Shader: Remove an unneeded assert, the negate bit is implemented for conversion instructions.

Added pred-condition GreaterThanWithNan

src/common/hex_util.cpp

@@ -4,6 +4,8 @@
 
 #include "common/hex_util.h"
 
+namespace Common {
+
 u8 ToHexNibble(char c1) {
     if (c1 >= 65 && c1 <= 70)
         return c1 - 55;
@@ -25,3 +27,5 @@ std::array<u8, 32> operator""_array32(const char* str, size_t len) {
         throw std::logic_error("Not of correct size.");
     return HexStringToArray<32>(str);
 }
+
+} // namespace Common

src/common/hex_util.h

@@ -10,6 +10,8 @@
 #include <fmt/format.h>
 #include "common/common_types.h"
 
+namespace Common {
+
 u8 ToHexNibble(char c1);
 
 template <size_t Size, bool le = false>
@@ -35,3 +37,5 @@ std::string HexArrayToString(std::array<u8, Size> array, bool upper = true) {
 
 std::array<u8, 0x10> operator"" _array16(const char* str, size_t len);
 std::array<u8, 0x20> operator"" _array32(const char* str, size_t len);
+
+} // namespace Common

src/core/arm/dynarmic/arm_dynarmic.cpp

@@ -134,6 +134,9 @@ std::unique_ptr<Dynarmic::A64::Jit> ARM_Dynarmic::MakeJit() const {
     config.dczid_el0 = 4;
     config.ctr_el0 = 0x8444c004;
 
+    // Unpredictable instructions
+    config.define_unpredictable_behaviour = true;
+
     return std::make_unique<Dynarmic::A64::Jit>(config);
 }
 

src/core/core.h

@@ -40,6 +40,12 @@ namespace Core {
 
 class System {
 public:
+    System(const System&) = delete;
+    System& operator=(const System&) = delete;
+
+    System(System&&) = delete;
+    System& operator=(System&&) = delete;
+
     ~System();
 
     /**

src/core/crypto/key_manager.cpp

@@ -52,20 +52,20 @@ void KeyManager::LoadFromFile(const std::string& filename, bool is_title_keys) {
         out[1].erase(std::remove(out[1].begin(), out[1].end(), ' '), out[1].end());
 
         if (is_title_keys) {
-            auto rights_id_raw = HexStringToArray<16>(out[0]);
+            auto rights_id_raw = Common::HexStringToArray<16>(out[0]);
             u128 rights_id{};
             std::memcpy(rights_id.data(), rights_id_raw.data(), rights_id_raw.size());
-            Key128 key = HexStringToArray<16>(out[1]);
+            Key128 key = Common::HexStringToArray<16>(out[1]);
             SetKey(S128KeyType::Titlekey, key, rights_id[1], rights_id[0]);
         } else {
             std::transform(out[0].begin(), out[0].end(), out[0].begin(), ::tolower);
             if (s128_file_id.find(out[0]) != s128_file_id.end()) {
                 const auto index = s128_file_id.at(out[0]);
-                Key128 key = HexStringToArray<16>(out[1]);
+                Key128 key = Common::HexStringToArray<16>(out[1]);
                 SetKey(index.type, key, index.field1, index.field2);
             } else if (s256_file_id.find(out[0]) != s256_file_id.end()) {
                 const auto index = s256_file_id.at(out[0]);
-                Key256 key = HexStringToArray<32>(out[1]);
+                Key256 key = Common::HexStringToArray<32>(out[1]);
                 SetKey(index.type, key, index.field1, index.field2);
             }
         }

src/core/file_sys/registered_cache.cpp

@@ -37,11 +37,12 @@ static bool FollowsNcaIdFormat(std::string_view name) {
 static std::string GetRelativePathFromNcaID(const std::array<u8, 16>& nca_id, bool second_hex_upper,
                                             bool within_two_digit) {
     if (!within_two_digit)
-        return fmt::format("/{}.nca", HexArrayToString(nca_id, second_hex_upper));
+        return fmt::format("/{}.nca", Common::HexArrayToString(nca_id, second_hex_upper));
 
