GPU: Use the bf bit in FSET to determine whether to write 0xFFFFFFFF or 1.0f.

am: Implement CreateStorage, PushInData, etc.

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

@@ -3,6 +3,7 @@
 // Refer to the license.txt file included.
 
 #include <cinttypes>
+#include <stack>
 #include "core/file_sys/filesystem.h"
 #include "core/hle/ipc_helpers.h"
 #include "core/hle/kernel/event.h"
@@ -348,6 +349,87 @@ void ICommonStateGetter::GetPerformanceMode(Kernel::HLERequestContext& ctx) {
     NGLOG_WARNING(Service_AM, "(STUBBED) called");
 }
 
+class IStorageAccessor final : public ServiceFramework<IStorageAccessor> {
+public:
+    explicit IStorageAccessor(std::vector<u8> buffer)
+        : ServiceFramework("IStorageAccessor"), buffer(std::move(buffer)) {
+        static const FunctionInfo functions[] = {
+            {0, &IStorageAccessor::GetSize, "GetSize"},
+            {10, &IStorageAccessor::Write, "Write"},
+            {11, &IStorageAccessor::Read, "Read"},
+        };
+        RegisterHandlers(functions);
+    }
+
+private:
+    std::vector<u8> buffer;
+
+    void GetSize(Kernel::HLERequestContext& ctx) {
+        IPC::ResponseBuilder rb{ctx, 4};
+
+        rb.Push(RESULT_SUCCESS);
+        rb.Push(static_cast<u64>(buffer.size()));
+
+        NGLOG_DEBUG(Service_AM, "called");
+    }
+
+    void Write(Kernel::HLERequestContext& ctx) {
+        IPC::RequestParser rp{ctx};
+
+        const u64 offset{rp.Pop<u64>()};
+        const std::vector<u8> data{ctx.ReadBuffer()};
+
+        ASSERT(offset + data.size() <= buffer.size());
+
+        std::memcpy(&buffer[offset], data.data(), data.size());
+
+        IPC::ResponseBuilder rb{rp.MakeBuilder(2, 0, 0)};
+        rb.Push(RESULT_SUCCESS);
+
+        NGLOG_DEBUG(Service_AM, "called, offset={}", offset);
+    }
+
+    void Read(Kernel::HLERequestContext& ctx) {
+        IPC::RequestParser rp{ctx};
+
+        const u64 offset{rp.Pop<u64>()};
+        const size_t size{ctx.GetWriteBufferSize()};
+
+        ASSERT(offset + size <= buffer.size());
+
+        ctx.WriteBuffer(buffer.data() + offset, size);
+
+        IPC::ResponseBuilder rb{rp.MakeBuilder(2, 0, 0)};
+        rb.Push(RESULT_SUCCESS);
+
+        NGLOG_DEBUG(Service_AM, "called, offset={}", offset);
+    }
+};
+
+class IStorage final : public ServiceFramework<IStorage> {
+public:
+    explicit IStorage(std::vector<u8> buffer)
+        : ServiceFramework("IStorage"), buffer(std::move(buffer)) {
+        static const FunctionInfo functions[] = {
+            {0, &IStorage::Open, "Open"},
+            {1, nullptr, "OpenTransferStorage"},
+        };
+        RegisterHandlers(functions);
+    }
+
+private:
+    std::vector<u8> buffer;
+
+    void Open(Kernel::HLERequestContext& ctx) {
+        IPC::ResponseBuilder rb{ctx, 2, 0, 1};
+
+        rb.Push(RESULT_SUCCESS);
+        rb.PushIpcInterface<AM::IStorageAccessor>(buffer);
+
+        NGLOG_DEBUG(Service_AM, "called");
+    }
+};
+
 class ILibraryAppletAccessor final : public ServiceFramework<ILibraryAppletAccessor> {
 public:
     explicit ILibraryAppletAccessor() : ServiceFramework("ILibraryAppletAccessor") {
@@ -359,7 +441,7 @@ public:
             {25, nullptr, "Terminate"},
             {30, nullptr, "GetResult"},
             {50, nullptr, "SetOutOfFocusApplicationSuspendingEnabled"},
-            {100, nullptr, "PushInData"},
+            {100, &ILibraryAppletAccessor::PushInData, "PushInData"},
             {101, nullptr, "PopOutData"},
             {102, nullptr, "PushExtraStorage"},
             {103, nullptr, "PushInteractiveInData"},
@@ -388,6 +470,17 @@ private:
         NGLOG_WARNING(Service_AM, "(STUBBED) called");
     }
 
