Merge pull request #3742 from FernandoS27/command-list

Optimize GPU Command Lists and Introduce Fast GPU Time Option

src/core/settings.h

@@ -446,6 +446,7 @@ struct Values {
     bool use_asynchronous_gpu_emulation;
     bool use_vsync;
     bool force_30fps_mode;
+    bool use_fast_gpu_time;
 
     float bg_red;
     float bg_green;

src/video_core/dma_pusher.cpp

@@ -71,16 +71,22 @@ bool DmaPusher::Step() {
     gpu.MemoryManager().ReadBlockUnsafe(dma_get, command_headers.data(),
                                         command_list_header.size * sizeof(u32));
 
-    for (const CommandHeader& command_header : command_headers) {
+    for (std::size_t index = 0; index < command_headers.size();) {
+        const CommandHeader& command_header = command_headers[index];
 
-        // now, see if we're in the middle of a command
+        if (dma_state.method_count) {
-        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) {
+                const u32 max_write = static_cast<u32>(
+                    std::min<std::size_t>(index + dma_state.method_count, command_headers.size()) -
+                    index);
+                CallMultiMethod(&command_header.argument, max_write);
+                dma_state.method_count -= max_write;
+                index += max_write;
+                continue;
+            } else {
+                CallMethod(command_header.argument);
+            }
 
             if (!dma_state.non_incrementing) {
                 dma_state.method++;
@@ -120,6 +126,7 @@ bool DmaPusher::Step() {
                 break;
             }
         }
+        index++;
     }
 
     if (!non_main) {
@@ -140,4 +147,9 @@ void DmaPusher::CallMethod(u32 argument) const {
     gpu.CallMethod({dma_state.method, argument, dma_state.subchannel, dma_state.method_count});
 }
 
+void DmaPusher::CallMultiMethod(const u32* base_start, u32 num_methods) const {
+    gpu.CallMultiMethod(dma_state.method, dma_state.subchannel, base_start, num_methods,
+                        dma_state.method_count);
+}
+
 } // namespace Tegra

src/video_core/dma_pusher.h

@@ -75,6 +75,7 @@ private:
     void SetState(const CommandHeader& command_header);
 
     void CallMethod(u32 argument) const;
+    void CallMultiMethod(const u32* base_start, u32 num_methods) const;
 
     std::vector<CommandHeader> command_headers; ///< Buffer for list of commands fetched at once
 

src/video_core/engines/fermi_2d.cpp

@@ -28,6 +28,12 @@ void Fermi2D::CallMethod(const GPU::MethodCall& method_call) {
     }
 }
 
+void Fermi2D::CallMultiMethod(u32 method, const u32* base_start, u32 amount, u32 methods_pending) {
+    for (std::size_t i = 0; i < amount; i++) {
+        CallMethod({method, base_start[i], 0, methods_pending - static_cast<u32>(i)});
+    }
+}
+
 static std::pair<u32, u32> DelimitLine(u32 src_1, u32 src_2, u32 dst_1, u32 dst_2, u32 src_line) {
     const u32 line_a = src_2 - src_1;
     const u32 line_b = dst_2 - dst_1;

src/video_core/engines/fermi_2d.h

@@ -39,6 +39,9 @@ public:
     /// Write the value to the register identified by method.
     void CallMethod(const GPU::MethodCall& method_call);
 
+    /// Write multiple values to the register identified by method.
+    void CallMultiMethod(u32 method, const u32* base_start, u32 amount, u32 methods_pending);
+
     enum class Origin : u32 {
         Center = 0,
         Corner = 1,

src/video_core/engines/kepler_compute.cpp

@@ -51,6 +51,13 @@ void KeplerCompute::CallMethod(const GPU::MethodCall& method_call) {
     }
 }
 
+void KeplerCompute::CallMultiMethod(u32 method, const u32* base_start, u32 amount,
+                                    u32 methods_pending) {
+    for (std::size_t i = 0; i < amount; i++) {
+        CallMethod({method, base_start[i], 0, methods_pending - static_cast<u32>(i)});
+    }
+}
+
 Texture::FullTextureInfo KeplerCompute::GetTexture(std::size_t offset) const {
     const std::bitset<8> cbuf_mask = launch_description.const_buffer_enable_mask.Value();
     ASSERT(cbuf_mask[regs.tex_cb_index]);

src/video_core/engines/kepler_compute.h

@@ -202,6 +202,9 @@ public:
     /// Write the value to the register identified by method.
     void CallMethod(const GPU::MethodCall& method_call);
 
