GPU: Implement channel scheduling.

src/video_core/control/scheduler.cpp

@@ -13,20 +13,102 @@ Scheduler::Scheduler(GPU& gpu_) : gpu{gpu_} {}
 
 Scheduler::~Scheduler() = default;
 
+void Scheduler::Init() {
+    master_control = Common::Fiber::ThreadToFiber();
+}
+
+void Scheduler::Resume() {
+    bool nothing_pending;
+    do {
+        nothing_pending = true;
+        current_fifo = nullptr;
+        {
+            std::unique_lock lk(scheduling_guard);
+            size_t num_iters = gpfifos.size();
+            for (size_t i = 0; i < num_iters; i++) {
+                size_t current_id = (current_fifo_rotation_id + i) % gpfifos.size();
+                auto& fifo = gpfifos[current_id];
+                if (!fifo.is_active) {
+                    continue;
+                }
+                std::scoped_lock lk2(fifo.guard);
+                if (!fifo.pending_work.empty() || fifo.working.load(std::memory_order_acquire)) {
+                    current_fifo = &fifo;
+                    current_fifo_rotation_id = current_id;
+                    nothing_pending = false;
+                    break;
+                }
+            }
+        }
+        if (current_fifo) {
+            Common::Fiber::YieldTo(master_control, *current_fifo->context);
+            current_fifo = nullptr;
+        }
+    } while (!nothing_pending);
+}
+
+void Scheduler::Yield() {
+    ASSERT(current_fifo != nullptr);
+    Common::Fiber::YieldTo(current_fifo->context, *master_control);
+    gpu.BindChannel(current_fifo->bind_id);
+}
+
 void Scheduler::Push(s32 channel, CommandList&& entries) {
     std::unique_lock lk(scheduling_guard);
-    auto it = channels.find(channel);
+    auto it = channel_gpfifo_ids.find(channel);
-    ASSERT(it != channels.end());
+    ASSERT(it != channel_gpfifo_ids.end());
-    auto channel_state = it->second;
+    auto gpfifo_id = it->second;
-    gpu.BindChannel(channel_state->bind_id);
+    auto& fifo = gpfifos[gpfifo_id];
-    channel_state->dma_pusher->Push(std::move(entries));
+    {
-    channel_state->dma_pusher->DispatchCalls();
+        std::scoped_lock lk2(fifo.guard);
+        fifo.pending_work.emplace_back(std::move(entries));
+    }
+}
+
+void Scheduler::ChannelLoop(size_t gpfifo_id, s32 channel_id) {
+    gpu.BindChannel(channel_id);
+    auto& fifo = gpfifos[gpfifo_id];
+    while (true) {
+        auto* channel_state = channels[channel_id].get();
+        fifo.guard.lock();
+        while (!fifo.pending_work.empty()) {
+            {
+
+                fifo.working.store(true, std::memory_order_release);
+                CommandList&& entries = std::move(fifo.pending_work.front());
+                channel_state->dma_pusher->Push(std::move(entries));
+                fifo.pending_work.pop_front();
+            }
+            fifo.guard.unlock();
+            channel_state->dma_pusher->DispatchCalls();
+            fifo.guard.lock();
+        }
+        fifo.working.store(false, std::memory_order_relaxed);
+        fifo.guard.unlock();
+        Common::Fiber::YieldTo(fifo.context, *master_control);
+        gpu.BindChannel(channel_id);
+    }
 }
 
 void Scheduler::DeclareChannel(std::shared_ptr<ChannelState> new_channel) {
     s32 channel = new_channel->bind_id;
     std::unique_lock lk(scheduling_guard);
     channels.emplace(channel, new_channel);
+    size_t new_fifo_id;
+    if (!free_fifos.empty()) {
+        new_fifo_id = free_fifos.front();
+        free_fifos.pop_front();
+    } else {
+        new_fifo_id = gpfifos.size();
+        gpfifos.emplace_back();
+    }
+    auto& new_fifo = gpfifos[new_fifo_id];
+    channel_gpfifo_ids[channel] = new_fifo_id;
+    new_fifo.is_active = true;
+    new_fifo.bind_id = channel;
+    new_fifo.pending_work.clear();
+    std::function<void()> callback = std::bind(&Scheduler::ChannelLoop, this, new_fifo_id, channel);
+    new_fifo.context = std::make_shared<Common::Fiber>(std::move(callback));
 }
 