+    void PushInData(Kernel::HLERequestContext& ctx) {
+        IPC::RequestParser rp{ctx};
+        storage_stack.push(rp.PopIpcInterface<AM::IStorage>());
+
+        IPC::ResponseBuilder rb{rp.MakeBuilder(2, 0, 0)};
+        rb.Push(RESULT_SUCCESS);
+
+        NGLOG_DEBUG(Service_AM, "called");
+    }
+
+    std::stack<std::shared_ptr<AM::IStorage>> storage_stack;
     Kernel::SharedPtr<Kernel::Event> state_changed_event;
 };
 
@@ -396,7 +489,7 @@ ILibraryAppletCreator::ILibraryAppletCreator() : ServiceFramework("ILibraryApple
         {0, &ILibraryAppletCreator::CreateLibraryApplet, "CreateLibraryApplet"},
         {1, nullptr, "TerminateAllLibraryApplets"},
         {2, nullptr, "AreAnyLibraryAppletsLeft"},
-        {10, nullptr, "CreateStorage"},
+        {10, &ILibraryAppletCreator::CreateStorage, "CreateStorage"},
         {11, nullptr, "CreateTransferMemoryStorage"},
         {12, nullptr, "CreateHandleStorage"},
     };
@@ -412,72 +505,17 @@ void ILibraryAppletCreator::CreateLibraryApplet(Kernel::HLERequestContext& ctx)
     NGLOG_DEBUG(Service_AM, "called");
 }
 
-class IStorageAccessor final : public ServiceFramework<IStorageAccessor> {
-public:
-    explicit IStorageAccessor(std::vector<u8> buffer)
-        : ServiceFramework("IStorageAccessor"), buffer(std::move(buffer)) {
-        static const FunctionInfo functions[] = {
-            {0, &IStorageAccessor::GetSize, "GetSize"},
-            {10, nullptr, "Write"},
-            {11, &IStorageAccessor::Read, "Read"},
-        };
-        RegisterHandlers(functions);
-    }
+void ILibraryAppletCreator::CreateStorage(Kernel::HLERequestContext& ctx) {
+    IPC::RequestParser rp{ctx};
+    const u64 size{rp.Pop<u64>()};
+    std::vector<u8> buffer(size);
 
-private:
-    std::vector<u8> buffer;
+    IPC::ResponseBuilder rb{rp.MakeBuilder(2, 0, 1)};
+    rb.Push(RESULT_SUCCESS);
+    rb.PushIpcInterface<AM::IStorage>(std::move(buffer));
 
-    void GetSize(Kernel::HLERequestContext& ctx) {
-        IPC::ResponseBuilder rb{ctx, 4};
-
-        rb.Push(RESULT_SUCCESS);
-        rb.Push(static_cast<u64>(buffer.size()));
-
-        NGLOG_DEBUG(Service_AM, "called");
-    }
-
-    void Read(Kernel::HLERequestContext& ctx) {
-        IPC::RequestParser rp{ctx};
-
-        u64 offset = rp.Pop<u64>();
-
-        const size_t size{ctx.GetWriteBufferSize()};
-
-        ASSERT(offset + size <= buffer.size());
-
-        ctx.WriteBuffer(buffer.data() + offset, size);
-
-        IPC::ResponseBuilder rb{ctx, 2};
-
-        rb.Push(RESULT_SUCCESS);
-
-        NGLOG_DEBUG(Service_AM, "called");
-    }
-};
-
-class IStorage final : public ServiceFramework<IStorage> {
-public:
-    explicit IStorage(std::vector<u8> buffer)
-        : ServiceFramework("IStorage"), buffer(std::move(buffer)) {
-        static const FunctionInfo functions[] = {
-            {0, &IStorage::Open, "Open"},
-            {1, nullptr, "OpenTransferStorage"},
-        };
-        RegisterHandlers(functions);
-    }
-
-private:
-    std::vector<u8> buffer;
-
-    void Open(Kernel::HLERequestContext& ctx) {
-        IPC::ResponseBuilder rb{ctx, 2, 0, 1};
-
-        rb.Push(RESULT_SUCCESS);
-        rb.PushIpcInterface<AM::IStorageAccessor>(buffer);
-
-        NGLOG_DEBUG(Service_AM, "called");
-    }
-};
+    NGLOG_DEBUG(Service_AM, "called, size={}", size);
+}
 
