gpu: Rewrite GPU command list processing with DmaPusher class.

- More accurate impl., fixes Undertale (among other games).

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

@@ -8,7 +8,6 @@
 #include "core/core.h"
 #include "core/hle/service/nvdrv/devices/nvhost_gpu.h"
 #include "core/memory.h"
-#include "video_core/command_processor.h"
 #include "video_core/gpu.h"
 #include "video_core/memory_manager.h"
 
@@ -129,6 +128,14 @@ u32 nvhost_gpu::AllocateObjectContext(const std::vector<u8>& input, std::vector<
     return 0;
 }
 
+static void PushGPUEntries(const std::vector<Tegra::CommandListHeader>& entries) {
+    auto& dma_pusher{Core::System::GetInstance().GPU().DmaPusher()};
+    for (const auto& entry : entries) {
+        dma_pusher.Push(entry);
+    }
+    dma_pusher.DispatchCalls();
+}
+
 u32 nvhost_gpu::SubmitGPFIFO(const std::vector<u8>& input, std::vector<u8>& output) {
     if (input.size() < sizeof(IoctlSubmitGpfifo)) {
         UNIMPLEMENTED();
@@ -146,7 +153,7 @@ u32 nvhost_gpu::SubmitGPFIFO(const std::vector<u8>& input, std::vector<u8>& outp
     std::memcpy(entries.data(), &input[sizeof(IoctlSubmitGpfifo)],
                 params.num_entries * sizeof(Tegra::CommandListHeader));
 
-    Core::System::GetInstance().GPU().ProcessCommandLists(entries);
+    PushGPUEntries(entries);
 
     params.fence_out.id = 0;
     params.fence_out.value = 0;
@@ -167,7 +174,7 @@ u32 nvhost_gpu::KickoffPB(const std::vector<u8>& input, std::vector<u8>& output)
     Memory::ReadBlock(params.address, entries.data(),
                       params.num_entries * sizeof(Tegra::CommandListHeader));
 
-    Core::System::GetInstance().GPU().ProcessCommandLists(entries);
+    PushGPUEntries(entries);
 
     params.fence_out.id = 0;
     params.fence_out.value = 0;

src/video_core/CMakeLists.txt

@@ -1,6 +1,6 @@
 add_library(video_core STATIC
-    command_processor.cpp
-    command_processor.h
+    dma_pusher.cpp
+    dma_pusher.h
     debug_utils/debug_utils.cpp
     debug_utils/debug_utils.h
     engines/fermi_2d.cpp

src/video_core/command_processor.h

@@ -1,53 +0,0 @@
-// Copyright 2018 yuzu Emulator Project
-// Licensed under GPLv2 or any later version
-// Refer to the license.txt file included.
-
-#pragma once
-
-#include <type_traits>
-#include "common/bit_field.h"
-#include "common/common_types.h"
-#include "video_core/memory_manager.h"
-
-namespace Tegra {
-
-enum class SubmissionMode : u32 {
-    IncreasingOld = 0,
-    Increasing = 1,
-    NonIncreasingOld = 2,
-    NonIncreasing = 3,
-    Inline = 4,
-    IncreaseOnce = 5
-};
-
-struct CommandListHeader {
-    u32 entry0; // gpu_va_lo
-    union {
-        u32 entry1; // gpu_va_hi | (unk_0x02 << 0x08) | (size << 0x0A) | (unk_0x01 << 0x1F)
-        BitField<0, 8, u32> gpu_va_hi;
-        BitField<8, 2, u32> unk1;
-        BitField<10, 21, u32> sz;
-        BitField<31, 1, u32> unk2;
-    };
-
-    GPUVAddr Address() const {
-        return (static_cast<GPUVAddr>(gpu_va_hi) << 32) | entry0;
-    }
-};
-static_assert(sizeof(CommandListHeader) == 8, "CommandListHeader is incorrect size");
-
-union CommandHeader {
-    u32 hex;
-
-    BitField<0, 13, u32> method;
-    BitField<13, 3, u32> subchannel;
-
-    BitField<16, 13, u32> arg_count;
-    BitField<16, 13, u32> inline_data;
-
-    BitField<29, 3, SubmissionMode> mode;
-};
-static_assert(std::is_standard_layout_v<CommandHeader>, "CommandHeader is not standard layout");
-static_assert(sizeof(CommandHeader) == sizeof(u32), "CommandHeader has incorrect size!");
-
-} // namespace Tegra