     Core::Crypto::SHA256Hash hash{};
     mbedtls_sha256(nca_id.data(), nca_id.size(), hash.data(), 0);
-    return fmt::format("/000000{:02X}/{}.nca", hash[0], HexArrayToString(nca_id, second_hex_upper));
+    return fmt::format("/000000{:02X}/{}.nca", hash[0],
+                       Common::HexArrayToString(nca_id, second_hex_upper));
 }
 
 static std::string GetCNMTName(TitleType type, u64 title_id) {
@@ -170,7 +171,7 @@ std::vector<NcaID> RegisteredCache::AccumulateFiles() const {
     std::vector<NcaID> ids;
     for (const auto& d2_dir : dir->GetSubdirectories()) {
         if (FollowsNcaIdFormat(d2_dir->GetName())) {
-            ids.push_back(HexStringToArray<0x10, true>(d2_dir->GetName().substr(0, 0x20)));
+            ids.push_back(Common::HexStringToArray<0x10, true>(d2_dir->GetName().substr(0, 0x20)));
             continue;
         }
 
@@ -181,20 +182,21 @@ std::vector<NcaID> RegisteredCache::AccumulateFiles() const {
             if (!FollowsNcaIdFormat(nca_dir->GetName()))
                 continue;
 
-            ids.push_back(HexStringToArray<0x10, true>(nca_dir->GetName().substr(0, 0x20)));
+            ids.push_back(Common::HexStringToArray<0x10, true>(nca_dir->GetName().substr(0, 0x20)));
         }
 
         for (const auto& nca_file : d2_dir->GetFiles()) {
             if (!FollowsNcaIdFormat(nca_file->GetName()))
                 continue;
 
-            ids.push_back(HexStringToArray<0x10, true>(nca_file->GetName().substr(0, 0x20)));
+            ids.push_back(
+                Common::HexStringToArray<0x10, true>(nca_file->GetName().substr(0, 0x20)));
         }
     }
 
     for (const auto& d2_file : dir->GetFiles()) {
         if (FollowsNcaIdFormat(d2_file->GetName()))
-            ids.push_back(HexStringToArray<0x10, true>(d2_file->GetName().substr(0, 0x20)));
+            ids.push_back(Common::HexStringToArray<0x10, true>(d2_file->GetName().substr(0, 0x20)));
     }
     return ids;
 }
@@ -339,7 +341,7 @@ std::vector<RegisteredCacheEntry> RegisteredCache::ListEntriesFilter(
 }
 
 static std::shared_ptr<NCA> GetNCAFromXCIForID(std::shared_ptr<XCI> xci, const NcaID& id) {
-    const auto filename = fmt::format("{}.nca", HexArrayToString(id, false));
+    const auto filename = fmt::format("{}.nca", Common::HexArrayToString(id, false));
     const auto iter =
         std::find_if(xci->GetNCAs().begin(), xci->GetNCAs().end(),
                      [&filename](std::shared_ptr<NCA> nca) { return nca->GetName() == filename; });
@@ -361,7 +363,7 @@ InstallResult RegisteredCache::InstallEntry(std::shared_ptr<XCI> xci, bool overw
 
     // Install Metadata File
     const auto meta_id_raw = (*meta_iter)->GetName().substr(0, 32);
-    const auto meta_id = HexStringToArray<16>(meta_id_raw);
+    const auto meta_id = Common::HexStringToArray<16>(meta_id_raw);
 
     const auto res = RawInstallNCA(*meta_iter, copy, overwrite_if_exists, meta_id);
     if (res != InstallResult::Success)

