GPU: Implemented a gpu macro interpreter.

The Ryujinx macro interpreter and envydis were used as reference.

Macros are programs that are uploaded by the games during boot and can later be called by writing to their method id in a GPU command buffer.

src/video_core/CMakeLists.txt

@@ -11,6 +11,8 @@ add_library(video_core STATIC
     engines/maxwell_compute.h
     gpu.cpp
     gpu.h
+    macro_interpreter.cpp
+    macro_interpreter.h
     memory_manager.cpp
     memory_manager.h
     rasterizer_interface.h

src/video_core/engines/maxwell_3d.cpp

@@ -386,5 +386,10 @@ std::vector<Texture::FullTextureInfo> Maxwell3D::GetStageTextures(Regs::ShaderSt
     return textures;
 }
 
+u32 Maxwell3D::GetRegisterValue(u32 method) const {
+    ASSERT_MSG(method < Regs::NUM_REGS, "Invalid Maxwell3D register");
+    return regs.reg_array[method];
+}
+
 } // namespace Engines
 } // namespace Tegra

src/video_core/engines/maxwell_3d.h

@@ -514,6 +514,9 @@ public:
 
     State state{};
 
+    /// Reads a register value located at the input method address
+    u32 GetRegisterValue(u32 method) const;
+
     /// Write the value to the register identified by method.
     void WriteReg(u32 method, u32 value, u32 remaining_params);
 

