,
         BRA,
         LD_A,
         LD_C,
@@ -444,6 +456,9 @@ public:
         FMNMX_C,
         FMNMX_R,
         FMNMX_IMM,
+        IMNMX_C,
+        IMNMX_R,
+        IMNMX_IMM,
         FSETP_C, // Set Predicate
         FSETP_R,
         FSETP_IMM,
@@ -454,6 +469,10 @@ public:
         ISETP_IMM,
         ISETP_R,
         PSETP,
+        XMAD_IMM,
+        XMAD_CR,
+        XMAD_RC,
+        XMAD_RR,
     };
 
     enum class Type {
@@ -565,6 +584,9 @@ 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("0100110000000---", Id::BFE_C, Type::Flow, "BFE_C"),
+            INST("0101110000000---", Id::BFE_R, Type::Flow, "BFE_R"),
+            INST("0011100-00000---", Id::BFE_IMM, Type::Flow, "BFE_IMM"),
             INST("111000100100----", Id::BRA, Type::Flow, "BRA"),
             INST("1110111111011---", Id::LD_A, Type::Memory, "LD_A"),
             INST("1110111110010---", Id::LD_C, Type::Memory, "LD_C"),
@@ -606,6 +628,9 @@ private:
             INST("0100110001100---", Id::FMNMX_C, Type::Arithmetic, "FMNMX_C"),
             INST("0101110001100---", Id::FMNMX_R, Type::Arithmetic, "FMNMX_R"),
             INST("0011100-01100---", Id::FMNMX_IMM, Type::Arithmetic, "FMNMX_IMM"),
+            INST("0100110000100---", Id::IMNMX_C, Type::Arithmetic, "FMNMX_IMM"),
+            INST("0101110000100---", Id::IMNMX_R, Type::Arithmetic, "FMNMX_IMM"),
+            INST("0011100-00100---", Id::IMNMX_IMM, Type::Arithmetic, "FMNMX_IMM"),
             INST("000001----------", Id::LOP32I, Type::Logic, "LOP32I"),
             INST("0100110001001---", Id::SHL_C, Type::Shift, "SHL_C"),
             INST("0101110001001---", Id::SHL_R, Type::Shift, "SHL_R"),
@@ -629,6 +654,10 @@ private:
             INST("010110110110----", Id::ISETP_R, Type::IntegerSetPredicate, "ISETP_R"),
             INST("0011011-0110----", Id::ISETP_IMM, Type::IntegerSetPredicate, "ISETP_IMM"),
             INST("0101000010010---", Id::PSETP, Type::PredicateSetPredicate, "PSETP"),
+            INST("0011011-00------", Id::XMAD_IMM, Type::Arithmetic, "XMAD_IMM"),
+            INST("0100111---------", Id::XMAD_CR, Type::Arithmetic, "XMAD_CR"),
+            INST("010100010-------", Id::XMAD_RC, Type::Arithmetic, "XMAD_RC"),
+            INST("0101101100------", Id::XMAD_RR, Type::Arithmetic, "XMAD_RR"),
         };
 #undef INST
         std::stable_sort(table.begin(), table.end(), [](const auto& a, const auto& b) {
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index b23b8fb29..2e90ebcf4 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -681,6 +681,14 @@ u32 RasterizerOpenGL::SetupTextures(Maxwell::ShaderStage stage, GLuint program,
         Surface surface = res_cache.GetTextureSurface(texture);
         if (surface != nullptr) {
             state.texture_units[current_bindpoint].texture_2d = surface->texture.handle;
+            state.texture_units[current_bindpoint].swizzle.r =
+                MaxwellToGL::SwizzleSource(texture.tic.x_source);
+            state.texture_units[current_bindpoint].swizzle.g =
+                MaxwellToGL::SwizzleSource(texture.tic.y_source);
+            state.texture_units[current_bindpoint].swizzle.b =
+                MaxwellToGL::SwizzleSource(texture.tic.z_source);
+            state.texture_units[current_bindpoint].swizzle.a =
+                MaxwellToGL::SwizzleSource(texture.tic.w_source);
         } else {
             // Can occur when texture addr is null or its memory is unmapped/invalid
             state.texture_units[current_bindpoint].texture_2d = 0;
diff --git a/src/video_core/renderer_opengl/gl_shader_decompiler.cpp b/src/video_core/renderer_opengl/gl_shader_decompiler.cpp
index 3067ce3b3..8e249584f 100644
--- a/src/video_core/renderer_opengl/gl_shader_decompiler.cpp
+++ b/src/video_core/renderer_opengl/gl_shader_decompiler.cpp
@@ -1056,10 +1056,27 @@ private:
                 break;
             }
             case OpCode::Id::F2F_R: {
-                // TODO(Subv): Implement rounding operations.
-                ASSERT_MSG(instr.conversion.f2f.rounding == 0, "Unimplemented rounding operation");
                 std::string op_a = regs.GetRegisterAsFloat(instr.gpr20);
 