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

@@ -145,8 +145,8 @@ ISelfController::ISelfController(std::shared_ptr<NVFlinger::NVFlinger> nvflinger
         {51, nullptr, "ApproveToDisplay"},
         {60, nullptr, "OverrideAutoSleepTimeAndDimmingTime"},
         {61, nullptr, "SetMediaPlaybackState"},
-        {62, nullptr, "SetIdleTimeDetectionExtension"},
-        {63, nullptr, "GetIdleTimeDetectionExtension"},
+        {62, &ISelfController::SetIdleTimeDetectionExtension, "SetIdleTimeDetectionExtension"},
+        {63, &ISelfController::GetIdleTimeDetectionExtension, "GetIdleTimeDetectionExtension"},
         {64, nullptr, "SetInputDetectionSourceSet"},
         {65, nullptr, "ReportUserIsActive"},
         {66, nullptr, "GetCurrentIlluminance"},
@@ -281,6 +281,23 @@ void ISelfController::SetHandlesRequestToDisplay(Kernel::HLERequestContext& ctx)
     LOG_WARNING(Service_AM, "(STUBBED) called");
 }
 
+void ISelfController::SetIdleTimeDetectionExtension(Kernel::HLERequestContext& ctx) {
+    IPC::RequestParser rp{ctx};
+    idle_time_detection_extension = rp.Pop<u32>();
+    IPC::ResponseBuilder rb{ctx, 2};
+    rb.Push(RESULT_SUCCESS);
+
+    LOG_WARNING(Service_AM, "(STUBBED) called");
+}
+
+void ISelfController::GetIdleTimeDetectionExtension(Kernel::HLERequestContext& ctx) {
+    IPC::ResponseBuilder rb{ctx, 3};
+    rb.Push(RESULT_SUCCESS);
+    rb.Push<u32>(idle_time_detection_extension);
+
+    LOG_WARNING(Service_AM, "(STUBBED) called");
+}
+
 ICommonStateGetter::ICommonStateGetter() : ServiceFramework("ICommonStateGetter") {
     static const FunctionInfo functions[] = {
         {0, &ICommonStateGetter::GetEventHandle, "GetEventHandle"},
@@ -306,7 +323,8 @@ ICommonStateGetter::ICommonStateGetter() : ServiceFramework("ICommonStateGetter"
         {52, nullptr, "SwitchLcdBacklight"},
         {55, nullptr, "IsInControllerFirmwareUpdateSection"},
         {60, nullptr, "GetDefaultDisplayResolution"},
-        {61, nullptr, "GetDefaultDisplayResolutionChangeEvent"},
+        {61, &ICommonStateGetter::GetDefaultDisplayResolutionChangeEvent,
+         "GetDefaultDisplayResolutionChangeEvent"},
         {62, nullptr, "GetHdcpAuthenticationState"},
         {63, nullptr, "GetHdcpAuthenticationStateChangeEvent"},
     };
@@ -341,6 +359,16 @@ void ICommonStateGetter::GetCurrentFocusState(Kernel::HLERequestContext& ctx) {
     LOG_WARNING(Service_AM, "(STUBBED) called");
 }
 
+void ICommonStateGetter::GetDefaultDisplayResolutionChangeEvent(Kernel::HLERequestContext& ctx) {
+    event->Signal();
+
+    IPC::ResponseBuilder rb{ctx, 2, 1};
+    rb.Push(RESULT_SUCCESS);
+    rb.PushCopyObjects(event);
+
+    LOG_WARNING(Service_AM, "(STUBBED) called");
+}
+
 void ICommonStateGetter::GetOperationMode(Kernel::HLERequestContext& ctx) {
     const bool use_docked_mode{Settings::values.use_docked_mode};
     IPC::ResponseBuilder rb{ctx, 3};

src/core/hle/service/am/am.h

@@ -87,9 +87,12 @@ private:
     void CreateManagedDisplayLayer(Kernel::HLERequestContext& ctx);
     void SetScreenShotPermission(Kernel::HLERequestContext& ctx);
     void SetHandlesRequestToDisplay(Kernel::HLERequestContext& ctx);
+    void SetIdleTimeDetectionExtension(Kernel::HLERequestContext& ctx);
+    void GetIdleTimeDetectionExtension(Kernel::HLERequestContext& ctx);
 