src/video_core/dma_pusher.cpp

@@ -0,0 +1,110 @@
+// Copyright 2018 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include "core/core.h"
+#include "core/memory.h"
+#include "video_core/dma_pusher.h"
+#include "video_core/engines/maxwell_3d.h"
+#include "video_core/gpu.h"
+
+namespace Tegra {
+
+DmaPusher::DmaPusher(GPU& gpu) : gpu(gpu) {}
+
+DmaPusher::~DmaPusher() = default;
+
+void DmaPusher::DispatchCalls() {
+    // On entering GPU code, assume all memory may be touched by the ARM core.
+    gpu.Maxwell3D().dirty_flags.OnMemoryWrite();
+
+    while (Core::System::GetInstance().IsPoweredOn()) {
+        if (!Step()) {
+            break;
+        }
+    }
+}
+
+bool DmaPusher::Step() {
+    if (dma_get != dma_put) {
+        // Push buffer non-empty, read a word
+        const CommandHeader command_header{
+            Memory::Read32(*gpu.MemoryManager().GpuToCpuAddress(dma_get))};
+
+        dma_get += sizeof(u32);
+
+        if (!non_main) {
+            dma_mget = dma_get;
+        }
+
+        // now, see if we're in the middle of a command
+        if (dma_state.length_pending) {
+            // Second word of long non-inc methods command - method count
+            dma_state.length_pending = 0;
+            dma_state.method_count = command_header.method_count_;
+        } else if (dma_state.method_count) {
+            // Data word of methods command
+            CallMethod(command_header.argument);
+
+            if (!dma_state.non_incrementing) {
+                dma_state.method++;
+            }
+
+            if (dma_increment_once) {
+                dma_state.non_incrementing = true;
+            }
+
+            dma_state.method_count--;
+        } else {
+            // No command active - this is the first word of a new one
+            switch (command_header.mode) {
+            case SubmissionMode::Increasing:
+                SetState(command_header);
+                dma_state.non_incrementing = false;
+                dma_increment_once = false;
+                break;
+            case SubmissionMode::NonIncreasing:
+                SetState(command_header);
+                dma_state.non_incrementing = true;
+                dma_increment_once = false;
+                break;
+            case SubmissionMode::Inline:
+                dma_state.method = command_header.method;
+                dma_state.subchannel = command_header.subchannel;
+                CallMethod(command_header.arg_count);
+                dma_state.non_incrementing = true;
+                dma_increment_once = false;
+                break;
+            case SubmissionMode::IncreaseOnce:
+                SetState(command_header);
+                dma_state.non_incrementing = false;
+                dma_increment_once = true;
+                break;
+            }
+        }
+    } else if (ib_enable && !dma_pushbuffer.empty()) {
+        // Current pushbuffer empty, but we have more IB entries to read
+        const CommandListHeader& command_list_header{dma_pushbuffer.front()};
+        dma_get = command_list_header.addr;
+        dma_put = dma_get + command_list_header.size * sizeof(u32);
+        non_main = command_list_header.is_non_main;
+        dma_pushbuffer.pop();
+    } else {
+        // Otherwise, pushbuffer empty and IB empty or nonexistent - nothing to do
+        return {};
+    }
+
+    return true;
+}
+
+void DmaPusher::SetState(const CommandHeader& command_header) {
+    dma_state.method = command_header.method;
+    dma_state.subchannel = command_header.subchannel;
+    dma_state.method_count = command_header.method_count;
+}
+
+void DmaPusher::CallMethod(u32 argument) const {
+    gpu.CallMethod({dma_state.method, argument, dma_state.subchannel, dma_state.method_count});
+}
+
+} // namespace Tegra