+    /// Write multiple values to the register identified by method.
+    void CallMultiMethod(u32 method, const u32* base_start, u32 amount, u32 methods_pending);
+
     Texture::FullTextureInfo GetTexture(std::size_t offset) const;
 
     /// Given a texture handle, returns the TSC and TIC entries.

src/video_core/engines/kepler_memory.cpp

@@ -41,4 +41,11 @@ void KeplerMemory::CallMethod(const GPU::MethodCall& method_call) {
     }
 }
 
+void KeplerMemory::CallMultiMethod(u32 method, const u32* base_start, u32 amount,
+                                   u32 methods_pending) {
+    for (std::size_t i = 0; i < amount; i++) {
+        CallMethod({method, base_start[i], 0, methods_pending - static_cast<u32>(i)});
+    }
+}
+
 } // namespace Tegra::Engines

src/video_core/engines/kepler_memory.h

@@ -40,6 +40,9 @@ public:
     /// Write the value to the register identified by method.
     void CallMethod(const GPU::MethodCall& method_call);
 
+    /// Write multiple values to the register identified by method.
+    void CallMultiMethod(u32 method, const u32* base_start, u32 amount, u32 methods_pending);
+
     struct Regs {
         static constexpr size_t NUM_REGS = 0x7F;
 

src/video_core/engines/maxwell_3d.cpp

@@ -280,6 +280,58 @@ void Maxwell3D::CallMethod(const GPU::MethodCall& method_call) {
     }
 }
 
+void Maxwell3D::CallMultiMethod(u32 method, const u32* base_start, u32 amount,
+                                u32 methods_pending) {
+    // Methods after 0xE00 are special, they're actually triggers for some microcode that was
+    // uploaded to the GPU during initialization.
+    if (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,
+                       "Can't start macro execution by writing to the ARGS register");
+            executing_macro = method;
+        }
+
+        for (std::size_t i = 0; i < amount; i++) {
+            macro_params.push_back(base_start[i]);
+        }
+
+        // Call the macro when there are no more parameters in the command buffer
+        if (amount == methods_pending) {
+            CallMacroMethod(executing_macro, macro_params.size(), macro_params.data());
+            macro_params.clear();
+        }
+        return;
+    }
+    switch (method) {
+    case MAXWELL3D_REG_INDEX(const_buffer.cb_data[0]):
+    case MAXWELL3D_REG_INDEX(const_buffer.cb_data[1]):
+    case MAXWELL3D_REG_INDEX(const_buffer.cb_data[2]):
+    case MAXWELL3D_REG_INDEX(const_buffer.cb_data[3]):
+    case MAXWELL3D_REG_INDEX(const_buffer.cb_data[4]):
+    case MAXWELL3D_REG_INDEX(const_buffer.cb_data[5]):
+    case MAXWELL3D_REG_INDEX(const_buffer.cb_data[6]):
+    case MAXWELL3D_REG_INDEX(const_buffer.cb_data[7]):
+    case MAXWELL3D_REG_INDEX(const_buffer.cb_data[8]):
+    case MAXWELL3D_REG_INDEX(const_buffer.cb_data[9]):
+    case MAXWELL3D_REG_INDEX(const_buffer.cb_data[10]):
+    case MAXWELL3D_REG_INDEX(const_buffer.cb_data[11]):
+    case MAXWELL3D_REG_INDEX(const_buffer.cb_data[12]):
+    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]): {
+        ProcessCBMultiData(method, base_start, amount);
+        break;
+    }
+    default: {
+        for (std::size_t i = 0; i < amount; i++) {
+            CallMethod({method, base_start[i], 0, methods_pending - static_cast<u32>(i)});
+        }
+    }
+    }
+}
+
 void Maxwell3D::StepInstance(const MMEDrawMode expected_mode, const u32 count) {
     if (mme_draw.current_mode == MMEDrawMode::Undefined) {
         if (mme_draw.gl_begin_consume) {
@@ -570,6 +622,28 @@ void Maxwell3D::StartCBData(u32 method) {
     ProcessCBData(regs.const_buffer.cb_data[cb_data_state.id]);
 }
 