 } // namespace Tegra::Control

src/video_core/control/scheduler.h

@@ -3,10 +3,13 @@
 
 #pragma once
 
+#include <atomic>
+#include <deque>
 #include <memory>
 #include <mutex>
 #include <unordered_map>
 
+#include "common/fiber.h"
 #include "video_core/dma_pusher.h"
 
 namespace Tegra {
@@ -22,14 +25,36 @@ public:
     explicit Scheduler(GPU& gpu_);
     ~Scheduler();
 
+    void Init();
+
+    void Resume();
+
+    void Yield();
+
     void Push(s32 channel, CommandList&& entries);
 
     void DeclareChannel(std::shared_ptr<ChannelState> new_channel);
 
 private:
+    void ChannelLoop(size_t gpfifo_id, s32 channel_id);
+
     std::unordered_map<s32, std::shared_ptr<ChannelState>> channels;
+    std::unordered_map<s32, size_t> channel_gpfifo_ids;
     std::mutex scheduling_guard;
+    std::shared_ptr<Common::Fiber> master_control;
+    struct GPFifoContext {
+        bool is_active;
+        std::shared_ptr<Common::Fiber> context;
+        std::deque<CommandList> pending_work;
+        std::atomic<bool> working{};
+        std::mutex guard;
+        s32 bind_id;
+    };
+    std::deque<GPFifoContext> gpfifos;
+    std::deque<size_t> free_fifos;
     GPU& gpu;
+    size_t current_fifo_rotation_id{};
+    GPFifoContext* current_fifo{};
 };
 
 } // namespace Control

src/video_core/engines/puller.cpp

@@ -6,6 +6,7 @@
 #include "common/settings.h"
 #include "core/core.h"
 #include "video_core/control/channel_state.h"
+#include "video_core/control/scheduler.h"
 #include "video_core/dma_pusher.h"
 #include "video_core/engines/fermi_2d.h"
 #include "video_core/engines/kepler_compute.h"
@@ -14,6 +15,8 @@
 #include "video_core/engines/maxwell_dma.h"
 #include "video_core/engines/puller.h"
 #include "video_core/gpu.h"
+#include "video_core/host1x/host1x.h"
+#include "video_core/host1x/syncpoint_manager.h"
 #include "video_core/memory_manager.h"
 #include "video_core/rasterizer_interface.h"
 
@@ -60,11 +63,14 @@ void Puller::ProcessBindMethod(const MethodCall& method_call) {
 }
 
 void Puller::ProcessFenceActionMethod() {
+    auto& syncpoint_manager = gpu.Host1x().GetSyncpointManager();
     switch (regs.fence_action.op) {
     case Puller::FenceOperation::Acquire:
-        // UNIMPLEMENTED_MSG("Channel Scheduling pending.");
+        while (regs.fence_value >
-        // WaitFence(regs.fence_action.syncpoint_id, regs.fence_value);
+               syncpoint_manager.GetGuestSyncpointValue(regs.fence_action.syncpoint_id)) {
-        rasterizer->ReleaseFences();
+            rasterizer->ReleaseFences();
+            gpu.Scheduler().Yield();
+        }
         break;
     case Puller::FenceOperation::Increment:
         rasterizer->SignalSyncPoint(regs.fence_action.syncpoint_id);

src/video_core/gpu.cpp

@@ -401,6 +401,14 @@ std::shared_ptr<Control::ChannelState> GPU::AllocateChannel() {
     return impl->AllocateChannel();
 }
 