 IApplicationFunctions::IApplicationFunctions() : ServiceFramework("IApplicationFunctions") {
     static const FunctionInfo functions[] = {

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

@@ -121,6 +121,7 @@ public:
 
 private:
     void CreateLibraryApplet(Kernel::HLERequestContext& ctx);
+    void CreateStorage(Kernel::HLERequestContext& ctx);
 };
 
 class IApplicationFunctions final : public ServiceFramework<IApplicationFunctions> {

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

@@ -26,19 +26,11 @@ 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,
@@ -64,12 +56,6 @@ 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;
     };
@@ -90,45 +76,12 @@ 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.cpp

@@ -156,16 +156,15 @@ void Maxwell3D::ProcessQueryGet() {
     // TODO(Subv): Support the other query units.
     ASSERT_MSG(regs.query.query_get.unit == Regs::QueryUnit::Crop,
                "Units other than CROP are unimplemented");
-    ASSERT_MSG(regs.query.query_get.short_query,
-               "Writing the entire query result structure is unimplemented");
 
     u32 value = Memory::Read32(*address);
-    u32 result = 0;
+    u64 result = 0;
 
     // TODO(Subv): Support the other query variables
     switch (regs.query.query_get.select) {
     case Regs::QuerySelect::Zero:
-        result = 0;
+        // This seems to actually write the query sequence to the query address.
+        result = regs.query.query_sequence;
         break;
     default:
         UNIMPLEMENTED_MSG("Unimplemented query select type {}",
@@ -174,15 +173,31 @@ void Maxwell3D::ProcessQueryGet() {
 
     // TODO(Subv): Research and implement how query sync conditions work.
 
+    struct LongQueryResult {
+        u64_le value;
+        u64_le timestamp;
+    };
+    static_assert(sizeof(LongQueryResult) == 16, "LongQueryResult has wrong size");
+
     switch (regs.query.query_get.mode) {
     case Regs::QueryMode::Write:
     case Regs::QueryMode::Write2: {
-        // Write the current query sequence to the sequence address.
         u32 sequence = regs.query.query_sequence;
-        Memory::Write32(*address, sequence);
-
-        // TODO(Subv): Write the proper query response structure to the address when not using short
-        // mode.
+        if (regs.query.query_get.short_query) {
+            // Write the current query sequence to the sequence address.
+            // TODO(Subv): Find out what happens if you use a long query type but mark it as a short
+            // query.
+            Memory::Write32(*address, sequence);
+        } else {
+            // Write the 128-bit result structure in long mode. Note: We emulate an infinitely fast
+            // GPU, this command may actually take a while to complete in real hardware due to GPU
+            // wait queues.
+            LongQueryResult query_result{};
+            query_result.value = result;
+            // TODO(Subv): Generate a real GPU timestamp and write it here instead of 0
+            query_result.timestamp = 0;
+            Memory::WriteBlock(*address, &query_result, sizeof(query_result));
+        }
         break;
     }
     default:

src/video_core/engines/shader_bytecode.h

@@ -245,9 +245,9 @@ union Instruction {
         BitField<44, 1, u64> abs_b;
         BitField<45, 2, PredOperation> op;
         BitField<48, 4, PredCondition> cond;
+        BitField<52, 1, u64> bf;
         BitField<53, 1, u64> neg_b;
         BitField<54, 1, u64> abs_a;
-        BitField<52, 1, u64> bf;
         BitField<55, 1, u64> ftz;
         BitField<56, 1, u64> neg_imm;
     } fset;
@@ -270,8 +270,22 @@ union Instruction {
     } tex;
 