+                switch (instr.conversion.f2f.rounding) {
+                case Tegra::Shader::F2fRoundingOp::None:
+                    break;
+                case Tegra::Shader::F2fRoundingOp::Floor:
+                    op_a = "floor(" + op_a + ')';
+                    break;
+                case Tegra::Shader::F2fRoundingOp::Ceil:
+                    op_a = "ceil(" + op_a + ')';
+                    break;
+                case Tegra::Shader::F2fRoundingOp::Trunc:
+                    op_a = "trunc(" + op_a + ')';
+                    break;
+                default:
+                    NGLOG_CRITICAL(HW_GPU, "Unimplemented f2f rounding mode {}",
+                                   static_cast<u32>(instr.conversion.f2f.rounding.Value()));
+                    UNREACHABLE();
+                    break;
+                }
+
                 if (instr.conversion.abs_a) {
                     op_a = "abs(" + op_a + ')';
                 }
@@ -1074,17 +1091,16 @@ private:
                     op_a = "abs(" + op_a + ')';
                 }
 
-                using Tegra::Shader::FloatRoundingOp;
                 switch (instr.conversion.f2i.rounding) {
-                case FloatRoundingOp::None:
+                case Tegra::Shader::F2iRoundingOp::None:
                     break;
-                case FloatRoundingOp::Floor:
+                case Tegra::Shader::F2iRoundingOp::Floor:
                     op_a = "floor(" + op_a + ')';
                     break;
-                case FloatRoundingOp::Ceil:
+                case Tegra::Shader::F2iRoundingOp::Ceil:
                     op_a = "ceil(" + op_a + ')';
                     break;
-                case FloatRoundingOp::Trunc:
+                case Tegra::Shader::F2iRoundingOp::Trunc:
                     op_a = "trunc(" + op_a + ')';
                     break;
                 default:
@@ -1112,13 +1128,11 @@ private:
             break;
         }
         case OpCode::Type::Memory: {
-            const Attribute::Index attribute = instr.attribute.fmt20.index;
-
             switch (opcode->GetId()) {
             case OpCode::Id::LD_A: {
                 ASSERT_MSG(instr.attribute.fmt20.size == 0, "untested");
                 regs.SetRegisterToInputAttibute(instr.gpr0, instr.attribute.fmt20.element,
-                                                attribute);
+                                                instr.attribute.fmt20.index);
                 break;
             }
             case OpCode::Id::LD_C: {
@@ -1150,12 +1164,11 @@ private:
             }
             case OpCode::Id::ST_A: {
                 ASSERT_MSG(instr.attribute.fmt20.size == 0, "untested");
-                regs.SetOutputAttributeToRegister(attribute, instr.attribute.fmt20.element,
-                                                  instr.gpr0);
+                regs.SetOutputAttributeToRegister(instr.attribute.fmt20.index,
+                                                  instr.attribute.fmt20.element, instr.gpr0);
                 break;
             }
             case OpCode::Id::TEX: {
-                ASSERT_MSG(instr.attribute.fmt20.size == 4, "untested");
                 const std::string op_a = regs.GetRegisterAsFloat(instr.gpr8);
                 const std::string op_b = regs.GetRegisterAsFloat(instr.gpr8.Value() + 1);
                 const std::string sampler = GetSampler(instr.sampler);
@@ -1168,7 +1181,7 @@ private:
                 const std::string texture = "texture(" + sampler + ", coords)";
 