+Tegra::Control::Scheduler& GPU::Scheduler() {
+    return *impl->scheduler;
+}
+
+const Tegra::Control::Scheduler& GPU::Scheduler() const {
+    return *impl->scheduler;
+}
+
 void GPU::InitChannel(Control::ChannelState& to_init) {
     impl->InitChannel(to_init);
 }

src/video_core/gpu.h

@@ -124,7 +124,8 @@ class KeplerCompute;
 
 namespace Control {
 struct ChannelState;
-}
+class Scheduler;
+} // namespace Control
 
 namespace Host1x {
 class Host1x;
@@ -204,6 +205,12 @@ public:
     /// Returns a const reference to the shader notifier.
     [[nodiscard]] const VideoCore::ShaderNotify& ShaderNotify() const;
 
+    /// Returns GPU Channel Scheduler.
+    [[nodiscard]] Tegra::Control::Scheduler& Scheduler();
+
+    /// Returns GPU Channel Scheduler.
+    [[nodiscard]] const Tegra::Control::Scheduler& Scheduler() const;
+
     [[nodiscard]] u64 GetTicks() const;
 
     [[nodiscard]] bool IsAsync() const;

src/video_core/gpu_thread.cpp

@@ -33,13 +33,15 @@ static void RunThread(std::stop_token stop_token, Core::System& system,
 
     CommandDataContainer next;
 
+    scheduler.Init();
+
     while (!stop_token.stop_requested()) {
         state.queue.PopWait(next, stop_token);
         if (stop_token.stop_requested()) {
             break;
         }
-        if (auto* submit_list = std::get_if<SubmitListCommand>(&next.data)) {
+        if (std::holds_alternative<SubmitListCommand>(next.data)) {
-            scheduler.Push(submit_list->channel, std::move(submit_list->entries));
+            scheduler.Resume();
         } else if (std::holds_alternative<GPUTickCommand>(next.data)) {
             system.GPU().TickWork();
         } else if (const auto* flush = std::get_if<FlushRegionCommand>(&next.data)) {
@@ -66,14 +68,16 @@ ThreadManager::~ThreadManager() = default;
 
 void ThreadManager::StartThread(VideoCore::RendererBase& renderer,
                                 Core::Frontend::GraphicsContext& context,
-                                Tegra::Control::Scheduler& scheduler) {
+                                Tegra::Control::Scheduler& scheduler_) {
     rasterizer = renderer.ReadRasterizer();
+    scheduler = &scheduler_;
     thread = std::jthread(RunThread, std::ref(system), std::ref(renderer), std::ref(context),
-                          std::ref(scheduler), std::ref(state));
+                          std::ref(scheduler_), std::ref(state));
 }
 
 void ThreadManager::SubmitList(s32 channel, Tegra::CommandList&& entries) {
-    PushCommand(SubmitListCommand(channel, std::move(entries)));
+    scheduler->Push(channel, std::move(entries));
+    PushCommand(SubmitListCommand());
 }
 
 void ThreadManager::FlushRegion(DAddr addr, u64 size) {

src/video_core/gpu_thread.h

@@ -36,13 +36,7 @@ class RendererBase;
 namespace VideoCommon::GPUThread {
 
 /// Command to signal to the GPU thread that a command list is ready for processing
-struct SubmitListCommand final {
+struct SubmitListCommand final {};
-    explicit SubmitListCommand(s32 channel_, Tegra::CommandList&& entries_)
-        : channel{channel_}, entries{std::move(entries_)} {}
-
-    s32 channel;
-    Tegra::CommandList entries;
-};
 
 /// Command to signal to the GPU thread to flush a region
 struct FlushRegionCommand final {
@@ -124,6 +118,7 @@ public:
 private:
     /// Pushes a command to be executed by the GPU thread
     u64 PushCommand(CommandData&& command_data, bool block = false);
+    Tegra::Control::Scheduler* scheduler;
 
     Core::System& system;
     const bool is_async;