src/video_core/macro_interpreter.cpp

@@ -0,0 +1,257 @@
+// Copyright 2018 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include "common/assert.h"
+#include "common/logging/log.h"
+#include "video_core/engines/maxwell_3d.h"
+#include "video_core/macro_interpreter.h"
+
+namespace Tegra {
+
+MacroInterpreter::MacroInterpreter(Engines::Maxwell3D& maxwell3d) : maxwell3d(maxwell3d) {}
+
+void MacroInterpreter::Execute(const std::vector<u32>& code, std::vector<u32> parameters) {
+    Reset();
+    registers[1] = parameters[0];
+    this->parameters = std::move(parameters);
+
+    // Execute the code until we hit an exit condition.
+    bool keep_executing = true;
+    while (keep_executing) {
+        keep_executing = Step(code, false);
+    }
+
+    // Assert the the macro used all the input parameters
+    ASSERT(next_parameter_index == this->parameters.size());
+}
+
+void MacroInterpreter::Reset() {
+    registers = {};
+    pc = 0;
+    delayed_pc = boost::none;
+    method_address.raw = 0;
+    parameters.clear();
+    // The next parameter index starts at 1, because $r1 already has the value of the first
+    // parameter.
+    next_parameter_index = 1;
+}
+
+bool MacroInterpreter::Step(const std::vector<u32>& code, bool is_delay_slot) {
+    u32 base_address = pc;
+
+    Opcode opcode = GetOpcode(code);
+    pc += 4;
+
+    // Update the program counter if we were delayed
+    if (delayed_pc != boost::none) {
+        ASSERT(is_delay_slot);
+        pc = *delayed_pc;
+        delayed_pc = boost::none;
+    }
+
+    switch (opcode.operation) {
+    case Operation::ALU: {
+        u32 result = GetALUResult(opcode.alu_operation, GetRegister(opcode.src_a),
+                                  GetRegister(opcode.src_b));
+        ProcessResult(opcode.result_operation, opcode.dst, result);
+        break;
+    }
+    case Operation::AddImmediate: {
+        ProcessResult(opcode.result_operation, opcode.dst,
+                      GetRegister(opcode.src_a) + opcode.immediate);
+        break;
+    }
+    case Operation::ExtractInsert: {
+        u32 dst = GetRegister(opcode.src_a);
+        u32 src = GetRegister(opcode.src_b);
+
+        src = (src >> opcode.bf_src_bit) & opcode.GetBitfieldMask();
+        dst &= ~(opcode.GetBitfieldMask() << opcode.bf_dst_bit);
+        dst |= src << opcode.bf_dst_bit;
+        ProcessResult(opcode.result_operation, opcode.dst, dst);
+        break;
+    }
+    case Operation::ExtractShiftLeftImmediate: {
+        u32 dst = GetRegister(opcode.src_a);
+        u32 src = GetRegister(opcode.src_b);
+
+        u32 result = ((src >> dst) & opcode.GetBitfieldMask()) << opcode.bf_dst_bit;
+
+        ProcessResult(opcode.result_operation, opcode.dst, result);
+        break;
+    }
+    case Operation::ExtractShiftLeftRegister: {
+        u32 dst = GetRegister(opcode.src_a);
+        u32 src = GetRegister(opcode.src_b);
+
+        u32 result = ((src >> opcode.bf_src_bit) & opcode.GetBitfieldMask()) << dst;
+
+        ProcessResult(opcode.result_operation, opcode.dst, result);
+        break;
+    }
+    case Operation::Read: {
+        u32 result = Read(GetRegister(opcode.src_a) + opcode.immediate);
+        ProcessResult(opcode.result_operation, opcode.dst, result);
+        break;
+    }
+    case Operation::Branch: {
+        ASSERT_MSG(!is_delay_slot, "Executing a branch in a delay slot is not valid");
+        u32 value = GetRegister(opcode.src_a);
+        bool taken = EvaluateBranchCondition(opcode.branch_condition, value);
+        if (taken) {
+            // Ignore the delay slot if the branch has the annul bit.
+            if (opcode.branch_annul) {
+                pc = base_address + (opcode.immediate << 2);
+                return true;
+            }
+
+            delayed_pc = base_address + (opcode.immediate << 2);
+            // Execute one more instruction due to the delay slot.
+            return Step(code, true);
+        }
+        break;
+    }
+    default:
+        UNIMPLEMENTED_MSG("Unimplemented macro operation %u",
+                          static_cast<u32>(opcode.operation.Value()));
+    }
+
+    if (opcode.is_exit) {
+        // Exit has a delay slot, execute the next instruction
+        // Note: Executing an exit during a branch delay slot will cause the instruction at the