     std::shared_ptr<NVFlinger::NVFlinger> nvflinger;
     Kernel::SharedPtr<Kernel::Event> launchable_event;
+    u32 idle_time_detection_extension = 0;
 };
 
 class ICommonStateGetter final : public ServiceFramework<ICommonStateGetter> {
@@ -110,6 +113,7 @@ private:
     void GetEventHandle(Kernel::HLERequestContext& ctx);
     void ReceiveMessage(Kernel::HLERequestContext& ctx);
     void GetCurrentFocusState(Kernel::HLERequestContext& ctx);
+    void GetDefaultDisplayResolutionChangeEvent(Kernel::HLERequestContext& ctx);
     void GetOperationMode(Kernel::HLERequestContext& ctx);
     void GetPerformanceMode(Kernel::HLERequestContext& ctx);
 

src/core/loader/nca.cpp

@@ -3,28 +3,22 @@
 // Refer to the license.txt file included.
 
 #include <utility>
-#include <vector>
 
 #include "common/file_util.h"
 #include "common/logging/log.h"
-#include "common/string_util.h"
-#include "common/swap.h"
-#include "core/core.h"
 #include "core/file_sys/content_archive.h"
-#include "core/file_sys/program_metadata.h"
-#include "core/gdbstub/gdbstub.h"
 #include "core/hle/kernel/process.h"
-#include "core/hle/kernel/resource_limit.h"
 #include "core/hle/service/filesystem/filesystem.h"
+#include "core/loader/deconstructed_rom_directory.h"
 #include "core/loader/nca.h"
-#include "core/loader/nso.h"
-#include "core/memory.h"
 
 namespace Loader {
 
 AppLoader_NCA::AppLoader_NCA(FileSys::VirtualFile file_)
     : AppLoader(std::move(file_)), nca(std::make_unique<FileSys::NCA>(file)) {}
 
+AppLoader_NCA::~AppLoader_NCA() = default;
+
 FileType AppLoader_NCA::IdentifyType(const FileSys::VirtualFile& file) {
     FileSys::NCA nca(file);
 
@@ -83,6 +77,4 @@ ResultStatus AppLoader_NCA::ReadProgramId(u64& out_program_id) {
     return ResultStatus::Success;
 }
 
-AppLoader_NCA::~AppLoader_NCA() = default;
-
 } // namespace Loader

src/core/loader/nca.h

@@ -4,20 +4,24 @@
 
 #pragma once
 
-#include <string>
 #include "common/common_types.h"
-#include "core/file_sys/content_archive.h"
-#include "core/file_sys/program_metadata.h"
+#include "core/file_sys/vfs.h"
 #include "core/hle/kernel/object.h"
 #include "core/loader/loader.h"
-#include "deconstructed_rom_directory.h"
+
+namespace FileSys {
+class NCA;
+}
 
 namespace Loader {
 
+class AppLoader_DeconstructedRomDirectory;
+
 /// Loads an NCA file
 class AppLoader_NCA final : public AppLoader {
 public:
     explicit AppLoader_NCA(FileSys::VirtualFile file);
+    ~AppLoader_NCA() override;
 
     /**
      * Returns the type of the file
@@ -35,12 +39,7 @@ public:
     ResultStatus ReadRomFS(FileSys::VirtualFile& dir) override;
     ResultStatus ReadProgramId(u64& out_program_id) override;
 