     union {
-        // TODO(bunnei): This is just a guess, needs to be verified
-        BitField<52, 1, u64> enable_g_component;
+        BitField<50, 3, u64> component_mask_selector;
+        BitField<28, 8, Register> gpr28;
+
+        bool HasTwoDestinations() const {
+            return gpr28.Value() != Register::ZeroIndex;
+        }
+
+        bool IsComponentEnabled(size_t component) const {
+            static constexpr std::array<size_t, 5> one_dest_mask{0x1, 0x2, 0x4, 0x8, 0x3};
+            static constexpr std::array<size_t, 5> two_dest_mask{0x7, 0xb, 0xd, 0xe, 0xf};
+            const auto& mask{HasTwoDestinations() ? two_dest_mask : one_dest_mask};
+
+            ASSERT(component_mask_selector < mask.size());
+
+            return ((1 << component) & mask[component_mask_selector]) != 0;
+        }
     } texs;
 
     BitField<61, 1, u64> is_b_imm;

src/video_core/renderer_opengl/gl_shader_decompiler.cpp

@@ -849,8 +849,7 @@ private:
             ASSERT_MSG(!instr.conversion.saturate_a, "Unimplemented");
 
             switch (opcode->GetId()) {
-            case OpCode::Id::I2I_R:
-            case OpCode::Id::I2F_R: {
+            case OpCode::Id::I2I_R: {
                 ASSERT_MSG(!instr.conversion.selector, "Unimplemented");
 
                 std::string op_a =
@@ -863,6 +862,17 @@ private:
                 regs.SetRegisterToInteger(instr.gpr0, instr.conversion.is_signed, 0, op_a, 1, 1);
                 break;
             }
+            case OpCode::Id::I2F_R: {
+                std::string op_a =
+                    regs.GetRegisterAsInteger(instr.gpr20, 0, instr.conversion.is_signed);
+
+                if (instr.conversion.abs_a) {
+                    op_a = "abs(" + op_a + ')';
+                }
+
+                regs.SetRegisterToFloat(instr.gpr0, 0, op_a, 1, 1);
+                break;
+            }
             case OpCode::Id::F2F_R: {
                 std::string op_a = regs.GetRegisterAsFloat(instr.gpr20);
 
@@ -938,18 +948,21 @@ private:
                 // TEXS has two destination registers. RG goes into gpr0+0 and gpr0+1, and BA goes
                 // into gpr28+0 and gpr28+1
                 size_t offset{};
+
                 for (const auto& dest : {instr.gpr0.Value(), instr.gpr28.Value()}) {
                     for (unsigned elem = 0; elem < 2; ++elem) {
-                        if (dest + elem >= Register::ZeroIndex) {
-                            // Skip invalid register values
-                            break;
+                        if (!instr.texs.IsComponentEnabled(elem)) {
+                            // Skip disabled components
+                            continue;
                         }
                         regs.SetRegisterToFloat(dest, elem + offset, texture, 1, 4, false, elem);
-                        if (!instr.texs.enable_g_component) {
-                            // Skip the second component
-                            break;
-                        }
                     }
+
+                    if (!instr.texs.HasTwoDestinations()) {
+                        // Skip the second destination
+                        break;
+                    }
+
                     offset += 2;
                 }
                 --shader.scope;
@@ -1053,7 +1066,12 @@ private:
             std::string predicate = "(((" + op_a + ") " + comparator + " (" + op_b + ")) " +
                                     combiner + " (" + second_pred + "))";
 
-            regs.SetRegisterToFloat(instr.gpr0, 0, predicate + " ? 1.0 : 0.0", 1, 1);
+            if (instr.fset.bf) {
+                regs.SetRegisterToFloat(instr.gpr0, 0, predicate + " ? 1.0 : 0.0", 1, 1);
+            } else {
+                regs.SetRegisterToInteger(instr.gpr0, false, 0, predicate + " ? 0xFFFFFFFF : 0", 1,
+                                          1);
+            }
             break;
         }
         default: {