+void Maxwell3D::ProcessCBMultiData(u32 method, const u32* start_base, u32 amount) {
+    if (cb_data_state.current != method) {
+        if (cb_data_state.current != null_cb_data) {
+            FinishCBData();
+        }
+        constexpr u32 first_cb_data = MAXWELL3D_REG_INDEX(const_buffer.cb_data[0]);
+        cb_data_state.start_pos = regs.const_buffer.cb_pos;
+        cb_data_state.id = method - first_cb_data;
+        cb_data_state.current = method;
+        cb_data_state.counter = 0;
+    }
+    const std::size_t id = cb_data_state.id;
+    const std::size_t size = amount;
+    std::size_t i = 0;
+    for (; i < size; i++) {
+        cb_data_state.buffer[id][cb_data_state.counter] = start_base[i];
+        cb_data_state.counter++;
+    }
+    // Increment the current buffer position.
+    regs.const_buffer.cb_pos = regs.const_buffer.cb_pos + 4 * amount;
+}
+
 void Maxwell3D::FinishCBData() {
     // Write the input value to the current const buffer at the current position.
     const GPUVAddr buffer_address = regs.const_buffer.BufferAddress();

src/video_core/engines/maxwell_3d.h

@@ -1359,6 +1359,9 @@ public:
     /// Write the value to the register identified by method.
     void CallMethod(const GPU::MethodCall& method_call);
 
+    /// Write multiple values to the register identified by method.
+    void CallMultiMethod(u32 method, const u32* base_start, u32 amount, u32 methods_pending);
+
     /// Write the value to the register identified by method.
     void CallMethodFromMME(const GPU::MethodCall& method_call);
 
@@ -1512,6 +1515,7 @@ private:
     /// Handles a write to the CB_DATA[i] register.
     void StartCBData(u32 method);
     void ProcessCBData(u32 value);
+    void ProcessCBMultiData(u32 method, const u32* start_base, u32 amount);
     void FinishCBData();
 
     /// Handles a write to the CB_BIND register.

src/video_core/engines/maxwell_dma.cpp

@@ -36,6 +36,13 @@ void MaxwellDMA::CallMethod(const GPU::MethodCall& method_call) {
 #undef MAXWELLDMA_REG_INDEX
 }
 
+void MaxwellDMA::CallMultiMethod(u32 method, const u32* base_start, u32 amount,
+                                 u32 methods_pending) {
+    for (std::size_t i = 0; i < amount; i++) {
+        CallMethod({method, base_start[i], 0, methods_pending - static_cast<u32>(i)});
+    }
+}
+
 void MaxwellDMA::HandleCopy() {
     LOG_TRACE(HW_GPU, "Requested a DMA copy");
 

src/video_core/engines/maxwell_dma.h

@@ -35,6 +35,9 @@ public:
     /// Write the value to the register identified by method.
     void CallMethod(const GPU::MethodCall& method_call);
 
+    /// Write multiple values to the register identified by method.
+    void CallMultiMethod(u32 method, const u32* base_start, u32 amount, u32 methods_pending);
+
     struct Regs {
         static constexpr std::size_t NUM_REGS = 0x1D6;
 

src/video_core/gpu.cpp

@@ -9,6 +9,7 @@
 #include "core/core_timing_util.h"
 #include "core/frontend/emu_window.h"
 #include "core/memory.h"
+#include "core/settings.h"
 #include "video_core/engines/fermi_2d.h"
 #include "video_core/engines/kepler_compute.h"
 #include "video_core/engines/kepler_memory.h"
@@ -154,7 +155,10 @@ u64 GPU::GetTicks() const {
     constexpr u64 gpu_ticks_den = 625;
 
     const u64 cpu_ticks = system.CoreTiming().GetTicks();
-    const u64 nanoseconds = Core::Timing::CyclesToNs(cpu_ticks).count();
+    u64 nanoseconds = Core::Timing::CyclesToNs(cpu_ticks).count();
+    if (Settings::values.use_fast_gpu_time) {
+        nanoseconds /= 256;
+    }
     const u64 nanoseconds_num = nanoseconds / gpu_ticks_den;
     const u64 nanoseconds_rem = nanoseconds % gpu_ticks_den;
     return nanoseconds_num * gpu_ticks_num + (nanoseconds_rem * gpu_ticks_num) / gpu_ticks_den;
@@ -209,16 +213,32 @@ void GPU::CallMethod(const MethodCall& method_call) {
 