-    ~AppLoader_NCA();
-
 private:
-    FileSys::ProgramMetadata metadata;
-
-    FileSys::NCAHeader header;
     std::unique_ptr<FileSys::NCA> nca;
     std::unique_ptr<AppLoader_DeconstructedRomDirectory> directory_loader;
 };

src/core/loader/xci.cpp

@@ -4,22 +4,14 @@
 
 #include <vector>
 
-#include "common/file_util.h"
-#include "common/logging/log.h"
-#include "common/string_util.h"
-#include "common/swap.h"
-#include "core/core.h"
+#include "common/common_types.h"
+#include "core/file_sys/card_image.h"
 #include "core/file_sys/content_archive.h"
 #include "core/file_sys/control_metadata.h"
-#include "core/file_sys/program_metadata.h"
 #include "core/file_sys/romfs.h"
-#include "core/gdbstub/gdbstub.h"
 #include "core/hle/kernel/process.h"
-#include "core/hle/kernel/resource_limit.h"
-#include "core/hle/service/filesystem/filesystem.h"
-#include "core/loader/nso.h"
+#include "core/loader/nca.h"
 #include "core/loader/xci.h"
-#include "core/memory.h"
 
 namespace Loader {
 

src/core/loader/xci.h

@@ -6,12 +6,18 @@
 
 #include <memory>
 #include "common/common_types.h"
-#include "core/file_sys/card_image.h"
+#include "core/file_sys/vfs.h"
 #include "core/loader/loader.h"
-#include "core/loader/nca.h"
+
+namespace FileSys {
+class NACP;
+class XCI;
+} // namespace FileSys
 
 namespace Loader {
 
+class AppLoader_NCA;
+
 /// Loads an XCI file
 class AppLoader_XCI final : public AppLoader {
 public:
@@ -37,8 +43,6 @@ public:
     ResultStatus ReadTitle(std::string& title) override;
 
 private:
-    FileSys::ProgramMetadata metadata;
-
     std::unique_ptr<FileSys::XCI> xci;
     std::unique_ptr<AppLoader_NCA> nca_loader;
 

src/video_core/engines/maxwell_3d.cpp

@@ -222,6 +222,18 @@ void Maxwell3D::DrawArrays() {
         debug_context->OnEvent(Tegra::DebugContext::Event::FinishedPrimitiveBatch, nullptr);
     }
 
+    // Both instance configuration registers can not be set at the same time.
+    ASSERT_MSG(!regs.draw.instance_next || !regs.draw.instance_cont,
+               "Illegal combination of instancing parameters");
+
+    if (regs.draw.instance_next) {
+        // Increment the current instance *before* drawing.
+        state.current_instance += 1;
+    } else if (!regs.draw.instance_cont) {
+        // Reset the current instance to 0.
+        state.current_instance = 0;
+    }
+
     const bool is_indexed{regs.index_array.count && !regs.vertex_buffer.count};
     rasterizer.AccelerateDrawBatch(is_indexed);
 

src/video_core/engines/maxwell_3d.h

@@ -638,6 +638,8 @@ public:
                     union {
                         u32 vertex_begin_gl;
                         BitField<0, 16, PrimitiveTopology> topology;
+                        BitField<26, 1, u32> instance_next;
+                        BitField<27, 1, u32> instance_cont;
                     };
                 } draw;
 
@@ -830,6 +832,7 @@ public:
         };
 
         std::array<ShaderStageInfo, Regs::MaxShaderStage> shader_stages;
+        u32 current_instance = 0; ///< Current instance to be used to simulate instanced rendering.
     };
 
     State state{};

src/video_core/engines/shader_bytecode.h

@@ -141,6 +141,7 @@ enum class PredCondition : u64 {
     NotEqual = 5,
     GreaterEqual = 6,
     LessThanWithNan = 9,
+    GreaterThanWithNan = 12,
     NotEqualWithNan = 13,
     // TODO(Subv): Other condition types
 };

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -124,7 +124,7 @@ std::pair<u8*, GLintptr> RasterizerOpenGL::SetupVertexArrays(u8* array_ptr,
         glBindVertexBuffer(index, stream_buffer.GetHandle(), vertex_buffer_offset,
                            vertex_array.stride);
 
-        ASSERT_MSG(vertex_array.divisor == 0, "Vertex buffer divisor unimplemented");
+        ASSERT_MSG(vertex_array.divisor == 0, "Instanced vertex arrays are not supported");
     }
 
     // Use the vertex array as-is, assumes that the data is formatted correctly for OpenGL.

