GPU: Partially implemented the shader BRA instruction.

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

@@ -3,7 +3,6 @@
 // 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"
@@ -349,87 +348,6 @@ 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") {
@@ -441,7 +359,7 @@ public:
             {25, nullptr, "Terminate"},
             {30, nullptr, "GetResult"},
             {50, nullptr, "SetOutOfFocusApplicationSuspendingEnabled"},
-            {100, &ILibraryAppletAccessor::PushInData, "PushInData"},
+            {100, nullptr, "PushInData"},
             {101, nullptr, "PopOutData"},
             {102, nullptr, "PushExtraStorage"},
             {103, nullptr, "PushInteractiveInData"},
@@ -470,17 +388,6 @@ 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;
 };
 
@@ -489,7 +396,7 @@ ILibraryAppletCreator::ILibraryAppletCreator() : ServiceFramework("ILibraryApple
         {0, &ILibraryAppletCreator::CreateLibraryApplet, "CreateLibraryApplet"},
         {1, nullptr, "TerminateAllLibraryApplets"},
         {2, nullptr, "AreAnyLibraryAppletsLeft"},
-        {10, &ILibraryAppletCreator::CreateStorage, "CreateStorage"},
+        {10, nullptr, "CreateStorage"},
         {11, nullptr, "CreateTransferMemoryStorage"},
         {12, nullptr, "CreateHandleStorage"},
     };
@@ -505,17 +412,72 @@ void ILibraryAppletCreator::CreateLibraryApplet(Kernel::HLERequestContext& ctx)
     NGLOG_DEBUG(Service_AM, "called");
 }
 
-void ILibraryAppletCreator::CreateStorage(Kernel::HLERequestContext& ctx) {
-    IPC::RequestParser rp{ctx};
-    const u64 size{rp.Pop<u64>()};
-    std::vector<u8> buffer(size);
+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);
+    }
 
-    IPC::ResponseBuilder rb{rp.MakeBuilder(2, 0, 1)};
-    rb.Push(RESULT_SUCCESS);
-    rb.PushIpcInterface<AM::IStorage>(std::move(buffer));
+private:
+    std::vector<u8> buffer;
 
-    NGLOG_DEBUG(Service_AM, "called, size={}", size);
-}
+    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");
+    }
+};
 
 IApplicationFunctions::IApplicationFunctions() : ServiceFramework("IApplicationFunctions") {
     static const FunctionInfo functions[] = {

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

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

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;
@@ -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_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] =
@@ -97,11 +111,19 @@ 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;
                 }
+                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);
+                }
                 }
             }
         }
@@ -849,7 +871,8 @@ private:
             ASSERT_MSG(!instr.conversion.saturate_a, "Unimplemented");
 
             switch (opcode->GetId()) {
-            case OpCode::Id::I2I_R: {
+            case OpCode::Id::I2I_R:
+            case OpCode::Id::I2F_R: {
                 ASSERT_MSG(!instr.conversion.selector, "Unimplemented");
 
                 std::string op_a =
@@ -862,17 +885,6 @@ 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);
 
@@ -1066,12 +1078,7 @@ private:
             std::string predicate = "(((" + op_a + ") " + comparator + " (" + op_b + ")) " +
                                     combiner + " (" + second_pred + "))";
 
-            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);
-            }
+            regs.SetRegisterToFloat(instr.gpr0, 0, predicate + " ? 1.0 : 0.0", 1, 1);
             break;
         }
         default: {
@@ -1096,6 +1103,13 @@ private:
                 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);