+        // branch target to be executed before exiting.
+        Step(code, true);
+        return false;
+    }
+
+    return true;
+}
+
+MacroInterpreter::Opcode MacroInterpreter::GetOpcode(const std::vector<u32>& code) const {
+    ASSERT((pc % sizeof(u32)) == 0);
+    ASSERT(pc < code.size() * sizeof(u32));
+    return {code[pc / sizeof(u32)]};
+}
+
+u32 MacroInterpreter::GetALUResult(ALUOperation operation, u32 src_a, u32 src_b) const {
+    switch (operation) {
+    case ALUOperation::Add:
+        return src_a + src_b;
+    // TODO(Subv): Implement AddWithCarry
+    case ALUOperation::Subtract:
+        return src_a - src_b;
+    // TODO(Subv): Implement SubtractWithBorrow
+    case ALUOperation::Xor:
+        return src_a ^ src_b;
+    case ALUOperation::Or:
+        return src_a | src_b;
+    case ALUOperation::And:
+        return src_a & src_b;
+    case ALUOperation::AndNot:
+        return src_a & ~src_b;
+    case ALUOperation::Nand:
+        return ~(src_a & src_b);
+
+    default:
+        UNIMPLEMENTED_MSG("Unimplemented ALU operation %u", static_cast<u32>(operation));
+    }
+}
+
+void MacroInterpreter::ProcessResult(ResultOperation operation, u32 reg, u32 result) {
+    switch (operation) {
+    case ResultOperation::IgnoreAndFetch:
+        // Fetch parameter and ignore result.
+        SetRegister(reg, FetchParameter());
+        break;
+    case ResultOperation::Move:
+        // Move result.
+        SetRegister(reg, result);
+        break;
+    case ResultOperation::MoveAndSetMethod:
+        // Move result and use as Method Address.
+        SetRegister(reg, result);
+        SetMethodAddress(result);
+        break;
+    case ResultOperation::FetchAndSend:
+        // Fetch parameter and send result.
+        SetRegister(reg, FetchParameter());
+        Send(result);
+        break;
+    case ResultOperation::MoveAndSend:
+        // Move and send result.
+        SetRegister(reg, result);
+        Send(result);
+        break;
+    case ResultOperation::FetchAndSetMethod:
+        // Fetch parameter and use result as Method Address.
+        SetRegister(reg, FetchParameter());
+        SetMethodAddress(result);
+        break;
+    case ResultOperation::MoveAndSetMethodFetchAndSend:
+        // Move result and use as Method Address, then fetch and send parameter.
+        SetRegister(reg, result);
+        SetMethodAddress(result);
+        Send(FetchParameter());
+        break;
+    case ResultOperation::MoveAndSetMethodSend:
+        // Move result and use as Method Address, then send bits 12:17 of result.
+        SetRegister(reg, result);
+        SetMethodAddress(result);
+        Send((result >> 12) & 0b111111);
+        break;
+    default:
+        UNIMPLEMENTED_MSG("Unimplemented result operation %u", static_cast<u32>(operation));
+    }
+}
+
+u32 MacroInterpreter::FetchParameter() {
+    ASSERT(next_parameter_index < parameters.size());
+    return parameters[next_parameter_index++];
+}
+
+u32 MacroInterpreter::GetRegister(u32 register_id) const {
+    // Register 0 is supposed to always return 0.
+    if (register_id == 0)
+        return 0;
+
+    ASSERT(register_id < registers.size());
+    return registers[register_id];
+}
+
+void MacroInterpreter::SetRegister(u32 register_id, u32 value) {
+    // Register 0 is supposed to always return 0. NOP is implemented as a store to the zero
+    // register.
+    if (register_id == 0)
+        return;
+
+    ASSERT(register_id < registers.size());
+    registers[register_id] = value;
+}
+
+void MacroInterpreter::SetMethodAddress(u32 address) {
+    method_address.raw = address;
+}
+
+void MacroInterpreter::Send(u32 value) {
+    maxwell3d.WriteReg(method_address.address, value, 0);
+    // Increment the method address by the method increment.
+    method_address.address.Assign(method_address.address.Value() +
+                                  method_address.increment.Value());
+}
+
+u32 MacroInterpreter::Read(u32 method) const {
+    return maxwell3d.GetRegisterValue(method);
+}
+
+bool MacroInterpreter::EvaluateBranchCondition(BranchCondition cond, u32 value) const {
+    switch (cond) {
+    case BranchCondition::Zero:
+        return value == 0;
+    case BranchCondition::NotZero:
+        return value != 0;
+    }
+    UNREACHABLE();
+}
+
+} // namespace Tegra