src/video_core/renderer_opengl/gl_rasterizer_cache.cpp

@@ -142,14 +142,16 @@ static constexpr std::array<FormatTuple, SurfaceParams::MaxPixelFormat> tex_form
     {GL_RG32UI, GL_RG_INTEGER, GL_UNSIGNED_INT, ComponentType::UInt, false},              // RG32UI
     {GL_R32UI, GL_RED_INTEGER, GL_UNSIGNED_INT, ComponentType::UInt, false},              // R32UI
 
+    // Depth formats
+    {GL_DEPTH_COMPONENT32F, GL_DEPTH_COMPONENT, GL_FLOAT, ComponentType::Float, false}, // Z32F
+    {GL_DEPTH_COMPONENT16, GL_DEPTH_COMPONENT, GL_UNSIGNED_SHORT, ComponentType::UNorm,
+     false}, // Z16
+
     // DepthStencil formats
     {GL_DEPTH24_STENCIL8, GL_DEPTH_STENCIL, GL_UNSIGNED_INT_24_8, ComponentType::UNorm,
      false}, // Z24S8
     {GL_DEPTH24_STENCIL8, GL_DEPTH_STENCIL, GL_UNSIGNED_INT_24_8, ComponentType::UNorm,
-     false},                                                                            // S8Z24
-    {GL_DEPTH_COMPONENT32F, GL_DEPTH_COMPONENT, GL_FLOAT, ComponentType::Float, false}, // Z32F
-    {GL_DEPTH_COMPONENT16, GL_DEPTH_COMPONENT, GL_UNSIGNED_SHORT, ComponentType::UNorm,
-     false}, // Z16
+     false}, // S8Z24
     {GL_DEPTH32F_STENCIL8, GL_DEPTH_STENCIL, GL_FLOAT_32_UNSIGNED_INT_24_8_REV,
      ComponentType::Float, false}, // Z32FS8
 }};
@@ -283,10 +285,10 @@ static constexpr std::array<void (*)(u32, u32, u32, std::vector<u8>&, Tegra::GPU
         MortonCopy<true, PixelFormat::RG8S>,
         MortonCopy<true, PixelFormat::RG32UI>,
         MortonCopy<true, PixelFormat::R32UI>,
-        MortonCopy<true, PixelFormat::Z24S8>,
-        MortonCopy<true, PixelFormat::S8Z24>,
         MortonCopy<true, PixelFormat::Z32F>,
         MortonCopy<true, PixelFormat::Z16>,
+        MortonCopy<true, PixelFormat::Z24S8>,
+        MortonCopy<true, PixelFormat::S8Z24>,
         MortonCopy<true, PixelFormat::Z32FS8>,
         // clang-format on
 };
@@ -339,10 +341,10 @@ static constexpr std::array<void (*)(u32, u32, u32, std::vector<u8>&, Tegra::GPU
         MortonCopy<false, PixelFormat::RG8S>,
         MortonCopy<false, PixelFormat::RG32UI>,
         MortonCopy<false, PixelFormat::R32UI>,
-        MortonCopy<false, PixelFormat::Z24S8>,
-        MortonCopy<false, PixelFormat::S8Z24>,
         MortonCopy<false, PixelFormat::Z32F>,
         MortonCopy<false, PixelFormat::Z16>,
+        MortonCopy<false, PixelFormat::Z24S8>,
+        MortonCopy<false, PixelFormat::S8Z24>,
         MortonCopy<false, PixelFormat::Z32FS8>,
         // clang-format on
 };
@@ -788,8 +790,6 @@ Surface RasterizerCacheOpenGL::RecreateSurface(const Surface& surface,
     // Verify surface is compatible for blitting
     const auto& params{surface->GetSurfaceParams()};
     ASSERT(params.type == new_params.type);
-    ASSERT(params.pixel_format == new_params.pixel_format);
-    ASSERT(params.component_type == new_params.component_type);
 
