gstreamer/subprojects/gst-plugins-bad/gst-libs/gst/d3d12/gstd3d12commandqueue.cpp

/* GStreamer
 * Copyright (C) 2023 Seungha Yang <seungha@centricular.com>
 *
 * This library is free software; you cln redistribute it and/or
 * modify it under the terms of the GNU Library General Public
 * License as published by the Free Software Foundation; either
 * version 2 of the License, or (at your option) any later version.
 *
 * This library is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * Library General Public License for more details.
 *
 * You should have received a copy of the GNU Library General Public
 * License along with this library; if not, write to the
 * Free Software Foundation, Inc., 51 Franklin St, Fifth Floor,
 * Boston, MA 02110-1301, USA.
 */

#ifdef HAVE_CONFIG_H
#include "config.h"
#endif

#include "gstd3d12.h"
#include "gstd3d12-private.h"
#include <wrl.h>
#include <queue>
#include <mutex>
#include <memory>
#include <condition_variable>

GST_DEBUG_CATEGORY_STATIC (gst_d3d12_command_queue_debug);
#define GST_CAT_DEFAULT gst_d3d12_command_queue_debug

/* *INDENT-OFF* */
using namespace Microsoft::WRL;

struct GCData
{
  GCData (gpointer user_data, GDestroyNotify destroy_notify,
      guint64 fence_value) : data (user_data),
          notify(destroy_notify), fence_val (fence_value) {}
  ~GCData ()
  {
    if (notify)
      notify (data);
  }

  gpointer data = nullptr;
  GDestroyNotify notify = nullptr;
  guint64 fence_val = 0;
};

typedef std::shared_ptr<GCData> GCDataPtr;

struct gc_cmp {
  bool operator()(const GCDataPtr &a, const GCDataPtr &b)
  {
    return a->fence_val > b->fence_val;
  }
};

struct _GstD3D12CommandQueuePrivate
{
  _GstD3D12CommandQueuePrivate ()
  {
    event_handle = CreateEventEx (nullptr, nullptr, 0, EVENT_ALL_ACCESS);
  }

  ~_GstD3D12CommandQueuePrivate ()
  {
    {
      std::lock_guard <std::mutex> lk (lock);
      shutdown = true;
      cond.notify_one ();
    }

    g_clear_pointer (&gc_thread, g_thread_join);

    auto completed = fence->GetCompletedValue ();
    if (fence_val > completed) {
      auto hr = fence->SetEventOnCompletion (completed, event_handle);
      if (SUCCEEDED (hr))
        WaitForSingleObjectEx (event_handle, INFINITE, FALSE);
    }

    CloseHandle (event_handle);
  }

  D3D12_COMMAND_QUEUE_DESC desc;

  ComPtr<ID3D12Device> device;
  ComPtr<ID3D12CommandQueue> cq;
  ComPtr<ID3D12Fence> fence;
  HANDLE event_handle;
  guint64 fence_val = 0;

  GThread *gc_thread = nullptr;
  std::priority_queue<GCDataPtr, std::vector<GCDataPtr>, gc_cmp> gc_list;

  std::mutex execute_lock;
  std::mutex lock;
  std::condition_variable cond;
  bool shutdown = false;
  size_t queue_size = 0;
};
/* *INDENT-ON* */

static void gst_d3d12_command_queue_finalize (GObject * object);

#define gst_d3d12_command_queue_parent_class parent_class
G_DEFINE_TYPE (GstD3D12CommandQueue, gst_d3d12_command_queue, GST_TYPE_OBJECT);

static void
gst_d3d12_command_queue_class_init (GstD3D12CommandQueueClass * klass)
{
  auto object_class = G_OBJECT_CLASS (klass);

  object_class->finalize = gst_d3d12_command_queue_finalize;
}

static void
gst_d3d12_command_queue_init (GstD3D12CommandQueue * self)
{
  self->priv = new GstD3D12CommandQueuePrivate ();
}

static void
gst_d3d12_command_queue_finalize (GObject * object)
{
  auto self = GST_D3D12_COMMAND_QUEUE (object);

  delete self->priv;