src/video_core/dma_pusher.h

@@ -0,0 +1,95 @@
+// Copyright 2018 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include <queue>
+
+#include "common/bit_field.h"
+#include "common/common_types.h"
+#include "video_core/memory_manager.h"
+
+namespace Tegra {
+
+enum class SubmissionMode : u32 {
+    IncreasingOld = 0,
+    Increasing = 1,
+    NonIncreasingOld = 2,
+    NonIncreasing = 3,
+    Inline = 4,
+    IncreaseOnce = 5
+};
+
+struct CommandListHeader {
+    union {
+        u64 raw;
+        BitField<0, 40, GPUVAddr> addr;
+        BitField<41, 1, u64> is_non_main;
+        BitField<42, 21, u64> size;
+    };
+};
+static_assert(sizeof(CommandListHeader) == sizeof(u64), "CommandListHeader is incorrect size");
+
+union CommandHeader {
+    u32 argument;
+    BitField<0, 13, u32> method;
+    BitField<0, 24, u32> method_count_;
+    BitField<13, 3, u32> subchannel;
+    BitField<16, 13, u32> arg_count;
+    BitField<16, 13, u32> method_count;
+    BitField<29, 3, SubmissionMode> mode;
+};
+static_assert(std::is_standard_layout_v<CommandHeader>, "CommandHeader is not standard layout");
+static_assert(sizeof(CommandHeader) == sizeof(u32), "CommandHeader has incorrect size!");
+
+class GPU;
+
+/**
+ * The DmaPusher class implements DMA submission to FIFOs, providing an area of memory that the
+ * emulated app fills with commands and tells PFIFO to process. The pushbuffers are then assembled
+ * into a "command stream" consisting of 32-bit words that make up "commands".
+ * See https://envytools.readthedocs.io/en/latest/hw/fifo/dma-pusher.html#fifo-dma-pusher for
+ * details on this implementation.
+ */
+class DmaPusher {
+public:
+    explicit DmaPusher(GPU& gpu);
+    ~DmaPusher();
+
+    void Push(const CommandListHeader& command_list_header) {
+        dma_pushbuffer.push(command_list_header);
+    }
+
+    void DispatchCalls();
+
+private:
+    bool Step();
+
+    void SetState(const CommandHeader& command_header);
+
+    void CallMethod(u32 argument) const;
+
+    GPU& gpu;
+
+    std::queue<CommandListHeader> dma_pushbuffer;
+
+    struct DmaState {
+        u32 method;            ///< Current method
+        u32 subchannel;        ///< Current subchannel
+        u32 method_count;      ///< Current method count
+        u32 length_pending;    ///< Large NI command length pending
+        bool non_incrementing; ///< Current command’s NI flag
+    };
+
+    DmaState dma_state{};
+    bool dma_increment_once{};
+
+    GPUVAddr dma_put{};   ///< pushbuffer current end address
+    GPUVAddr dma_get{};   ///< pushbuffer current read address
+    GPUVAddr dma_mget{};  ///< main pushbuffer last read address
+    bool ib_enable{true}; ///< IB mode enabled
+    bool non_main{};      ///< non-main pushbuffer active
+};
+
+} // namespace Tegra

src/video_core/engines/fermi_2d.cpp

@@ -14,13 +14,13 @@ namespace Tegra::Engines {
 Fermi2D::Fermi2D(VideoCore::RasterizerInterface& rasterizer, MemoryManager& memory_manager)
     : memory_manager(memory_manager), rasterizer{rasterizer} {}
 
-void Fermi2D::WriteReg(u32 method, u32 value) {
-    ASSERT_MSG(method < Regs::NUM_REGS,
+void Fermi2D::CallMethod(const GPU::MethodCall& method_call) {
+    ASSERT_MSG(method_call.method < Regs::NUM_REGS,
                "Invalid Fermi2D register, increase the size of the Regs structure");
 
-    regs.reg_array[method] = value;
+    regs.reg_array[method_call.method] = method_call.argument;
 
-    switch (method) {
+    switch (method_call.method) {
     case FERMI2D_REG_INDEX(trigger): {
         HandleSurfaceCopy();
         break;

src/video_core/engines/fermi_2d.h

@@ -27,7 +27,7 @@ public:
     ~Fermi2D() = default;
 
     /// Write the value to the register identified by method.
-    void WriteReg(u32 method, u32 value);
+    void CallMethod(const GPU::MethodCall& method_call);
 
     struct Regs {
         static constexpr std::size_t NUM_REGS = 0x258;

src/video_core/engines/kepler_memory.cpp

@@ -17,19 +17,19 @@ KeplerMemory::KeplerMemory(VideoCore::RasterizerInterface& rasterizer,
 
