6 files changed, 39 insertions, 18 deletions

diff --git a/src/core/hle/service/nvdrv/devices/nvhost_nvdec.cpp b/src/core/hle/service/nvdrv/devices/nvhost_nvdec.cpp
index d8735491c..36970f828 100644
--- a/src/core/hle/service/nvdrv/devices/nvhost_nvdec.cpp
+++ b/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,
diff --git a/src/core/hle/service/nvdrv/devices/nvhost_nvdec.h b/src/core/hle/service/nvdrv/devices/nvhost_nvdec.h
index 79b8b6de1..77ef53cdd 100644
--- a/src/core/hle/service/nvdrv/devices/nvhost_nvdec.h
+++ b/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;
diff --git a/src/core/hle/service/nvdrv/devices/nvhost_nvdec_common.cpp b/src/core/hle/service/nvdrv/devices/nvhost_nvdec_common.cpp
index b49cecb42..64370ad4c 100644
--- a/src/core/hle/service/nvdrv/devices/nvhost_nvdec_common.cpp
+++ b/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;
diff --git a/src/core/hle/service/nvdrv/devices/nvhost_nvdec_common.h b/src/core/hle/service/nvdrv/devices/nvhost_nvdec_common.h
index d9f95ba58..4c9d4ba41 100644
--- a/src/core/hle/service/nvdrv/devices/nvhost_nvdec_common.h
+++ b/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
diff --git a/src/core/hle/service/nvdrv/devices/nvhost_vic.cpp b/src/core/hle/service/nvdrv/devices/nvhost_vic.cpp
index 805fe86ae..72499654c 100644
--- a/src/core/hle/service/nvdrv/devices/nvhost_vic.cpp
+++ b/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;
 
diff --git a/src/core/hle/service/nvdrv/devices/nvhost_vic.h b/src/core/hle/service/nvdrv/devices/nvhost_vic.h
index b2e11f4d4..f401c61fa 100644
--- a/src/core/hle/service/nvdrv/devices/nvhost_vic.h
+++ b/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;