  G_OBJECT_CLASS (parent_class)->finalize (object);
}

/**
 * gst_d3d12_command_queue_new:
 * @device: a #GstD3D12Device
 * @desc: a D3D12_COMMAND_QUEUE_DESC
 * @fence_flags: a D3D12_FENCE_FLAGS
 * @queue_size: command queue size, Sets zero for unlimited queue size
 *
 * Creates GstD3D12CommandQueue with given parameters.
 *
 * Returns: (transfer full): a new #GstD3D12CommandQueue instance
 *
 * Since: 1.26
 */
GstD3D12CommandQueue *
gst_d3d12_command_queue_new (ID3D12Device * device,
    const D3D12_COMMAND_QUEUE_DESC * desc, D3D12_FENCE_FLAGS fence_flags,
    guint queue_size)
{
  g_return_val_if_fail (device, nullptr);
  g_return_val_if_fail (desc, nullptr);

  GST_D3D12_CALL_ONCE_BEGIN {
    GST_DEBUG_CATEGORY_INIT (gst_d3d12_command_queue_debug,
        "d3d12commandqueue", 0, "d3d12commandqueue");
  } GST_D3D12_CALL_ONCE_END;

  ComPtr < ID3D12CommandQueue > cq;
  auto hr = device->CreateCommandQueue (desc, IID_PPV_ARGS (&cq));
  if (FAILED (hr)) {
    GST_ERROR ("Couldn't create command queue, hr: 0x%x", (guint) hr);
    return nullptr;
  }

  ComPtr < ID3D12Fence > fence;
  hr = device->CreateFence (0, fence_flags, IID_PPV_ARGS (&fence));
  if (FAILED (hr)) {
    GST_ERROR ("Couldn't create fence, hr: 0x%x", (guint) hr);
    return nullptr;
  }

  auto self = (GstD3D12CommandQueue *)
      g_object_new (GST_TYPE_D3D12_COMMAND_QUEUE, nullptr);
  gst_object_ref_sink (self);

  auto priv = self->priv;
  priv->device = device;
  priv->cq = cq;
  priv->fence = fence;
  priv->queue_size = queue_size;

  return self;
}

/**
 * gst_d3d12_command_queue_get_handle:
 * @queue: a #GstD3D12CommandQueue
 *
 * Gets command queue handle
 *
 * Return: (transfer none): ID3D12CommandQueue handle
 *
 * Since: 1.26
 */
ID3D12CommandQueue *
gst_d3d12_command_queue_get_handle (GstD3D12CommandQueue * queue)
{
  g_return_val_if_fail (GST_IS_D3D12_COMMAND_QUEUE (queue), nullptr);

  return queue->priv->cq.Get ();
}

/**
 * gst_d3d12_command_queue_get_fence_handle:
 * @queue: a #GstD3D12CommandQueue
 *
 * Gets fence handle handle
 *
 * Return: (transfer none): ID3D12Fence handle
 *
 * Since: 1.26
 */
ID3D12Fence *
gst_d3d12_command_queue_get_fence_handle (GstD3D12CommandQueue * queue)
{
  g_return_val_if_fail (GST_IS_D3D12_COMMAND_QUEUE (queue), nullptr);

  return queue->priv->fence.Get ();
}

/**
 * gst_d3d12_command_queue_execute_command_lists:
 * @queue: a #GstD3D12CommandQueue
 * @num_command_lists: command list size
 * @command_lists: array of ID3D12CommandList
 * @fence_value: (out) (optional): fence value of submitted command
 *
 * Executes command list and signals queue. If @num_command_lists is zero,
 * Only fence signal is executed with fence value increment.
 *
 * Return: HRESULT code
 *
 * Since: 1.26
 */
HRESULT
gst_d3d12_command_queue_execute_command_lists (GstD3D12CommandQueue * queue,
    guint num_command_lists, ID3D12CommandList ** command_lists,
    guint64 * fence_value)
{
  g_return_val_if_fail (GST_IS_D3D12_COMMAND_QUEUE (queue), E_INVALIDARG);

  auto priv = queue->priv;
  HRESULT hr;

  std::lock_guard < std::mutex > lk (priv->execute_lock);
  priv->fence_val++;
  if (num_command_lists)
    priv->cq->ExecuteCommandLists (num_command_lists, command_lists);
  hr = priv->cq->Signal (priv->fence.Get (), priv->fence_val);
  if (FAILED (hr)) {
    GST_ERROR_OBJECT (queue, "Signal failed");
    priv->fence_val--;
  } else if (fence_value) {
    *fence_value = priv->fence_val;
  }