src/video_core/macro_interpreter.h

@@ -0,0 +1,164 @@
+// Copyright 2018 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include <array>
+#include <vector>
+#include <boost/optional.hpp>
+#include "common/bit_field.h"
+#include "common/common_types.h"
+
+namespace Tegra {
+namespace Engines {
+class Maxwell3D;
+}
+
+class MacroInterpreter final {
+public:
+    explicit MacroInterpreter(Engines::Maxwell3D& maxwell3d);
+
+    /**
+     * Executes the macro code with the specified input parameters.
+     * @param code The macro byte code to execute
+     * @param parameters The parameters of the macro
+     */
+    void Execute(const std::vector<u32>& code, std::vector<u32> parameters);
+
+private:
+    enum class Operation : u32 {
+        ALU = 0,
+        AddImmediate = 1,
+        ExtractInsert = 2,
+        ExtractShiftLeftImmediate = 3,
+        ExtractShiftLeftRegister = 4,
+        Read = 5,
+        Unused = 6, // This operation doesn't seem to be a valid encoding.
+        Branch = 7,
+    };
+
+    enum class ALUOperation : u32 {
+        Add = 0,
+        AddWithCarry = 1,
+        Subtract = 2,
+        SubtractWithBorrow = 3,
+        // Operations 4-7 don't seem to be valid encodings.
+        Xor = 8,
+        Or = 9,
+        And = 10,
+        AndNot = 11,
+        Nand = 12
+    };
+
+    enum class ResultOperation : u32 {
+        IgnoreAndFetch = 0,
+        Move = 1,
+        MoveAndSetMethod = 2,
+        FetchAndSend = 3,
+        MoveAndSend = 4,
+        FetchAndSetMethod = 5,
+        MoveAndSetMethodFetchAndSend = 6,
+        MoveAndSetMethodSend = 7
+    };
+
+    enum class BranchCondition : u32 {
+        Zero = 0,
+        NotZero = 1,
+    };
+
+    union Opcode {
+        u32 raw;
+        BitField<0, 3, Operation> operation;
+        BitField<4, 3, ResultOperation> result_operation;
+        BitField<4, 1, BranchCondition> branch_condition;
+        BitField<5, 1, u32>
+            branch_annul; // If set on a branch, then the branch doesn't have a delay slot.
+        BitField<7, 1, u32> is_exit;
+        BitField<8, 3, u32> dst;
+        BitField<11, 3, u32> src_a;
+        BitField<14, 3, u32> src_b;
+        // The signed immediate overlaps the second source operand and the alu operation.
+        BitField<14, 18, s32> immediate;
+
+        BitField<17, 5, ALUOperation> alu_operation;
+
+        // Bitfield instructions data
+        BitField<17, 5, u32> bf_src_bit;
+        BitField<22, 5, u32> bf_size;
+        BitField<27, 5, u32> bf_dst_bit;
+
+        u32 GetBitfieldMask() const {
+            return (1 << bf_size) - 1;
+        }
+    };
+
+    union MethodAddress {
+        u32 raw;
+        BitField<0, 12, u32> address;
+        BitField<12, 6, u32> increment;
+    };
+
+    /// Resets the execution engine state, zeroing registers, etc.
+    void Reset();
+
+    /**
+     * Executes a single macro instruction located at the current program counter. Returns whether
+     * the interpreter should keep running.
+     * @param code The macro code to execute.
+     * @param is_delay_slot Whether the current step is being executed due to a delay slot in a
+     * previous instruction.
+     */
+    bool Step(const std::vector<u32>& code, bool is_delay_slot);
+
+    /// Calculates the result of an ALU operation. src_a OP src_b;
+    u32 GetALUResult(ALUOperation operation, u32 src_a, u32 src_b) const;
+
+    /// Performs the result operation on the input result and stores it in the specified register
+    /// (if necessary).
+    void ProcessResult(ResultOperation operation, u32 reg, u32 result);
+
+    /// Evaluates the branch condition and returns whether the branch should be taken or not.
+    bool EvaluateBranchCondition(BranchCondition cond, u32 value) const;
+
+    /// Reads an opcode at the current program counter location.
+    Opcode GetOpcode(const std::vector<u32>& code) const;
+
+    /// Returns the specified register's value. Register 0 is hardcoded to always return 0.
+    u32 GetRegister(u32 register_id) const;
+
+    /// Sets the register to the input value.
+    void SetRegister(u32 register_id, u32 value);
+
+    /// Sets the method address to use for the next Send instruction.
+    void SetMethodAddress(u32 address);
+
+    /// Calls a GPU Engine method with the input parameter.
+    void Send(u32 value);
+
+    /// Reads a GPU register located at the method address.
+    u32 Read(u32 method) const;
+
+    /// Returns the next parameter in the parameter queue.
+    u32 FetchParameter();
+
+    Engines::Maxwell3D& maxwell3d;
+
+    u32 pc; ///< Current program counter
+    boost::optional<u32>
+        delayed_pc; ///< Program counter to execute at after the delay slot is executed.
+
+    static constexpr size_t NumMacroRegisters = 8;
+
+    /// General purpose macro registers.
+    std::array<u32, NumMacroRegisters> registers = {};
+
+    /// Method address to use for the next Send instruction.
+    MethodAddress method_address = {};
+
+    /// Input parameters of the current macro.
+    std::vector<u32> parameters;
+    /// Index of the next parameter that will be fetched by the 'parm' instruction.
+    u32 next_parameter_index = 0;
+};
+} // namespace Tegra