     // Create a new surface with the new parameters, and blit the previous surface to it
     Surface new_surface{std::make_shared<CachedSurface>(new_params)};

src/video_core/renderer_opengl/gl_rasterizer_cache.h

@@ -68,11 +68,15 @@ struct SurfaceParams {
 
         MaxColorFormat,
 
+        // Depth formats
+        Z32F = 42,
+        Z16 = 43,
+
+        MaxDepthFormat,
+
         // DepthStencil formats
-        Z24S8 = 42,
-        S8Z24 = 43,
-        Z32F = 44,
-        Z16 = 45,
+        Z24S8 = 44,
+        S8Z24 = 45,
         Z32FS8 = 46,
 
         MaxDepthStencilFormat,
@@ -153,10 +157,10 @@ struct SurfaceParams {
             1, // RG8S
             1, // RG32UI
             1, // R32UI
-            1, // Z24S8
-            1, // S8Z24
             1, // Z32F
             1, // Z16
+            1, // Z24S8
+            1, // S8Z24
             1, // Z32FS8
         }};
 
@@ -211,10 +215,10 @@ struct SurfaceParams {
             16,  // RG8S
             64,  // RG32UI
             32,  // R32UI
-            32,  // Z24S8
-            32,  // S8Z24
             32,  // Z32F
             16,  // Z16
+            32,  // Z24S8
+            32,  // S8Z24
             64,  // Z32FS8
         }};
 
@@ -587,6 +591,10 @@ struct SurfaceParams {
             return SurfaceType::ColorTexture;
         }
 
+        if (static_cast<size_t>(pixel_format) < static_cast<size_t>(PixelFormat::MaxDepthFormat)) {
+            return SurfaceType::Depth;
+        }
+
         if (static_cast<size_t>(pixel_format) <
             static_cast<size_t>(PixelFormat::MaxDepthStencilFormat)) {
             return SurfaceType::DepthStencil;

src/video_core/renderer_opengl/gl_shader_decompiler.cpp

@@ -541,7 +541,7 @@ private:
             // vertex shader, and what's the value of the fourth element when inside a Tess Eval
             // shader.
             ASSERT(stage == Maxwell3D::Regs::ShaderStage::Vertex);
-            return "vec4(0, 0, uintBitsToFloat(gl_InstanceID), uintBitsToFloat(gl_VertexID))";
+            return "vec4(0, 0, uintBitsToFloat(instance_id.x), uintBitsToFloat(gl_VertexID))";
         default:
             const u32 index{static_cast<u32>(attribute) -
                             static_cast<u32>(Attribute::Index::Attribute_0)};
@@ -703,10 +703,11 @@ private:
                                        const std::string& op_a, const std::string& op_b) const {
         using Tegra::Shader::PredCondition;
         static const std::unordered_map<PredCondition, const char*> PredicateComparisonStrings = {
-            {PredCondition::LessThan, "<"},        {PredCondition::Equal, "=="},
-            {PredCondition::LessEqual, "<="},      {PredCondition::GreaterThan, ">"},
-            {PredCondition::NotEqual, "!="},       {PredCondition::GreaterEqual, ">="},
-            {PredCondition::LessThanWithNan, "<"}, {PredCondition::NotEqualWithNan, "!="},
+            {PredCondition::LessThan, "<"},           {PredCondition::Equal, "=="},
+            {PredCondition::LessEqual, "<="},         {PredCondition::GreaterThan, ">"},
+            {PredCondition::NotEqual, "!="},          {PredCondition::GreaterEqual, ">="},
+            {PredCondition::LessThanWithNan, "<"},    {PredCondition::NotEqualWithNan, "!="},
+            {PredCondition::GreaterThanWithNan, ">"},
         };
 
         const auto& comparison{PredicateComparisonStrings.find(condition)};
@@ -715,7 +716,8 @@ private:
 
         std::string predicate{'(' + op_a + ") " + comparison->second + " (" + op_b + ')'};
         if (condition == PredCondition::LessThanWithNan ||
-            condition == PredCondition::NotEqualWithNan) {
+            condition == PredCondition::NotEqualWithNan ||
+            condition == PredCondition::GreaterThanWithNan) {
             predicate += " || isnan(" + op_a + ") || isnan(" + op_b + ')';
         }
 