  if (priv->queue_size > 0) {
    auto completed = priv->fence->GetCompletedValue ();
    if (completed + priv->queue_size < priv->fence_val) {
      hr = priv->fence->SetEventOnCompletion (priv->fence_val -
          priv->queue_size, priv->event_handle);
      if (FAILED (hr)) {
        GST_ERROR_OBJECT (queue, "SetEventOnCompletion failed");
        return hr;
      }
      WaitForSingleObjectEx (priv->event_handle, INFINITE, FALSE);
    }
  }

  return hr;
}

/**
 * gst_d3d12_command_queue_execute_wait:
 * @queue: a #GstD3D12CommandQueue
 * @fence: a ID3D12Fence
 * @fence_value: fence value to wait
 *
 * Exectues ID3D12CommandQueue::Wait() operation
 *
 * Return: %TRUE if successful
 *
 * Since: 1.26
 */
HRESULT
gst_d3d12_command_queue_execute_wait (GstD3D12CommandQueue * queue,
    ID3D12Fence * fence, guint64 fence_value)
{
  g_return_val_if_fail (GST_IS_D3D12_COMMAND_QUEUE (queue), E_INVALIDARG);
  g_return_val_if_fail (fence, E_INVALIDARG);

  auto priv = queue->priv;

  return priv->cq->Wait (fence, fence_value);
}

/**
 * gst_d3d12_command_queue_get_completed_value:
 * @queue: a #GstD3D12CommandQueue
 *
 * Gets completed fence value
 *
 * Returns: Completed fence value
 *
 * Since: 1.26
 */
guint64
gst_d3d12_command_queue_get_completed_value (GstD3D12CommandQueue * queue)
{
  g_return_val_if_fail (GST_IS_D3D12_COMMAND_QUEUE (queue), G_MAXUINT64);

  return queue->priv->fence->GetCompletedValue ();
}

/**
 * gst_d3d12_command_queue_fence_wait:
 * @queue: a #GstD3D12CommandQueue
 * @fence_value: fence value to wait
 * @handle: (nullable) (transfer none): event handle used for fence wait
 *
 * Blocks calling CPU thread until command corresponding @fence_value
 * is completed. If @fence_value is %G_MAXUINT64, this method will block
 * calling thread until all pending GPU operations are completed
 *
 * Returns: HRESULT code
 *
 * Since: 1.26
 */
HRESULT
gst_d3d12_command_queue_fence_wait (GstD3D12CommandQueue * queue,
    guint64 fence_value, HANDLE event_handle)
{
  g_return_val_if_fail (GST_IS_D3D12_COMMAND_QUEUE (queue), E_INVALIDARG);

  auto priv = queue->priv;
  guint64 fence_to_wait = fence_value;
  HRESULT hr;
  if (fence_value == G_MAXUINT64) {
    std::lock_guard < std::mutex > lk (priv->execute_lock);
    priv->fence_val++;
    hr = priv->cq->Signal (priv->fence.Get (), priv->fence_val);
    if (FAILED (hr)) {
      GST_ERROR_OBJECT (queue, "Signal failed");
      priv->fence_val--;
      return hr;
    }

    fence_to_wait = priv->fence_val;
  }

  auto completed = priv->fence->GetCompletedValue ();
  if (completed < fence_to_wait) {
    bool close_handle = false;
    if (!event_handle) {
      event_handle = CreateEventEx (nullptr, nullptr, 0, EVENT_ALL_ACCESS);
      close_handle = true;
    }

    hr = priv->fence->SetEventOnCompletion (fence_to_wait, event_handle);
    if (FAILED (hr)) {
      GST_ERROR_OBJECT (queue, "SetEventOnCompletion failed");
      if (close_handle)
        CloseHandle (event_handle);

      return hr;
    }

    WaitForSingleObjectEx (event_handle, INFINITE, FALSE);
    if (close_handle)
      CloseHandle (event_handle);
  }

  return S_OK;
}

static gpointer
gst_d3d12_command_queue_gc_thread (GstD3D12CommandQueue * self)
{
  auto priv = self->priv;

  GST_INFO_OBJECT (self, "Entering GC thread");

  HANDLE event_handle = CreateEventEx (nullptr, nullptr, 0, EVENT_ALL_ACCESS);

  while (true) {
    GCDataPtr gc_data;

    {
      std::unique_lock < std::mutex > lk (priv->lock);
      while (!priv->shutdown && priv->gc_list.empty ())
        priv->cond.wait (lk);

      if (priv->shutdown)
        break;

      auto completed = priv->fence->GetCompletedValue ();
      while (!priv->gc_list.empty ()) {
        auto top = priv->gc_list.top ();
        if (top->fence_val > completed) {
          gc_data = top;
          priv->gc_list.pop ();
          break;
        }