 KeplerMemory::~KeplerMemory() = default;
 
-void KeplerMemory::WriteReg(u32 method, u32 value) {
-    ASSERT_MSG(method < Regs::NUM_REGS,
+void KeplerMemory::CallMethod(const GPU::MethodCall& method_call) {
+    ASSERT_MSG(method_call.method < Regs::NUM_REGS,
                "Invalid KeplerMemory register, increase the size of the Regs structure");
 
-    regs.reg_array[method] = value;
+    regs.reg_array[method_call.method] = method_call.argument;
 
-    switch (method) {
+    switch (method_call.method) {
     case KEPLERMEMORY_REG_INDEX(exec): {
         state.write_offset = 0;
         break;
     }
     case KEPLERMEMORY_REG_INDEX(data): {
-        ProcessData(value);
+        ProcessData(method_call.argument);
         break;
     }
     }

src/video_core/engines/kepler_memory.h

@@ -9,6 +9,7 @@
 #include "common/bit_field.h"
 #include "common/common_funcs.h"
 #include "common/common_types.h"
+#include "video_core/gpu.h"
 #include "video_core/memory_manager.h"
 
 namespace VideoCore {
@@ -26,7 +27,7 @@ public:
     ~KeplerMemory();
 
     /// Write the value to the register identified by method.
-    void WriteReg(u32 method, u32 value);
+    void CallMethod(const GPU::MethodCall& method_call);
 
     struct Regs {
         static constexpr size_t NUM_REGS = 0x7F;

src/video_core/engines/maxwell_3d.cpp

@@ -97,71 +97,74 @@ void Maxwell3D::CallMacroMethod(u32 method, std::vector<u32> parameters) {
     macro_interpreter.Execute(search->second, std::move(parameters));
 }
 
-void Maxwell3D::WriteReg(u32 method, u32 value, u32 remaining_params) {
+void Maxwell3D::CallMethod(const GPU::MethodCall& method_call) {
     auto debug_context = Core::System::GetInstance().GetGPUDebugContext();
 
     // It is an error to write to a register other than the current macro's ARG register before it
     // has finished execution.
     if (executing_macro != 0) {
-        ASSERT(method == executing_macro + 1);
+        ASSERT(method_call.method == executing_macro + 1);
     }
 
     // Methods after 0xE00 are special, they're actually triggers for some microcode that was
     // uploaded to the GPU during initialization.
-    if (method >= MacroRegistersStart) {
+    if (method_call.method >= MacroRegistersStart) {
         // We're trying to execute a macro
         if (executing_macro == 0) {
             // A macro call must begin by writing the macro method's register, not its argument.
-            ASSERT_MSG((method % 2) == 0,
+            ASSERT_MSG((method_call.method % 2) == 0,
                        "Can't start macro execution by writing to the ARGS register");
-            executing_macro = method;
+            executing_macro = method_call.method;
         }
 
-        macro_params.push_back(value);
+        macro_params.push_back(method_call.argument);
 
         // Call the macro when there are no more parameters in the command buffer
-        if (remaining_params == 0) {
+        if (method_call.IsLastCall()) {
             CallMacroMethod(executing_macro, std::move(macro_params));
         }
         return;
     }
 
-    ASSERT_MSG(method < Regs::NUM_REGS,
+    ASSERT_MSG(method_call.method < Regs::NUM_REGS,
                "Invalid Maxwell3D register, increase the size of the Regs structure");
 
     if (debug_context) {
         debug_context->OnEvent(Tegra::DebugContext::Event::MaxwellCommandLoaded, nullptr);
     }
 
-    if (regs.reg_array[method] != value) {
-        regs.reg_array[method] = value;
+    if (regs.reg_array[method_call.method] != method_call.argument) {
+        regs.reg_array[method_call.method] = method_call.argument;
         // Vertex format
-        if (method >= MAXWELL3D_REG_INDEX(vertex_attrib_format) &&
-            method < MAXWELL3D_REG_INDEX(vertex_attrib_format) + regs.vertex_attrib_format.size()) {
+        if (method_call.method >= MAXWELL3D_REG_INDEX(vertex_attrib_format) &&
+            method_call.method <
+                MAXWELL3D_REG_INDEX(vertex_attrib_format) + regs.vertex_attrib_format.size()) {
             dirty_flags.vertex_attrib_format = true;
         }
 