     ASSERT(method_call.subchannel < bound_engines.size());
 
-    if (ExecuteMethodOnEngine(method_call)) {
+    if (ExecuteMethodOnEngine(method_call.method)) {
         CallEngineMethod(method_call);
     } else {
         CallPullerMethod(method_call);
     }
 }
 
-bool GPU::ExecuteMethodOnEngine(const MethodCall& method_call) {
+void GPU::CallMultiMethod(u32 method, u32 subchannel, const u32* base_start, u32 amount,
-    const auto method = static_cast<BufferMethods>(method_call.method);
+                          u32 methods_pending) {
-    return method >= BufferMethods::NonPullerMethods;
+    LOG_TRACE(HW_GPU, "Processing method {:08X} on subchannel {}", method, subchannel);
+
+    ASSERT(subchannel < bound_engines.size());
+
+    if (ExecuteMethodOnEngine(method)) {
+        CallEngineMultiMethod(method, subchannel, base_start, amount, methods_pending);
+    } else {
+        for (std::size_t i = 0; i < amount; i++) {
+            CallPullerMethod(
+                {method, base_start[i], subchannel, methods_pending - static_cast<u32>(i)});
+        }
+    }
+}
+
+bool GPU::ExecuteMethodOnEngine(u32 method) {
+    const auto buffer_method = static_cast<BufferMethods>(method);
+    return buffer_method >= BufferMethods::NonPullerMethods;
 }
 
 void GPU::CallPullerMethod(const MethodCall& method_call) {
@@ -298,6 +318,31 @@ void GPU::CallEngineMethod(const MethodCall& method_call) {
     }
 }
 
+void GPU::CallEngineMultiMethod(u32 method, u32 subchannel, const u32* base_start, u32 amount,
+                                u32 methods_pending) {
+    const EngineID engine = bound_engines[subchannel];
+
+    switch (engine) {
+    case EngineID::FERMI_TWOD_A:
+        fermi_2d->CallMultiMethod(method, base_start, amount, methods_pending);
+        break;
+    case EngineID::MAXWELL_B:
+        maxwell_3d->CallMultiMethod(method, base_start, amount, methods_pending);
+        break;
+    case EngineID::KEPLER_COMPUTE_B:
+        kepler_compute->CallMultiMethod(method, base_start, amount, methods_pending);
+        break;
+    case EngineID::MAXWELL_DMA_COPY_A:
+        maxwell_dma->CallMultiMethod(method, base_start, amount, methods_pending);
+        break;
+    case EngineID::KEPLER_INLINE_TO_MEMORY_B:
+        kepler_memory->CallMultiMethod(method, base_start, amount, methods_pending);
+        break;
+    default:
+        UNIMPLEMENTED_MSG("Unimplemented engine");
+    }
+}
+
 void GPU::ProcessBindMethod(const MethodCall& method_call) {
     // Bind the current subchannel to the desired engine id.
     LOG_DEBUG(HW_GPU, "Binding subchannel {} to engine {}", method_call.subchannel,

src/video_core/gpu.h

@@ -155,6 +155,10 @@ public:
     /// Calls a GPU method.
     void CallMethod(const MethodCall& method_call);
 
+    /// Calls a GPU multivalue method.
+    void CallMultiMethod(u32 method, u32 subchannel, const u32* base_start, u32 amount,
+                         u32 methods_pending);
+
     /// Flush all current written commands into the host GPU for execution.
     void FlushCommands();
     /// Synchronizes CPU writes with Host GPU memory.
@@ -309,8 +313,12 @@ private:
     /// Calls a GPU engine method.
     void CallEngineMethod(const MethodCall& method_call);
 
+    /// Calls a GPU engine multivalue method.
+    void CallEngineMultiMethod(u32 method, u32 subchannel, const u32* base_start, u32 amount,
+                               u32 methods_pending);
+
     /// Determines where the method should be executed.
-    bool ExecuteMethodOnEngine(const MethodCall& method_call);
+    bool ExecuteMethodOnEngine(u32 method);
 
 protected:
     std::unique_ptr<Tegra::DmaPusher> dma_pusher;

src/yuzu/configuration/config.cpp

@@ -643,6 +643,8 @@ void Config::ReadRendererValues() {
     Settings::values.use_asynchronous_gpu_emulation =
         ReadSetting(QStringLiteral("use_asynchronous_gpu_emulation"), false).toBool();
     Settings::values.use_vsync = ReadSetting(QStringLiteral("use_vsync"), true).toBool();
+    Settings::values.use_fast_gpu_time =
+        ReadSetting(QStringLiteral("use_fast_gpu_time"), true).toBool();
     Settings::values.force_30fps_mode =
         ReadSetting(QStringLiteral("force_30fps_mode"), false).toBool();
 
