nvdec syncpt incorporation

laying the groundwork for async gpu, although this does not fully implement async nvdec operations

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

@@ -11,8 +11,9 @@
 
 namespace Service::Nvidia::Devices {
 
-nvhost_nvdec::nvhost_nvdec(Core::System& system, std::shared_ptr<nvmap> nvmap_dev)
-    : nvhost_nvdec_common(system, std::move(nvmap_dev)) {}
+nvhost_nvdec::nvhost_nvdec(Core::System& system, std::shared_ptr<nvmap> nvmap_dev,
+                           SyncpointManager& syncpoint_manager)
+    : nvhost_nvdec_common(system, std::move(nvmap_dev), syncpoint_manager) {}
 nvhost_nvdec::~nvhost_nvdec() = default;
 
 NvResult nvhost_nvdec::Ioctl1(Ioctl command, const std::vector<u8>& input,

src/core/hle/service/nvdrv/devices/nvhost_nvdec.h

@@ -11,7 +11,8 @@ namespace Service::Nvidia::Devices {
 
 class nvhost_nvdec final : public nvhost_nvdec_common {
 public:
-    explicit nvhost_nvdec(Core::System& system, std::shared_ptr<nvmap> nvmap_dev);
+    explicit nvhost_nvdec(Core::System& system, std::shared_ptr<nvmap> nvmap_dev,
+                          SyncpointManager& syncpoint_manager);
     ~nvhost_nvdec() override;
 
     NvResult Ioctl1(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output) override;

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

@@ -11,6 +11,7 @@
 #include "core/core.h"
 #include "core/hle/service/nvdrv/devices/nvhost_nvdec_common.h"
 #include "core/hle/service/nvdrv/devices/nvmap.h"
+#include "core/hle/service/nvdrv/syncpoint_manager.h"
 #include "core/memory.h"
 #include "video_core/memory_manager.h"
 #include "video_core/renderer_base.h"
@@ -36,8 +37,9 @@ std::size_t WriteVectors(std::vector<u8>& dst, const std::vector<T>& src, std::s
 }
 } // Anonymous namespace
 
-nvhost_nvdec_common::nvhost_nvdec_common(Core::System& system, std::shared_ptr<nvmap> nvmap_dev)
-    : nvdevice(system), nvmap_dev(std::move(nvmap_dev)) {}
+nvhost_nvdec_common::nvhost_nvdec_common(Core::System& system, std::shared_ptr<nvmap> nvmap_dev,
+                                         SyncpointManager& syncpoint_manager)
+    : nvdevice(system), nvmap_dev(std::move(nvmap_dev)), syncpoint_manager(syncpoint_manager) {}
 nvhost_nvdec_common::~nvhost_nvdec_common() = default;
 
 NvResult nvhost_nvdec_common::SetNVMAPfd(const std::vector<u8>& input) {
@@ -71,10 +73,14 @@ NvResult nvhost_nvdec_common::Submit(const std::vector<u8>& input, std::vector<u
     offset = SpliceVectors(input, wait_checks, params.syncpoint_count, offset);
     offset = SpliceVectors(input, fences, params.fence_count, offset);
 
-    // TODO(ameerj): For async gpu, utilize fences for syncpoint 'max' increment
+    for (std::size_t i = 0; i < syncpt_increments.size(); i++) {
+        SyncptIncr syncpt_incr = syncpt_increments[i];
 
+        fences[i].id = syncpt_incr.id;
+        fences[i].value =
+            syncpoint_manager.IncreaseSyncpoint(syncpt_incr.id, syncpt_incr.increments);
+    }
     auto& gpu = system.GPU();
-
     for (const auto& cmd_buffer : command_buffers) {
         auto object = nvmap_dev->GetObject(cmd_buffer.memory_id);
         ASSERT_OR_EXECUTE(object, return NvResult::InvalidState;);
@@ -89,6 +95,10 @@ NvResult nvhost_nvdec_common::Submit(const std::vector<u8>& input, std::vector<u
                                       cmdlist.size() * sizeof(u32));
         gpu.PushCommandBuffer(cmdlist);
     }
+    fences[0].value = syncpoint_manager.IncreaseSyncpoint(fences[0].id, 1);
+
+    Tegra::ChCommandHeaderList cmdlist{{(4 << 28) | fences[0].id}};
+    gpu.PushCommandBuffer(cmdlist);
 
     std::memcpy(output.data(), &params, sizeof(IoctlSubmit));
     // Some games expect command_buffers to be written back
@@ -98,6 +108,7 @@ NvResult nvhost_nvdec_common::Submit(const std::vector<u8>& input, std::vector<u
     offset = WriteVectors(output, reloc_shifts, offset);
     offset = WriteVectors(output, syncpt_increments, offset);
     offset = WriteVectors(output, wait_checks, offset);
+    offset = WriteVectors(output, fences, offset);
 
     return NvResult::Success;
 }
@@ -107,9 +118,10 @@ NvResult nvhost_nvdec_common::GetSyncpoint(const std::vector<u8>& input, std::ve
     std::memcpy(&params, input.data(), sizeof(IoctlGetSyncpoint));
     LOG_DEBUG(Service_NVDRV, "called GetSyncpoint, id={}", params.param);
 