         // Vertex buffer
-        if (method >= MAXWELL3D_REG_INDEX(vertex_array) &&
-            method < MAXWELL3D_REG_INDEX(vertex_array) + 4 * 32) {
-            dirty_flags.vertex_array |= 1u << ((method - MAXWELL3D_REG_INDEX(vertex_array)) >> 2);
-        } else if (method >= MAXWELL3D_REG_INDEX(vertex_array_limit) &&
-                   method < MAXWELL3D_REG_INDEX(vertex_array_limit) + 2 * 32) {
+        if (method_call.method >= MAXWELL3D_REG_INDEX(vertex_array) &&
+            method_call.method < MAXWELL3D_REG_INDEX(vertex_array) + 4 * 32) {
             dirty_flags.vertex_array |=
-                1u << ((method - MAXWELL3D_REG_INDEX(vertex_array_limit)) >> 1);
-        } else if (method >= MAXWELL3D_REG_INDEX(instanced_arrays) &&
-                   method < MAXWELL3D_REG_INDEX(instanced_arrays) + 32) {
-            dirty_flags.vertex_array |= 1u << (method - MAXWELL3D_REG_INDEX(instanced_arrays));
+                1u << ((method_call.method - MAXWELL3D_REG_INDEX(vertex_array)) >> 2);
+        } else if (method_call.method >= MAXWELL3D_REG_INDEX(vertex_array_limit) &&
+                   method_call.method < MAXWELL3D_REG_INDEX(vertex_array_limit) + 2 * 32) {
+            dirty_flags.vertex_array |=
+                1u << ((method_call.method - MAXWELL3D_REG_INDEX(vertex_array_limit)) >> 1);
+        } else if (method_call.method >= MAXWELL3D_REG_INDEX(instanced_arrays) &&
+                   method_call.method < MAXWELL3D_REG_INDEX(instanced_arrays) + 32) {
+            dirty_flags.vertex_array |=
+                1u << (method_call.method - MAXWELL3D_REG_INDEX(instanced_arrays));
         }
     }
 
-    switch (method) {
+    switch (method_call.method) {
     case MAXWELL3D_REG_INDEX(macros.data): {
-        ProcessMacroUpload(value);
+        ProcessMacroUpload(method_call.argument);
         break;
     }
     case MAXWELL3D_REG_INDEX(macros.bind): {
-        ProcessMacroBind(value);
+        ProcessMacroBind(method_call.argument);
         break;
     }
     case MAXWELL3D_REG_INDEX(const_buffer.cb_data[0]):
@@ -180,7 +183,7 @@ void Maxwell3D::WriteReg(u32 method, u32 value, u32 remaining_params) {
     case MAXWELL3D_REG_INDEX(const_buffer.cb_data[13]):
     case MAXWELL3D_REG_INDEX(const_buffer.cb_data[14]):
     case MAXWELL3D_REG_INDEX(const_buffer.cb_data[15]): {
-        ProcessCBData(value);
+        ProcessCBData(method_call.argument);
         break;
     }
     case MAXWELL3D_REG_INDEX(cb_bind[0].raw_config): {

src/video_core/engines/maxwell_3d.h

@@ -1080,7 +1080,7 @@ public:
     u32 GetRegisterValue(u32 method) const;
 
     /// Write the value to the register identified by method.
-    void WriteReg(u32 method, u32 value, u32 remaining_params);
+    void CallMethod(const GPU::MethodCall& method_call);
 
     /// Returns a list of enabled textures for the specified shader stage.
     std::vector<Texture::FullTextureInfo> GetStageTextures(Regs::ShaderStage stage) const;

src/video_core/engines/maxwell_compute.cpp

@@ -8,13 +8,13 @@
 
 namespace Tegra::Engines {
 
-void MaxwellCompute::WriteReg(u32 method, u32 value) {
-    ASSERT_MSG(method < Regs::NUM_REGS,
+void MaxwellCompute::CallMethod(const GPU::MethodCall& method_call) {
+    ASSERT_MSG(method_call.method < Regs::NUM_REGS,
                "Invalid MaxwellCompute register, increase the size of the Regs structure");
 