                 size_t dest_elem{};
-                for (size_t elem = 0; elem < instr.attribute.fmt20.size; ++elem) {
+                for (size_t elem = 0; elem < 4; ++elem) {
                     if (!instr.tex.IsComponentEnabled(elem)) {
                         // Skip disabled components
                         continue;
@@ -1181,7 +1194,6 @@ private:
                 break;
             }
             case OpCode::Id::TEXS: {
-                ASSERT_MSG(instr.attribute.fmt20.size == 4, "untested");
                 const std::string op_a = regs.GetRegisterAsFloat(instr.gpr8);
                 const std::string op_b = regs.GetRegisterAsFloat(instr.gpr20);
                 const std::string sampler = GetSampler(instr.sampler);
@@ -1278,16 +1290,17 @@ private:
         }
         case OpCode::Type::IntegerSetPredicate: {
             std::string op_a = regs.GetRegisterAsInteger(instr.gpr8, 0, instr.isetp.is_signed);
+            std::string op_b;
 
-            std::string op_b{};
-
-            ASSERT_MSG(!instr.is_b_imm, "ISETP_IMM not implemented");
-
-            if (instr.is_b_gpr) {
-                op_b += regs.GetRegisterAsInteger(instr.gpr20, 0, instr.isetp.is_signed);
+            if (instr.is_b_imm) {
+                op_b += '(' + std::to_string(instr.alu.GetSignedImm20_20()) + ')';
             } else {
-                op_b += regs.GetUniform(instr.cbuf34.index, instr.cbuf34.offset,
-                                        GLSLRegister::Type::Integer);
+                if (instr.is_b_gpr) {
+                    op_b += regs.GetRegisterAsInteger(instr.gpr20, 0, instr.isetp.is_signed);
+                } else {
+                    op_b += regs.GetUniform(instr.cbuf34.index, instr.cbuf34.offset,
+                                            GLSLRegister::Type::Integer);
+                }
             }
 
             using Tegra::Shader::Pred;
diff --git a/src/video_core/renderer_opengl/gl_state.cpp b/src/video_core/renderer_opengl/gl_state.cpp
index f91dfe36a..44f0c8a01 100644
--- a/src/video_core/renderer_opengl/gl_state.cpp
+++ b/src/video_core/renderer_opengl/gl_state.cpp
@@ -50,6 +50,10 @@ OpenGLState::OpenGLState() {
     for (auto& texture_unit : texture_units) {
         texture_unit.texture_2d = 0;
         texture_unit.sampler = 0;
+        texture_unit.swizzle.r = GL_RED;
+        texture_unit.swizzle.g = GL_GREEN;
+        texture_unit.swizzle.b = GL_BLUE;
+        texture_unit.swizzle.a = GL_ALPHA;
     }
 
     lighting_lut.texture_buffer = 0;
@@ -200,6 +204,15 @@ void OpenGLState::Apply() const {
         if (texture_units[i].sampler != cur_state.texture_units[i].sampler) {
             glBindSampler(i, texture_units[i].sampler);
         }
+        // Update the texture swizzle
+        if (texture_units[i].swizzle.r != cur_state.texture_units[i].swizzle.r ||
+            texture_units[i].swizzle.g != cur_state.texture_units[i].swizzle.g ||
+            texture_units[i].swizzle.b != cur_state.texture_units[i].swizzle.b ||
+            texture_units[i].swizzle.a != cur_state.texture_units[i].swizzle.a) {
+            std::array<GLint, 4> mask = {texture_units[i].swizzle.r, texture_units[i].swizzle.g,
+                                         texture_units[i].swizzle.b, texture_units[i].swizzle.a};
+            glTexParameteriv(GL_TEXTURE_2D, GL_TEXTURE_SWIZZLE_RGBA, mask.data());
+        }
     }
 