        GST_LOG_OBJECT (self, "Releasing fence data, completed %"
            G_GUINT64_FORMAT ", fence value %" G_GUINT64_FORMAT,
            completed, top->fence_val);

        priv->gc_list.pop ();
      }
    }

    if (gc_data) {
      GST_LOG_OBJECT (self, "Waiting for fence data %" G_GUINT64_FORMAT,
          gc_data->fence_val);
      auto hr =
          priv->fence->SetEventOnCompletion (gc_data->fence_val, event_handle);
      if (FAILED (hr)) {
        GST_ERROR_OBJECT (self, "SetEventOnCompletion failed");
      } else {
        WaitForSingleObjectEx (event_handle, INFINITE, FALSE);
        GST_LOG_OBJECT (self, "Waiting done, %" G_GUINT64_FORMAT,
            gc_data->fence_val);
      }
    }
  }

  GST_INFO_OBJECT (self, "Leaving GC thread");

  CloseHandle (event_handle);

  return nullptr;
}

/**
 * gst_d3d12_command_queue_set_notify:
 * @queue: a #GstD3D12CommandQueue
 * @fence_value: target fence value
 * @fence_data: user data
 * @notify: a #GDestroyNotify
 *
 * Schedules oneshot @notify callback.
 *
 * This method is designed for garbage collection task.
 * Users can construct a storage which holds graphics command associated
 * resources (e.g., command allocator, descriptors, and textures) and pass
 * the storage with destructor, in order to keep resources alive during
 * command execution.
 *
 * GstD3D12CommandQueue launches internal worker thread to monitor fence value
 * and once it reaches the scheduled value, @notify will be called with @fence_data
 *
 * Since: 1.26
 */
void
gst_d3d12_command_queue_set_notify (GstD3D12CommandQueue * queue,
    guint64 fence_value, gpointer fence_data, GDestroyNotify notify)
{
  g_return_if_fail (GST_IS_D3D12_COMMAND_QUEUE (queue));

  auto priv = queue->priv;

  std::lock_guard < std::mutex > elk (priv->execute_lock);
  auto gc_data = std::make_shared < GCData > (fence_data, notify, fence_value);
  if (!priv->gc_thread) {
    priv->gc_thread = g_thread_new ("GstD3D12Gc",
        (GThreadFunc) gst_d3d12_command_queue_gc_thread, queue);
  }

  GST_LOG_OBJECT (queue, "Pushing GC data %" G_GUINT64_FORMAT, fence_value);

  std::lock_guard < std::mutex > lk (priv->lock);
  priv->gc_list.push (std::move (gc_data));
  priv->cond.notify_one ();
}

/**
 * gst_d3d12_command_queue_drain:
 * @queue: a #GstD3D12CommandQueue
 *
 * Waits for all scheduled GPU commands to be finished
 *
 * Returns: HRESULT code
 *
 * Since: 1.26
 */
HRESULT
gst_d3d12_command_queue_drain (GstD3D12CommandQueue * queue)
{
  g_return_val_if_fail (GST_IS_D3D12_COMMAND_QUEUE (queue), E_INVALIDARG);

  auto priv = queue->priv;

  std::priority_queue < GCDataPtr, std::vector < GCDataPtr >, gc_cmp > gc_list;

  {
    std::lock_guard < std::mutex > elk (priv->execute_lock);
    priv->fence_val++;
    auto hr = priv->cq->Signal (priv->fence.Get (), priv->fence_val);
    if (FAILED (hr)) {
      GST_ERROR_OBJECT (queue, "Signal failed");
      priv->fence_val--;
      return hr;
    }

    auto completed = priv->fence->GetCompletedValue ();
    if (completed < priv->fence_val) {
      auto event_handle = CreateEventEx (nullptr, nullptr, 0, EVENT_ALL_ACCESS);
      hr = priv->fence->SetEventOnCompletion (priv->fence_val, event_handle);
      if (FAILED (hr)) {
        GST_ERROR_OBJECT (queue, "SetEventOnCompletion failed");
        CloseHandle (event_handle);
        return hr;
      }

      WaitForSingleObjectEx (event_handle, INFINITE, FALSE);
      CloseHandle (event_handle);
    }

    {
      std::lock_guard < std::mutex > lk (priv->lock);
      gc_list = priv->gc_list;
      priv->gc_list = { };
    }
  }

  return S_OK;
}