-    regs.reg_array[method] = value;
+    regs.reg_array[method_call.method] = method_call.argument;
 
-    switch (method) {
+    switch (method_call.method) {
     case MAXWELL_COMPUTE_REG_INDEX(compute): {
         LOG_CRITICAL(HW_GPU, "Compute shaders are not implemented");
         UNREACHABLE();

src/video_core/engines/maxwell_compute.h

@@ -9,6 +9,7 @@
 #include "common/bit_field.h"
 #include "common/common_funcs.h"
 #include "common/common_types.h"
+#include "video_core/gpu.h"
 
 namespace Tegra::Engines {
 
@@ -42,7 +43,7 @@ public:
                   "MaxwellCompute Regs has wrong size");
 
     /// Write the value to the register identified by method.
-    void WriteReg(u32 method, u32 value);
+    void CallMethod(const GPU::MethodCall& method_call);
 };
 
 #define ASSERT_REG_POSITION(field_name, position)                                                  \

src/video_core/engines/maxwell_dma.cpp

@@ -14,16 +14,16 @@ namespace Tegra::Engines {
 MaxwellDMA::MaxwellDMA(VideoCore::RasterizerInterface& rasterizer, MemoryManager& memory_manager)
     : memory_manager(memory_manager), rasterizer{rasterizer} {}
 
-void MaxwellDMA::WriteReg(u32 method, u32 value) {
-    ASSERT_MSG(method < Regs::NUM_REGS,
+void MaxwellDMA::CallMethod(const GPU::MethodCall& method_call) {
+    ASSERT_MSG(method_call.method < Regs::NUM_REGS,
                "Invalid MaxwellDMA register, increase the size of the Regs structure");
 
-    regs.reg_array[method] = value;
+    regs.reg_array[method_call.method] = method_call.argument;
 
 #define MAXWELLDMA_REG_INDEX(field_name)                                                           \
     (offsetof(Tegra::Engines::MaxwellDMA::Regs, field_name) / sizeof(u32))
 
-    switch (method) {
+    switch (method_call.method) {
     case MAXWELLDMA_REG_INDEX(exec): {
         HandleCopy();
         break;

src/video_core/engines/maxwell_dma.h

@@ -24,7 +24,7 @@ public:
     ~MaxwellDMA() = default;
 
     /// Write the value to the register identified by method.
-    void WriteReg(u32 method, u32 value);
+    void CallMethod(const GPU::MethodCall& method_call);
 
     struct Regs {
         static constexpr std::size_t NUM_REGS = 0x1D6;

src/video_core/gpu.cpp

@@ -3,6 +3,7 @@
 // Refer to the license.txt file included.
 
 #include "common/assert.h"
+#include "common/microprofile.h"
 #include "video_core/engines/fermi_2d.h"
 #include "video_core/engines/kepler_memory.h"
 #include "video_core/engines/maxwell_3d.h"
@@ -26,6 +27,7 @@ u32 FramebufferConfig::BytesPerPixel(PixelFormat format) {
 
 GPU::GPU(VideoCore::RasterizerInterface& rasterizer) {
     memory_manager = std::make_unique<Tegra::MemoryManager>();
+    dma_pusher = std::make_unique<Tegra::DmaPusher>(*this);
     maxwell_3d = std::make_unique<Engines::Maxwell3D>(rasterizer, *memory_manager);
     fermi_2d = std::make_unique<Engines::Fermi2D>(rasterizer, *memory_manager);
     maxwell_compute = std::make_unique<Engines::MaxwellCompute>();
@@ -51,6 +53,14 @@ const MemoryManager& GPU::MemoryManager() const {
     return *memory_manager;
 }
 
+DmaPusher& GPU::DmaPusher() {
+    return *dma_pusher;
+}
+
+const DmaPusher& GPU::DmaPusher() const {
+    return *dma_pusher;
+}
+
 u32 RenderTargetBytesPerPixel(RenderTargetFormat format) {
     ASSERT(format != RenderTargetFormat::NONE);
 
@@ -113,4 +123,52 @@ u32 DepthFormatBytesPerPixel(DepthFormat format) {
     }
 }
 