     // Constbuffers
diff --git a/src/video_core/renderer_opengl/gl_state.h b/src/video_core/renderer_opengl/gl_state.h
index 75c08e645..839e50e93 100644
--- a/src/video_core/renderer_opengl/gl_state.h
+++ b/src/video_core/renderer_opengl/gl_state.h
@@ -85,6 +85,12 @@ public:
     struct {
         GLuint texture_2d; // GL_TEXTURE_BINDING_2D
         GLuint sampler;    // GL_SAMPLER_BINDING
+        struct {
+            GLint r; // GL_TEXTURE_SWIZZLE_R
+            GLint g; // GL_TEXTURE_SWIZZLE_G
+            GLint b; // GL_TEXTURE_SWIZZLE_B
+            GLint a; // GL_TEXTURE_SWIZZLE_A
+        } swizzle;
     } texture_units[32];
 
     struct {
diff --git a/src/video_core/renderer_opengl/maxwell_to_gl.h b/src/video_core/renderer_opengl/maxwell_to_gl.h
index cf11983cf..2155fb019 100644
--- a/src/video_core/renderer_opengl/maxwell_to_gl.h
+++ b/src/video_core/renderer_opengl/maxwell_to_gl.h
@@ -180,4 +180,25 @@ inline GLenum BlendFunc(Maxwell::Blend::Factor factor) {
     return {};
 }
 
+inline GLenum SwizzleSource(Tegra::Texture::SwizzleSource source) {
+    switch (source) {
+    case Tegra::Texture::SwizzleSource::Zero:
+        return GL_ZERO;
+    case Tegra::Texture::SwizzleSource::R:
+        return GL_RED;
+    case Tegra::Texture::SwizzleSource::G:
+        return GL_GREEN;
+    case Tegra::Texture::SwizzleSource::B:
+        return GL_BLUE;
+    case Tegra::Texture::SwizzleSource::A:
+        return GL_ALPHA;
+    case Tegra::Texture::SwizzleSource::OneInt:
+    case Tegra::Texture::SwizzleSource::OneFloat:
+        return GL_ONE;
+    }
+    NGLOG_CRITICAL(Render_OpenGL, "Unimplemented swizzle source={}", static_cast<u32>(source));
+    UNREACHABLE();
+    return {};
+}
+
 } // namespace MaxwellToGL
diff --git a/src/video_core/renderer_opengl/renderer_opengl.cpp b/src/video_core/renderer_opengl/renderer_opengl.cpp
index 3440d2190..f33766bfd 100644
--- a/src/video_core/renderer_opengl/renderer_opengl.cpp
+++ b/src/video_core/renderer_opengl/renderer_opengl.cpp
@@ -316,6 +316,7 @@ void RendererOpenGL::DrawScreenTriangles(const ScreenInfo& screen_info, float x,
     }};
 
     state.texture_units[0].texture_2d = screen_info.display_texture;
+    state.texture_units[0].swizzle = {GL_RED, GL_GREEN, GL_BLUE, GL_ALPHA};
     state.Apply();
 
     glBufferSubData(GL_ARRAY_BUFFER, 0, sizeof(vertices), vertices.data());
diff --git a/src/video_core/textures/texture.h b/src/video_core/textures/texture.h
index f48ca30b8..a17eaf19d 100644
--- a/src/video_core/textures/texture.h
+++ b/src/video_core/textures/texture.h
@@ -122,6 +122,17 @@ enum class ComponentType : u32 {
     FLOAT = 7
 };
 
+enum class SwizzleSource : u32 {
+    Zero = 0,
+
+    R = 2,
+    G = 3,
+    B = 4,
+    A = 5,
+    OneInt = 6,
+    OneFloat = 7,
+};
+
 union TextureHandle {
     u32 raw;
     BitField<0, 20, u32> tic_id;
@@ -139,6 +150,11 @@ struct TICEntry {
         BitField<10, 3, ComponentType> g_type;
         BitField<13, 3, ComponentType> b_type;
         BitField<16, 3, ComponentType> a_type;
+
+        BitField<19, 3, SwizzleSource> x_source;
+        BitField<22, 3, SwizzleSource> y_source;
+        BitField<25, 3, SwizzleSource> z_source;
+        BitField<28, 3, SwizzleSource> w_source;
     };
     u32 address_low;
     union {