@@ -1084,6 +1086,7 @@ void Config::SaveRendererValues() {
     WriteSetting(QStringLiteral("use_asynchronous_gpu_emulation"),
                  Settings::values.use_asynchronous_gpu_emulation, false);
     WriteSetting(QStringLiteral("use_vsync"), Settings::values.use_vsync, true);
+    WriteSetting(QStringLiteral("use_fast_gpu_time"), Settings::values.use_fast_gpu_time, true);
     WriteSetting(QStringLiteral("force_30fps_mode"), Settings::values.force_30fps_mode, false);
 
     // Cast to double because Qt's written float values are not human-readable

src/yuzu/configuration/configure_graphics_advanced.cpp

@@ -22,6 +22,7 @@ void ConfigureGraphicsAdvanced::SetConfiguration() {
     ui->gpu_accuracy->setCurrentIndex(static_cast<int>(Settings::values.gpu_accuracy));
     ui->use_vsync->setEnabled(runtime_lock);
     ui->use_vsync->setChecked(Settings::values.use_vsync);
+    ui->use_fast_gpu_time->setChecked(Settings::values.use_fast_gpu_time);
     ui->force_30fps_mode->setEnabled(runtime_lock);
     ui->force_30fps_mode->setChecked(Settings::values.force_30fps_mode);
     ui->anisotropic_filtering_combobox->setEnabled(runtime_lock);
@@ -32,6 +33,7 @@ void ConfigureGraphicsAdvanced::ApplyConfiguration() {
     auto gpu_accuracy = static_cast<Settings::GPUAccuracy>(ui->gpu_accuracy->currentIndex());
     Settings::values.gpu_accuracy = gpu_accuracy;
     Settings::values.use_vsync = ui->use_vsync->isChecked();
+    Settings::values.use_fast_gpu_time = ui->use_fast_gpu_time->isChecked();
     Settings::values.force_30fps_mode = ui->force_30fps_mode->isChecked();
     Settings::values.max_anisotropy = ui->anisotropic_filtering_combobox->currentIndex();
 }

src/yuzu/configuration/configure_graphics_advanced.ui

@@ -69,6 +69,13 @@
           </property>
          </widget>
         </item>
+        <item>
+         <widget class="QCheckBox" name="use_fast_gpu_time">
+          <property name="text">
+           <string>Use Fast GPU Time</string>
+          </property>
+         </widget>
+        </item>
         <item>
          <layout class="QHBoxLayout" name="horizontalLayout_1">
           <item>

src/yuzu_cmd/config.cpp

@@ -394,6 +394,8 @@ void Config::ReadValues() {
         sdl2_config->GetBoolean("Renderer", "use_asynchronous_gpu_emulation", false);
     Settings::values.use_vsync =
         static_cast<u16>(sdl2_config->GetInteger("Renderer", "use_vsync", 1));
+    Settings::values.use_fast_gpu_time =
+        sdl2_config->GetBoolean("Renderer", "use_fast_gpu_time", true);
 
     Settings::values.bg_red = static_cast<float>(sdl2_config->GetReal("Renderer", "bg_red", 0.0));
     Settings::values.bg_green =

src/yuzu_tester/config.cpp

@@ -130,6 +130,8 @@ void Config::ReadValues() {
     Settings::values.gpu_accuracy = static_cast<Settings::GPUAccuracy>(gpu_accuracy_level);
     Settings::values.use_asynchronous_gpu_emulation =
         sdl2_config->GetBoolean("Renderer", "use_asynchronous_gpu_emulation", false);
+    Settings::values.use_fast_gpu_time =
+        sdl2_config->GetBoolean("Renderer", "use_fast_gpu_time", true);
 
     Settings::values.bg_red = static_cast<float>(sdl2_config->GetReal("Renderer", "bg_red", 0.0));
     Settings::values.bg_green =