-    // We found that implementing this causes deadlocks with async gpu, along with degraded
-    // performance. TODO: RE the nvdec async implementation
-    params.value = 0;
+    if (device_syncpoints[params.param] == 0) {
+        device_syncpoints[params.param] = syncpoint_manager.AllocateSyncpoint();
+    }
+    params.value = device_syncpoints[params.param];
     std::memcpy(output.data(), &params, sizeof(IoctlGetSyncpoint));
 
     return NvResult::Success;

src/core/hle/service/nvdrv/devices/nvhost_nvdec_common.h

@@ -10,12 +10,16 @@
 #include "common/swap.h"
 #include "core/hle/service/nvdrv/devices/nvdevice.h"
 
-namespace Service::Nvidia::Devices {
+namespace Service::Nvidia {
+class SyncpointManager;
+
+namespace Devices {
 class nvmap;
 
 class nvhost_nvdec_common : public nvdevice {
 public:
-    explicit nvhost_nvdec_common(Core::System& system, std::shared_ptr<nvmap> nvmap_dev);
+    explicit nvhost_nvdec_common(Core::System& system, std::shared_ptr<nvmap> nvmap_dev,
+                                 SyncpointManager& syncpoint_manager);
     ~nvhost_nvdec_common() override;
 
 protected:
@@ -157,8 +161,10 @@ protected:
     s32_le nvmap_fd{};
     u32_le submit_timeout{};
     std::shared_ptr<nvmap> nvmap_dev;
-
+    SyncpointManager& syncpoint_manager;
+    std::array<u32, MaxSyncPoints> device_syncpoints{};
     // This is expected to be ordered, therefore we must use a map, not unordered_map
     std::map<GPUVAddr, BufferMap> buffer_mappings;
 };
-}; // namespace Service::Nvidia::Devices
+}; // namespace Devices
+} // namespace Service::Nvidia

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

@@ -10,8 +10,9 @@
 #include "video_core/renderer_base.h"
 
 namespace Service::Nvidia::Devices {
-nvhost_vic::nvhost_vic(Core::System& system, std::shared_ptr<nvmap> nvmap_dev)
-    : nvhost_nvdec_common(system, std::move(nvmap_dev)) {}
+nvhost_vic::nvhost_vic(Core::System& system, std::shared_ptr<nvmap> nvmap_dev,
+                       SyncpointManager& syncpoint_manager)
+    : nvhost_nvdec_common(system, std::move(nvmap_dev), syncpoint_manager) {}
 
 nvhost_vic::~nvhost_vic() = default;
 

src/core/hle/service/nvdrv/devices/nvhost_vic.h

@@ -7,11 +7,11 @@
 #include "core/hle/service/nvdrv/devices/nvhost_nvdec_common.h"
 
 namespace Service::Nvidia::Devices {
-class nvmap;
 
 class nvhost_vic final : public nvhost_nvdec_common {
 public:
-    explicit nvhost_vic(Core::System& system, std::shared_ptr<nvmap> nvmap_dev);
+    explicit nvhost_vic(Core::System& system, std::shared_ptr<nvmap> nvmap_dev,
+                        SyncpointManager& syncpoint_manager);
     ~nvhost_vic();
 
     NvResult Ioctl1(Ioctl command, const std::vector<u8>& input, std::vector<u8>& output) override;

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

@@ -55,9 +55,11 @@ Module::Module(Core::System& system) : syncpoint_manager{system.GPU()} {
     devices["/dev/nvdisp_disp0"] = std::make_shared<Devices::nvdisp_disp0>(system, nvmap_dev);
     devices["/dev/nvhost-ctrl"] =
         std::make_shared<Devices::nvhost_ctrl>(system, events_interface, syncpoint_manager);
-    devices["/dev/nvhost-nvdec"] = std::make_shared<Devices::nvhost_nvdec>(system, nvmap_dev);
+    devices["/dev/nvhost-nvdec"] =
+        std::make_shared<Devices::nvhost_nvdec>(system, nvmap_dev, syncpoint_manager);
     devices["/dev/nvhost-nvjpg"] = std::make_shared<Devices::nvhost_nvjpg>(system);
-    devices["/dev/nvhost-vic"] = std::make_shared<Devices::nvhost_vic>(system, nvmap_dev);
+    devices["/dev/nvhost-vic"] =
+        std::make_shared<Devices::nvhost_vic>(system, nvmap_dev, syncpoint_manager);
 }
 