@@ -1237,8 +1239,6 @@ private:
             break;
         }
         case OpCode::Type::Conversion: {
-            ASSERT_MSG(!instr.conversion.negate_a, "Unimplemented");
-
             switch (opcode->GetId()) {
             case OpCode::Id::I2I_R: {
                 ASSERT_MSG(!instr.conversion.selector, "Unimplemented");

src/video_core/renderer_opengl/gl_shader_gen.cpp

@@ -38,6 +38,7 @@ out vec4 position;
 
 layout (std140) uniform vs_config {
     vec4 viewport_flip;
+    uvec4 instance_id;
 };
 
 void main() {
@@ -90,6 +91,7 @@ out vec4 color;
 
 layout (std140) uniform fs_config {
     vec4 viewport_flip;
+    uvec4 instance_id;
 };
 
 void main() {

src/video_core/renderer_opengl/gl_shader_manager.cpp

@@ -37,11 +37,16 @@ void SetShaderUniformBlockBindings(GLuint shader) {
 } // namespace Impl
 
 void MaxwellUniformData::SetFromRegs(const Maxwell3D::State::ShaderStageInfo& shader_stage) {
-    const auto& regs = Core::System::GetInstance().GPU().Maxwell3D().regs;
+    const auto& gpu = Core::System::GetInstance().GPU().Maxwell3D();
+    const auto& regs = gpu.regs;
+    const auto& state = gpu.state;
 
     // TODO(bunnei): Support more than one viewport
     viewport_flip[0] = regs.viewport_transform[0].scale_x < 0.0 ? -1.0f : 1.0f;
     viewport_flip[1] = regs.viewport_transform[0].scale_y < 0.0 ? -1.0f : 1.0f;
+
+    // We only assign the instance to the first component of the vector, the rest is just padding.
+    instance_id[0] = state.current_instance;
 }
 
 } // namespace GLShader

src/video_core/renderer_opengl/gl_shader_manager.h

@@ -24,14 +24,15 @@ void SetShaderUniformBlockBindings(GLuint shader);
 } // namespace Impl
 
 /// Uniform structure for the Uniform Buffer Object, all vectors must be 16-byte aligned
-// NOTE: Always keep a vec4 at the end. The GL spec is not clear wether the alignment at
+// NOTE: Always keep a vec4 at the end. The GL spec is not clear whether the alignment at
 //       the end of a uniform block is included in UNIFORM_BLOCK_DATA_SIZE or not.
 //       Not following that rule will cause problems on some AMD drivers.
 struct MaxwellUniformData {
     void SetFromRegs(const Maxwell3D::State::ShaderStageInfo& shader_stage);
     alignas(16) GLvec4 viewport_flip;
+    alignas(16) GLuvec4 instance_id;
 };
-static_assert(sizeof(MaxwellUniformData) == 16, "MaxwellUniformData structure size is incorrect");
+static_assert(sizeof(MaxwellUniformData) == 32, "MaxwellUniformData structure size is incorrect");
 static_assert(sizeof(MaxwellUniformData) < 16384,
               "MaxwellUniformData structure must be less than 16kb as per the OpenGL spec");
 

src/video_core/renderer_opengl/renderer_opengl.cpp

@@ -425,7 +425,7 @@ static void APIENTRY DebugHandler(GLenum source, GLenum type, GLuint id, GLenum
 
     switch (severity) {
     case GL_DEBUG_SEVERITY_HIGH:
-        LOG_ERROR(Render_OpenGL, format, str_source, str_type, id, message);
+        LOG_CRITICAL(Render_OpenGL, format, str_source, str_type, id, message);
         break;
     case GL_DEBUG_SEVERITY_MEDIUM:
         LOG_WARNING(Render_OpenGL, format, str_source, str_type, id, message);