+enum class BufferMethods {
+    BindObject = 0,
+    CountBufferMethods = 0x40,
+};
+
+MICROPROFILE_DEFINE(ProcessCommandLists, "GPU", "Execute command buffer", MP_RGB(128, 128, 192));
+
+void GPU::CallMethod(const MethodCall& method_call) {
+    MICROPROFILE_SCOPE(ProcessCommandLists);
+
+    LOG_TRACE(HW_GPU,
+              "Processing method {:08X} on subchannel {} value "
+              "{:08X} remaining params {}",
+              MethCall.method, MethCall.subchannel, value, remaining_params);
+
+    ASSERT(method_call.subchannel < bound_engines.size());
+
+    if (method_call.method == static_cast<u32>(BufferMethods::BindObject)) {
+        // Bind the current subchannel to the desired engine id.
+        LOG_DEBUG(HW_GPU, "Binding subchannel {} to engine {}", method_call.subchannel,
+                  method_call.argument);
+        bound_engines[method_call.subchannel] = static_cast<EngineID>(method_call.argument);
+        return;
+    }
+
+    const EngineID engine = bound_engines[method_call.subchannel];
+
+    switch (engine) {
+    case EngineID::FERMI_TWOD_A:
+        fermi_2d->CallMethod(method_call);
+        break;
+    case EngineID::MAXWELL_B:
+        maxwell_3d->CallMethod(method_call);
+        break;
+    case EngineID::MAXWELL_COMPUTE_B:
+        maxwell_compute->CallMethod(method_call);
+        break;
+    case EngineID::MAXWELL_DMA_COPY_A:
+        maxwell_dma->CallMethod(method_call);
+        break;
+    case EngineID::KEPLER_INLINE_TO_MEMORY_B:
+        kepler_memory->CallMethod(method_call);
+        break;
+    default:
+        UNIMPLEMENTED_MSG("Unimplemented engine");
+    }
+}
+
 } // namespace Tegra

src/video_core/gpu.h

@@ -9,6 +9,7 @@
 #include <vector>
 #include "common/common_types.h"
 #include "core/hle/service/nvflinger/buffer_queue.h"
+#include "video_core/dma_pusher.h"
 #include "video_core/memory_manager.h"
 
 namespace VideoCore {
@@ -119,8 +120,23 @@ public:
     explicit GPU(VideoCore::RasterizerInterface& rasterizer);
     ~GPU();
 
-    /// Processes a command list stored at the specified address in GPU memory.
-    void ProcessCommandLists(const std::vector<CommandListHeader>& commands);
+    struct MethodCall {
+        u32 method{};
+        u32 argument{};
+        u32 subchannel{};
+        u32 method_count{};
+
+        bool IsLastCall() const {
+            return method_count <= 1;
+        }
+
+        MethodCall(u32 method, u32 argument, u32 subchannel = 0, u32 method_count = 0)
+            : method(method), argument(argument), subchannel(subchannel),
+              method_count(method_count) {}
+    };
+
+    /// Calls a GPU method.
+    void CallMethod(const MethodCall& method_call);
 
     /// Returns a reference to the Maxwell3D GPU engine.
     Engines::Maxwell3D& Maxwell3D();
@@ -134,7 +150,14 @@ public:
     /// Returns a const reference to the GPU memory manager.
     const Tegra::MemoryManager& MemoryManager() const;
 
+    /// Returns a reference to the GPU DMA pusher.
+    Tegra::DmaPusher& DmaPusher();
+
+    /// Returns a const reference to the GPU DMA pusher.
+    const Tegra::DmaPusher& DmaPusher() const;
+
 private:
+    std::unique_ptr<Tegra::DmaPusher> dma_pusher;
     std::unique_ptr<Tegra::MemoryManager> memory_manager;
 
     /// Mapping of command subchannels to their bound engine ids.

src/video_core/macro_interpreter.cpp

@@ -250,7 +250,7 @@ void MacroInterpreter::SetMethodAddress(u32 address) {
 }
 
 void MacroInterpreter::Send(u32 value) {
-    maxwell3d.WriteReg(method_address.address, value, 0);
+    maxwell3d.CallMethod({method_address.address, value});
     // Increment the method address by the method increment.
     method_address.address.Assign(method_address.address.Value() +
                                   method_address.increment.Value());