 Module::~Module() = default;

src/video_core/cdma_pusher.cpp

@@ -33,8 +33,7 @@ CDmaPusher::CDmaPusher(GPU& gpu_)
     : gpu{gpu_}, nvdec_processor(std::make_shared<Nvdec>(gpu)),
       vic_processor(std::make_unique<Vic>(gpu, nvdec_processor)),
       host1x_processor(std::make_unique<Host1x>(gpu)),
-      nvdec_sync(std::make_unique<SyncptIncrManager>(gpu)),
-      vic_sync(std::make_unique<SyncptIncrManager>(gpu)) {}
+      sync_manager(std::make_unique<SyncptIncrManager>(gpu)) {}
 
 CDmaPusher::~CDmaPusher() = default;
 
@@ -110,10 +109,10 @@ void CDmaPusher::ExecuteCommand(u32 state_offset, u32 data) {
             const auto syncpoint_id = static_cast<u32>(data & 0xFF);
             const auto cond = static_cast<u32>((data >> 8) & 0xFF);
             if (cond == 0) {
-                nvdec_sync->Increment(syncpoint_id);
+                sync_manager->Increment(syncpoint_id);
             } else {
-                nvdec_sync->IncrementWhenDone(static_cast<u32>(current_class), syncpoint_id);
-                nvdec_sync->SignalDone(syncpoint_id);
+                sync_manager->SignalDone(
+                    sync_manager->IncrementWhenDone(static_cast<u32>(current_class), syncpoint_id));
             }
             break;
         }
@@ -135,10 +134,10 @@ void CDmaPusher::ExecuteCommand(u32 state_offset, u32 data) {
             const auto syncpoint_id = static_cast<u32>(data & 0xFF);
             const auto cond = static_cast<u32>((data >> 8) & 0xFF);
             if (cond == 0) {
-                vic_sync->Increment(syncpoint_id);
+                sync_manager->Increment(syncpoint_id);
             } else {
-                vic_sync->IncrementWhenDone(static_cast<u32>(current_class), syncpoint_id);
-                vic_sync->SignalDone(syncpoint_id);
+                sync_manager->SignalDone(
+                    sync_manager->IncrementWhenDone(static_cast<u32>(current_class), syncpoint_id));
             }
             break;
         }

src/video_core/cdma_pusher.h

@@ -116,12 +116,10 @@ private:
     void ThiStateWrite(ThiRegisters& state, u32 state_offset, const std::vector<u32>& arguments);
 
     GPU& gpu;
-
-    std::shared_ptr<Nvdec> nvdec_processor;
-    std::unique_ptr<Vic> vic_processor;
-    std::unique_ptr<Host1x> host1x_processor;
-    std::unique_ptr<SyncptIncrManager> nvdec_sync;
-    std::unique_ptr<SyncptIncrManager> vic_sync;
+    std::shared_ptr<Tegra::Nvdec> nvdec_processor;
+    std::unique_ptr<Tegra::Vic> vic_processor;
+    std::unique_ptr<Tegra::Host1x> host1x_processor;
+    std::unique_ptr<SyncptIncrManager> sync_manager;
     ChClassId current_class{};
     ThiRegisters vic_thi_state{};
     ThiRegisters nvdec_thi_state{};

src/video_core/command_classes/host1x.cpp

@@ -34,6 +34,8 @@ void Tegra::Host1x::ProcessMethod(Method method, const std::vector<u32>& argumen
 }
 
 void Tegra::Host1x::Execute(u32 data) {
-    // This method waits on a valid syncpoint.
-    // TODO: Implement when proper Async is in place
+    u32 syncpointId = (data & 0xFF);
+    u32 threshold = state.load_syncpoint_payload32;
+
+    gpu.WaitFence(syncpointId, threshold);
 }

src/video_core/command_classes/sync_manager.cpp

@@ -38,7 +38,7 @@ u32 SyncptIncrManager::IncrementWhenDone(u32 class_id, u32 id) {
 }
 
 void SyncptIncrManager::SignalDone(u32 handle) {
-    const auto done_incr =
+    const auto& done_incr =
         std::find_if(increments.begin(), increments.end(),
                      [handle](const SyncptIncr& incr) { return incr.id == handle; });
     if (done_incr != increments.cend()) {